1 // SPDX-License-Identifier: GPL-2.0
3 * Scheduler topology setup/handling methods
5 #include <linux/sched.h>
6 #include <linux/mutex.h>
10 DEFINE_MUTEX(sched_domains_mutex
);
12 /* Protected by sched_domains_mutex: */
13 cpumask_var_t sched_domains_tmpmask
;
14 cpumask_var_t sched_domains_tmpmask2
;
16 #ifdef CONFIG_SCHED_DEBUG
18 static int __init
sched_debug_setup(char *str
)
20 sched_debug_enabled
= true;
24 early_param("sched_debug", sched_debug_setup
);
26 static inline bool sched_debug(void)
28 return sched_debug_enabled
;
31 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
32 struct cpumask
*groupmask
)
34 struct sched_group
*group
= sd
->groups
;
36 cpumask_clear(groupmask
);
38 printk(KERN_DEBUG
"%*s domain-%d: ", level
, "", level
);
40 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
41 printk("does not load-balance\n");
43 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
48 printk(KERN_CONT
"span=%*pbl level=%s\n",
49 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
51 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
52 printk(KERN_ERR
"ERROR: domain->span does not contain "
55 if (!cpumask_test_cpu(cpu
, sched_group_span(group
))) {
56 printk(KERN_ERR
"ERROR: domain->groups does not contain"
60 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
64 printk(KERN_ERR
"ERROR: group is NULL\n");
68 if (!cpumask_weight(sched_group_span(group
))) {
69 printk(KERN_CONT
"\n");
70 printk(KERN_ERR
"ERROR: empty group\n");
74 if (!(sd
->flags
& SD_OVERLAP
) &&
75 cpumask_intersects(groupmask
, sched_group_span(group
))) {
76 printk(KERN_CONT
"\n");
77 printk(KERN_ERR
"ERROR: repeated CPUs\n");
81 cpumask_or(groupmask
, groupmask
, sched_group_span(group
));
83 printk(KERN_CONT
" %d:{ span=%*pbl",
85 cpumask_pr_args(sched_group_span(group
)));
87 if ((sd
->flags
& SD_OVERLAP
) &&
88 !cpumask_equal(group_balance_mask(group
), sched_group_span(group
))) {
89 printk(KERN_CONT
" mask=%*pbl",
90 cpumask_pr_args(group_balance_mask(group
)));
93 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
)
94 printk(KERN_CONT
" cap=%lu", group
->sgc
->capacity
);
96 if (group
== sd
->groups
&& sd
->child
&&
97 !cpumask_equal(sched_domain_span(sd
->child
),
98 sched_group_span(group
))) {
99 printk(KERN_ERR
"ERROR: domain->groups does not match domain->child\n");
102 printk(KERN_CONT
" }");
106 if (group
!= sd
->groups
)
107 printk(KERN_CONT
",");
109 } while (group
!= sd
->groups
);
110 printk(KERN_CONT
"\n");
112 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
113 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
116 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
117 printk(KERN_ERR
"ERROR: parent span is not a superset "
118 "of domain->span\n");
122 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
126 if (!sched_debug_enabled
)
130 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
134 printk(KERN_DEBUG
"CPU%d attaching sched-domain(s):\n", cpu
);
137 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
145 #else /* !CONFIG_SCHED_DEBUG */
147 # define sched_debug_enabled 0
148 # define sched_domain_debug(sd, cpu) do { } while (0)
149 static inline bool sched_debug(void)
153 #endif /* CONFIG_SCHED_DEBUG */
155 static int sd_degenerate(struct sched_domain
*sd
)
157 if (cpumask_weight(sched_domain_span(sd
)) == 1)
160 /* Following flags need at least 2 groups */
161 if (sd
->flags
& (SD_LOAD_BALANCE
|
165 SD_SHARE_CPUCAPACITY
|
166 SD_ASYM_CPUCAPACITY
|
167 SD_SHARE_PKG_RESOURCES
|
168 SD_SHARE_POWERDOMAIN
)) {
169 if (sd
->groups
!= sd
->groups
->next
)
173 /* Following flags don't use groups */
174 if (sd
->flags
& (SD_WAKE_AFFINE
))
181 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
183 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
185 if (sd_degenerate(parent
))
188 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
191 /* Flags needing groups don't count if only 1 group in parent */
192 if (parent
->groups
== parent
->groups
->next
) {
193 pflags
&= ~(SD_LOAD_BALANCE
|
197 SD_ASYM_CPUCAPACITY
|
198 SD_SHARE_CPUCAPACITY
|
199 SD_SHARE_PKG_RESOURCES
|
201 SD_SHARE_POWERDOMAIN
);
202 if (nr_node_ids
== 1)
203 pflags
&= ~SD_SERIALIZE
;
205 if (~cflags
& pflags
)
211 static void free_rootdomain(struct rcu_head
*rcu
)
213 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
215 cpupri_cleanup(&rd
->cpupri
);
216 cpudl_cleanup(&rd
->cpudl
);
217 free_cpumask_var(rd
->dlo_mask
);
218 free_cpumask_var(rd
->rto_mask
);
219 free_cpumask_var(rd
->online
);
220 free_cpumask_var(rd
->span
);
224 void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
226 struct root_domain
*old_rd
= NULL
;
229 raw_spin_lock_irqsave(&rq
->lock
, flags
);
234 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
237 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
240 * If we dont want to free the old_rd yet then
241 * set old_rd to NULL to skip the freeing later
244 if (!atomic_dec_and_test(&old_rd
->refcount
))
248 atomic_inc(&rd
->refcount
);
251 cpumask_set_cpu(rq
->cpu
, rd
->span
);
252 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
255 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
258 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
261 static int init_rootdomain(struct root_domain
*rd
)
263 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
265 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
267 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
269 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
272 init_dl_bw(&rd
->dl_bw
);
273 if (cpudl_init(&rd
->cpudl
) != 0)
276 if (cpupri_init(&rd
->cpupri
) != 0)
281 cpudl_cleanup(&rd
->cpudl
);
283 free_cpumask_var(rd
->rto_mask
);
285 free_cpumask_var(rd
->dlo_mask
);
287 free_cpumask_var(rd
->online
);
289 free_cpumask_var(rd
->span
);
295 * By default the system creates a single root-domain with all CPUs as
296 * members (mimicking the global state we have today).
298 struct root_domain def_root_domain
;
300 void init_defrootdomain(void)
302 init_rootdomain(&def_root_domain
);
304 atomic_set(&def_root_domain
.refcount
, 1);
307 static struct root_domain
*alloc_rootdomain(void)
309 struct root_domain
*rd
;
311 rd
= kzalloc(sizeof(*rd
), GFP_KERNEL
);
315 if (init_rootdomain(rd
) != 0) {
323 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
325 struct sched_group
*tmp
, *first
;
334 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
337 if (atomic_dec_and_test(&sg
->ref
))
340 } while (sg
!= first
);
343 static void destroy_sched_domain(struct sched_domain
*sd
)
346 * A normal sched domain may have multiple group references, an
347 * overlapping domain, having private groups, only one. Iterate,
348 * dropping group/capacity references, freeing where none remain.
350 free_sched_groups(sd
->groups
, 1);
352 if (sd
->shared
&& atomic_dec_and_test(&sd
->shared
->ref
))
357 static void destroy_sched_domains_rcu(struct rcu_head
*rcu
)
359 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
362 struct sched_domain
*parent
= sd
->parent
;
363 destroy_sched_domain(sd
);
368 static void destroy_sched_domains(struct sched_domain
*sd
)
371 call_rcu(&sd
->rcu
, destroy_sched_domains_rcu
);
375 * Keep a special pointer to the highest sched_domain that has
376 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
377 * allows us to avoid some pointer chasing select_idle_sibling().
379 * Also keep a unique ID per domain (we use the first CPU number in
380 * the cpumask of the domain), this allows us to quickly tell if
381 * two CPUs are in the same cache domain, see cpus_share_cache().
383 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
384 DEFINE_PER_CPU(int, sd_llc_size
);
385 DEFINE_PER_CPU(int, sd_llc_id
);
386 DEFINE_PER_CPU(struct sched_domain_shared
*, sd_llc_shared
);
387 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
388 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
390 static void update_top_cache_domain(int cpu
)
392 struct sched_domain_shared
*sds
= NULL
;
393 struct sched_domain
*sd
;
397 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
399 id
= cpumask_first(sched_domain_span(sd
));
400 size
= cpumask_weight(sched_domain_span(sd
));
404 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
405 per_cpu(sd_llc_size
, cpu
) = size
;
406 per_cpu(sd_llc_id
, cpu
) = id
;
407 rcu_assign_pointer(per_cpu(sd_llc_shared
, cpu
), sds
);
409 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
410 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
412 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
413 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
417 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
418 * hold the hotplug lock.
421 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
423 struct rq
*rq
= cpu_rq(cpu
);
424 struct sched_domain
*tmp
;
426 /* Remove the sched domains which do not contribute to scheduling. */
427 for (tmp
= sd
; tmp
; ) {
428 struct sched_domain
*parent
= tmp
->parent
;
432 if (sd_parent_degenerate(tmp
, parent
)) {
433 tmp
->parent
= parent
->parent
;
435 parent
->parent
->child
= tmp
;
437 * Transfer SD_PREFER_SIBLING down in case of a
438 * degenerate parent; the spans match for this
439 * so the property transfers.
441 if (parent
->flags
& SD_PREFER_SIBLING
)
442 tmp
->flags
|= SD_PREFER_SIBLING
;
443 destroy_sched_domain(parent
);
448 if (sd
&& sd_degenerate(sd
)) {
451 destroy_sched_domain(tmp
);
456 sched_domain_debug(sd
, cpu
);
458 rq_attach_root(rq
, rd
);
460 rcu_assign_pointer(rq
->sd
, sd
);
461 dirty_sched_domain_sysctl(cpu
);
462 destroy_sched_domains(tmp
);
464 update_top_cache_domain(cpu
);
467 /* Setup the mask of CPUs configured for isolated domains */
468 static int __init
isolated_cpu_setup(char *str
)
472 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
473 ret
= cpulist_parse(str
, cpu_isolated_map
);
475 pr_err("sched: Error, all isolcpus= values must be between 0 and %u\n", nr_cpu_ids
);
480 __setup("isolcpus=", isolated_cpu_setup
);
483 struct sched_domain
** __percpu sd
;
484 struct root_domain
*rd
;
495 * Return the canonical balance CPU for this group, this is the first CPU
496 * of this group that's also in the balance mask.
498 * The balance mask are all those CPUs that could actually end up at this
499 * group. See build_balance_mask().
501 * Also see should_we_balance().
503 int group_balance_cpu(struct sched_group
*sg
)
505 return cpumask_first(group_balance_mask(sg
));
510 * NUMA topology (first read the regular topology blurb below)
512 * Given a node-distance table, for example:
520 * which represents a 4 node ring topology like:
528 * We want to construct domains and groups to represent this. The way we go
529 * about doing this is to build the domains on 'hops'. For each NUMA level we
530 * construct the mask of all nodes reachable in @level hops.
532 * For the above NUMA topology that gives 3 levels:
534 * NUMA-2 0-3 0-3 0-3 0-3
535 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
537 * NUMA-1 0-1,3 0-2 1-3 0,2-3
538 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
543 * As can be seen; things don't nicely line up as with the regular topology.
544 * When we iterate a domain in child domain chunks some nodes can be
545 * represented multiple times -- hence the "overlap" naming for this part of
548 * In order to minimize this overlap, we only build enough groups to cover the
549 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
553 * - the first group of each domain is its child domain; this
554 * gets us the first 0-1,3
555 * - the only uncovered node is 2, who's child domain is 1-3.
557 * However, because of the overlap, computing a unique CPU for each group is
558 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
559 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
560 * end up at those groups (they would end up in group: 0-1,3).
562 * To correct this we have to introduce the group balance mask. This mask
563 * will contain those CPUs in the group that can reach this group given the
564 * (child) domain tree.
566 * With this we can once again compute balance_cpu and sched_group_capacity
569 * XXX include words on how balance_cpu is unique and therefore can be
570 * used for sched_group_capacity links.
573 * Another 'interesting' topology is:
581 * Which looks a little like:
589 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
592 * This leads to a few particularly weird cases where the sched_domain's are
593 * not of the same number for each cpu. Consider:
596 * groups: {0-2},{1-3} {1-3},{0-2}
598 * NUMA-1 0-2 0-3 0-3 1-3
606 * Build the balance mask; it contains only those CPUs that can arrive at this
607 * group and should be considered to continue balancing.
609 * We do this during the group creation pass, therefore the group information
610 * isn't complete yet, however since each group represents a (child) domain we
611 * can fully construct this using the sched_domain bits (which are already
615 build_balance_mask(struct sched_domain
*sd
, struct sched_group
*sg
, struct cpumask
*mask
)
617 const struct cpumask
*sg_span
= sched_group_span(sg
);
618 struct sd_data
*sdd
= sd
->private;
619 struct sched_domain
*sibling
;
624 for_each_cpu(i
, sg_span
) {
625 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
628 * Can happen in the asymmetric case, where these siblings are
629 * unused. The mask will not be empty because those CPUs that
630 * do have the top domain _should_ span the domain.
635 /* If we would not end up here, we can't continue from here */
636 if (!cpumask_equal(sg_span
, sched_domain_span(sibling
->child
)))
639 cpumask_set_cpu(i
, mask
);
642 /* We must not have empty masks here */
643 WARN_ON_ONCE(cpumask_empty(mask
));
647 * XXX: This creates per-node group entries; since the load-balancer will
648 * immediately access remote memory to construct this group's load-balance
649 * statistics having the groups node local is of dubious benefit.
651 static struct sched_group
*
652 build_group_from_child_sched_domain(struct sched_domain
*sd
, int cpu
)
654 struct sched_group
*sg
;
655 struct cpumask
*sg_span
;
657 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
658 GFP_KERNEL
, cpu_to_node(cpu
));
663 sg_span
= sched_group_span(sg
);
665 cpumask_copy(sg_span
, sched_domain_span(sd
->child
));
667 cpumask_copy(sg_span
, sched_domain_span(sd
));
669 atomic_inc(&sg
->ref
);
673 static void init_overlap_sched_group(struct sched_domain
*sd
,
674 struct sched_group
*sg
)
676 struct cpumask
*mask
= sched_domains_tmpmask2
;
677 struct sd_data
*sdd
= sd
->private;
678 struct cpumask
*sg_span
;
681 build_balance_mask(sd
, sg
, mask
);
682 cpu
= cpumask_first_and(sched_group_span(sg
), mask
);
684 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
685 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
686 cpumask_copy(group_balance_mask(sg
), mask
);
688 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg
), mask
));
691 * Initialize sgc->capacity such that even if we mess up the
692 * domains and no possible iteration will get us here, we won't
695 sg_span
= sched_group_span(sg
);
696 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
697 sg
->sgc
->min_capacity
= SCHED_CAPACITY_SCALE
;
701 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
703 struct sched_group
*first
= NULL
, *last
= NULL
, *sg
;
704 const struct cpumask
*span
= sched_domain_span(sd
);
705 struct cpumask
*covered
= sched_domains_tmpmask
;
706 struct sd_data
*sdd
= sd
->private;
707 struct sched_domain
*sibling
;
710 cpumask_clear(covered
);
712 for_each_cpu_wrap(i
, span
, cpu
) {
713 struct cpumask
*sg_span
;
715 if (cpumask_test_cpu(i
, covered
))
718 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
721 * Asymmetric node setups can result in situations where the
722 * domain tree is of unequal depth, make sure to skip domains
723 * that already cover the entire range.
725 * In that case build_sched_domains() will have terminated the
726 * iteration early and our sibling sd spans will be empty.
727 * Domains should always include the CPU they're built on, so
730 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
733 sg
= build_group_from_child_sched_domain(sibling
, cpu
);
737 sg_span
= sched_group_span(sg
);
738 cpumask_or(covered
, covered
, sg_span
);
740 init_overlap_sched_group(sd
, sg
);
754 free_sched_groups(first
, 0);
761 * Package topology (also see the load-balance blurb in fair.c)
763 * The scheduler builds a tree structure to represent a number of important
764 * topology features. By default (default_topology[]) these include:
766 * - Simultaneous multithreading (SMT)
767 * - Multi-Core Cache (MC)
770 * Where the last one more or less denotes everything up to a NUMA node.
772 * The tree consists of 3 primary data structures:
774 * sched_domain -> sched_group -> sched_group_capacity
778 * The sched_domains are per-cpu and have a two way link (parent & child) and
779 * denote the ever growing mask of CPUs belonging to that level of topology.
781 * Each sched_domain has a circular (double) linked list of sched_group's, each
782 * denoting the domains of the level below (or individual CPUs in case of the
783 * first domain level). The sched_group linked by a sched_domain includes the
784 * CPU of that sched_domain [*].
786 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
788 * CPU 0 1 2 3 4 5 6 7
792 * SMT [ ] [ ] [ ] [ ]
796 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
797 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
798 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
800 * CPU 0 1 2 3 4 5 6 7
802 * One way to think about it is: sched_domain moves you up and down among these
803 * topology levels, while sched_group moves you sideways through it, at child
804 * domain granularity.
806 * sched_group_capacity ensures each unique sched_group has shared storage.
808 * There are two related construction problems, both require a CPU that
809 * uniquely identify each group (for a given domain):
811 * - The first is the balance_cpu (see should_we_balance() and the
812 * load-balance blub in fair.c); for each group we only want 1 CPU to
813 * continue balancing at a higher domain.
815 * - The second is the sched_group_capacity; we want all identical groups
816 * to share a single sched_group_capacity.
818 * Since these topologies are exclusive by construction. That is, its
819 * impossible for an SMT thread to belong to multiple cores, and cores to
820 * be part of multiple caches. There is a very clear and unique location
821 * for each CPU in the hierarchy.
823 * Therefore computing a unique CPU for each group is trivial (the iteration
824 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
825 * group), we can simply pick the first CPU in each group.
828 * [*] in other words, the first group of each domain is its child domain.
831 static struct sched_group
*get_group(int cpu
, struct sd_data
*sdd
)
833 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
834 struct sched_domain
*child
= sd
->child
;
835 struct sched_group
*sg
;
838 cpu
= cpumask_first(sched_domain_span(child
));
840 sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
841 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
843 /* For claim_allocations: */
844 atomic_inc(&sg
->ref
);
845 atomic_inc(&sg
->sgc
->ref
);
848 cpumask_copy(sched_group_span(sg
), sched_domain_span(child
));
849 cpumask_copy(group_balance_mask(sg
), sched_group_span(sg
));
851 cpumask_set_cpu(cpu
, sched_group_span(sg
));
852 cpumask_set_cpu(cpu
, group_balance_mask(sg
));
855 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sched_group_span(sg
));
856 sg
->sgc
->min_capacity
= SCHED_CAPACITY_SCALE
;
862 * build_sched_groups will build a circular linked list of the groups
863 * covered by the given span, and will set each group's ->cpumask correctly,
864 * and ->cpu_capacity to 0.
866 * Assumes the sched_domain tree is fully constructed
869 build_sched_groups(struct sched_domain
*sd
, int cpu
)
871 struct sched_group
*first
= NULL
, *last
= NULL
;
872 struct sd_data
*sdd
= sd
->private;
873 const struct cpumask
*span
= sched_domain_span(sd
);
874 struct cpumask
*covered
;
877 lockdep_assert_held(&sched_domains_mutex
);
878 covered
= sched_domains_tmpmask
;
880 cpumask_clear(covered
);
882 for_each_cpu_wrap(i
, span
, cpu
) {
883 struct sched_group
*sg
;
885 if (cpumask_test_cpu(i
, covered
))
888 sg
= get_group(i
, sdd
);
890 cpumask_or(covered
, covered
, sched_group_span(sg
));
905 * Initialize sched groups cpu_capacity.
907 * cpu_capacity indicates the capacity of sched group, which is used while
908 * distributing the load between different sched groups in a sched domain.
909 * Typically cpu_capacity for all the groups in a sched domain will be same
910 * unless there are asymmetries in the topology. If there are asymmetries,
911 * group having more cpu_capacity will pickup more load compared to the
912 * group having less cpu_capacity.
914 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
916 struct sched_group
*sg
= sd
->groups
;
921 int cpu
, max_cpu
= -1;
923 sg
->group_weight
= cpumask_weight(sched_group_span(sg
));
925 if (!(sd
->flags
& SD_ASYM_PACKING
))
928 for_each_cpu(cpu
, sched_group_span(sg
)) {
931 else if (sched_asym_prefer(cpu
, max_cpu
))
934 sg
->asym_prefer_cpu
= max_cpu
;
938 } while (sg
!= sd
->groups
);
940 if (cpu
!= group_balance_cpu(sg
))
943 update_group_capacity(sd
, cpu
);
947 * Initializers for schedule domains
948 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
951 static int default_relax_domain_level
= -1;
952 int sched_domain_level_max
;
954 static int __init
setup_relax_domain_level(char *str
)
956 if (kstrtoint(str
, 0, &default_relax_domain_level
))
957 pr_warn("Unable to set relax_domain_level\n");
961 __setup("relax_domain_level=", setup_relax_domain_level
);
963 static void set_domain_attribute(struct sched_domain
*sd
,
964 struct sched_domain_attr
*attr
)
968 if (!attr
|| attr
->relax_domain_level
< 0) {
969 if (default_relax_domain_level
< 0)
972 request
= default_relax_domain_level
;
974 request
= attr
->relax_domain_level
;
975 if (request
< sd
->level
) {
976 /* Turn off idle balance on this domain: */
977 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
979 /* Turn on idle balance on this domain: */
980 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
984 static void __sdt_free(const struct cpumask
*cpu_map
);
985 static int __sdt_alloc(const struct cpumask
*cpu_map
);
987 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
988 const struct cpumask
*cpu_map
)
992 if (!atomic_read(&d
->rd
->refcount
))
993 free_rootdomain(&d
->rd
->rcu
);
1007 __visit_domain_allocation_hell(struct s_data
*d
, const struct cpumask
*cpu_map
)
1009 memset(d
, 0, sizeof(*d
));
1011 if (__sdt_alloc(cpu_map
))
1012 return sa_sd_storage
;
1013 d
->sd
= alloc_percpu(struct sched_domain
*);
1015 return sa_sd_storage
;
1016 d
->rd
= alloc_rootdomain();
1019 return sa_rootdomain
;
1023 * NULL the sd_data elements we've used to build the sched_domain and
1024 * sched_group structure so that the subsequent __free_domain_allocs()
1025 * will not free the data we're using.
1027 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
1029 struct sd_data
*sdd
= sd
->private;
1031 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
1032 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
1034 if (atomic_read(&(*per_cpu_ptr(sdd
->sds
, cpu
))->ref
))
1035 *per_cpu_ptr(sdd
->sds
, cpu
) = NULL
;
1037 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
1038 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
1040 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
1041 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
1045 static int sched_domains_numa_levels
;
1046 enum numa_topology_type sched_numa_topology_type
;
1047 static int *sched_domains_numa_distance
;
1048 int sched_max_numa_distance
;
1049 static struct cpumask
***sched_domains_numa_masks
;
1050 static int sched_domains_curr_level
;
1054 * SD_flags allowed in topology descriptions.
1056 * These flags are purely descriptive of the topology and do not prescribe
1057 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1060 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1061 * SD_SHARE_PKG_RESOURCES - describes shared caches
1062 * SD_NUMA - describes NUMA topologies
1063 * SD_SHARE_POWERDOMAIN - describes shared power domain
1064 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
1066 * Odd one out, which beside describing the topology has a quirk also
1067 * prescribes the desired behaviour that goes along with it:
1069 * SD_ASYM_PACKING - describes SMT quirks
1071 #define TOPOLOGY_SD_FLAGS \
1072 (SD_SHARE_CPUCAPACITY | \
1073 SD_SHARE_PKG_RESOURCES | \
1076 SD_ASYM_CPUCAPACITY | \
1077 SD_SHARE_POWERDOMAIN)
1079 static struct sched_domain
*
1080 sd_init(struct sched_domain_topology_level
*tl
,
1081 const struct cpumask
*cpu_map
,
1082 struct sched_domain
*child
, int cpu
)
1084 struct sd_data
*sdd
= &tl
->data
;
1085 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
1086 int sd_id
, sd_weight
, sd_flags
= 0;
1090 * Ugly hack to pass state to sd_numa_mask()...
1092 sched_domains_curr_level
= tl
->numa_level
;
1095 sd_weight
= cpumask_weight(tl
->mask(cpu
));
1098 sd_flags
= (*tl
->sd_flags
)();
1099 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
1100 "wrong sd_flags in topology description\n"))
1101 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
1103 *sd
= (struct sched_domain
){
1104 .min_interval
= sd_weight
,
1105 .max_interval
= 2*sd_weight
,
1107 .imbalance_pct
= 125,
1109 .cache_nice_tries
= 0,
1116 .flags
= 1*SD_LOAD_BALANCE
1117 | 1*SD_BALANCE_NEWIDLE
1122 | 0*SD_SHARE_CPUCAPACITY
1123 | 0*SD_SHARE_PKG_RESOURCES
1125 | 0*SD_PREFER_SIBLING
1130 .last_balance
= jiffies
,
1131 .balance_interval
= sd_weight
,
1133 .max_newidle_lb_cost
= 0,
1134 .next_decay_max_lb_cost
= jiffies
,
1136 #ifdef CONFIG_SCHED_DEBUG
1141 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
1142 sd_id
= cpumask_first(sched_domain_span(sd
));
1145 * Convert topological properties into behaviour.
1148 if (sd
->flags
& SD_ASYM_CPUCAPACITY
) {
1149 struct sched_domain
*t
= sd
;
1151 for_each_lower_domain(t
)
1152 t
->flags
|= SD_BALANCE_WAKE
;
1155 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
1156 sd
->flags
|= SD_PREFER_SIBLING
;
1157 sd
->imbalance_pct
= 110;
1158 sd
->smt_gain
= 1178; /* ~15% */
1160 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
1161 sd
->imbalance_pct
= 117;
1162 sd
->cache_nice_tries
= 1;
1166 } else if (sd
->flags
& SD_NUMA
) {
1167 sd
->cache_nice_tries
= 2;
1171 sd
->flags
|= SD_SERIALIZE
;
1172 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
1173 sd
->flags
&= ~(SD_BALANCE_EXEC
|
1180 sd
->flags
|= SD_PREFER_SIBLING
;
1181 sd
->cache_nice_tries
= 1;
1187 * For all levels sharing cache; connect a sched_domain_shared
1190 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
1191 sd
->shared
= *per_cpu_ptr(sdd
->sds
, sd_id
);
1192 atomic_inc(&sd
->shared
->ref
);
1193 atomic_set(&sd
->shared
->nr_busy_cpus
, sd_weight
);
1202 * Topology list, bottom-up.
1204 static struct sched_domain_topology_level default_topology
[] = {
1205 #ifdef CONFIG_SCHED_SMT
1206 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
1208 #ifdef CONFIG_SCHED_MC
1209 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
1211 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
1215 static struct sched_domain_topology_level
*sched_domain_topology
=
1218 #define for_each_sd_topology(tl) \
1219 for (tl = sched_domain_topology; tl->mask; tl++)
1221 void set_sched_topology(struct sched_domain_topology_level
*tl
)
1223 if (WARN_ON_ONCE(sched_smp_initialized
))
1226 sched_domain_topology
= tl
;
1231 static const struct cpumask
*sd_numa_mask(int cpu
)
1233 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
1236 static void sched_numa_warn(const char *str
)
1238 static int done
= false;
1246 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
1248 for (i
= 0; i
< nr_node_ids
; i
++) {
1249 printk(KERN_WARNING
" ");
1250 for (j
= 0; j
< nr_node_ids
; j
++)
1251 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
1252 printk(KERN_CONT
"\n");
1254 printk(KERN_WARNING
"\n");
1257 bool find_numa_distance(int distance
)
1261 if (distance
== node_distance(0, 0))
1264 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
1265 if (sched_domains_numa_distance
[i
] == distance
)
1273 * A system can have three types of NUMA topology:
1274 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1275 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1276 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1278 * The difference between a glueless mesh topology and a backplane
1279 * topology lies in whether communication between not directly
1280 * connected nodes goes through intermediary nodes (where programs
1281 * could run), or through backplane controllers. This affects
1282 * placement of programs.
1284 * The type of topology can be discerned with the following tests:
1285 * - If the maximum distance between any nodes is 1 hop, the system
1286 * is directly connected.
1287 * - If for two nodes A and B, located N > 1 hops away from each other,
1288 * there is an intermediary node C, which is < N hops away from both
1289 * nodes A and B, the system is a glueless mesh.
1291 static void init_numa_topology_type(void)
1295 n
= sched_max_numa_distance
;
1297 if (sched_domains_numa_levels
<= 1) {
1298 sched_numa_topology_type
= NUMA_DIRECT
;
1302 for_each_online_node(a
) {
1303 for_each_online_node(b
) {
1304 /* Find two nodes furthest removed from each other. */
1305 if (node_distance(a
, b
) < n
)
1308 /* Is there an intermediary node between a and b? */
1309 for_each_online_node(c
) {
1310 if (node_distance(a
, c
) < n
&&
1311 node_distance(b
, c
) < n
) {
1312 sched_numa_topology_type
=
1318 sched_numa_topology_type
= NUMA_BACKPLANE
;
1324 void sched_init_numa(void)
1326 int next_distance
, curr_distance
= node_distance(0, 0);
1327 struct sched_domain_topology_level
*tl
;
1331 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
1332 if (!sched_domains_numa_distance
)
1336 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1337 * unique distances in the node_distance() table.
1339 * Assumes node_distance(0,j) includes all distances in
1340 * node_distance(i,j) in order to avoid cubic time.
1342 next_distance
= curr_distance
;
1343 for (i
= 0; i
< nr_node_ids
; i
++) {
1344 for (j
= 0; j
< nr_node_ids
; j
++) {
1345 for (k
= 0; k
< nr_node_ids
; k
++) {
1346 int distance
= node_distance(i
, k
);
1348 if (distance
> curr_distance
&&
1349 (distance
< next_distance
||
1350 next_distance
== curr_distance
))
1351 next_distance
= distance
;
1354 * While not a strong assumption it would be nice to know
1355 * about cases where if node A is connected to B, B is not
1356 * equally connected to A.
1358 if (sched_debug() && node_distance(k
, i
) != distance
)
1359 sched_numa_warn("Node-distance not symmetric");
1361 if (sched_debug() && i
&& !find_numa_distance(distance
))
1362 sched_numa_warn("Node-0 not representative");
1364 if (next_distance
!= curr_distance
) {
1365 sched_domains_numa_distance
[level
++] = next_distance
;
1366 sched_domains_numa_levels
= level
;
1367 curr_distance
= next_distance
;
1372 * In case of sched_debug() we verify the above assumption.
1382 * 'level' contains the number of unique distances, excluding the
1383 * identity distance node_distance(i,i).
1385 * The sched_domains_numa_distance[] array includes the actual distance
1390 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1391 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1392 * the array will contain less then 'level' members. This could be
1393 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1394 * in other functions.
1396 * We reset it to 'level' at the end of this function.
1398 sched_domains_numa_levels
= 0;
1400 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
1401 if (!sched_domains_numa_masks
)
1405 * Now for each level, construct a mask per node which contains all
1406 * CPUs of nodes that are that many hops away from us.
1408 for (i
= 0; i
< level
; i
++) {
1409 sched_domains_numa_masks
[i
] =
1410 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
1411 if (!sched_domains_numa_masks
[i
])
1414 for (j
= 0; j
< nr_node_ids
; j
++) {
1415 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
1419 sched_domains_numa_masks
[i
][j
] = mask
;
1422 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
1425 cpumask_or(mask
, mask
, cpumask_of_node(k
));
1430 /* Compute default topology size */
1431 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
1433 tl
= kzalloc((i
+ level
+ 1) *
1434 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
1439 * Copy the default topology bits..
1441 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
1442 tl
[i
] = sched_domain_topology
[i
];
1445 * .. and append 'j' levels of NUMA goodness.
1447 for (j
= 0; j
< level
; i
++, j
++) {
1448 tl
[i
] = (struct sched_domain_topology_level
){
1449 .mask
= sd_numa_mask
,
1450 .sd_flags
= cpu_numa_flags
,
1451 .flags
= SDTL_OVERLAP
,
1457 sched_domain_topology
= tl
;
1459 sched_domains_numa_levels
= level
;
1460 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
1462 init_numa_topology_type();
1465 void sched_domains_numa_masks_set(unsigned int cpu
)
1467 int node
= cpu_to_node(cpu
);
1470 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
1471 for (j
= 0; j
< nr_node_ids
; j
++) {
1472 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
1473 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
1478 void sched_domains_numa_masks_clear(unsigned int cpu
)
1482 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
1483 for (j
= 0; j
< nr_node_ids
; j
++)
1484 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
1488 #endif /* CONFIG_NUMA */
1490 static int __sdt_alloc(const struct cpumask
*cpu_map
)
1492 struct sched_domain_topology_level
*tl
;
1495 for_each_sd_topology(tl
) {
1496 struct sd_data
*sdd
= &tl
->data
;
1498 sdd
->sd
= alloc_percpu(struct sched_domain
*);
1502 sdd
->sds
= alloc_percpu(struct sched_domain_shared
*);
1506 sdd
->sg
= alloc_percpu(struct sched_group
*);
1510 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
1514 for_each_cpu(j
, cpu_map
) {
1515 struct sched_domain
*sd
;
1516 struct sched_domain_shared
*sds
;
1517 struct sched_group
*sg
;
1518 struct sched_group_capacity
*sgc
;
1520 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
1521 GFP_KERNEL
, cpu_to_node(j
));
1525 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
1527 sds
= kzalloc_node(sizeof(struct sched_domain_shared
),
1528 GFP_KERNEL
, cpu_to_node(j
));
1532 *per_cpu_ptr(sdd
->sds
, j
) = sds
;
1534 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
1535 GFP_KERNEL
, cpu_to_node(j
));
1541 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
1543 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
1544 GFP_KERNEL
, cpu_to_node(j
));
1548 #ifdef CONFIG_SCHED_DEBUG
1552 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
1559 static void __sdt_free(const struct cpumask
*cpu_map
)
1561 struct sched_domain_topology_level
*tl
;
1564 for_each_sd_topology(tl
) {
1565 struct sd_data
*sdd
= &tl
->data
;
1567 for_each_cpu(j
, cpu_map
) {
1568 struct sched_domain
*sd
;
1571 sd
= *per_cpu_ptr(sdd
->sd
, j
);
1572 if (sd
&& (sd
->flags
& SD_OVERLAP
))
1573 free_sched_groups(sd
->groups
, 0);
1574 kfree(*per_cpu_ptr(sdd
->sd
, j
));
1578 kfree(*per_cpu_ptr(sdd
->sds
, j
));
1580 kfree(*per_cpu_ptr(sdd
->sg
, j
));
1582 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
1584 free_percpu(sdd
->sd
);
1586 free_percpu(sdd
->sds
);
1588 free_percpu(sdd
->sg
);
1590 free_percpu(sdd
->sgc
);
1595 static struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
1596 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
1597 struct sched_domain
*child
, int cpu
)
1599 struct sched_domain
*sd
= sd_init(tl
, cpu_map
, child
, cpu
);
1602 sd
->level
= child
->level
+ 1;
1603 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
1606 if (!cpumask_subset(sched_domain_span(child
),
1607 sched_domain_span(sd
))) {
1608 pr_err("BUG: arch topology borken\n");
1609 #ifdef CONFIG_SCHED_DEBUG
1610 pr_err(" the %s domain not a subset of the %s domain\n",
1611 child
->name
, sd
->name
);
1613 /* Fixup, ensure @sd has at least @child cpus. */
1614 cpumask_or(sched_domain_span(sd
),
1615 sched_domain_span(sd
),
1616 sched_domain_span(child
));
1620 set_domain_attribute(sd
, attr
);
1626 * Build sched domains for a given set of CPUs and attach the sched domains
1627 * to the individual CPUs
1630 build_sched_domains(const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
)
1632 enum s_alloc alloc_state
;
1633 struct sched_domain
*sd
;
1635 struct rq
*rq
= NULL
;
1636 int i
, ret
= -ENOMEM
;
1638 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
1639 if (alloc_state
!= sa_rootdomain
)
1642 /* Set up domains for CPUs specified by the cpu_map: */
1643 for_each_cpu(i
, cpu_map
) {
1644 struct sched_domain_topology_level
*tl
;
1647 for_each_sd_topology(tl
) {
1648 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
1649 if (tl
== sched_domain_topology
)
1650 *per_cpu_ptr(d
.sd
, i
) = sd
;
1651 if (tl
->flags
& SDTL_OVERLAP
)
1652 sd
->flags
|= SD_OVERLAP
;
1653 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
1658 /* Build the groups for the domains */
1659 for_each_cpu(i
, cpu_map
) {
1660 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
1661 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
1662 if (sd
->flags
& SD_OVERLAP
) {
1663 if (build_overlap_sched_groups(sd
, i
))
1666 if (build_sched_groups(sd
, i
))
1672 /* Calculate CPU capacity for physical packages and nodes */
1673 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
1674 if (!cpumask_test_cpu(i
, cpu_map
))
1677 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
1678 claim_allocations(i
, sd
);
1679 init_sched_groups_capacity(i
, sd
);
1683 /* Attach the domains */
1685 for_each_cpu(i
, cpu_map
) {
1687 sd
= *per_cpu_ptr(d
.sd
, i
);
1689 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
1690 if (rq
->cpu_capacity_orig
> READ_ONCE(d
.rd
->max_cpu_capacity
))
1691 WRITE_ONCE(d
.rd
->max_cpu_capacity
, rq
->cpu_capacity_orig
);
1693 cpu_attach_domain(sd
, d
.rd
, i
);
1697 if (rq
&& sched_debug_enabled
) {
1698 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
1699 cpumask_pr_args(cpu_map
), rq
->rd
->max_cpu_capacity
);
1704 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
1708 /* Current sched domains: */
1709 static cpumask_var_t
*doms_cur
;
1711 /* Number of sched domains in 'doms_cur': */
1712 static int ndoms_cur
;
1714 /* Attribues of custom domains in 'doms_cur' */
1715 static struct sched_domain_attr
*dattr_cur
;
1718 * Special case: If a kmalloc() of a doms_cur partition (array of
1719 * cpumask) fails, then fallback to a single sched domain,
1720 * as determined by the single cpumask fallback_doms.
1722 static cpumask_var_t fallback_doms
;
1725 * arch_update_cpu_topology lets virtualized architectures update the
1726 * CPU core maps. It is supposed to return 1 if the topology changed
1727 * or 0 if it stayed the same.
1729 int __weak
arch_update_cpu_topology(void)
1734 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
1737 cpumask_var_t
*doms
;
1739 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
1742 for (i
= 0; i
< ndoms
; i
++) {
1743 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
1744 free_sched_domains(doms
, i
);
1751 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
1754 for (i
= 0; i
< ndoms
; i
++)
1755 free_cpumask_var(doms
[i
]);
1760 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
1761 * For now this just excludes isolated CPUs, but could be used to
1762 * exclude other special cases in the future.
1764 int sched_init_domains(const struct cpumask
*cpu_map
)
1768 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_KERNEL
);
1769 zalloc_cpumask_var(&sched_domains_tmpmask2
, GFP_KERNEL
);
1770 zalloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
1772 arch_update_cpu_topology();
1774 doms_cur
= alloc_sched_domains(ndoms_cur
);
1776 doms_cur
= &fallback_doms
;
1777 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
1778 err
= build_sched_domains(doms_cur
[0], NULL
);
1779 register_sched_domain_sysctl();
1785 * Detach sched domains from a group of CPUs specified in cpu_map
1786 * These CPUs will now be attached to the NULL domain
1788 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
1793 for_each_cpu(i
, cpu_map
)
1794 cpu_attach_domain(NULL
, &def_root_domain
, i
);
1798 /* handle null as "default" */
1799 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
1800 struct sched_domain_attr
*new, int idx_new
)
1802 struct sched_domain_attr tmp
;
1809 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
1810 new ? (new + idx_new
) : &tmp
,
1811 sizeof(struct sched_domain_attr
));
1815 * Partition sched domains as specified by the 'ndoms_new'
1816 * cpumasks in the array doms_new[] of cpumasks. This compares
1817 * doms_new[] to the current sched domain partitioning, doms_cur[].
1818 * It destroys each deleted domain and builds each new domain.
1820 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
1821 * The masks don't intersect (don't overlap.) We should setup one
1822 * sched domain for each mask. CPUs not in any of the cpumasks will
1823 * not be load balanced. If the same cpumask appears both in the
1824 * current 'doms_cur' domains and in the new 'doms_new', we can leave
1827 * The passed in 'doms_new' should be allocated using
1828 * alloc_sched_domains. This routine takes ownership of it and will
1829 * free_sched_domains it when done with it. If the caller failed the
1830 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
1831 * and partition_sched_domains() will fallback to the single partition
1832 * 'fallback_doms', it also forces the domains to be rebuilt.
1834 * If doms_new == NULL it will be replaced with cpu_online_mask.
1835 * ndoms_new == 0 is a special case for destroying existing domains,
1836 * and it will not create the default domain.
1838 * Call with hotplug lock held
1840 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
1841 struct sched_domain_attr
*dattr_new
)
1846 mutex_lock(&sched_domains_mutex
);
1848 /* Always unregister in case we don't destroy any domains: */
1849 unregister_sched_domain_sysctl();
1851 /* Let the architecture update CPU core mappings: */
1852 new_topology
= arch_update_cpu_topology();
1855 WARN_ON_ONCE(dattr_new
);
1857 doms_new
= alloc_sched_domains(1);
1860 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
1866 /* Destroy deleted domains: */
1867 for (i
= 0; i
< ndoms_cur
; i
++) {
1868 for (j
= 0; j
< n
&& !new_topology
; j
++) {
1869 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
1870 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
1873 /* No match - a current sched domain not in new doms_new[] */
1874 detach_destroy_domains(doms_cur
[i
]);
1882 doms_new
= &fallback_doms
;
1883 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
1886 /* Build new domains: */
1887 for (i
= 0; i
< ndoms_new
; i
++) {
1888 for (j
= 0; j
< n
&& !new_topology
; j
++) {
1889 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
1890 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
1893 /* No match - add a new doms_new */
1894 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
1899 /* Remember the new sched domains: */
1900 if (doms_cur
!= &fallback_doms
)
1901 free_sched_domains(doms_cur
, ndoms_cur
);
1904 doms_cur
= doms_new
;
1905 dattr_cur
= dattr_new
;
1906 ndoms_cur
= ndoms_new
;
1908 register_sched_domain_sysctl();
1910 mutex_unlock(&sched_domains_mutex
);