1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Scheduler topology setup/handling methods
4 */
5 #include "sched.h"
6
7 DEFINE_MUTEX(sched_domains_mutex);
8
9 /* Protected by sched_domains_mutex: */
10 static cpumask_var_t sched_domains_tmpmask;
11 static cpumask_var_t sched_domains_tmpmask2;
12
13 #ifdef CONFIG_SCHED_DEBUG
14
sched_debug_setup(char * str)15 static int __init sched_debug_setup(char *str)
16 {
17 sched_debug_enabled = true;
18
19 return 0;
20 }
21 early_param("sched_debug", sched_debug_setup);
22
sched_debug(void)23 static inline bool sched_debug(void)
24 {
25 return sched_debug_enabled;
26 }
27
sched_domain_debug_one(struct sched_domain * sd,int cpu,int level,struct cpumask * groupmask)28 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
29 struct cpumask *groupmask)
30 {
31 struct sched_group *group = sd->groups;
32
33 cpumask_clear(groupmask);
34
35 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
36
37 if (!(sd->flags & SD_LOAD_BALANCE)) {
38 printk("does not load-balance\n");
39 if (sd->parent)
40 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
41 return -1;
42 }
43
44 printk(KERN_CONT "span=%*pbl level=%s\n",
45 cpumask_pr_args(sched_domain_span(sd)), sd->name);
46
47 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
48 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
49 }
50 if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
51 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
52 }
53
54 printk(KERN_DEBUG "%*s groups:", level + 1, "");
55 do {
56 if (!group) {
57 printk("\n");
58 printk(KERN_ERR "ERROR: group is NULL\n");
59 break;
60 }
61
62 if (!cpumask_weight(sched_group_span(group))) {
63 printk(KERN_CONT "\n");
64 printk(KERN_ERR "ERROR: empty group\n");
65 break;
66 }
67
68 if (!(sd->flags & SD_OVERLAP) &&
69 cpumask_intersects(groupmask, sched_group_span(group))) {
70 printk(KERN_CONT "\n");
71 printk(KERN_ERR "ERROR: repeated CPUs\n");
72 break;
73 }
74
75 cpumask_or(groupmask, groupmask, sched_group_span(group));
76
77 printk(KERN_CONT " %d:{ span=%*pbl",
78 group->sgc->id,
79 cpumask_pr_args(sched_group_span(group)));
80
81 if ((sd->flags & SD_OVERLAP) &&
82 !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
83 printk(KERN_CONT " mask=%*pbl",
84 cpumask_pr_args(group_balance_mask(group)));
85 }
86
87 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
88 printk(KERN_CONT " cap=%lu", group->sgc->capacity);
89
90 if (group == sd->groups && sd->child &&
91 !cpumask_equal(sched_domain_span(sd->child),
92 sched_group_span(group))) {
93 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
94 }
95
96 printk(KERN_CONT " }");
97
98 group = group->next;
99
100 if (group != sd->groups)
101 printk(KERN_CONT ",");
102
103 } while (group != sd->groups);
104 printk(KERN_CONT "\n");
105
106 if (!cpumask_equal(sched_domain_span(sd), groupmask))
107 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
108
109 if (sd->parent &&
110 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
111 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
112 return 0;
113 }
114
sched_domain_debug(struct sched_domain * sd,int cpu)115 static void sched_domain_debug(struct sched_domain *sd, int cpu)
116 {
117 int level = 0;
118
119 if (!sched_debug_enabled)
120 return;
121
122 if (!sd) {
123 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
124 return;
125 }
126
127 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
128
129 for (;;) {
130 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
131 break;
132 level++;
133 sd = sd->parent;
134 if (!sd)
135 break;
136 }
137 }
138 #else /* !CONFIG_SCHED_DEBUG */
139
140 # define sched_debug_enabled 0
141 # define sched_domain_debug(sd, cpu) do { } while (0)
sched_debug(void)142 static inline bool sched_debug(void)
143 {
144 return false;
145 }
146 #endif /* CONFIG_SCHED_DEBUG */
147
sd_degenerate(struct sched_domain * sd)148 static int sd_degenerate(struct sched_domain *sd)
149 {
150 if (cpumask_weight(sched_domain_span(sd)) == 1)
151 return 1;
152
153 /* Following flags need at least 2 groups */
154 if (sd->flags & (SD_LOAD_BALANCE |
155 SD_BALANCE_NEWIDLE |
156 SD_BALANCE_FORK |
157 SD_BALANCE_EXEC |
158 SD_SHARE_CPUCAPACITY |
159 SD_ASYM_CPUCAPACITY |
160 SD_SHARE_PKG_RESOURCES |
161 SD_SHARE_POWERDOMAIN)) {
162 if (sd->groups != sd->groups->next)
163 return 0;
164 }
165
166 /* Following flags don't use groups */
167 if (sd->flags & (SD_WAKE_AFFINE))
168 return 0;
169
170 return 1;
171 }
172
173 static int
sd_parent_degenerate(struct sched_domain * sd,struct sched_domain * parent)174 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
175 {
176 unsigned long cflags = sd->flags, pflags = parent->flags;
177
178 if (sd_degenerate(parent))
179 return 1;
180
181 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
182 return 0;
183
184 /* Flags needing groups don't count if only 1 group in parent */
185 if (parent->groups == parent->groups->next) {
186 pflags &= ~(SD_LOAD_BALANCE |
187 SD_BALANCE_NEWIDLE |
188 SD_BALANCE_FORK |
189 SD_BALANCE_EXEC |
190 SD_ASYM_CPUCAPACITY |
191 SD_SHARE_CPUCAPACITY |
192 SD_SHARE_PKG_RESOURCES |
193 SD_PREFER_SIBLING |
194 SD_SHARE_POWERDOMAIN);
195 if (nr_node_ids == 1)
196 pflags &= ~SD_SERIALIZE;
197 }
198 if (~cflags & pflags)
199 return 0;
200
201 return 1;
202 }
203
204 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
205 DEFINE_STATIC_KEY_FALSE(sched_energy_present);
206 unsigned int sysctl_sched_energy_aware = 1;
207 DEFINE_MUTEX(sched_energy_mutex);
208 bool sched_energy_update;
209
210 #ifdef CONFIG_PROC_SYSCTL
sched_energy_aware_handler(struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)211 int sched_energy_aware_handler(struct ctl_table *table, int write,
212 void __user *buffer, size_t *lenp, loff_t *ppos)
213 {
214 int ret, state;
215
216 if (write && !capable(CAP_SYS_ADMIN))
217 return -EPERM;
218
219 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
220 if (!ret && write) {
221 state = static_branch_unlikely(&sched_energy_present);
222 if (state != sysctl_sched_energy_aware) {
223 mutex_lock(&sched_energy_mutex);
224 sched_energy_update = 1;
225 rebuild_sched_domains();
226 sched_energy_update = 0;
227 mutex_unlock(&sched_energy_mutex);
228 }
229 }
230
231 return ret;
232 }
233 #endif
234
free_pd(struct perf_domain * pd)235 static void free_pd(struct perf_domain *pd)
236 {
237 struct perf_domain *tmp;
238
239 while (pd) {
240 tmp = pd->next;
241 kfree(pd);
242 pd = tmp;
243 }
244 }
245
find_pd(struct perf_domain * pd,int cpu)246 static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
247 {
248 while (pd) {
249 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
250 return pd;
251 pd = pd->next;
252 }
253
254 return NULL;
255 }
256
pd_init(int cpu)257 static struct perf_domain *pd_init(int cpu)
258 {
259 struct em_perf_domain *obj = em_cpu_get(cpu);
260 struct perf_domain *pd;
261
262 if (!obj) {
263 if (sched_debug())
264 pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
265 return NULL;
266 }
267
268 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
269 if (!pd)
270 return NULL;
271 pd->em_pd = obj;
272
273 return pd;
274 }
275
perf_domain_debug(const struct cpumask * cpu_map,struct perf_domain * pd)276 static void perf_domain_debug(const struct cpumask *cpu_map,
277 struct perf_domain *pd)
278 {
279 if (!sched_debug() || !pd)
280 return;
281
282 printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
283
284 while (pd) {
285 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_cstate=%d }",
286 cpumask_first(perf_domain_span(pd)),
287 cpumask_pr_args(perf_domain_span(pd)),
288 em_pd_nr_cap_states(pd->em_pd));
289 pd = pd->next;
290 }
291
292 printk(KERN_CONT "\n");
293 }
294
destroy_perf_domain_rcu(struct rcu_head * rp)295 static void destroy_perf_domain_rcu(struct rcu_head *rp)
296 {
297 struct perf_domain *pd;
298
299 pd = container_of(rp, struct perf_domain, rcu);
300 free_pd(pd);
301 }
302
sched_energy_set(bool has_eas)303 static void sched_energy_set(bool has_eas)
304 {
305 if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
306 if (sched_debug())
307 pr_info("%s: stopping EAS\n", __func__);
308 static_branch_disable_cpuslocked(&sched_energy_present);
309 } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
310 if (sched_debug())
311 pr_info("%s: starting EAS\n", __func__);
312 static_branch_enable_cpuslocked(&sched_energy_present);
313 }
314 }
315
316 /*
317 * EAS can be used on a root domain if it meets all the following conditions:
318 * 1. an Energy Model (EM) is available;
319 * 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
320 * 3. the EM complexity is low enough to keep scheduling overheads low;
321 * 4. schedutil is driving the frequency of all CPUs of the rd;
322 *
323 * The complexity of the Energy Model is defined as:
324 *
325 * C = nr_pd * (nr_cpus + nr_cs)
326 *
327 * with parameters defined as:
328 * - nr_pd: the number of performance domains
329 * - nr_cpus: the number of CPUs
330 * - nr_cs: the sum of the number of capacity states of all performance
331 * domains (for example, on a system with 2 performance domains,
332 * with 10 capacity states each, nr_cs = 2 * 10 = 20).
333 *
334 * It is generally not a good idea to use such a model in the wake-up path on
335 * very complex platforms because of the associated scheduling overheads. The
336 * arbitrary constraint below prevents that. It makes EAS usable up to 16 CPUs
337 * with per-CPU DVFS and less than 8 capacity states each, for example.
338 */
339 #define EM_MAX_COMPLEXITY 2048
340
341 extern struct cpufreq_governor schedutil_gov;
build_perf_domains(const struct cpumask * cpu_map)342 static bool build_perf_domains(const struct cpumask *cpu_map)
343 {
344 int i, nr_pd = 0, nr_cs = 0, nr_cpus = cpumask_weight(cpu_map);
345 struct perf_domain *pd = NULL, *tmp;
346 int cpu = cpumask_first(cpu_map);
347 struct root_domain *rd = cpu_rq(cpu)->rd;
348 struct cpufreq_policy *policy;
349 struct cpufreq_governor *gov;
350
351 if (!sysctl_sched_energy_aware)
352 goto free;
353
354 /* EAS is enabled for asymmetric CPU capacity topologies. */
355 if (!per_cpu(sd_asym_cpucapacity, cpu)) {
356 if (sched_debug()) {
357 pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
358 cpumask_pr_args(cpu_map));
359 }
360 goto free;
361 }
362
363 for_each_cpu(i, cpu_map) {
364 /* Skip already covered CPUs. */
365 if (find_pd(pd, i))
366 continue;
367
368 /* Do not attempt EAS if schedutil is not being used. */
369 policy = cpufreq_cpu_get(i);
370 if (!policy)
371 goto free;
372 gov = policy->governor;
373 cpufreq_cpu_put(policy);
374 if (gov != &schedutil_gov) {
375 if (rd->pd)
376 pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
377 cpumask_pr_args(cpu_map));
378 goto free;
379 }
380
381 /* Create the new pd and add it to the local list. */
382 tmp = pd_init(i);
383 if (!tmp)
384 goto free;
385 tmp->next = pd;
386 pd = tmp;
387
388 /*
389 * Count performance domains and capacity states for the
390 * complexity check.
391 */
392 nr_pd++;
393 nr_cs += em_pd_nr_cap_states(pd->em_pd);
394 }
395
396 /* Bail out if the Energy Model complexity is too high. */
397 if (nr_pd * (nr_cs + nr_cpus) > EM_MAX_COMPLEXITY) {
398 WARN(1, "rd %*pbl: Failed to start EAS, EM complexity is too high\n",
399 cpumask_pr_args(cpu_map));
400 goto free;
401 }
402
403 perf_domain_debug(cpu_map, pd);
404
405 /* Attach the new list of performance domains to the root domain. */
406 tmp = rd->pd;
407 rcu_assign_pointer(rd->pd, pd);
408 if (tmp)
409 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
410
411 return !!pd;
412
413 free:
414 free_pd(pd);
415 tmp = rd->pd;
416 rcu_assign_pointer(rd->pd, NULL);
417 if (tmp)
418 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
419
420 return false;
421 }
422 #else
free_pd(struct perf_domain * pd)423 static void free_pd(struct perf_domain *pd) { }
424 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
425
free_rootdomain(struct rcu_head * rcu)426 static void free_rootdomain(struct rcu_head *rcu)
427 {
428 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
429
430 cpupri_cleanup(&rd->cpupri);
431 cpudl_cleanup(&rd->cpudl);
432 free_cpumask_var(rd->dlo_mask);
433 free_cpumask_var(rd->rto_mask);
434 free_cpumask_var(rd->online);
435 free_cpumask_var(rd->span);
436 free_pd(rd->pd);
437 kfree(rd);
438 }
439
rq_attach_root(struct rq * rq,struct root_domain * rd)440 void rq_attach_root(struct rq *rq, struct root_domain *rd)
441 {
442 struct root_domain *old_rd = NULL;
443 unsigned long flags;
444
445 raw_spin_lock_irqsave(&rq->lock, flags);
446
447 if (rq->rd) {
448 old_rd = rq->rd;
449
450 if (cpumask_test_cpu(rq->cpu, old_rd->online))
451 set_rq_offline(rq);
452
453 cpumask_clear_cpu(rq->cpu, old_rd->span);
454
455 /*
456 * If we dont want to free the old_rd yet then
457 * set old_rd to NULL to skip the freeing later
458 * in this function:
459 */
460 if (!atomic_dec_and_test(&old_rd->refcount))
461 old_rd = NULL;
462 }
463
464 atomic_inc(&rd->refcount);
465 rq->rd = rd;
466
467 cpumask_set_cpu(rq->cpu, rd->span);
468 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
469 set_rq_online(rq);
470
471 raw_spin_unlock_irqrestore(&rq->lock, flags);
472
473 if (old_rd)
474 call_rcu(&old_rd->rcu, free_rootdomain);
475 }
476
sched_get_rd(struct root_domain * rd)477 void sched_get_rd(struct root_domain *rd)
478 {
479 atomic_inc(&rd->refcount);
480 }
481
sched_put_rd(struct root_domain * rd)482 void sched_put_rd(struct root_domain *rd)
483 {
484 if (!atomic_dec_and_test(&rd->refcount))
485 return;
486
487 call_rcu(&rd->rcu, free_rootdomain);
488 }
489
init_rootdomain(struct root_domain * rd)490 static int init_rootdomain(struct root_domain *rd)
491 {
492 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
493 goto out;
494 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
495 goto free_span;
496 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
497 goto free_online;
498 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
499 goto free_dlo_mask;
500
501 #ifdef HAVE_RT_PUSH_IPI
502 rd->rto_cpu = -1;
503 raw_spin_lock_init(&rd->rto_lock);
504 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
505 #endif
506
507 init_dl_bw(&rd->dl_bw);
508 if (cpudl_init(&rd->cpudl) != 0)
509 goto free_rto_mask;
510
511 if (cpupri_init(&rd->cpupri) != 0)
512 goto free_cpudl;
513
514 init_max_cpu_capacity(&rd->max_cpu_capacity);
515
516 return 0;
517
518 free_cpudl:
519 cpudl_cleanup(&rd->cpudl);
520 free_rto_mask:
521 free_cpumask_var(rd->rto_mask);
522 free_dlo_mask:
523 free_cpumask_var(rd->dlo_mask);
524 free_online:
525 free_cpumask_var(rd->online);
526 free_span:
527 free_cpumask_var(rd->span);
528 out:
529 return -ENOMEM;
530 }
531
532 /*
533 * By default the system creates a single root-domain with all CPUs as
534 * members (mimicking the global state we have today).
535 */
536 struct root_domain def_root_domain;
537
init_defrootdomain(void)538 void init_defrootdomain(void)
539 {
540 init_rootdomain(&def_root_domain);
541
542 atomic_set(&def_root_domain.refcount, 1);
543 }
544
alloc_rootdomain(void)545 static struct root_domain *alloc_rootdomain(void)
546 {
547 struct root_domain *rd;
548
549 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
550 if (!rd)
551 return NULL;
552
553 if (init_rootdomain(rd) != 0) {
554 kfree(rd);
555 return NULL;
556 }
557
558 return rd;
559 }
560
free_sched_groups(struct sched_group * sg,int free_sgc)561 static void free_sched_groups(struct sched_group *sg, int free_sgc)
562 {
563 struct sched_group *tmp, *first;
564
565 if (!sg)
566 return;
567
568 first = sg;
569 do {
570 tmp = sg->next;
571
572 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
573 kfree(sg->sgc);
574
575 if (atomic_dec_and_test(&sg->ref))
576 kfree(sg);
577 sg = tmp;
578 } while (sg != first);
579 }
580
destroy_sched_domain(struct sched_domain * sd)581 static void destroy_sched_domain(struct sched_domain *sd)
582 {
583 /*
584 * A normal sched domain may have multiple group references, an
585 * overlapping domain, having private groups, only one. Iterate,
586 * dropping group/capacity references, freeing where none remain.
587 */
588 free_sched_groups(sd->groups, 1);
589
590 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
591 kfree(sd->shared);
592 kfree(sd);
593 }
594
destroy_sched_domains_rcu(struct rcu_head * rcu)595 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
596 {
597 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
598
599 while (sd) {
600 struct sched_domain *parent = sd->parent;
601 destroy_sched_domain(sd);
602 sd = parent;
603 }
604 }
605
destroy_sched_domains(struct sched_domain * sd)606 static void destroy_sched_domains(struct sched_domain *sd)
607 {
608 if (sd)
609 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
610 }
611
612 /*
613 * Keep a special pointer to the highest sched_domain that has
614 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
615 * allows us to avoid some pointer chasing select_idle_sibling().
616 *
617 * Also keep a unique ID per domain (we use the first CPU number in
618 * the cpumask of the domain), this allows us to quickly tell if
619 * two CPUs are in the same cache domain, see cpus_share_cache().
620 */
621 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
622 DEFINE_PER_CPU(int, sd_llc_size);
623 DEFINE_PER_CPU(int, sd_llc_id);
624 DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
625 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
626 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
627 DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
628 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
629
update_top_cache_domain(int cpu)630 static void update_top_cache_domain(int cpu)
631 {
632 struct sched_domain_shared *sds = NULL;
633 struct sched_domain *sd;
634 int id = cpu;
635 int size = 1;
636
637 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
638 if (sd) {
639 id = cpumask_first(sched_domain_span(sd));
640 size = cpumask_weight(sched_domain_span(sd));
641 sds = sd->shared;
642 }
643
644 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
645 per_cpu(sd_llc_size, cpu) = size;
646 per_cpu(sd_llc_id, cpu) = id;
647 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
648
649 sd = lowest_flag_domain(cpu, SD_NUMA);
650 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
651
652 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
653 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
654
655 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY);
656 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
657 }
658
659 /*
660 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
661 * hold the hotplug lock.
662 */
663 static void
cpu_attach_domain(struct sched_domain * sd,struct root_domain * rd,int cpu)664 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
665 {
666 struct rq *rq = cpu_rq(cpu);
667 struct sched_domain *tmp;
668
669 /* Remove the sched domains which do not contribute to scheduling. */
670 for (tmp = sd; tmp; ) {
671 struct sched_domain *parent = tmp->parent;
672 if (!parent)
673 break;
674
675 if (sd_parent_degenerate(tmp, parent)) {
676 tmp->parent = parent->parent;
677 if (parent->parent)
678 parent->parent->child = tmp;
679 /*
680 * Transfer SD_PREFER_SIBLING down in case of a
681 * degenerate parent; the spans match for this
682 * so the property transfers.
683 */
684 if (parent->flags & SD_PREFER_SIBLING)
685 tmp->flags |= SD_PREFER_SIBLING;
686 destroy_sched_domain(parent);
687 } else
688 tmp = tmp->parent;
689 }
690
691 if (sd && sd_degenerate(sd)) {
692 tmp = sd;
693 sd = sd->parent;
694 destroy_sched_domain(tmp);
695 if (sd)
696 sd->child = NULL;
697 }
698
699 sched_domain_debug(sd, cpu);
700
701 rq_attach_root(rq, rd);
702 tmp = rq->sd;
703 rcu_assign_pointer(rq->sd, sd);
704 dirty_sched_domain_sysctl(cpu);
705 destroy_sched_domains(tmp);
706
707 update_top_cache_domain(cpu);
708 }
709
710 struct s_data {
711 struct sched_domain * __percpu *sd;
712 struct root_domain *rd;
713 };
714
715 enum s_alloc {
716 sa_rootdomain,
717 sa_sd,
718 sa_sd_storage,
719 sa_none,
720 };
721
722 /*
723 * Return the canonical balance CPU for this group, this is the first CPU
724 * of this group that's also in the balance mask.
725 *
726 * The balance mask are all those CPUs that could actually end up at this
727 * group. See build_balance_mask().
728 *
729 * Also see should_we_balance().
730 */
group_balance_cpu(struct sched_group * sg)731 int group_balance_cpu(struct sched_group *sg)
732 {
733 return cpumask_first(group_balance_mask(sg));
734 }
735
736
737 /*
738 * NUMA topology (first read the regular topology blurb below)
739 *
740 * Given a node-distance table, for example:
741 *
742 * node 0 1 2 3
743 * 0: 10 20 30 20
744 * 1: 20 10 20 30
745 * 2: 30 20 10 20
746 * 3: 20 30 20 10
747 *
748 * which represents a 4 node ring topology like:
749 *
750 * 0 ----- 1
751 * | |
752 * | |
753 * | |
754 * 3 ----- 2
755 *
756 * We want to construct domains and groups to represent this. The way we go
757 * about doing this is to build the domains on 'hops'. For each NUMA level we
758 * construct the mask of all nodes reachable in @level hops.
759 *
760 * For the above NUMA topology that gives 3 levels:
761 *
762 * NUMA-2 0-3 0-3 0-3 0-3
763 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
764 *
765 * NUMA-1 0-1,3 0-2 1-3 0,2-3
766 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
767 *
768 * NUMA-0 0 1 2 3
769 *
770 *
771 * As can be seen; things don't nicely line up as with the regular topology.
772 * When we iterate a domain in child domain chunks some nodes can be
773 * represented multiple times -- hence the "overlap" naming for this part of
774 * the topology.
775 *
776 * In order to minimize this overlap, we only build enough groups to cover the
777 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
778 *
779 * Because:
780 *
781 * - the first group of each domain is its child domain; this
782 * gets us the first 0-1,3
783 * - the only uncovered node is 2, who's child domain is 1-3.
784 *
785 * However, because of the overlap, computing a unique CPU for each group is
786 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
787 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
788 * end up at those groups (they would end up in group: 0-1,3).
789 *
790 * To correct this we have to introduce the group balance mask. This mask
791 * will contain those CPUs in the group that can reach this group given the
792 * (child) domain tree.
793 *
794 * With this we can once again compute balance_cpu and sched_group_capacity
795 * relations.
796 *
797 * XXX include words on how balance_cpu is unique and therefore can be
798 * used for sched_group_capacity links.
799 *
800 *
801 * Another 'interesting' topology is:
802 *
803 * node 0 1 2 3
804 * 0: 10 20 20 30
805 * 1: 20 10 20 20
806 * 2: 20 20 10 20
807 * 3: 30 20 20 10
808 *
809 * Which looks a little like:
810 *
811 * 0 ----- 1
812 * | / |
813 * | / |
814 * | / |
815 * 2 ----- 3
816 *
817 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
818 * are not.
819 *
820 * This leads to a few particularly weird cases where the sched_domain's are
821 * not of the same number for each CPU. Consider:
822 *
823 * NUMA-2 0-3 0-3
824 * groups: {0-2},{1-3} {1-3},{0-2}
825 *
826 * NUMA-1 0-2 0-3 0-3 1-3
827 *
828 * NUMA-0 0 1 2 3
829 *
830 */
831
832
833 /*
834 * Build the balance mask; it contains only those CPUs that can arrive at this
835 * group and should be considered to continue balancing.
836 *
837 * We do this during the group creation pass, therefore the group information
838 * isn't complete yet, however since each group represents a (child) domain we
839 * can fully construct this using the sched_domain bits (which are already
840 * complete).
841 */
842 static void
build_balance_mask(struct sched_domain * sd,struct sched_group * sg,struct cpumask * mask)843 build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
844 {
845 const struct cpumask *sg_span = sched_group_span(sg);
846 struct sd_data *sdd = sd->private;
847 struct sched_domain *sibling;
848 int i;
849
850 cpumask_clear(mask);
851
852 for_each_cpu(i, sg_span) {
853 sibling = *per_cpu_ptr(sdd->sd, i);
854
855 /*
856 * Can happen in the asymmetric case, where these siblings are
857 * unused. The mask will not be empty because those CPUs that
858 * do have the top domain _should_ span the domain.
859 */
860 if (!sibling->child)
861 continue;
862
863 /* If we would not end up here, we can't continue from here */
864 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
865 continue;
866
867 cpumask_set_cpu(i, mask);
868 }
869
870 /* We must not have empty masks here */
871 WARN_ON_ONCE(cpumask_empty(mask));
872 }
873
874 /*
875 * XXX: This creates per-node group entries; since the load-balancer will
876 * immediately access remote memory to construct this group's load-balance
877 * statistics having the groups node local is of dubious benefit.
878 */
879 static struct sched_group *
build_group_from_child_sched_domain(struct sched_domain * sd,int cpu)880 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
881 {
882 struct sched_group *sg;
883 struct cpumask *sg_span;
884
885 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
886 GFP_KERNEL, cpu_to_node(cpu));
887
888 if (!sg)
889 return NULL;
890
891 sg_span = sched_group_span(sg);
892 if (sd->child)
893 cpumask_copy(sg_span, sched_domain_span(sd->child));
894 else
895 cpumask_copy(sg_span, sched_domain_span(sd));
896
897 atomic_inc(&sg->ref);
898 return sg;
899 }
900
init_overlap_sched_group(struct sched_domain * sd,struct sched_group * sg)901 static void init_overlap_sched_group(struct sched_domain *sd,
902 struct sched_group *sg)
903 {
904 struct cpumask *mask = sched_domains_tmpmask2;
905 struct sd_data *sdd = sd->private;
906 struct cpumask *sg_span;
907 int cpu;
908
909 build_balance_mask(sd, sg, mask);
910 cpu = cpumask_first_and(sched_group_span(sg), mask);
911
912 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
913 if (atomic_inc_return(&sg->sgc->ref) == 1)
914 cpumask_copy(group_balance_mask(sg), mask);
915 else
916 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
917
918 /*
919 * Initialize sgc->capacity such that even if we mess up the
920 * domains and no possible iteration will get us here, we won't
921 * die on a /0 trap.
922 */
923 sg_span = sched_group_span(sg);
924 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
925 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
926 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
927 }
928
929 static int
build_overlap_sched_groups(struct sched_domain * sd,int cpu)930 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
931 {
932 struct sched_group *first = NULL, *last = NULL, *sg;
933 const struct cpumask *span = sched_domain_span(sd);
934 struct cpumask *covered = sched_domains_tmpmask;
935 struct sd_data *sdd = sd->private;
936 struct sched_domain *sibling;
937 int i;
938
939 cpumask_clear(covered);
940
941 for_each_cpu_wrap(i, span, cpu) {
942 struct cpumask *sg_span;
943
944 if (cpumask_test_cpu(i, covered))
945 continue;
946
947 sibling = *per_cpu_ptr(sdd->sd, i);
948
949 /*
950 * Asymmetric node setups can result in situations where the
951 * domain tree is of unequal depth, make sure to skip domains
952 * that already cover the entire range.
953 *
954 * In that case build_sched_domains() will have terminated the
955 * iteration early and our sibling sd spans will be empty.
956 * Domains should always include the CPU they're built on, so
957 * check that.
958 */
959 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
960 continue;
961
962 sg = build_group_from_child_sched_domain(sibling, cpu);
963 if (!sg)
964 goto fail;
965
966 sg_span = sched_group_span(sg);
967 cpumask_or(covered, covered, sg_span);
968
969 init_overlap_sched_group(sd, sg);
970
971 if (!first)
972 first = sg;
973 if (last)
974 last->next = sg;
975 last = sg;
976 last->next = first;
977 }
978 sd->groups = first;
979
980 return 0;
981
982 fail:
983 free_sched_groups(first, 0);
984
985 return -ENOMEM;
986 }
987
988
989 /*
990 * Package topology (also see the load-balance blurb in fair.c)
991 *
992 * The scheduler builds a tree structure to represent a number of important
993 * topology features. By default (default_topology[]) these include:
994 *
995 * - Simultaneous multithreading (SMT)
996 * - Multi-Core Cache (MC)
997 * - Package (DIE)
998 *
999 * Where the last one more or less denotes everything up to a NUMA node.
1000 *
1001 * The tree consists of 3 primary data structures:
1002 *
1003 * sched_domain -> sched_group -> sched_group_capacity
1004 * ^ ^ ^ ^
1005 * `-' `-'
1006 *
1007 * The sched_domains are per-CPU and have a two way link (parent & child) and
1008 * denote the ever growing mask of CPUs belonging to that level of topology.
1009 *
1010 * Each sched_domain has a circular (double) linked list of sched_group's, each
1011 * denoting the domains of the level below (or individual CPUs in case of the
1012 * first domain level). The sched_group linked by a sched_domain includes the
1013 * CPU of that sched_domain [*].
1014 *
1015 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1016 *
1017 * CPU 0 1 2 3 4 5 6 7
1018 *
1019 * DIE [ ]
1020 * MC [ ] [ ]
1021 * SMT [ ] [ ] [ ] [ ]
1022 *
1023 * - or -
1024 *
1025 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1026 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1027 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1028 *
1029 * CPU 0 1 2 3 4 5 6 7
1030 *
1031 * One way to think about it is: sched_domain moves you up and down among these
1032 * topology levels, while sched_group moves you sideways through it, at child
1033 * domain granularity.
1034 *
1035 * sched_group_capacity ensures each unique sched_group has shared storage.
1036 *
1037 * There are two related construction problems, both require a CPU that
1038 * uniquely identify each group (for a given domain):
1039 *
1040 * - The first is the balance_cpu (see should_we_balance() and the
1041 * load-balance blub in fair.c); for each group we only want 1 CPU to
1042 * continue balancing at a higher domain.
1043 *
1044 * - The second is the sched_group_capacity; we want all identical groups
1045 * to share a single sched_group_capacity.
1046 *
1047 * Since these topologies are exclusive by construction. That is, its
1048 * impossible for an SMT thread to belong to multiple cores, and cores to
1049 * be part of multiple caches. There is a very clear and unique location
1050 * for each CPU in the hierarchy.
1051 *
1052 * Therefore computing a unique CPU for each group is trivial (the iteration
1053 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1054 * group), we can simply pick the first CPU in each group.
1055 *
1056 *
1057 * [*] in other words, the first group of each domain is its child domain.
1058 */
1059
get_group(int cpu,struct sd_data * sdd)1060 static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1061 {
1062 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1063 struct sched_domain *child = sd->child;
1064 struct sched_group *sg;
1065 bool already_visited;
1066
1067 if (child)
1068 cpu = cpumask_first(sched_domain_span(child));
1069
1070 sg = *per_cpu_ptr(sdd->sg, cpu);
1071 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1072
1073 /* Increase refcounts for claim_allocations: */
1074 already_visited = atomic_inc_return(&sg->ref) > 1;
1075 /* sgc visits should follow a similar trend as sg */
1076 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1077
1078 /* If we have already visited that group, it's already initialized. */
1079 if (already_visited)
1080 return sg;
1081
1082 if (child) {
1083 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1084 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1085 } else {
1086 cpumask_set_cpu(cpu, sched_group_span(sg));
1087 cpumask_set_cpu(cpu, group_balance_mask(sg));
1088 }
1089
1090 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1091 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1092 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1093
1094 return sg;
1095 }
1096
1097 /*
1098 * build_sched_groups will build a circular linked list of the groups
1099 * covered by the given span, will set each group's ->cpumask correctly,
1100 * and will initialize their ->sgc.
1101 *
1102 * Assumes the sched_domain tree is fully constructed
1103 */
1104 static int
build_sched_groups(struct sched_domain * sd,int cpu)1105 build_sched_groups(struct sched_domain *sd, int cpu)
1106 {
1107 struct sched_group *first = NULL, *last = NULL;
1108 struct sd_data *sdd = sd->private;
1109 const struct cpumask *span = sched_domain_span(sd);
1110 struct cpumask *covered;
1111 int i;
1112
1113 lockdep_assert_held(&sched_domains_mutex);
1114 covered = sched_domains_tmpmask;
1115
1116 cpumask_clear(covered);
1117
1118 for_each_cpu_wrap(i, span, cpu) {
1119 struct sched_group *sg;
1120
1121 if (cpumask_test_cpu(i, covered))
1122 continue;
1123
1124 sg = get_group(i, sdd);
1125
1126 cpumask_or(covered, covered, sched_group_span(sg));
1127
1128 if (!first)
1129 first = sg;
1130 if (last)
1131 last->next = sg;
1132 last = sg;
1133 }
1134 last->next = first;
1135 sd->groups = first;
1136
1137 return 0;
1138 }
1139
1140 /*
1141 * Initialize sched groups cpu_capacity.
1142 *
1143 * cpu_capacity indicates the capacity of sched group, which is used while
1144 * distributing the load between different sched groups in a sched domain.
1145 * Typically cpu_capacity for all the groups in a sched domain will be same
1146 * unless there are asymmetries in the topology. If there are asymmetries,
1147 * group having more cpu_capacity will pickup more load compared to the
1148 * group having less cpu_capacity.
1149 */
init_sched_groups_capacity(int cpu,struct sched_domain * sd)1150 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1151 {
1152 struct sched_group *sg = sd->groups;
1153
1154 WARN_ON(!sg);
1155
1156 do {
1157 int cpu, max_cpu = -1;
1158
1159 sg->group_weight = cpumask_weight(sched_group_span(sg));
1160
1161 if (!(sd->flags & SD_ASYM_PACKING))
1162 goto next;
1163
1164 for_each_cpu(cpu, sched_group_span(sg)) {
1165 if (max_cpu < 0)
1166 max_cpu = cpu;
1167 else if (sched_asym_prefer(cpu, max_cpu))
1168 max_cpu = cpu;
1169 }
1170 sg->asym_prefer_cpu = max_cpu;
1171
1172 next:
1173 sg = sg->next;
1174 } while (sg != sd->groups);
1175
1176 if (cpu != group_balance_cpu(sg))
1177 return;
1178
1179 update_group_capacity(sd, cpu);
1180 }
1181
1182 /*
1183 * Initializers for schedule domains
1184 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1185 */
1186
1187 static int default_relax_domain_level = -1;
1188 int sched_domain_level_max;
1189
setup_relax_domain_level(char * str)1190 static int __init setup_relax_domain_level(char *str)
1191 {
1192 if (kstrtoint(str, 0, &default_relax_domain_level))
1193 pr_warn("Unable to set relax_domain_level\n");
1194
1195 return 1;
1196 }
1197 __setup("relax_domain_level=", setup_relax_domain_level);
1198
set_domain_attribute(struct sched_domain * sd,struct sched_domain_attr * attr)1199 static void set_domain_attribute(struct sched_domain *sd,
1200 struct sched_domain_attr *attr)
1201 {
1202 int request;
1203
1204 if (!attr || attr->relax_domain_level < 0) {
1205 if (default_relax_domain_level < 0)
1206 return;
1207 else
1208 request = default_relax_domain_level;
1209 } else
1210 request = attr->relax_domain_level;
1211 if (request < sd->level) {
1212 /* Turn off idle balance on this domain: */
1213 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1214 } else {
1215 /* Turn on idle balance on this domain: */
1216 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1217 }
1218 }
1219
1220 static void __sdt_free(const struct cpumask *cpu_map);
1221 static int __sdt_alloc(const struct cpumask *cpu_map);
1222
__free_domain_allocs(struct s_data * d,enum s_alloc what,const struct cpumask * cpu_map)1223 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1224 const struct cpumask *cpu_map)
1225 {
1226 switch (what) {
1227 case sa_rootdomain:
1228 if (!atomic_read(&d->rd->refcount))
1229 free_rootdomain(&d->rd->rcu);
1230 /* Fall through */
1231 case sa_sd:
1232 free_percpu(d->sd);
1233 /* Fall through */
1234 case sa_sd_storage:
1235 __sdt_free(cpu_map);
1236 /* Fall through */
1237 case sa_none:
1238 break;
1239 }
1240 }
1241
1242 static enum s_alloc
__visit_domain_allocation_hell(struct s_data * d,const struct cpumask * cpu_map)1243 __visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1244 {
1245 memset(d, 0, sizeof(*d));
1246
1247 if (__sdt_alloc(cpu_map))
1248 return sa_sd_storage;
1249 d->sd = alloc_percpu(struct sched_domain *);
1250 if (!d->sd)
1251 return sa_sd_storage;
1252 d->rd = alloc_rootdomain();
1253 if (!d->rd)
1254 return sa_sd;
1255
1256 return sa_rootdomain;
1257 }
1258
1259 /*
1260 * NULL the sd_data elements we've used to build the sched_domain and
1261 * sched_group structure so that the subsequent __free_domain_allocs()
1262 * will not free the data we're using.
1263 */
claim_allocations(int cpu,struct sched_domain * sd)1264 static void claim_allocations(int cpu, struct sched_domain *sd)
1265 {
1266 struct sd_data *sdd = sd->private;
1267
1268 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1269 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1270
1271 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1272 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1273
1274 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1275 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1276
1277 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1278 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1279 }
1280
1281 #ifdef CONFIG_NUMA
1282 enum numa_topology_type sched_numa_topology_type;
1283
1284 static int sched_domains_numa_levels;
1285 static int sched_domains_curr_level;
1286
1287 int sched_max_numa_distance;
1288 static int *sched_domains_numa_distance;
1289 static struct cpumask ***sched_domains_numa_masks;
1290 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
1291 #endif
1292
1293 /*
1294 * SD_flags allowed in topology descriptions.
1295 *
1296 * These flags are purely descriptive of the topology and do not prescribe
1297 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1298 * function:
1299 *
1300 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1301 * SD_SHARE_PKG_RESOURCES - describes shared caches
1302 * SD_NUMA - describes NUMA topologies
1303 * SD_SHARE_POWERDOMAIN - describes shared power domain
1304 *
1305 * Odd one out, which beside describing the topology has a quirk also
1306 * prescribes the desired behaviour that goes along with it:
1307 *
1308 * SD_ASYM_PACKING - describes SMT quirks
1309 */
1310 #define TOPOLOGY_SD_FLAGS \
1311 (SD_SHARE_CPUCAPACITY | \
1312 SD_SHARE_PKG_RESOURCES | \
1313 SD_NUMA | \
1314 SD_ASYM_PACKING | \
1315 SD_SHARE_POWERDOMAIN)
1316
1317 static struct sched_domain *
sd_init(struct sched_domain_topology_level * tl,const struct cpumask * cpu_map,struct sched_domain * child,int dflags,int cpu)1318 sd_init(struct sched_domain_topology_level *tl,
1319 const struct cpumask *cpu_map,
1320 struct sched_domain *child, int dflags, int cpu)
1321 {
1322 struct sd_data *sdd = &tl->data;
1323 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1324 int sd_id, sd_weight, sd_flags = 0;
1325
1326 #ifdef CONFIG_NUMA
1327 /*
1328 * Ugly hack to pass state to sd_numa_mask()...
1329 */
1330 sched_domains_curr_level = tl->numa_level;
1331 #endif
1332
1333 sd_weight = cpumask_weight(tl->mask(cpu));
1334
1335 if (tl->sd_flags)
1336 sd_flags = (*tl->sd_flags)();
1337 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1338 "wrong sd_flags in topology description\n"))
1339 sd_flags &= TOPOLOGY_SD_FLAGS;
1340
1341 /* Apply detected topology flags */
1342 sd_flags |= dflags;
1343
1344 *sd = (struct sched_domain){
1345 .min_interval = sd_weight,
1346 .max_interval = 2*sd_weight,
1347 .busy_factor = 32,
1348 .imbalance_pct = 125,
1349
1350 .cache_nice_tries = 0,
1351
1352 .flags = 1*SD_LOAD_BALANCE
1353 | 1*SD_BALANCE_NEWIDLE
1354 | 1*SD_BALANCE_EXEC
1355 | 1*SD_BALANCE_FORK
1356 | 0*SD_BALANCE_WAKE
1357 | 1*SD_WAKE_AFFINE
1358 | 0*SD_SHARE_CPUCAPACITY
1359 | 0*SD_SHARE_PKG_RESOURCES
1360 | 0*SD_SERIALIZE
1361 | 1*SD_PREFER_SIBLING
1362 | 0*SD_NUMA
1363 | sd_flags
1364 ,
1365
1366 .last_balance = jiffies,
1367 .balance_interval = sd_weight,
1368 .max_newidle_lb_cost = 0,
1369 .next_decay_max_lb_cost = jiffies,
1370 .child = child,
1371 #ifdef CONFIG_SCHED_DEBUG
1372 .name = tl->name,
1373 #endif
1374 };
1375
1376 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1377 sd_id = cpumask_first(sched_domain_span(sd));
1378
1379 /*
1380 * Convert topological properties into behaviour.
1381 */
1382
1383 /* Don't attempt to spread across CPUs of different capacities. */
1384 if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1385 sd->child->flags &= ~SD_PREFER_SIBLING;
1386
1387 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1388 sd->imbalance_pct = 110;
1389
1390 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1391 sd->imbalance_pct = 117;
1392 sd->cache_nice_tries = 1;
1393
1394 #ifdef CONFIG_NUMA
1395 } else if (sd->flags & SD_NUMA) {
1396 sd->cache_nice_tries = 2;
1397
1398 sd->flags &= ~SD_PREFER_SIBLING;
1399 sd->flags |= SD_SERIALIZE;
1400 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1401 sd->flags &= ~(SD_BALANCE_EXEC |
1402 SD_BALANCE_FORK |
1403 SD_WAKE_AFFINE);
1404 }
1405
1406 #endif
1407 } else {
1408 sd->cache_nice_tries = 1;
1409 }
1410
1411 /*
1412 * For all levels sharing cache; connect a sched_domain_shared
1413 * instance.
1414 */
1415 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1416 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1417 atomic_inc(&sd->shared->ref);
1418 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1419 }
1420
1421 sd->private = sdd;
1422
1423 return sd;
1424 }
1425
1426 /*
1427 * Topology list, bottom-up.
1428 */
1429 static struct sched_domain_topology_level default_topology[] = {
1430 #ifdef CONFIG_SCHED_SMT
1431 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1432 #endif
1433 #ifdef CONFIG_SCHED_MC
1434 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1435 #endif
1436 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1437 { NULL, },
1438 };
1439
1440 static struct sched_domain_topology_level *sched_domain_topology =
1441 default_topology;
1442
1443 #define for_each_sd_topology(tl) \
1444 for (tl = sched_domain_topology; tl->mask; tl++)
1445
set_sched_topology(struct sched_domain_topology_level * tl)1446 void set_sched_topology(struct sched_domain_topology_level *tl)
1447 {
1448 if (WARN_ON_ONCE(sched_smp_initialized))
1449 return;
1450
1451 sched_domain_topology = tl;
1452 }
1453
1454 #ifdef CONFIG_NUMA
1455
sd_numa_mask(int cpu)1456 static const struct cpumask *sd_numa_mask(int cpu)
1457 {
1458 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1459 }
1460
sched_numa_warn(const char * str)1461 static void sched_numa_warn(const char *str)
1462 {
1463 static int done = false;
1464 int i,j;
1465
1466 if (done)
1467 return;
1468
1469 done = true;
1470
1471 printk(KERN_WARNING "ERROR: %s\n\n", str);
1472
1473 for (i = 0; i < nr_node_ids; i++) {
1474 printk(KERN_WARNING " ");
1475 for (j = 0; j < nr_node_ids; j++)
1476 printk(KERN_CONT "%02d ", node_distance(i,j));
1477 printk(KERN_CONT "\n");
1478 }
1479 printk(KERN_WARNING "\n");
1480 }
1481
find_numa_distance(int distance)1482 bool find_numa_distance(int distance)
1483 {
1484 int i;
1485
1486 if (distance == node_distance(0, 0))
1487 return true;
1488
1489 for (i = 0; i < sched_domains_numa_levels; i++) {
1490 if (sched_domains_numa_distance[i] == distance)
1491 return true;
1492 }
1493
1494 return false;
1495 }
1496
1497 /*
1498 * A system can have three types of NUMA topology:
1499 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1500 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1501 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1502 *
1503 * The difference between a glueless mesh topology and a backplane
1504 * topology lies in whether communication between not directly
1505 * connected nodes goes through intermediary nodes (where programs
1506 * could run), or through backplane controllers. This affects
1507 * placement of programs.
1508 *
1509 * The type of topology can be discerned with the following tests:
1510 * - If the maximum distance between any nodes is 1 hop, the system
1511 * is directly connected.
1512 * - If for two nodes A and B, located N > 1 hops away from each other,
1513 * there is an intermediary node C, which is < N hops away from both
1514 * nodes A and B, the system is a glueless mesh.
1515 */
init_numa_topology_type(void)1516 static void init_numa_topology_type(void)
1517 {
1518 int a, b, c, n;
1519
1520 n = sched_max_numa_distance;
1521
1522 if (sched_domains_numa_levels <= 2) {
1523 sched_numa_topology_type = NUMA_DIRECT;
1524 return;
1525 }
1526
1527 for_each_online_node(a) {
1528 for_each_online_node(b) {
1529 /* Find two nodes furthest removed from each other. */
1530 if (node_distance(a, b) < n)
1531 continue;
1532
1533 /* Is there an intermediary node between a and b? */
1534 for_each_online_node(c) {
1535 if (node_distance(a, c) < n &&
1536 node_distance(b, c) < n) {
1537 sched_numa_topology_type =
1538 NUMA_GLUELESS_MESH;
1539 return;
1540 }
1541 }
1542
1543 sched_numa_topology_type = NUMA_BACKPLANE;
1544 return;
1545 }
1546 }
1547 }
1548
1549
1550 #define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1551
sched_init_numa(void)1552 void sched_init_numa(void)
1553 {
1554 struct sched_domain_topology_level *tl;
1555 unsigned long *distance_map;
1556 int nr_levels = 0;
1557 int i, j;
1558
1559 /*
1560 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1561 * unique distances in the node_distance() table.
1562 */
1563 distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1564 if (!distance_map)
1565 return;
1566
1567 bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1568 for (i = 0; i < nr_node_ids; i++) {
1569 for (j = 0; j < nr_node_ids; j++) {
1570 int distance = node_distance(i, j);
1571
1572 if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1573 sched_numa_warn("Invalid distance value range");
1574 return;
1575 }
1576
1577 bitmap_set(distance_map, distance, 1);
1578 }
1579 }
1580 /*
1581 * We can now figure out how many unique distance values there are and
1582 * allocate memory accordingly.
1583 */
1584 nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1585
1586 sched_domains_numa_distance = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1587 if (!sched_domains_numa_distance) {
1588 bitmap_free(distance_map);
1589 return;
1590 }
1591
1592 for (i = 0, j = 0; i < nr_levels; i++, j++) {
1593 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1594 sched_domains_numa_distance[i] = j;
1595 }
1596
1597 bitmap_free(distance_map);
1598
1599 /*
1600 * 'nr_levels' contains the number of unique distances
1601 *
1602 * The sched_domains_numa_distance[] array includes the actual distance
1603 * numbers.
1604 */
1605
1606 /*
1607 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1608 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1609 * the array will contain less then 'nr_levels' members. This could be
1610 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1611 * in other functions.
1612 *
1613 * We reset it to 'nr_levels' at the end of this function.
1614 */
1615 sched_domains_numa_levels = 0;
1616
1617 sched_domains_numa_masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1618 if (!sched_domains_numa_masks)
1619 return;
1620
1621 /*
1622 * Now for each level, construct a mask per node which contains all
1623 * CPUs of nodes that are that many hops away from us.
1624 */
1625 for (i = 0; i < nr_levels; i++) {
1626 sched_domains_numa_masks[i] =
1627 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1628 if (!sched_domains_numa_masks[i])
1629 return;
1630
1631 for (j = 0; j < nr_node_ids; j++) {
1632 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1633 int k;
1634
1635 if (!mask)
1636 return;
1637
1638 sched_domains_numa_masks[i][j] = mask;
1639
1640 for_each_node(k) {
1641 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1642 sched_numa_warn("Node-distance not symmetric");
1643
1644 if (node_distance(j, k) > sched_domains_numa_distance[i])
1645 continue;
1646
1647 cpumask_or(mask, mask, cpumask_of_node(k));
1648 }
1649 }
1650 }
1651
1652 /* Compute default topology size */
1653 for (i = 0; sched_domain_topology[i].mask; i++);
1654
1655 tl = kzalloc((i + nr_levels + 1) *
1656 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1657 if (!tl)
1658 return;
1659
1660 /*
1661 * Copy the default topology bits..
1662 */
1663 for (i = 0; sched_domain_topology[i].mask; i++)
1664 tl[i] = sched_domain_topology[i];
1665
1666 /*
1667 * Add the NUMA identity distance, aka single NODE.
1668 */
1669 tl[i++] = (struct sched_domain_topology_level){
1670 .mask = sd_numa_mask,
1671 .numa_level = 0,
1672 SD_INIT_NAME(NODE)
1673 };
1674
1675 /*
1676 * .. and append 'j' levels of NUMA goodness.
1677 */
1678 for (j = 1; j < nr_levels; i++, j++) {
1679 tl[i] = (struct sched_domain_topology_level){
1680 .mask = sd_numa_mask,
1681 .sd_flags = cpu_numa_flags,
1682 .flags = SDTL_OVERLAP,
1683 .numa_level = j,
1684 SD_INIT_NAME(NUMA)
1685 };
1686 }
1687
1688 sched_domain_topology = tl;
1689
1690 sched_domains_numa_levels = nr_levels;
1691 sched_max_numa_distance = sched_domains_numa_distance[nr_levels - 1];
1692
1693 init_numa_topology_type();
1694 }
1695
sched_domains_numa_masks_set(unsigned int cpu)1696 void sched_domains_numa_masks_set(unsigned int cpu)
1697 {
1698 int node = cpu_to_node(cpu);
1699 int i, j;
1700
1701 for (i = 0; i < sched_domains_numa_levels; i++) {
1702 for (j = 0; j < nr_node_ids; j++) {
1703 if (node_distance(j, node) <= sched_domains_numa_distance[i])
1704 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1705 }
1706 }
1707 }
1708
sched_domains_numa_masks_clear(unsigned int cpu)1709 void sched_domains_numa_masks_clear(unsigned int cpu)
1710 {
1711 int i, j;
1712
1713 for (i = 0; i < sched_domains_numa_levels; i++) {
1714 for (j = 0; j < nr_node_ids; j++)
1715 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1716 }
1717 }
1718
1719 /*
1720 * sched_numa_find_closest() - given the NUMA topology, find the cpu
1721 * closest to @cpu from @cpumask.
1722 * cpumask: cpumask to find a cpu from
1723 * cpu: cpu to be close to
1724 *
1725 * returns: cpu, or nr_cpu_ids when nothing found.
1726 */
sched_numa_find_closest(const struct cpumask * cpus,int cpu)1727 int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
1728 {
1729 int i, j = cpu_to_node(cpu);
1730
1731 for (i = 0; i < sched_domains_numa_levels; i++) {
1732 cpu = cpumask_any_and(cpus, sched_domains_numa_masks[i][j]);
1733 if (cpu < nr_cpu_ids)
1734 return cpu;
1735 }
1736 return nr_cpu_ids;
1737 }
1738
1739 #endif /* CONFIG_NUMA */
1740
__sdt_alloc(const struct cpumask * cpu_map)1741 static int __sdt_alloc(const struct cpumask *cpu_map)
1742 {
1743 struct sched_domain_topology_level *tl;
1744 int j;
1745
1746 for_each_sd_topology(tl) {
1747 struct sd_data *sdd = &tl->data;
1748
1749 sdd->sd = alloc_percpu(struct sched_domain *);
1750 if (!sdd->sd)
1751 return -ENOMEM;
1752
1753 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1754 if (!sdd->sds)
1755 return -ENOMEM;
1756
1757 sdd->sg = alloc_percpu(struct sched_group *);
1758 if (!sdd->sg)
1759 return -ENOMEM;
1760
1761 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1762 if (!sdd->sgc)
1763 return -ENOMEM;
1764
1765 for_each_cpu(j, cpu_map) {
1766 struct sched_domain *sd;
1767 struct sched_domain_shared *sds;
1768 struct sched_group *sg;
1769 struct sched_group_capacity *sgc;
1770
1771 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1772 GFP_KERNEL, cpu_to_node(j));
1773 if (!sd)
1774 return -ENOMEM;
1775
1776 *per_cpu_ptr(sdd->sd, j) = sd;
1777
1778 sds = kzalloc_node(sizeof(struct sched_domain_shared),
1779 GFP_KERNEL, cpu_to_node(j));
1780 if (!sds)
1781 return -ENOMEM;
1782
1783 *per_cpu_ptr(sdd->sds, j) = sds;
1784
1785 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1786 GFP_KERNEL, cpu_to_node(j));
1787 if (!sg)
1788 return -ENOMEM;
1789
1790 sg->next = sg;
1791
1792 *per_cpu_ptr(sdd->sg, j) = sg;
1793
1794 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1795 GFP_KERNEL, cpu_to_node(j));
1796 if (!sgc)
1797 return -ENOMEM;
1798
1799 #ifdef CONFIG_SCHED_DEBUG
1800 sgc->id = j;
1801 #endif
1802
1803 *per_cpu_ptr(sdd->sgc, j) = sgc;
1804 }
1805 }
1806
1807 return 0;
1808 }
1809
__sdt_free(const struct cpumask * cpu_map)1810 static void __sdt_free(const struct cpumask *cpu_map)
1811 {
1812 struct sched_domain_topology_level *tl;
1813 int j;
1814
1815 for_each_sd_topology(tl) {
1816 struct sd_data *sdd = &tl->data;
1817
1818 for_each_cpu(j, cpu_map) {
1819 struct sched_domain *sd;
1820
1821 if (sdd->sd) {
1822 sd = *per_cpu_ptr(sdd->sd, j);
1823 if (sd && (sd->flags & SD_OVERLAP))
1824 free_sched_groups(sd->groups, 0);
1825 kfree(*per_cpu_ptr(sdd->sd, j));
1826 }
1827
1828 if (sdd->sds)
1829 kfree(*per_cpu_ptr(sdd->sds, j));
1830 if (sdd->sg)
1831 kfree(*per_cpu_ptr(sdd->sg, j));
1832 if (sdd->sgc)
1833 kfree(*per_cpu_ptr(sdd->sgc, j));
1834 }
1835 free_percpu(sdd->sd);
1836 sdd->sd = NULL;
1837 free_percpu(sdd->sds);
1838 sdd->sds = NULL;
1839 free_percpu(sdd->sg);
1840 sdd->sg = NULL;
1841 free_percpu(sdd->sgc);
1842 sdd->sgc = NULL;
1843 }
1844 }
1845
build_sched_domain(struct sched_domain_topology_level * tl,const struct cpumask * cpu_map,struct sched_domain_attr * attr,struct sched_domain * child,int dflags,int cpu)1846 static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1847 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1848 struct sched_domain *child, int dflags, int cpu)
1849 {
1850 struct sched_domain *sd = sd_init(tl, cpu_map, child, dflags, cpu);
1851
1852 if (child) {
1853 sd->level = child->level + 1;
1854 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1855 child->parent = sd;
1856
1857 if (!cpumask_subset(sched_domain_span(child),
1858 sched_domain_span(sd))) {
1859 pr_err("BUG: arch topology borken\n");
1860 #ifdef CONFIG_SCHED_DEBUG
1861 pr_err(" the %s domain not a subset of the %s domain\n",
1862 child->name, sd->name);
1863 #endif
1864 /* Fixup, ensure @sd has at least @child CPUs. */
1865 cpumask_or(sched_domain_span(sd),
1866 sched_domain_span(sd),
1867 sched_domain_span(child));
1868 }
1869
1870 }
1871 set_domain_attribute(sd, attr);
1872
1873 return sd;
1874 }
1875
1876 /*
1877 * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
1878 * any two given CPUs at this (non-NUMA) topology level.
1879 */
topology_span_sane(struct sched_domain_topology_level * tl,const struct cpumask * cpu_map,int cpu)1880 static bool topology_span_sane(struct sched_domain_topology_level *tl,
1881 const struct cpumask *cpu_map, int cpu)
1882 {
1883 int i;
1884
1885 /* NUMA levels are allowed to overlap */
1886 if (tl->flags & SDTL_OVERLAP)
1887 return true;
1888
1889 /*
1890 * Non-NUMA levels cannot partially overlap - they must be either
1891 * completely equal or completely disjoint. Otherwise we can end up
1892 * breaking the sched_group lists - i.e. a later get_group() pass
1893 * breaks the linking done for an earlier span.
1894 */
1895 for_each_cpu(i, cpu_map) {
1896 if (i == cpu)
1897 continue;
1898 /*
1899 * We should 'and' all those masks with 'cpu_map' to exactly
1900 * match the topology we're about to build, but that can only
1901 * remove CPUs, which only lessens our ability to detect
1902 * overlaps
1903 */
1904 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
1905 cpumask_intersects(tl->mask(cpu), tl->mask(i)))
1906 return false;
1907 }
1908
1909 return true;
1910 }
1911
1912 /*
1913 * Find the sched_domain_topology_level where all CPU capacities are visible
1914 * for all CPUs.
1915 */
1916 static struct sched_domain_topology_level
asym_cpu_capacity_level(const struct cpumask * cpu_map)1917 *asym_cpu_capacity_level(const struct cpumask *cpu_map)
1918 {
1919 int i, j, asym_level = 0;
1920 bool asym = false;
1921 struct sched_domain_topology_level *tl, *asym_tl = NULL;
1922 unsigned long cap;
1923
1924 /* Is there any asymmetry? */
1925 cap = arch_scale_cpu_capacity(cpumask_first(cpu_map));
1926
1927 for_each_cpu(i, cpu_map) {
1928 if (arch_scale_cpu_capacity(i) != cap) {
1929 asym = true;
1930 break;
1931 }
1932 }
1933
1934 if (!asym)
1935 return NULL;
1936
1937 /*
1938 * Examine topology from all CPU's point of views to detect the lowest
1939 * sched_domain_topology_level where a highest capacity CPU is visible
1940 * to everyone.
1941 */
1942 for_each_cpu(i, cpu_map) {
1943 unsigned long max_capacity = arch_scale_cpu_capacity(i);
1944 int tl_id = 0;
1945
1946 for_each_sd_topology(tl) {
1947 if (tl_id < asym_level)
1948 goto next_level;
1949
1950 for_each_cpu_and(j, tl->mask(i), cpu_map) {
1951 unsigned long capacity;
1952
1953 capacity = arch_scale_cpu_capacity(j);
1954
1955 if (capacity <= max_capacity)
1956 continue;
1957
1958 max_capacity = capacity;
1959 asym_level = tl_id;
1960 asym_tl = tl;
1961 }
1962 next_level:
1963 tl_id++;
1964 }
1965 }
1966
1967 return asym_tl;
1968 }
1969
1970
1971 /*
1972 * Build sched domains for a given set of CPUs and attach the sched domains
1973 * to the individual CPUs
1974 */
1975 static int
build_sched_domains(const struct cpumask * cpu_map,struct sched_domain_attr * attr)1976 build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1977 {
1978 enum s_alloc alloc_state = sa_none;
1979 struct sched_domain *sd;
1980 struct s_data d;
1981 int i, ret = -ENOMEM;
1982 struct sched_domain_topology_level *tl_asym;
1983 bool has_asym = false;
1984
1985 if (WARN_ON(cpumask_empty(cpu_map)))
1986 goto error;
1987
1988 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1989 if (alloc_state != sa_rootdomain)
1990 goto error;
1991
1992 tl_asym = asym_cpu_capacity_level(cpu_map);
1993
1994 /* Set up domains for CPUs specified by the cpu_map: */
1995 for_each_cpu(i, cpu_map) {
1996 struct sched_domain_topology_level *tl;
1997
1998 sd = NULL;
1999 for_each_sd_topology(tl) {
2000 int dflags = 0;
2001
2002 if (tl == tl_asym) {
2003 dflags |= SD_ASYM_CPUCAPACITY;
2004 has_asym = true;
2005 }
2006
2007 if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2008 goto error;
2009
2010 sd = build_sched_domain(tl, cpu_map, attr, sd, dflags, i);
2011
2012 if (tl == sched_domain_topology)
2013 *per_cpu_ptr(d.sd, i) = sd;
2014 if (tl->flags & SDTL_OVERLAP)
2015 sd->flags |= SD_OVERLAP;
2016 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2017 break;
2018 }
2019 }
2020
2021 /* Build the groups for the domains */
2022 for_each_cpu(i, cpu_map) {
2023 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2024 sd->span_weight = cpumask_weight(sched_domain_span(sd));
2025 if (sd->flags & SD_OVERLAP) {
2026 if (build_overlap_sched_groups(sd, i))
2027 goto error;
2028 } else {
2029 if (build_sched_groups(sd, i))
2030 goto error;
2031 }
2032 }
2033 }
2034
2035 /* Calculate CPU capacity for physical packages and nodes */
2036 for (i = nr_cpumask_bits-1; i >= 0; i--) {
2037 if (!cpumask_test_cpu(i, cpu_map))
2038 continue;
2039
2040 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2041 claim_allocations(i, sd);
2042 init_sched_groups_capacity(i, sd);
2043 }
2044 }
2045
2046 /* Attach the domains */
2047 rcu_read_lock();
2048 for_each_cpu(i, cpu_map) {
2049 sd = *per_cpu_ptr(d.sd, i);
2050 cpu_attach_domain(sd, d.rd, i);
2051 }
2052 rcu_read_unlock();
2053
2054 if (has_asym)
2055 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2056
2057 ret = 0;
2058 error:
2059 __free_domain_allocs(&d, alloc_state, cpu_map);
2060
2061 return ret;
2062 }
2063
2064 /* Current sched domains: */
2065 static cpumask_var_t *doms_cur;
2066
2067 /* Number of sched domains in 'doms_cur': */
2068 static int ndoms_cur;
2069
2070 /* Attribues of custom domains in 'doms_cur' */
2071 static struct sched_domain_attr *dattr_cur;
2072
2073 /*
2074 * Special case: If a kmalloc() of a doms_cur partition (array of
2075 * cpumask) fails, then fallback to a single sched domain,
2076 * as determined by the single cpumask fallback_doms.
2077 */
2078 static cpumask_var_t fallback_doms;
2079
2080 /*
2081 * arch_update_cpu_topology lets virtualized architectures update the
2082 * CPU core maps. It is supposed to return 1 if the topology changed
2083 * or 0 if it stayed the same.
2084 */
arch_update_cpu_topology(void)2085 int __weak arch_update_cpu_topology(void)
2086 {
2087 return 0;
2088 }
2089
alloc_sched_domains(unsigned int ndoms)2090 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2091 {
2092 int i;
2093 cpumask_var_t *doms;
2094
2095 doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2096 if (!doms)
2097 return NULL;
2098 for (i = 0; i < ndoms; i++) {
2099 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2100 free_sched_domains(doms, i);
2101 return NULL;
2102 }
2103 }
2104 return doms;
2105 }
2106
free_sched_domains(cpumask_var_t doms[],unsigned int ndoms)2107 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2108 {
2109 unsigned int i;
2110 for (i = 0; i < ndoms; i++)
2111 free_cpumask_var(doms[i]);
2112 kfree(doms);
2113 }
2114
2115 /*
2116 * Set up scheduler domains and groups. For now this just excludes isolated
2117 * CPUs, but could be used to exclude other special cases in the future.
2118 */
sched_init_domains(const struct cpumask * cpu_map)2119 int sched_init_domains(const struct cpumask *cpu_map)
2120 {
2121 int err;
2122
2123 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2124 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2125 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2126
2127 arch_update_cpu_topology();
2128 ndoms_cur = 1;
2129 doms_cur = alloc_sched_domains(ndoms_cur);
2130 if (!doms_cur)
2131 doms_cur = &fallback_doms;
2132 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
2133 err = build_sched_domains(doms_cur[0], NULL);
2134 register_sched_domain_sysctl();
2135
2136 return err;
2137 }
2138
2139 /*
2140 * Detach sched domains from a group of CPUs specified in cpu_map
2141 * These CPUs will now be attached to the NULL domain
2142 */
detach_destroy_domains(const struct cpumask * cpu_map)2143 static void detach_destroy_domains(const struct cpumask *cpu_map)
2144 {
2145 unsigned int cpu = cpumask_any(cpu_map);
2146 int i;
2147
2148 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2149 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2150
2151 rcu_read_lock();
2152 for_each_cpu(i, cpu_map)
2153 cpu_attach_domain(NULL, &def_root_domain, i);
2154 rcu_read_unlock();
2155 }
2156
2157 /* handle null as "default" */
dattrs_equal(struct sched_domain_attr * cur,int idx_cur,struct sched_domain_attr * new,int idx_new)2158 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2159 struct sched_domain_attr *new, int idx_new)
2160 {
2161 struct sched_domain_attr tmp;
2162
2163 /* Fast path: */
2164 if (!new && !cur)
2165 return 1;
2166
2167 tmp = SD_ATTR_INIT;
2168
2169 return !memcmp(cur ? (cur + idx_cur) : &tmp,
2170 new ? (new + idx_new) : &tmp,
2171 sizeof(struct sched_domain_attr));
2172 }
2173
2174 /*
2175 * Partition sched domains as specified by the 'ndoms_new'
2176 * cpumasks in the array doms_new[] of cpumasks. This compares
2177 * doms_new[] to the current sched domain partitioning, doms_cur[].
2178 * It destroys each deleted domain and builds each new domain.
2179 *
2180 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2181 * The masks don't intersect (don't overlap.) We should setup one
2182 * sched domain for each mask. CPUs not in any of the cpumasks will
2183 * not be load balanced. If the same cpumask appears both in the
2184 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2185 * it as it is.
2186 *
2187 * The passed in 'doms_new' should be allocated using
2188 * alloc_sched_domains. This routine takes ownership of it and will
2189 * free_sched_domains it when done with it. If the caller failed the
2190 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2191 * and partition_sched_domains() will fallback to the single partition
2192 * 'fallback_doms', it also forces the domains to be rebuilt.
2193 *
2194 * If doms_new == NULL it will be replaced with cpu_online_mask.
2195 * ndoms_new == 0 is a special case for destroying existing domains,
2196 * and it will not create the default domain.
2197 *
2198 * Call with hotplug lock and sched_domains_mutex held
2199 */
partition_sched_domains_locked(int ndoms_new,cpumask_var_t doms_new[],struct sched_domain_attr * dattr_new)2200 void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2201 struct sched_domain_attr *dattr_new)
2202 {
2203 bool __maybe_unused has_eas = false;
2204 int i, j, n;
2205 int new_topology;
2206
2207 lockdep_assert_held(&sched_domains_mutex);
2208
2209 /* Always unregister in case we don't destroy any domains: */
2210 unregister_sched_domain_sysctl();
2211
2212 /* Let the architecture update CPU core mappings: */
2213 new_topology = arch_update_cpu_topology();
2214
2215 if (!doms_new) {
2216 WARN_ON_ONCE(dattr_new);
2217 n = 0;
2218 doms_new = alloc_sched_domains(1);
2219 if (doms_new) {
2220 n = 1;
2221 cpumask_and(doms_new[0], cpu_active_mask,
2222 housekeeping_cpumask(HK_FLAG_DOMAIN));
2223 }
2224 } else {
2225 n = ndoms_new;
2226 }
2227
2228 /* Destroy deleted domains: */
2229 for (i = 0; i < ndoms_cur; i++) {
2230 for (j = 0; j < n && !new_topology; j++) {
2231 if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2232 dattrs_equal(dattr_cur, i, dattr_new, j)) {
2233 struct root_domain *rd;
2234
2235 /*
2236 * This domain won't be destroyed and as such
2237 * its dl_bw->total_bw needs to be cleared. It
2238 * will be recomputed in function
2239 * update_tasks_root_domain().
2240 */
2241 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2242 dl_clear_root_domain(rd);
2243 goto match1;
2244 }
2245 }
2246 /* No match - a current sched domain not in new doms_new[] */
2247 detach_destroy_domains(doms_cur[i]);
2248 match1:
2249 ;
2250 }
2251
2252 n = ndoms_cur;
2253 if (!doms_new) {
2254 n = 0;
2255 doms_new = &fallback_doms;
2256 cpumask_and(doms_new[0], cpu_active_mask,
2257 housekeeping_cpumask(HK_FLAG_DOMAIN));
2258 }
2259
2260 /* Build new domains: */
2261 for (i = 0; i < ndoms_new; i++) {
2262 for (j = 0; j < n && !new_topology; j++) {
2263 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2264 dattrs_equal(dattr_new, i, dattr_cur, j))
2265 goto match2;
2266 }
2267 /* No match - add a new doms_new */
2268 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2269 match2:
2270 ;
2271 }
2272
2273 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2274 /* Build perf. domains: */
2275 for (i = 0; i < ndoms_new; i++) {
2276 for (j = 0; j < n && !sched_energy_update; j++) {
2277 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2278 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2279 has_eas = true;
2280 goto match3;
2281 }
2282 }
2283 /* No match - add perf. domains for a new rd */
2284 has_eas |= build_perf_domains(doms_new[i]);
2285 match3:
2286 ;
2287 }
2288 sched_energy_set(has_eas);
2289 #endif
2290
2291 /* Remember the new sched domains: */
2292 if (doms_cur != &fallback_doms)
2293 free_sched_domains(doms_cur, ndoms_cur);
2294
2295 kfree(dattr_cur);
2296 doms_cur = doms_new;
2297 dattr_cur = dattr_new;
2298 ndoms_cur = ndoms_new;
2299
2300 register_sched_domain_sysctl();
2301 }
2302
2303 /*
2304 * Call with hotplug lock held
2305 */
partition_sched_domains(int ndoms_new,cpumask_var_t doms_new[],struct sched_domain_attr * dattr_new)2306 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2307 struct sched_domain_attr *dattr_new)
2308 {
2309 mutex_lock(&sched_domains_mutex);
2310 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2311 mutex_unlock(&sched_domains_mutex);
2312 }
2313