1 // SPDX-License-Identifier: GPL-2.0
3 * Scheduler topology setup/handling methods
7 DEFINE_MUTEX(sched_domains_mutex
);
9 /* Protected by sched_domains_mutex: */
10 static cpumask_var_t sched_domains_tmpmask
;
11 static cpumask_var_t sched_domains_tmpmask2
;
13 #ifdef CONFIG_SCHED_DEBUG
15 static int __init
sched_debug_setup(char *str
)
17 sched_debug_enabled
= true;
21 early_param("sched_debug", sched_debug_setup
);
23 static inline bool sched_debug(void)
25 return sched_debug_enabled
;
28 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
29 struct cpumask
*groupmask
)
31 struct sched_group
*group
= sd
->groups
;
33 cpumask_clear(groupmask
);
35 printk(KERN_DEBUG
"%*s domain-%d: ", level
, "", level
);
37 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
38 printk("does not load-balance\n");
40 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain has parent");
44 printk(KERN_CONT
"span=%*pbl level=%s\n",
45 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
47 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
48 printk(KERN_ERR
"ERROR: domain->span does not contain CPU%d\n", cpu
);
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
);
54 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
58 printk(KERN_ERR
"ERROR: group is NULL\n");
62 if (!cpumask_weight(sched_group_span(group
))) {
63 printk(KERN_CONT
"\n");
64 printk(KERN_ERR
"ERROR: empty group\n");
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");
75 cpumask_or(groupmask
, groupmask
, sched_group_span(group
));
77 printk(KERN_CONT
" %d:{ span=%*pbl",
79 cpumask_pr_args(sched_group_span(group
)));
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
)));
87 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
)
88 printk(KERN_CONT
" cap=%lu", group
->sgc
->capacity
);
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");
96 printk(KERN_CONT
" }");
100 if (group
!= sd
->groups
)
101 printk(KERN_CONT
",");
103 } while (group
!= sd
->groups
);
104 printk(KERN_CONT
"\n");
106 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
107 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
110 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
111 printk(KERN_ERR
"ERROR: parent span is not a superset of domain->span\n");
115 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
119 if (!sched_debug_enabled
)
123 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
127 printk(KERN_DEBUG
"CPU%d attaching sched-domain(s):\n", cpu
);
130 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
138 #else /* !CONFIG_SCHED_DEBUG */
140 # define sched_debug_enabled 0
141 # define sched_domain_debug(sd, cpu) do { } while (0)
142 static inline bool sched_debug(void)
146 #endif /* CONFIG_SCHED_DEBUG */
148 static int sd_degenerate(struct sched_domain
*sd
)
150 if (cpumask_weight(sched_domain_span(sd
)) == 1)
153 /* Following flags need at least 2 groups */
154 if (sd
->flags
& (SD_LOAD_BALANCE
|
158 SD_SHARE_CPUCAPACITY
|
159 SD_ASYM_CPUCAPACITY
|
160 SD_SHARE_PKG_RESOURCES
|
161 SD_SHARE_POWERDOMAIN
)) {
162 if (sd
->groups
!= sd
->groups
->next
)
166 /* Following flags don't use groups */
167 if (sd
->flags
& (SD_WAKE_AFFINE
))
174 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
176 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
178 if (sd_degenerate(parent
))
181 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
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
|
190 SD_ASYM_CPUCAPACITY
|
191 SD_SHARE_CPUCAPACITY
|
192 SD_SHARE_PKG_RESOURCES
|
194 SD_SHARE_POWERDOMAIN
);
195 if (nr_node_ids
== 1)
196 pflags
&= ~SD_SERIALIZE
;
198 if (~cflags
& pflags
)
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
;
210 #ifdef CONFIG_PROC_SYSCTL
211 int sched_energy_aware_handler(struct ctl_table
*table
, int write
,
212 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
216 if (write
&& !capable(CAP_SYS_ADMIN
))
219 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
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
);
235 static void free_pd(struct perf_domain
*pd
)
237 struct perf_domain
*tmp
;
246 static struct perf_domain
*find_pd(struct perf_domain
*pd
, int cpu
)
249 if (cpumask_test_cpu(cpu
, perf_domain_span(pd
)))
257 static struct perf_domain
*pd_init(int cpu
)
259 struct em_perf_domain
*obj
= em_cpu_get(cpu
);
260 struct perf_domain
*pd
;
264 pr_info("%s: no EM found for CPU%d\n", __func__
, cpu
);
268 pd
= kzalloc(sizeof(*pd
), GFP_KERNEL
);
276 static void perf_domain_debug(const struct cpumask
*cpu_map
,
277 struct perf_domain
*pd
)
279 if (!sched_debug() || !pd
)
282 printk(KERN_DEBUG
"root_domain %*pbl:", cpumask_pr_args(cpu_map
));
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
));
292 printk(KERN_CONT
"\n");
295 static void destroy_perf_domain_rcu(struct rcu_head
*rp
)
297 struct perf_domain
*pd
;
299 pd
= container_of(rp
, struct perf_domain
, rcu
);
303 static void sched_energy_set(bool has_eas
)
305 if (!has_eas
&& static_branch_unlikely(&sched_energy_present
)) {
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
)) {
311 pr_info("%s: starting EAS\n", __func__
);
312 static_branch_enable_cpuslocked(&sched_energy_present
);
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;
323 * The complexity of the Energy Model is defined as:
325 * C = nr_pd * (nr_cpus + nr_cs)
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).
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.
339 #define EM_MAX_COMPLEXITY 2048
341 extern struct cpufreq_governor schedutil_gov
;
342 static bool build_perf_domains(const struct cpumask
*cpu_map
)
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
;
351 if (!sysctl_sched_energy_aware
)
354 /* EAS is enabled for asymmetric CPU capacity topologies. */
355 if (!per_cpu(sd_asym_cpucapacity
, cpu
)) {
357 pr_info("rd %*pbl: CPUs do not have asymmetric capacities\n",
358 cpumask_pr_args(cpu_map
));
363 for_each_cpu(i
, cpu_map
) {
364 /* Skip already covered CPUs. */
368 /* Do not attempt EAS if schedutil is not being used. */
369 policy
= cpufreq_cpu_get(i
);
372 gov
= policy
->governor
;
373 cpufreq_cpu_put(policy
);
374 if (gov
!= &schedutil_gov
) {
376 pr_warn("rd %*pbl: Disabling EAS, schedutil is mandatory\n",
377 cpumask_pr_args(cpu_map
));
381 /* Create the new pd and add it to the local list. */
389 * Count performance domains and capacity states for the
393 nr_cs
+= em_pd_nr_cap_states(pd
->em_pd
);
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
));
403 perf_domain_debug(cpu_map
, pd
);
405 /* Attach the new list of performance domains to the root domain. */
407 rcu_assign_pointer(rd
->pd
, pd
);
409 call_rcu(&tmp
->rcu
, destroy_perf_domain_rcu
);
416 rcu_assign_pointer(rd
->pd
, NULL
);
418 call_rcu(&tmp
->rcu
, destroy_perf_domain_rcu
);
423 static void free_pd(struct perf_domain
*pd
) { }
424 #endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
426 static void free_rootdomain(struct rcu_head
*rcu
)
428 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
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
);
440 void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
442 struct root_domain
*old_rd
= NULL
;
445 raw_spin_lock_irqsave(&rq
->lock
, flags
);
450 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
453 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
456 * If we dont want to free the old_rd yet then
457 * set old_rd to NULL to skip the freeing later
460 if (!atomic_dec_and_test(&old_rd
->refcount
))
464 atomic_inc(&rd
->refcount
);
467 cpumask_set_cpu(rq
->cpu
, rd
->span
);
468 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
471 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
474 call_rcu(&old_rd
->rcu
, free_rootdomain
);
477 void sched_get_rd(struct root_domain
*rd
)
479 atomic_inc(&rd
->refcount
);
482 void sched_put_rd(struct root_domain
*rd
)
484 if (!atomic_dec_and_test(&rd
->refcount
))
487 call_rcu(&rd
->rcu
, free_rootdomain
);
490 static int init_rootdomain(struct root_domain
*rd
)
492 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
494 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
496 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
498 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
501 #ifdef HAVE_RT_PUSH_IPI
503 raw_spin_lock_init(&rd
->rto_lock
);
504 init_irq_work(&rd
->rto_push_work
, rto_push_irq_work_func
);
507 init_dl_bw(&rd
->dl_bw
);
508 if (cpudl_init(&rd
->cpudl
) != 0)
511 if (cpupri_init(&rd
->cpupri
) != 0)
516 cpudl_cleanup(&rd
->cpudl
);
518 free_cpumask_var(rd
->rto_mask
);
520 free_cpumask_var(rd
->dlo_mask
);
522 free_cpumask_var(rd
->online
);
524 free_cpumask_var(rd
->span
);
530 * By default the system creates a single root-domain with all CPUs as
531 * members (mimicking the global state we have today).
533 struct root_domain def_root_domain
;
535 void init_defrootdomain(void)
537 init_rootdomain(&def_root_domain
);
539 atomic_set(&def_root_domain
.refcount
, 1);
542 static struct root_domain
*alloc_rootdomain(void)
544 struct root_domain
*rd
;
546 rd
= kzalloc(sizeof(*rd
), GFP_KERNEL
);
550 if (init_rootdomain(rd
) != 0) {
558 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
560 struct sched_group
*tmp
, *first
;
569 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
572 if (atomic_dec_and_test(&sg
->ref
))
575 } while (sg
!= first
);
578 static void destroy_sched_domain(struct sched_domain
*sd
)
581 * A normal sched domain may have multiple group references, an
582 * overlapping domain, having private groups, only one. Iterate,
583 * dropping group/capacity references, freeing where none remain.
585 free_sched_groups(sd
->groups
, 1);
587 if (sd
->shared
&& atomic_dec_and_test(&sd
->shared
->ref
))
592 static void destroy_sched_domains_rcu(struct rcu_head
*rcu
)
594 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
597 struct sched_domain
*parent
= sd
->parent
;
598 destroy_sched_domain(sd
);
603 static void destroy_sched_domains(struct sched_domain
*sd
)
606 call_rcu(&sd
->rcu
, destroy_sched_domains_rcu
);
610 * Keep a special pointer to the highest sched_domain that has
611 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
612 * allows us to avoid some pointer chasing select_idle_sibling().
614 * Also keep a unique ID per domain (we use the first CPU number in
615 * the cpumask of the domain), this allows us to quickly tell if
616 * two CPUs are in the same cache domain, see cpus_share_cache().
618 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
619 DEFINE_PER_CPU(int, sd_llc_size
);
620 DEFINE_PER_CPU(int, sd_llc_id
);
621 DEFINE_PER_CPU(struct sched_domain_shared
*, sd_llc_shared
);
622 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
623 DEFINE_PER_CPU(struct sched_domain
*, sd_asym_packing
);
624 DEFINE_PER_CPU(struct sched_domain
*, sd_asym_cpucapacity
);
625 DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity
);
627 static void update_top_cache_domain(int cpu
)
629 struct sched_domain_shared
*sds
= NULL
;
630 struct sched_domain
*sd
;
634 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
636 id
= cpumask_first(sched_domain_span(sd
));
637 size
= cpumask_weight(sched_domain_span(sd
));
641 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
642 per_cpu(sd_llc_size
, cpu
) = size
;
643 per_cpu(sd_llc_id
, cpu
) = id
;
644 rcu_assign_pointer(per_cpu(sd_llc_shared
, cpu
), sds
);
646 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
647 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
649 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
650 rcu_assign_pointer(per_cpu(sd_asym_packing
, cpu
), sd
);
652 sd
= lowest_flag_domain(cpu
, SD_ASYM_CPUCAPACITY
);
653 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity
, cpu
), sd
);
657 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
658 * hold the hotplug lock.
661 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
663 struct rq
*rq
= cpu_rq(cpu
);
664 struct sched_domain
*tmp
;
666 /* Remove the sched domains which do not contribute to scheduling. */
667 for (tmp
= sd
; tmp
; ) {
668 struct sched_domain
*parent
= tmp
->parent
;
672 if (sd_parent_degenerate(tmp
, parent
)) {
673 tmp
->parent
= parent
->parent
;
675 parent
->parent
->child
= tmp
;
677 * Transfer SD_PREFER_SIBLING down in case of a
678 * degenerate parent; the spans match for this
679 * so the property transfers.
681 if (parent
->flags
& SD_PREFER_SIBLING
)
682 tmp
->flags
|= SD_PREFER_SIBLING
;
683 destroy_sched_domain(parent
);
688 if (sd
&& sd_degenerate(sd
)) {
691 destroy_sched_domain(tmp
);
696 sched_domain_debug(sd
, cpu
);
698 rq_attach_root(rq
, rd
);
700 rcu_assign_pointer(rq
->sd
, sd
);
701 dirty_sched_domain_sysctl(cpu
);
702 destroy_sched_domains(tmp
);
704 update_top_cache_domain(cpu
);
708 struct sched_domain
* __percpu
*sd
;
709 struct root_domain
*rd
;
720 * Return the canonical balance CPU for this group, this is the first CPU
721 * of this group that's also in the balance mask.
723 * The balance mask are all those CPUs that could actually end up at this
724 * group. See build_balance_mask().
726 * Also see should_we_balance().
728 int group_balance_cpu(struct sched_group
*sg
)
730 return cpumask_first(group_balance_mask(sg
));
735 * NUMA topology (first read the regular topology blurb below)
737 * Given a node-distance table, for example:
745 * which represents a 4 node ring topology like:
753 * We want to construct domains and groups to represent this. The way we go
754 * about doing this is to build the domains on 'hops'. For each NUMA level we
755 * construct the mask of all nodes reachable in @level hops.
757 * For the above NUMA topology that gives 3 levels:
759 * NUMA-2 0-3 0-3 0-3 0-3
760 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
762 * NUMA-1 0-1,3 0-2 1-3 0,2-3
763 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
768 * As can be seen; things don't nicely line up as with the regular topology.
769 * When we iterate a domain in child domain chunks some nodes can be
770 * represented multiple times -- hence the "overlap" naming for this part of
773 * In order to minimize this overlap, we only build enough groups to cover the
774 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
778 * - the first group of each domain is its child domain; this
779 * gets us the first 0-1,3
780 * - the only uncovered node is 2, who's child domain is 1-3.
782 * However, because of the overlap, computing a unique CPU for each group is
783 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
784 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
785 * end up at those groups (they would end up in group: 0-1,3).
787 * To correct this we have to introduce the group balance mask. This mask
788 * will contain those CPUs in the group that can reach this group given the
789 * (child) domain tree.
791 * With this we can once again compute balance_cpu and sched_group_capacity
794 * XXX include words on how balance_cpu is unique and therefore can be
795 * used for sched_group_capacity links.
798 * Another 'interesting' topology is:
806 * Which looks a little like:
814 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
817 * This leads to a few particularly weird cases where the sched_domain's are
818 * not of the same number for each CPU. Consider:
821 * groups: {0-2},{1-3} {1-3},{0-2}
823 * NUMA-1 0-2 0-3 0-3 1-3
831 * Build the balance mask; it contains only those CPUs that can arrive at this
832 * group and should be considered to continue balancing.
834 * We do this during the group creation pass, therefore the group information
835 * isn't complete yet, however since each group represents a (child) domain we
836 * can fully construct this using the sched_domain bits (which are already
840 build_balance_mask(struct sched_domain
*sd
, struct sched_group
*sg
, struct cpumask
*mask
)
842 const struct cpumask
*sg_span
= sched_group_span(sg
);
843 struct sd_data
*sdd
= sd
->private;
844 struct sched_domain
*sibling
;
849 for_each_cpu(i
, sg_span
) {
850 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
853 * Can happen in the asymmetric case, where these siblings are
854 * unused. The mask will not be empty because those CPUs that
855 * do have the top domain _should_ span the domain.
860 /* If we would not end up here, we can't continue from here */
861 if (!cpumask_equal(sg_span
, sched_domain_span(sibling
->child
)))
864 cpumask_set_cpu(i
, mask
);
867 /* We must not have empty masks here */
868 WARN_ON_ONCE(cpumask_empty(mask
));
872 * XXX: This creates per-node group entries; since the load-balancer will
873 * immediately access remote memory to construct this group's load-balance
874 * statistics having the groups node local is of dubious benefit.
876 static struct sched_group
*
877 build_group_from_child_sched_domain(struct sched_domain
*sd
, int cpu
)
879 struct sched_group
*sg
;
880 struct cpumask
*sg_span
;
882 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
883 GFP_KERNEL
, cpu_to_node(cpu
));
888 sg_span
= sched_group_span(sg
);
890 cpumask_copy(sg_span
, sched_domain_span(sd
->child
));
892 cpumask_copy(sg_span
, sched_domain_span(sd
));
894 atomic_inc(&sg
->ref
);
898 static void init_overlap_sched_group(struct sched_domain
*sd
,
899 struct sched_group
*sg
)
901 struct cpumask
*mask
= sched_domains_tmpmask2
;
902 struct sd_data
*sdd
= sd
->private;
903 struct cpumask
*sg_span
;
906 build_balance_mask(sd
, sg
, mask
);
907 cpu
= cpumask_first_and(sched_group_span(sg
), mask
);
909 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
910 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
911 cpumask_copy(group_balance_mask(sg
), mask
);
913 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg
), mask
));
916 * Initialize sgc->capacity such that even if we mess up the
917 * domains and no possible iteration will get us here, we won't
920 sg_span
= sched_group_span(sg
);
921 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
922 sg
->sgc
->min_capacity
= SCHED_CAPACITY_SCALE
;
923 sg
->sgc
->max_capacity
= SCHED_CAPACITY_SCALE
;
927 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
929 struct sched_group
*first
= NULL
, *last
= NULL
, *sg
;
930 const struct cpumask
*span
= sched_domain_span(sd
);
931 struct cpumask
*covered
= sched_domains_tmpmask
;
932 struct sd_data
*sdd
= sd
->private;
933 struct sched_domain
*sibling
;
936 cpumask_clear(covered
);
938 for_each_cpu_wrap(i
, span
, cpu
) {
939 struct cpumask
*sg_span
;
941 if (cpumask_test_cpu(i
, covered
))
944 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
947 * Asymmetric node setups can result in situations where the
948 * domain tree is of unequal depth, make sure to skip domains
949 * that already cover the entire range.
951 * In that case build_sched_domains() will have terminated the
952 * iteration early and our sibling sd spans will be empty.
953 * Domains should always include the CPU they're built on, so
956 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
959 sg
= build_group_from_child_sched_domain(sibling
, cpu
);
963 sg_span
= sched_group_span(sg
);
964 cpumask_or(covered
, covered
, sg_span
);
966 init_overlap_sched_group(sd
, sg
);
980 free_sched_groups(first
, 0);
987 * Package topology (also see the load-balance blurb in fair.c)
989 * The scheduler builds a tree structure to represent a number of important
990 * topology features. By default (default_topology[]) these include:
992 * - Simultaneous multithreading (SMT)
993 * - Multi-Core Cache (MC)
996 * Where the last one more or less denotes everything up to a NUMA node.
998 * The tree consists of 3 primary data structures:
1000 * sched_domain -> sched_group -> sched_group_capacity
1004 * The sched_domains are per-CPU and have a two way link (parent & child) and
1005 * denote the ever growing mask of CPUs belonging to that level of topology.
1007 * Each sched_domain has a circular (double) linked list of sched_group's, each
1008 * denoting the domains of the level below (or individual CPUs in case of the
1009 * first domain level). The sched_group linked by a sched_domain includes the
1010 * CPU of that sched_domain [*].
1012 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1014 * CPU 0 1 2 3 4 5 6 7
1018 * SMT [ ] [ ] [ ] [ ]
1022 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1023 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1024 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1026 * CPU 0 1 2 3 4 5 6 7
1028 * One way to think about it is: sched_domain moves you up and down among these
1029 * topology levels, while sched_group moves you sideways through it, at child
1030 * domain granularity.
1032 * sched_group_capacity ensures each unique sched_group has shared storage.
1034 * There are two related construction problems, both require a CPU that
1035 * uniquely identify each group (for a given domain):
1037 * - The first is the balance_cpu (see should_we_balance() and the
1038 * load-balance blub in fair.c); for each group we only want 1 CPU to
1039 * continue balancing at a higher domain.
1041 * - The second is the sched_group_capacity; we want all identical groups
1042 * to share a single sched_group_capacity.
1044 * Since these topologies are exclusive by construction. That is, its
1045 * impossible for an SMT thread to belong to multiple cores, and cores to
1046 * be part of multiple caches. There is a very clear and unique location
1047 * for each CPU in the hierarchy.
1049 * Therefore computing a unique CPU for each group is trivial (the iteration
1050 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1051 * group), we can simply pick the first CPU in each group.
1054 * [*] in other words, the first group of each domain is its child domain.
1057 static struct sched_group
*get_group(int cpu
, struct sd_data
*sdd
)
1059 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
1060 struct sched_domain
*child
= sd
->child
;
1061 struct sched_group
*sg
;
1064 cpu
= cpumask_first(sched_domain_span(child
));
1066 sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
1067 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
1069 /* For claim_allocations: */
1070 atomic_inc(&sg
->ref
);
1071 atomic_inc(&sg
->sgc
->ref
);
1074 cpumask_copy(sched_group_span(sg
), sched_domain_span(child
));
1075 cpumask_copy(group_balance_mask(sg
), sched_group_span(sg
));
1077 cpumask_set_cpu(cpu
, sched_group_span(sg
));
1078 cpumask_set_cpu(cpu
, group_balance_mask(sg
));
1081 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sched_group_span(sg
));
1082 sg
->sgc
->min_capacity
= SCHED_CAPACITY_SCALE
;
1083 sg
->sgc
->max_capacity
= SCHED_CAPACITY_SCALE
;
1089 * build_sched_groups will build a circular linked list of the groups
1090 * covered by the given span, and will set each group's ->cpumask correctly,
1091 * and ->cpu_capacity to 0.
1093 * Assumes the sched_domain tree is fully constructed
1096 build_sched_groups(struct sched_domain
*sd
, int cpu
)
1098 struct sched_group
*first
= NULL
, *last
= NULL
;
1099 struct sd_data
*sdd
= sd
->private;
1100 const struct cpumask
*span
= sched_domain_span(sd
);
1101 struct cpumask
*covered
;
1104 lockdep_assert_held(&sched_domains_mutex
);
1105 covered
= sched_domains_tmpmask
;
1107 cpumask_clear(covered
);
1109 for_each_cpu_wrap(i
, span
, cpu
) {
1110 struct sched_group
*sg
;
1112 if (cpumask_test_cpu(i
, covered
))
1115 sg
= get_group(i
, sdd
);
1117 cpumask_or(covered
, covered
, sched_group_span(sg
));
1132 * Initialize sched groups cpu_capacity.
1134 * cpu_capacity indicates the capacity of sched group, which is used while
1135 * distributing the load between different sched groups in a sched domain.
1136 * Typically cpu_capacity for all the groups in a sched domain will be same
1137 * unless there are asymmetries in the topology. If there are asymmetries,
1138 * group having more cpu_capacity will pickup more load compared to the
1139 * group having less cpu_capacity.
1141 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
1143 struct sched_group
*sg
= sd
->groups
;
1148 int cpu
, max_cpu
= -1;
1150 sg
->group_weight
= cpumask_weight(sched_group_span(sg
));
1152 if (!(sd
->flags
& SD_ASYM_PACKING
))
1155 for_each_cpu(cpu
, sched_group_span(sg
)) {
1158 else if (sched_asym_prefer(cpu
, max_cpu
))
1161 sg
->asym_prefer_cpu
= max_cpu
;
1165 } while (sg
!= sd
->groups
);
1167 if (cpu
!= group_balance_cpu(sg
))
1170 update_group_capacity(sd
, cpu
);
1174 * Initializers for schedule domains
1175 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1178 static int default_relax_domain_level
= -1;
1179 int sched_domain_level_max
;
1181 static int __init
setup_relax_domain_level(char *str
)
1183 if (kstrtoint(str
, 0, &default_relax_domain_level
))
1184 pr_warn("Unable to set relax_domain_level\n");
1188 __setup("relax_domain_level=", setup_relax_domain_level
);
1190 static void set_domain_attribute(struct sched_domain
*sd
,
1191 struct sched_domain_attr
*attr
)
1195 if (!attr
|| attr
->relax_domain_level
< 0) {
1196 if (default_relax_domain_level
< 0)
1199 request
= default_relax_domain_level
;
1201 request
= attr
->relax_domain_level
;
1202 if (request
< sd
->level
) {
1203 /* Turn off idle balance on this domain: */
1204 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
1206 /* Turn on idle balance on this domain: */
1207 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
1211 static void __sdt_free(const struct cpumask
*cpu_map
);
1212 static int __sdt_alloc(const struct cpumask
*cpu_map
);
1214 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
1215 const struct cpumask
*cpu_map
)
1219 if (!atomic_read(&d
->rd
->refcount
))
1220 free_rootdomain(&d
->rd
->rcu
);
1226 __sdt_free(cpu_map
);
1234 __visit_domain_allocation_hell(struct s_data
*d
, const struct cpumask
*cpu_map
)
1236 memset(d
, 0, sizeof(*d
));
1238 if (__sdt_alloc(cpu_map
))
1239 return sa_sd_storage
;
1240 d
->sd
= alloc_percpu(struct sched_domain
*);
1242 return sa_sd_storage
;
1243 d
->rd
= alloc_rootdomain();
1247 return sa_rootdomain
;
1251 * NULL the sd_data elements we've used to build the sched_domain and
1252 * sched_group structure so that the subsequent __free_domain_allocs()
1253 * will not free the data we're using.
1255 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
1257 struct sd_data
*sdd
= sd
->private;
1259 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
1260 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
1262 if (atomic_read(&(*per_cpu_ptr(sdd
->sds
, cpu
))->ref
))
1263 *per_cpu_ptr(sdd
->sds
, cpu
) = NULL
;
1265 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
1266 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
1268 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
1269 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
1273 enum numa_topology_type sched_numa_topology_type
;
1275 static int sched_domains_numa_levels
;
1276 static int sched_domains_curr_level
;
1278 int sched_max_numa_distance
;
1279 static int *sched_domains_numa_distance
;
1280 static struct cpumask
***sched_domains_numa_masks
;
1284 * SD_flags allowed in topology descriptions.
1286 * These flags are purely descriptive of the topology and do not prescribe
1287 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1290 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1291 * SD_SHARE_PKG_RESOURCES - describes shared caches
1292 * SD_NUMA - describes NUMA topologies
1293 * SD_SHARE_POWERDOMAIN - describes shared power domain
1295 * Odd one out, which beside describing the topology has a quirk also
1296 * prescribes the desired behaviour that goes along with it:
1298 * SD_ASYM_PACKING - describes SMT quirks
1300 #define TOPOLOGY_SD_FLAGS \
1301 (SD_SHARE_CPUCAPACITY | \
1302 SD_SHARE_PKG_RESOURCES | \
1305 SD_SHARE_POWERDOMAIN)
1307 static struct sched_domain
*
1308 sd_init(struct sched_domain_topology_level
*tl
,
1309 const struct cpumask
*cpu_map
,
1310 struct sched_domain
*child
, int dflags
, int cpu
)
1312 struct sd_data
*sdd
= &tl
->data
;
1313 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
1314 int sd_id
, sd_weight
, sd_flags
= 0;
1318 * Ugly hack to pass state to sd_numa_mask()...
1320 sched_domains_curr_level
= tl
->numa_level
;
1323 sd_weight
= cpumask_weight(tl
->mask(cpu
));
1326 sd_flags
= (*tl
->sd_flags
)();
1327 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
1328 "wrong sd_flags in topology description\n"))
1329 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
1331 /* Apply detected topology flags */
1334 *sd
= (struct sched_domain
){
1335 .min_interval
= sd_weight
,
1336 .max_interval
= 2*sd_weight
,
1338 .imbalance_pct
= 125,
1340 .cache_nice_tries
= 0,
1347 .flags
= 1*SD_LOAD_BALANCE
1348 | 1*SD_BALANCE_NEWIDLE
1353 | 0*SD_SHARE_CPUCAPACITY
1354 | 0*SD_SHARE_PKG_RESOURCES
1356 | 1*SD_PREFER_SIBLING
1361 .last_balance
= jiffies
,
1362 .balance_interval
= sd_weight
,
1363 .max_newidle_lb_cost
= 0,
1364 .next_decay_max_lb_cost
= jiffies
,
1366 #ifdef CONFIG_SCHED_DEBUG
1371 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
1372 sd_id
= cpumask_first(sched_domain_span(sd
));
1375 * Convert topological properties into behaviour.
1378 if (sd
->flags
& SD_ASYM_CPUCAPACITY
) {
1379 struct sched_domain
*t
= sd
;
1382 * Don't attempt to spread across CPUs of different capacities.
1385 sd
->child
->flags
&= ~SD_PREFER_SIBLING
;
1387 for_each_lower_domain(t
)
1388 t
->flags
|= SD_BALANCE_WAKE
;
1391 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
1392 sd
->imbalance_pct
= 110;
1394 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
1395 sd
->imbalance_pct
= 117;
1396 sd
->cache_nice_tries
= 1;
1400 } else if (sd
->flags
& SD_NUMA
) {
1401 sd
->cache_nice_tries
= 2;
1405 sd
->flags
&= ~SD_PREFER_SIBLING
;
1406 sd
->flags
|= SD_SERIALIZE
;
1407 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
1408 sd
->flags
&= ~(SD_BALANCE_EXEC
|
1415 sd
->cache_nice_tries
= 1;
1421 * For all levels sharing cache; connect a sched_domain_shared
1424 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
1425 sd
->shared
= *per_cpu_ptr(sdd
->sds
, sd_id
);
1426 atomic_inc(&sd
->shared
->ref
);
1427 atomic_set(&sd
->shared
->nr_busy_cpus
, sd_weight
);
1436 * Topology list, bottom-up.
1438 static struct sched_domain_topology_level default_topology
[] = {
1439 #ifdef CONFIG_SCHED_SMT
1440 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
1442 #ifdef CONFIG_SCHED_MC
1443 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
1445 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
1449 static struct sched_domain_topology_level
*sched_domain_topology
=
1452 #define for_each_sd_topology(tl) \
1453 for (tl = sched_domain_topology; tl->mask; tl++)
1455 void set_sched_topology(struct sched_domain_topology_level
*tl
)
1457 if (WARN_ON_ONCE(sched_smp_initialized
))
1460 sched_domain_topology
= tl
;
1465 static const struct cpumask
*sd_numa_mask(int cpu
)
1467 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
1470 static void sched_numa_warn(const char *str
)
1472 static int done
= false;
1480 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
1482 for (i
= 0; i
< nr_node_ids
; i
++) {
1483 printk(KERN_WARNING
" ");
1484 for (j
= 0; j
< nr_node_ids
; j
++)
1485 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
1486 printk(KERN_CONT
"\n");
1488 printk(KERN_WARNING
"\n");
1491 bool find_numa_distance(int distance
)
1495 if (distance
== node_distance(0, 0))
1498 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
1499 if (sched_domains_numa_distance
[i
] == distance
)
1507 * A system can have three types of NUMA topology:
1508 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1509 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1510 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1512 * The difference between a glueless mesh topology and a backplane
1513 * topology lies in whether communication between not directly
1514 * connected nodes goes through intermediary nodes (where programs
1515 * could run), or through backplane controllers. This affects
1516 * placement of programs.
1518 * The type of topology can be discerned with the following tests:
1519 * - If the maximum distance between any nodes is 1 hop, the system
1520 * is directly connected.
1521 * - If for two nodes A and B, located N > 1 hops away from each other,
1522 * there is an intermediary node C, which is < N hops away from both
1523 * nodes A and B, the system is a glueless mesh.
1525 static void init_numa_topology_type(void)
1529 n
= sched_max_numa_distance
;
1531 if (sched_domains_numa_levels
<= 2) {
1532 sched_numa_topology_type
= NUMA_DIRECT
;
1536 for_each_online_node(a
) {
1537 for_each_online_node(b
) {
1538 /* Find two nodes furthest removed from each other. */
1539 if (node_distance(a
, b
) < n
)
1542 /* Is there an intermediary node between a and b? */
1543 for_each_online_node(c
) {
1544 if (node_distance(a
, c
) < n
&&
1545 node_distance(b
, c
) < n
) {
1546 sched_numa_topology_type
=
1552 sched_numa_topology_type
= NUMA_BACKPLANE
;
1558 void sched_init_numa(void)
1560 int next_distance
, curr_distance
= node_distance(0, 0);
1561 struct sched_domain_topology_level
*tl
;
1565 sched_domains_numa_distance
= kzalloc(sizeof(int) * (nr_node_ids
+ 1), GFP_KERNEL
);
1566 if (!sched_domains_numa_distance
)
1569 /* Includes NUMA identity node at level 0. */
1570 sched_domains_numa_distance
[level
++] = curr_distance
;
1571 sched_domains_numa_levels
= level
;
1574 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1575 * unique distances in the node_distance() table.
1577 * Assumes node_distance(0,j) includes all distances in
1578 * node_distance(i,j) in order to avoid cubic time.
1580 next_distance
= curr_distance
;
1581 for (i
= 0; i
< nr_node_ids
; i
++) {
1582 for (j
= 0; j
< nr_node_ids
; j
++) {
1583 for (k
= 0; k
< nr_node_ids
; k
++) {
1584 int distance
= node_distance(i
, k
);
1586 if (distance
> curr_distance
&&
1587 (distance
< next_distance
||
1588 next_distance
== curr_distance
))
1589 next_distance
= distance
;
1592 * While not a strong assumption it would be nice to know
1593 * about cases where if node A is connected to B, B is not
1594 * equally connected to A.
1596 if (sched_debug() && node_distance(k
, i
) != distance
)
1597 sched_numa_warn("Node-distance not symmetric");
1599 if (sched_debug() && i
&& !find_numa_distance(distance
))
1600 sched_numa_warn("Node-0 not representative");
1602 if (next_distance
!= curr_distance
) {
1603 sched_domains_numa_distance
[level
++] = next_distance
;
1604 sched_domains_numa_levels
= level
;
1605 curr_distance
= next_distance
;
1610 * In case of sched_debug() we verify the above assumption.
1617 * 'level' contains the number of unique distances
1619 * The sched_domains_numa_distance[] array includes the actual distance
1624 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1625 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1626 * the array will contain less then 'level' members. This could be
1627 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1628 * in other functions.
1630 * We reset it to 'level' at the end of this function.
1632 sched_domains_numa_levels
= 0;
1634 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
1635 if (!sched_domains_numa_masks
)
1639 * Now for each level, construct a mask per node which contains all
1640 * CPUs of nodes that are that many hops away from us.
1642 for (i
= 0; i
< level
; i
++) {
1643 sched_domains_numa_masks
[i
] =
1644 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
1645 if (!sched_domains_numa_masks
[i
])
1648 for (j
= 0; j
< nr_node_ids
; j
++) {
1649 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
1653 sched_domains_numa_masks
[i
][j
] = mask
;
1656 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
1659 cpumask_or(mask
, mask
, cpumask_of_node(k
));
1664 /* Compute default topology size */
1665 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
1667 tl
= kzalloc((i
+ level
+ 1) *
1668 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
1673 * Copy the default topology bits..
1675 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
1676 tl
[i
] = sched_domain_topology
[i
];
1679 * Add the NUMA identity distance, aka single NODE.
1681 tl
[i
++] = (struct sched_domain_topology_level
){
1682 .mask
= sd_numa_mask
,
1688 * .. and append 'j' levels of NUMA goodness.
1690 for (j
= 1; j
< level
; i
++, j
++) {
1691 tl
[i
] = (struct sched_domain_topology_level
){
1692 .mask
= sd_numa_mask
,
1693 .sd_flags
= cpu_numa_flags
,
1694 .flags
= SDTL_OVERLAP
,
1700 sched_domain_topology
= tl
;
1702 sched_domains_numa_levels
= level
;
1703 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
1705 init_numa_topology_type();
1708 void sched_domains_numa_masks_set(unsigned int cpu
)
1710 int node
= cpu_to_node(cpu
);
1713 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
1714 for (j
= 0; j
< nr_node_ids
; j
++) {
1715 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
1716 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
1721 void sched_domains_numa_masks_clear(unsigned int cpu
)
1725 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
1726 for (j
= 0; j
< nr_node_ids
; j
++)
1727 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
1731 #endif /* CONFIG_NUMA */
1733 static int __sdt_alloc(const struct cpumask
*cpu_map
)
1735 struct sched_domain_topology_level
*tl
;
1738 for_each_sd_topology(tl
) {
1739 struct sd_data
*sdd
= &tl
->data
;
1741 sdd
->sd
= alloc_percpu(struct sched_domain
*);
1745 sdd
->sds
= alloc_percpu(struct sched_domain_shared
*);
1749 sdd
->sg
= alloc_percpu(struct sched_group
*);
1753 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
1757 for_each_cpu(j
, cpu_map
) {
1758 struct sched_domain
*sd
;
1759 struct sched_domain_shared
*sds
;
1760 struct sched_group
*sg
;
1761 struct sched_group_capacity
*sgc
;
1763 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
1764 GFP_KERNEL
, cpu_to_node(j
));
1768 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
1770 sds
= kzalloc_node(sizeof(struct sched_domain_shared
),
1771 GFP_KERNEL
, cpu_to_node(j
));
1775 *per_cpu_ptr(sdd
->sds
, j
) = sds
;
1777 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
1778 GFP_KERNEL
, cpu_to_node(j
));
1784 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
1786 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
1787 GFP_KERNEL
, cpu_to_node(j
));
1791 #ifdef CONFIG_SCHED_DEBUG
1795 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
1802 static void __sdt_free(const struct cpumask
*cpu_map
)
1804 struct sched_domain_topology_level
*tl
;
1807 for_each_sd_topology(tl
) {
1808 struct sd_data
*sdd
= &tl
->data
;
1810 for_each_cpu(j
, cpu_map
) {
1811 struct sched_domain
*sd
;
1814 sd
= *per_cpu_ptr(sdd
->sd
, j
);
1815 if (sd
&& (sd
->flags
& SD_OVERLAP
))
1816 free_sched_groups(sd
->groups
, 0);
1817 kfree(*per_cpu_ptr(sdd
->sd
, j
));
1821 kfree(*per_cpu_ptr(sdd
->sds
, j
));
1823 kfree(*per_cpu_ptr(sdd
->sg
, j
));
1825 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
1827 free_percpu(sdd
->sd
);
1829 free_percpu(sdd
->sds
);
1831 free_percpu(sdd
->sg
);
1833 free_percpu(sdd
->sgc
);
1838 static struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
1839 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
1840 struct sched_domain
*child
, int dflags
, int cpu
)
1842 struct sched_domain
*sd
= sd_init(tl
, cpu_map
, child
, dflags
, cpu
);
1845 sd
->level
= child
->level
+ 1;
1846 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
1849 if (!cpumask_subset(sched_domain_span(child
),
1850 sched_domain_span(sd
))) {
1851 pr_err("BUG: arch topology borken\n");
1852 #ifdef CONFIG_SCHED_DEBUG
1853 pr_err(" the %s domain not a subset of the %s domain\n",
1854 child
->name
, sd
->name
);
1856 /* Fixup, ensure @sd has at least @child CPUs. */
1857 cpumask_or(sched_domain_span(sd
),
1858 sched_domain_span(sd
),
1859 sched_domain_span(child
));
1863 set_domain_attribute(sd
, attr
);
1869 * Find the sched_domain_topology_level where all CPU capacities are visible
1872 static struct sched_domain_topology_level
1873 *asym_cpu_capacity_level(const struct cpumask
*cpu_map
)
1875 int i
, j
, asym_level
= 0;
1877 struct sched_domain_topology_level
*tl
, *asym_tl
= NULL
;
1880 /* Is there any asymmetry? */
1881 cap
= arch_scale_cpu_capacity(NULL
, cpumask_first(cpu_map
));
1883 for_each_cpu(i
, cpu_map
) {
1884 if (arch_scale_cpu_capacity(NULL
, i
) != cap
) {
1894 * Examine topology from all CPU's point of views to detect the lowest
1895 * sched_domain_topology_level where a highest capacity CPU is visible
1898 for_each_cpu(i
, cpu_map
) {
1899 unsigned long max_capacity
= arch_scale_cpu_capacity(NULL
, i
);
1902 for_each_sd_topology(tl
) {
1903 if (tl_id
< asym_level
)
1906 for_each_cpu_and(j
, tl
->mask(i
), cpu_map
) {
1907 unsigned long capacity
;
1909 capacity
= arch_scale_cpu_capacity(NULL
, j
);
1911 if (capacity
<= max_capacity
)
1914 max_capacity
= capacity
;
1928 * Build sched domains for a given set of CPUs and attach the sched domains
1929 * to the individual CPUs
1932 build_sched_domains(const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
)
1934 enum s_alloc alloc_state
;
1935 struct sched_domain
*sd
;
1937 struct rq
*rq
= NULL
;
1938 int i
, ret
= -ENOMEM
;
1939 struct sched_domain_topology_level
*tl_asym
;
1940 bool has_asym
= false;
1942 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
1943 if (alloc_state
!= sa_rootdomain
)
1946 tl_asym
= asym_cpu_capacity_level(cpu_map
);
1948 /* Set up domains for CPUs specified by the cpu_map: */
1949 for_each_cpu(i
, cpu_map
) {
1950 struct sched_domain_topology_level
*tl
;
1953 for_each_sd_topology(tl
) {
1956 if (tl
== tl_asym
) {
1957 dflags
|= SD_ASYM_CPUCAPACITY
;
1961 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, dflags
, i
);
1963 if (tl
== sched_domain_topology
)
1964 *per_cpu_ptr(d
.sd
, i
) = sd
;
1965 if (tl
->flags
& SDTL_OVERLAP
)
1966 sd
->flags
|= SD_OVERLAP
;
1967 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
1972 /* Build the groups for the domains */
1973 for_each_cpu(i
, cpu_map
) {
1974 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
1975 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
1976 if (sd
->flags
& SD_OVERLAP
) {
1977 if (build_overlap_sched_groups(sd
, i
))
1980 if (build_sched_groups(sd
, i
))
1986 /* Calculate CPU capacity for physical packages and nodes */
1987 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
1988 if (!cpumask_test_cpu(i
, cpu_map
))
1991 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
1992 claim_allocations(i
, sd
);
1993 init_sched_groups_capacity(i
, sd
);
1997 /* Attach the domains */
1999 for_each_cpu(i
, cpu_map
) {
2001 sd
= *per_cpu_ptr(d
.sd
, i
);
2003 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
2004 if (rq
->cpu_capacity_orig
> READ_ONCE(d
.rd
->max_cpu_capacity
))
2005 WRITE_ONCE(d
.rd
->max_cpu_capacity
, rq
->cpu_capacity_orig
);
2007 cpu_attach_domain(sd
, d
.rd
, i
);
2012 static_branch_enable_cpuslocked(&sched_asym_cpucapacity
);
2014 if (rq
&& sched_debug_enabled
) {
2015 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
2016 cpumask_pr_args(cpu_map
), rq
->rd
->max_cpu_capacity
);
2021 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
2026 /* Current sched domains: */
2027 static cpumask_var_t
*doms_cur
;
2029 /* Number of sched domains in 'doms_cur': */
2030 static int ndoms_cur
;
2032 /* Attribues of custom domains in 'doms_cur' */
2033 static struct sched_domain_attr
*dattr_cur
;
2036 * Special case: If a kmalloc() of a doms_cur partition (array of
2037 * cpumask) fails, then fallback to a single sched domain,
2038 * as determined by the single cpumask fallback_doms.
2040 static cpumask_var_t fallback_doms
;
2043 * arch_update_cpu_topology lets virtualized architectures update the
2044 * CPU core maps. It is supposed to return 1 if the topology changed
2045 * or 0 if it stayed the same.
2047 int __weak
arch_update_cpu_topology(void)
2052 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
2055 cpumask_var_t
*doms
;
2057 doms
= kmalloc_array(ndoms
, sizeof(*doms
), GFP_KERNEL
);
2060 for (i
= 0; i
< ndoms
; i
++) {
2061 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
2062 free_sched_domains(doms
, i
);
2069 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
2072 for (i
= 0; i
< ndoms
; i
++)
2073 free_cpumask_var(doms
[i
]);
2078 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
2079 * For now this just excludes isolated CPUs, but could be used to
2080 * exclude other special cases in the future.
2082 int sched_init_domains(const struct cpumask
*cpu_map
)
2086 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_KERNEL
);
2087 zalloc_cpumask_var(&sched_domains_tmpmask2
, GFP_KERNEL
);
2088 zalloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
2090 arch_update_cpu_topology();
2092 doms_cur
= alloc_sched_domains(ndoms_cur
);
2094 doms_cur
= &fallback_doms
;
2095 cpumask_and(doms_cur
[0], cpu_map
, housekeeping_cpumask(HK_FLAG_DOMAIN
));
2096 err
= build_sched_domains(doms_cur
[0], NULL
);
2097 register_sched_domain_sysctl();
2103 * Detach sched domains from a group of CPUs specified in cpu_map
2104 * These CPUs will now be attached to the NULL domain
2106 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
2111 for_each_cpu(i
, cpu_map
)
2112 cpu_attach_domain(NULL
, &def_root_domain
, i
);
2116 /* handle null as "default" */
2117 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
2118 struct sched_domain_attr
*new, int idx_new
)
2120 struct sched_domain_attr tmp
;
2128 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
2129 new ? (new + idx_new
) : &tmp
,
2130 sizeof(struct sched_domain_attr
));
2134 * Partition sched domains as specified by the 'ndoms_new'
2135 * cpumasks in the array doms_new[] of cpumasks. This compares
2136 * doms_new[] to the current sched domain partitioning, doms_cur[].
2137 * It destroys each deleted domain and builds each new domain.
2139 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2140 * The masks don't intersect (don't overlap.) We should setup one
2141 * sched domain for each mask. CPUs not in any of the cpumasks will
2142 * not be load balanced. If the same cpumask appears both in the
2143 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2146 * The passed in 'doms_new' should be allocated using
2147 * alloc_sched_domains. This routine takes ownership of it and will
2148 * free_sched_domains it when done with it. If the caller failed the
2149 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2150 * and partition_sched_domains() will fallback to the single partition
2151 * 'fallback_doms', it also forces the domains to be rebuilt.
2153 * If doms_new == NULL it will be replaced with cpu_online_mask.
2154 * ndoms_new == 0 is a special case for destroying existing domains,
2155 * and it will not create the default domain.
2157 * Call with hotplug lock held
2159 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
2160 struct sched_domain_attr
*dattr_new
)
2162 bool __maybe_unused has_eas
= false;
2166 mutex_lock(&sched_domains_mutex
);
2168 /* Always unregister in case we don't destroy any domains: */
2169 unregister_sched_domain_sysctl();
2171 /* Let the architecture update CPU core mappings: */
2172 new_topology
= arch_update_cpu_topology();
2175 WARN_ON_ONCE(dattr_new
);
2177 doms_new
= alloc_sched_domains(1);
2180 cpumask_and(doms_new
[0], cpu_active_mask
,
2181 housekeeping_cpumask(HK_FLAG_DOMAIN
));
2187 /* Destroy deleted domains: */
2188 for (i
= 0; i
< ndoms_cur
; i
++) {
2189 for (j
= 0; j
< n
&& !new_topology
; j
++) {
2190 if (cpumask_equal(doms_cur
[i
], doms_new
[j
]) &&
2191 dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
2194 /* No match - a current sched domain not in new doms_new[] */
2195 detach_destroy_domains(doms_cur
[i
]);
2203 doms_new
= &fallback_doms
;
2204 cpumask_and(doms_new
[0], cpu_active_mask
,
2205 housekeeping_cpumask(HK_FLAG_DOMAIN
));
2208 /* Build new domains: */
2209 for (i
= 0; i
< ndoms_new
; i
++) {
2210 for (j
= 0; j
< n
&& !new_topology
; j
++) {
2211 if (cpumask_equal(doms_new
[i
], doms_cur
[j
]) &&
2212 dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
2215 /* No match - add a new doms_new */
2216 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
2221 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2222 /* Build perf. domains: */
2223 for (i
= 0; i
< ndoms_new
; i
++) {
2224 for (j
= 0; j
< n
&& !sched_energy_update
; j
++) {
2225 if (cpumask_equal(doms_new
[i
], doms_cur
[j
]) &&
2226 cpu_rq(cpumask_first(doms_cur
[j
]))->rd
->pd
) {
2231 /* No match - add perf. domains for a new rd */
2232 has_eas
|= build_perf_domains(doms_new
[i
]);
2236 sched_energy_set(has_eas
);
2239 /* Remember the new sched domains: */
2240 if (doms_cur
!= &fallback_doms
)
2241 free_sched_domains(doms_cur
, ndoms_cur
);
2244 doms_cur
= doms_new
;
2245 dattr_cur
= dattr_new
;
2246 ndoms_cur
= ndoms_new
;
2248 register_sched_domain_sysctl();
2250 mutex_unlock(&sched_domains_mutex
);