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,const struct load_weight * lw)3794 void reweight_task(struct task_struct *p, const struct load_weight *lw)
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
3800 reweight_entity(cfs_rq, se, lw->weight);
3801 load->inv_weight = lw->inv_weight;
3802 }
3803
3804 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3805
3806 #ifdef CONFIG_FAIR_GROUP_SCHED
3807 #ifdef CONFIG_SMP
3808 /*
3809 * All this does is approximate the hierarchical proportion which includes that
3810 * global sum we all love to hate.
3811 *
3812 * That is, the weight of a group entity, is the proportional share of the
3813 * group weight based on the group runqueue weights. That is:
3814 *
3815 * tg->weight * grq->load.weight
3816 * ge->load.weight = ----------------------------- (1)
3817 * \Sum grq->load.weight
3818 *
3819 * Now, because computing that sum is prohibitively expensive to compute (been
3820 * there, done that) we approximate it with this average stuff. The average
3821 * moves slower and therefore the approximation is cheaper and more stable.
3822 *
3823 * So instead of the above, we substitute:
3824 *
3825 * grq->load.weight -> grq->avg.load_avg (2)
3826 *
3827 * which yields the following:
3828 *
3829 * tg->weight * grq->avg.load_avg
3830 * ge->load.weight = ------------------------------ (3)
3831 * tg->load_avg
3832 *
3833 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3834 *
3835 * That is shares_avg, and it is right (given the approximation (2)).
3836 *
3837 * The problem with it is that because the average is slow -- it was designed
3838 * to be exactly that of course -- this leads to transients in boundary
3839 * conditions. In specific, the case where the group was idle and we start the
3840 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3841 * yielding bad latency etc..
3842 *
3843 * Now, in that special case (1) reduces to:
3844 *
3845 * tg->weight * grq->load.weight
3846 * ge->load.weight = ----------------------------- = tg->weight (4)
3847 * grp->load.weight
3848 *
3849 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3850 *
3851 * So what we do is modify our approximation (3) to approach (4) in the (near)
3852 * UP case, like:
3853 *
3854 * ge->load.weight =
3855 *
3856 * tg->weight * grq->load.weight
3857 * --------------------------------------------------- (5)
3858 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3859 *
3860 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3861 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3862 *
3863 *
3864 * tg->weight * grq->load.weight
3865 * ge->load.weight = ----------------------------- (6)
3866 * tg_load_avg'
3867 *
3868 * Where:
3869 *
3870 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3871 * max(grq->load.weight, grq->avg.load_avg)
3872 *
3873 * And that is shares_weight and is icky. In the (near) UP case it approaches
3874 * (4) while in the normal case it approaches (3). It consistently
3875 * overestimates the ge->load.weight and therefore:
3876 *
3877 * \Sum ge->load.weight >= tg->weight
3878 *
3879 * hence icky!
3880 */
calc_group_shares(struct cfs_rq * cfs_rq)3881 static long calc_group_shares(struct cfs_rq *cfs_rq)
3882 {
3883 long tg_weight, tg_shares, load, shares;
3884 struct task_group *tg = cfs_rq->tg;
3885
3886 tg_shares = READ_ONCE(tg->shares);
3887
3888 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3889
3890 tg_weight = atomic_long_read(&tg->load_avg);
3891
3892 /* Ensure tg_weight >= load */
3893 tg_weight -= cfs_rq->tg_load_avg_contrib;
3894 tg_weight += load;
3895
3896 shares = (tg_shares * load);
3897 if (tg_weight)
3898 shares /= tg_weight;
3899
3900 /*
3901 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3902 * of a group with small tg->shares value. It is a floor value which is
3903 * assigned as a minimum load.weight to the sched_entity representing
3904 * the group on a CPU.
3905 *
3906 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3907 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3908 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3909 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3910 * instead of 0.
3911 */
3912 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3913 }
3914 #endif /* CONFIG_SMP */
3915
3916 /*
3917 * Recomputes the group entity based on the current state of its group
3918 * runqueue.
3919 */
update_cfs_group(struct sched_entity * se)3920 static void update_cfs_group(struct sched_entity *se)
3921 {
3922 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3923 long shares;
3924
3925 if (!gcfs_rq)
3926 return;
3927
3928 if (throttled_hierarchy(gcfs_rq))
3929 return;
3930
3931 #ifndef CONFIG_SMP
3932 shares = READ_ONCE(gcfs_rq->tg->shares);
3933 #else
3934 shares = calc_group_shares(gcfs_rq);
3935 #endif
3936 if (unlikely(se->load.weight != shares))
3937 reweight_entity(cfs_rq_of(se), se, shares);
3938 }
3939
3940 #else /* CONFIG_FAIR_GROUP_SCHED */
update_cfs_group(struct sched_entity * se)3941 static inline void update_cfs_group(struct sched_entity *se)
3942 {
3943 }
3944 #endif /* CONFIG_FAIR_GROUP_SCHED */
3945
cfs_rq_util_change(struct cfs_rq * cfs_rq,int flags)3946 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3947 {
3948 struct rq *rq = rq_of(cfs_rq);
3949
3950 if (&rq->cfs == cfs_rq) {
3951 /*
3952 * There are a few boundary cases this might miss but it should
3953 * get called often enough that that should (hopefully) not be
3954 * a real problem.
3955 *
3956 * It will not get called when we go idle, because the idle
3957 * thread is a different class (!fair), nor will the utilization
3958 * number include things like RT tasks.
3959 *
3960 * As is, the util number is not freq-invariant (we'd have to
3961 * implement arch_scale_freq_capacity() for that).
3962 *
3963 * See cpu_util_cfs().
3964 */
3965 cpufreq_update_util(rq, flags);
3966 }
3967 }
3968
3969 #ifdef CONFIG_SMP
load_avg_is_decayed(struct sched_avg * sa)3970 static inline bool load_avg_is_decayed(struct sched_avg *sa)
3971 {
3972 if (sa->load_sum)
3973 return false;
3974
3975 if (sa->util_sum)
3976 return false;
3977
3978 if (sa->runnable_sum)
3979 return false;
3980
3981 /*
3982 * _avg must be null when _sum are null because _avg = _sum / divider
3983 * Make sure that rounding and/or propagation of PELT values never
3984 * break this.
3985 */
3986 SCHED_WARN_ON(sa->load_avg ||
3987 sa->util_avg ||
3988 sa->runnable_avg);
3989
3990 return true;
3991 }
3992
cfs_rq_last_update_time(struct cfs_rq * cfs_rq)3993 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3994 {
3995 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
3996 cfs_rq->last_update_time_copy);
3997 }
3998 #ifdef CONFIG_FAIR_GROUP_SCHED
3999 /*
4000 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
4001 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
4002 * bottom-up, we only have to test whether the cfs_rq before us on the list
4003 * is our child.
4004 * If cfs_rq is not on the list, test whether a child needs its to be added to
4005 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
4006 */
child_cfs_rq_on_list(struct cfs_rq * cfs_rq)4007 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
4008 {
4009 struct cfs_rq *prev_cfs_rq;
4010 struct list_head *prev;
4011
4012 if (cfs_rq->on_list) {
4013 prev = cfs_rq->leaf_cfs_rq_list.prev;
4014 } else {
4015 struct rq *rq = rq_of(cfs_rq);
4016
4017 prev = rq->tmp_alone_branch;
4018 }
4019
4020 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
4021
4022 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
4023 }
4024
cfs_rq_is_decayed(struct cfs_rq * cfs_rq)4025 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4026 {
4027 if (cfs_rq->load.weight)
4028 return false;
4029
4030 if (!load_avg_is_decayed(&cfs_rq->avg))
4031 return false;
4032
4033 if (child_cfs_rq_on_list(cfs_rq))
4034 return false;
4035
4036 return true;
4037 }
4038
4039 /**
4040 * update_tg_load_avg - update the tg's load avg
4041 * @cfs_rq: the cfs_rq whose avg changed
4042 *
4043 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
4044 * However, because tg->load_avg is a global value there are performance
4045 * considerations.
4046 *
4047 * In order to avoid having to look at the other cfs_rq's, we use a
4048 * differential update where we store the last value we propagated. This in
4049 * turn allows skipping updates if the differential is 'small'.
4050 *
4051 * Updating tg's load_avg is necessary before update_cfs_share().
4052 */
update_tg_load_avg(struct cfs_rq * cfs_rq)4053 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
4054 {
4055 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
4056
4057 /*
4058 * No need to update load_avg for root_task_group as it is not used.
4059 */
4060 if (cfs_rq->tg == &root_task_group)
4061 return;
4062
4063 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
4064 atomic_long_add(delta, &cfs_rq->tg->load_avg);
4065 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
4066 }
4067 }
4068
4069 /*
4070 * Called within set_task_rq() right before setting a task's CPU. The
4071 * caller only guarantees p->pi_lock is held; no other assumptions,
4072 * including the state of rq->lock, should be made.
4073 */
set_task_rq_fair(struct sched_entity * se,struct cfs_rq * prev,struct cfs_rq * next)4074 void set_task_rq_fair(struct sched_entity *se,
4075 struct cfs_rq *prev, struct cfs_rq *next)
4076 {
4077 u64 p_last_update_time;
4078 u64 n_last_update_time;
4079
4080 if (!sched_feat(ATTACH_AGE_LOAD))
4081 return;
4082
4083 /*
4084 * We are supposed to update the task to "current" time, then its up to
4085 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
4086 * getting what current time is, so simply throw away the out-of-date
4087 * time. This will result in the wakee task is less decayed, but giving
4088 * the wakee more load sounds not bad.
4089 */
4090 if (!(se->avg.last_update_time && prev))
4091 return;
4092
4093 p_last_update_time = cfs_rq_last_update_time(prev);
4094 n_last_update_time = cfs_rq_last_update_time(next);
4095
4096 __update_load_avg_blocked_se(p_last_update_time, se);
4097 se->avg.last_update_time = n_last_update_time;
4098 }
4099
4100 /*
4101 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
4102 * propagate its contribution. The key to this propagation is the invariant
4103 * that for each group:
4104 *
4105 * ge->avg == grq->avg (1)
4106 *
4107 * _IFF_ we look at the pure running and runnable sums. Because they
4108 * represent the very same entity, just at different points in the hierarchy.
4109 *
4110 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
4111 * and simply copies the running/runnable sum over (but still wrong, because
4112 * the group entity and group rq do not have their PELT windows aligned).
4113 *
4114 * However, update_tg_cfs_load() is more complex. So we have:
4115 *
4116 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
4117 *
4118 * And since, like util, the runnable part should be directly transferable,
4119 * the following would _appear_ to be the straight forward approach:
4120 *
4121 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
4122 *
4123 * And per (1) we have:
4124 *
4125 * ge->avg.runnable_avg == grq->avg.runnable_avg
4126 *
4127 * Which gives:
4128 *
4129 * ge->load.weight * grq->avg.load_avg
4130 * ge->avg.load_avg = ----------------------------------- (4)
4131 * grq->load.weight
4132 *
4133 * Except that is wrong!
4134 *
4135 * Because while for entities historical weight is not important and we
4136 * really only care about our future and therefore can consider a pure
4137 * runnable sum, runqueues can NOT do this.
4138 *
4139 * We specifically want runqueues to have a load_avg that includes
4140 * historical weights. Those represent the blocked load, the load we expect
4141 * to (shortly) return to us. This only works by keeping the weights as
4142 * integral part of the sum. We therefore cannot decompose as per (3).
4143 *
4144 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4145 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
4146 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
4147 * runnable section of these tasks overlap (or not). If they were to perfectly
4148 * align the rq as a whole would be runnable 2/3 of the time. If however we
4149 * always have at least 1 runnable task, the rq as a whole is always runnable.
4150 *
4151 * So we'll have to approximate.. :/
4152 *
4153 * Given the constraint:
4154 *
4155 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4156 *
4157 * We can construct a rule that adds runnable to a rq by assuming minimal
4158 * overlap.
4159 *
4160 * On removal, we'll assume each task is equally runnable; which yields:
4161 *
4162 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4163 *
4164 * XXX: only do this for the part of runnable > running ?
4165 *
4166 */
4167 static inline void
update_tg_cfs_util(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4168 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4169 {
4170 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
4171 u32 new_sum, divider;
4172
4173 /* Nothing to update */
4174 if (!delta_avg)
4175 return;
4176
4177 /*
4178 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4179 * See ___update_load_avg() for details.
4180 */
4181 divider = get_pelt_divider(&cfs_rq->avg);
4182
4183
4184 /* Set new sched_entity's utilization */
4185 se->avg.util_avg = gcfs_rq->avg.util_avg;
4186 new_sum = se->avg.util_avg * divider;
4187 delta_sum = (long)new_sum - (long)se->avg.util_sum;
4188 se->avg.util_sum = new_sum;
4189
4190 /* Update parent cfs_rq utilization */
4191 add_positive(&cfs_rq->avg.util_avg, delta_avg);
4192 add_positive(&cfs_rq->avg.util_sum, delta_sum);
4193
4194 /* See update_cfs_rq_load_avg() */
4195 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4196 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4197 }
4198
4199 static inline void
update_tg_cfs_runnable(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4200 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4201 {
4202 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
4203 u32 new_sum, divider;
4204
4205 /* Nothing to update */
4206 if (!delta_avg)
4207 return;
4208
4209 /*
4210 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4211 * See ___update_load_avg() for details.
4212 */
4213 divider = get_pelt_divider(&cfs_rq->avg);
4214
4215 /* Set new sched_entity's runnable */
4216 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
4217 new_sum = se->avg.runnable_avg * divider;
4218 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
4219 se->avg.runnable_sum = new_sum;
4220
4221 /* Update parent cfs_rq runnable */
4222 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
4223 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
4224 /* See update_cfs_rq_load_avg() */
4225 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4226 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4227 }
4228
4229 static inline void
update_tg_cfs_load(struct cfs_rq * cfs_rq,struct sched_entity * se,struct cfs_rq * gcfs_rq)4230 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4231 {
4232 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
4233 unsigned long load_avg;
4234 u64 load_sum = 0;
4235 s64 delta_sum;
4236 u32 divider;
4237
4238 if (!runnable_sum)
4239 return;
4240
4241 gcfs_rq->prop_runnable_sum = 0;
4242
4243 /*
4244 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4245 * See ___update_load_avg() for details.
4246 */
4247 divider = get_pelt_divider(&cfs_rq->avg);
4248
4249 if (runnable_sum >= 0) {
4250 /*
4251 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
4252 * the CPU is saturated running == runnable.
4253 */
4254 runnable_sum += se->avg.load_sum;
4255 runnable_sum = min_t(long, runnable_sum, divider);
4256 } else {
4257 /*
4258 * Estimate the new unweighted runnable_sum of the gcfs_rq by
4259 * assuming all tasks are equally runnable.
4260 */
4261 if (scale_load_down(gcfs_rq->load.weight)) {
4262 load_sum = div_u64(gcfs_rq->avg.load_sum,
4263 scale_load_down(gcfs_rq->load.weight));
4264 }
4265
4266 /* But make sure to not inflate se's runnable */
4267 runnable_sum = min(se->avg.load_sum, load_sum);
4268 }
4269
4270 /*
4271 * runnable_sum can't be lower than running_sum
4272 * Rescale running sum to be in the same range as runnable sum
4273 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
4274 * runnable_sum is in [0 : LOAD_AVG_MAX]
4275 */
4276 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
4277 runnable_sum = max(runnable_sum, running_sum);
4278
4279 load_sum = se_weight(se) * runnable_sum;
4280 load_avg = div_u64(load_sum, divider);
4281
4282 delta_avg = load_avg - se->avg.load_avg;
4283 if (!delta_avg)
4284 return;
4285
4286 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
4287
4288 se->avg.load_sum = runnable_sum;
4289 se->avg.load_avg = load_avg;
4290 add_positive(&cfs_rq->avg.load_avg, delta_avg);
4291 add_positive(&cfs_rq->avg.load_sum, delta_sum);
4292 /* See update_cfs_rq_load_avg() */
4293 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
4294 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
4295 }
4296
add_tg_cfs_propagate(struct cfs_rq * cfs_rq,long runnable_sum)4297 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
4298 {
4299 cfs_rq->propagate = 1;
4300 cfs_rq->prop_runnable_sum += runnable_sum;
4301 }
4302
4303 /* Update task and its cfs_rq load average */
propagate_entity_load_avg(struct sched_entity * se)4304 static inline int propagate_entity_load_avg(struct sched_entity *se)
4305 {
4306 struct cfs_rq *cfs_rq, *gcfs_rq;
4307
4308 if (entity_is_task(se))
4309 return 0;
4310
4311 gcfs_rq = group_cfs_rq(se);
4312 if (!gcfs_rq->propagate)
4313 return 0;
4314
4315 gcfs_rq->propagate = 0;
4316
4317 cfs_rq = cfs_rq_of(se);
4318
4319 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
4320
4321 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
4322 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
4323 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
4324
4325 trace_pelt_cfs_tp(cfs_rq);
4326 trace_pelt_se_tp(se);
4327
4328 return 1;
4329 }
4330
4331 /*
4332 * Check if we need to update the load and the utilization of a blocked
4333 * group_entity:
4334 */
skip_blocked_update(struct sched_entity * se)4335 static inline bool skip_blocked_update(struct sched_entity *se)
4336 {
4337 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
4338
4339 /*
4340 * If sched_entity still have not zero load or utilization, we have to
4341 * decay it:
4342 */
4343 if (se->avg.load_avg || se->avg.util_avg)
4344 return false;
4345
4346 /*
4347 * If there is a pending propagation, we have to update the load and
4348 * the utilization of the sched_entity:
4349 */
4350 if (gcfs_rq->propagate)
4351 return false;
4352
4353 /*
4354 * Otherwise, the load and the utilization of the sched_entity is
4355 * already zero and there is no pending propagation, so it will be a
4356 * waste of time to try to decay it:
4357 */
4358 return true;
4359 }
4360
4361 #else /* CONFIG_FAIR_GROUP_SCHED */
4362
update_tg_load_avg(struct cfs_rq * cfs_rq)4363 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
4364
propagate_entity_load_avg(struct sched_entity * se)4365 static inline int propagate_entity_load_avg(struct sched_entity *se)
4366 {
4367 return 0;
4368 }
4369
add_tg_cfs_propagate(struct cfs_rq * cfs_rq,long runnable_sum)4370 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
4371
4372 #endif /* CONFIG_FAIR_GROUP_SCHED */
4373
4374 #ifdef CONFIG_NO_HZ_COMMON
migrate_se_pelt_lag(struct sched_entity * se)4375 static inline void migrate_se_pelt_lag(struct sched_entity *se)
4376 {
4377 u64 throttled = 0, now, lut;
4378 struct cfs_rq *cfs_rq;
4379 struct rq *rq;
4380 bool is_idle;
4381
4382 if (load_avg_is_decayed(&se->avg))
4383 return;
4384
4385 cfs_rq = cfs_rq_of(se);
4386 rq = rq_of(cfs_rq);
4387
4388 rcu_read_lock();
4389 is_idle = is_idle_task(rcu_dereference(rq->curr));
4390 rcu_read_unlock();
4391
4392 /*
4393 * The lag estimation comes with a cost we don't want to pay all the
4394 * time. Hence, limiting to the case where the source CPU is idle and
4395 * we know we are at the greatest risk to have an outdated clock.
4396 */
4397 if (!is_idle)
4398 return;
4399
4400 /*
4401 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
4402 *
4403 * last_update_time (the cfs_rq's last_update_time)
4404 * = cfs_rq_clock_pelt()@cfs_rq_idle
4405 * = rq_clock_pelt()@cfs_rq_idle
4406 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
4407 *
4408 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
4409 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
4410 *
4411 * rq_idle_lag (delta between now and rq's update)
4412 * = sched_clock_cpu() - rq_clock()@rq_idle
4413 *
4414 * We can then write:
4415 *
4416 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
4417 * sched_clock_cpu() - rq_clock()@rq_idle
4418 * Where:
4419 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
4420 * rq_clock()@rq_idle is rq->clock_idle
4421 * cfs->throttled_clock_pelt_time@cfs_rq_idle
4422 * is cfs_rq->throttled_pelt_idle
4423 */
4424
4425 #ifdef CONFIG_CFS_BANDWIDTH
4426 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
4427 /* The clock has been stopped for throttling */
4428 if (throttled == U64_MAX)
4429 return;
4430 #endif
4431 now = u64_u32_load(rq->clock_pelt_idle);
4432 /*
4433 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
4434 * is observed the old clock_pelt_idle value and the new clock_idle,
4435 * which lead to an underestimation. The opposite would lead to an
4436 * overestimation.
4437 */
4438 smp_rmb();
4439 lut = cfs_rq_last_update_time(cfs_rq);
4440
4441 now -= throttled;
4442 if (now < lut)
4443 /*
4444 * cfs_rq->avg.last_update_time is more recent than our
4445 * estimation, let's use it.
4446 */
4447 now = lut;
4448 else
4449 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4450
4451 __update_load_avg_blocked_se(now, se);
4452 }
4453 #else
migrate_se_pelt_lag(struct sched_entity * se)4454 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4455 #endif
4456
4457 /**
4458 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4459 * @now: current time, as per cfs_rq_clock_pelt()
4460 * @cfs_rq: cfs_rq to update
4461 *
4462 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4463 * avg. The immediate corollary is that all (fair) tasks must be attached.
4464 *
4465 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4466 *
4467 * Return: true if the load decayed or we removed load.
4468 *
4469 * Since both these conditions indicate a changed cfs_rq->avg.load we should
4470 * call update_tg_load_avg() when this function returns true.
4471 */
4472 static inline int
update_cfs_rq_load_avg(u64 now,struct cfs_rq * cfs_rq)4473 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4474 {
4475 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4476 struct sched_avg *sa = &cfs_rq->avg;
4477 int decayed = 0;
4478
4479 if (cfs_rq->removed.nr) {
4480 unsigned long r;
4481 u32 divider = get_pelt_divider(&cfs_rq->avg);
4482
4483 raw_spin_lock(&cfs_rq->removed.lock);
4484 swap(cfs_rq->removed.util_avg, removed_util);
4485 swap(cfs_rq->removed.load_avg, removed_load);
4486 swap(cfs_rq->removed.runnable_avg, removed_runnable);
4487 cfs_rq->removed.nr = 0;
4488 raw_spin_unlock(&cfs_rq->removed.lock);
4489
4490 r = removed_load;
4491 sub_positive(&sa->load_avg, r);
4492 sub_positive(&sa->load_sum, r * divider);
4493 /* See sa->util_sum below */
4494 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4495
4496 r = removed_util;
4497 sub_positive(&sa->util_avg, r);
4498 sub_positive(&sa->util_sum, r * divider);
4499 /*
4500 * Because of rounding, se->util_sum might ends up being +1 more than
4501 * cfs->util_sum. Although this is not a problem by itself, detaching
4502 * a lot of tasks with the rounding problem between 2 updates of
4503 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4504 * cfs_util_avg is not.
4505 * Check that util_sum is still above its lower bound for the new
4506 * util_avg. Given that period_contrib might have moved since the last
4507 * sync, we are only sure that util_sum must be above or equal to
4508 * util_avg * minimum possible divider
4509 */
4510 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4511
4512 r = removed_runnable;
4513 sub_positive(&sa->runnable_avg, r);
4514 sub_positive(&sa->runnable_sum, r * divider);
4515 /* See sa->util_sum above */
4516 sa->runnable_sum = max_t(u32, sa->runnable_sum,
4517 sa->runnable_avg * PELT_MIN_DIVIDER);
4518
4519 /*
4520 * removed_runnable is the unweighted version of removed_load so we
4521 * can use it to estimate removed_load_sum.
4522 */
4523 add_tg_cfs_propagate(cfs_rq,
4524 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4525
4526 decayed = 1;
4527 }
4528
4529 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4530 u64_u32_store_copy(sa->last_update_time,
4531 cfs_rq->last_update_time_copy,
4532 sa->last_update_time);
4533 return decayed;
4534 }
4535
4536 /**
4537 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4538 * @cfs_rq: cfs_rq to attach to
4539 * @se: sched_entity to attach
4540 *
4541 * Must call update_cfs_rq_load_avg() before this, since we rely on
4542 * cfs_rq->avg.last_update_time being current.
4543 */
attach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)4544 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4545 {
4546 /*
4547 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4548 * See ___update_load_avg() for details.
4549 */
4550 u32 divider = get_pelt_divider(&cfs_rq->avg);
4551
4552 /*
4553 * When we attach the @se to the @cfs_rq, we must align the decay
4554 * window because without that, really weird and wonderful things can
4555 * happen.
4556 *
4557 * XXX illustrate
4558 */
4559 se->avg.last_update_time = cfs_rq->avg.last_update_time;
4560 se->avg.period_contrib = cfs_rq->avg.period_contrib;
4561
4562 /*
4563 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4564 * period_contrib. This isn't strictly correct, but since we're
4565 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4566 * _sum a little.
4567 */
4568 se->avg.util_sum = se->avg.util_avg * divider;
4569
4570 se->avg.runnable_sum = se->avg.runnable_avg * divider;
4571
4572 se->avg.load_sum = se->avg.load_avg * divider;
4573 if (se_weight(se) < se->avg.load_sum)
4574 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4575 else
4576 se->avg.load_sum = 1;
4577
4578 enqueue_load_avg(cfs_rq, se);
4579 cfs_rq->avg.util_avg += se->avg.util_avg;
4580 cfs_rq->avg.util_sum += se->avg.util_sum;
4581 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4582 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4583
4584 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4585
4586 cfs_rq_util_change(cfs_rq, 0);
4587
4588 trace_pelt_cfs_tp(cfs_rq);
4589 }
4590
4591 /**
4592 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4593 * @cfs_rq: cfs_rq to detach from
4594 * @se: sched_entity to detach
4595 *
4596 * Must call update_cfs_rq_load_avg() before this, since we rely on
4597 * cfs_rq->avg.last_update_time being current.
4598 */
detach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)4599 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4600 {
4601 dequeue_load_avg(cfs_rq, se);
4602 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4603 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4604 /* See update_cfs_rq_load_avg() */
4605 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4606 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4607
4608 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4609 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4610 /* See update_cfs_rq_load_avg() */
4611 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4612 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4613
4614 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4615
4616 cfs_rq_util_change(cfs_rq, 0);
4617
4618 trace_pelt_cfs_tp(cfs_rq);
4619 }
4620
4621 /*
4622 * Optional action to be done while updating the load average
4623 */
4624 #define UPDATE_TG 0x1
4625 #define SKIP_AGE_LOAD 0x2
4626 #define DO_ATTACH 0x4
4627 #define DO_DETACH 0x8
4628
4629 /* Update task and its cfs_rq load average */
update_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)4630 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4631 {
4632 u64 now = cfs_rq_clock_pelt(cfs_rq);
4633 int decayed;
4634
4635 /*
4636 * Track task load average for carrying it to new CPU after migrated, and
4637 * track group sched_entity load average for task_h_load calc in migration
4638 */
4639 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4640 __update_load_avg_se(now, cfs_rq, se);
4641
4642 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4643 decayed |= propagate_entity_load_avg(se);
4644
4645 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4646
4647 /*
4648 * DO_ATTACH means we're here from enqueue_entity().
4649 * !last_update_time means we've passed through
4650 * migrate_task_rq_fair() indicating we migrated.
4651 *
4652 * IOW we're enqueueing a task on a new CPU.
4653 */
4654 attach_entity_load_avg(cfs_rq, se);
4655 update_tg_load_avg(cfs_rq);
4656
4657 } else if (flags & DO_DETACH) {
4658 /*
4659 * DO_DETACH means we're here from dequeue_entity()
4660 * and we are migrating task out of the CPU.
4661 */
4662 detach_entity_load_avg(cfs_rq, se);
4663 update_tg_load_avg(cfs_rq);
4664 } else if (decayed) {
4665 cfs_rq_util_change(cfs_rq, 0);
4666
4667 if (flags & UPDATE_TG)
4668 update_tg_load_avg(cfs_rq);
4669 }
4670 }
4671
4672 /*
4673 * Synchronize entity load avg of dequeued entity without locking
4674 * the previous rq.
4675 */
sync_entity_load_avg(struct sched_entity * se)4676 static void sync_entity_load_avg(struct sched_entity *se)
4677 {
4678 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4679 u64 last_update_time;
4680
4681 last_update_time = cfs_rq_last_update_time(cfs_rq);
4682 __update_load_avg_blocked_se(last_update_time, se);
4683 }
4684
4685 /*
4686 * Task first catches up with cfs_rq, and then subtract
4687 * itself from the cfs_rq (task must be off the queue now).
4688 */
remove_entity_load_avg(struct sched_entity * se)4689 static void remove_entity_load_avg(struct sched_entity *se)
4690 {
4691 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4692 unsigned long flags;
4693
4694 /*
4695 * tasks cannot exit without having gone through wake_up_new_task() ->
4696 * enqueue_task_fair() which will have added things to the cfs_rq,
4697 * so we can remove unconditionally.
4698 */
4699
4700 sync_entity_load_avg(se);
4701
4702 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4703 ++cfs_rq->removed.nr;
4704 cfs_rq->removed.util_avg += se->avg.util_avg;
4705 cfs_rq->removed.load_avg += se->avg.load_avg;
4706 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4707 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4708 }
4709
cfs_rq_runnable_avg(struct cfs_rq * cfs_rq)4710 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4711 {
4712 return cfs_rq->avg.runnable_avg;
4713 }
4714
cfs_rq_load_avg(struct cfs_rq * cfs_rq)4715 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4716 {
4717 return cfs_rq->avg.load_avg;
4718 }
4719
4720 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4721
task_util(struct task_struct * p)4722 static inline unsigned long task_util(struct task_struct *p)
4723 {
4724 return READ_ONCE(p->se.avg.util_avg);
4725 }
4726
_task_util_est(struct task_struct * p)4727 static inline unsigned long _task_util_est(struct task_struct *p)
4728 {
4729 struct util_est ue = READ_ONCE(p->se.avg.util_est);
4730
4731 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4732 }
4733
task_util_est(struct task_struct * p)4734 static inline unsigned long task_util_est(struct task_struct *p)
4735 {
4736 return max(task_util(p), _task_util_est(p));
4737 }
4738
util_est_enqueue(struct cfs_rq * cfs_rq,struct task_struct * p)4739 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4740 struct task_struct *p)
4741 {
4742 unsigned int enqueued;
4743
4744 if (!sched_feat(UTIL_EST))
4745 return;
4746
4747 /* Update root cfs_rq's estimated utilization */
4748 enqueued = cfs_rq->avg.util_est.enqueued;
4749 enqueued += _task_util_est(p);
4750 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4751
4752 trace_sched_util_est_cfs_tp(cfs_rq);
4753 }
4754
util_est_dequeue(struct cfs_rq * cfs_rq,struct task_struct * p)4755 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4756 struct task_struct *p)
4757 {
4758 unsigned int enqueued;
4759
4760 if (!sched_feat(UTIL_EST))
4761 return;
4762
4763 /* Update root cfs_rq's estimated utilization */
4764 enqueued = cfs_rq->avg.util_est.enqueued;
4765 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4766 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4767
4768 trace_sched_util_est_cfs_tp(cfs_rq);
4769 }
4770
4771 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4772
4773 /*
4774 * Check if a (signed) value is within a specified (unsigned) margin,
4775 * based on the observation that:
4776 *
4777 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4778 *
4779 * NOTE: this only works when value + margin < INT_MAX.
4780 */
within_margin(int value,int margin)4781 static inline bool within_margin(int value, int margin)
4782 {
4783 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4784 }
4785
util_est_update(struct cfs_rq * cfs_rq,struct task_struct * p,bool task_sleep)4786 static inline void util_est_update(struct cfs_rq *cfs_rq,
4787 struct task_struct *p,
4788 bool task_sleep)
4789 {
4790 long last_ewma_diff, last_enqueued_diff;
4791 struct util_est ue;
4792
4793 if (!sched_feat(UTIL_EST))
4794 return;
4795
4796 /*
4797 * Skip update of task's estimated utilization when the task has not
4798 * yet completed an activation, e.g. being migrated.
4799 */
4800 if (!task_sleep)
4801 return;
4802
4803 /*
4804 * If the PELT values haven't changed since enqueue time,
4805 * skip the util_est update.
4806 */
4807 ue = p->se.avg.util_est;
4808 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4809 return;
4810
4811 last_enqueued_diff = ue.enqueued;
4812
4813 /*
4814 * Reset EWMA on utilization increases, the moving average is used only
4815 * to smooth utilization decreases.
4816 */
4817 ue.enqueued = task_util(p);
4818 if (sched_feat(UTIL_EST_FASTUP)) {
4819 if (ue.ewma < ue.enqueued) {
4820 ue.ewma = ue.enqueued;
4821 goto done;
4822 }
4823 }
4824
4825 /*
4826 * Skip update of task's estimated utilization when its members are
4827 * already ~1% close to its last activation value.
4828 */
4829 last_ewma_diff = ue.enqueued - ue.ewma;
4830 last_enqueued_diff -= ue.enqueued;
4831 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4832 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4833 goto done;
4834
4835 return;
4836 }
4837
4838 /*
4839 * To avoid overestimation of actual task utilization, skip updates if
4840 * we cannot grant there is idle time in this CPU.
4841 */
4842 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4843 return;
4844
4845 /*
4846 * Update Task's estimated utilization
4847 *
4848 * When *p completes an activation we can consolidate another sample
4849 * of the task size. This is done by storing the current PELT value
4850 * as ue.enqueued and by using this value to update the Exponential
4851 * Weighted Moving Average (EWMA):
4852 *
4853 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4854 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4855 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4856 * = w * ( last_ewma_diff ) + ewma(t-1)
4857 * = w * (last_ewma_diff + ewma(t-1) / w)
4858 *
4859 * Where 'w' is the weight of new samples, which is configured to be
4860 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4861 */
4862 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4863 ue.ewma += last_ewma_diff;
4864 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4865 done:
4866 ue.enqueued |= UTIL_AVG_UNCHANGED;
4867 WRITE_ONCE(p->se.avg.util_est, ue);
4868
4869 trace_sched_util_est_se_tp(&p->se);
4870 }
4871
util_fits_cpu(unsigned long util,unsigned long uclamp_min,unsigned long uclamp_max,int cpu)4872 static inline int util_fits_cpu(unsigned long util,
4873 unsigned long uclamp_min,
4874 unsigned long uclamp_max,
4875 int cpu)
4876 {
4877 unsigned long capacity_orig, capacity_orig_thermal;
4878 unsigned long capacity = capacity_of(cpu);
4879 bool fits, uclamp_max_fits;
4880
4881 /*
4882 * Check if the real util fits without any uclamp boost/cap applied.
4883 */
4884 fits = fits_capacity(util, capacity);
4885
4886 if (!uclamp_is_used())
4887 return fits;
4888
4889 /*
4890 * We must use capacity_orig_of() for comparing against uclamp_min and
4891 * uclamp_max. We only care about capacity pressure (by using
4892 * capacity_of()) for comparing against the real util.
4893 *
4894 * If a task is boosted to 1024 for example, we don't want a tiny
4895 * pressure to skew the check whether it fits a CPU or not.
4896 *
4897 * Similarly if a task is capped to capacity_orig_of(little_cpu), it
4898 * should fit a little cpu even if there's some pressure.
4899 *
4900 * Only exception is for thermal pressure since it has a direct impact
4901 * on available OPP of the system.
4902 *
4903 * We honour it for uclamp_min only as a drop in performance level
4904 * could result in not getting the requested minimum performance level.
4905 *
4906 * For uclamp_max, we can tolerate a drop in performance level as the
4907 * goal is to cap the task. So it's okay if it's getting less.
4908 */
4909 capacity_orig = capacity_orig_of(cpu);
4910 capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
4911
4912 /*
4913 * We want to force a task to fit a cpu as implied by uclamp_max.
4914 * But we do have some corner cases to cater for..
4915 *
4916 *
4917 * C=z
4918 * | ___
4919 * | C=y | |
4920 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4921 * | C=x | | | |
4922 * | ___ | | | |
4923 * | | | | | | | (util somewhere in this region)
4924 * | | | | | | |
4925 * | | | | | | |
4926 * +----------------------------------------
4927 * cpu0 cpu1 cpu2
4928 *
4929 * In the above example if a task is capped to a specific performance
4930 * point, y, then when:
4931 *
4932 * * util = 80% of x then it does not fit on cpu0 and should migrate
4933 * to cpu1
4934 * * util = 80% of y then it is forced to fit on cpu1 to honour
4935 * uclamp_max request.
4936 *
4937 * which is what we're enforcing here. A task always fits if
4938 * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
4939 * the normal upmigration rules should withhold still.
4940 *
4941 * Only exception is when we are on max capacity, then we need to be
4942 * careful not to block overutilized state. This is so because:
4943 *
4944 * 1. There's no concept of capping at max_capacity! We can't go
4945 * beyond this performance level anyway.
4946 * 2. The system is being saturated when we're operating near
4947 * max capacity, it doesn't make sense to block overutilized.
4948 */
4949 uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
4950 uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
4951 fits = fits || uclamp_max_fits;
4952
4953 /*
4954 *
4955 * C=z
4956 * | ___ (region a, capped, util >= uclamp_max)
4957 * | C=y | |
4958 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4959 * | C=x | | | |
4960 * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max)
4961 * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
4962 * | | | | | | |
4963 * | | | | | | | (region c, boosted, util < uclamp_min)
4964 * +----------------------------------------
4965 * cpu0 cpu1 cpu2
4966 *
4967 * a) If util > uclamp_max, then we're capped, we don't care about
4968 * actual fitness value here. We only care if uclamp_max fits
4969 * capacity without taking margin/pressure into account.
4970 * See comment above.
4971 *
4972 * b) If uclamp_min <= util <= uclamp_max, then the normal
4973 * fits_capacity() rules apply. Except we need to ensure that we
4974 * enforce we remain within uclamp_max, see comment above.
4975 *
4976 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
4977 * need to take into account the boosted value fits the CPU without
4978 * taking margin/pressure into account.
4979 *
4980 * Cases (a) and (b) are handled in the 'fits' variable already. We
4981 * just need to consider an extra check for case (c) after ensuring we
4982 * handle the case uclamp_min > uclamp_max.
4983 */
4984 uclamp_min = min(uclamp_min, uclamp_max);
4985 if (fits && (util < uclamp_min) && (uclamp_min > capacity_orig_thermal))
4986 return -1;
4987
4988 return fits;
4989 }
4990
task_fits_cpu(struct task_struct * p,int cpu)4991 static inline int task_fits_cpu(struct task_struct *p, int cpu)
4992 {
4993 unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
4994 unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
4995 unsigned long util = task_util_est(p);
4996 /*
4997 * Return true only if the cpu fully fits the task requirements, which
4998 * include the utilization but also the performance hints.
4999 */
5000 return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
5001 }
5002
update_misfit_status(struct task_struct * p,struct rq * rq)5003 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
5004 {
5005 if (!sched_asym_cpucap_active())
5006 return;
5007
5008 if (!p || p->nr_cpus_allowed == 1) {
5009 rq->misfit_task_load = 0;
5010 return;
5011 }
5012
5013 if (task_fits_cpu(p, cpu_of(rq))) {
5014 rq->misfit_task_load = 0;
5015 return;
5016 }
5017
5018 /*
5019 * Make sure that misfit_task_load will not be null even if
5020 * task_h_load() returns 0.
5021 */
5022 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
5023 }
5024
5025 #else /* CONFIG_SMP */
5026
cfs_rq_is_decayed(struct cfs_rq * cfs_rq)5027 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
5028 {
5029 return !cfs_rq->nr_running;
5030 }
5031
5032 #define UPDATE_TG 0x0
5033 #define SKIP_AGE_LOAD 0x0
5034 #define DO_ATTACH 0x0
5035 #define DO_DETACH 0x0
5036
update_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se,int not_used1)5037 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
5038 {
5039 cfs_rq_util_change(cfs_rq, 0);
5040 }
5041
remove_entity_load_avg(struct sched_entity * se)5042 static inline void remove_entity_load_avg(struct sched_entity *se) {}
5043
5044 static inline void
attach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)5045 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5046 static inline void
detach_entity_load_avg(struct cfs_rq * cfs_rq,struct sched_entity * se)5047 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5048
newidle_balance(struct rq * rq,struct rq_flags * rf)5049 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
5050 {
5051 return 0;
5052 }
5053
5054 static inline void
util_est_enqueue(struct cfs_rq * cfs_rq,struct task_struct * p)5055 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5056
5057 static inline void
util_est_dequeue(struct cfs_rq * cfs_rq,struct task_struct * p)5058 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5059
5060 static inline void
util_est_update(struct cfs_rq * cfs_rq,struct task_struct * p,bool task_sleep)5061 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
5062 bool task_sleep) {}
update_misfit_status(struct task_struct * p,struct rq * rq)5063 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
5064
5065 #endif /* CONFIG_SMP */
5066
5067 static void
place_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5068 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5069 {
5070 u64 vslice, vruntime = avg_vruntime(cfs_rq);
5071 s64 lag = 0;
5072
5073 se->slice = sysctl_sched_base_slice;
5074 vslice = calc_delta_fair(se->slice, se);
5075
5076 /*
5077 * Due to how V is constructed as the weighted average of entities,
5078 * adding tasks with positive lag, or removing tasks with negative lag
5079 * will move 'time' backwards, this can screw around with the lag of
5080 * other tasks.
5081 *
5082 * EEVDF: placement strategy #1 / #2
5083 */
5084 if (sched_feat(PLACE_LAG) && cfs_rq->nr_running) {
5085 struct sched_entity *curr = cfs_rq->curr;
5086 unsigned long load;
5087
5088 lag = se->vlag;
5089
5090 /*
5091 * If we want to place a task and preserve lag, we have to
5092 * consider the effect of the new entity on the weighted
5093 * average and compensate for this, otherwise lag can quickly
5094 * evaporate.
5095 *
5096 * Lag is defined as:
5097 *
5098 * lag_i = S - s_i = w_i * (V - v_i)
5099 *
5100 * To avoid the 'w_i' term all over the place, we only track
5101 * the virtual lag:
5102 *
5103 * vl_i = V - v_i <=> v_i = V - vl_i
5104 *
5105 * And we take V to be the weighted average of all v:
5106 *
5107 * V = (\Sum w_j*v_j) / W
5108 *
5109 * Where W is: \Sum w_j
5110 *
5111 * Then, the weighted average after adding an entity with lag
5112 * vl_i is given by:
5113 *
5114 * V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i)
5115 * = (W*V + w_i*(V - vl_i)) / (W + w_i)
5116 * = (W*V + w_i*V - w_i*vl_i) / (W + w_i)
5117 * = (V*(W + w_i) - w_i*l) / (W + w_i)
5118 * = V - w_i*vl_i / (W + w_i)
5119 *
5120 * And the actual lag after adding an entity with vl_i is:
5121 *
5122 * vl'_i = V' - v_i
5123 * = V - w_i*vl_i / (W + w_i) - (V - vl_i)
5124 * = vl_i - w_i*vl_i / (W + w_i)
5125 *
5126 * Which is strictly less than vl_i. So in order to preserve lag
5127 * we should inflate the lag before placement such that the
5128 * effective lag after placement comes out right.
5129 *
5130 * As such, invert the above relation for vl'_i to get the vl_i
5131 * we need to use such that the lag after placement is the lag
5132 * we computed before dequeue.
5133 *
5134 * vl'_i = vl_i - w_i*vl_i / (W + w_i)
5135 * = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)
5136 *
5137 * (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i
5138 * = W*vl_i
5139 *
5140 * vl_i = (W + w_i)*vl'_i / W
5141 */
5142 load = cfs_rq->avg_load;
5143 if (curr && curr->on_rq)
5144 load += scale_load_down(curr->load.weight);
5145
5146 lag *= load + scale_load_down(se->load.weight);
5147 if (WARN_ON_ONCE(!load))
5148 load = 1;
5149 lag = div_s64(lag, load);
5150 }
5151
5152 se->vruntime = vruntime - lag;
5153
5154 /*
5155 * When joining the competition; the exisiting tasks will be,
5156 * on average, halfway through their slice, as such start tasks
5157 * off with half a slice to ease into the competition.
5158 */
5159 if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL))
5160 vslice /= 2;
5161
5162 /*
5163 * EEVDF: vd_i = ve_i + r_i/w_i
5164 */
5165 se->deadline = se->vruntime + vslice;
5166 }
5167
5168 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
5169 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
5170
5171 static inline bool cfs_bandwidth_used(void);
5172
5173 static void
enqueue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5174 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5175 {
5176 bool curr = cfs_rq->curr == se;
5177
5178 /*
5179 * If we're the current task, we must renormalise before calling
5180 * update_curr().
5181 */
5182 if (curr)
5183 place_entity(cfs_rq, se, flags);
5184
5185 update_curr(cfs_rq);
5186
5187 /*
5188 * When enqueuing a sched_entity, we must:
5189 * - Update loads to have both entity and cfs_rq synced with now.
5190 * - For group_entity, update its runnable_weight to reflect the new
5191 * h_nr_running of its group cfs_rq.
5192 * - For group_entity, update its weight to reflect the new share of
5193 * its group cfs_rq
5194 * - Add its new weight to cfs_rq->load.weight
5195 */
5196 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
5197 se_update_runnable(se);
5198 /*
5199 * XXX update_load_avg() above will have attached us to the pelt sum;
5200 * but update_cfs_group() here will re-adjust the weight and have to
5201 * undo/redo all that. Seems wasteful.
5202 */
5203 update_cfs_group(se);
5204
5205 /*
5206 * XXX now that the entity has been re-weighted, and it's lag adjusted,
5207 * we can place the entity.
5208 */
5209 if (!curr)
5210 place_entity(cfs_rq, se, flags);
5211
5212 account_entity_enqueue(cfs_rq, se);
5213
5214 /* Entity has migrated, no longer consider this task hot */
5215 if (flags & ENQUEUE_MIGRATED)
5216 se->exec_start = 0;
5217
5218 check_schedstat_required();
5219 update_stats_enqueue_fair(cfs_rq, se, flags);
5220 if (!curr)
5221 __enqueue_entity(cfs_rq, se);
5222 se->on_rq = 1;
5223
5224 if (cfs_rq->nr_running == 1) {
5225 check_enqueue_throttle(cfs_rq);
5226 if (!throttled_hierarchy(cfs_rq)) {
5227 list_add_leaf_cfs_rq(cfs_rq);
5228 } else {
5229 #ifdef CONFIG_CFS_BANDWIDTH
5230 struct rq *rq = rq_of(cfs_rq);
5231
5232 if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)
5233 cfs_rq->throttled_clock = rq_clock(rq);
5234 if (!cfs_rq->throttled_clock_self)
5235 cfs_rq->throttled_clock_self = rq_clock(rq);
5236 #endif
5237 }
5238 }
5239 }
5240
__clear_buddies_next(struct sched_entity * se)5241 static void __clear_buddies_next(struct sched_entity *se)
5242 {
5243 for_each_sched_entity(se) {
5244 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5245 if (cfs_rq->next != se)
5246 break;
5247
5248 cfs_rq->next = NULL;
5249 }
5250 }
5251
clear_buddies(struct cfs_rq * cfs_rq,struct sched_entity * se)5252 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
5253 {
5254 if (cfs_rq->next == se)
5255 __clear_buddies_next(se);
5256 }
5257
5258 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5259
5260 static void
dequeue_entity(struct cfs_rq * cfs_rq,struct sched_entity * se,int flags)5261 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5262 {
5263 int action = UPDATE_TG;
5264
5265 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
5266 action |= DO_DETACH;
5267
5268 /*
5269 * Update run-time statistics of the 'current'.
5270 */
5271 update_curr(cfs_rq);
5272
5273 /*
5274 * When dequeuing a sched_entity, we must:
5275 * - Update loads to have both entity and cfs_rq synced with now.
5276 * - For group_entity, update its runnable_weight to reflect the new
5277 * h_nr_running of its group cfs_rq.
5278 * - Subtract its previous weight from cfs_rq->load.weight.
5279 * - For group entity, update its weight to reflect the new share
5280 * of its group cfs_rq.
5281 */
5282 update_load_avg(cfs_rq, se, action);
5283 se_update_runnable(se);
5284
5285 update_stats_dequeue_fair(cfs_rq, se, flags);
5286
5287 clear_buddies(cfs_rq, se);
5288
5289 update_entity_lag(cfs_rq, se);
5290 if (se != cfs_rq->curr)
5291 __dequeue_entity(cfs_rq, se);
5292 se->on_rq = 0;
5293 account_entity_dequeue(cfs_rq, se);
5294
5295 /* return excess runtime on last dequeue */
5296 return_cfs_rq_runtime(cfs_rq);
5297
5298 update_cfs_group(se);
5299
5300 /*
5301 * Now advance min_vruntime if @se was the entity holding it back,
5302 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
5303 * put back on, and if we advance min_vruntime, we'll be placed back
5304 * further than we started -- ie. we'll be penalized.
5305 */
5306 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
5307 update_min_vruntime(cfs_rq);
5308
5309 if (cfs_rq->nr_running == 0)
5310 update_idle_cfs_rq_clock_pelt(cfs_rq);
5311 }
5312
5313 static void
set_next_entity(struct cfs_rq * cfs_rq,struct sched_entity * se)5314 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
5315 {
5316 clear_buddies(cfs_rq, se);
5317
5318 /* 'current' is not kept within the tree. */
5319 if (se->on_rq) {
5320 /*
5321 * Any task has to be enqueued before it get to execute on
5322 * a CPU. So account for the time it spent waiting on the
5323 * runqueue.
5324 */
5325 update_stats_wait_end_fair(cfs_rq, se);
5326 __dequeue_entity(cfs_rq, se);
5327 update_load_avg(cfs_rq, se, UPDATE_TG);
5328 /*
5329 * HACK, stash a copy of deadline at the point of pick in vlag,
5330 * which isn't used until dequeue.
5331 */
5332 se->vlag = se->deadline;
5333 }
5334
5335 update_stats_curr_start(cfs_rq, se);
5336 cfs_rq->curr = se;
5337
5338 /*
5339 * Track our maximum slice length, if the CPU's load is at
5340 * least twice that of our own weight (i.e. dont track it
5341 * when there are only lesser-weight tasks around):
5342 */
5343 if (schedstat_enabled() &&
5344 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
5345 struct sched_statistics *stats;
5346
5347 stats = __schedstats_from_se(se);
5348 __schedstat_set(stats->slice_max,
5349 max((u64)stats->slice_max,
5350 se->sum_exec_runtime - se->prev_sum_exec_runtime));
5351 }
5352
5353 se->prev_sum_exec_runtime = se->sum_exec_runtime;
5354 }
5355
5356 /*
5357 * Pick the next process, keeping these things in mind, in this order:
5358 * 1) keep things fair between processes/task groups
5359 * 2) pick the "next" process, since someone really wants that to run
5360 * 3) pick the "last" process, for cache locality
5361 * 4) do not run the "skip" process, if something else is available
5362 */
5363 static struct sched_entity *
pick_next_entity(struct cfs_rq * cfs_rq,struct sched_entity * curr)5364 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
5365 {
5366 /*
5367 * Enabling NEXT_BUDDY will affect latency but not fairness.
5368 */
5369 if (sched_feat(NEXT_BUDDY) &&
5370 cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next))
5371 return cfs_rq->next;
5372
5373 return pick_eevdf(cfs_rq);
5374 }
5375
5376 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5377
put_prev_entity(struct cfs_rq * cfs_rq,struct sched_entity * prev)5378 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5379 {
5380 /*
5381 * If still on the runqueue then deactivate_task()
5382 * was not called and update_curr() has to be done:
5383 */
5384 if (prev->on_rq)
5385 update_curr(cfs_rq);
5386
5387 /* throttle cfs_rqs exceeding runtime */
5388 check_cfs_rq_runtime(cfs_rq);
5389
5390 if (prev->on_rq) {
5391 update_stats_wait_start_fair(cfs_rq, prev);
5392 /* Put 'current' back into the tree. */
5393 __enqueue_entity(cfs_rq, prev);
5394 /* in !on_rq case, update occurred at dequeue */
5395 update_load_avg(cfs_rq, prev, 0);
5396 }
5397 cfs_rq->curr = NULL;
5398 }
5399
5400 static void
entity_tick(struct cfs_rq * cfs_rq,struct sched_entity * curr,int queued)5401 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5402 {
5403 /*
5404 * Update run-time statistics of the 'current'.
5405 */
5406 update_curr(cfs_rq);
5407
5408 /*
5409 * Ensure that runnable average is periodically updated.
5410 */
5411 update_load_avg(cfs_rq, curr, UPDATE_TG);
5412 update_cfs_group(curr);
5413
5414 #ifdef CONFIG_SCHED_HRTICK
5415 /*
5416 * queued ticks are scheduled to match the slice, so don't bother
5417 * validating it and just reschedule.
5418 */
5419 if (queued) {
5420 resched_curr(rq_of(cfs_rq));
5421 return;
5422 }
5423 /*
5424 * don't let the period tick interfere with the hrtick preemption
5425 */
5426 if (!sched_feat(DOUBLE_TICK) &&
5427 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5428 return;
5429 #endif
5430 }
5431
5432
5433 /**************************************************
5434 * CFS bandwidth control machinery
5435 */
5436
5437 #ifdef CONFIG_CFS_BANDWIDTH
5438
5439 #ifdef CONFIG_JUMP_LABEL
5440 static struct static_key __cfs_bandwidth_used;
5441
cfs_bandwidth_used(void)5442 static inline bool cfs_bandwidth_used(void)
5443 {
5444 return static_key_false(&__cfs_bandwidth_used);
5445 }
5446
cfs_bandwidth_usage_inc(void)5447 void cfs_bandwidth_usage_inc(void)
5448 {
5449 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5450 }
5451
cfs_bandwidth_usage_dec(void)5452 void cfs_bandwidth_usage_dec(void)
5453 {
5454 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5455 }
5456 #else /* CONFIG_JUMP_LABEL */
cfs_bandwidth_used(void)5457 static bool cfs_bandwidth_used(void)
5458 {
5459 return true;
5460 }
5461
cfs_bandwidth_usage_inc(void)5462 void cfs_bandwidth_usage_inc(void) {}
cfs_bandwidth_usage_dec(void)5463 void cfs_bandwidth_usage_dec(void) {}
5464 #endif /* CONFIG_JUMP_LABEL */
5465
5466 /*
5467 * default period for cfs group bandwidth.
5468 * default: 0.1s, units: nanoseconds
5469 */
default_cfs_period(void)5470 static inline u64 default_cfs_period(void)
5471 {
5472 return 100000000ULL;
5473 }
5474
sched_cfs_bandwidth_slice(void)5475 static inline u64 sched_cfs_bandwidth_slice(void)
5476 {
5477 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5478 }
5479
5480 /*
5481 * Replenish runtime according to assigned quota. We use sched_clock_cpu
5482 * directly instead of rq->clock to avoid adding additional synchronization
5483 * around rq->lock.
5484 *
5485 * requires cfs_b->lock
5486 */
__refill_cfs_bandwidth_runtime(struct cfs_bandwidth * cfs_b)5487 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5488 {
5489 s64 runtime;
5490
5491 if (unlikely(cfs_b->quota == RUNTIME_INF))
5492 return;
5493
5494 cfs_b->runtime += cfs_b->quota;
5495 runtime = cfs_b->runtime_snap - cfs_b->runtime;
5496 if (runtime > 0) {
5497 cfs_b->burst_time += runtime;
5498 cfs_b->nr_burst++;
5499 }
5500
5501 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5502 cfs_b->runtime_snap = cfs_b->runtime;
5503 }
5504
tg_cfs_bandwidth(struct task_group * tg)5505 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5506 {
5507 return &tg->cfs_bandwidth;
5508 }
5509
5510 /* returns 0 on failure to allocate runtime */
__assign_cfs_rq_runtime(struct cfs_bandwidth * cfs_b,struct cfs_rq * cfs_rq,u64 target_runtime)5511 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5512 struct cfs_rq *cfs_rq, u64 target_runtime)
5513 {
5514 u64 min_amount, amount = 0;
5515
5516 lockdep_assert_held(&cfs_b->lock);
5517
5518 /* note: this is a positive sum as runtime_remaining <= 0 */
5519 min_amount = target_runtime - cfs_rq->runtime_remaining;
5520
5521 if (cfs_b->quota == RUNTIME_INF)
5522 amount = min_amount;
5523 else {
5524 start_cfs_bandwidth(cfs_b);
5525
5526 if (cfs_b->runtime > 0) {
5527 amount = min(cfs_b->runtime, min_amount);
5528 cfs_b->runtime -= amount;
5529 cfs_b->idle = 0;
5530 }
5531 }
5532
5533 cfs_rq->runtime_remaining += amount;
5534
5535 return cfs_rq->runtime_remaining > 0;
5536 }
5537
5538 /* returns 0 on failure to allocate runtime */
assign_cfs_rq_runtime(struct cfs_rq * cfs_rq)5539 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5540 {
5541 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5542 int ret;
5543
5544 raw_spin_lock(&cfs_b->lock);
5545 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5546 raw_spin_unlock(&cfs_b->lock);
5547
5548 return ret;
5549 }
5550
__account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)5551 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5552 {
5553 /* dock delta_exec before expiring quota (as it could span periods) */
5554 cfs_rq->runtime_remaining -= delta_exec;
5555
5556 if (likely(cfs_rq->runtime_remaining > 0))
5557 return;
5558
5559 if (cfs_rq->throttled)
5560 return;
5561 /*
5562 * if we're unable to extend our runtime we resched so that the active
5563 * hierarchy can be throttled
5564 */
5565 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5566 resched_curr(rq_of(cfs_rq));
5567 }
5568
5569 static __always_inline
account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)5570 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5571 {
5572 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5573 return;
5574
5575 __account_cfs_rq_runtime(cfs_rq, delta_exec);
5576 }
5577
cfs_rq_throttled(struct cfs_rq * cfs_rq)5578 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5579 {
5580 return cfs_bandwidth_used() && cfs_rq->throttled;
5581 }
5582
5583 /* check whether cfs_rq, or any parent, is throttled */
throttled_hierarchy(struct cfs_rq * cfs_rq)5584 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5585 {
5586 return cfs_bandwidth_used() && cfs_rq->throttle_count;
5587 }
5588
5589 /*
5590 * Ensure that neither of the group entities corresponding to src_cpu or
5591 * dest_cpu are members of a throttled hierarchy when performing group
5592 * load-balance operations.
5593 */
throttled_lb_pair(struct task_group * tg,int src_cpu,int dest_cpu)5594 static inline int throttled_lb_pair(struct task_group *tg,
5595 int src_cpu, int dest_cpu)
5596 {
5597 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5598
5599 src_cfs_rq = tg->cfs_rq[src_cpu];
5600 dest_cfs_rq = tg->cfs_rq[dest_cpu];
5601
5602 return throttled_hierarchy(src_cfs_rq) ||
5603 throttled_hierarchy(dest_cfs_rq);
5604 }
5605
tg_unthrottle_up(struct task_group * tg,void * data)5606 static int tg_unthrottle_up(struct task_group *tg, void *data)
5607 {
5608 struct rq *rq = data;
5609 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5610
5611 cfs_rq->throttle_count--;
5612 if (!cfs_rq->throttle_count) {
5613 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5614 cfs_rq->throttled_clock_pelt;
5615
5616 /* Add cfs_rq with load or one or more already running entities to the list */
5617 if (!cfs_rq_is_decayed(cfs_rq))
5618 list_add_leaf_cfs_rq(cfs_rq);
5619
5620 if (cfs_rq->throttled_clock_self) {
5621 u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;
5622
5623 cfs_rq->throttled_clock_self = 0;
5624
5625 if (SCHED_WARN_ON((s64)delta < 0))
5626 delta = 0;
5627
5628 cfs_rq->throttled_clock_self_time += delta;
5629 }
5630 }
5631
5632 return 0;
5633 }
5634
tg_throttle_down(struct task_group * tg,void * data)5635 static int tg_throttle_down(struct task_group *tg, void *data)
5636 {
5637 struct rq *rq = data;
5638 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5639
5640 /* group is entering throttled state, stop time */
5641 if (!cfs_rq->throttle_count) {
5642 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5643 list_del_leaf_cfs_rq(cfs_rq);
5644
5645 SCHED_WARN_ON(cfs_rq->throttled_clock_self);
5646 if (cfs_rq->nr_running)
5647 cfs_rq->throttled_clock_self = rq_clock(rq);
5648 }
5649 cfs_rq->throttle_count++;
5650
5651 return 0;
5652 }
5653
throttle_cfs_rq(struct cfs_rq * cfs_rq)5654 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5655 {
5656 struct rq *rq = rq_of(cfs_rq);
5657 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5658 struct sched_entity *se;
5659 long task_delta, idle_task_delta, dequeue = 1;
5660
5661 raw_spin_lock(&cfs_b->lock);
5662 /* This will start the period timer if necessary */
5663 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5664 /*
5665 * We have raced with bandwidth becoming available, and if we
5666 * actually throttled the timer might not unthrottle us for an
5667 * entire period. We additionally needed to make sure that any
5668 * subsequent check_cfs_rq_runtime calls agree not to throttle
5669 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5670 * for 1ns of runtime rather than just check cfs_b.
5671 */
5672 dequeue = 0;
5673 } else {
5674 list_add_tail_rcu(&cfs_rq->throttled_list,
5675 &cfs_b->throttled_cfs_rq);
5676 }
5677 raw_spin_unlock(&cfs_b->lock);
5678
5679 if (!dequeue)
5680 return false; /* Throttle no longer required. */
5681
5682 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5683
5684 /* freeze hierarchy runnable averages while throttled */
5685 rcu_read_lock();
5686 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5687 rcu_read_unlock();
5688
5689 task_delta = cfs_rq->h_nr_running;
5690 idle_task_delta = cfs_rq->idle_h_nr_running;
5691 for_each_sched_entity(se) {
5692 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5693 /* throttled entity or throttle-on-deactivate */
5694 if (!se->on_rq)
5695 goto done;
5696
5697 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5698
5699 if (cfs_rq_is_idle(group_cfs_rq(se)))
5700 idle_task_delta = cfs_rq->h_nr_running;
5701
5702 qcfs_rq->h_nr_running -= task_delta;
5703 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5704
5705 if (qcfs_rq->load.weight) {
5706 /* Avoid re-evaluating load for this entity: */
5707 se = parent_entity(se);
5708 break;
5709 }
5710 }
5711
5712 for_each_sched_entity(se) {
5713 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5714 /* throttled entity or throttle-on-deactivate */
5715 if (!se->on_rq)
5716 goto done;
5717
5718 update_load_avg(qcfs_rq, se, 0);
5719 se_update_runnable(se);
5720
5721 if (cfs_rq_is_idle(group_cfs_rq(se)))
5722 idle_task_delta = cfs_rq->h_nr_running;
5723
5724 qcfs_rq->h_nr_running -= task_delta;
5725 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5726 }
5727
5728 /* At this point se is NULL and we are at root level*/
5729 sub_nr_running(rq, task_delta);
5730
5731 done:
5732 /*
5733 * Note: distribution will already see us throttled via the
5734 * throttled-list. rq->lock protects completion.
5735 */
5736 cfs_rq->throttled = 1;
5737 SCHED_WARN_ON(cfs_rq->throttled_clock);
5738 if (cfs_rq->nr_running)
5739 cfs_rq->throttled_clock = rq_clock(rq);
5740 return true;
5741 }
5742
unthrottle_cfs_rq(struct cfs_rq * cfs_rq)5743 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5744 {
5745 struct rq *rq = rq_of(cfs_rq);
5746 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5747 struct sched_entity *se;
5748 long task_delta, idle_task_delta;
5749
5750 se = cfs_rq->tg->se[cpu_of(rq)];
5751
5752 cfs_rq->throttled = 0;
5753
5754 update_rq_clock(rq);
5755
5756 raw_spin_lock(&cfs_b->lock);
5757 if (cfs_rq->throttled_clock) {
5758 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5759 cfs_rq->throttled_clock = 0;
5760 }
5761 list_del_rcu(&cfs_rq->throttled_list);
5762 raw_spin_unlock(&cfs_b->lock);
5763
5764 /* update hierarchical throttle state */
5765 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5766
5767 if (!cfs_rq->load.weight) {
5768 if (!cfs_rq->on_list)
5769 return;
5770 /*
5771 * Nothing to run but something to decay (on_list)?
5772 * Complete the branch.
5773 */
5774 for_each_sched_entity(se) {
5775 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5776 break;
5777 }
5778 goto unthrottle_throttle;
5779 }
5780
5781 task_delta = cfs_rq->h_nr_running;
5782 idle_task_delta = cfs_rq->idle_h_nr_running;
5783 for_each_sched_entity(se) {
5784 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5785
5786 if (se->on_rq)
5787 break;
5788 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5789
5790 if (cfs_rq_is_idle(group_cfs_rq(se)))
5791 idle_task_delta = cfs_rq->h_nr_running;
5792
5793 qcfs_rq->h_nr_running += task_delta;
5794 qcfs_rq->idle_h_nr_running += idle_task_delta;
5795
5796 /* end evaluation on encountering a throttled cfs_rq */
5797 if (cfs_rq_throttled(qcfs_rq))
5798 goto unthrottle_throttle;
5799 }
5800
5801 for_each_sched_entity(se) {
5802 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5803
5804 update_load_avg(qcfs_rq, se, UPDATE_TG);
5805 se_update_runnable(se);
5806
5807 if (cfs_rq_is_idle(group_cfs_rq(se)))
5808 idle_task_delta = cfs_rq->h_nr_running;
5809
5810 qcfs_rq->h_nr_running += task_delta;
5811 qcfs_rq->idle_h_nr_running += idle_task_delta;
5812
5813 /* end evaluation on encountering a throttled cfs_rq */
5814 if (cfs_rq_throttled(qcfs_rq))
5815 goto unthrottle_throttle;
5816 }
5817
5818 /* At this point se is NULL and we are at root level*/
5819 add_nr_running(rq, task_delta);
5820
5821 unthrottle_throttle:
5822 assert_list_leaf_cfs_rq(rq);
5823
5824 /* Determine whether we need to wake up potentially idle CPU: */
5825 if (rq->curr == rq->idle && rq->cfs.nr_running)
5826 resched_curr(rq);
5827 }
5828
5829 #ifdef CONFIG_SMP
__cfsb_csd_unthrottle(void * arg)5830 static void __cfsb_csd_unthrottle(void *arg)
5831 {
5832 struct cfs_rq *cursor, *tmp;
5833 struct rq *rq = arg;
5834 struct rq_flags rf;
5835
5836 rq_lock(rq, &rf);
5837
5838 /*
5839 * Iterating over the list can trigger several call to
5840 * update_rq_clock() in unthrottle_cfs_rq().
5841 * Do it once and skip the potential next ones.
5842 */
5843 update_rq_clock(rq);
5844 rq_clock_start_loop_update(rq);
5845
5846 /*
5847 * Since we hold rq lock we're safe from concurrent manipulation of
5848 * the CSD list. However, this RCU critical section annotates the
5849 * fact that we pair with sched_free_group_rcu(), so that we cannot
5850 * race with group being freed in the window between removing it
5851 * from the list and advancing to the next entry in the list.
5852 */
5853 rcu_read_lock();
5854
5855 list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
5856 throttled_csd_list) {
5857 list_del_init(&cursor->throttled_csd_list);
5858
5859 if (cfs_rq_throttled(cursor))
5860 unthrottle_cfs_rq(cursor);
5861 }
5862
5863 rcu_read_unlock();
5864
5865 rq_clock_stop_loop_update(rq);
5866 rq_unlock(rq, &rf);
5867 }
5868
__unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)5869 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5870 {
5871 struct rq *rq = rq_of(cfs_rq);
5872 bool first;
5873
5874 if (rq == this_rq()) {
5875 unthrottle_cfs_rq(cfs_rq);
5876 return;
5877 }
5878
5879 /* Already enqueued */
5880 if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))
5881 return;
5882
5883 first = list_empty(&rq->cfsb_csd_list);
5884 list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
5885 if (first)
5886 smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
5887 }
5888 #else
__unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)5889 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5890 {
5891 unthrottle_cfs_rq(cfs_rq);
5892 }
5893 #endif
5894
unthrottle_cfs_rq_async(struct cfs_rq * cfs_rq)5895 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5896 {
5897 lockdep_assert_rq_held(rq_of(cfs_rq));
5898
5899 if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) ||
5900 cfs_rq->runtime_remaining <= 0))
5901 return;
5902
5903 __unthrottle_cfs_rq_async(cfs_rq);
5904 }
5905
distribute_cfs_runtime(struct cfs_bandwidth * cfs_b)5906 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5907 {
5908 struct cfs_rq *local_unthrottle = NULL;
5909 int this_cpu = smp_processor_id();
5910 u64 runtime, remaining = 1;
5911 bool throttled = false;
5912 struct cfs_rq *cfs_rq;
5913 struct rq_flags rf;
5914 struct rq *rq;
5915
5916 rcu_read_lock();
5917 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5918 throttled_list) {
5919 rq = rq_of(cfs_rq);
5920
5921 if (!remaining) {
5922 throttled = true;
5923 break;
5924 }
5925
5926 rq_lock_irqsave(rq, &rf);
5927 if (!cfs_rq_throttled(cfs_rq))
5928 goto next;
5929
5930 #ifdef CONFIG_SMP
5931 /* Already queued for async unthrottle */
5932 if (!list_empty(&cfs_rq->throttled_csd_list))
5933 goto next;
5934 #endif
5935
5936 /* By the above checks, this should never be true */
5937 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5938
5939 raw_spin_lock(&cfs_b->lock);
5940 runtime = -cfs_rq->runtime_remaining + 1;
5941 if (runtime > cfs_b->runtime)
5942 runtime = cfs_b->runtime;
5943 cfs_b->runtime -= runtime;
5944 remaining = cfs_b->runtime;
5945 raw_spin_unlock(&cfs_b->lock);
5946
5947 cfs_rq->runtime_remaining += runtime;
5948
5949 /* we check whether we're throttled above */
5950 if (cfs_rq->runtime_remaining > 0) {
5951 if (cpu_of(rq) != this_cpu ||
5952 SCHED_WARN_ON(local_unthrottle))
5953 unthrottle_cfs_rq_async(cfs_rq);
5954 else
5955 local_unthrottle = cfs_rq;
5956 } else {
5957 throttled = true;
5958 }
5959
5960 next:
5961 rq_unlock_irqrestore(rq, &rf);
5962 }
5963 rcu_read_unlock();
5964
5965 if (local_unthrottle) {
5966 rq = cpu_rq(this_cpu);
5967 rq_lock_irqsave(rq, &rf);
5968 if (cfs_rq_throttled(local_unthrottle))
5969 unthrottle_cfs_rq(local_unthrottle);
5970 rq_unlock_irqrestore(rq, &rf);
5971 }
5972
5973 return throttled;
5974 }
5975
5976 /*
5977 * Responsible for refilling a task_group's bandwidth and unthrottling its
5978 * cfs_rqs as appropriate. If there has been no activity within the last
5979 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5980 * used to track this state.
5981 */
do_sched_cfs_period_timer(struct cfs_bandwidth * cfs_b,int overrun,unsigned long flags)5982 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5983 {
5984 int throttled;
5985
5986 /* no need to continue the timer with no bandwidth constraint */
5987 if (cfs_b->quota == RUNTIME_INF)
5988 goto out_deactivate;
5989
5990 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5991 cfs_b->nr_periods += overrun;
5992
5993 /* Refill extra burst quota even if cfs_b->idle */
5994 __refill_cfs_bandwidth_runtime(cfs_b);
5995
5996 /*
5997 * idle depends on !throttled (for the case of a large deficit), and if
5998 * we're going inactive then everything else can be deferred
5999 */
6000 if (cfs_b->idle && !throttled)
6001 goto out_deactivate;
6002
6003 if (!throttled) {
6004 /* mark as potentially idle for the upcoming period */
6005 cfs_b->idle = 1;
6006 return 0;
6007 }
6008
6009 /* account preceding periods in which throttling occurred */
6010 cfs_b->nr_throttled += overrun;
6011
6012 /*
6013 * This check is repeated as we release cfs_b->lock while we unthrottle.
6014 */
6015 while (throttled && cfs_b->runtime > 0) {
6016 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6017 /* we can't nest cfs_b->lock while distributing bandwidth */
6018 throttled = distribute_cfs_runtime(cfs_b);
6019 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6020 }
6021
6022 /*
6023 * While we are ensured activity in the period following an
6024 * unthrottle, this also covers the case in which the new bandwidth is
6025 * insufficient to cover the existing bandwidth deficit. (Forcing the
6026 * timer to remain active while there are any throttled entities.)
6027 */
6028 cfs_b->idle = 0;
6029
6030 return 0;
6031
6032 out_deactivate:
6033 return 1;
6034 }
6035
6036 /* a cfs_rq won't donate quota below this amount */
6037 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
6038 /* minimum remaining period time to redistribute slack quota */
6039 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
6040 /* how long we wait to gather additional slack before distributing */
6041 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
6042
6043 /*
6044 * Are we near the end of the current quota period?
6045 *
6046 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
6047 * hrtimer base being cleared by hrtimer_start. In the case of
6048 * migrate_hrtimers, base is never cleared, so we are fine.
6049 */
runtime_refresh_within(struct cfs_bandwidth * cfs_b,u64 min_expire)6050 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
6051 {
6052 struct hrtimer *refresh_timer = &cfs_b->period_timer;
6053 s64 remaining;
6054
6055 /* if the call-back is running a quota refresh is already occurring */
6056 if (hrtimer_callback_running(refresh_timer))
6057 return 1;
6058
6059 /* is a quota refresh about to occur? */
6060 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
6061 if (remaining < (s64)min_expire)
6062 return 1;
6063
6064 return 0;
6065 }
6066
start_cfs_slack_bandwidth(struct cfs_bandwidth * cfs_b)6067 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
6068 {
6069 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
6070
6071 /* if there's a quota refresh soon don't bother with slack */
6072 if (runtime_refresh_within(cfs_b, min_left))
6073 return;
6074
6075 /* don't push forwards an existing deferred unthrottle */
6076 if (cfs_b->slack_started)
6077 return;
6078 cfs_b->slack_started = true;
6079
6080 hrtimer_start(&cfs_b->slack_timer,
6081 ns_to_ktime(cfs_bandwidth_slack_period),
6082 HRTIMER_MODE_REL);
6083 }
6084
6085 /* we know any runtime found here is valid as update_curr() precedes return */
__return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6086 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6087 {
6088 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6089 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
6090
6091 if (slack_runtime <= 0)
6092 return;
6093
6094 raw_spin_lock(&cfs_b->lock);
6095 if (cfs_b->quota != RUNTIME_INF) {
6096 cfs_b->runtime += slack_runtime;
6097
6098 /* we are under rq->lock, defer unthrottling using a timer */
6099 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
6100 !list_empty(&cfs_b->throttled_cfs_rq))
6101 start_cfs_slack_bandwidth(cfs_b);
6102 }
6103 raw_spin_unlock(&cfs_b->lock);
6104
6105 /* even if it's not valid for return we don't want to try again */
6106 cfs_rq->runtime_remaining -= slack_runtime;
6107 }
6108
return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6109 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6110 {
6111 if (!cfs_bandwidth_used())
6112 return;
6113
6114 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
6115 return;
6116
6117 __return_cfs_rq_runtime(cfs_rq);
6118 }
6119
6120 /*
6121 * This is done with a timer (instead of inline with bandwidth return) since
6122 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6123 */
do_sched_cfs_slack_timer(struct cfs_bandwidth * cfs_b)6124 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
6125 {
6126 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
6127 unsigned long flags;
6128
6129 /* confirm we're still not at a refresh boundary */
6130 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6131 cfs_b->slack_started = false;
6132
6133 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
6134 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6135 return;
6136 }
6137
6138 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
6139 runtime = cfs_b->runtime;
6140
6141 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6142
6143 if (!runtime)
6144 return;
6145
6146 distribute_cfs_runtime(cfs_b);
6147 }
6148
6149 /*
6150 * When a group wakes up we want to make sure that its quota is not already
6151 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
6152 * runtime as update_curr() throttling can not trigger until it's on-rq.
6153 */
check_enqueue_throttle(struct cfs_rq * cfs_rq)6154 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
6155 {
6156 if (!cfs_bandwidth_used())
6157 return;
6158
6159 /* an active group must be handled by the update_curr()->put() path */
6160 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
6161 return;
6162
6163 /* ensure the group is not already throttled */
6164 if (cfs_rq_throttled(cfs_rq))
6165 return;
6166
6167 /* update runtime allocation */
6168 account_cfs_rq_runtime(cfs_rq, 0);
6169 if (cfs_rq->runtime_remaining <= 0)
6170 throttle_cfs_rq(cfs_rq);
6171 }
6172
sync_throttle(struct task_group * tg,int cpu)6173 static void sync_throttle(struct task_group *tg, int cpu)
6174 {
6175 struct cfs_rq *pcfs_rq, *cfs_rq;
6176
6177 if (!cfs_bandwidth_used())
6178 return;
6179
6180 if (!tg->parent)
6181 return;
6182
6183 cfs_rq = tg->cfs_rq[cpu];
6184 pcfs_rq = tg->parent->cfs_rq[cpu];
6185
6186 cfs_rq->throttle_count = pcfs_rq->throttle_count;
6187 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
6188 }
6189
6190 /* conditionally throttle active cfs_rq's from put_prev_entity() */
check_cfs_rq_runtime(struct cfs_rq * cfs_rq)6191 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6192 {
6193 if (!cfs_bandwidth_used())
6194 return false;
6195
6196 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
6197 return false;
6198
6199 /*
6200 * it's possible for a throttled entity to be forced into a running
6201 * state (e.g. set_curr_task), in this case we're finished.
6202 */
6203 if (cfs_rq_throttled(cfs_rq))
6204 return true;
6205
6206 return throttle_cfs_rq(cfs_rq);
6207 }
6208
sched_cfs_slack_timer(struct hrtimer * timer)6209 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
6210 {
6211 struct cfs_bandwidth *cfs_b =
6212 container_of(timer, struct cfs_bandwidth, slack_timer);
6213
6214 do_sched_cfs_slack_timer(cfs_b);
6215
6216 return HRTIMER_NORESTART;
6217 }
6218
6219 extern const u64 max_cfs_quota_period;
6220
sched_cfs_period_timer(struct hrtimer * timer)6221 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
6222 {
6223 struct cfs_bandwidth *cfs_b =
6224 container_of(timer, struct cfs_bandwidth, period_timer);
6225 unsigned long flags;
6226 int overrun;
6227 int idle = 0;
6228 int count = 0;
6229
6230 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6231 for (;;) {
6232 overrun = hrtimer_forward_now(timer, cfs_b->period);
6233 if (!overrun)
6234 break;
6235
6236 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
6237
6238 if (++count > 3) {
6239 u64 new, old = ktime_to_ns(cfs_b->period);
6240
6241 /*
6242 * Grow period by a factor of 2 to avoid losing precision.
6243 * Precision loss in the quota/period ratio can cause __cfs_schedulable
6244 * to fail.
6245 */
6246 new = old * 2;
6247 if (new < max_cfs_quota_period) {
6248 cfs_b->period = ns_to_ktime(new);
6249 cfs_b->quota *= 2;
6250 cfs_b->burst *= 2;
6251
6252 pr_warn_ratelimited(
6253 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6254 smp_processor_id(),
6255 div_u64(new, NSEC_PER_USEC),
6256 div_u64(cfs_b->quota, NSEC_PER_USEC));
6257 } else {
6258 pr_warn_ratelimited(
6259 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6260 smp_processor_id(),
6261 div_u64(old, NSEC_PER_USEC),
6262 div_u64(cfs_b->quota, NSEC_PER_USEC));
6263 }
6264
6265 /* reset count so we don't come right back in here */
6266 count = 0;
6267 }
6268 }
6269 if (idle)
6270 cfs_b->period_active = 0;
6271 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6272
6273 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
6274 }
6275
init_cfs_bandwidth(struct cfs_bandwidth * cfs_b,struct cfs_bandwidth * parent)6276 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent)
6277 {
6278 raw_spin_lock_init(&cfs_b->lock);
6279 cfs_b->runtime = 0;
6280 cfs_b->quota = RUNTIME_INF;
6281 cfs_b->period = ns_to_ktime(default_cfs_period());
6282 cfs_b->burst = 0;
6283 cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;
6284
6285 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
6286 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
6287 cfs_b->period_timer.function = sched_cfs_period_timer;
6288
6289 /* Add a random offset so that timers interleave */
6290 hrtimer_set_expires(&cfs_b->period_timer,
6291 get_random_u32_below(cfs_b->period));
6292 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6293 cfs_b->slack_timer.function = sched_cfs_slack_timer;
6294 cfs_b->slack_started = false;
6295 }
6296
init_cfs_rq_runtime(struct cfs_rq * cfs_rq)6297 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6298 {
6299 cfs_rq->runtime_enabled = 0;
6300 INIT_LIST_HEAD(&cfs_rq->throttled_list);
6301 #ifdef CONFIG_SMP
6302 INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
6303 #endif
6304 }
6305
start_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6306 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6307 {
6308 lockdep_assert_held(&cfs_b->lock);
6309
6310 if (cfs_b->period_active)
6311 return;
6312
6313 cfs_b->period_active = 1;
6314 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
6315 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
6316 }
6317
destroy_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6318 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6319 {
6320 int __maybe_unused i;
6321
6322 /* init_cfs_bandwidth() was not called */
6323 if (!cfs_b->throttled_cfs_rq.next)
6324 return;
6325
6326 hrtimer_cancel(&cfs_b->period_timer);
6327 hrtimer_cancel(&cfs_b->slack_timer);
6328
6329 /*
6330 * It is possible that we still have some cfs_rq's pending on a CSD
6331 * list, though this race is very rare. In order for this to occur, we
6332 * must have raced with the last task leaving the group while there
6333 * exist throttled cfs_rq(s), and the period_timer must have queued the
6334 * CSD item but the remote cpu has not yet processed it. To handle this,
6335 * we can simply flush all pending CSD work inline here. We're
6336 * guaranteed at this point that no additional cfs_rq of this group can
6337 * join a CSD list.
6338 */
6339 #ifdef CONFIG_SMP
6340 for_each_possible_cpu(i) {
6341 struct rq *rq = cpu_rq(i);
6342 unsigned long flags;
6343
6344 if (list_empty(&rq->cfsb_csd_list))
6345 continue;
6346
6347 local_irq_save(flags);
6348 __cfsb_csd_unthrottle(rq);
6349 local_irq_restore(flags);
6350 }
6351 #endif
6352 }
6353
6354 /*
6355 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
6356 *
6357 * The race is harmless, since modifying bandwidth settings of unhooked group
6358 * bits doesn't do much.
6359 */
6360
6361 /* cpu online callback */
update_runtime_enabled(struct rq * rq)6362 static void __maybe_unused update_runtime_enabled(struct rq *rq)
6363 {
6364 struct task_group *tg;
6365
6366 lockdep_assert_rq_held(rq);
6367
6368 rcu_read_lock();
6369 list_for_each_entry_rcu(tg, &task_groups, list) {
6370 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6371 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6372
6373 raw_spin_lock(&cfs_b->lock);
6374 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6375 raw_spin_unlock(&cfs_b->lock);
6376 }
6377 rcu_read_unlock();
6378 }
6379
6380 /* cpu offline callback */
unthrottle_offline_cfs_rqs(struct rq * rq)6381 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6382 {
6383 struct task_group *tg;
6384
6385 lockdep_assert_rq_held(rq);
6386
6387 /*
6388 * The rq clock has already been updated in the
6389 * set_rq_offline(), so we should skip updating
6390 * the rq clock again in unthrottle_cfs_rq().
6391 */
6392 rq_clock_start_loop_update(rq);
6393
6394 rcu_read_lock();
6395 list_for_each_entry_rcu(tg, &task_groups, list) {
6396 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6397
6398 if (!cfs_rq->runtime_enabled)
6399 continue;
6400
6401 /*
6402 * clock_task is not advancing so we just need to make sure
6403 * there's some valid quota amount
6404 */
6405 cfs_rq->runtime_remaining = 1;
6406 /*
6407 * Offline rq is schedulable till CPU is completely disabled
6408 * in take_cpu_down(), so we prevent new cfs throttling here.
6409 */
6410 cfs_rq->runtime_enabled = 0;
6411
6412 if (cfs_rq_throttled(cfs_rq))
6413 unthrottle_cfs_rq(cfs_rq);
6414 }
6415 rcu_read_unlock();
6416
6417 rq_clock_stop_loop_update(rq);
6418 }
6419
cfs_task_bw_constrained(struct task_struct * p)6420 bool cfs_task_bw_constrained(struct task_struct *p)
6421 {
6422 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6423
6424 if (!cfs_bandwidth_used())
6425 return false;
6426
6427 if (cfs_rq->runtime_enabled ||
6428 tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)
6429 return true;
6430
6431 return false;
6432 }
6433
6434 #ifdef CONFIG_NO_HZ_FULL
6435 /* called from pick_next_task_fair() */
sched_fair_update_stop_tick(struct rq * rq,struct task_struct * p)6436 static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
6437 {
6438 int cpu = cpu_of(rq);
6439
6440 if (!sched_feat(HZ_BW) || !cfs_bandwidth_used())
6441 return;
6442
6443 if (!tick_nohz_full_cpu(cpu))
6444 return;
6445
6446 if (rq->nr_running != 1)
6447 return;
6448
6449 /*
6450 * We know there is only one task runnable and we've just picked it. The
6451 * normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
6452 * be otherwise able to stop the tick. Just need to check if we are using
6453 * bandwidth control.
6454 */
6455 if (cfs_task_bw_constrained(p))
6456 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
6457 }
6458 #endif
6459
6460 #else /* CONFIG_CFS_BANDWIDTH */
6461
cfs_bandwidth_used(void)6462 static inline bool cfs_bandwidth_used(void)
6463 {
6464 return false;
6465 }
6466
account_cfs_rq_runtime(struct cfs_rq * cfs_rq,u64 delta_exec)6467 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
check_cfs_rq_runtime(struct cfs_rq * cfs_rq)6468 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
check_enqueue_throttle(struct cfs_rq * cfs_rq)6469 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
sync_throttle(struct task_group * tg,int cpu)6470 static inline void sync_throttle(struct task_group *tg, int cpu) {}
return_cfs_rq_runtime(struct cfs_rq * cfs_rq)6471 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6472
cfs_rq_throttled(struct cfs_rq * cfs_rq)6473 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6474 {
6475 return 0;
6476 }
6477
throttled_hierarchy(struct cfs_rq * cfs_rq)6478 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6479 {
6480 return 0;
6481 }
6482
throttled_lb_pair(struct task_group * tg,int src_cpu,int dest_cpu)6483 static inline int throttled_lb_pair(struct task_group *tg,
6484 int src_cpu, int dest_cpu)
6485 {
6486 return 0;
6487 }
6488
6489 #ifdef CONFIG_FAIR_GROUP_SCHED
init_cfs_bandwidth(struct cfs_bandwidth * cfs_b,struct cfs_bandwidth * parent)6490 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {}
init_cfs_rq_runtime(struct cfs_rq * cfs_rq)6491 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6492 #endif
6493
tg_cfs_bandwidth(struct task_group * tg)6494 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6495 {
6496 return NULL;
6497 }
destroy_cfs_bandwidth(struct cfs_bandwidth * cfs_b)6498 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
update_runtime_enabled(struct rq * rq)6499 static inline void update_runtime_enabled(struct rq *rq) {}
unthrottle_offline_cfs_rqs(struct rq * rq)6500 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6501 #ifdef CONFIG_CGROUP_SCHED
cfs_task_bw_constrained(struct task_struct * p)6502 bool cfs_task_bw_constrained(struct task_struct *p)
6503 {
6504 return false;
6505 }
6506 #endif
6507 #endif /* CONFIG_CFS_BANDWIDTH */
6508
6509 #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL)
sched_fair_update_stop_tick(struct rq * rq,struct task_struct * p)6510 static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {}
6511 #endif
6512
6513 /**************************************************
6514 * CFS operations on tasks:
6515 */
6516
6517 #ifdef CONFIG_SCHED_HRTICK
hrtick_start_fair(struct rq * rq,struct task_struct * p)6518 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6519 {
6520 struct sched_entity *se = &p->se;
6521
6522 SCHED_WARN_ON(task_rq(p) != rq);
6523
6524 if (rq->cfs.h_nr_running > 1) {
6525 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6526 u64 slice = se->slice;
6527 s64 delta = slice - ran;
6528
6529 if (delta < 0) {
6530 if (task_current(rq, p))
6531 resched_curr(rq);
6532 return;
6533 }
6534 hrtick_start(rq, delta);
6535 }
6536 }
6537
6538 /*
6539 * called from enqueue/dequeue and updates the hrtick when the
6540 * current task is from our class and nr_running is low enough
6541 * to matter.
6542 */
hrtick_update(struct rq * rq)6543 static void hrtick_update(struct rq *rq)
6544 {
6545 struct task_struct *curr = rq->curr;
6546
6547 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
6548 return;
6549
6550 hrtick_start_fair(rq, curr);
6551 }
6552 #else /* !CONFIG_SCHED_HRTICK */
6553 static inline void
hrtick_start_fair(struct rq * rq,struct task_struct * p)6554 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6555 {
6556 }
6557
hrtick_update(struct rq * rq)6558 static inline void hrtick_update(struct rq *rq)
6559 {
6560 }
6561 #endif
6562
6563 #ifdef CONFIG_SMP
cpu_overutilized(int cpu)6564 static inline bool cpu_overutilized(int cpu)
6565 {
6566 unsigned long rq_util_min, rq_util_max;
6567
6568 if (!sched_energy_enabled())
6569 return false;
6570
6571 rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6572 rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6573
6574 /* Return true only if the utilization doesn't fit CPU's capacity */
6575 return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6576 }
6577
set_rd_overutilized_status(struct root_domain * rd,unsigned int status)6578 static inline void set_rd_overutilized_status(struct root_domain *rd,
6579 unsigned int status)
6580 {
6581 if (!sched_energy_enabled())
6582 return;
6583
6584 WRITE_ONCE(rd->overutilized, status);
6585 trace_sched_overutilized_tp(rd, !!status);
6586 }
6587
check_update_overutilized_status(struct rq * rq)6588 static inline void check_update_overutilized_status(struct rq *rq)
6589 {
6590 /*
6591 * overutilized field is used for load balancing decisions only
6592 * if energy aware scheduler is being used
6593 */
6594 if (!sched_energy_enabled())
6595 return;
6596
6597 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu))
6598 set_rd_overutilized_status(rq->rd, SG_OVERUTILIZED);
6599 }
6600 #else
check_update_overutilized_status(struct rq * rq)6601 static inline void check_update_overutilized_status(struct rq *rq) { }
6602 #endif
6603
6604 /* Runqueue only has SCHED_IDLE tasks enqueued */
sched_idle_rq(struct rq * rq)6605 static int sched_idle_rq(struct rq *rq)
6606 {
6607 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6608 rq->nr_running);
6609 }
6610
6611 #ifdef CONFIG_SMP
sched_idle_cpu(int cpu)6612 static int sched_idle_cpu(int cpu)
6613 {
6614 return sched_idle_rq(cpu_rq(cpu));
6615 }
6616 #endif
6617
6618 /*
6619 * The enqueue_task method is called before nr_running is
6620 * increased. Here we update the fair scheduling stats and
6621 * then put the task into the rbtree:
6622 */
6623 static void
enqueue_task_fair(struct rq * rq,struct task_struct * p,int flags)6624 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6625 {
6626 struct cfs_rq *cfs_rq;
6627 struct sched_entity *se = &p->se;
6628 int idle_h_nr_running = task_has_idle_policy(p);
6629 int task_new = !(flags & ENQUEUE_WAKEUP);
6630
6631 /*
6632 * The code below (indirectly) updates schedutil which looks at
6633 * the cfs_rq utilization to select a frequency.
6634 * Let's add the task's estimated utilization to the cfs_rq's
6635 * estimated utilization, before we update schedutil.
6636 */
6637 util_est_enqueue(&rq->cfs, p);
6638
6639 /*
6640 * If in_iowait is set, the code below may not trigger any cpufreq
6641 * utilization updates, so do it here explicitly with the IOWAIT flag
6642 * passed.
6643 */
6644 if (p->in_iowait)
6645 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6646
6647 for_each_sched_entity(se) {
6648 if (se->on_rq)
6649 break;
6650 cfs_rq = cfs_rq_of(se);
6651 enqueue_entity(cfs_rq, se, flags);
6652
6653 cfs_rq->h_nr_running++;
6654 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6655
6656 if (cfs_rq_is_idle(cfs_rq))
6657 idle_h_nr_running = 1;
6658
6659 /* end evaluation on encountering a throttled cfs_rq */
6660 if (cfs_rq_throttled(cfs_rq))
6661 goto enqueue_throttle;
6662
6663 flags = ENQUEUE_WAKEUP;
6664 }
6665
6666 for_each_sched_entity(se) {
6667 cfs_rq = cfs_rq_of(se);
6668
6669 update_load_avg(cfs_rq, se, UPDATE_TG);
6670 se_update_runnable(se);
6671 update_cfs_group(se);
6672
6673 cfs_rq->h_nr_running++;
6674 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6675
6676 if (cfs_rq_is_idle(cfs_rq))
6677 idle_h_nr_running = 1;
6678
6679 /* end evaluation on encountering a throttled cfs_rq */
6680 if (cfs_rq_throttled(cfs_rq))
6681 goto enqueue_throttle;
6682 }
6683
6684 /* At this point se is NULL and we are at root level*/
6685 add_nr_running(rq, 1);
6686
6687 /*
6688 * Since new tasks are assigned an initial util_avg equal to
6689 * half of the spare capacity of their CPU, tiny tasks have the
6690 * ability to cross the overutilized threshold, which will
6691 * result in the load balancer ruining all the task placement
6692 * done by EAS. As a way to mitigate that effect, do not account
6693 * for the first enqueue operation of new tasks during the
6694 * overutilized flag detection.
6695 *
6696 * A better way of solving this problem would be to wait for
6697 * the PELT signals of tasks to converge before taking them
6698 * into account, but that is not straightforward to implement,
6699 * and the following generally works well enough in practice.
6700 */
6701 if (!task_new)
6702 check_update_overutilized_status(rq);
6703
6704 enqueue_throttle:
6705 assert_list_leaf_cfs_rq(rq);
6706
6707 hrtick_update(rq);
6708 }
6709
6710 static void set_next_buddy(struct sched_entity *se);
6711
6712 /*
6713 * The dequeue_task method is called before nr_running is
6714 * decreased. We remove the task from the rbtree and
6715 * update the fair scheduling stats:
6716 */
dequeue_task_fair(struct rq * rq,struct task_struct * p,int flags)6717 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6718 {
6719 struct cfs_rq *cfs_rq;
6720 struct sched_entity *se = &p->se;
6721 int task_sleep = flags & DEQUEUE_SLEEP;
6722 int idle_h_nr_running = task_has_idle_policy(p);
6723 bool was_sched_idle = sched_idle_rq(rq);
6724
6725 util_est_dequeue(&rq->cfs, p);
6726
6727 for_each_sched_entity(se) {
6728 cfs_rq = cfs_rq_of(se);
6729 dequeue_entity(cfs_rq, se, flags);
6730
6731 cfs_rq->h_nr_running--;
6732 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6733
6734 if (cfs_rq_is_idle(cfs_rq))
6735 idle_h_nr_running = 1;
6736
6737 /* end evaluation on encountering a throttled cfs_rq */
6738 if (cfs_rq_throttled(cfs_rq))
6739 goto dequeue_throttle;
6740
6741 /* Don't dequeue parent if it has other entities besides us */
6742 if (cfs_rq->load.weight) {
6743 /* Avoid re-evaluating load for this entity: */
6744 se = parent_entity(se);
6745 /*
6746 * Bias pick_next to pick a task from this cfs_rq, as
6747 * p is sleeping when it is within its sched_slice.
6748 */
6749 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6750 set_next_buddy(se);
6751 break;
6752 }
6753 flags |= DEQUEUE_SLEEP;
6754 }
6755
6756 for_each_sched_entity(se) {
6757 cfs_rq = cfs_rq_of(se);
6758
6759 update_load_avg(cfs_rq, se, UPDATE_TG);
6760 se_update_runnable(se);
6761 update_cfs_group(se);
6762
6763 cfs_rq->h_nr_running--;
6764 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6765
6766 if (cfs_rq_is_idle(cfs_rq))
6767 idle_h_nr_running = 1;
6768
6769 /* end evaluation on encountering a throttled cfs_rq */
6770 if (cfs_rq_throttled(cfs_rq))
6771 goto dequeue_throttle;
6772
6773 }
6774
6775 /* At this point se is NULL and we are at root level*/
6776 sub_nr_running(rq, 1);
6777
6778 /* balance early to pull high priority tasks */
6779 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6780 rq->next_balance = jiffies;
6781
6782 dequeue_throttle:
6783 util_est_update(&rq->cfs, p, task_sleep);
6784 hrtick_update(rq);
6785 }
6786
6787 #ifdef CONFIG_SMP
6788
6789 /* Working cpumask for: load_balance, load_balance_newidle. */
6790 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6791 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6792 static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask);
6793
6794 #ifdef CONFIG_NO_HZ_COMMON
6795
6796 static struct {
6797 cpumask_var_t idle_cpus_mask;
6798 atomic_t nr_cpus;
6799 int has_blocked; /* Idle CPUS has blocked load */
6800 int needs_update; /* Newly idle CPUs need their next_balance collated */
6801 unsigned long next_balance; /* in jiffy units */
6802 unsigned long next_blocked; /* Next update of blocked load in jiffies */
6803 } nohz ____cacheline_aligned;
6804
6805 #endif /* CONFIG_NO_HZ_COMMON */
6806
cpu_load(struct rq * rq)6807 static unsigned long cpu_load(struct rq *rq)
6808 {
6809 return cfs_rq_load_avg(&rq->cfs);
6810 }
6811
6812 /*
6813 * cpu_load_without - compute CPU load without any contributions from *p
6814 * @cpu: the CPU which load is requested
6815 * @p: the task which load should be discounted
6816 *
6817 * The load of a CPU is defined by the load of tasks currently enqueued on that
6818 * CPU as well as tasks which are currently sleeping after an execution on that
6819 * CPU.
6820 *
6821 * This method returns the load of the specified CPU by discounting the load of
6822 * the specified task, whenever the task is currently contributing to the CPU
6823 * load.
6824 */
cpu_load_without(struct rq * rq,struct task_struct * p)6825 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6826 {
6827 struct cfs_rq *cfs_rq;
6828 unsigned int load;
6829
6830 /* Task has no contribution or is new */
6831 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6832 return cpu_load(rq);
6833
6834 cfs_rq = &rq->cfs;
6835 load = READ_ONCE(cfs_rq->avg.load_avg);
6836
6837 /* Discount task's util from CPU's util */
6838 lsub_positive(&load, task_h_load(p));
6839
6840 return load;
6841 }
6842
cpu_runnable(struct rq * rq)6843 static unsigned long cpu_runnable(struct rq *rq)
6844 {
6845 return cfs_rq_runnable_avg(&rq->cfs);
6846 }
6847
cpu_runnable_without(struct rq * rq,struct task_struct * p)6848 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6849 {
6850 struct cfs_rq *cfs_rq;
6851 unsigned int runnable;
6852
6853 /* Task has no contribution or is new */
6854 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6855 return cpu_runnable(rq);
6856
6857 cfs_rq = &rq->cfs;
6858 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6859
6860 /* Discount task's runnable from CPU's runnable */
6861 lsub_positive(&runnable, p->se.avg.runnable_avg);
6862
6863 return runnable;
6864 }
6865
capacity_of(int cpu)6866 static unsigned long capacity_of(int cpu)
6867 {
6868 return cpu_rq(cpu)->cpu_capacity;
6869 }
6870
record_wakee(struct task_struct * p)6871 static void record_wakee(struct task_struct *p)
6872 {
6873 /*
6874 * Only decay a single time; tasks that have less then 1 wakeup per
6875 * jiffy will not have built up many flips.
6876 */
6877 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6878 current->wakee_flips >>= 1;
6879 current->wakee_flip_decay_ts = jiffies;
6880 }
6881
6882 if (current->last_wakee != p) {
6883 current->last_wakee = p;
6884 current->wakee_flips++;
6885 }
6886 }
6887
6888 /*
6889 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6890 *
6891 * A waker of many should wake a different task than the one last awakened
6892 * at a frequency roughly N times higher than one of its wakees.
6893 *
6894 * In order to determine whether we should let the load spread vs consolidating
6895 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6896 * partner, and a factor of lls_size higher frequency in the other.
6897 *
6898 * With both conditions met, we can be relatively sure that the relationship is
6899 * non-monogamous, with partner count exceeding socket size.
6900 *
6901 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6902 * whatever is irrelevant, spread criteria is apparent partner count exceeds
6903 * socket size.
6904 */
wake_wide(struct task_struct * p)6905 static int wake_wide(struct task_struct *p)
6906 {
6907 unsigned int master = current->wakee_flips;
6908 unsigned int slave = p->wakee_flips;
6909 int factor = __this_cpu_read(sd_llc_size);
6910
6911 if (master < slave)
6912 swap(master, slave);
6913 if (slave < factor || master < slave * factor)
6914 return 0;
6915 return 1;
6916 }
6917
6918 /*
6919 * The purpose of wake_affine() is to quickly determine on which CPU we can run
6920 * soonest. For the purpose of speed we only consider the waking and previous
6921 * CPU.
6922 *
6923 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
6924 * cache-affine and is (or will be) idle.
6925 *
6926 * wake_affine_weight() - considers the weight to reflect the average
6927 * scheduling latency of the CPUs. This seems to work
6928 * for the overloaded case.
6929 */
6930 static int
wake_affine_idle(int this_cpu,int prev_cpu,int sync)6931 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
6932 {
6933 /*
6934 * If this_cpu is idle, it implies the wakeup is from interrupt
6935 * context. Only allow the move if cache is shared. Otherwise an
6936 * interrupt intensive workload could force all tasks onto one
6937 * node depending on the IO topology or IRQ affinity settings.
6938 *
6939 * If the prev_cpu is idle and cache affine then avoid a migration.
6940 * There is no guarantee that the cache hot data from an interrupt
6941 * is more important than cache hot data on the prev_cpu and from
6942 * a cpufreq perspective, it's better to have higher utilisation
6943 * on one CPU.
6944 */
6945 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
6946 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
6947
6948 if (sync && cpu_rq(this_cpu)->nr_running == 1)
6949 return this_cpu;
6950
6951 if (available_idle_cpu(prev_cpu))
6952 return prev_cpu;
6953
6954 return nr_cpumask_bits;
6955 }
6956
6957 static int
wake_affine_weight(struct sched_domain * sd,struct task_struct * p,int this_cpu,int prev_cpu,int sync)6958 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
6959 int this_cpu, int prev_cpu, int sync)
6960 {
6961 s64 this_eff_load, prev_eff_load;
6962 unsigned long task_load;
6963
6964 this_eff_load = cpu_load(cpu_rq(this_cpu));
6965
6966 if (sync) {
6967 unsigned long current_load = task_h_load(current);
6968
6969 if (current_load > this_eff_load)
6970 return this_cpu;
6971
6972 this_eff_load -= current_load;
6973 }
6974
6975 task_load = task_h_load(p);
6976
6977 this_eff_load += task_load;
6978 if (sched_feat(WA_BIAS))
6979 this_eff_load *= 100;
6980 this_eff_load *= capacity_of(prev_cpu);
6981
6982 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
6983 prev_eff_load -= task_load;
6984 if (sched_feat(WA_BIAS))
6985 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
6986 prev_eff_load *= capacity_of(this_cpu);
6987
6988 /*
6989 * If sync, adjust the weight of prev_eff_load such that if
6990 * prev_eff == this_eff that select_idle_sibling() will consider
6991 * stacking the wakee on top of the waker if no other CPU is
6992 * idle.
6993 */
6994 if (sync)
6995 prev_eff_load += 1;
6996
6997 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
6998 }
6999
wake_affine(struct sched_domain * sd,struct task_struct * p,int this_cpu,int prev_cpu,int sync)7000 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
7001 int this_cpu, int prev_cpu, int sync)
7002 {
7003 int target = nr_cpumask_bits;
7004
7005 if (sched_feat(WA_IDLE))
7006 target = wake_affine_idle(this_cpu, prev_cpu, sync);
7007
7008 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
7009 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
7010
7011 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
7012 if (target != this_cpu)
7013 return prev_cpu;
7014
7015 schedstat_inc(sd->ttwu_move_affine);
7016 schedstat_inc(p->stats.nr_wakeups_affine);
7017 return target;
7018 }
7019
7020 static struct sched_group *
7021 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
7022
7023 /*
7024 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
7025 */
7026 static int
find_idlest_group_cpu(struct sched_group * group,struct task_struct * p,int this_cpu)7027 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
7028 {
7029 unsigned long load, min_load = ULONG_MAX;
7030 unsigned int min_exit_latency = UINT_MAX;
7031 u64 latest_idle_timestamp = 0;
7032 int least_loaded_cpu = this_cpu;
7033 int shallowest_idle_cpu = -1;
7034 int i;
7035
7036 /* Check if we have any choice: */
7037 if (group->group_weight == 1)
7038 return cpumask_first(sched_group_span(group));
7039
7040 /* Traverse only the allowed CPUs */
7041 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
7042 struct rq *rq = cpu_rq(i);
7043
7044 if (!sched_core_cookie_match(rq, p))
7045 continue;
7046
7047 if (sched_idle_cpu(i))
7048 return i;
7049
7050 if (available_idle_cpu(i)) {
7051 struct cpuidle_state *idle = idle_get_state(rq);
7052 if (idle && idle->exit_latency < min_exit_latency) {
7053 /*
7054 * We give priority to a CPU whose idle state
7055 * has the smallest exit latency irrespective
7056 * of any idle timestamp.
7057 */
7058 min_exit_latency = idle->exit_latency;
7059 latest_idle_timestamp = rq->idle_stamp;
7060 shallowest_idle_cpu = i;
7061 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
7062 rq->idle_stamp > latest_idle_timestamp) {
7063 /*
7064 * If equal or no active idle state, then
7065 * the most recently idled CPU might have
7066 * a warmer cache.
7067 */
7068 latest_idle_timestamp = rq->idle_stamp;
7069 shallowest_idle_cpu = i;
7070 }
7071 } else if (shallowest_idle_cpu == -1) {
7072 load = cpu_load(cpu_rq(i));
7073 if (load < min_load) {
7074 min_load = load;
7075 least_loaded_cpu = i;
7076 }
7077 }
7078 }
7079
7080 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
7081 }
7082
find_idlest_cpu(struct sched_domain * sd,struct task_struct * p,int cpu,int prev_cpu,int sd_flag)7083 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
7084 int cpu, int prev_cpu, int sd_flag)
7085 {
7086 int new_cpu = cpu;
7087
7088 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
7089 return prev_cpu;
7090
7091 /*
7092 * We need task's util for cpu_util_without, sync it up to
7093 * prev_cpu's last_update_time.
7094 */
7095 if (!(sd_flag & SD_BALANCE_FORK))
7096 sync_entity_load_avg(&p->se);
7097
7098 while (sd) {
7099 struct sched_group *group;
7100 struct sched_domain *tmp;
7101 int weight;
7102
7103 if (!(sd->flags & sd_flag)) {
7104 sd = sd->child;
7105 continue;
7106 }
7107
7108 group = find_idlest_group(sd, p, cpu);
7109 if (!group) {
7110 sd = sd->child;
7111 continue;
7112 }
7113
7114 new_cpu = find_idlest_group_cpu(group, p, cpu);
7115 if (new_cpu == cpu) {
7116 /* Now try balancing at a lower domain level of 'cpu': */
7117 sd = sd->child;
7118 continue;
7119 }
7120
7121 /* Now try balancing at a lower domain level of 'new_cpu': */
7122 cpu = new_cpu;
7123 weight = sd->span_weight;
7124 sd = NULL;
7125 for_each_domain(cpu, tmp) {
7126 if (weight <= tmp->span_weight)
7127 break;
7128 if (tmp->flags & sd_flag)
7129 sd = tmp;
7130 }
7131 }
7132
7133 return new_cpu;
7134 }
7135
__select_idle_cpu(int cpu,struct task_struct * p)7136 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
7137 {
7138 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
7139 sched_cpu_cookie_match(cpu_rq(cpu), p))
7140 return cpu;
7141
7142 return -1;
7143 }
7144
7145 #ifdef CONFIG_SCHED_SMT
7146 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7147 EXPORT_SYMBOL_GPL(sched_smt_present);
7148
set_idle_cores(int cpu,int val)7149 static inline void set_idle_cores(int cpu, int val)
7150 {
7151 struct sched_domain_shared *sds;
7152
7153 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7154 if (sds)
7155 WRITE_ONCE(sds->has_idle_cores, val);
7156 }
7157
test_idle_cores(int cpu)7158 static inline bool test_idle_cores(int cpu)
7159 {
7160 struct sched_domain_shared *sds;
7161
7162 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7163 if (sds)
7164 return READ_ONCE(sds->has_idle_cores);
7165
7166 return false;
7167 }
7168
7169 /*
7170 * Scans the local SMT mask to see if the entire core is idle, and records this
7171 * information in sd_llc_shared->has_idle_cores.
7172 *
7173 * Since SMT siblings share all cache levels, inspecting this limited remote
7174 * state should be fairly cheap.
7175 */
__update_idle_core(struct rq * rq)7176 void __update_idle_core(struct rq *rq)
7177 {
7178 int core = cpu_of(rq);
7179 int cpu;
7180
7181 rcu_read_lock();
7182 if (test_idle_cores(core))
7183 goto unlock;
7184
7185 for_each_cpu(cpu, cpu_smt_mask(core)) {
7186 if (cpu == core)
7187 continue;
7188
7189 if (!available_idle_cpu(cpu))
7190 goto unlock;
7191 }
7192
7193 set_idle_cores(core, 1);
7194 unlock:
7195 rcu_read_unlock();
7196 }
7197
7198 /*
7199 * Scan the entire LLC domain for idle cores; this dynamically switches off if
7200 * there are no idle cores left in the system; tracked through
7201 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7202 */
select_idle_core(struct task_struct * p,int core,struct cpumask * cpus,int * idle_cpu)7203 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7204 {
7205 bool idle = true;
7206 int cpu;
7207
7208 for_each_cpu(cpu, cpu_smt_mask(core)) {
7209 if (!available_idle_cpu(cpu)) {
7210 idle = false;
7211 if (*idle_cpu == -1) {
7212 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, cpus)) {
7213 *idle_cpu = cpu;
7214 break;
7215 }
7216 continue;
7217 }
7218 break;
7219 }
7220 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, cpus))
7221 *idle_cpu = cpu;
7222 }
7223
7224 if (idle)
7225 return core;
7226
7227 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
7228 return -1;
7229 }
7230
7231 /*
7232 * Scan the local SMT mask for idle CPUs.
7233 */
select_idle_smt(struct task_struct * p,struct sched_domain * sd,int target)7234 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7235 {
7236 int cpu;
7237
7238 for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
7239 if (cpu == target)
7240 continue;
7241 /*
7242 * Check if the CPU is in the LLC scheduling domain of @target.
7243 * Due to isolcpus, there is no guarantee that all the siblings are in the domain.
7244 */
7245 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7246 continue;
7247 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
7248 return cpu;
7249 }
7250
7251 return -1;
7252 }
7253
7254 #else /* CONFIG_SCHED_SMT */
7255
set_idle_cores(int cpu,int val)7256 static inline void set_idle_cores(int cpu, int val)
7257 {
7258 }
7259
test_idle_cores(int cpu)7260 static inline bool test_idle_cores(int cpu)
7261 {
7262 return false;
7263 }
7264
select_idle_core(struct task_struct * p,int core,struct cpumask * cpus,int * idle_cpu)7265 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7266 {
7267 return __select_idle_cpu(core, p);
7268 }
7269
select_idle_smt(struct task_struct * p,struct sched_domain * sd,int target)7270 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7271 {
7272 return -1;
7273 }
7274
7275 #endif /* CONFIG_SCHED_SMT */
7276
7277 /*
7278 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
7279 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7280 * average idle time for this rq (as found in rq->avg_idle).
7281 */
select_idle_cpu(struct task_struct * p,struct sched_domain * sd,bool has_idle_core,int target)7282 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
7283 {
7284 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7285 int i, cpu, idle_cpu = -1, nr = INT_MAX;
7286 struct sched_domain_shared *sd_share;
7287 struct rq *this_rq = this_rq();
7288 int this = smp_processor_id();
7289 struct sched_domain *this_sd = NULL;
7290 u64 time = 0;
7291
7292 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7293
7294 if (sched_feat(SIS_PROP) && !has_idle_core) {
7295 u64 avg_cost, avg_idle, span_avg;
7296 unsigned long now = jiffies;
7297
7298 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
7299 if (!this_sd)
7300 return -1;
7301
7302 /*
7303 * If we're busy, the assumption that the last idle period
7304 * predicts the future is flawed; age away the remaining
7305 * predicted idle time.
7306 */
7307 if (unlikely(this_rq->wake_stamp < now)) {
7308 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
7309 this_rq->wake_stamp++;
7310 this_rq->wake_avg_idle >>= 1;
7311 }
7312 }
7313
7314 avg_idle = this_rq->wake_avg_idle;
7315 avg_cost = this_sd->avg_scan_cost + 1;
7316
7317 span_avg = sd->span_weight * avg_idle;
7318 if (span_avg > 4*avg_cost)
7319 nr = div_u64(span_avg, avg_cost);
7320 else
7321 nr = 4;
7322
7323 time = cpu_clock(this);
7324 }
7325
7326 if (sched_feat(SIS_UTIL)) {
7327 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
7328 if (sd_share) {
7329 /* because !--nr is the condition to stop scan */
7330 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
7331 /* overloaded LLC is unlikely to have idle cpu/core */
7332 if (nr == 1)
7333 return -1;
7334 }
7335 }
7336
7337 for_each_cpu_wrap(cpu, cpus, target + 1) {
7338 if (has_idle_core) {
7339 i = select_idle_core(p, cpu, cpus, &idle_cpu);
7340 if ((unsigned int)i < nr_cpumask_bits)
7341 return i;
7342
7343 } else {
7344 if (!--nr)
7345 return -1;
7346 idle_cpu = __select_idle_cpu(cpu, p);
7347 if ((unsigned int)idle_cpu < nr_cpumask_bits)
7348 break;
7349 }
7350 }
7351
7352 if (has_idle_core)
7353 set_idle_cores(target, false);
7354
7355 if (sched_feat(SIS_PROP) && this_sd && !has_idle_core) {
7356 time = cpu_clock(this) - time;
7357
7358 /*
7359 * Account for the scan cost of wakeups against the average
7360 * idle time.
7361 */
7362 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
7363
7364 update_avg(&this_sd->avg_scan_cost, time);
7365 }
7366
7367 return idle_cpu;
7368 }
7369
7370 /*
7371 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
7372 * the task fits. If no CPU is big enough, but there are idle ones, try to
7373 * maximize capacity.
7374 */
7375 static int
select_idle_capacity(struct task_struct * p,struct sched_domain * sd,int target)7376 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
7377 {
7378 unsigned long task_util, util_min, util_max, best_cap = 0;
7379 int fits, best_fits = 0;
7380 int cpu, best_cpu = -1;
7381 struct cpumask *cpus;
7382
7383 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7384 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7385
7386 task_util = task_util_est(p);
7387 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7388 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7389
7390 for_each_cpu_wrap(cpu, cpus, target) {
7391 unsigned long cpu_cap = capacity_of(cpu);
7392
7393 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
7394 continue;
7395
7396 fits = util_fits_cpu(task_util, util_min, util_max, cpu);
7397
7398 /* This CPU fits with all requirements */
7399 if (fits > 0)
7400 return cpu;
7401 /*
7402 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
7403 * Look for the CPU with best capacity.
7404 */
7405 else if (fits < 0)
7406 cpu_cap = capacity_orig_of(cpu) - thermal_load_avg(cpu_rq(cpu));
7407
7408 /*
7409 * First, select CPU which fits better (-1 being better than 0).
7410 * Then, select the one with best capacity at same level.
7411 */
7412 if ((fits < best_fits) ||
7413 ((fits == best_fits) && (cpu_cap > best_cap))) {
7414 best_cap = cpu_cap;
7415 best_cpu = cpu;
7416 best_fits = fits;
7417 }
7418 }
7419
7420 return best_cpu;
7421 }
7422
asym_fits_cpu(unsigned long util,unsigned long util_min,unsigned long util_max,int cpu)7423 static inline bool asym_fits_cpu(unsigned long util,
7424 unsigned long util_min,
7425 unsigned long util_max,
7426 int cpu)
7427 {
7428 if (sched_asym_cpucap_active())
7429 /*
7430 * Return true only if the cpu fully fits the task requirements
7431 * which include the utilization and the performance hints.
7432 */
7433 return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7434
7435 return true;
7436 }
7437
7438 /*
7439 * Try and locate an idle core/thread in the LLC cache domain.
7440 */
select_idle_sibling(struct task_struct * p,int prev,int target)7441 static int select_idle_sibling(struct task_struct *p, int prev, int target)
7442 {
7443 bool has_idle_core = false;
7444 struct sched_domain *sd;
7445 unsigned long task_util, util_min, util_max;
7446 int i, recent_used_cpu;
7447
7448 /*
7449 * On asymmetric system, update task utilization because we will check
7450 * that the task fits with cpu's capacity.
7451 */
7452 if (sched_asym_cpucap_active()) {
7453 sync_entity_load_avg(&p->se);
7454 task_util = task_util_est(p);
7455 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7456 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7457 }
7458
7459 /*
7460 * per-cpu select_rq_mask usage
7461 */
7462 lockdep_assert_irqs_disabled();
7463
7464 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7465 asym_fits_cpu(task_util, util_min, util_max, target))
7466 return target;
7467
7468 /*
7469 * If the previous CPU is cache affine and idle, don't be stupid:
7470 */
7471 if (prev != target && cpus_share_cache(prev, target) &&
7472 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7473 asym_fits_cpu(task_util, util_min, util_max, prev))
7474 return prev;
7475
7476 /*
7477 * Allow a per-cpu kthread to stack with the wakee if the
7478 * kworker thread and the tasks previous CPUs are the same.
7479 * The assumption is that the wakee queued work for the
7480 * per-cpu kthread that is now complete and the wakeup is
7481 * essentially a sync wakeup. An obvious example of this
7482 * pattern is IO completions.
7483 */
7484 if (is_per_cpu_kthread(current) &&
7485 in_task() &&
7486 prev == smp_processor_id() &&
7487 this_rq()->nr_running <= 1 &&
7488 asym_fits_cpu(task_util, util_min, util_max, prev)) {
7489 return prev;
7490 }
7491
7492 /* Check a recently used CPU as a potential idle candidate: */
7493 recent_used_cpu = p->recent_used_cpu;
7494 p->recent_used_cpu = prev;
7495 if (recent_used_cpu != prev &&
7496 recent_used_cpu != target &&
7497 cpus_share_cache(recent_used_cpu, target) &&
7498 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7499 cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&
7500 asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7501 return recent_used_cpu;
7502 }
7503
7504 /*
7505 * For asymmetric CPU capacity systems, our domain of interest is
7506 * sd_asym_cpucapacity rather than sd_llc.
7507 */
7508 if (sched_asym_cpucap_active()) {
7509 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7510 /*
7511 * On an asymmetric CPU capacity system where an exclusive
7512 * cpuset defines a symmetric island (i.e. one unique
7513 * capacity_orig value through the cpuset), the key will be set
7514 * but the CPUs within that cpuset will not have a domain with
7515 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7516 * capacity path.
7517 */
7518 if (sd) {
7519 i = select_idle_capacity(p, sd, target);
7520 return ((unsigned)i < nr_cpumask_bits) ? i : target;
7521 }
7522 }
7523
7524 sd = rcu_dereference(per_cpu(sd_llc, target));
7525 if (!sd)
7526 return target;
7527
7528 if (sched_smt_active()) {
7529 has_idle_core = test_idle_cores(target);
7530
7531 if (!has_idle_core && cpus_share_cache(prev, target)) {
7532 i = select_idle_smt(p, sd, prev);
7533 if ((unsigned int)i < nr_cpumask_bits)
7534 return i;
7535 }
7536 }
7537
7538 i = select_idle_cpu(p, sd, has_idle_core, target);
7539 if ((unsigned)i < nr_cpumask_bits)
7540 return i;
7541
7542 return target;
7543 }
7544
7545 /**
7546 * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7547 * @cpu: the CPU to get the utilization for
7548 * @p: task for which the CPU utilization should be predicted or NULL
7549 * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7550 * @boost: 1 to enable boosting, otherwise 0
7551 *
7552 * The unit of the return value must be the same as the one of CPU capacity
7553 * so that CPU utilization can be compared with CPU capacity.
7554 *
7555 * CPU utilization is the sum of running time of runnable tasks plus the
7556 * recent utilization of currently non-runnable tasks on that CPU.
7557 * It represents the amount of CPU capacity currently used by CFS tasks in
7558 * the range [0..max CPU capacity] with max CPU capacity being the CPU
7559 * capacity at f_max.
7560 *
7561 * The estimated CPU utilization is defined as the maximum between CPU
7562 * utilization and sum of the estimated utilization of the currently
7563 * runnable tasks on that CPU. It preserves a utilization "snapshot" of
7564 * previously-executed tasks, which helps better deduce how busy a CPU will
7565 * be when a long-sleeping task wakes up. The contribution to CPU utilization
7566 * of such a task would be significantly decayed at this point of time.
7567 *
7568 * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization).
7569 * CPU contention for CFS tasks can be detected by CPU runnable > CPU
7570 * utilization. Boosting is implemented in cpu_util() so that internal
7571 * users (e.g. EAS) can use it next to external users (e.g. schedutil),
7572 * latter via cpu_util_cfs_boost().
7573 *
7574 * CPU utilization can be higher than the current CPU capacity
7575 * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because
7576 * of rounding errors as well as task migrations or wakeups of new tasks.
7577 * CPU utilization has to be capped to fit into the [0..max CPU capacity]
7578 * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%)
7579 * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7580 * capacity. CPU utilization is allowed to overshoot current CPU capacity
7581 * though since this is useful for predicting the CPU capacity required
7582 * after task migrations (scheduler-driven DVFS).
7583 *
7584 * Return: (Boosted) (estimated) utilization for the specified CPU.
7585 */
7586 static unsigned long
cpu_util(int cpu,struct task_struct * p,int dst_cpu,int boost)7587 cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
7588 {
7589 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
7590 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
7591 unsigned long runnable;
7592
7593 if (boost) {
7594 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
7595 util = max(util, runnable);
7596 }
7597
7598 /*
7599 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
7600 * contribution. If @p migrates from another CPU to @cpu add its
7601 * contribution. In all the other cases @cpu is not impacted by the
7602 * migration so its util_avg is already correct.
7603 */
7604 if (p && task_cpu(p) == cpu && dst_cpu != cpu)
7605 lsub_positive(&util, task_util(p));
7606 else if (p && task_cpu(p) != cpu && dst_cpu == cpu)
7607 util += task_util(p);
7608
7609 if (sched_feat(UTIL_EST)) {
7610 unsigned long util_est;
7611
7612 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
7613
7614 /*
7615 * During wake-up @p isn't enqueued yet and doesn't contribute
7616 * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
7617 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
7618 * has been enqueued.
7619 *
7620 * During exec (@dst_cpu = -1) @p is enqueued and does
7621 * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
7622 * Remove it to "simulate" cpu_util without @p's contribution.
7623 *
7624 * Despite the task_on_rq_queued(@p) check there is still a
7625 * small window for a possible race when an exec
7626 * select_task_rq_fair() races with LB's detach_task().
7627 *
7628 * detach_task()
7629 * deactivate_task()
7630 * p->on_rq = TASK_ON_RQ_MIGRATING;
7631 * -------------------------------- A
7632 * dequeue_task() \
7633 * dequeue_task_fair() + Race Time
7634 * util_est_dequeue() /
7635 * -------------------------------- B
7636 *
7637 * The additional check "current == p" is required to further
7638 * reduce the race window.
7639 */
7640 if (dst_cpu == cpu)
7641 util_est += _task_util_est(p);
7642 else if (p && unlikely(task_on_rq_queued(p) || current == p))
7643 lsub_positive(&util_est, _task_util_est(p));
7644
7645 util = max(util, util_est);
7646 }
7647
7648 return min(util, capacity_orig_of(cpu));
7649 }
7650
cpu_util_cfs(int cpu)7651 unsigned long cpu_util_cfs(int cpu)
7652 {
7653 return cpu_util(cpu, NULL, -1, 0);
7654 }
7655
cpu_util_cfs_boost(int cpu)7656 unsigned long cpu_util_cfs_boost(int cpu)
7657 {
7658 return cpu_util(cpu, NULL, -1, 1);
7659 }
7660
7661 /*
7662 * cpu_util_without: compute cpu utilization without any contributions from *p
7663 * @cpu: the CPU which utilization is requested
7664 * @p: the task which utilization should be discounted
7665 *
7666 * The utilization of a CPU is defined by the utilization of tasks currently
7667 * enqueued on that CPU as well as tasks which are currently sleeping after an
7668 * execution on that CPU.
7669 *
7670 * This method returns the utilization of the specified CPU by discounting the
7671 * utilization of the specified task, whenever the task is currently
7672 * contributing to the CPU utilization.
7673 */
cpu_util_without(int cpu,struct task_struct * p)7674 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7675 {
7676 /* Task has no contribution or is new */
7677 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7678 p = NULL;
7679
7680 return cpu_util(cpu, p, -1, 0);
7681 }
7682
7683 /*
7684 * energy_env - Utilization landscape for energy estimation.
7685 * @task_busy_time: Utilization contribution by the task for which we test the
7686 * placement. Given by eenv_task_busy_time().
7687 * @pd_busy_time: Utilization of the whole perf domain without the task
7688 * contribution. Given by eenv_pd_busy_time().
7689 * @cpu_cap: Maximum CPU capacity for the perf domain.
7690 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7691 */
7692 struct energy_env {
7693 unsigned long task_busy_time;
7694 unsigned long pd_busy_time;
7695 unsigned long cpu_cap;
7696 unsigned long pd_cap;
7697 };
7698
7699 /*
7700 * Compute the task busy time for compute_energy(). This time cannot be
7701 * injected directly into effective_cpu_util() because of the IRQ scaling.
7702 * The latter only makes sense with the most recent CPUs where the task has
7703 * run.
7704 */
eenv_task_busy_time(struct energy_env * eenv,struct task_struct * p,int prev_cpu)7705 static inline void eenv_task_busy_time(struct energy_env *eenv,
7706 struct task_struct *p, int prev_cpu)
7707 {
7708 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7709 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7710
7711 if (unlikely(irq >= max_cap))
7712 busy_time = max_cap;
7713 else
7714 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7715
7716 eenv->task_busy_time = busy_time;
7717 }
7718
7719 /*
7720 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7721 * utilization for each @pd_cpus, it however doesn't take into account
7722 * clamping since the ratio (utilization / cpu_capacity) is already enough to
7723 * scale the EM reported power consumption at the (eventually clamped)
7724 * cpu_capacity.
7725 *
7726 * The contribution of the task @p for which we want to estimate the
7727 * energy cost is removed (by cpu_util()) and must be calculated
7728 * separately (see eenv_task_busy_time). This ensures:
7729 *
7730 * - A stable PD utilization, no matter which CPU of that PD we want to place
7731 * the task on.
7732 *
7733 * - A fair comparison between CPUs as the task contribution (task_util())
7734 * will always be the same no matter which CPU utilization we rely on
7735 * (util_avg or util_est).
7736 *
7737 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7738 * exceed @eenv->pd_cap.
7739 */
eenv_pd_busy_time(struct energy_env * eenv,struct cpumask * pd_cpus,struct task_struct * p)7740 static inline void eenv_pd_busy_time(struct energy_env *eenv,
7741 struct cpumask *pd_cpus,
7742 struct task_struct *p)
7743 {
7744 unsigned long busy_time = 0;
7745 int cpu;
7746
7747 for_each_cpu(cpu, pd_cpus) {
7748 unsigned long util = cpu_util(cpu, p, -1, 0);
7749
7750 busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
7751 }
7752
7753 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7754 }
7755
7756 /*
7757 * Compute the maximum utilization for compute_energy() when the task @p
7758 * is placed on the cpu @dst_cpu.
7759 *
7760 * Returns the maximum utilization among @eenv->cpus. This utilization can't
7761 * exceed @eenv->cpu_cap.
7762 */
7763 static inline unsigned long
eenv_pd_max_util(struct energy_env * eenv,struct cpumask * pd_cpus,struct task_struct * p,int dst_cpu)7764 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7765 struct task_struct *p, int dst_cpu)
7766 {
7767 unsigned long max_util = 0;
7768 int cpu;
7769
7770 for_each_cpu(cpu, pd_cpus) {
7771 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7772 unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
7773 unsigned long eff_util;
7774
7775 /*
7776 * Performance domain frequency: utilization clamping
7777 * must be considered since it affects the selection
7778 * of the performance domain frequency.
7779 * NOTE: in case RT tasks are running, by default the
7780 * FREQUENCY_UTIL's utilization can be max OPP.
7781 */
7782 eff_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
7783 max_util = max(max_util, eff_util);
7784 }
7785
7786 return min(max_util, eenv->cpu_cap);
7787 }
7788
7789 /*
7790 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7791 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7792 * contribution is ignored.
7793 */
7794 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)7795 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7796 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7797 {
7798 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7799 unsigned long busy_time = eenv->pd_busy_time;
7800
7801 if (dst_cpu >= 0)
7802 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7803
7804 return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7805 }
7806
7807 /*
7808 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7809 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7810 * spare capacity in each performance domain and uses it as a potential
7811 * candidate to execute the task. Then, it uses the Energy Model to figure
7812 * out which of the CPU candidates is the most energy-efficient.
7813 *
7814 * The rationale for this heuristic is as follows. In a performance domain,
7815 * all the most energy efficient CPU candidates (according to the Energy
7816 * Model) are those for which we'll request a low frequency. When there are
7817 * several CPUs for which the frequency request will be the same, we don't
7818 * have enough data to break the tie between them, because the Energy Model
7819 * only includes active power costs. With this model, if we assume that
7820 * frequency requests follow utilization (e.g. using schedutil), the CPU with
7821 * the maximum spare capacity in a performance domain is guaranteed to be among
7822 * the best candidates of the performance domain.
7823 *
7824 * In practice, it could be preferable from an energy standpoint to pack
7825 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7826 * but that could also hurt our chances to go cluster idle, and we have no
7827 * ways to tell with the current Energy Model if this is actually a good
7828 * idea or not. So, find_energy_efficient_cpu() basically favors
7829 * cluster-packing, and spreading inside a cluster. That should at least be
7830 * a good thing for latency, and this is consistent with the idea that most
7831 * of the energy savings of EAS come from the asymmetry of the system, and
7832 * not so much from breaking the tie between identical CPUs. That's also the
7833 * reason why EAS is enabled in the topology code only for systems where
7834 * SD_ASYM_CPUCAPACITY is set.
7835 *
7836 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7837 * they don't have any useful utilization data yet and it's not possible to
7838 * forecast their impact on energy consumption. Consequently, they will be
7839 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7840 * to be energy-inefficient in some use-cases. The alternative would be to
7841 * bias new tasks towards specific types of CPUs first, or to try to infer
7842 * their util_avg from the parent task, but those heuristics could hurt
7843 * other use-cases too. So, until someone finds a better way to solve this,
7844 * let's keep things simple by re-using the existing slow path.
7845 */
find_energy_efficient_cpu(struct task_struct * p,int prev_cpu)7846 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7847 {
7848 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7849 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7850 unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7851 unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7852 struct root_domain *rd = this_rq()->rd;
7853 int cpu, best_energy_cpu, target = -1;
7854 int prev_fits = -1, best_fits = -1;
7855 unsigned long best_thermal_cap = 0;
7856 unsigned long prev_thermal_cap = 0;
7857 struct sched_domain *sd;
7858 struct perf_domain *pd;
7859 struct energy_env eenv;
7860
7861 rcu_read_lock();
7862 pd = rcu_dereference(rd->pd);
7863 if (!pd || READ_ONCE(rd->overutilized))
7864 goto unlock;
7865
7866 /*
7867 * Energy-aware wake-up happens on the lowest sched_domain starting
7868 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7869 */
7870 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7871 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7872 sd = sd->parent;
7873 if (!sd)
7874 goto unlock;
7875
7876 target = prev_cpu;
7877
7878 sync_entity_load_avg(&p->se);
7879 if (!task_util_est(p) && p_util_min == 0)
7880 goto unlock;
7881
7882 eenv_task_busy_time(&eenv, p, prev_cpu);
7883
7884 for (; pd; pd = pd->next) {
7885 unsigned long util_min = p_util_min, util_max = p_util_max;
7886 unsigned long cpu_cap, cpu_thermal_cap, util;
7887 long prev_spare_cap = -1, max_spare_cap = -1;
7888 unsigned long rq_util_min, rq_util_max;
7889 unsigned long cur_delta, base_energy;
7890 int max_spare_cap_cpu = -1;
7891 int fits, max_fits = -1;
7892
7893 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
7894
7895 if (cpumask_empty(cpus))
7896 continue;
7897
7898 /* Account thermal pressure for the energy estimation */
7899 cpu = cpumask_first(cpus);
7900 cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
7901 cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
7902
7903 eenv.cpu_cap = cpu_thermal_cap;
7904 eenv.pd_cap = 0;
7905
7906 for_each_cpu(cpu, cpus) {
7907 struct rq *rq = cpu_rq(cpu);
7908
7909 eenv.pd_cap += cpu_thermal_cap;
7910
7911 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7912 continue;
7913
7914 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
7915 continue;
7916
7917 util = cpu_util(cpu, p, cpu, 0);
7918 cpu_cap = capacity_of(cpu);
7919
7920 /*
7921 * Skip CPUs that cannot satisfy the capacity request.
7922 * IOW, placing the task there would make the CPU
7923 * overutilized. Take uclamp into account to see how
7924 * much capacity we can get out of the CPU; this is
7925 * aligned with sched_cpu_util().
7926 */
7927 if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
7928 /*
7929 * Open code uclamp_rq_util_with() except for
7930 * the clamp() part. Ie: apply max aggregation
7931 * only. util_fits_cpu() logic requires to
7932 * operate on non clamped util but must use the
7933 * max-aggregated uclamp_{min, max}.
7934 */
7935 rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
7936 rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
7937
7938 util_min = max(rq_util_min, p_util_min);
7939 util_max = max(rq_util_max, p_util_max);
7940 }
7941
7942 fits = util_fits_cpu(util, util_min, util_max, cpu);
7943 if (!fits)
7944 continue;
7945
7946 lsub_positive(&cpu_cap, util);
7947
7948 if (cpu == prev_cpu) {
7949 /* Always use prev_cpu as a candidate. */
7950 prev_spare_cap = cpu_cap;
7951 prev_fits = fits;
7952 } else if ((fits > max_fits) ||
7953 ((fits == max_fits) && ((long)cpu_cap > max_spare_cap))) {
7954 /*
7955 * Find the CPU with the maximum spare capacity
7956 * among the remaining CPUs in the performance
7957 * domain.
7958 */
7959 max_spare_cap = cpu_cap;
7960 max_spare_cap_cpu = cpu;
7961 max_fits = fits;
7962 }
7963 }
7964
7965 if (max_spare_cap_cpu < 0 && prev_spare_cap < 0)
7966 continue;
7967
7968 eenv_pd_busy_time(&eenv, cpus, p);
7969 /* Compute the 'base' energy of the pd, without @p */
7970 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
7971
7972 /* Evaluate the energy impact of using prev_cpu. */
7973 if (prev_spare_cap > -1) {
7974 prev_delta = compute_energy(&eenv, pd, cpus, p,
7975 prev_cpu);
7976 /* CPU utilization has changed */
7977 if (prev_delta < base_energy)
7978 goto unlock;
7979 prev_delta -= base_energy;
7980 prev_thermal_cap = cpu_thermal_cap;
7981 best_delta = min(best_delta, prev_delta);
7982 }
7983
7984 /* Evaluate the energy impact of using max_spare_cap_cpu. */
7985 if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
7986 /* Current best energy cpu fits better */
7987 if (max_fits < best_fits)
7988 continue;
7989
7990 /*
7991 * Both don't fit performance hint (i.e. uclamp_min)
7992 * but best energy cpu has better capacity.
7993 */
7994 if ((max_fits < 0) &&
7995 (cpu_thermal_cap <= best_thermal_cap))
7996 continue;
7997
7998 cur_delta = compute_energy(&eenv, pd, cpus, p,
7999 max_spare_cap_cpu);
8000 /* CPU utilization has changed */
8001 if (cur_delta < base_energy)
8002 goto unlock;
8003 cur_delta -= base_energy;
8004
8005 /*
8006 * Both fit for the task but best energy cpu has lower
8007 * energy impact.
8008 */
8009 if ((max_fits > 0) && (best_fits > 0) &&
8010 (cur_delta >= best_delta))
8011 continue;
8012
8013 best_delta = cur_delta;
8014 best_energy_cpu = max_spare_cap_cpu;
8015 best_fits = max_fits;
8016 best_thermal_cap = cpu_thermal_cap;
8017 }
8018 }
8019 rcu_read_unlock();
8020
8021 if ((best_fits > prev_fits) ||
8022 ((best_fits > 0) && (best_delta < prev_delta)) ||
8023 ((best_fits < 0) && (best_thermal_cap > prev_thermal_cap)))
8024 target = best_energy_cpu;
8025
8026 return target;
8027
8028 unlock:
8029 rcu_read_unlock();
8030
8031 return target;
8032 }
8033
8034 /*
8035 * select_task_rq_fair: Select target runqueue for the waking task in domains
8036 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
8037 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
8038 *
8039 * Balances load by selecting the idlest CPU in the idlest group, or under
8040 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
8041 *
8042 * Returns the target CPU number.
8043 */
8044 static int
select_task_rq_fair(struct task_struct * p,int prev_cpu,int wake_flags)8045 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
8046 {
8047 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
8048 struct sched_domain *tmp, *sd = NULL;
8049 int cpu = smp_processor_id();
8050 int new_cpu = prev_cpu;
8051 int want_affine = 0;
8052 /* SD_flags and WF_flags share the first nibble */
8053 int sd_flag = wake_flags & 0xF;
8054
8055 /*
8056 * required for stable ->cpus_allowed
8057 */
8058 lockdep_assert_held(&p->pi_lock);
8059 if (wake_flags & WF_TTWU) {
8060 record_wakee(p);
8061
8062 if ((wake_flags & WF_CURRENT_CPU) &&
8063 cpumask_test_cpu(cpu, p->cpus_ptr))
8064 return cpu;
8065
8066 if (sched_energy_enabled()) {
8067 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
8068 if (new_cpu >= 0)
8069 return new_cpu;
8070 new_cpu = prev_cpu;
8071 }
8072
8073 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
8074 }
8075
8076 rcu_read_lock();
8077 for_each_domain(cpu, tmp) {
8078 /*
8079 * If both 'cpu' and 'prev_cpu' are part of this domain,
8080 * cpu is a valid SD_WAKE_AFFINE target.
8081 */
8082 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
8083 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
8084 if (cpu != prev_cpu)
8085 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
8086
8087 sd = NULL; /* Prefer wake_affine over balance flags */
8088 break;
8089 }
8090
8091 /*
8092 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
8093 * usually do not have SD_BALANCE_WAKE set. That means wakeup
8094 * will usually go to the fast path.
8095 */
8096 if (tmp->flags & sd_flag)
8097 sd = tmp;
8098 else if (!want_affine)
8099 break;
8100 }
8101
8102 if (unlikely(sd)) {
8103 /* Slow path */
8104 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
8105 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
8106 /* Fast path */
8107 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
8108 }
8109 rcu_read_unlock();
8110
8111 return new_cpu;
8112 }
8113
8114 /*
8115 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
8116 * cfs_rq_of(p) references at time of call are still valid and identify the
8117 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
8118 */
migrate_task_rq_fair(struct task_struct * p,int new_cpu)8119 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
8120 {
8121 struct sched_entity *se = &p->se;
8122
8123 if (!task_on_rq_migrating(p)) {
8124 remove_entity_load_avg(se);
8125
8126 /*
8127 * Here, the task's PELT values have been updated according to
8128 * the current rq's clock. But if that clock hasn't been
8129 * updated in a while, a substantial idle time will be missed,
8130 * leading to an inflation after wake-up on the new rq.
8131 *
8132 * Estimate the missing time from the cfs_rq last_update_time
8133 * and update sched_avg to improve the PELT continuity after
8134 * migration.
8135 */
8136 migrate_se_pelt_lag(se);
8137 }
8138
8139 /* Tell new CPU we are migrated */
8140 se->avg.last_update_time = 0;
8141
8142 update_scan_period(p, new_cpu);
8143 }
8144
task_dead_fair(struct task_struct * p)8145 static void task_dead_fair(struct task_struct *p)
8146 {
8147 remove_entity_load_avg(&p->se);
8148 }
8149
8150 static int
balance_fair(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)8151 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8152 {
8153 if (rq->nr_running)
8154 return 1;
8155
8156 return newidle_balance(rq, rf) != 0;
8157 }
8158 #endif /* CONFIG_SMP */
8159
set_next_buddy(struct sched_entity * se)8160 static void set_next_buddy(struct sched_entity *se)
8161 {
8162 for_each_sched_entity(se) {
8163 if (SCHED_WARN_ON(!se->on_rq))
8164 return;
8165 if (se_is_idle(se))
8166 return;
8167 cfs_rq_of(se)->next = se;
8168 }
8169 }
8170
8171 /*
8172 * Preempt the current task with a newly woken task if needed:
8173 */
check_preempt_wakeup(struct rq * rq,struct task_struct * p,int wake_flags)8174 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
8175 {
8176 struct task_struct *curr = rq->curr;
8177 struct sched_entity *se = &curr->se, *pse = &p->se;
8178 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8179 int next_buddy_marked = 0;
8180 int cse_is_idle, pse_is_idle;
8181
8182 if (unlikely(se == pse))
8183 return;
8184
8185 /*
8186 * This is possible from callers such as attach_tasks(), in which we
8187 * unconditionally check_preempt_curr() after an enqueue (which may have
8188 * lead to a throttle). This both saves work and prevents false
8189 * next-buddy nomination below.
8190 */
8191 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
8192 return;
8193
8194 if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) {
8195 set_next_buddy(pse);
8196 next_buddy_marked = 1;
8197 }
8198
8199 /*
8200 * We can come here with TIF_NEED_RESCHED already set from new task
8201 * wake up path.
8202 *
8203 * Note: this also catches the edge-case of curr being in a throttled
8204 * group (e.g. via set_curr_task), since update_curr() (in the
8205 * enqueue of curr) will have resulted in resched being set. This
8206 * prevents us from potentially nominating it as a false LAST_BUDDY
8207 * below.
8208 */
8209 if (test_tsk_need_resched(curr))
8210 return;
8211
8212 /* Idle tasks are by definition preempted by non-idle tasks. */
8213 if (unlikely(task_has_idle_policy(curr)) &&
8214 likely(!task_has_idle_policy(p)))
8215 goto preempt;
8216
8217 /*
8218 * Batch and idle tasks do not preempt non-idle tasks (their preemption
8219 * is driven by the tick):
8220 */
8221 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
8222 return;
8223
8224 find_matching_se(&se, &pse);
8225 WARN_ON_ONCE(!pse);
8226
8227 cse_is_idle = se_is_idle(se);
8228 pse_is_idle = se_is_idle(pse);
8229
8230 /*
8231 * Preempt an idle group in favor of a non-idle group (and don't preempt
8232 * in the inverse case).
8233 */
8234 if (cse_is_idle && !pse_is_idle)
8235 goto preempt;
8236 if (cse_is_idle != pse_is_idle)
8237 return;
8238
8239 cfs_rq = cfs_rq_of(se);
8240 update_curr(cfs_rq);
8241
8242 /*
8243 * XXX pick_eevdf(cfs_rq) != se ?
8244 */
8245 if (pick_eevdf(cfs_rq) == pse)
8246 goto preempt;
8247
8248 return;
8249
8250 preempt:
8251 resched_curr(rq);
8252 }
8253
8254 #ifdef CONFIG_SMP
pick_task_fair(struct rq * rq)8255 static struct task_struct *pick_task_fair(struct rq *rq)
8256 {
8257 struct sched_entity *se;
8258 struct cfs_rq *cfs_rq;
8259
8260 again:
8261 cfs_rq = &rq->cfs;
8262 if (!cfs_rq->nr_running)
8263 return NULL;
8264
8265 do {
8266 struct sched_entity *curr = cfs_rq->curr;
8267
8268 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
8269 if (curr) {
8270 if (curr->on_rq)
8271 update_curr(cfs_rq);
8272 else
8273 curr = NULL;
8274
8275 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
8276 goto again;
8277 }
8278
8279 se = pick_next_entity(cfs_rq, curr);
8280 cfs_rq = group_cfs_rq(se);
8281 } while (cfs_rq);
8282
8283 return task_of(se);
8284 }
8285 #endif
8286
8287 struct task_struct *
pick_next_task_fair(struct rq * rq,struct task_struct * prev,struct rq_flags * rf)8288 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8289 {
8290 struct cfs_rq *cfs_rq = &rq->cfs;
8291 struct sched_entity *se;
8292 struct task_struct *p;
8293 int new_tasks;
8294
8295 again:
8296 if (!sched_fair_runnable(rq))
8297 goto idle;
8298
8299 #ifdef CONFIG_FAIR_GROUP_SCHED
8300 if (!prev || prev->sched_class != &fair_sched_class)
8301 goto simple;
8302
8303 /*
8304 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
8305 * likely that a next task is from the same cgroup as the current.
8306 *
8307 * Therefore attempt to avoid putting and setting the entire cgroup
8308 * hierarchy, only change the part that actually changes.
8309 */
8310
8311 do {
8312 struct sched_entity *curr = cfs_rq->curr;
8313
8314 /*
8315 * Since we got here without doing put_prev_entity() we also
8316 * have to consider cfs_rq->curr. If it is still a runnable
8317 * entity, update_curr() will update its vruntime, otherwise
8318 * forget we've ever seen it.
8319 */
8320 if (curr) {
8321 if (curr->on_rq)
8322 update_curr(cfs_rq);
8323 else
8324 curr = NULL;
8325
8326 /*
8327 * This call to check_cfs_rq_runtime() will do the
8328 * throttle and dequeue its entity in the parent(s).
8329 * Therefore the nr_running test will indeed
8330 * be correct.
8331 */
8332 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
8333 cfs_rq = &rq->cfs;
8334
8335 if (!cfs_rq->nr_running)
8336 goto idle;
8337
8338 goto simple;
8339 }
8340 }
8341
8342 se = pick_next_entity(cfs_rq, curr);
8343 cfs_rq = group_cfs_rq(se);
8344 } while (cfs_rq);
8345
8346 p = task_of(se);
8347
8348 /*
8349 * Since we haven't yet done put_prev_entity and if the selected task
8350 * is a different task than we started out with, try and touch the
8351 * least amount of cfs_rqs.
8352 */
8353 if (prev != p) {
8354 struct sched_entity *pse = &prev->se;
8355
8356 while (!(cfs_rq = is_same_group(se, pse))) {
8357 int se_depth = se->depth;
8358 int pse_depth = pse->depth;
8359
8360 if (se_depth <= pse_depth) {
8361 put_prev_entity(cfs_rq_of(pse), pse);
8362 pse = parent_entity(pse);
8363 }
8364 if (se_depth >= pse_depth) {
8365 set_next_entity(cfs_rq_of(se), se);
8366 se = parent_entity(se);
8367 }
8368 }
8369
8370 put_prev_entity(cfs_rq, pse);
8371 set_next_entity(cfs_rq, se);
8372 }
8373
8374 goto done;
8375 simple:
8376 #endif
8377 if (prev)
8378 put_prev_task(rq, prev);
8379
8380 do {
8381 se = pick_next_entity(cfs_rq, NULL);
8382 set_next_entity(cfs_rq, se);
8383 cfs_rq = group_cfs_rq(se);
8384 } while (cfs_rq);
8385
8386 p = task_of(se);
8387
8388 done: __maybe_unused;
8389 #ifdef CONFIG_SMP
8390 /*
8391 * Move the next running task to the front of
8392 * the list, so our cfs_tasks list becomes MRU
8393 * one.
8394 */
8395 list_move(&p->se.group_node, &rq->cfs_tasks);
8396 #endif
8397
8398 if (hrtick_enabled_fair(rq))
8399 hrtick_start_fair(rq, p);
8400
8401 update_misfit_status(p, rq);
8402 sched_fair_update_stop_tick(rq, p);
8403
8404 return p;
8405
8406 idle:
8407 if (!rf)
8408 return NULL;
8409
8410 new_tasks = newidle_balance(rq, rf);
8411
8412 /*
8413 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
8414 * possible for any higher priority task to appear. In that case we
8415 * must re-start the pick_next_entity() loop.
8416 */
8417 if (new_tasks < 0)
8418 return RETRY_TASK;
8419
8420 if (new_tasks > 0)
8421 goto again;
8422
8423 /*
8424 * rq is about to be idle, check if we need to update the
8425 * lost_idle_time of clock_pelt
8426 */
8427 update_idle_rq_clock_pelt(rq);
8428
8429 return NULL;
8430 }
8431
__pick_next_task_fair(struct rq * rq)8432 static struct task_struct *__pick_next_task_fair(struct rq *rq)
8433 {
8434 return pick_next_task_fair(rq, NULL, NULL);
8435 }
8436
8437 /*
8438 * Account for a descheduled task:
8439 */
put_prev_task_fair(struct rq * rq,struct task_struct * prev)8440 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
8441 {
8442 struct sched_entity *se = &prev->se;
8443 struct cfs_rq *cfs_rq;
8444
8445 for_each_sched_entity(se) {
8446 cfs_rq = cfs_rq_of(se);
8447 put_prev_entity(cfs_rq, se);
8448 }
8449 }
8450
8451 /*
8452 * sched_yield() is very simple
8453 */
yield_task_fair(struct rq * rq)8454 static void yield_task_fair(struct rq *rq)
8455 {
8456 struct task_struct *curr = rq->curr;
8457 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8458 struct sched_entity *se = &curr->se;
8459
8460 /*
8461 * Are we the only task in the tree?
8462 */
8463 if (unlikely(rq->nr_running == 1))
8464 return;
8465
8466 clear_buddies(cfs_rq, se);
8467
8468 update_rq_clock(rq);
8469 /*
8470 * Update run-time statistics of the 'current'.
8471 */
8472 update_curr(cfs_rq);
8473 /*
8474 * Tell update_rq_clock() that we've just updated,
8475 * so we don't do microscopic update in schedule()
8476 * and double the fastpath cost.
8477 */
8478 rq_clock_skip_update(rq);
8479
8480 se->deadline += calc_delta_fair(se->slice, se);
8481 }
8482
yield_to_task_fair(struct rq * rq,struct task_struct * p)8483 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
8484 {
8485 struct sched_entity *se = &p->se;
8486
8487 /* throttled hierarchies are not runnable */
8488 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
8489 return false;
8490
8491 /* Tell the scheduler that we'd really like pse to run next. */
8492 set_next_buddy(se);
8493
8494 yield_task_fair(rq);
8495
8496 return true;
8497 }
8498
8499 #ifdef CONFIG_SMP
8500 /**************************************************
8501 * Fair scheduling class load-balancing methods.
8502 *
8503 * BASICS
8504 *
8505 * The purpose of load-balancing is to achieve the same basic fairness the
8506 * per-CPU scheduler provides, namely provide a proportional amount of compute
8507 * time to each task. This is expressed in the following equation:
8508 *
8509 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
8510 *
8511 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
8512 * W_i,0 is defined as:
8513 *
8514 * W_i,0 = \Sum_j w_i,j (2)
8515 *
8516 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
8517 * is derived from the nice value as per sched_prio_to_weight[].
8518 *
8519 * The weight average is an exponential decay average of the instantaneous
8520 * weight:
8521 *
8522 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
8523 *
8524 * C_i is the compute capacity of CPU i, typically it is the
8525 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
8526 * can also include other factors [XXX].
8527 *
8528 * To achieve this balance we define a measure of imbalance which follows
8529 * directly from (1):
8530 *
8531 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
8532 *
8533 * We them move tasks around to minimize the imbalance. In the continuous
8534 * function space it is obvious this converges, in the discrete case we get
8535 * a few fun cases generally called infeasible weight scenarios.
8536 *
8537 * [XXX expand on:
8538 * - infeasible weights;
8539 * - local vs global optima in the discrete case. ]
8540 *
8541 *
8542 * SCHED DOMAINS
8543 *
8544 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
8545 * for all i,j solution, we create a tree of CPUs that follows the hardware
8546 * topology where each level pairs two lower groups (or better). This results
8547 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
8548 * tree to only the first of the previous level and we decrease the frequency
8549 * of load-balance at each level inv. proportional to the number of CPUs in
8550 * the groups.
8551 *
8552 * This yields:
8553 *
8554 * log_2 n 1 n
8555 * \Sum { --- * --- * 2^i } = O(n) (5)
8556 * i = 0 2^i 2^i
8557 * `- size of each group
8558 * | | `- number of CPUs doing load-balance
8559 * | `- freq
8560 * `- sum over all levels
8561 *
8562 * Coupled with a limit on how many tasks we can migrate every balance pass,
8563 * this makes (5) the runtime complexity of the balancer.
8564 *
8565 * An important property here is that each CPU is still (indirectly) connected
8566 * to every other CPU in at most O(log n) steps:
8567 *
8568 * The adjacency matrix of the resulting graph is given by:
8569 *
8570 * log_2 n
8571 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
8572 * k = 0
8573 *
8574 * And you'll find that:
8575 *
8576 * A^(log_2 n)_i,j != 0 for all i,j (7)
8577 *
8578 * Showing there's indeed a path between every CPU in at most O(log n) steps.
8579 * The task movement gives a factor of O(m), giving a convergence complexity
8580 * of:
8581 *
8582 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
8583 *
8584 *
8585 * WORK CONSERVING
8586 *
8587 * In order to avoid CPUs going idle while there's still work to do, new idle
8588 * balancing is more aggressive and has the newly idle CPU iterate up the domain
8589 * tree itself instead of relying on other CPUs to bring it work.
8590 *
8591 * This adds some complexity to both (5) and (8) but it reduces the total idle
8592 * time.
8593 *
8594 * [XXX more?]
8595 *
8596 *
8597 * CGROUPS
8598 *
8599 * Cgroups make a horror show out of (2), instead of a simple sum we get:
8600 *
8601 * s_k,i
8602 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
8603 * S_k
8604 *
8605 * Where
8606 *
8607 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
8608 *
8609 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8610 *
8611 * The big problem is S_k, its a global sum needed to compute a local (W_i)
8612 * property.
8613 *
8614 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8615 * rewrite all of this once again.]
8616 */
8617
8618 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8619
8620 enum fbq_type { regular, remote, all };
8621
8622 /*
8623 * 'group_type' describes the group of CPUs at the moment of load balancing.
8624 *
8625 * The enum is ordered by pulling priority, with the group with lowest priority
8626 * first so the group_type can simply be compared when selecting the busiest
8627 * group. See update_sd_pick_busiest().
8628 */
8629 enum group_type {
8630 /* The group has spare capacity that can be used to run more tasks. */
8631 group_has_spare = 0,
8632 /*
8633 * The group is fully used and the tasks don't compete for more CPU
8634 * cycles. Nevertheless, some tasks might wait before running.
8635 */
8636 group_fully_busy,
8637 /*
8638 * One task doesn't fit with CPU's capacity and must be migrated to a
8639 * more powerful CPU.
8640 */
8641 group_misfit_task,
8642 /*
8643 * Balance SMT group that's fully busy. Can benefit from migration
8644 * a task on SMT with busy sibling to another CPU on idle core.
8645 */
8646 group_smt_balance,
8647 /*
8648 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8649 * and the task should be migrated to it instead of running on the
8650 * current CPU.
8651 */
8652 group_asym_packing,
8653 /*
8654 * The tasks' affinity constraints previously prevented the scheduler
8655 * from balancing the load across the system.
8656 */
8657 group_imbalanced,
8658 /*
8659 * The CPU is overloaded and can't provide expected CPU cycles to all
8660 * tasks.
8661 */
8662 group_overloaded
8663 };
8664
8665 enum migration_type {
8666 migrate_load = 0,
8667 migrate_util,
8668 migrate_task,
8669 migrate_misfit
8670 };
8671
8672 #define LBF_ALL_PINNED 0x01
8673 #define LBF_NEED_BREAK 0x02
8674 #define LBF_DST_PINNED 0x04
8675 #define LBF_SOME_PINNED 0x08
8676 #define LBF_ACTIVE_LB 0x10
8677
8678 struct lb_env {
8679 struct sched_domain *sd;
8680
8681 struct rq *src_rq;
8682 int src_cpu;
8683
8684 int dst_cpu;
8685 struct rq *dst_rq;
8686
8687 struct cpumask *dst_grpmask;
8688 int new_dst_cpu;
8689 enum cpu_idle_type idle;
8690 long imbalance;
8691 /* The set of CPUs under consideration for load-balancing */
8692 struct cpumask *cpus;
8693
8694 unsigned int flags;
8695
8696 unsigned int loop;
8697 unsigned int loop_break;
8698 unsigned int loop_max;
8699
8700 enum fbq_type fbq_type;
8701 enum migration_type migration_type;
8702 struct list_head tasks;
8703 };
8704
8705 /*
8706 * Is this task likely cache-hot:
8707 */
task_hot(struct task_struct * p,struct lb_env * env)8708 static int task_hot(struct task_struct *p, struct lb_env *env)
8709 {
8710 s64 delta;
8711
8712 lockdep_assert_rq_held(env->src_rq);
8713
8714 if (p->sched_class != &fair_sched_class)
8715 return 0;
8716
8717 if (unlikely(task_has_idle_policy(p)))
8718 return 0;
8719
8720 /* SMT siblings share cache */
8721 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8722 return 0;
8723
8724 /*
8725 * Buddy candidates are cache hot:
8726 */
8727 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8728 (&p->se == cfs_rq_of(&p->se)->next))
8729 return 1;
8730
8731 if (sysctl_sched_migration_cost == -1)
8732 return 1;
8733
8734 /*
8735 * Don't migrate task if the task's cookie does not match
8736 * with the destination CPU's core cookie.
8737 */
8738 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8739 return 1;
8740
8741 if (sysctl_sched_migration_cost == 0)
8742 return 0;
8743
8744 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8745
8746 return delta < (s64)sysctl_sched_migration_cost;
8747 }
8748
8749 #ifdef CONFIG_NUMA_BALANCING
8750 /*
8751 * Returns 1, if task migration degrades locality
8752 * Returns 0, if task migration improves locality i.e migration preferred.
8753 * Returns -1, if task migration is not affected by locality.
8754 */
migrate_degrades_locality(struct task_struct * p,struct lb_env * env)8755 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8756 {
8757 struct numa_group *numa_group = rcu_dereference(p->numa_group);
8758 unsigned long src_weight, dst_weight;
8759 int src_nid, dst_nid, dist;
8760
8761 if (!static_branch_likely(&sched_numa_balancing))
8762 return -1;
8763
8764 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8765 return -1;
8766
8767 src_nid = cpu_to_node(env->src_cpu);
8768 dst_nid = cpu_to_node(env->dst_cpu);
8769
8770 if (src_nid == dst_nid)
8771 return -1;
8772
8773 /* Migrating away from the preferred node is always bad. */
8774 if (src_nid == p->numa_preferred_nid) {
8775 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8776 return 1;
8777 else
8778 return -1;
8779 }
8780
8781 /* Encourage migration to the preferred node. */
8782 if (dst_nid == p->numa_preferred_nid)
8783 return 0;
8784
8785 /* Leaving a core idle is often worse than degrading locality. */
8786 if (env->idle == CPU_IDLE)
8787 return -1;
8788
8789 dist = node_distance(src_nid, dst_nid);
8790 if (numa_group) {
8791 src_weight = group_weight(p, src_nid, dist);
8792 dst_weight = group_weight(p, dst_nid, dist);
8793 } else {
8794 src_weight = task_weight(p, src_nid, dist);
8795 dst_weight = task_weight(p, dst_nid, dist);
8796 }
8797
8798 return dst_weight < src_weight;
8799 }
8800
8801 #else
migrate_degrades_locality(struct task_struct * p,struct lb_env * env)8802 static inline int migrate_degrades_locality(struct task_struct *p,
8803 struct lb_env *env)
8804 {
8805 return -1;
8806 }
8807 #endif
8808
8809 /*
8810 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8811 */
8812 static
can_migrate_task(struct task_struct * p,struct lb_env * env)8813 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8814 {
8815 int tsk_cache_hot;
8816
8817 lockdep_assert_rq_held(env->src_rq);
8818
8819 /*
8820 * We do not migrate tasks that are:
8821 * 1) throttled_lb_pair, or
8822 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8823 * 3) running (obviously), or
8824 * 4) are cache-hot on their current CPU.
8825 */
8826 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8827 return 0;
8828
8829 /* Disregard pcpu kthreads; they are where they need to be. */
8830 if (kthread_is_per_cpu(p))
8831 return 0;
8832
8833 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8834 int cpu;
8835
8836 schedstat_inc(p->stats.nr_failed_migrations_affine);
8837
8838 env->flags |= LBF_SOME_PINNED;
8839
8840 /*
8841 * Remember if this task can be migrated to any other CPU in
8842 * our sched_group. We may want to revisit it if we couldn't
8843 * meet load balance goals by pulling other tasks on src_cpu.
8844 *
8845 * Avoid computing new_dst_cpu
8846 * - for NEWLY_IDLE
8847 * - if we have already computed one in current iteration
8848 * - if it's an active balance
8849 */
8850 if (env->idle == CPU_NEWLY_IDLE ||
8851 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8852 return 0;
8853
8854 /* Prevent to re-select dst_cpu via env's CPUs: */
8855 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8856 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8857 env->flags |= LBF_DST_PINNED;
8858 env->new_dst_cpu = cpu;
8859 break;
8860 }
8861 }
8862
8863 return 0;
8864 }
8865
8866 /* Record that we found at least one task that could run on dst_cpu */
8867 env->flags &= ~LBF_ALL_PINNED;
8868
8869 if (task_on_cpu(env->src_rq, p)) {
8870 schedstat_inc(p->stats.nr_failed_migrations_running);
8871 return 0;
8872 }
8873
8874 /*
8875 * Aggressive migration if:
8876 * 1) active balance
8877 * 2) destination numa is preferred
8878 * 3) task is cache cold, or
8879 * 4) too many balance attempts have failed.
8880 */
8881 if (env->flags & LBF_ACTIVE_LB)
8882 return 1;
8883
8884 tsk_cache_hot = migrate_degrades_locality(p, env);
8885 if (tsk_cache_hot == -1)
8886 tsk_cache_hot = task_hot(p, env);
8887
8888 if (tsk_cache_hot <= 0 ||
8889 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
8890 if (tsk_cache_hot == 1) {
8891 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
8892 schedstat_inc(p->stats.nr_forced_migrations);
8893 }
8894 return 1;
8895 }
8896
8897 schedstat_inc(p->stats.nr_failed_migrations_hot);
8898 return 0;
8899 }
8900
8901 /*
8902 * detach_task() -- detach the task for the migration specified in env
8903 */
detach_task(struct task_struct * p,struct lb_env * env)8904 static void detach_task(struct task_struct *p, struct lb_env *env)
8905 {
8906 lockdep_assert_rq_held(env->src_rq);
8907
8908 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
8909 set_task_cpu(p, env->dst_cpu);
8910 }
8911
8912 /*
8913 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
8914 * part of active balancing operations within "domain".
8915 *
8916 * Returns a task if successful and NULL otherwise.
8917 */
detach_one_task(struct lb_env * env)8918 static struct task_struct *detach_one_task(struct lb_env *env)
8919 {
8920 struct task_struct *p;
8921
8922 lockdep_assert_rq_held(env->src_rq);
8923
8924 list_for_each_entry_reverse(p,
8925 &env->src_rq->cfs_tasks, se.group_node) {
8926 if (!can_migrate_task(p, env))
8927 continue;
8928
8929 detach_task(p, env);
8930
8931 /*
8932 * Right now, this is only the second place where
8933 * lb_gained[env->idle] is updated (other is detach_tasks)
8934 * so we can safely collect stats here rather than
8935 * inside detach_tasks().
8936 */
8937 schedstat_inc(env->sd->lb_gained[env->idle]);
8938 return p;
8939 }
8940 return NULL;
8941 }
8942
8943 /*
8944 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
8945 * busiest_rq, as part of a balancing operation within domain "sd".
8946 *
8947 * Returns number of detached tasks if successful and 0 otherwise.
8948 */
detach_tasks(struct lb_env * env)8949 static int detach_tasks(struct lb_env *env)
8950 {
8951 struct list_head *tasks = &env->src_rq->cfs_tasks;
8952 unsigned long util, load;
8953 struct task_struct *p;
8954 int detached = 0;
8955
8956 lockdep_assert_rq_held(env->src_rq);
8957
8958 /*
8959 * Source run queue has been emptied by another CPU, clear
8960 * LBF_ALL_PINNED flag as we will not test any task.
8961 */
8962 if (env->src_rq->nr_running <= 1) {
8963 env->flags &= ~LBF_ALL_PINNED;
8964 return 0;
8965 }
8966
8967 if (env->imbalance <= 0)
8968 return 0;
8969
8970 while (!list_empty(tasks)) {
8971 /*
8972 * We don't want to steal all, otherwise we may be treated likewise,
8973 * which could at worst lead to a livelock crash.
8974 */
8975 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
8976 break;
8977
8978 env->loop++;
8979 /* We've more or less seen every task there is, call it quits */
8980 if (env->loop > env->loop_max)
8981 break;
8982
8983 /* take a breather every nr_migrate tasks */
8984 if (env->loop > env->loop_break) {
8985 env->loop_break += SCHED_NR_MIGRATE_BREAK;
8986 env->flags |= LBF_NEED_BREAK;
8987 break;
8988 }
8989
8990 p = list_last_entry(tasks, struct task_struct, se.group_node);
8991
8992 if (!can_migrate_task(p, env))
8993 goto next;
8994
8995 switch (env->migration_type) {
8996 case migrate_load:
8997 /*
8998 * Depending of the number of CPUs and tasks and the
8999 * cgroup hierarchy, task_h_load() can return a null
9000 * value. Make sure that env->imbalance decreases
9001 * otherwise detach_tasks() will stop only after
9002 * detaching up to loop_max tasks.
9003 */
9004 load = max_t(unsigned long, task_h_load(p), 1);
9005
9006 if (sched_feat(LB_MIN) &&
9007 load < 16 && !env->sd->nr_balance_failed)
9008 goto next;
9009
9010 /*
9011 * Make sure that we don't migrate too much load.
9012 * Nevertheless, let relax the constraint if
9013 * scheduler fails to find a good waiting task to
9014 * migrate.
9015 */
9016 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
9017 goto next;
9018
9019 env->imbalance -= load;
9020 break;
9021
9022 case migrate_util:
9023 util = task_util_est(p);
9024
9025 if (util > env->imbalance)
9026 goto next;
9027
9028 env->imbalance -= util;
9029 break;
9030
9031 case migrate_task:
9032 env->imbalance--;
9033 break;
9034
9035 case migrate_misfit:
9036 /* This is not a misfit task */
9037 if (task_fits_cpu(p, env->src_cpu))
9038 goto next;
9039
9040 env->imbalance = 0;
9041 break;
9042 }
9043
9044 detach_task(p, env);
9045 list_add(&p->se.group_node, &env->tasks);
9046
9047 detached++;
9048
9049 #ifdef CONFIG_PREEMPTION
9050 /*
9051 * NEWIDLE balancing is a source of latency, so preemptible
9052 * kernels will stop after the first task is detached to minimize
9053 * the critical section.
9054 */
9055 if (env->idle == CPU_NEWLY_IDLE)
9056 break;
9057 #endif
9058
9059 /*
9060 * We only want to steal up to the prescribed amount of
9061 * load/util/tasks.
9062 */
9063 if (env->imbalance <= 0)
9064 break;
9065
9066 continue;
9067 next:
9068 list_move(&p->se.group_node, tasks);
9069 }
9070
9071 /*
9072 * Right now, this is one of only two places we collect this stat
9073 * so we can safely collect detach_one_task() stats here rather
9074 * than inside detach_one_task().
9075 */
9076 schedstat_add(env->sd->lb_gained[env->idle], detached);
9077
9078 return detached;
9079 }
9080
9081 /*
9082 * attach_task() -- attach the task detached by detach_task() to its new rq.
9083 */
attach_task(struct rq * rq,struct task_struct * p)9084 static void attach_task(struct rq *rq, struct task_struct *p)
9085 {
9086 lockdep_assert_rq_held(rq);
9087
9088 WARN_ON_ONCE(task_rq(p) != rq);
9089 activate_task(rq, p, ENQUEUE_NOCLOCK);
9090 check_preempt_curr(rq, p, 0);
9091 }
9092
9093 /*
9094 * attach_one_task() -- attaches the task returned from detach_one_task() to
9095 * its new rq.
9096 */
attach_one_task(struct rq * rq,struct task_struct * p)9097 static void attach_one_task(struct rq *rq, struct task_struct *p)
9098 {
9099 struct rq_flags rf;
9100
9101 rq_lock(rq, &rf);
9102 update_rq_clock(rq);
9103 attach_task(rq, p);
9104 rq_unlock(rq, &rf);
9105 }
9106
9107 /*
9108 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
9109 * new rq.
9110 */
attach_tasks(struct lb_env * env)9111 static void attach_tasks(struct lb_env *env)
9112 {
9113 struct list_head *tasks = &env->tasks;
9114 struct task_struct *p;
9115 struct rq_flags rf;
9116
9117 rq_lock(env->dst_rq, &rf);
9118 update_rq_clock(env->dst_rq);
9119
9120 while (!list_empty(tasks)) {
9121 p = list_first_entry(tasks, struct task_struct, se.group_node);
9122 list_del_init(&p->se.group_node);
9123
9124 attach_task(env->dst_rq, p);
9125 }
9126
9127 rq_unlock(env->dst_rq, &rf);
9128 }
9129
9130 #ifdef CONFIG_NO_HZ_COMMON
cfs_rq_has_blocked(struct cfs_rq * cfs_rq)9131 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
9132 {
9133 if (cfs_rq->avg.load_avg)
9134 return true;
9135
9136 if (cfs_rq->avg.util_avg)
9137 return true;
9138
9139 return false;
9140 }
9141
others_have_blocked(struct rq * rq)9142 static inline bool others_have_blocked(struct rq *rq)
9143 {
9144 if (READ_ONCE(rq->avg_rt.util_avg))
9145 return true;
9146
9147 if (READ_ONCE(rq->avg_dl.util_avg))
9148 return true;
9149
9150 if (thermal_load_avg(rq))
9151 return true;
9152
9153 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
9154 if (READ_ONCE(rq->avg_irq.util_avg))
9155 return true;
9156 #endif
9157
9158 return false;
9159 }
9160
update_blocked_load_tick(struct rq * rq)9161 static inline void update_blocked_load_tick(struct rq *rq)
9162 {
9163 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
9164 }
9165
update_blocked_load_status(struct rq * rq,bool has_blocked)9166 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
9167 {
9168 if (!has_blocked)
9169 rq->has_blocked_load = 0;
9170 }
9171 #else
cfs_rq_has_blocked(struct cfs_rq * cfs_rq)9172 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
others_have_blocked(struct rq * rq)9173 static inline bool others_have_blocked(struct rq *rq) { return false; }
update_blocked_load_tick(struct rq * rq)9174 static inline void update_blocked_load_tick(struct rq *rq) {}
update_blocked_load_status(struct rq * rq,bool has_blocked)9175 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
9176 #endif
9177
__update_blocked_others(struct rq * rq,bool * done)9178 static bool __update_blocked_others(struct rq *rq, bool *done)
9179 {
9180 const struct sched_class *curr_class;
9181 u64 now = rq_clock_pelt(rq);
9182 unsigned long thermal_pressure;
9183 bool decayed;
9184
9185 /*
9186 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
9187 * DL and IRQ signals have been updated before updating CFS.
9188 */
9189 curr_class = rq->curr->sched_class;
9190
9191 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
9192
9193 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
9194 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
9195 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
9196 update_irq_load_avg(rq, 0);
9197
9198 if (others_have_blocked(rq))
9199 *done = false;
9200
9201 return decayed;
9202 }
9203
9204 #ifdef CONFIG_FAIR_GROUP_SCHED
9205
__update_blocked_fair(struct rq * rq,bool * done)9206 static bool __update_blocked_fair(struct rq *rq, bool *done)
9207 {
9208 struct cfs_rq *cfs_rq, *pos;
9209 bool decayed = false;
9210 int cpu = cpu_of(rq);
9211
9212 /*
9213 * Iterates the task_group tree in a bottom up fashion, see
9214 * list_add_leaf_cfs_rq() for details.
9215 */
9216 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
9217 struct sched_entity *se;
9218
9219 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
9220 update_tg_load_avg(cfs_rq);
9221
9222 if (cfs_rq->nr_running == 0)
9223 update_idle_cfs_rq_clock_pelt(cfs_rq);
9224
9225 if (cfs_rq == &rq->cfs)
9226 decayed = true;
9227 }
9228
9229 /* Propagate pending load changes to the parent, if any: */
9230 se = cfs_rq->tg->se[cpu];
9231 if (se && !skip_blocked_update(se))
9232 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9233
9234 /*
9235 * There can be a lot of idle CPU cgroups. Don't let fully
9236 * decayed cfs_rqs linger on the list.
9237 */
9238 if (cfs_rq_is_decayed(cfs_rq))
9239 list_del_leaf_cfs_rq(cfs_rq);
9240
9241 /* Don't need periodic decay once load/util_avg are null */
9242 if (cfs_rq_has_blocked(cfs_rq))
9243 *done = false;
9244 }
9245
9246 return decayed;
9247 }
9248
9249 /*
9250 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
9251 * This needs to be done in a top-down fashion because the load of a child
9252 * group is a fraction of its parents load.
9253 */
update_cfs_rq_h_load(struct cfs_rq * cfs_rq)9254 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
9255 {
9256 struct rq *rq = rq_of(cfs_rq);
9257 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
9258 unsigned long now = jiffies;
9259 unsigned long load;
9260
9261 if (cfs_rq->last_h_load_update == now)
9262 return;
9263
9264 WRITE_ONCE(cfs_rq->h_load_next, NULL);
9265 for_each_sched_entity(se) {
9266 cfs_rq = cfs_rq_of(se);
9267 WRITE_ONCE(cfs_rq->h_load_next, se);
9268 if (cfs_rq->last_h_load_update == now)
9269 break;
9270 }
9271
9272 if (!se) {
9273 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
9274 cfs_rq->last_h_load_update = now;
9275 }
9276
9277 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
9278 load = cfs_rq->h_load;
9279 load = div64_ul(load * se->avg.load_avg,
9280 cfs_rq_load_avg(cfs_rq) + 1);
9281 cfs_rq = group_cfs_rq(se);
9282 cfs_rq->h_load = load;
9283 cfs_rq->last_h_load_update = now;
9284 }
9285 }
9286
task_h_load(struct task_struct * p)9287 static unsigned long task_h_load(struct task_struct *p)
9288 {
9289 struct cfs_rq *cfs_rq = task_cfs_rq(p);
9290
9291 update_cfs_rq_h_load(cfs_rq);
9292 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
9293 cfs_rq_load_avg(cfs_rq) + 1);
9294 }
9295 #else
__update_blocked_fair(struct rq * rq,bool * done)9296 static bool __update_blocked_fair(struct rq *rq, bool *done)
9297 {
9298 struct cfs_rq *cfs_rq = &rq->cfs;
9299 bool decayed;
9300
9301 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
9302 if (cfs_rq_has_blocked(cfs_rq))
9303 *done = false;
9304
9305 return decayed;
9306 }
9307
task_h_load(struct task_struct * p)9308 static unsigned long task_h_load(struct task_struct *p)
9309 {
9310 return p->se.avg.load_avg;
9311 }
9312 #endif
9313
update_blocked_averages(int cpu)9314 static void update_blocked_averages(int cpu)
9315 {
9316 bool decayed = false, done = true;
9317 struct rq *rq = cpu_rq(cpu);
9318 struct rq_flags rf;
9319
9320 rq_lock_irqsave(rq, &rf);
9321 update_blocked_load_tick(rq);
9322 update_rq_clock(rq);
9323
9324 decayed |= __update_blocked_others(rq, &done);
9325 decayed |= __update_blocked_fair(rq, &done);
9326
9327 update_blocked_load_status(rq, !done);
9328 if (decayed)
9329 cpufreq_update_util(rq, 0);
9330 rq_unlock_irqrestore(rq, &rf);
9331 }
9332
9333 /********** Helpers for find_busiest_group ************************/
9334
9335 /*
9336 * sg_lb_stats - stats of a sched_group required for load_balancing
9337 */
9338 struct sg_lb_stats {
9339 unsigned long avg_load; /*Avg load across the CPUs of the group */
9340 unsigned long group_load; /* Total load over the CPUs of the group */
9341 unsigned long group_capacity;
9342 unsigned long group_util; /* Total utilization over the CPUs of the group */
9343 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
9344 unsigned int sum_nr_running; /* Nr of tasks running in the group */
9345 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
9346 unsigned int idle_cpus;
9347 unsigned int group_weight;
9348 enum group_type group_type;
9349 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
9350 unsigned int group_smt_balance; /* Task on busy SMT be moved */
9351 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
9352 #ifdef CONFIG_NUMA_BALANCING
9353 unsigned int nr_numa_running;
9354 unsigned int nr_preferred_running;
9355 #endif
9356 };
9357
9358 /*
9359 * sd_lb_stats - Structure to store the statistics of a sched_domain
9360 * during load balancing.
9361 */
9362 struct sd_lb_stats {
9363 struct sched_group *busiest; /* Busiest group in this sd */
9364 struct sched_group *local; /* Local group in this sd */
9365 unsigned long total_load; /* Total load of all groups in sd */
9366 unsigned long total_capacity; /* Total capacity of all groups in sd */
9367 unsigned long avg_load; /* Average load across all groups in sd */
9368 unsigned int prefer_sibling; /* tasks should go to sibling first */
9369
9370 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
9371 struct sg_lb_stats local_stat; /* Statistics of the local group */
9372 };
9373
init_sd_lb_stats(struct sd_lb_stats * sds)9374 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
9375 {
9376 /*
9377 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
9378 * local_stat because update_sg_lb_stats() does a full clear/assignment.
9379 * We must however set busiest_stat::group_type and
9380 * busiest_stat::idle_cpus to the worst busiest group because
9381 * update_sd_pick_busiest() reads these before assignment.
9382 */
9383 *sds = (struct sd_lb_stats){
9384 .busiest = NULL,
9385 .local = NULL,
9386 .total_load = 0UL,
9387 .total_capacity = 0UL,
9388 .busiest_stat = {
9389 .idle_cpus = UINT_MAX,
9390 .group_type = group_has_spare,
9391 },
9392 };
9393 }
9394
scale_rt_capacity(int cpu)9395 static unsigned long scale_rt_capacity(int cpu)
9396 {
9397 struct rq *rq = cpu_rq(cpu);
9398 unsigned long max = arch_scale_cpu_capacity(cpu);
9399 unsigned long used, free;
9400 unsigned long irq;
9401
9402 irq = cpu_util_irq(rq);
9403
9404 if (unlikely(irq >= max))
9405 return 1;
9406
9407 /*
9408 * avg_rt.util_avg and avg_dl.util_avg track binary signals
9409 * (running and not running) with weights 0 and 1024 respectively.
9410 * avg_thermal.load_avg tracks thermal pressure and the weighted
9411 * average uses the actual delta max capacity(load).
9412 */
9413 used = READ_ONCE(rq->avg_rt.util_avg);
9414 used += READ_ONCE(rq->avg_dl.util_avg);
9415 used += thermal_load_avg(rq);
9416
9417 if (unlikely(used >= max))
9418 return 1;
9419
9420 free = max - used;
9421
9422 return scale_irq_capacity(free, irq, max);
9423 }
9424
update_cpu_capacity(struct sched_domain * sd,int cpu)9425 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
9426 {
9427 unsigned long capacity = scale_rt_capacity(cpu);
9428 struct sched_group *sdg = sd->groups;
9429
9430 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
9431
9432 if (!capacity)
9433 capacity = 1;
9434
9435 cpu_rq(cpu)->cpu_capacity = capacity;
9436 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
9437
9438 sdg->sgc->capacity = capacity;
9439 sdg->sgc->min_capacity = capacity;
9440 sdg->sgc->max_capacity = capacity;
9441 }
9442
update_group_capacity(struct sched_domain * sd,int cpu)9443 void update_group_capacity(struct sched_domain *sd, int cpu)
9444 {
9445 struct sched_domain *child = sd->child;
9446 struct sched_group *group, *sdg = sd->groups;
9447 unsigned long capacity, min_capacity, max_capacity;
9448 unsigned long interval;
9449
9450 interval = msecs_to_jiffies(sd->balance_interval);
9451 interval = clamp(interval, 1UL, max_load_balance_interval);
9452 sdg->sgc->next_update = jiffies + interval;
9453
9454 if (!child) {
9455 update_cpu_capacity(sd, cpu);
9456 return;
9457 }
9458
9459 capacity = 0;
9460 min_capacity = ULONG_MAX;
9461 max_capacity = 0;
9462
9463 if (child->flags & SD_OVERLAP) {
9464 /*
9465 * SD_OVERLAP domains cannot assume that child groups
9466 * span the current group.
9467 */
9468
9469 for_each_cpu(cpu, sched_group_span(sdg)) {
9470 unsigned long cpu_cap = capacity_of(cpu);
9471
9472 capacity += cpu_cap;
9473 min_capacity = min(cpu_cap, min_capacity);
9474 max_capacity = max(cpu_cap, max_capacity);
9475 }
9476 } else {
9477 /*
9478 * !SD_OVERLAP domains can assume that child groups
9479 * span the current group.
9480 */
9481
9482 group = child->groups;
9483 do {
9484 struct sched_group_capacity *sgc = group->sgc;
9485
9486 capacity += sgc->capacity;
9487 min_capacity = min(sgc->min_capacity, min_capacity);
9488 max_capacity = max(sgc->max_capacity, max_capacity);
9489 group = group->next;
9490 } while (group != child->groups);
9491 }
9492
9493 sdg->sgc->capacity = capacity;
9494 sdg->sgc->min_capacity = min_capacity;
9495 sdg->sgc->max_capacity = max_capacity;
9496 }
9497
9498 /*
9499 * Check whether the capacity of the rq has been noticeably reduced by side
9500 * activity. The imbalance_pct is used for the threshold.
9501 * Return true is the capacity is reduced
9502 */
9503 static inline int
check_cpu_capacity(struct rq * rq,struct sched_domain * sd)9504 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
9505 {
9506 return ((rq->cpu_capacity * sd->imbalance_pct) <
9507 (rq->cpu_capacity_orig * 100));
9508 }
9509
9510 /*
9511 * Check whether a rq has a misfit task and if it looks like we can actually
9512 * help that task: we can migrate the task to a CPU of higher capacity, or
9513 * the task's current CPU is heavily pressured.
9514 */
check_misfit_status(struct rq * rq,struct sched_domain * sd)9515 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
9516 {
9517 return rq->misfit_task_load &&
9518 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
9519 check_cpu_capacity(rq, sd));
9520 }
9521
9522 /*
9523 * Group imbalance indicates (and tries to solve) the problem where balancing
9524 * groups is inadequate due to ->cpus_ptr constraints.
9525 *
9526 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9527 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9528 * Something like:
9529 *
9530 * { 0 1 2 3 } { 4 5 6 7 }
9531 * * * * *
9532 *
9533 * If we were to balance group-wise we'd place two tasks in the first group and
9534 * two tasks in the second group. Clearly this is undesired as it will overload
9535 * cpu 3 and leave one of the CPUs in the second group unused.
9536 *
9537 * The current solution to this issue is detecting the skew in the first group
9538 * by noticing the lower domain failed to reach balance and had difficulty
9539 * moving tasks due to affinity constraints.
9540 *
9541 * When this is so detected; this group becomes a candidate for busiest; see
9542 * update_sd_pick_busiest(). And calculate_imbalance() and
9543 * find_busiest_group() avoid some of the usual balance conditions to allow it
9544 * to create an effective group imbalance.
9545 *
9546 * This is a somewhat tricky proposition since the next run might not find the
9547 * group imbalance and decide the groups need to be balanced again. A most
9548 * subtle and fragile situation.
9549 */
9550
sg_imbalanced(struct sched_group * group)9551 static inline int sg_imbalanced(struct sched_group *group)
9552 {
9553 return group->sgc->imbalance;
9554 }
9555
9556 /*
9557 * group_has_capacity returns true if the group has spare capacity that could
9558 * be used by some tasks.
9559 * We consider that a group has spare capacity if the number of task is
9560 * smaller than the number of CPUs or if the utilization is lower than the
9561 * available capacity for CFS tasks.
9562 * For the latter, we use a threshold to stabilize the state, to take into
9563 * account the variance of the tasks' load and to return true if the available
9564 * capacity in meaningful for the load balancer.
9565 * As an example, an available capacity of 1% can appear but it doesn't make
9566 * any benefit for the load balance.
9567 */
9568 static inline bool
group_has_capacity(unsigned int imbalance_pct,struct sg_lb_stats * sgs)9569 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9570 {
9571 if (sgs->sum_nr_running < sgs->group_weight)
9572 return true;
9573
9574 if ((sgs->group_capacity * imbalance_pct) <
9575 (sgs->group_runnable * 100))
9576 return false;
9577
9578 if ((sgs->group_capacity * 100) >
9579 (sgs->group_util * imbalance_pct))
9580 return true;
9581
9582 return false;
9583 }
9584
9585 /*
9586 * group_is_overloaded returns true if the group has more tasks than it can
9587 * handle.
9588 * group_is_overloaded is not equals to !group_has_capacity because a group
9589 * with the exact right number of tasks, has no more spare capacity but is not
9590 * overloaded so both group_has_capacity and group_is_overloaded return
9591 * false.
9592 */
9593 static inline bool
group_is_overloaded(unsigned int imbalance_pct,struct sg_lb_stats * sgs)9594 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9595 {
9596 if (sgs->sum_nr_running <= sgs->group_weight)
9597 return false;
9598
9599 if ((sgs->group_capacity * 100) <
9600 (sgs->group_util * imbalance_pct))
9601 return true;
9602
9603 if ((sgs->group_capacity * imbalance_pct) <
9604 (sgs->group_runnable * 100))
9605 return true;
9606
9607 return false;
9608 }
9609
9610 static inline enum
group_classify(unsigned int imbalance_pct,struct sched_group * group,struct sg_lb_stats * sgs)9611 group_type group_classify(unsigned int imbalance_pct,
9612 struct sched_group *group,
9613 struct sg_lb_stats *sgs)
9614 {
9615 if (group_is_overloaded(imbalance_pct, sgs))
9616 return group_overloaded;
9617
9618 if (sg_imbalanced(group))
9619 return group_imbalanced;
9620
9621 if (sgs->group_asym_packing)
9622 return group_asym_packing;
9623
9624 if (sgs->group_smt_balance)
9625 return group_smt_balance;
9626
9627 if (sgs->group_misfit_task_load)
9628 return group_misfit_task;
9629
9630 if (!group_has_capacity(imbalance_pct, sgs))
9631 return group_fully_busy;
9632
9633 return group_has_spare;
9634 }
9635
9636 /**
9637 * sched_use_asym_prio - Check whether asym_packing priority must be used
9638 * @sd: The scheduling domain of the load balancing
9639 * @cpu: A CPU
9640 *
9641 * Always use CPU priority when balancing load between SMT siblings. When
9642 * balancing load between cores, it is not sufficient that @cpu is idle. Only
9643 * use CPU priority if the whole core is idle.
9644 *
9645 * Returns: True if the priority of @cpu must be followed. False otherwise.
9646 */
sched_use_asym_prio(struct sched_domain * sd,int cpu)9647 static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
9648 {
9649 if (!sched_smt_active())
9650 return true;
9651
9652 return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);
9653 }
9654
9655 /**
9656 * sched_asym - Check if the destination CPU can do asym_packing load balance
9657 * @env: The load balancing environment
9658 * @sds: Load-balancing data with statistics of the local group
9659 * @sgs: Load-balancing statistics of the candidate busiest group
9660 * @group: The candidate busiest group
9661 *
9662 * @env::dst_cpu can do asym_packing if it has higher priority than the
9663 * preferred CPU of @group.
9664 *
9665 * SMT is a special case. If we are balancing load between cores, @env::dst_cpu
9666 * can do asym_packing balance only if all its SMT siblings are idle. Also, it
9667 * can only do it if @group is an SMT group and has exactly on busy CPU. Larger
9668 * imbalances in the number of CPUS are dealt with in find_busiest_group().
9669 *
9670 * If we are balancing load within an SMT core, or at DIE domain level, always
9671 * proceed.
9672 *
9673 * Return: true if @env::dst_cpu can do with asym_packing load balance. False
9674 * otherwise.
9675 */
9676 static inline bool
sched_asym(struct lb_env * env,struct sd_lb_stats * sds,struct sg_lb_stats * sgs,struct sched_group * group)9677 sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs,
9678 struct sched_group *group)
9679 {
9680 /* Ensure that the whole local core is idle, if applicable. */
9681 if (!sched_use_asym_prio(env->sd, env->dst_cpu))
9682 return false;
9683
9684 /*
9685 * CPU priorities does not make sense for SMT cores with more than one
9686 * busy sibling.
9687 */
9688 if (group->flags & SD_SHARE_CPUCAPACITY) {
9689 if (sgs->group_weight - sgs->idle_cpus != 1)
9690 return false;
9691 }
9692
9693 return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
9694 }
9695
9696 /* 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)9697 static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1,
9698 struct sched_group *sg2)
9699 {
9700 if (!sg1 || !sg2)
9701 return false;
9702
9703 return (sg1->flags & SD_SHARE_CPUCAPACITY) !=
9704 (sg2->flags & SD_SHARE_CPUCAPACITY);
9705 }
9706
smt_balance(struct lb_env * env,struct sg_lb_stats * sgs,struct sched_group * group)9707 static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs,
9708 struct sched_group *group)
9709 {
9710 if (env->idle == CPU_NOT_IDLE)
9711 return false;
9712
9713 /*
9714 * For SMT source group, it is better to move a task
9715 * to a CPU that doesn't have multiple tasks sharing its CPU capacity.
9716 * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY
9717 * will not be on.
9718 */
9719 if (group->flags & SD_SHARE_CPUCAPACITY &&
9720 sgs->sum_h_nr_running > 1)
9721 return true;
9722
9723 return false;
9724 }
9725
sibling_imbalance(struct lb_env * env,struct sd_lb_stats * sds,struct sg_lb_stats * busiest,struct sg_lb_stats * local)9726 static inline long sibling_imbalance(struct lb_env *env,
9727 struct sd_lb_stats *sds,
9728 struct sg_lb_stats *busiest,
9729 struct sg_lb_stats *local)
9730 {
9731 int ncores_busiest, ncores_local;
9732 long imbalance;
9733
9734 if (env->idle == CPU_NOT_IDLE || !busiest->sum_nr_running)
9735 return 0;
9736
9737 ncores_busiest = sds->busiest->cores;
9738 ncores_local = sds->local->cores;
9739
9740 if (ncores_busiest == ncores_local) {
9741 imbalance = busiest->sum_nr_running;
9742 lsub_positive(&imbalance, local->sum_nr_running);
9743 return imbalance;
9744 }
9745
9746 /* Balance such that nr_running/ncores ratio are same on both groups */
9747 imbalance = ncores_local * busiest->sum_nr_running;
9748 lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);
9749 /* Normalize imbalance and do rounding on normalization */
9750 imbalance = 2 * imbalance + ncores_local + ncores_busiest;
9751 imbalance /= ncores_local + ncores_busiest;
9752
9753 /* Take advantage of resource in an empty sched group */
9754 if (imbalance <= 1 && local->sum_nr_running == 0 &&
9755 busiest->sum_nr_running > 1)
9756 imbalance = 2;
9757
9758 return imbalance;
9759 }
9760
9761 static inline bool
sched_reduced_capacity(struct rq * rq,struct sched_domain * sd)9762 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9763 {
9764 /*
9765 * When there is more than 1 task, the group_overloaded case already
9766 * takes care of cpu with reduced capacity
9767 */
9768 if (rq->cfs.h_nr_running != 1)
9769 return false;
9770
9771 return check_cpu_capacity(rq, sd);
9772 }
9773
9774 /**
9775 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9776 * @env: The load balancing environment.
9777 * @sds: Load-balancing data with statistics of the local group.
9778 * @group: sched_group whose statistics are to be updated.
9779 * @sgs: variable to hold the statistics for this group.
9780 * @sg_status: Holds flag indicating the status of the sched_group
9781 */
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)9782 static inline void update_sg_lb_stats(struct lb_env *env,
9783 struct sd_lb_stats *sds,
9784 struct sched_group *group,
9785 struct sg_lb_stats *sgs,
9786 int *sg_status)
9787 {
9788 int i, nr_running, local_group;
9789
9790 memset(sgs, 0, sizeof(*sgs));
9791
9792 local_group = group == sds->local;
9793
9794 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9795 struct rq *rq = cpu_rq(i);
9796 unsigned long load = cpu_load(rq);
9797
9798 sgs->group_load += load;
9799 sgs->group_util += cpu_util_cfs(i);
9800 sgs->group_runnable += cpu_runnable(rq);
9801 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9802
9803 nr_running = rq->nr_running;
9804 sgs->sum_nr_running += nr_running;
9805
9806 if (nr_running > 1)
9807 *sg_status |= SG_OVERLOAD;
9808
9809 if (cpu_overutilized(i))
9810 *sg_status |= SG_OVERUTILIZED;
9811
9812 #ifdef CONFIG_NUMA_BALANCING
9813 sgs->nr_numa_running += rq->nr_numa_running;
9814 sgs->nr_preferred_running += rq->nr_preferred_running;
9815 #endif
9816 /*
9817 * No need to call idle_cpu() if nr_running is not 0
9818 */
9819 if (!nr_running && idle_cpu(i)) {
9820 sgs->idle_cpus++;
9821 /* Idle cpu can't have misfit task */
9822 continue;
9823 }
9824
9825 if (local_group)
9826 continue;
9827
9828 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9829 /* Check for a misfit task on the cpu */
9830 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9831 sgs->group_misfit_task_load = rq->misfit_task_load;
9832 *sg_status |= SG_OVERLOAD;
9833 }
9834 } else if ((env->idle != CPU_NOT_IDLE) &&
9835 sched_reduced_capacity(rq, env->sd)) {
9836 /* Check for a task running on a CPU with reduced capacity */
9837 if (sgs->group_misfit_task_load < load)
9838 sgs->group_misfit_task_load = load;
9839 }
9840 }
9841
9842 sgs->group_capacity = group->sgc->capacity;
9843
9844 sgs->group_weight = group->group_weight;
9845
9846 /* Check if dst CPU is idle and preferred to this group */
9847 if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
9848 env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9849 sched_asym(env, sds, sgs, group)) {
9850 sgs->group_asym_packing = 1;
9851 }
9852
9853 /* Check for loaded SMT group to be balanced to dst CPU */
9854 if (!local_group && smt_balance(env, sgs, group))
9855 sgs->group_smt_balance = 1;
9856
9857 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9858
9859 /* Computing avg_load makes sense only when group is overloaded */
9860 if (sgs->group_type == group_overloaded)
9861 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9862 sgs->group_capacity;
9863 }
9864
9865 /**
9866 * update_sd_pick_busiest - return 1 on busiest group
9867 * @env: The load balancing environment.
9868 * @sds: sched_domain statistics
9869 * @sg: sched_group candidate to be checked for being the busiest
9870 * @sgs: sched_group statistics
9871 *
9872 * Determine if @sg is a busier group than the previously selected
9873 * busiest group.
9874 *
9875 * Return: %true if @sg is a busier group than the previously selected
9876 * busiest group. %false otherwise.
9877 */
update_sd_pick_busiest(struct lb_env * env,struct sd_lb_stats * sds,struct sched_group * sg,struct sg_lb_stats * sgs)9878 static bool update_sd_pick_busiest(struct lb_env *env,
9879 struct sd_lb_stats *sds,
9880 struct sched_group *sg,
9881 struct sg_lb_stats *sgs)
9882 {
9883 struct sg_lb_stats *busiest = &sds->busiest_stat;
9884
9885 /* Make sure that there is at least one task to pull */
9886 if (!sgs->sum_h_nr_running)
9887 return false;
9888
9889 /*
9890 * Don't try to pull misfit tasks we can't help.
9891 * We can use max_capacity here as reduction in capacity on some
9892 * CPUs in the group should either be possible to resolve
9893 * internally or be covered by avg_load imbalance (eventually).
9894 */
9895 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9896 (sgs->group_type == group_misfit_task) &&
9897 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
9898 sds->local_stat.group_type != group_has_spare))
9899 return false;
9900
9901 if (sgs->group_type > busiest->group_type)
9902 return true;
9903
9904 if (sgs->group_type < busiest->group_type)
9905 return false;
9906
9907 /*
9908 * The candidate and the current busiest group are the same type of
9909 * group. Let check which one is the busiest according to the type.
9910 */
9911
9912 switch (sgs->group_type) {
9913 case group_overloaded:
9914 /* Select the overloaded group with highest avg_load. */
9915 if (sgs->avg_load <= busiest->avg_load)
9916 return false;
9917 break;
9918
9919 case group_imbalanced:
9920 /*
9921 * Select the 1st imbalanced group as we don't have any way to
9922 * choose one more than another.
9923 */
9924 return false;
9925
9926 case group_asym_packing:
9927 /* Prefer to move from lowest priority CPU's work */
9928 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
9929 return false;
9930 break;
9931
9932 case group_misfit_task:
9933 /*
9934 * If we have more than one misfit sg go with the biggest
9935 * misfit.
9936 */
9937 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
9938 return false;
9939 break;
9940
9941 case group_smt_balance:
9942 /*
9943 * Check if we have spare CPUs on either SMT group to
9944 * choose has spare or fully busy handling.
9945 */
9946 if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)
9947 goto has_spare;
9948
9949 fallthrough;
9950
9951 case group_fully_busy:
9952 /*
9953 * Select the fully busy group with highest avg_load. In
9954 * theory, there is no need to pull task from such kind of
9955 * group because tasks have all compute capacity that they need
9956 * but we can still improve the overall throughput by reducing
9957 * contention when accessing shared HW resources.
9958 *
9959 * XXX for now avg_load is not computed and always 0 so we
9960 * select the 1st one, except if @sg is composed of SMT
9961 * siblings.
9962 */
9963
9964 if (sgs->avg_load < busiest->avg_load)
9965 return false;
9966
9967 if (sgs->avg_load == busiest->avg_load) {
9968 /*
9969 * SMT sched groups need more help than non-SMT groups.
9970 * If @sg happens to also be SMT, either choice is good.
9971 */
9972 if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)
9973 return false;
9974 }
9975
9976 break;
9977
9978 case group_has_spare:
9979 /*
9980 * Do not pick sg with SMT CPUs over sg with pure CPUs,
9981 * as we do not want to pull task off SMT core with one task
9982 * and make the core idle.
9983 */
9984 if (smt_vs_nonsmt_groups(sds->busiest, sg)) {
9985 if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)
9986 return false;
9987 else
9988 return true;
9989 }
9990 has_spare:
9991
9992 /*
9993 * Select not overloaded group with lowest number of idle cpus
9994 * and highest number of running tasks. We could also compare
9995 * the spare capacity which is more stable but it can end up
9996 * that the group has less spare capacity but finally more idle
9997 * CPUs which means less opportunity to pull tasks.
9998 */
9999 if (sgs->idle_cpus > busiest->idle_cpus)
10000 return false;
10001 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
10002 (sgs->sum_nr_running <= busiest->sum_nr_running))
10003 return false;
10004
10005 break;
10006 }
10007
10008 /*
10009 * Candidate sg has no more than one task per CPU and has higher
10010 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
10011 * throughput. Maximize throughput, power/energy consequences are not
10012 * considered.
10013 */
10014 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
10015 (sgs->group_type <= group_fully_busy) &&
10016 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
10017 return false;
10018
10019 return true;
10020 }
10021
10022 #ifdef CONFIG_NUMA_BALANCING
fbq_classify_group(struct sg_lb_stats * sgs)10023 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10024 {
10025 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
10026 return regular;
10027 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
10028 return remote;
10029 return all;
10030 }
10031
fbq_classify_rq(struct rq * rq)10032 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10033 {
10034 if (rq->nr_running > rq->nr_numa_running)
10035 return regular;
10036 if (rq->nr_running > rq->nr_preferred_running)
10037 return remote;
10038 return all;
10039 }
10040 #else
fbq_classify_group(struct sg_lb_stats * sgs)10041 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10042 {
10043 return all;
10044 }
10045
fbq_classify_rq(struct rq * rq)10046 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10047 {
10048 return regular;
10049 }
10050 #endif /* CONFIG_NUMA_BALANCING */
10051
10052
10053 struct sg_lb_stats;
10054
10055 /*
10056 * task_running_on_cpu - return 1 if @p is running on @cpu.
10057 */
10058
task_running_on_cpu(int cpu,struct task_struct * p)10059 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
10060 {
10061 /* Task has no contribution or is new */
10062 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
10063 return 0;
10064
10065 if (task_on_rq_queued(p))
10066 return 1;
10067
10068 return 0;
10069 }
10070
10071 /**
10072 * idle_cpu_without - would a given CPU be idle without p ?
10073 * @cpu: the processor on which idleness is tested.
10074 * @p: task which should be ignored.
10075 *
10076 * Return: 1 if the CPU would be idle. 0 otherwise.
10077 */
idle_cpu_without(int cpu,struct task_struct * p)10078 static int idle_cpu_without(int cpu, struct task_struct *p)
10079 {
10080 struct rq *rq = cpu_rq(cpu);
10081
10082 if (rq->curr != rq->idle && rq->curr != p)
10083 return 0;
10084
10085 /*
10086 * rq->nr_running can't be used but an updated version without the
10087 * impact of p on cpu must be used instead. The updated nr_running
10088 * be computed and tested before calling idle_cpu_without().
10089 */
10090
10091 #ifdef CONFIG_SMP
10092 if (rq->ttwu_pending)
10093 return 0;
10094 #endif
10095
10096 return 1;
10097 }
10098
10099 /*
10100 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
10101 * @sd: The sched_domain level to look for idlest group.
10102 * @group: sched_group whose statistics are to be updated.
10103 * @sgs: variable to hold the statistics for this group.
10104 * @p: The task for which we look for the idlest group/CPU.
10105 */
update_sg_wakeup_stats(struct sched_domain * sd,struct sched_group * group,struct sg_lb_stats * sgs,struct task_struct * p)10106 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
10107 struct sched_group *group,
10108 struct sg_lb_stats *sgs,
10109 struct task_struct *p)
10110 {
10111 int i, nr_running;
10112
10113 memset(sgs, 0, sizeof(*sgs));
10114
10115 /* Assume that task can't fit any CPU of the group */
10116 if (sd->flags & SD_ASYM_CPUCAPACITY)
10117 sgs->group_misfit_task_load = 1;
10118
10119 for_each_cpu(i, sched_group_span(group)) {
10120 struct rq *rq = cpu_rq(i);
10121 unsigned int local;
10122
10123 sgs->group_load += cpu_load_without(rq, p);
10124 sgs->group_util += cpu_util_without(i, p);
10125 sgs->group_runnable += cpu_runnable_without(rq, p);
10126 local = task_running_on_cpu(i, p);
10127 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
10128
10129 nr_running = rq->nr_running - local;
10130 sgs->sum_nr_running += nr_running;
10131
10132 /*
10133 * No need to call idle_cpu_without() if nr_running is not 0
10134 */
10135 if (!nr_running && idle_cpu_without(i, p))
10136 sgs->idle_cpus++;
10137
10138 /* Check if task fits in the CPU */
10139 if (sd->flags & SD_ASYM_CPUCAPACITY &&
10140 sgs->group_misfit_task_load &&
10141 task_fits_cpu(p, i))
10142 sgs->group_misfit_task_load = 0;
10143
10144 }
10145
10146 sgs->group_capacity = group->sgc->capacity;
10147
10148 sgs->group_weight = group->group_weight;
10149
10150 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
10151
10152 /*
10153 * Computing avg_load makes sense only when group is fully busy or
10154 * overloaded
10155 */
10156 if (sgs->group_type == group_fully_busy ||
10157 sgs->group_type == group_overloaded)
10158 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10159 sgs->group_capacity;
10160 }
10161
update_pick_idlest(struct sched_group * idlest,struct sg_lb_stats * idlest_sgs,struct sched_group * group,struct sg_lb_stats * sgs)10162 static bool update_pick_idlest(struct sched_group *idlest,
10163 struct sg_lb_stats *idlest_sgs,
10164 struct sched_group *group,
10165 struct sg_lb_stats *sgs)
10166 {
10167 if (sgs->group_type < idlest_sgs->group_type)
10168 return true;
10169
10170 if (sgs->group_type > idlest_sgs->group_type)
10171 return false;
10172
10173 /*
10174 * The candidate and the current idlest group are the same type of
10175 * group. Let check which one is the idlest according to the type.
10176 */
10177
10178 switch (sgs->group_type) {
10179 case group_overloaded:
10180 case group_fully_busy:
10181 /* Select the group with lowest avg_load. */
10182 if (idlest_sgs->avg_load <= sgs->avg_load)
10183 return false;
10184 break;
10185
10186 case group_imbalanced:
10187 case group_asym_packing:
10188 case group_smt_balance:
10189 /* Those types are not used in the slow wakeup path */
10190 return false;
10191
10192 case group_misfit_task:
10193 /* Select group with the highest max capacity */
10194 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
10195 return false;
10196 break;
10197
10198 case group_has_spare:
10199 /* Select group with most idle CPUs */
10200 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
10201 return false;
10202
10203 /* Select group with lowest group_util */
10204 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
10205 idlest_sgs->group_util <= sgs->group_util)
10206 return false;
10207
10208 break;
10209 }
10210
10211 return true;
10212 }
10213
10214 /*
10215 * find_idlest_group() finds and returns the least busy CPU group within the
10216 * domain.
10217 *
10218 * Assumes p is allowed on at least one CPU in sd.
10219 */
10220 static struct sched_group *
find_idlest_group(struct sched_domain * sd,struct task_struct * p,int this_cpu)10221 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
10222 {
10223 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
10224 struct sg_lb_stats local_sgs, tmp_sgs;
10225 struct sg_lb_stats *sgs;
10226 unsigned long imbalance;
10227 struct sg_lb_stats idlest_sgs = {
10228 .avg_load = UINT_MAX,
10229 .group_type = group_overloaded,
10230 };
10231
10232 do {
10233 int local_group;
10234
10235 /* Skip over this group if it has no CPUs allowed */
10236 if (!cpumask_intersects(sched_group_span(group),
10237 p->cpus_ptr))
10238 continue;
10239
10240 /* Skip over this group if no cookie matched */
10241 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
10242 continue;
10243
10244 local_group = cpumask_test_cpu(this_cpu,
10245 sched_group_span(group));
10246
10247 if (local_group) {
10248 sgs = &local_sgs;
10249 local = group;
10250 } else {
10251 sgs = &tmp_sgs;
10252 }
10253
10254 update_sg_wakeup_stats(sd, group, sgs, p);
10255
10256 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
10257 idlest = group;
10258 idlest_sgs = *sgs;
10259 }
10260
10261 } while (group = group->next, group != sd->groups);
10262
10263
10264 /* There is no idlest group to push tasks to */
10265 if (!idlest)
10266 return NULL;
10267
10268 /* The local group has been skipped because of CPU affinity */
10269 if (!local)
10270 return idlest;
10271
10272 /*
10273 * If the local group is idler than the selected idlest group
10274 * don't try and push the task.
10275 */
10276 if (local_sgs.group_type < idlest_sgs.group_type)
10277 return NULL;
10278
10279 /*
10280 * If the local group is busier than the selected idlest group
10281 * try and push the task.
10282 */
10283 if (local_sgs.group_type > idlest_sgs.group_type)
10284 return idlest;
10285
10286 switch (local_sgs.group_type) {
10287 case group_overloaded:
10288 case group_fully_busy:
10289
10290 /* Calculate allowed imbalance based on load */
10291 imbalance = scale_load_down(NICE_0_LOAD) *
10292 (sd->imbalance_pct-100) / 100;
10293
10294 /*
10295 * When comparing groups across NUMA domains, it's possible for
10296 * the local domain to be very lightly loaded relative to the
10297 * remote domains but "imbalance" skews the comparison making
10298 * remote CPUs look much more favourable. When considering
10299 * cross-domain, add imbalance to the load on the remote node
10300 * and consider staying local.
10301 */
10302
10303 if ((sd->flags & SD_NUMA) &&
10304 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
10305 return NULL;
10306
10307 /*
10308 * If the local group is less loaded than the selected
10309 * idlest group don't try and push any tasks.
10310 */
10311 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
10312 return NULL;
10313
10314 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
10315 return NULL;
10316 break;
10317
10318 case group_imbalanced:
10319 case group_asym_packing:
10320 case group_smt_balance:
10321 /* Those type are not used in the slow wakeup path */
10322 return NULL;
10323
10324 case group_misfit_task:
10325 /* Select group with the highest max capacity */
10326 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
10327 return NULL;
10328 break;
10329
10330 case group_has_spare:
10331 #ifdef CONFIG_NUMA
10332 if (sd->flags & SD_NUMA) {
10333 int imb_numa_nr = sd->imb_numa_nr;
10334 #ifdef CONFIG_NUMA_BALANCING
10335 int idlest_cpu;
10336 /*
10337 * If there is spare capacity at NUMA, try to select
10338 * the preferred node
10339 */
10340 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
10341 return NULL;
10342
10343 idlest_cpu = cpumask_first(sched_group_span(idlest));
10344 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
10345 return idlest;
10346 #endif /* CONFIG_NUMA_BALANCING */
10347 /*
10348 * Otherwise, keep the task close to the wakeup source
10349 * and improve locality if the number of running tasks
10350 * would remain below threshold where an imbalance is
10351 * allowed while accounting for the possibility the
10352 * task is pinned to a subset of CPUs. If there is a
10353 * real need of migration, periodic load balance will
10354 * take care of it.
10355 */
10356 if (p->nr_cpus_allowed != NR_CPUS) {
10357 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
10358
10359 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
10360 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
10361 }
10362
10363 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
10364 if (!adjust_numa_imbalance(imbalance,
10365 local_sgs.sum_nr_running + 1,
10366 imb_numa_nr)) {
10367 return NULL;
10368 }
10369 }
10370 #endif /* CONFIG_NUMA */
10371
10372 /*
10373 * Select group with highest number of idle CPUs. We could also
10374 * compare the utilization which is more stable but it can end
10375 * up that the group has less spare capacity but finally more
10376 * idle CPUs which means more opportunity to run task.
10377 */
10378 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
10379 return NULL;
10380 break;
10381 }
10382
10383 return idlest;
10384 }
10385
update_idle_cpu_scan(struct lb_env * env,unsigned long sum_util)10386 static void update_idle_cpu_scan(struct lb_env *env,
10387 unsigned long sum_util)
10388 {
10389 struct sched_domain_shared *sd_share;
10390 int llc_weight, pct;
10391 u64 x, y, tmp;
10392 /*
10393 * Update the number of CPUs to scan in LLC domain, which could
10394 * be used as a hint in select_idle_cpu(). The update of sd_share
10395 * could be expensive because it is within a shared cache line.
10396 * So the write of this hint only occurs during periodic load
10397 * balancing, rather than CPU_NEWLY_IDLE, because the latter
10398 * can fire way more frequently than the former.
10399 */
10400 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
10401 return;
10402
10403 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
10404 if (env->sd->span_weight != llc_weight)
10405 return;
10406
10407 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
10408 if (!sd_share)
10409 return;
10410
10411 /*
10412 * The number of CPUs to search drops as sum_util increases, when
10413 * sum_util hits 85% or above, the scan stops.
10414 * The reason to choose 85% as the threshold is because this is the
10415 * imbalance_pct(117) when a LLC sched group is overloaded.
10416 *
10417 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
10418 * and y'= y / SCHED_CAPACITY_SCALE
10419 *
10420 * x is the ratio of sum_util compared to the CPU capacity:
10421 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
10422 * y' is the ratio of CPUs to be scanned in the LLC domain,
10423 * and the number of CPUs to scan is calculated by:
10424 *
10425 * nr_scan = llc_weight * y' [2]
10426 *
10427 * When x hits the threshold of overloaded, AKA, when
10428 * x = 100 / pct, y drops to 0. According to [1],
10429 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
10430 *
10431 * Scale x by SCHED_CAPACITY_SCALE:
10432 * x' = sum_util / llc_weight; [3]
10433 *
10434 * and finally [1] becomes:
10435 * y = SCHED_CAPACITY_SCALE -
10436 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
10437 *
10438 */
10439 /* equation [3] */
10440 x = sum_util;
10441 do_div(x, llc_weight);
10442
10443 /* equation [4] */
10444 pct = env->sd->imbalance_pct;
10445 tmp = x * x * pct * pct;
10446 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
10447 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
10448 y = SCHED_CAPACITY_SCALE - tmp;
10449
10450 /* equation [2] */
10451 y *= llc_weight;
10452 do_div(y, SCHED_CAPACITY_SCALE);
10453 if ((int)y != sd_share->nr_idle_scan)
10454 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
10455 }
10456
10457 /**
10458 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
10459 * @env: The load balancing environment.
10460 * @sds: variable to hold the statistics for this sched_domain.
10461 */
10462
update_sd_lb_stats(struct lb_env * env,struct sd_lb_stats * sds)10463 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
10464 {
10465 struct sched_group *sg = env->sd->groups;
10466 struct sg_lb_stats *local = &sds->local_stat;
10467 struct sg_lb_stats tmp_sgs;
10468 unsigned long sum_util = 0;
10469 int sg_status = 0;
10470
10471 do {
10472 struct sg_lb_stats *sgs = &tmp_sgs;
10473 int local_group;
10474
10475 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
10476 if (local_group) {
10477 sds->local = sg;
10478 sgs = local;
10479
10480 if (env->idle != CPU_NEWLY_IDLE ||
10481 time_after_eq(jiffies, sg->sgc->next_update))
10482 update_group_capacity(env->sd, env->dst_cpu);
10483 }
10484
10485 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
10486
10487 if (local_group)
10488 goto next_group;
10489
10490
10491 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
10492 sds->busiest = sg;
10493 sds->busiest_stat = *sgs;
10494 }
10495
10496 next_group:
10497 /* Now, start updating sd_lb_stats */
10498 sds->total_load += sgs->group_load;
10499 sds->total_capacity += sgs->group_capacity;
10500
10501 sum_util += sgs->group_util;
10502 sg = sg->next;
10503 } while (sg != env->sd->groups);
10504
10505 /*
10506 * Indicate that the child domain of the busiest group prefers tasks
10507 * go to a child's sibling domains first. NB the flags of a sched group
10508 * are those of the child domain.
10509 */
10510 if (sds->busiest)
10511 sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);
10512
10513
10514 if (env->sd->flags & SD_NUMA)
10515 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
10516
10517 if (!env->sd->parent) {
10518 /* update overload indicator if we are at root domain */
10519 WRITE_ONCE(env->dst_rq->rd->overload, sg_status & SG_OVERLOAD);
10520
10521 /* Update over-utilization (tipping point, U >= 0) indicator */
10522 set_rd_overutilized_status(env->dst_rq->rd,
10523 sg_status & SG_OVERUTILIZED);
10524 } else if (sg_status & SG_OVERUTILIZED) {
10525 set_rd_overutilized_status(env->dst_rq->rd, SG_OVERUTILIZED);
10526 }
10527
10528 update_idle_cpu_scan(env, sum_util);
10529 }
10530
10531 /**
10532 * calculate_imbalance - Calculate the amount of imbalance present within the
10533 * groups of a given sched_domain during load balance.
10534 * @env: load balance environment
10535 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
10536 */
calculate_imbalance(struct lb_env * env,struct sd_lb_stats * sds)10537 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
10538 {
10539 struct sg_lb_stats *local, *busiest;
10540
10541 local = &sds->local_stat;
10542 busiest = &sds->busiest_stat;
10543
10544 if (busiest->group_type == group_misfit_task) {
10545 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
10546 /* Set imbalance to allow misfit tasks to be balanced. */
10547 env->migration_type = migrate_misfit;
10548 env->imbalance = 1;
10549 } else {
10550 /*
10551 * Set load imbalance to allow moving task from cpu
10552 * with reduced capacity.
10553 */
10554 env->migration_type = migrate_load;
10555 env->imbalance = busiest->group_misfit_task_load;
10556 }
10557 return;
10558 }
10559
10560 if (busiest->group_type == group_asym_packing) {
10561 /*
10562 * In case of asym capacity, we will try to migrate all load to
10563 * the preferred CPU.
10564 */
10565 env->migration_type = migrate_task;
10566 env->imbalance = busiest->sum_h_nr_running;
10567 return;
10568 }
10569
10570 if (busiest->group_type == group_smt_balance) {
10571 /* Reduce number of tasks sharing CPU capacity */
10572 env->migration_type = migrate_task;
10573 env->imbalance = 1;
10574 return;
10575 }
10576
10577 if (busiest->group_type == group_imbalanced) {
10578 /*
10579 * In the group_imb case we cannot rely on group-wide averages
10580 * to ensure CPU-load equilibrium, try to move any task to fix
10581 * the imbalance. The next load balance will take care of
10582 * balancing back the system.
10583 */
10584 env->migration_type = migrate_task;
10585 env->imbalance = 1;
10586 return;
10587 }
10588
10589 /*
10590 * Try to use spare capacity of local group without overloading it or
10591 * emptying busiest.
10592 */
10593 if (local->group_type == group_has_spare) {
10594 if ((busiest->group_type > group_fully_busy) &&
10595 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
10596 /*
10597 * If busiest is overloaded, try to fill spare
10598 * capacity. This might end up creating spare capacity
10599 * in busiest or busiest still being overloaded but
10600 * there is no simple way to directly compute the
10601 * amount of load to migrate in order to balance the
10602 * system.
10603 */
10604 env->migration_type = migrate_util;
10605 env->imbalance = max(local->group_capacity, local->group_util) -
10606 local->group_util;
10607
10608 /*
10609 * In some cases, the group's utilization is max or even
10610 * higher than capacity because of migrations but the
10611 * local CPU is (newly) idle. There is at least one
10612 * waiting task in this overloaded busiest group. Let's
10613 * try to pull it.
10614 */
10615 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
10616 env->migration_type = migrate_task;
10617 env->imbalance = 1;
10618 }
10619
10620 return;
10621 }
10622
10623 if (busiest->group_weight == 1 || sds->prefer_sibling) {
10624 /*
10625 * When prefer sibling, evenly spread running tasks on
10626 * groups.
10627 */
10628 env->migration_type = migrate_task;
10629 env->imbalance = sibling_imbalance(env, sds, busiest, local);
10630 } else {
10631
10632 /*
10633 * If there is no overload, we just want to even the number of
10634 * idle cpus.
10635 */
10636 env->migration_type = migrate_task;
10637 env->imbalance = max_t(long, 0,
10638 (local->idle_cpus - busiest->idle_cpus));
10639 }
10640
10641 #ifdef CONFIG_NUMA
10642 /* Consider allowing a small imbalance between NUMA groups */
10643 if (env->sd->flags & SD_NUMA) {
10644 env->imbalance = adjust_numa_imbalance(env->imbalance,
10645 local->sum_nr_running + 1,
10646 env->sd->imb_numa_nr);
10647 }
10648 #endif
10649
10650 /* Number of tasks to move to restore balance */
10651 env->imbalance >>= 1;
10652
10653 return;
10654 }
10655
10656 /*
10657 * Local is fully busy but has to take more load to relieve the
10658 * busiest group
10659 */
10660 if (local->group_type < group_overloaded) {
10661 /*
10662 * Local will become overloaded so the avg_load metrics are
10663 * finally needed.
10664 */
10665
10666 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10667 local->group_capacity;
10668
10669 /*
10670 * If the local group is more loaded than the selected
10671 * busiest group don't try to pull any tasks.
10672 */
10673 if (local->avg_load >= busiest->avg_load) {
10674 env->imbalance = 0;
10675 return;
10676 }
10677
10678 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10679 sds->total_capacity;
10680
10681 /*
10682 * If the local group is more loaded than the average system
10683 * load, don't try to pull any tasks.
10684 */
10685 if (local->avg_load >= sds->avg_load) {
10686 env->imbalance = 0;
10687 return;
10688 }
10689
10690 }
10691
10692 /*
10693 * Both group are or will become overloaded and we're trying to get all
10694 * the CPUs to the average_load, so we don't want to push ourselves
10695 * above the average load, nor do we wish to reduce the max loaded CPU
10696 * below the average load. At the same time, we also don't want to
10697 * reduce the group load below the group capacity. Thus we look for
10698 * the minimum possible imbalance.
10699 */
10700 env->migration_type = migrate_load;
10701 env->imbalance = min(
10702 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10703 (sds->avg_load - local->avg_load) * local->group_capacity
10704 ) / SCHED_CAPACITY_SCALE;
10705 }
10706
10707 /******* find_busiest_group() helpers end here *********************/
10708
10709 /*
10710 * Decision matrix according to the local and busiest group type:
10711 *
10712 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10713 * has_spare nr_idle balanced N/A N/A balanced balanced
10714 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
10715 * misfit_task force N/A N/A N/A N/A N/A
10716 * asym_packing force force N/A N/A force force
10717 * imbalanced force force N/A N/A force force
10718 * overloaded force force N/A N/A force avg_load
10719 *
10720 * N/A : Not Applicable because already filtered while updating
10721 * statistics.
10722 * balanced : The system is balanced for these 2 groups.
10723 * force : Calculate the imbalance as load migration is probably needed.
10724 * avg_load : Only if imbalance is significant enough.
10725 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
10726 * different in groups.
10727 */
10728
10729 /**
10730 * find_busiest_group - Returns the busiest group within the sched_domain
10731 * if there is an imbalance.
10732 * @env: The load balancing environment.
10733 *
10734 * Also calculates the amount of runnable load which should be moved
10735 * to restore balance.
10736 *
10737 * Return: - The busiest group if imbalance exists.
10738 */
find_busiest_group(struct lb_env * env)10739 static struct sched_group *find_busiest_group(struct lb_env *env)
10740 {
10741 struct sg_lb_stats *local, *busiest;
10742 struct sd_lb_stats sds;
10743
10744 init_sd_lb_stats(&sds);
10745
10746 /*
10747 * Compute the various statistics relevant for load balancing at
10748 * this level.
10749 */
10750 update_sd_lb_stats(env, &sds);
10751
10752 /* There is no busy sibling group to pull tasks from */
10753 if (!sds.busiest)
10754 goto out_balanced;
10755
10756 busiest = &sds.busiest_stat;
10757
10758 /* Misfit tasks should be dealt with regardless of the avg load */
10759 if (busiest->group_type == group_misfit_task)
10760 goto force_balance;
10761
10762 if (sched_energy_enabled()) {
10763 struct root_domain *rd = env->dst_rq->rd;
10764
10765 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
10766 goto out_balanced;
10767 }
10768
10769 /* ASYM feature bypasses nice load balance check */
10770 if (busiest->group_type == group_asym_packing)
10771 goto force_balance;
10772
10773 /*
10774 * If the busiest group is imbalanced the below checks don't
10775 * work because they assume all things are equal, which typically
10776 * isn't true due to cpus_ptr constraints and the like.
10777 */
10778 if (busiest->group_type == group_imbalanced)
10779 goto force_balance;
10780
10781 local = &sds.local_stat;
10782 /*
10783 * If the local group is busier than the selected busiest group
10784 * don't try and pull any tasks.
10785 */
10786 if (local->group_type > busiest->group_type)
10787 goto out_balanced;
10788
10789 /*
10790 * When groups are overloaded, use the avg_load to ensure fairness
10791 * between tasks.
10792 */
10793 if (local->group_type == group_overloaded) {
10794 /*
10795 * If the local group is more loaded than the selected
10796 * busiest group don't try to pull any tasks.
10797 */
10798 if (local->avg_load >= busiest->avg_load)
10799 goto out_balanced;
10800
10801 /* XXX broken for overlapping NUMA groups */
10802 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10803 sds.total_capacity;
10804
10805 /*
10806 * Don't pull any tasks if this group is already above the
10807 * domain average load.
10808 */
10809 if (local->avg_load >= sds.avg_load)
10810 goto out_balanced;
10811
10812 /*
10813 * If the busiest group is more loaded, use imbalance_pct to be
10814 * conservative.
10815 */
10816 if (100 * busiest->avg_load <=
10817 env->sd->imbalance_pct * local->avg_load)
10818 goto out_balanced;
10819 }
10820
10821 /*
10822 * Try to move all excess tasks to a sibling domain of the busiest
10823 * group's child domain.
10824 */
10825 if (sds.prefer_sibling && local->group_type == group_has_spare &&
10826 sibling_imbalance(env, &sds, busiest, local) > 1)
10827 goto force_balance;
10828
10829 if (busiest->group_type != group_overloaded) {
10830 if (env->idle == CPU_NOT_IDLE) {
10831 /*
10832 * If the busiest group is not overloaded (and as a
10833 * result the local one too) but this CPU is already
10834 * busy, let another idle CPU try to pull task.
10835 */
10836 goto out_balanced;
10837 }
10838
10839 if (busiest->group_type == group_smt_balance &&
10840 smt_vs_nonsmt_groups(sds.local, sds.busiest)) {
10841 /* Let non SMT CPU pull from SMT CPU sharing with sibling */
10842 goto force_balance;
10843 }
10844
10845 if (busiest->group_weight > 1 &&
10846 local->idle_cpus <= (busiest->idle_cpus + 1)) {
10847 /*
10848 * If the busiest group is not overloaded
10849 * and there is no imbalance between this and busiest
10850 * group wrt idle CPUs, it is balanced. The imbalance
10851 * becomes significant if the diff is greater than 1
10852 * otherwise we might end up to just move the imbalance
10853 * on another group. Of course this applies only if
10854 * there is more than 1 CPU per group.
10855 */
10856 goto out_balanced;
10857 }
10858
10859 if (busiest->sum_h_nr_running == 1) {
10860 /*
10861 * busiest doesn't have any tasks waiting to run
10862 */
10863 goto out_balanced;
10864 }
10865 }
10866
10867 force_balance:
10868 /* Looks like there is an imbalance. Compute it */
10869 calculate_imbalance(env, &sds);
10870 return env->imbalance ? sds.busiest : NULL;
10871
10872 out_balanced:
10873 env->imbalance = 0;
10874 return NULL;
10875 }
10876
10877 /*
10878 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10879 */
find_busiest_queue(struct lb_env * env,struct sched_group * group)10880 static struct rq *find_busiest_queue(struct lb_env *env,
10881 struct sched_group *group)
10882 {
10883 struct rq *busiest = NULL, *rq;
10884 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
10885 unsigned int busiest_nr = 0;
10886 int i;
10887
10888 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10889 unsigned long capacity, load, util;
10890 unsigned int nr_running;
10891 enum fbq_type rt;
10892
10893 rq = cpu_rq(i);
10894 rt = fbq_classify_rq(rq);
10895
10896 /*
10897 * We classify groups/runqueues into three groups:
10898 * - regular: there are !numa tasks
10899 * - remote: there are numa tasks that run on the 'wrong' node
10900 * - all: there is no distinction
10901 *
10902 * In order to avoid migrating ideally placed numa tasks,
10903 * ignore those when there's better options.
10904 *
10905 * If we ignore the actual busiest queue to migrate another
10906 * task, the next balance pass can still reduce the busiest
10907 * queue by moving tasks around inside the node.
10908 *
10909 * If we cannot move enough load due to this classification
10910 * the next pass will adjust the group classification and
10911 * allow migration of more tasks.
10912 *
10913 * Both cases only affect the total convergence complexity.
10914 */
10915 if (rt > env->fbq_type)
10916 continue;
10917
10918 nr_running = rq->cfs.h_nr_running;
10919 if (!nr_running)
10920 continue;
10921
10922 capacity = capacity_of(i);
10923
10924 /*
10925 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
10926 * eventually lead to active_balancing high->low capacity.
10927 * Higher per-CPU capacity is considered better than balancing
10928 * average load.
10929 */
10930 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
10931 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
10932 nr_running == 1)
10933 continue;
10934
10935 /*
10936 * Make sure we only pull tasks from a CPU of lower priority
10937 * when balancing between SMT siblings.
10938 *
10939 * If balancing between cores, let lower priority CPUs help
10940 * SMT cores with more than one busy sibling.
10941 */
10942 if ((env->sd->flags & SD_ASYM_PACKING) &&
10943 sched_use_asym_prio(env->sd, i) &&
10944 sched_asym_prefer(i, env->dst_cpu) &&
10945 nr_running == 1)
10946 continue;
10947
10948 switch (env->migration_type) {
10949 case migrate_load:
10950 /*
10951 * When comparing with load imbalance, use cpu_load()
10952 * which is not scaled with the CPU capacity.
10953 */
10954 load = cpu_load(rq);
10955
10956 if (nr_running == 1 && load > env->imbalance &&
10957 !check_cpu_capacity(rq, env->sd))
10958 break;
10959
10960 /*
10961 * For the load comparisons with the other CPUs,
10962 * consider the cpu_load() scaled with the CPU
10963 * capacity, so that the load can be moved away
10964 * from the CPU that is potentially running at a
10965 * lower capacity.
10966 *
10967 * Thus we're looking for max(load_i / capacity_i),
10968 * crosswise multiplication to rid ourselves of the
10969 * division works out to:
10970 * load_i * capacity_j > load_j * capacity_i;
10971 * where j is our previous maximum.
10972 */
10973 if (load * busiest_capacity > busiest_load * capacity) {
10974 busiest_load = load;
10975 busiest_capacity = capacity;
10976 busiest = rq;
10977 }
10978 break;
10979
10980 case migrate_util:
10981 util = cpu_util_cfs_boost(i);
10982
10983 /*
10984 * Don't try to pull utilization from a CPU with one
10985 * running task. Whatever its utilization, we will fail
10986 * detach the task.
10987 */
10988 if (nr_running <= 1)
10989 continue;
10990
10991 if (busiest_util < util) {
10992 busiest_util = util;
10993 busiest = rq;
10994 }
10995 break;
10996
10997 case migrate_task:
10998 if (busiest_nr < nr_running) {
10999 busiest_nr = nr_running;
11000 busiest = rq;
11001 }
11002 break;
11003
11004 case migrate_misfit:
11005 /*
11006 * For ASYM_CPUCAPACITY domains with misfit tasks we
11007 * simply seek the "biggest" misfit task.
11008 */
11009 if (rq->misfit_task_load > busiest_load) {
11010 busiest_load = rq->misfit_task_load;
11011 busiest = rq;
11012 }
11013
11014 break;
11015
11016 }
11017 }
11018
11019 return busiest;
11020 }
11021
11022 /*
11023 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
11024 * so long as it is large enough.
11025 */
11026 #define MAX_PINNED_INTERVAL 512
11027
11028 static inline bool
asym_active_balance(struct lb_env * env)11029 asym_active_balance(struct lb_env *env)
11030 {
11031 /*
11032 * ASYM_PACKING needs to force migrate tasks from busy but lower
11033 * priority CPUs in order to pack all tasks in the highest priority
11034 * CPUs. When done between cores, do it only if the whole core if the
11035 * whole core is idle.
11036 *
11037 * If @env::src_cpu is an SMT core with busy siblings, let
11038 * the lower priority @env::dst_cpu help it. Do not follow
11039 * CPU priority.
11040 */
11041 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
11042 sched_use_asym_prio(env->sd, env->dst_cpu) &&
11043 (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
11044 !sched_use_asym_prio(env->sd, env->src_cpu));
11045 }
11046
11047 static inline bool
imbalanced_active_balance(struct lb_env * env)11048 imbalanced_active_balance(struct lb_env *env)
11049 {
11050 struct sched_domain *sd = env->sd;
11051
11052 /*
11053 * The imbalanced case includes the case of pinned tasks preventing a fair
11054 * distribution of the load on the system but also the even distribution of the
11055 * threads on a system with spare capacity
11056 */
11057 if ((env->migration_type == migrate_task) &&
11058 (sd->nr_balance_failed > sd->cache_nice_tries+2))
11059 return 1;
11060
11061 return 0;
11062 }
11063
need_active_balance(struct lb_env * env)11064 static int need_active_balance(struct lb_env *env)
11065 {
11066 struct sched_domain *sd = env->sd;
11067
11068 if (asym_active_balance(env))
11069 return 1;
11070
11071 if (imbalanced_active_balance(env))
11072 return 1;
11073
11074 /*
11075 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
11076 * It's worth migrating the task if the src_cpu's capacity is reduced
11077 * because of other sched_class or IRQs if more capacity stays
11078 * available on dst_cpu.
11079 */
11080 if ((env->idle != CPU_NOT_IDLE) &&
11081 (env->src_rq->cfs.h_nr_running == 1)) {
11082 if ((check_cpu_capacity(env->src_rq, sd)) &&
11083 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
11084 return 1;
11085 }
11086
11087 if (env->migration_type == migrate_misfit)
11088 return 1;
11089
11090 return 0;
11091 }
11092
11093 static int active_load_balance_cpu_stop(void *data);
11094
should_we_balance(struct lb_env * env)11095 static int should_we_balance(struct lb_env *env)
11096 {
11097 struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask);
11098 struct sched_group *sg = env->sd->groups;
11099 int cpu, idle_smt = -1;
11100
11101 /*
11102 * Ensure the balancing environment is consistent; can happen
11103 * when the softirq triggers 'during' hotplug.
11104 */
11105 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
11106 return 0;
11107
11108 /*
11109 * In the newly idle case, we will allow all the CPUs
11110 * to do the newly idle load balance.
11111 *
11112 * However, we bail out if we already have tasks or a wakeup pending,
11113 * to optimize wakeup latency.
11114 */
11115 if (env->idle == CPU_NEWLY_IDLE) {
11116 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
11117 return 0;
11118 return 1;
11119 }
11120
11121 cpumask_copy(swb_cpus, group_balance_mask(sg));
11122 /* Try to find first idle CPU */
11123 for_each_cpu_and(cpu, swb_cpus, env->cpus) {
11124 if (!idle_cpu(cpu))
11125 continue;
11126
11127 /*
11128 * Don't balance to idle SMT in busy core right away when
11129 * balancing cores, but remember the first idle SMT CPU for
11130 * later consideration. Find CPU on an idle core first.
11131 */
11132 if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) {
11133 if (idle_smt == -1)
11134 idle_smt = cpu;
11135 /*
11136 * If the core is not idle, and first SMT sibling which is
11137 * idle has been found, then its not needed to check other
11138 * SMT siblings for idleness:
11139 */
11140 #ifdef CONFIG_SCHED_SMT
11141 cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu));
11142 #endif
11143 continue;
11144 }
11145
11146 /*
11147 * Are we the first idle core in a non-SMT domain or higher,
11148 * or the first idle CPU in a SMT domain?
11149 */
11150 return cpu == env->dst_cpu;
11151 }
11152
11153 /* Are we the first idle CPU with busy siblings? */
11154 if (idle_smt != -1)
11155 return idle_smt == env->dst_cpu;
11156
11157 /* Are we the first CPU of this group ? */
11158 return group_balance_cpu(sg) == env->dst_cpu;
11159 }
11160
11161 /*
11162 * Check this_cpu to ensure it is balanced within domain. Attempt to move
11163 * tasks if there is an imbalance.
11164 */
load_balance(int this_cpu,struct rq * this_rq,struct sched_domain * sd,enum cpu_idle_type idle,int * continue_balancing)11165 static int load_balance(int this_cpu, struct rq *this_rq,
11166 struct sched_domain *sd, enum cpu_idle_type idle,
11167 int *continue_balancing)
11168 {
11169 int ld_moved, cur_ld_moved, active_balance = 0;
11170 struct sched_domain *sd_parent = sd->parent;
11171 struct sched_group *group;
11172 struct rq *busiest;
11173 struct rq_flags rf;
11174 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
11175 struct lb_env env = {
11176 .sd = sd,
11177 .dst_cpu = this_cpu,
11178 .dst_rq = this_rq,
11179 .dst_grpmask = group_balance_mask(sd->groups),
11180 .idle = idle,
11181 .loop_break = SCHED_NR_MIGRATE_BREAK,
11182 .cpus = cpus,
11183 .fbq_type = all,
11184 .tasks = LIST_HEAD_INIT(env.tasks),
11185 };
11186
11187 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
11188
11189 schedstat_inc(sd->lb_count[idle]);
11190
11191 redo:
11192 if (!should_we_balance(&env)) {
11193 *continue_balancing = 0;
11194 goto out_balanced;
11195 }
11196
11197 group = find_busiest_group(&env);
11198 if (!group) {
11199 schedstat_inc(sd->lb_nobusyg[idle]);
11200 goto out_balanced;
11201 }
11202
11203 busiest = find_busiest_queue(&env, group);
11204 if (!busiest) {
11205 schedstat_inc(sd->lb_nobusyq[idle]);
11206 goto out_balanced;
11207 }
11208
11209 WARN_ON_ONCE(busiest == env.dst_rq);
11210
11211 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
11212
11213 env.src_cpu = busiest->cpu;
11214 env.src_rq = busiest;
11215
11216 ld_moved = 0;
11217 /* Clear this flag as soon as we find a pullable task */
11218 env.flags |= LBF_ALL_PINNED;
11219 if (busiest->nr_running > 1) {
11220 /*
11221 * Attempt to move tasks. If find_busiest_group has found
11222 * an imbalance but busiest->nr_running <= 1, the group is
11223 * still unbalanced. ld_moved simply stays zero, so it is
11224 * correctly treated as an imbalance.
11225 */
11226 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
11227
11228 more_balance:
11229 rq_lock_irqsave(busiest, &rf);
11230 update_rq_clock(busiest);
11231
11232 /*
11233 * cur_ld_moved - load moved in current iteration
11234 * ld_moved - cumulative load moved across iterations
11235 */
11236 cur_ld_moved = detach_tasks(&env);
11237
11238 /*
11239 * We've detached some tasks from busiest_rq. Every
11240 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
11241 * unlock busiest->lock, and we are able to be sure
11242 * that nobody can manipulate the tasks in parallel.
11243 * See task_rq_lock() family for the details.
11244 */
11245
11246 rq_unlock(busiest, &rf);
11247
11248 if (cur_ld_moved) {
11249 attach_tasks(&env);
11250 ld_moved += cur_ld_moved;
11251 }
11252
11253 local_irq_restore(rf.flags);
11254
11255 if (env.flags & LBF_NEED_BREAK) {
11256 env.flags &= ~LBF_NEED_BREAK;
11257 goto more_balance;
11258 }
11259
11260 /*
11261 * Revisit (affine) tasks on src_cpu that couldn't be moved to
11262 * us and move them to an alternate dst_cpu in our sched_group
11263 * where they can run. The upper limit on how many times we
11264 * iterate on same src_cpu is dependent on number of CPUs in our
11265 * sched_group.
11266 *
11267 * This changes load balance semantics a bit on who can move
11268 * load to a given_cpu. In addition to the given_cpu itself
11269 * (or a ilb_cpu acting on its behalf where given_cpu is
11270 * nohz-idle), we now have balance_cpu in a position to move
11271 * load to given_cpu. In rare situations, this may cause
11272 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
11273 * _independently_ and at _same_ time to move some load to
11274 * given_cpu) causing excess load to be moved to given_cpu.
11275 * This however should not happen so much in practice and
11276 * moreover subsequent load balance cycles should correct the
11277 * excess load moved.
11278 */
11279 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
11280
11281 /* Prevent to re-select dst_cpu via env's CPUs */
11282 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
11283
11284 env.dst_rq = cpu_rq(env.new_dst_cpu);
11285 env.dst_cpu = env.new_dst_cpu;
11286 env.flags &= ~LBF_DST_PINNED;
11287 env.loop = 0;
11288 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11289
11290 /*
11291 * Go back to "more_balance" rather than "redo" since we
11292 * need to continue with same src_cpu.
11293 */
11294 goto more_balance;
11295 }
11296
11297 /*
11298 * We failed to reach balance because of affinity.
11299 */
11300 if (sd_parent) {
11301 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11302
11303 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
11304 *group_imbalance = 1;
11305 }
11306
11307 /* All tasks on this runqueue were pinned by CPU affinity */
11308 if (unlikely(env.flags & LBF_ALL_PINNED)) {
11309 __cpumask_clear_cpu(cpu_of(busiest), cpus);
11310 /*
11311 * Attempting to continue load balancing at the current
11312 * sched_domain level only makes sense if there are
11313 * active CPUs remaining as possible busiest CPUs to
11314 * pull load from which are not contained within the
11315 * destination group that is receiving any migrated
11316 * load.
11317 */
11318 if (!cpumask_subset(cpus, env.dst_grpmask)) {
11319 env.loop = 0;
11320 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11321 goto redo;
11322 }
11323 goto out_all_pinned;
11324 }
11325 }
11326
11327 if (!ld_moved) {
11328 schedstat_inc(sd->lb_failed[idle]);
11329 /*
11330 * Increment the failure counter only on periodic balance.
11331 * We do not want newidle balance, which can be very
11332 * frequent, pollute the failure counter causing
11333 * excessive cache_hot migrations and active balances.
11334 */
11335 if (idle != CPU_NEWLY_IDLE)
11336 sd->nr_balance_failed++;
11337
11338 if (need_active_balance(&env)) {
11339 unsigned long flags;
11340
11341 raw_spin_rq_lock_irqsave(busiest, flags);
11342
11343 /*
11344 * Don't kick the active_load_balance_cpu_stop,
11345 * if the curr task on busiest CPU can't be
11346 * moved to this_cpu:
11347 */
11348 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
11349 raw_spin_rq_unlock_irqrestore(busiest, flags);
11350 goto out_one_pinned;
11351 }
11352
11353 /* Record that we found at least one task that could run on this_cpu */
11354 env.flags &= ~LBF_ALL_PINNED;
11355
11356 /*
11357 * ->active_balance synchronizes accesses to
11358 * ->active_balance_work. Once set, it's cleared
11359 * only after active load balance is finished.
11360 */
11361 if (!busiest->active_balance) {
11362 busiest->active_balance = 1;
11363 busiest->push_cpu = this_cpu;
11364 active_balance = 1;
11365 }
11366
11367 preempt_disable();
11368 raw_spin_rq_unlock_irqrestore(busiest, flags);
11369 if (active_balance) {
11370 stop_one_cpu_nowait(cpu_of(busiest),
11371 active_load_balance_cpu_stop, busiest,
11372 &busiest->active_balance_work);
11373 }
11374 preempt_enable();
11375 }
11376 } else {
11377 sd->nr_balance_failed = 0;
11378 }
11379
11380 if (likely(!active_balance) || need_active_balance(&env)) {
11381 /* We were unbalanced, so reset the balancing interval */
11382 sd->balance_interval = sd->min_interval;
11383 }
11384
11385 goto out;
11386
11387 out_balanced:
11388 /*
11389 * We reach balance although we may have faced some affinity
11390 * constraints. Clear the imbalance flag only if other tasks got
11391 * a chance to move and fix the imbalance.
11392 */
11393 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
11394 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11395
11396 if (*group_imbalance)
11397 *group_imbalance = 0;
11398 }
11399
11400 out_all_pinned:
11401 /*
11402 * We reach balance because all tasks are pinned at this level so
11403 * we can't migrate them. Let the imbalance flag set so parent level
11404 * can try to migrate them.
11405 */
11406 schedstat_inc(sd->lb_balanced[idle]);
11407
11408 sd->nr_balance_failed = 0;
11409
11410 out_one_pinned:
11411 ld_moved = 0;
11412
11413 /*
11414 * newidle_balance() disregards balance intervals, so we could
11415 * repeatedly reach this code, which would lead to balance_interval
11416 * skyrocketing in a short amount of time. Skip the balance_interval
11417 * increase logic to avoid that.
11418 */
11419 if (env.idle == CPU_NEWLY_IDLE)
11420 goto out;
11421
11422 /* tune up the balancing interval */
11423 if ((env.flags & LBF_ALL_PINNED &&
11424 sd->balance_interval < MAX_PINNED_INTERVAL) ||
11425 sd->balance_interval < sd->max_interval)
11426 sd->balance_interval *= 2;
11427 out:
11428 return ld_moved;
11429 }
11430
11431 static inline unsigned long
get_sd_balance_interval(struct sched_domain * sd,int cpu_busy)11432 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
11433 {
11434 unsigned long interval = sd->balance_interval;
11435
11436 if (cpu_busy)
11437 interval *= sd->busy_factor;
11438
11439 /* scale ms to jiffies */
11440 interval = msecs_to_jiffies(interval);
11441
11442 /*
11443 * Reduce likelihood of busy balancing at higher domains racing with
11444 * balancing at lower domains by preventing their balancing periods
11445 * from being multiples of each other.
11446 */
11447 if (cpu_busy)
11448 interval -= 1;
11449
11450 interval = clamp(interval, 1UL, max_load_balance_interval);
11451
11452 return interval;
11453 }
11454
11455 static inline void
update_next_balance(struct sched_domain * sd,unsigned long * next_balance)11456 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
11457 {
11458 unsigned long interval, next;
11459
11460 /* used by idle balance, so cpu_busy = 0 */
11461 interval = get_sd_balance_interval(sd, 0);
11462 next = sd->last_balance + interval;
11463
11464 if (time_after(*next_balance, next))
11465 *next_balance = next;
11466 }
11467
11468 /*
11469 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
11470 * running tasks off the busiest CPU onto idle CPUs. It requires at
11471 * least 1 task to be running on each physical CPU where possible, and
11472 * avoids physical / logical imbalances.
11473 */
active_load_balance_cpu_stop(void * data)11474 static int active_load_balance_cpu_stop(void *data)
11475 {
11476 struct rq *busiest_rq = data;
11477 int busiest_cpu = cpu_of(busiest_rq);
11478 int target_cpu = busiest_rq->push_cpu;
11479 struct rq *target_rq = cpu_rq(target_cpu);
11480 struct sched_domain *sd;
11481 struct task_struct *p = NULL;
11482 struct rq_flags rf;
11483
11484 rq_lock_irq(busiest_rq, &rf);
11485 /*
11486 * Between queueing the stop-work and running it is a hole in which
11487 * CPUs can become inactive. We should not move tasks from or to
11488 * inactive CPUs.
11489 */
11490 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
11491 goto out_unlock;
11492
11493 /* Make sure the requested CPU hasn't gone down in the meantime: */
11494 if (unlikely(busiest_cpu != smp_processor_id() ||
11495 !busiest_rq->active_balance))
11496 goto out_unlock;
11497
11498 /* Is there any task to move? */
11499 if (busiest_rq->nr_running <= 1)
11500 goto out_unlock;
11501
11502 /*
11503 * This condition is "impossible", if it occurs
11504 * we need to fix it. Originally reported by
11505 * Bjorn Helgaas on a 128-CPU setup.
11506 */
11507 WARN_ON_ONCE(busiest_rq == target_rq);
11508
11509 /* Search for an sd spanning us and the target CPU. */
11510 rcu_read_lock();
11511 for_each_domain(target_cpu, sd) {
11512 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
11513 break;
11514 }
11515
11516 if (likely(sd)) {
11517 struct lb_env env = {
11518 .sd = sd,
11519 .dst_cpu = target_cpu,
11520 .dst_rq = target_rq,
11521 .src_cpu = busiest_rq->cpu,
11522 .src_rq = busiest_rq,
11523 .idle = CPU_IDLE,
11524 .flags = LBF_ACTIVE_LB,
11525 };
11526
11527 schedstat_inc(sd->alb_count);
11528 update_rq_clock(busiest_rq);
11529
11530 p = detach_one_task(&env);
11531 if (p) {
11532 schedstat_inc(sd->alb_pushed);
11533 /* Active balancing done, reset the failure counter. */
11534 sd->nr_balance_failed = 0;
11535 } else {
11536 schedstat_inc(sd->alb_failed);
11537 }
11538 }
11539 rcu_read_unlock();
11540 out_unlock:
11541 busiest_rq->active_balance = 0;
11542 rq_unlock(busiest_rq, &rf);
11543
11544 if (p)
11545 attach_one_task(target_rq, p);
11546
11547 local_irq_enable();
11548
11549 return 0;
11550 }
11551
11552 static DEFINE_SPINLOCK(balancing);
11553
11554 /*
11555 * Scale the max load_balance interval with the number of CPUs in the system.
11556 * This trades load-balance latency on larger machines for less cross talk.
11557 */
update_max_interval(void)11558 void update_max_interval(void)
11559 {
11560 max_load_balance_interval = HZ*num_online_cpus()/10;
11561 }
11562
update_newidle_cost(struct sched_domain * sd,u64 cost)11563 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
11564 {
11565 if (cost > sd->max_newidle_lb_cost) {
11566 /*
11567 * Track max cost of a domain to make sure to not delay the
11568 * next wakeup on the CPU.
11569 */
11570 sd->max_newidle_lb_cost = cost;
11571 sd->last_decay_max_lb_cost = jiffies;
11572 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
11573 /*
11574 * Decay the newidle max times by ~1% per second to ensure that
11575 * it is not outdated and the current max cost is actually
11576 * shorter.
11577 */
11578 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
11579 sd->last_decay_max_lb_cost = jiffies;
11580
11581 return true;
11582 }
11583
11584 return false;
11585 }
11586
11587 /*
11588 * It checks each scheduling domain to see if it is due to be balanced,
11589 * and initiates a balancing operation if so.
11590 *
11591 * Balancing parameters are set up in init_sched_domains.
11592 */
rebalance_domains(struct rq * rq,enum cpu_idle_type idle)11593 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
11594 {
11595 int continue_balancing = 1;
11596 int cpu = rq->cpu;
11597 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11598 unsigned long interval;
11599 struct sched_domain *sd;
11600 /* Earliest time when we have to do rebalance again */
11601 unsigned long next_balance = jiffies + 60*HZ;
11602 int update_next_balance = 0;
11603 int need_serialize, need_decay = 0;
11604 u64 max_cost = 0;
11605
11606 rcu_read_lock();
11607 for_each_domain(cpu, sd) {
11608 /*
11609 * Decay the newidle max times here because this is a regular
11610 * visit to all the domains.
11611 */
11612 need_decay = update_newidle_cost(sd, 0);
11613 max_cost += sd->max_newidle_lb_cost;
11614
11615 /*
11616 * Stop the load balance at this level. There is another
11617 * CPU in our sched group which is doing load balancing more
11618 * actively.
11619 */
11620 if (!continue_balancing) {
11621 if (need_decay)
11622 continue;
11623 break;
11624 }
11625
11626 interval = get_sd_balance_interval(sd, busy);
11627
11628 need_serialize = sd->flags & SD_SERIALIZE;
11629 if (need_serialize) {
11630 if (!spin_trylock(&balancing))
11631 goto out;
11632 }
11633
11634 if (time_after_eq(jiffies, sd->last_balance + interval)) {
11635 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
11636 /*
11637 * The LBF_DST_PINNED logic could have changed
11638 * env->dst_cpu, so we can't know our idle
11639 * state even if we migrated tasks. Update it.
11640 */
11641 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
11642 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11643 }
11644 sd->last_balance = jiffies;
11645 interval = get_sd_balance_interval(sd, busy);
11646 }
11647 if (need_serialize)
11648 spin_unlock(&balancing);
11649 out:
11650 if (time_after(next_balance, sd->last_balance + interval)) {
11651 next_balance = sd->last_balance + interval;
11652 update_next_balance = 1;
11653 }
11654 }
11655 if (need_decay) {
11656 /*
11657 * Ensure the rq-wide value also decays but keep it at a
11658 * reasonable floor to avoid funnies with rq->avg_idle.
11659 */
11660 rq->max_idle_balance_cost =
11661 max((u64)sysctl_sched_migration_cost, max_cost);
11662 }
11663 rcu_read_unlock();
11664
11665 /*
11666 * next_balance will be updated only when there is a need.
11667 * When the cpu is attached to null domain for ex, it will not be
11668 * updated.
11669 */
11670 if (likely(update_next_balance))
11671 rq->next_balance = next_balance;
11672
11673 }
11674
on_null_domain(struct rq * rq)11675 static inline int on_null_domain(struct rq *rq)
11676 {
11677 return unlikely(!rcu_dereference_sched(rq->sd));
11678 }
11679
11680 #ifdef CONFIG_NO_HZ_COMMON
11681 /*
11682 * idle load balancing details
11683 * - When one of the busy CPUs notice that there may be an idle rebalancing
11684 * needed, they will kick the idle load balancer, which then does idle
11685 * load balancing for all the idle CPUs.
11686 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
11687 * anywhere yet.
11688 */
11689
find_new_ilb(void)11690 static inline int find_new_ilb(void)
11691 {
11692 int ilb;
11693 const struct cpumask *hk_mask;
11694
11695 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11696
11697 for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
11698
11699 if (ilb == smp_processor_id())
11700 continue;
11701
11702 if (idle_cpu(ilb))
11703 return ilb;
11704 }
11705
11706 return nr_cpu_ids;
11707 }
11708
11709 /*
11710 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
11711 * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11712 */
kick_ilb(unsigned int flags)11713 static void kick_ilb(unsigned int flags)
11714 {
11715 int ilb_cpu;
11716
11717 /*
11718 * Increase nohz.next_balance only when if full ilb is triggered but
11719 * not if we only update stats.
11720 */
11721 if (flags & NOHZ_BALANCE_KICK)
11722 nohz.next_balance = jiffies+1;
11723
11724 ilb_cpu = find_new_ilb();
11725
11726 if (ilb_cpu >= nr_cpu_ids)
11727 return;
11728
11729 /*
11730 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11731 * the first flag owns it; cleared by nohz_csd_func().
11732 */
11733 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11734 if (flags & NOHZ_KICK_MASK)
11735 return;
11736
11737 /*
11738 * This way we generate an IPI on the target CPU which
11739 * is idle. And the softirq performing nohz idle load balance
11740 * will be run before returning from the IPI.
11741 */
11742 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11743 }
11744
11745 /*
11746 * Current decision point for kicking the idle load balancer in the presence
11747 * of idle CPUs in the system.
11748 */
nohz_balancer_kick(struct rq * rq)11749 static void nohz_balancer_kick(struct rq *rq)
11750 {
11751 unsigned long now = jiffies;
11752 struct sched_domain_shared *sds;
11753 struct sched_domain *sd;
11754 int nr_busy, i, cpu = rq->cpu;
11755 unsigned int flags = 0;
11756
11757 if (unlikely(rq->idle_balance))
11758 return;
11759
11760 /*
11761 * We may be recently in ticked or tickless idle mode. At the first
11762 * busy tick after returning from idle, we will update the busy stats.
11763 */
11764 nohz_balance_exit_idle(rq);
11765
11766 /*
11767 * None are in tickless mode and hence no need for NOHZ idle load
11768 * balancing.
11769 */
11770 if (likely(!atomic_read(&nohz.nr_cpus)))
11771 return;
11772
11773 if (READ_ONCE(nohz.has_blocked) &&
11774 time_after(now, READ_ONCE(nohz.next_blocked)))
11775 flags = NOHZ_STATS_KICK;
11776
11777 if (time_before(now, nohz.next_balance))
11778 goto out;
11779
11780 if (rq->nr_running >= 2) {
11781 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11782 goto out;
11783 }
11784
11785 rcu_read_lock();
11786
11787 sd = rcu_dereference(rq->sd);
11788 if (sd) {
11789 /*
11790 * If there's a CFS task and the current CPU has reduced
11791 * capacity; kick the ILB to see if there's a better CPU to run
11792 * on.
11793 */
11794 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11795 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11796 goto unlock;
11797 }
11798 }
11799
11800 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11801 if (sd) {
11802 /*
11803 * When ASYM_PACKING; see if there's a more preferred CPU
11804 * currently idle; in which case, kick the ILB to move tasks
11805 * around.
11806 *
11807 * When balancing betwen cores, all the SMT siblings of the
11808 * preferred CPU must be idle.
11809 */
11810 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11811 if (sched_use_asym_prio(sd, i) &&
11812 sched_asym_prefer(i, cpu)) {
11813 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11814 goto unlock;
11815 }
11816 }
11817 }
11818
11819 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11820 if (sd) {
11821 /*
11822 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11823 * to run the misfit task on.
11824 */
11825 if (check_misfit_status(rq, sd)) {
11826 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11827 goto unlock;
11828 }
11829
11830 /*
11831 * For asymmetric systems, we do not want to nicely balance
11832 * cache use, instead we want to embrace asymmetry and only
11833 * ensure tasks have enough CPU capacity.
11834 *
11835 * Skip the LLC logic because it's not relevant in that case.
11836 */
11837 goto unlock;
11838 }
11839
11840 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
11841 if (sds) {
11842 /*
11843 * If there is an imbalance between LLC domains (IOW we could
11844 * increase the overall cache use), we need some less-loaded LLC
11845 * domain to pull some load. Likewise, we may need to spread
11846 * load within the current LLC domain (e.g. packed SMT cores but
11847 * other CPUs are idle). We can't really know from here how busy
11848 * the others are - so just get a nohz balance going if it looks
11849 * like this LLC domain has tasks we could move.
11850 */
11851 nr_busy = atomic_read(&sds->nr_busy_cpus);
11852 if (nr_busy > 1) {
11853 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11854 goto unlock;
11855 }
11856 }
11857 unlock:
11858 rcu_read_unlock();
11859 out:
11860 if (READ_ONCE(nohz.needs_update))
11861 flags |= NOHZ_NEXT_KICK;
11862
11863 if (flags)
11864 kick_ilb(flags);
11865 }
11866
set_cpu_sd_state_busy(int cpu)11867 static void set_cpu_sd_state_busy(int cpu)
11868 {
11869 struct sched_domain *sd;
11870
11871 rcu_read_lock();
11872 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11873
11874 if (!sd || !sd->nohz_idle)
11875 goto unlock;
11876 sd->nohz_idle = 0;
11877
11878 atomic_inc(&sd->shared->nr_busy_cpus);
11879 unlock:
11880 rcu_read_unlock();
11881 }
11882
nohz_balance_exit_idle(struct rq * rq)11883 void nohz_balance_exit_idle(struct rq *rq)
11884 {
11885 SCHED_WARN_ON(rq != this_rq());
11886
11887 if (likely(!rq->nohz_tick_stopped))
11888 return;
11889
11890 rq->nohz_tick_stopped = 0;
11891 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
11892 atomic_dec(&nohz.nr_cpus);
11893
11894 set_cpu_sd_state_busy(rq->cpu);
11895 }
11896
set_cpu_sd_state_idle(int cpu)11897 static void set_cpu_sd_state_idle(int cpu)
11898 {
11899 struct sched_domain *sd;
11900
11901 rcu_read_lock();
11902 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11903
11904 if (!sd || sd->nohz_idle)
11905 goto unlock;
11906 sd->nohz_idle = 1;
11907
11908 atomic_dec(&sd->shared->nr_busy_cpus);
11909 unlock:
11910 rcu_read_unlock();
11911 }
11912
11913 /*
11914 * This routine will record that the CPU is going idle with tick stopped.
11915 * This info will be used in performing idle load balancing in the future.
11916 */
nohz_balance_enter_idle(int cpu)11917 void nohz_balance_enter_idle(int cpu)
11918 {
11919 struct rq *rq = cpu_rq(cpu);
11920
11921 SCHED_WARN_ON(cpu != smp_processor_id());
11922
11923 /* If this CPU is going down, then nothing needs to be done: */
11924 if (!cpu_active(cpu))
11925 return;
11926
11927 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
11928 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
11929 return;
11930
11931 /*
11932 * Can be set safely without rq->lock held
11933 * If a clear happens, it will have evaluated last additions because
11934 * rq->lock is held during the check and the clear
11935 */
11936 rq->has_blocked_load = 1;
11937
11938 /*
11939 * The tick is still stopped but load could have been added in the
11940 * meantime. We set the nohz.has_blocked flag to trig a check of the
11941 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
11942 * of nohz.has_blocked can only happen after checking the new load
11943 */
11944 if (rq->nohz_tick_stopped)
11945 goto out;
11946
11947 /* If we're a completely isolated CPU, we don't play: */
11948 if (on_null_domain(rq))
11949 return;
11950
11951 rq->nohz_tick_stopped = 1;
11952
11953 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
11954 atomic_inc(&nohz.nr_cpus);
11955
11956 /*
11957 * Ensures that if nohz_idle_balance() fails to observe our
11958 * @idle_cpus_mask store, it must observe the @has_blocked
11959 * and @needs_update stores.
11960 */
11961 smp_mb__after_atomic();
11962
11963 set_cpu_sd_state_idle(cpu);
11964
11965 WRITE_ONCE(nohz.needs_update, 1);
11966 out:
11967 /*
11968 * Each time a cpu enter idle, we assume that it has blocked load and
11969 * enable the periodic update of the load of idle cpus
11970 */
11971 WRITE_ONCE(nohz.has_blocked, 1);
11972 }
11973
update_nohz_stats(struct rq * rq)11974 static bool update_nohz_stats(struct rq *rq)
11975 {
11976 unsigned int cpu = rq->cpu;
11977
11978 if (!rq->has_blocked_load)
11979 return false;
11980
11981 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
11982 return false;
11983
11984 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
11985 return true;
11986
11987 update_blocked_averages(cpu);
11988
11989 return rq->has_blocked_load;
11990 }
11991
11992 /*
11993 * Internal function that runs load balance for all idle cpus. The load balance
11994 * can be a simple update of blocked load or a complete load balance with
11995 * tasks movement depending of flags.
11996 */
_nohz_idle_balance(struct rq * this_rq,unsigned int flags)11997 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
11998 {
11999 /* Earliest time when we have to do rebalance again */
12000 unsigned long now = jiffies;
12001 unsigned long next_balance = now + 60*HZ;
12002 bool has_blocked_load = false;
12003 int update_next_balance = 0;
12004 int this_cpu = this_rq->cpu;
12005 int balance_cpu;
12006 struct rq *rq;
12007
12008 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
12009
12010 /*
12011 * We assume there will be no idle load after this update and clear
12012 * the has_blocked flag. If a cpu enters idle in the mean time, it will
12013 * set the has_blocked flag and trigger another update of idle load.
12014 * Because a cpu that becomes idle, is added to idle_cpus_mask before
12015 * setting the flag, we are sure to not clear the state and not
12016 * check the load of an idle cpu.
12017 *
12018 * Same applies to idle_cpus_mask vs needs_update.
12019 */
12020 if (flags & NOHZ_STATS_KICK)
12021 WRITE_ONCE(nohz.has_blocked, 0);
12022 if (flags & NOHZ_NEXT_KICK)
12023 WRITE_ONCE(nohz.needs_update, 0);
12024
12025 /*
12026 * Ensures that if we miss the CPU, we must see the has_blocked
12027 * store from nohz_balance_enter_idle().
12028 */
12029 smp_mb();
12030
12031 /*
12032 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
12033 * chance for other idle cpu to pull load.
12034 */
12035 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
12036 if (!idle_cpu(balance_cpu))
12037 continue;
12038
12039 /*
12040 * If this CPU gets work to do, stop the load balancing
12041 * work being done for other CPUs. Next load
12042 * balancing owner will pick it up.
12043 */
12044 if (need_resched()) {
12045 if (flags & NOHZ_STATS_KICK)
12046 has_blocked_load = true;
12047 if (flags & NOHZ_NEXT_KICK)
12048 WRITE_ONCE(nohz.needs_update, 1);
12049 goto abort;
12050 }
12051
12052 rq = cpu_rq(balance_cpu);
12053
12054 if (flags & NOHZ_STATS_KICK)
12055 has_blocked_load |= update_nohz_stats(rq);
12056
12057 /*
12058 * If time for next balance is due,
12059 * do the balance.
12060 */
12061 if (time_after_eq(jiffies, rq->next_balance)) {
12062 struct rq_flags rf;
12063
12064 rq_lock_irqsave(rq, &rf);
12065 update_rq_clock(rq);
12066 rq_unlock_irqrestore(rq, &rf);
12067
12068 if (flags & NOHZ_BALANCE_KICK)
12069 rebalance_domains(rq, CPU_IDLE);
12070 }
12071
12072 if (time_after(next_balance, rq->next_balance)) {
12073 next_balance = rq->next_balance;
12074 update_next_balance = 1;
12075 }
12076 }
12077
12078 /*
12079 * next_balance will be updated only when there is a need.
12080 * When the CPU is attached to null domain for ex, it will not be
12081 * updated.
12082 */
12083 if (likely(update_next_balance))
12084 nohz.next_balance = next_balance;
12085
12086 if (flags & NOHZ_STATS_KICK)
12087 WRITE_ONCE(nohz.next_blocked,
12088 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
12089
12090 abort:
12091 /* There is still blocked load, enable periodic update */
12092 if (has_blocked_load)
12093 WRITE_ONCE(nohz.has_blocked, 1);
12094 }
12095
12096 /*
12097 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
12098 * rebalancing for all the cpus for whom scheduler ticks are stopped.
12099 */
nohz_idle_balance(struct rq * this_rq,enum cpu_idle_type idle)12100 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12101 {
12102 unsigned int flags = this_rq->nohz_idle_balance;
12103
12104 if (!flags)
12105 return false;
12106
12107 this_rq->nohz_idle_balance = 0;
12108
12109 if (idle != CPU_IDLE)
12110 return false;
12111
12112 _nohz_idle_balance(this_rq, flags);
12113
12114 return true;
12115 }
12116
12117 /*
12118 * Check if we need to run the ILB for updating blocked load before entering
12119 * idle state.
12120 */
nohz_run_idle_balance(int cpu)12121 void nohz_run_idle_balance(int cpu)
12122 {
12123 unsigned int flags;
12124
12125 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
12126
12127 /*
12128 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
12129 * (ie NOHZ_STATS_KICK set) and will do the same.
12130 */
12131 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
12132 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
12133 }
12134
nohz_newidle_balance(struct rq * this_rq)12135 static void nohz_newidle_balance(struct rq *this_rq)
12136 {
12137 int this_cpu = this_rq->cpu;
12138
12139 /*
12140 * This CPU doesn't want to be disturbed by scheduler
12141 * housekeeping
12142 */
12143 if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
12144 return;
12145
12146 /* Will wake up very soon. No time for doing anything else*/
12147 if (this_rq->avg_idle < sysctl_sched_migration_cost)
12148 return;
12149
12150 /* Don't need to update blocked load of idle CPUs*/
12151 if (!READ_ONCE(nohz.has_blocked) ||
12152 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
12153 return;
12154
12155 /*
12156 * Set the need to trigger ILB in order to update blocked load
12157 * before entering idle state.
12158 */
12159 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
12160 }
12161
12162 #else /* !CONFIG_NO_HZ_COMMON */
nohz_balancer_kick(struct rq * rq)12163 static inline void nohz_balancer_kick(struct rq *rq) { }
12164
nohz_idle_balance(struct rq * this_rq,enum cpu_idle_type idle)12165 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12166 {
12167 return false;
12168 }
12169
nohz_newidle_balance(struct rq * this_rq)12170 static inline void nohz_newidle_balance(struct rq *this_rq) { }
12171 #endif /* CONFIG_NO_HZ_COMMON */
12172
12173 /*
12174 * newidle_balance is called by schedule() if this_cpu is about to become
12175 * idle. Attempts to pull tasks from other CPUs.
12176 *
12177 * Returns:
12178 * < 0 - we released the lock and there are !fair tasks present
12179 * 0 - failed, no new tasks
12180 * > 0 - success, new (fair) tasks present
12181 */
newidle_balance(struct rq * this_rq,struct rq_flags * rf)12182 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
12183 {
12184 unsigned long next_balance = jiffies + HZ;
12185 int this_cpu = this_rq->cpu;
12186 u64 t0, t1, curr_cost = 0;
12187 struct sched_domain *sd;
12188 int pulled_task = 0;
12189
12190 update_misfit_status(NULL, this_rq);
12191
12192 /*
12193 * There is a task waiting to run. No need to search for one.
12194 * Return 0; the task will be enqueued when switching to idle.
12195 */
12196 if (this_rq->ttwu_pending)
12197 return 0;
12198
12199 /*
12200 * We must set idle_stamp _before_ calling idle_balance(), such that we
12201 * measure the duration of idle_balance() as idle time.
12202 */
12203 this_rq->idle_stamp = rq_clock(this_rq);
12204
12205 /*
12206 * Do not pull tasks towards !active CPUs...
12207 */
12208 if (!cpu_active(this_cpu))
12209 return 0;
12210
12211 /*
12212 * This is OK, because current is on_cpu, which avoids it being picked
12213 * for load-balance and preemption/IRQs are still disabled avoiding
12214 * further scheduler activity on it and we're being very careful to
12215 * re-start the picking loop.
12216 */
12217 rq_unpin_lock(this_rq, rf);
12218
12219 rcu_read_lock();
12220 sd = rcu_dereference_check_sched_domain(this_rq->sd);
12221
12222 if (!READ_ONCE(this_rq->rd->overload) ||
12223 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
12224
12225 if (sd)
12226 update_next_balance(sd, &next_balance);
12227 rcu_read_unlock();
12228
12229 goto out;
12230 }
12231 rcu_read_unlock();
12232
12233 raw_spin_rq_unlock(this_rq);
12234
12235 t0 = sched_clock_cpu(this_cpu);
12236 update_blocked_averages(this_cpu);
12237
12238 rcu_read_lock();
12239 for_each_domain(this_cpu, sd) {
12240 int continue_balancing = 1;
12241 u64 domain_cost;
12242
12243 update_next_balance(sd, &next_balance);
12244
12245 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
12246 break;
12247
12248 if (sd->flags & SD_BALANCE_NEWIDLE) {
12249
12250 pulled_task = load_balance(this_cpu, this_rq,
12251 sd, CPU_NEWLY_IDLE,
12252 &continue_balancing);
12253
12254 t1 = sched_clock_cpu(this_cpu);
12255 domain_cost = t1 - t0;
12256 update_newidle_cost(sd, domain_cost);
12257
12258 curr_cost += domain_cost;
12259 t0 = t1;
12260 }
12261
12262 /*
12263 * Stop searching for tasks to pull if there are
12264 * now runnable tasks on this rq.
12265 */
12266 if (pulled_task || this_rq->nr_running > 0 ||
12267 this_rq->ttwu_pending)
12268 break;
12269 }
12270 rcu_read_unlock();
12271
12272 raw_spin_rq_lock(this_rq);
12273
12274 if (curr_cost > this_rq->max_idle_balance_cost)
12275 this_rq->max_idle_balance_cost = curr_cost;
12276
12277 /*
12278 * While browsing the domains, we released the rq lock, a task could
12279 * have been enqueued in the meantime. Since we're not going idle,
12280 * pretend we pulled a task.
12281 */
12282 if (this_rq->cfs.h_nr_running && !pulled_task)
12283 pulled_task = 1;
12284
12285 /* Is there a task of a high priority class? */
12286 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
12287 pulled_task = -1;
12288
12289 out:
12290 /* Move the next balance forward */
12291 if (time_after(this_rq->next_balance, next_balance))
12292 this_rq->next_balance = next_balance;
12293
12294 if (pulled_task)
12295 this_rq->idle_stamp = 0;
12296 else
12297 nohz_newidle_balance(this_rq);
12298
12299 rq_repin_lock(this_rq, rf);
12300
12301 return pulled_task;
12302 }
12303
12304 /*
12305 * run_rebalance_domains is triggered when needed from the scheduler tick.
12306 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
12307 */
run_rebalance_domains(struct softirq_action * h)12308 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
12309 {
12310 struct rq *this_rq = this_rq();
12311 enum cpu_idle_type idle = this_rq->idle_balance ?
12312 CPU_IDLE : CPU_NOT_IDLE;
12313
12314 /*
12315 * If this CPU has a pending nohz_balance_kick, then do the
12316 * balancing on behalf of the other idle CPUs whose ticks are
12317 * stopped. Do nohz_idle_balance *before* rebalance_domains to
12318 * give the idle CPUs a chance to load balance. Else we may
12319 * load balance only within the local sched_domain hierarchy
12320 * and abort nohz_idle_balance altogether if we pull some load.
12321 */
12322 if (nohz_idle_balance(this_rq, idle))
12323 return;
12324
12325 /* normal load balance */
12326 update_blocked_averages(this_rq->cpu);
12327 rebalance_domains(this_rq, idle);
12328 }
12329
12330 /*
12331 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
12332 */
trigger_load_balance(struct rq * rq)12333 void trigger_load_balance(struct rq *rq)
12334 {
12335 /*
12336 * Don't need to rebalance while attached to NULL domain or
12337 * runqueue CPU is not active
12338 */
12339 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
12340 return;
12341
12342 if (time_after_eq(jiffies, rq->next_balance))
12343 raise_softirq(SCHED_SOFTIRQ);
12344
12345 nohz_balancer_kick(rq);
12346 }
12347
rq_online_fair(struct rq * rq)12348 static void rq_online_fair(struct rq *rq)
12349 {
12350 update_sysctl();
12351
12352 update_runtime_enabled(rq);
12353 }
12354
rq_offline_fair(struct rq * rq)12355 static void rq_offline_fair(struct rq *rq)
12356 {
12357 update_sysctl();
12358
12359 /* Ensure any throttled groups are reachable by pick_next_task */
12360 unthrottle_offline_cfs_rqs(rq);
12361 }
12362
12363 #endif /* CONFIG_SMP */
12364
12365 #ifdef CONFIG_SCHED_CORE
12366 static inline bool
__entity_slice_used(struct sched_entity * se,int min_nr_tasks)12367 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
12368 {
12369 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
12370 u64 slice = se->slice;
12371
12372 return (rtime * min_nr_tasks > slice);
12373 }
12374
12375 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
task_tick_core(struct rq * rq,struct task_struct * curr)12376 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
12377 {
12378 if (!sched_core_enabled(rq))
12379 return;
12380
12381 /*
12382 * If runqueue has only one task which used up its slice and
12383 * if the sibling is forced idle, then trigger schedule to
12384 * give forced idle task a chance.
12385 *
12386 * sched_slice() considers only this active rq and it gets the
12387 * whole slice. But during force idle, we have siblings acting
12388 * like a single runqueue and hence we need to consider runnable
12389 * tasks on this CPU and the forced idle CPU. Ideally, we should
12390 * go through the forced idle rq, but that would be a perf hit.
12391 * We can assume that the forced idle CPU has at least
12392 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
12393 * if we need to give up the CPU.
12394 */
12395 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
12396 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
12397 resched_curr(rq);
12398 }
12399
12400 /*
12401 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
12402 */
se_fi_update(const struct sched_entity * se,unsigned int fi_seq,bool forceidle)12403 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
12404 bool forceidle)
12405 {
12406 for_each_sched_entity(se) {
12407 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12408
12409 if (forceidle) {
12410 if (cfs_rq->forceidle_seq == fi_seq)
12411 break;
12412 cfs_rq->forceidle_seq = fi_seq;
12413 }
12414
12415 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
12416 }
12417 }
12418
task_vruntime_update(struct rq * rq,struct task_struct * p,bool in_fi)12419 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
12420 {
12421 struct sched_entity *se = &p->se;
12422
12423 if (p->sched_class != &fair_sched_class)
12424 return;
12425
12426 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
12427 }
12428
cfs_prio_less(const struct task_struct * a,const struct task_struct * b,bool in_fi)12429 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
12430 bool in_fi)
12431 {
12432 struct rq *rq = task_rq(a);
12433 const struct sched_entity *sea = &a->se;
12434 const struct sched_entity *seb = &b->se;
12435 struct cfs_rq *cfs_rqa;
12436 struct cfs_rq *cfs_rqb;
12437 s64 delta;
12438
12439 SCHED_WARN_ON(task_rq(b)->core != rq->core);
12440
12441 #ifdef CONFIG_FAIR_GROUP_SCHED
12442 /*
12443 * Find an se in the hierarchy for tasks a and b, such that the se's
12444 * are immediate siblings.
12445 */
12446 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
12447 int sea_depth = sea->depth;
12448 int seb_depth = seb->depth;
12449
12450 if (sea_depth >= seb_depth)
12451 sea = parent_entity(sea);
12452 if (sea_depth <= seb_depth)
12453 seb = parent_entity(seb);
12454 }
12455
12456 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
12457 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
12458
12459 cfs_rqa = sea->cfs_rq;
12460 cfs_rqb = seb->cfs_rq;
12461 #else
12462 cfs_rqa = &task_rq(a)->cfs;
12463 cfs_rqb = &task_rq(b)->cfs;
12464 #endif
12465
12466 /*
12467 * Find delta after normalizing se's vruntime with its cfs_rq's
12468 * min_vruntime_fi, which would have been updated in prior calls
12469 * to se_fi_update().
12470 */
12471 delta = (s64)(sea->vruntime - seb->vruntime) +
12472 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
12473
12474 return delta > 0;
12475 }
12476
task_is_throttled_fair(struct task_struct * p,int cpu)12477 static int task_is_throttled_fair(struct task_struct *p, int cpu)
12478 {
12479 struct cfs_rq *cfs_rq;
12480
12481 #ifdef CONFIG_FAIR_GROUP_SCHED
12482 cfs_rq = task_group(p)->cfs_rq[cpu];
12483 #else
12484 cfs_rq = &cpu_rq(cpu)->cfs;
12485 #endif
12486 return throttled_hierarchy(cfs_rq);
12487 }
12488 #else
task_tick_core(struct rq * rq,struct task_struct * curr)12489 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
12490 #endif
12491
12492 /*
12493 * scheduler tick hitting a task of our scheduling class.
12494 *
12495 * NOTE: This function can be called remotely by the tick offload that
12496 * goes along full dynticks. Therefore no local assumption can be made
12497 * and everything must be accessed through the @rq and @curr passed in
12498 * parameters.
12499 */
task_tick_fair(struct rq * rq,struct task_struct * curr,int queued)12500 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
12501 {
12502 struct cfs_rq *cfs_rq;
12503 struct sched_entity *se = &curr->se;
12504
12505 for_each_sched_entity(se) {
12506 cfs_rq = cfs_rq_of(se);
12507 entity_tick(cfs_rq, se, queued);
12508 }
12509
12510 if (static_branch_unlikely(&sched_numa_balancing))
12511 task_tick_numa(rq, curr);
12512
12513 update_misfit_status(curr, rq);
12514 check_update_overutilized_status(task_rq(curr));
12515
12516 task_tick_core(rq, curr);
12517 }
12518
12519 /*
12520 * called on fork with the child task as argument from the parent's context
12521 * - child not yet on the tasklist
12522 * - preemption disabled
12523 */
task_fork_fair(struct task_struct * p)12524 static void task_fork_fair(struct task_struct *p)
12525 {
12526 struct sched_entity *se = &p->se, *curr;
12527 struct cfs_rq *cfs_rq;
12528 struct rq *rq = this_rq();
12529 struct rq_flags rf;
12530
12531 rq_lock(rq, &rf);
12532 update_rq_clock(rq);
12533
12534 cfs_rq = task_cfs_rq(current);
12535 curr = cfs_rq->curr;
12536 if (curr)
12537 update_curr(cfs_rq);
12538 place_entity(cfs_rq, se, ENQUEUE_INITIAL);
12539 rq_unlock(rq, &rf);
12540 }
12541
12542 /*
12543 * Priority of the task has changed. Check to see if we preempt
12544 * the current task.
12545 */
12546 static void
prio_changed_fair(struct rq * rq,struct task_struct * p,int oldprio)12547 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
12548 {
12549 if (!task_on_rq_queued(p))
12550 return;
12551
12552 if (rq->cfs.nr_running == 1)
12553 return;
12554
12555 /*
12556 * Reschedule if we are currently running on this runqueue and
12557 * our priority decreased, or if we are not currently running on
12558 * this runqueue and our priority is higher than the current's
12559 */
12560 if (task_current(rq, p)) {
12561 if (p->prio > oldprio)
12562 resched_curr(rq);
12563 } else
12564 check_preempt_curr(rq, p, 0);
12565 }
12566
12567 #ifdef CONFIG_FAIR_GROUP_SCHED
12568 /*
12569 * Propagate the changes of the sched_entity across the tg tree to make it
12570 * visible to the root
12571 */
propagate_entity_cfs_rq(struct sched_entity * se)12572 static void propagate_entity_cfs_rq(struct sched_entity *se)
12573 {
12574 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12575
12576 if (cfs_rq_throttled(cfs_rq))
12577 return;
12578
12579 if (!throttled_hierarchy(cfs_rq))
12580 list_add_leaf_cfs_rq(cfs_rq);
12581
12582 /* Start to propagate at parent */
12583 se = se->parent;
12584
12585 for_each_sched_entity(se) {
12586 cfs_rq = cfs_rq_of(se);
12587
12588 update_load_avg(cfs_rq, se, UPDATE_TG);
12589
12590 if (cfs_rq_throttled(cfs_rq))
12591 break;
12592
12593 if (!throttled_hierarchy(cfs_rq))
12594 list_add_leaf_cfs_rq(cfs_rq);
12595 }
12596 }
12597 #else
propagate_entity_cfs_rq(struct sched_entity * se)12598 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
12599 #endif
12600
detach_entity_cfs_rq(struct sched_entity * se)12601 static void detach_entity_cfs_rq(struct sched_entity *se)
12602 {
12603 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12604
12605 #ifdef CONFIG_SMP
12606 /*
12607 * In case the task sched_avg hasn't been attached:
12608 * - A forked task which hasn't been woken up by wake_up_new_task().
12609 * - A task which has been woken up by try_to_wake_up() but is
12610 * waiting for actually being woken up by sched_ttwu_pending().
12611 */
12612 if (!se->avg.last_update_time)
12613 return;
12614 #endif
12615
12616 /* Catch up with the cfs_rq and remove our load when we leave */
12617 update_load_avg(cfs_rq, se, 0);
12618 detach_entity_load_avg(cfs_rq, se);
12619 update_tg_load_avg(cfs_rq);
12620 propagate_entity_cfs_rq(se);
12621 }
12622
attach_entity_cfs_rq(struct sched_entity * se)12623 static void attach_entity_cfs_rq(struct sched_entity *se)
12624 {
12625 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12626
12627 /* Synchronize entity with its cfs_rq */
12628 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
12629 attach_entity_load_avg(cfs_rq, se);
12630 update_tg_load_avg(cfs_rq);
12631 propagate_entity_cfs_rq(se);
12632 }
12633
detach_task_cfs_rq(struct task_struct * p)12634 static void detach_task_cfs_rq(struct task_struct *p)
12635 {
12636 struct sched_entity *se = &p->se;
12637
12638 detach_entity_cfs_rq(se);
12639 }
12640
attach_task_cfs_rq(struct task_struct * p)12641 static void attach_task_cfs_rq(struct task_struct *p)
12642 {
12643 struct sched_entity *se = &p->se;
12644
12645 attach_entity_cfs_rq(se);
12646 }
12647
switched_from_fair(struct rq * rq,struct task_struct * p)12648 static void switched_from_fair(struct rq *rq, struct task_struct *p)
12649 {
12650 detach_task_cfs_rq(p);
12651 }
12652
switched_to_fair(struct rq * rq,struct task_struct * p)12653 static void switched_to_fair(struct rq *rq, struct task_struct *p)
12654 {
12655 attach_task_cfs_rq(p);
12656
12657 if (task_on_rq_queued(p)) {
12658 /*
12659 * We were most likely switched from sched_rt, so
12660 * kick off the schedule if running, otherwise just see
12661 * if we can still preempt the current task.
12662 */
12663 if (task_current(rq, p))
12664 resched_curr(rq);
12665 else
12666 check_preempt_curr(rq, p, 0);
12667 }
12668 }
12669
12670 /* Account for a task changing its policy or group.
12671 *
12672 * This routine is mostly called to set cfs_rq->curr field when a task
12673 * migrates between groups/classes.
12674 */
set_next_task_fair(struct rq * rq,struct task_struct * p,bool first)12675 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12676 {
12677 struct sched_entity *se = &p->se;
12678
12679 #ifdef CONFIG_SMP
12680 if (task_on_rq_queued(p)) {
12681 /*
12682 * Move the next running task to the front of the list, so our
12683 * cfs_tasks list becomes MRU one.
12684 */
12685 list_move(&se->group_node, &rq->cfs_tasks);
12686 }
12687 #endif
12688
12689 for_each_sched_entity(se) {
12690 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12691
12692 set_next_entity(cfs_rq, se);
12693 /* ensure bandwidth has been allocated on our new cfs_rq */
12694 account_cfs_rq_runtime(cfs_rq, 0);
12695 }
12696 }
12697
init_cfs_rq(struct cfs_rq * cfs_rq)12698 void init_cfs_rq(struct cfs_rq *cfs_rq)
12699 {
12700 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12701 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12702 #ifdef CONFIG_SMP
12703 raw_spin_lock_init(&cfs_rq->removed.lock);
12704 #endif
12705 }
12706
12707 #ifdef CONFIG_FAIR_GROUP_SCHED
task_change_group_fair(struct task_struct * p)12708 static void task_change_group_fair(struct task_struct *p)
12709 {
12710 /*
12711 * We couldn't detach or attach a forked task which
12712 * hasn't been woken up by wake_up_new_task().
12713 */
12714 if (READ_ONCE(p->__state) == TASK_NEW)
12715 return;
12716
12717 detach_task_cfs_rq(p);
12718
12719 #ifdef CONFIG_SMP
12720 /* Tell se's cfs_rq has been changed -- migrated */
12721 p->se.avg.last_update_time = 0;
12722 #endif
12723 set_task_rq(p, task_cpu(p));
12724 attach_task_cfs_rq(p);
12725 }
12726
free_fair_sched_group(struct task_group * tg)12727 void free_fair_sched_group(struct task_group *tg)
12728 {
12729 int i;
12730
12731 for_each_possible_cpu(i) {
12732 if (tg->cfs_rq)
12733 kfree(tg->cfs_rq[i]);
12734 if (tg->se)
12735 kfree(tg->se[i]);
12736 }
12737
12738 kfree(tg->cfs_rq);
12739 kfree(tg->se);
12740 }
12741
alloc_fair_sched_group(struct task_group * tg,struct task_group * parent)12742 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12743 {
12744 struct sched_entity *se;
12745 struct cfs_rq *cfs_rq;
12746 int i;
12747
12748 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12749 if (!tg->cfs_rq)
12750 goto err;
12751 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12752 if (!tg->se)
12753 goto err;
12754
12755 tg->shares = NICE_0_LOAD;
12756
12757 init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent));
12758
12759 for_each_possible_cpu(i) {
12760 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12761 GFP_KERNEL, cpu_to_node(i));
12762 if (!cfs_rq)
12763 goto err;
12764
12765 se = kzalloc_node(sizeof(struct sched_entity_stats),
12766 GFP_KERNEL, cpu_to_node(i));
12767 if (!se)
12768 goto err_free_rq;
12769
12770 init_cfs_rq(cfs_rq);
12771 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12772 init_entity_runnable_average(se);
12773 }
12774
12775 return 1;
12776
12777 err_free_rq:
12778 kfree(cfs_rq);
12779 err:
12780 return 0;
12781 }
12782
online_fair_sched_group(struct task_group * tg)12783 void online_fair_sched_group(struct task_group *tg)
12784 {
12785 struct sched_entity *se;
12786 struct rq_flags rf;
12787 struct rq *rq;
12788 int i;
12789
12790 for_each_possible_cpu(i) {
12791 rq = cpu_rq(i);
12792 se = tg->se[i];
12793 rq_lock_irq(rq, &rf);
12794 update_rq_clock(rq);
12795 attach_entity_cfs_rq(se);
12796 sync_throttle(tg, i);
12797 rq_unlock_irq(rq, &rf);
12798 }
12799 }
12800
unregister_fair_sched_group(struct task_group * tg)12801 void unregister_fair_sched_group(struct task_group *tg)
12802 {
12803 unsigned long flags;
12804 struct rq *rq;
12805 int cpu;
12806
12807 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12808
12809 for_each_possible_cpu(cpu) {
12810 if (tg->se[cpu])
12811 remove_entity_load_avg(tg->se[cpu]);
12812
12813 /*
12814 * Only empty task groups can be destroyed; so we can speculatively
12815 * check on_list without danger of it being re-added.
12816 */
12817 if (!tg->cfs_rq[cpu]->on_list)
12818 continue;
12819
12820 rq = cpu_rq(cpu);
12821
12822 raw_spin_rq_lock_irqsave(rq, flags);
12823 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
12824 raw_spin_rq_unlock_irqrestore(rq, flags);
12825 }
12826 }
12827
init_tg_cfs_entry(struct task_group * tg,struct cfs_rq * cfs_rq,struct sched_entity * se,int cpu,struct sched_entity * parent)12828 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
12829 struct sched_entity *se, int cpu,
12830 struct sched_entity *parent)
12831 {
12832 struct rq *rq = cpu_rq(cpu);
12833
12834 cfs_rq->tg = tg;
12835 cfs_rq->rq = rq;
12836 init_cfs_rq_runtime(cfs_rq);
12837
12838 tg->cfs_rq[cpu] = cfs_rq;
12839 tg->se[cpu] = se;
12840
12841 /* se could be NULL for root_task_group */
12842 if (!se)
12843 return;
12844
12845 if (!parent) {
12846 se->cfs_rq = &rq->cfs;
12847 se->depth = 0;
12848 } else {
12849 se->cfs_rq = parent->my_q;
12850 se->depth = parent->depth + 1;
12851 }
12852
12853 se->my_q = cfs_rq;
12854 /* guarantee group entities always have weight */
12855 update_load_set(&se->load, NICE_0_LOAD);
12856 se->parent = parent;
12857 }
12858
12859 static DEFINE_MUTEX(shares_mutex);
12860
__sched_group_set_shares(struct task_group * tg,unsigned long shares)12861 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
12862 {
12863 int i;
12864
12865 lockdep_assert_held(&shares_mutex);
12866
12867 /*
12868 * We can't change the weight of the root cgroup.
12869 */
12870 if (!tg->se[0])
12871 return -EINVAL;
12872
12873 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12874
12875 if (tg->shares == shares)
12876 return 0;
12877
12878 tg->shares = shares;
12879 for_each_possible_cpu(i) {
12880 struct rq *rq = cpu_rq(i);
12881 struct sched_entity *se = tg->se[i];
12882 struct rq_flags rf;
12883
12884 /* Propagate contribution to hierarchy */
12885 rq_lock_irqsave(rq, &rf);
12886 update_rq_clock(rq);
12887 for_each_sched_entity(se) {
12888 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
12889 update_cfs_group(se);
12890 }
12891 rq_unlock_irqrestore(rq, &rf);
12892 }
12893
12894 return 0;
12895 }
12896
sched_group_set_shares(struct task_group * tg,unsigned long shares)12897 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
12898 {
12899 int ret;
12900
12901 mutex_lock(&shares_mutex);
12902 if (tg_is_idle(tg))
12903 ret = -EINVAL;
12904 else
12905 ret = __sched_group_set_shares(tg, shares);
12906 mutex_unlock(&shares_mutex);
12907
12908 return ret;
12909 }
12910
sched_group_set_idle(struct task_group * tg,long idle)12911 int sched_group_set_idle(struct task_group *tg, long idle)
12912 {
12913 int i;
12914
12915 if (tg == &root_task_group)
12916 return -EINVAL;
12917
12918 if (idle < 0 || idle > 1)
12919 return -EINVAL;
12920
12921 mutex_lock(&shares_mutex);
12922
12923 if (tg->idle == idle) {
12924 mutex_unlock(&shares_mutex);
12925 return 0;
12926 }
12927
12928 tg->idle = idle;
12929
12930 for_each_possible_cpu(i) {
12931 struct rq *rq = cpu_rq(i);
12932 struct sched_entity *se = tg->se[i];
12933 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
12934 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
12935 long idle_task_delta;
12936 struct rq_flags rf;
12937
12938 rq_lock_irqsave(rq, &rf);
12939
12940 grp_cfs_rq->idle = idle;
12941 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
12942 goto next_cpu;
12943
12944 if (se->on_rq) {
12945 parent_cfs_rq = cfs_rq_of(se);
12946 if (cfs_rq_is_idle(grp_cfs_rq))
12947 parent_cfs_rq->idle_nr_running++;
12948 else
12949 parent_cfs_rq->idle_nr_running--;
12950 }
12951
12952 idle_task_delta = grp_cfs_rq->h_nr_running -
12953 grp_cfs_rq->idle_h_nr_running;
12954 if (!cfs_rq_is_idle(grp_cfs_rq))
12955 idle_task_delta *= -1;
12956
12957 for_each_sched_entity(se) {
12958 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12959
12960 if (!se->on_rq)
12961 break;
12962
12963 cfs_rq->idle_h_nr_running += idle_task_delta;
12964
12965 /* Already accounted at parent level and above. */
12966 if (cfs_rq_is_idle(cfs_rq))
12967 break;
12968 }
12969
12970 next_cpu:
12971 rq_unlock_irqrestore(rq, &rf);
12972 }
12973
12974 /* Idle groups have minimum weight. */
12975 if (tg_is_idle(tg))
12976 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
12977 else
12978 __sched_group_set_shares(tg, NICE_0_LOAD);
12979
12980 mutex_unlock(&shares_mutex);
12981 return 0;
12982 }
12983
12984 #else /* CONFIG_FAIR_GROUP_SCHED */
12985
free_fair_sched_group(struct task_group * tg)12986 void free_fair_sched_group(struct task_group *tg) { }
12987
alloc_fair_sched_group(struct task_group * tg,struct task_group * parent)12988 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12989 {
12990 return 1;
12991 }
12992
online_fair_sched_group(struct task_group * tg)12993 void online_fair_sched_group(struct task_group *tg) { }
12994
unregister_fair_sched_group(struct task_group * tg)12995 void unregister_fair_sched_group(struct task_group *tg) { }
12996
12997 #endif /* CONFIG_FAIR_GROUP_SCHED */
12998
12999
get_rr_interval_fair(struct rq * rq,struct task_struct * task)13000 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
13001 {
13002 struct sched_entity *se = &task->se;
13003 unsigned int rr_interval = 0;
13004
13005 /*
13006 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
13007 * idle runqueue:
13008 */
13009 if (rq->cfs.load.weight)
13010 rr_interval = NS_TO_JIFFIES(se->slice);
13011
13012 return rr_interval;
13013 }
13014
13015 /*
13016 * All the scheduling class methods:
13017 */
13018 DEFINE_SCHED_CLASS(fair) = {
13019
13020 .enqueue_task = enqueue_task_fair,
13021 .dequeue_task = dequeue_task_fair,
13022 .yield_task = yield_task_fair,
13023 .yield_to_task = yield_to_task_fair,
13024
13025 .check_preempt_curr = check_preempt_wakeup,
13026
13027 .pick_next_task = __pick_next_task_fair,
13028 .put_prev_task = put_prev_task_fair,
13029 .set_next_task = set_next_task_fair,
13030
13031 #ifdef CONFIG_SMP
13032 .balance = balance_fair,
13033 .pick_task = pick_task_fair,
13034 .select_task_rq = select_task_rq_fair,
13035 .migrate_task_rq = migrate_task_rq_fair,
13036
13037 .rq_online = rq_online_fair,
13038 .rq_offline = rq_offline_fair,
13039
13040 .task_dead = task_dead_fair,
13041 .set_cpus_allowed = set_cpus_allowed_common,
13042 #endif
13043
13044 .task_tick = task_tick_fair,
13045 .task_fork = task_fork_fair,
13046
13047 .prio_changed = prio_changed_fair,
13048 .switched_from = switched_from_fair,
13049 .switched_to = switched_to_fair,
13050
13051 .get_rr_interval = get_rr_interval_fair,
13052
13053 .update_curr = update_curr_fair,
13054
13055 #ifdef CONFIG_FAIR_GROUP_SCHED
13056 .task_change_group = task_change_group_fair,
13057 #endif
13058
13059 #ifdef CONFIG_SCHED_CORE
13060 .task_is_throttled = task_is_throttled_fair,
13061 #endif
13062
13063 #ifdef CONFIG_UCLAMP_TASK
13064 .uclamp_enabled = 1,
13065 #endif
13066 };
13067
13068 #ifdef CONFIG_SCHED_DEBUG
print_cfs_stats(struct seq_file * m,int cpu)13069 void print_cfs_stats(struct seq_file *m, int cpu)
13070 {
13071 struct cfs_rq *cfs_rq, *pos;
13072
13073 rcu_read_lock();
13074 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
13075 print_cfs_rq(m, cpu, cfs_rq);
13076 rcu_read_unlock();
13077 }
13078
13079 #ifdef CONFIG_NUMA_BALANCING
show_numa_stats(struct task_struct * p,struct seq_file * m)13080 void show_numa_stats(struct task_struct *p, struct seq_file *m)
13081 {
13082 int node;
13083 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
13084 struct numa_group *ng;
13085
13086 rcu_read_lock();
13087 ng = rcu_dereference(p->numa_group);
13088 for_each_online_node(node) {
13089 if (p->numa_faults) {
13090 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
13091 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
13092 }
13093 if (ng) {
13094 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
13095 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
13096 }
13097 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
13098 }
13099 rcu_read_unlock();
13100 }
13101 #endif /* CONFIG_NUMA_BALANCING */
13102 #endif /* CONFIG_SCHED_DEBUG */
13103
init_sched_fair_class(void)13104 __init void init_sched_fair_class(void)
13105 {
13106 #ifdef CONFIG_SMP
13107 int i;
13108
13109 for_each_possible_cpu(i) {
13110 zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
13111 zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
13112 zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i),
13113 GFP_KERNEL, cpu_to_node(i));
13114
13115 #ifdef CONFIG_CFS_BANDWIDTH
13116 INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
13117 INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
13118 #endif
13119 }
13120
13121 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
13122
13123 #ifdef CONFIG_NO_HZ_COMMON
13124 nohz.next_balance = jiffies;
13125 nohz.next_blocked = jiffies;
13126 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
13127 #endif
13128 #endif /* SMP */
13129
13130 }
13131