2 * Scheduler topology setup/handling methods
4 #include <linux/sched.h>
5 #include <linux/mutex.h>
9 DEFINE_MUTEX(sched_domains_mutex
);
11 /* Protected by sched_domains_mutex: */
12 cpumask_var_t sched_domains_tmpmask
;
13 cpumask_var_t sched_domains_tmpmask2
;
15 #ifdef CONFIG_SCHED_DEBUG
17 static __read_mostly
int sched_debug_enabled
;
19 static int __init
sched_debug_setup(char *str
)
21 sched_debug_enabled
= 1;
25 early_param("sched_debug", sched_debug_setup
);
27 static inline bool sched_debug(void)
29 return sched_debug_enabled
;
32 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
33 struct cpumask
*groupmask
)
35 struct sched_group
*group
= sd
->groups
;
37 cpumask_clear(groupmask
);
39 printk(KERN_DEBUG
"%*s domain-%d: ", level
, "", level
);
41 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
42 printk("does not load-balance\n");
44 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
49 printk(KERN_CONT
"span=%*pbl level=%s\n",
50 cpumask_pr_args(sched_domain_span(sd
)), sd
->name
);
52 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
53 printk(KERN_ERR
"ERROR: domain->span does not contain "
56 if (!cpumask_test_cpu(cpu
, sched_group_span(group
))) {
57 printk(KERN_ERR
"ERROR: domain->groups does not contain"
61 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
65 printk(KERN_ERR
"ERROR: group is NULL\n");
69 if (!cpumask_weight(sched_group_span(group
))) {
70 printk(KERN_CONT
"\n");
71 printk(KERN_ERR
"ERROR: empty group\n");
75 if (!(sd
->flags
& SD_OVERLAP
) &&
76 cpumask_intersects(groupmask
, sched_group_span(group
))) {
77 printk(KERN_CONT
"\n");
78 printk(KERN_ERR
"ERROR: repeated CPUs\n");
82 cpumask_or(groupmask
, groupmask
, sched_group_span(group
));
84 printk(KERN_CONT
" %d:{ span=%*pbl",
86 cpumask_pr_args(sched_group_span(group
)));
88 if ((sd
->flags
& SD_OVERLAP
) &&
89 !cpumask_equal(group_balance_mask(group
), sched_group_span(group
))) {
90 printk(KERN_CONT
" mask=%*pbl",
91 cpumask_pr_args(group_balance_mask(group
)));
94 if (group
->sgc
->capacity
!= SCHED_CAPACITY_SCALE
)
95 printk(KERN_CONT
" cap=%lu", group
->sgc
->capacity
);
97 if (group
== sd
->groups
&& sd
->child
&&
98 !cpumask_equal(sched_domain_span(sd
->child
),
99 sched_group_span(group
))) {
100 printk(KERN_ERR
"ERROR: domain->groups does not match domain->child\n");
103 printk(KERN_CONT
" }");
107 if (group
!= sd
->groups
)
108 printk(KERN_CONT
",");
110 } while (group
!= sd
->groups
);
111 printk(KERN_CONT
"\n");
113 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
114 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
117 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
118 printk(KERN_ERR
"ERROR: parent span is not a superset "
119 "of domain->span\n");
123 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
127 if (!sched_debug_enabled
)
131 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
135 printk(KERN_DEBUG
"CPU%d attaching sched-domain(s):\n", cpu
);
138 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
146 #else /* !CONFIG_SCHED_DEBUG */
148 # define sched_debug_enabled 0
149 # define sched_domain_debug(sd, cpu) do { } while (0)
150 static inline bool sched_debug(void)
154 #endif /* CONFIG_SCHED_DEBUG */
156 static int sd_degenerate(struct sched_domain
*sd
)
158 if (cpumask_weight(sched_domain_span(sd
)) == 1)
161 /* Following flags need at least 2 groups */
162 if (sd
->flags
& (SD_LOAD_BALANCE
|
166 SD_SHARE_CPUCAPACITY
|
167 SD_ASYM_CPUCAPACITY
|
168 SD_SHARE_PKG_RESOURCES
|
169 SD_SHARE_POWERDOMAIN
)) {
170 if (sd
->groups
!= sd
->groups
->next
)
174 /* Following flags don't use groups */
175 if (sd
->flags
& (SD_WAKE_AFFINE
))
182 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
184 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
186 if (sd_degenerate(parent
))
189 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
192 /* Flags needing groups don't count if only 1 group in parent */
193 if (parent
->groups
== parent
->groups
->next
) {
194 pflags
&= ~(SD_LOAD_BALANCE
|
198 SD_ASYM_CPUCAPACITY
|
199 SD_SHARE_CPUCAPACITY
|
200 SD_SHARE_PKG_RESOURCES
|
202 SD_SHARE_POWERDOMAIN
);
203 if (nr_node_ids
== 1)
204 pflags
&= ~SD_SERIALIZE
;
206 if (~cflags
& pflags
)
212 static void free_rootdomain(struct rcu_head
*rcu
)
214 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
216 cpupri_cleanup(&rd
->cpupri
);
217 cpudl_cleanup(&rd
->cpudl
);
218 free_cpumask_var(rd
->dlo_mask
);
219 free_cpumask_var(rd
->rto_mask
);
220 free_cpumask_var(rd
->online
);
221 free_cpumask_var(rd
->span
);
225 void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
227 struct root_domain
*old_rd
= NULL
;
230 raw_spin_lock_irqsave(&rq
->lock
, flags
);
235 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
238 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
241 * If we dont want to free the old_rd yet then
242 * set old_rd to NULL to skip the freeing later
245 if (!atomic_dec_and_test(&old_rd
->refcount
))
249 atomic_inc(&rd
->refcount
);
252 cpumask_set_cpu(rq
->cpu
, rd
->span
);
253 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
256 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
259 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
262 static int init_rootdomain(struct root_domain
*rd
)
264 if (!zalloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
266 if (!zalloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
268 if (!zalloc_cpumask_var(&rd
->dlo_mask
, GFP_KERNEL
))
270 if (!zalloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
273 init_dl_bw(&rd
->dl_bw
);
274 if (cpudl_init(&rd
->cpudl
) != 0)
277 if (cpupri_init(&rd
->cpupri
) != 0)
282 cpudl_cleanup(&rd
->cpudl
);
284 free_cpumask_var(rd
->rto_mask
);
286 free_cpumask_var(rd
->dlo_mask
);
288 free_cpumask_var(rd
->online
);
290 free_cpumask_var(rd
->span
);
296 * By default the system creates a single root-domain with all CPUs as
297 * members (mimicking the global state we have today).
299 struct root_domain def_root_domain
;
301 void init_defrootdomain(void)
303 init_rootdomain(&def_root_domain
);
305 atomic_set(&def_root_domain
.refcount
, 1);
308 static struct root_domain
*alloc_rootdomain(void)
310 struct root_domain
*rd
;
312 rd
= kzalloc(sizeof(*rd
), GFP_KERNEL
);
316 if (init_rootdomain(rd
) != 0) {
324 static void free_sched_groups(struct sched_group
*sg
, int free_sgc
)
326 struct sched_group
*tmp
, *first
;
335 if (free_sgc
&& atomic_dec_and_test(&sg
->sgc
->ref
))
338 if (atomic_dec_and_test(&sg
->ref
))
341 } while (sg
!= first
);
344 static void destroy_sched_domain(struct sched_domain
*sd
)
347 * A normal sched domain may have multiple group references, an
348 * overlapping domain, having private groups, only one. Iterate,
349 * dropping group/capacity references, freeing where none remain.
351 free_sched_groups(sd
->groups
, 1);
353 if (sd
->shared
&& atomic_dec_and_test(&sd
->shared
->ref
))
358 static void destroy_sched_domains_rcu(struct rcu_head
*rcu
)
360 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
363 struct sched_domain
*parent
= sd
->parent
;
364 destroy_sched_domain(sd
);
369 static void destroy_sched_domains(struct sched_domain
*sd
)
372 call_rcu(&sd
->rcu
, destroy_sched_domains_rcu
);
376 * Keep a special pointer to the highest sched_domain that has
377 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
378 * allows us to avoid some pointer chasing select_idle_sibling().
380 * Also keep a unique ID per domain (we use the first CPU number in
381 * the cpumask of the domain), this allows us to quickly tell if
382 * two CPUs are in the same cache domain, see cpus_share_cache().
384 DEFINE_PER_CPU(struct sched_domain
*, sd_llc
);
385 DEFINE_PER_CPU(int, sd_llc_size
);
386 DEFINE_PER_CPU(int, sd_llc_id
);
387 DEFINE_PER_CPU(struct sched_domain_shared
*, sd_llc_shared
);
388 DEFINE_PER_CPU(struct sched_domain
*, sd_numa
);
389 DEFINE_PER_CPU(struct sched_domain
*, sd_asym
);
391 static void update_top_cache_domain(int cpu
)
393 struct sched_domain_shared
*sds
= NULL
;
394 struct sched_domain
*sd
;
398 sd
= highest_flag_domain(cpu
, SD_SHARE_PKG_RESOURCES
);
400 id
= cpumask_first(sched_domain_span(sd
));
401 size
= cpumask_weight(sched_domain_span(sd
));
405 rcu_assign_pointer(per_cpu(sd_llc
, cpu
), sd
);
406 per_cpu(sd_llc_size
, cpu
) = size
;
407 per_cpu(sd_llc_id
, cpu
) = id
;
408 rcu_assign_pointer(per_cpu(sd_llc_shared
, cpu
), sds
);
410 sd
= lowest_flag_domain(cpu
, SD_NUMA
);
411 rcu_assign_pointer(per_cpu(sd_numa
, cpu
), sd
);
413 sd
= highest_flag_domain(cpu
, SD_ASYM_PACKING
);
414 rcu_assign_pointer(per_cpu(sd_asym
, cpu
), sd
);
418 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
419 * hold the hotplug lock.
422 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
424 struct rq
*rq
= cpu_rq(cpu
);
425 struct sched_domain
*tmp
;
427 /* Remove the sched domains which do not contribute to scheduling. */
428 for (tmp
= sd
; tmp
; ) {
429 struct sched_domain
*parent
= tmp
->parent
;
433 if (sd_parent_degenerate(tmp
, parent
)) {
434 tmp
->parent
= parent
->parent
;
436 parent
->parent
->child
= tmp
;
438 * Transfer SD_PREFER_SIBLING down in case of a
439 * degenerate parent; the spans match for this
440 * so the property transfers.
442 if (parent
->flags
& SD_PREFER_SIBLING
)
443 tmp
->flags
|= SD_PREFER_SIBLING
;
444 destroy_sched_domain(parent
);
449 if (sd
&& sd_degenerate(sd
)) {
452 destroy_sched_domain(tmp
);
457 sched_domain_debug(sd
, cpu
);
459 rq_attach_root(rq
, rd
);
461 rcu_assign_pointer(rq
->sd
, sd
);
462 dirty_sched_domain_sysctl(cpu
);
463 destroy_sched_domains(tmp
);
465 update_top_cache_domain(cpu
);
468 /* Setup the mask of CPUs configured for isolated domains */
469 static int __init
isolated_cpu_setup(char *str
)
473 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
474 ret
= cpulist_parse(str
, cpu_isolated_map
);
476 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids
);
481 __setup("isolcpus=", isolated_cpu_setup
);
484 struct sched_domain
** __percpu sd
;
485 struct root_domain
*rd
;
496 * Return the canonical balance CPU for this group, this is the first CPU
497 * of this group that's also in the balance mask.
499 * The balance mask are all those CPUs that could actually end up at this
500 * group. See build_balance_mask().
502 * Also see should_we_balance().
504 int group_balance_cpu(struct sched_group
*sg
)
506 return cpumask_first(group_balance_mask(sg
));
511 * NUMA topology (first read the regular topology blurb below)
513 * Given a node-distance table, for example:
521 * which represents a 4 node ring topology like:
529 * We want to construct domains and groups to represent this. The way we go
530 * about doing this is to build the domains on 'hops'. For each NUMA level we
531 * construct the mask of all nodes reachable in @level hops.
533 * For the above NUMA topology that gives 3 levels:
535 * NUMA-2 0-3 0-3 0-3 0-3
536 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
538 * NUMA-1 0-1,3 0-2 1-3 0,2-3
539 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
544 * As can be seen; things don't nicely line up as with the regular topology.
545 * When we iterate a domain in child domain chunks some nodes can be
546 * represented multiple times -- hence the "overlap" naming for this part of
549 * In order to minimize this overlap, we only build enough groups to cover the
550 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
554 * - the first group of each domain is its child domain; this
555 * gets us the first 0-1,3
556 * - the only uncovered node is 2, who's child domain is 1-3.
558 * However, because of the overlap, computing a unique CPU for each group is
559 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
560 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
561 * end up at those groups (they would end up in group: 0-1,3).
563 * To correct this we have to introduce the group balance mask. This mask
564 * will contain those CPUs in the group that can reach this group given the
565 * (child) domain tree.
567 * With this we can once again compute balance_cpu and sched_group_capacity
570 * XXX include words on how balance_cpu is unique and therefore can be
571 * used for sched_group_capacity links.
574 * Another 'interesting' topology is:
582 * Which looks a little like:
590 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
593 * This leads to a few particularly weird cases where the sched_domain's are
594 * not of the same number for each cpu. Consider:
597 * groups: {0-2},{1-3} {1-3},{0-2}
599 * NUMA-1 0-2 0-3 0-3 1-3
607 * Build the balance mask; it contains only those CPUs that can arrive at this
608 * group and should be considered to continue balancing.
610 * We do this during the group creation pass, therefore the group information
611 * isn't complete yet, however since each group represents a (child) domain we
612 * can fully construct this using the sched_domain bits (which are already
616 build_balance_mask(struct sched_domain
*sd
, struct sched_group
*sg
, struct cpumask
*mask
)
618 const struct cpumask
*sg_span
= sched_group_span(sg
);
619 struct sd_data
*sdd
= sd
->private;
620 struct sched_domain
*sibling
;
625 for_each_cpu(i
, sg_span
) {
626 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
629 * Can happen in the asymmetric case, where these siblings are
630 * unused. The mask will not be empty because those CPUs that
631 * do have the top domain _should_ span the domain.
636 /* If we would not end up here, we can't continue from here */
637 if (!cpumask_equal(sg_span
, sched_domain_span(sibling
->child
)))
640 cpumask_set_cpu(i
, mask
);
643 /* We must not have empty masks here */
644 WARN_ON_ONCE(cpumask_empty(mask
));
648 * XXX: This creates per-node group entries; since the load-balancer will
649 * immediately access remote memory to construct this group's load-balance
650 * statistics having the groups node local is of dubious benefit.
652 static struct sched_group
*
653 build_group_from_child_sched_domain(struct sched_domain
*sd
, int cpu
)
655 struct sched_group
*sg
;
656 struct cpumask
*sg_span
;
658 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
659 GFP_KERNEL
, cpu_to_node(cpu
));
664 sg_span
= sched_group_span(sg
);
666 cpumask_copy(sg_span
, sched_domain_span(sd
->child
));
668 cpumask_copy(sg_span
, sched_domain_span(sd
));
670 atomic_inc(&sg
->ref
);
674 static void init_overlap_sched_group(struct sched_domain
*sd
,
675 struct sched_group
*sg
)
677 struct cpumask
*mask
= sched_domains_tmpmask2
;
678 struct sd_data
*sdd
= sd
->private;
679 struct cpumask
*sg_span
;
682 build_balance_mask(sd
, sg
, mask
);
683 cpu
= cpumask_first_and(sched_group_span(sg
), mask
);
685 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
686 if (atomic_inc_return(&sg
->sgc
->ref
) == 1)
687 cpumask_copy(group_balance_mask(sg
), mask
);
689 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg
), mask
));
692 * Initialize sgc->capacity such that even if we mess up the
693 * domains and no possible iteration will get us here, we won't
696 sg_span
= sched_group_span(sg
);
697 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sg_span
);
698 sg
->sgc
->min_capacity
= SCHED_CAPACITY_SCALE
;
702 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
704 struct sched_group
*first
= NULL
, *last
= NULL
, *sg
;
705 const struct cpumask
*span
= sched_domain_span(sd
);
706 struct cpumask
*covered
= sched_domains_tmpmask
;
707 struct sd_data
*sdd
= sd
->private;
708 struct sched_domain
*sibling
;
711 cpumask_clear(covered
);
713 for_each_cpu_wrap(i
, span
, cpu
) {
714 struct cpumask
*sg_span
;
716 if (cpumask_test_cpu(i
, covered
))
719 sibling
= *per_cpu_ptr(sdd
->sd
, i
);
722 * Asymmetric node setups can result in situations where the
723 * domain tree is of unequal depth, make sure to skip domains
724 * that already cover the entire range.
726 * In that case build_sched_domains() will have terminated the
727 * iteration early and our sibling sd spans will be empty.
728 * Domains should always include the CPU they're built on, so
731 if (!cpumask_test_cpu(i
, sched_domain_span(sibling
)))
734 sg
= build_group_from_child_sched_domain(sibling
, cpu
);
738 sg_span
= sched_group_span(sg
);
739 cpumask_or(covered
, covered
, sg_span
);
741 init_overlap_sched_group(sd
, sg
);
755 free_sched_groups(first
, 0);
762 * Package topology (also see the load-balance blurb in fair.c)
764 * The scheduler builds a tree structure to represent a number of important
765 * topology features. By default (default_topology[]) these include:
767 * - Simultaneous multithreading (SMT)
768 * - Multi-Core Cache (MC)
771 * Where the last one more or less denotes everything up to a NUMA node.
773 * The tree consists of 3 primary data structures:
775 * sched_domain -> sched_group -> sched_group_capacity
779 * The sched_domains are per-cpu and have a two way link (parent & child) and
780 * denote the ever growing mask of CPUs belonging to that level of topology.
782 * Each sched_domain has a circular (double) linked list of sched_group's, each
783 * denoting the domains of the level below (or individual CPUs in case of the
784 * first domain level). The sched_group linked by a sched_domain includes the
785 * CPU of that sched_domain [*].
787 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
789 * CPU 0 1 2 3 4 5 6 7
793 * SMT [ ] [ ] [ ] [ ]
797 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
798 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
799 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
801 * CPU 0 1 2 3 4 5 6 7
803 * One way to think about it is: sched_domain moves you up and down among these
804 * topology levels, while sched_group moves you sideways through it, at child
805 * domain granularity.
807 * sched_group_capacity ensures each unique sched_group has shared storage.
809 * There are two related construction problems, both require a CPU that
810 * uniquely identify each group (for a given domain):
812 * - The first is the balance_cpu (see should_we_balance() and the
813 * load-balance blub in fair.c); for each group we only want 1 CPU to
814 * continue balancing at a higher domain.
816 * - The second is the sched_group_capacity; we want all identical groups
817 * to share a single sched_group_capacity.
819 * Since these topologies are exclusive by construction. That is, its
820 * impossible for an SMT thread to belong to multiple cores, and cores to
821 * be part of multiple caches. There is a very clear and unique location
822 * for each CPU in the hierarchy.
824 * Therefore computing a unique CPU for each group is trivial (the iteration
825 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
826 * group), we can simply pick the first CPU in each group.
829 * [*] in other words, the first group of each domain is its child domain.
832 static struct sched_group
*get_group(int cpu
, struct sd_data
*sdd
)
834 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
835 struct sched_domain
*child
= sd
->child
;
836 struct sched_group
*sg
;
839 cpu
= cpumask_first(sched_domain_span(child
));
841 sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
842 sg
->sgc
= *per_cpu_ptr(sdd
->sgc
, cpu
);
844 /* For claim_allocations: */
845 atomic_inc(&sg
->ref
);
846 atomic_inc(&sg
->sgc
->ref
);
849 cpumask_copy(sched_group_span(sg
), sched_domain_span(child
));
850 cpumask_copy(group_balance_mask(sg
), sched_group_span(sg
));
852 cpumask_set_cpu(cpu
, sched_group_span(sg
));
853 cpumask_set_cpu(cpu
, group_balance_mask(sg
));
856 sg
->sgc
->capacity
= SCHED_CAPACITY_SCALE
* cpumask_weight(sched_group_span(sg
));
857 sg
->sgc
->min_capacity
= SCHED_CAPACITY_SCALE
;
863 * build_sched_groups will build a circular linked list of the groups
864 * covered by the given span, and will set each group's ->cpumask correctly,
865 * and ->cpu_capacity to 0.
867 * Assumes the sched_domain tree is fully constructed
870 build_sched_groups(struct sched_domain
*sd
, int cpu
)
872 struct sched_group
*first
= NULL
, *last
= NULL
;
873 struct sd_data
*sdd
= sd
->private;
874 const struct cpumask
*span
= sched_domain_span(sd
);
875 struct cpumask
*covered
;
878 lockdep_assert_held(&sched_domains_mutex
);
879 covered
= sched_domains_tmpmask
;
881 cpumask_clear(covered
);
883 for_each_cpu_wrap(i
, span
, cpu
) {
884 struct sched_group
*sg
;
886 if (cpumask_test_cpu(i
, covered
))
889 sg
= get_group(i
, sdd
);
891 cpumask_or(covered
, covered
, sched_group_span(sg
));
906 * Initialize sched groups cpu_capacity.
908 * cpu_capacity indicates the capacity of sched group, which is used while
909 * distributing the load between different sched groups in a sched domain.
910 * Typically cpu_capacity for all the groups in a sched domain will be same
911 * unless there are asymmetries in the topology. If there are asymmetries,
912 * group having more cpu_capacity will pickup more load compared to the
913 * group having less cpu_capacity.
915 static void init_sched_groups_capacity(int cpu
, struct sched_domain
*sd
)
917 struct sched_group
*sg
= sd
->groups
;
922 int cpu
, max_cpu
= -1;
924 sg
->group_weight
= cpumask_weight(sched_group_span(sg
));
926 if (!(sd
->flags
& SD_ASYM_PACKING
))
929 for_each_cpu(cpu
, sched_group_span(sg
)) {
932 else if (sched_asym_prefer(cpu
, max_cpu
))
935 sg
->asym_prefer_cpu
= max_cpu
;
939 } while (sg
!= sd
->groups
);
941 if (cpu
!= group_balance_cpu(sg
))
944 update_group_capacity(sd
, cpu
);
948 * Initializers for schedule domains
949 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
952 static int default_relax_domain_level
= -1;
953 int sched_domain_level_max
;
955 static int __init
setup_relax_domain_level(char *str
)
957 if (kstrtoint(str
, 0, &default_relax_domain_level
))
958 pr_warn("Unable to set relax_domain_level\n");
962 __setup("relax_domain_level=", setup_relax_domain_level
);
964 static void set_domain_attribute(struct sched_domain
*sd
,
965 struct sched_domain_attr
*attr
)
969 if (!attr
|| attr
->relax_domain_level
< 0) {
970 if (default_relax_domain_level
< 0)
973 request
= default_relax_domain_level
;
975 request
= attr
->relax_domain_level
;
976 if (request
< sd
->level
) {
977 /* Turn off idle balance on this domain: */
978 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
980 /* Turn on idle balance on this domain: */
981 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
985 static void __sdt_free(const struct cpumask
*cpu_map
);
986 static int __sdt_alloc(const struct cpumask
*cpu_map
);
988 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
989 const struct cpumask
*cpu_map
)
993 if (!atomic_read(&d
->rd
->refcount
))
994 free_rootdomain(&d
->rd
->rcu
);
1000 __sdt_free(cpu_map
);
1008 __visit_domain_allocation_hell(struct s_data
*d
, const struct cpumask
*cpu_map
)
1010 memset(d
, 0, sizeof(*d
));
1012 if (__sdt_alloc(cpu_map
))
1013 return sa_sd_storage
;
1014 d
->sd
= alloc_percpu(struct sched_domain
*);
1016 return sa_sd_storage
;
1017 d
->rd
= alloc_rootdomain();
1020 return sa_rootdomain
;
1024 * NULL the sd_data elements we've used to build the sched_domain and
1025 * sched_group structure so that the subsequent __free_domain_allocs()
1026 * will not free the data we're using.
1028 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
1030 struct sd_data
*sdd
= sd
->private;
1032 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
1033 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
1035 if (atomic_read(&(*per_cpu_ptr(sdd
->sds
, cpu
))->ref
))
1036 *per_cpu_ptr(sdd
->sds
, cpu
) = NULL
;
1038 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
1039 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
1041 if (atomic_read(&(*per_cpu_ptr(sdd
->sgc
, cpu
))->ref
))
1042 *per_cpu_ptr(sdd
->sgc
, cpu
) = NULL
;
1046 static int sched_domains_numa_levels
;
1047 enum numa_topology_type sched_numa_topology_type
;
1048 static int *sched_domains_numa_distance
;
1049 int sched_max_numa_distance
;
1050 static struct cpumask
***sched_domains_numa_masks
;
1051 static int sched_domains_curr_level
;
1055 * SD_flags allowed in topology descriptions.
1057 * These flags are purely descriptive of the topology and do not prescribe
1058 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1061 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1062 * SD_SHARE_PKG_RESOURCES - describes shared caches
1063 * SD_NUMA - describes NUMA topologies
1064 * SD_SHARE_POWERDOMAIN - describes shared power domain
1065 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
1067 * Odd one out, which beside describing the topology has a quirk also
1068 * prescribes the desired behaviour that goes along with it:
1070 * SD_ASYM_PACKING - describes SMT quirks
1072 #define TOPOLOGY_SD_FLAGS \
1073 (SD_SHARE_CPUCAPACITY | \
1074 SD_SHARE_PKG_RESOURCES | \
1077 SD_ASYM_CPUCAPACITY | \
1078 SD_SHARE_POWERDOMAIN)
1080 static struct sched_domain
*
1081 sd_init(struct sched_domain_topology_level
*tl
,
1082 const struct cpumask
*cpu_map
,
1083 struct sched_domain
*child
, int cpu
)
1085 struct sd_data
*sdd
= &tl
->data
;
1086 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
1087 int sd_id
, sd_weight
, sd_flags
= 0;
1091 * Ugly hack to pass state to sd_numa_mask()...
1093 sched_domains_curr_level
= tl
->numa_level
;
1096 sd_weight
= cpumask_weight(tl
->mask(cpu
));
1099 sd_flags
= (*tl
->sd_flags
)();
1100 if (WARN_ONCE(sd_flags
& ~TOPOLOGY_SD_FLAGS
,
1101 "wrong sd_flags in topology description\n"))
1102 sd_flags
&= ~TOPOLOGY_SD_FLAGS
;
1104 *sd
= (struct sched_domain
){
1105 .min_interval
= sd_weight
,
1106 .max_interval
= 2*sd_weight
,
1108 .imbalance_pct
= 125,
1110 .cache_nice_tries
= 0,
1117 .flags
= 1*SD_LOAD_BALANCE
1118 | 1*SD_BALANCE_NEWIDLE
1123 | 0*SD_SHARE_CPUCAPACITY
1124 | 0*SD_SHARE_PKG_RESOURCES
1126 | 0*SD_PREFER_SIBLING
1131 .last_balance
= jiffies
,
1132 .balance_interval
= sd_weight
,
1134 .max_newidle_lb_cost
= 0,
1135 .next_decay_max_lb_cost
= jiffies
,
1137 #ifdef CONFIG_SCHED_DEBUG
1142 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
1143 sd_id
= cpumask_first(sched_domain_span(sd
));
1146 * Convert topological properties into behaviour.
1149 if (sd
->flags
& SD_ASYM_CPUCAPACITY
) {
1150 struct sched_domain
*t
= sd
;
1152 for_each_lower_domain(t
)
1153 t
->flags
|= SD_BALANCE_WAKE
;
1156 if (sd
->flags
& SD_SHARE_CPUCAPACITY
) {
1157 sd
->flags
|= SD_PREFER_SIBLING
;
1158 sd
->imbalance_pct
= 110;
1159 sd
->smt_gain
= 1178; /* ~15% */
1161 } else if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
1162 sd
->imbalance_pct
= 117;
1163 sd
->cache_nice_tries
= 1;
1167 } else if (sd
->flags
& SD_NUMA
) {
1168 sd
->cache_nice_tries
= 2;
1172 sd
->flags
|= SD_SERIALIZE
;
1173 if (sched_domains_numa_distance
[tl
->numa_level
] > RECLAIM_DISTANCE
) {
1174 sd
->flags
&= ~(SD_BALANCE_EXEC
|
1181 sd
->flags
|= SD_PREFER_SIBLING
;
1182 sd
->cache_nice_tries
= 1;
1188 * For all levels sharing cache; connect a sched_domain_shared
1191 if (sd
->flags
& SD_SHARE_PKG_RESOURCES
) {
1192 sd
->shared
= *per_cpu_ptr(sdd
->sds
, sd_id
);
1193 atomic_inc(&sd
->shared
->ref
);
1194 atomic_set(&sd
->shared
->nr_busy_cpus
, sd_weight
);
1203 * Topology list, bottom-up.
1205 static struct sched_domain_topology_level default_topology
[] = {
1206 #ifdef CONFIG_SCHED_SMT
1207 { cpu_smt_mask
, cpu_smt_flags
, SD_INIT_NAME(SMT
) },
1209 #ifdef CONFIG_SCHED_MC
1210 { cpu_coregroup_mask
, cpu_core_flags
, SD_INIT_NAME(MC
) },
1212 { cpu_cpu_mask
, SD_INIT_NAME(DIE
) },
1216 static struct sched_domain_topology_level
*sched_domain_topology
=
1219 #define for_each_sd_topology(tl) \
1220 for (tl = sched_domain_topology; tl->mask; tl++)
1222 void set_sched_topology(struct sched_domain_topology_level
*tl
)
1224 if (WARN_ON_ONCE(sched_smp_initialized
))
1227 sched_domain_topology
= tl
;
1232 static const struct cpumask
*sd_numa_mask(int cpu
)
1234 return sched_domains_numa_masks
[sched_domains_curr_level
][cpu_to_node(cpu
)];
1237 static void sched_numa_warn(const char *str
)
1239 static int done
= false;
1247 printk(KERN_WARNING
"ERROR: %s\n\n", str
);
1249 for (i
= 0; i
< nr_node_ids
; i
++) {
1250 printk(KERN_WARNING
" ");
1251 for (j
= 0; j
< nr_node_ids
; j
++)
1252 printk(KERN_CONT
"%02d ", node_distance(i
,j
));
1253 printk(KERN_CONT
"\n");
1255 printk(KERN_WARNING
"\n");
1258 bool find_numa_distance(int distance
)
1262 if (distance
== node_distance(0, 0))
1265 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
1266 if (sched_domains_numa_distance
[i
] == distance
)
1274 * A system can have three types of NUMA topology:
1275 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1276 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1277 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1279 * The difference between a glueless mesh topology and a backplane
1280 * topology lies in whether communication between not directly
1281 * connected nodes goes through intermediary nodes (where programs
1282 * could run), or through backplane controllers. This affects
1283 * placement of programs.
1285 * The type of topology can be discerned with the following tests:
1286 * - If the maximum distance between any nodes is 1 hop, the system
1287 * is directly connected.
1288 * - If for two nodes A and B, located N > 1 hops away from each other,
1289 * there is an intermediary node C, which is < N hops away from both
1290 * nodes A and B, the system is a glueless mesh.
1292 static void init_numa_topology_type(void)
1296 n
= sched_max_numa_distance
;
1298 if (sched_domains_numa_levels
<= 1) {
1299 sched_numa_topology_type
= NUMA_DIRECT
;
1303 for_each_online_node(a
) {
1304 for_each_online_node(b
) {
1305 /* Find two nodes furthest removed from each other. */
1306 if (node_distance(a
, b
) < n
)
1309 /* Is there an intermediary node between a and b? */
1310 for_each_online_node(c
) {
1311 if (node_distance(a
, c
) < n
&&
1312 node_distance(b
, c
) < n
) {
1313 sched_numa_topology_type
=
1319 sched_numa_topology_type
= NUMA_BACKPLANE
;
1325 void sched_init_numa(void)
1327 int next_distance
, curr_distance
= node_distance(0, 0);
1328 struct sched_domain_topology_level
*tl
;
1332 sched_domains_numa_distance
= kzalloc(sizeof(int) * nr_node_ids
, GFP_KERNEL
);
1333 if (!sched_domains_numa_distance
)
1337 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1338 * unique distances in the node_distance() table.
1340 * Assumes node_distance(0,j) includes all distances in
1341 * node_distance(i,j) in order to avoid cubic time.
1343 next_distance
= curr_distance
;
1344 for (i
= 0; i
< nr_node_ids
; i
++) {
1345 for (j
= 0; j
< nr_node_ids
; j
++) {
1346 for (k
= 0; k
< nr_node_ids
; k
++) {
1347 int distance
= node_distance(i
, k
);
1349 if (distance
> curr_distance
&&
1350 (distance
< next_distance
||
1351 next_distance
== curr_distance
))
1352 next_distance
= distance
;
1355 * While not a strong assumption it would be nice to know
1356 * about cases where if node A is connected to B, B is not
1357 * equally connected to A.
1359 if (sched_debug() && node_distance(k
, i
) != distance
)
1360 sched_numa_warn("Node-distance not symmetric");
1362 if (sched_debug() && i
&& !find_numa_distance(distance
))
1363 sched_numa_warn("Node-0 not representative");
1365 if (next_distance
!= curr_distance
) {
1366 sched_domains_numa_distance
[level
++] = next_distance
;
1367 sched_domains_numa_levels
= level
;
1368 curr_distance
= next_distance
;
1373 * In case of sched_debug() we verify the above assumption.
1383 * 'level' contains the number of unique distances, excluding the
1384 * identity distance node_distance(i,i).
1386 * The sched_domains_numa_distance[] array includes the actual distance
1391 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1392 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1393 * the array will contain less then 'level' members. This could be
1394 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1395 * in other functions.
1397 * We reset it to 'level' at the end of this function.
1399 sched_domains_numa_levels
= 0;
1401 sched_domains_numa_masks
= kzalloc(sizeof(void *) * level
, GFP_KERNEL
);
1402 if (!sched_domains_numa_masks
)
1406 * Now for each level, construct a mask per node which contains all
1407 * CPUs of nodes that are that many hops away from us.
1409 for (i
= 0; i
< level
; i
++) {
1410 sched_domains_numa_masks
[i
] =
1411 kzalloc(nr_node_ids
* sizeof(void *), GFP_KERNEL
);
1412 if (!sched_domains_numa_masks
[i
])
1415 for (j
= 0; j
< nr_node_ids
; j
++) {
1416 struct cpumask
*mask
= kzalloc(cpumask_size(), GFP_KERNEL
);
1420 sched_domains_numa_masks
[i
][j
] = mask
;
1423 if (node_distance(j
, k
) > sched_domains_numa_distance
[i
])
1426 cpumask_or(mask
, mask
, cpumask_of_node(k
));
1431 /* Compute default topology size */
1432 for (i
= 0; sched_domain_topology
[i
].mask
; i
++);
1434 tl
= kzalloc((i
+ level
+ 1) *
1435 sizeof(struct sched_domain_topology_level
), GFP_KERNEL
);
1440 * Copy the default topology bits..
1442 for (i
= 0; sched_domain_topology
[i
].mask
; i
++)
1443 tl
[i
] = sched_domain_topology
[i
];
1446 * .. and append 'j' levels of NUMA goodness.
1448 for (j
= 0; j
< level
; i
++, j
++) {
1449 tl
[i
] = (struct sched_domain_topology_level
){
1450 .mask
= sd_numa_mask
,
1451 .sd_flags
= cpu_numa_flags
,
1452 .flags
= SDTL_OVERLAP
,
1458 sched_domain_topology
= tl
;
1460 sched_domains_numa_levels
= level
;
1461 sched_max_numa_distance
= sched_domains_numa_distance
[level
- 1];
1463 init_numa_topology_type();
1466 void sched_domains_numa_masks_set(unsigned int cpu
)
1468 int node
= cpu_to_node(cpu
);
1471 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
1472 for (j
= 0; j
< nr_node_ids
; j
++) {
1473 if (node_distance(j
, node
) <= sched_domains_numa_distance
[i
])
1474 cpumask_set_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
1479 void sched_domains_numa_masks_clear(unsigned int cpu
)
1483 for (i
= 0; i
< sched_domains_numa_levels
; i
++) {
1484 for (j
= 0; j
< nr_node_ids
; j
++)
1485 cpumask_clear_cpu(cpu
, sched_domains_numa_masks
[i
][j
]);
1489 #endif /* CONFIG_NUMA */
1491 static int __sdt_alloc(const struct cpumask
*cpu_map
)
1493 struct sched_domain_topology_level
*tl
;
1496 for_each_sd_topology(tl
) {
1497 struct sd_data
*sdd
= &tl
->data
;
1499 sdd
->sd
= alloc_percpu(struct sched_domain
*);
1503 sdd
->sds
= alloc_percpu(struct sched_domain_shared
*);
1507 sdd
->sg
= alloc_percpu(struct sched_group
*);
1511 sdd
->sgc
= alloc_percpu(struct sched_group_capacity
*);
1515 for_each_cpu(j
, cpu_map
) {
1516 struct sched_domain
*sd
;
1517 struct sched_domain_shared
*sds
;
1518 struct sched_group
*sg
;
1519 struct sched_group_capacity
*sgc
;
1521 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
1522 GFP_KERNEL
, cpu_to_node(j
));
1526 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
1528 sds
= kzalloc_node(sizeof(struct sched_domain_shared
),
1529 GFP_KERNEL
, cpu_to_node(j
));
1533 *per_cpu_ptr(sdd
->sds
, j
) = sds
;
1535 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
1536 GFP_KERNEL
, cpu_to_node(j
));
1542 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
1544 sgc
= kzalloc_node(sizeof(struct sched_group_capacity
) + cpumask_size(),
1545 GFP_KERNEL
, cpu_to_node(j
));
1549 #ifdef CONFIG_SCHED_DEBUG
1553 *per_cpu_ptr(sdd
->sgc
, j
) = sgc
;
1560 static void __sdt_free(const struct cpumask
*cpu_map
)
1562 struct sched_domain_topology_level
*tl
;
1565 for_each_sd_topology(tl
) {
1566 struct sd_data
*sdd
= &tl
->data
;
1568 for_each_cpu(j
, cpu_map
) {
1569 struct sched_domain
*sd
;
1572 sd
= *per_cpu_ptr(sdd
->sd
, j
);
1573 if (sd
&& (sd
->flags
& SD_OVERLAP
))
1574 free_sched_groups(sd
->groups
, 0);
1575 kfree(*per_cpu_ptr(sdd
->sd
, j
));
1579 kfree(*per_cpu_ptr(sdd
->sds
, j
));
1581 kfree(*per_cpu_ptr(sdd
->sg
, j
));
1583 kfree(*per_cpu_ptr(sdd
->sgc
, j
));
1585 free_percpu(sdd
->sd
);
1587 free_percpu(sdd
->sds
);
1589 free_percpu(sdd
->sg
);
1591 free_percpu(sdd
->sgc
);
1596 static struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
1597 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
1598 struct sched_domain
*child
, int cpu
)
1600 struct sched_domain
*sd
= sd_init(tl
, cpu_map
, child
, cpu
);
1603 sd
->level
= child
->level
+ 1;
1604 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
1607 if (!cpumask_subset(sched_domain_span(child
),
1608 sched_domain_span(sd
))) {
1609 pr_err("BUG: arch topology borken\n");
1610 #ifdef CONFIG_SCHED_DEBUG
1611 pr_err(" the %s domain not a subset of the %s domain\n",
1612 child
->name
, sd
->name
);
1614 /* Fixup, ensure @sd has at least @child cpus. */
1615 cpumask_or(sched_domain_span(sd
),
1616 sched_domain_span(sd
),
1617 sched_domain_span(child
));
1621 set_domain_attribute(sd
, attr
);
1627 * Build sched domains for a given set of CPUs and attach the sched domains
1628 * to the individual CPUs
1631 build_sched_domains(const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
)
1633 enum s_alloc alloc_state
;
1634 struct sched_domain
*sd
;
1636 struct rq
*rq
= NULL
;
1637 int i
, ret
= -ENOMEM
;
1639 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
1640 if (alloc_state
!= sa_rootdomain
)
1643 /* Set up domains for CPUs specified by the cpu_map: */
1644 for_each_cpu(i
, cpu_map
) {
1645 struct sched_domain_topology_level
*tl
;
1648 for_each_sd_topology(tl
) {
1649 sd
= build_sched_domain(tl
, cpu_map
, attr
, sd
, i
);
1650 if (tl
== sched_domain_topology
)
1651 *per_cpu_ptr(d
.sd
, i
) = sd
;
1652 if (tl
->flags
& SDTL_OVERLAP
)
1653 sd
->flags
|= SD_OVERLAP
;
1654 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
1659 /* Build the groups for the domains */
1660 for_each_cpu(i
, cpu_map
) {
1661 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
1662 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
1663 if (sd
->flags
& SD_OVERLAP
) {
1664 if (build_overlap_sched_groups(sd
, i
))
1667 if (build_sched_groups(sd
, i
))
1673 /* Calculate CPU capacity for physical packages and nodes */
1674 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
1675 if (!cpumask_test_cpu(i
, cpu_map
))
1678 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
1679 claim_allocations(i
, sd
);
1680 init_sched_groups_capacity(i
, sd
);
1684 /* Attach the domains */
1686 for_each_cpu(i
, cpu_map
) {
1688 sd
= *per_cpu_ptr(d
.sd
, i
);
1690 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
1691 if (rq
->cpu_capacity_orig
> READ_ONCE(d
.rd
->max_cpu_capacity
))
1692 WRITE_ONCE(d
.rd
->max_cpu_capacity
, rq
->cpu_capacity_orig
);
1694 cpu_attach_domain(sd
, d
.rd
, i
);
1698 if (rq
&& sched_debug_enabled
) {
1699 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
1700 cpumask_pr_args(cpu_map
), rq
->rd
->max_cpu_capacity
);
1705 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
1709 /* Current sched domains: */
1710 static cpumask_var_t
*doms_cur
;
1712 /* Number of sched domains in 'doms_cur': */
1713 static int ndoms_cur
;
1715 /* Attribues of custom domains in 'doms_cur' */
1716 static struct sched_domain_attr
*dattr_cur
;
1719 * Special case: If a kmalloc() of a doms_cur partition (array of
1720 * cpumask) fails, then fallback to a single sched domain,
1721 * as determined by the single cpumask fallback_doms.
1723 static cpumask_var_t fallback_doms
;
1726 * arch_update_cpu_topology lets virtualized architectures update the
1727 * CPU core maps. It is supposed to return 1 if the topology changed
1728 * or 0 if it stayed the same.
1730 int __weak
arch_update_cpu_topology(void)
1735 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
1738 cpumask_var_t
*doms
;
1740 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
1743 for (i
= 0; i
< ndoms
; i
++) {
1744 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
1745 free_sched_domains(doms
, i
);
1752 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
1755 for (i
= 0; i
< ndoms
; i
++)
1756 free_cpumask_var(doms
[i
]);
1761 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
1762 * For now this just excludes isolated CPUs, but could be used to
1763 * exclude other special cases in the future.
1765 int sched_init_domains(const struct cpumask
*cpu_map
)
1769 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_KERNEL
);
1770 zalloc_cpumask_var(&sched_domains_tmpmask2
, GFP_KERNEL
);
1771 zalloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
1773 arch_update_cpu_topology();
1775 doms_cur
= alloc_sched_domains(ndoms_cur
);
1777 doms_cur
= &fallback_doms
;
1778 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
1779 err
= build_sched_domains(doms_cur
[0], NULL
);
1780 register_sched_domain_sysctl();
1786 * Detach sched domains from a group of CPUs specified in cpu_map
1787 * These CPUs will now be attached to the NULL domain
1789 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
1794 for_each_cpu(i
, cpu_map
)
1795 cpu_attach_domain(NULL
, &def_root_domain
, i
);
1799 /* handle null as "default" */
1800 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
1801 struct sched_domain_attr
*new, int idx_new
)
1803 struct sched_domain_attr tmp
;
1810 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
1811 new ? (new + idx_new
) : &tmp
,
1812 sizeof(struct sched_domain_attr
));
1816 * Partition sched domains as specified by the 'ndoms_new'
1817 * cpumasks in the array doms_new[] of cpumasks. This compares
1818 * doms_new[] to the current sched domain partitioning, doms_cur[].
1819 * It destroys each deleted domain and builds each new domain.
1821 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
1822 * The masks don't intersect (don't overlap.) We should setup one
1823 * sched domain for each mask. CPUs not in any of the cpumasks will
1824 * not be load balanced. If the same cpumask appears both in the
1825 * current 'doms_cur' domains and in the new 'doms_new', we can leave
1828 * The passed in 'doms_new' should be allocated using
1829 * alloc_sched_domains. This routine takes ownership of it and will
1830 * free_sched_domains it when done with it. If the caller failed the
1831 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
1832 * and partition_sched_domains() will fallback to the single partition
1833 * 'fallback_doms', it also forces the domains to be rebuilt.
1835 * If doms_new == NULL it will be replaced with cpu_online_mask.
1836 * ndoms_new == 0 is a special case for destroying existing domains,
1837 * and it will not create the default domain.
1839 * Call with hotplug lock held
1841 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
1842 struct sched_domain_attr
*dattr_new
)
1847 mutex_lock(&sched_domains_mutex
);
1849 /* Always unregister in case we don't destroy any domains: */
1850 unregister_sched_domain_sysctl();
1852 /* Let the architecture update CPU core mappings: */
1853 new_topology
= arch_update_cpu_topology();
1856 WARN_ON_ONCE(dattr_new
);
1858 doms_new
= alloc_sched_domains(1);
1861 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
1867 /* Destroy deleted domains: */
1868 for (i
= 0; i
< ndoms_cur
; i
++) {
1869 for (j
= 0; j
< n
&& !new_topology
; j
++) {
1870 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
1871 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
1874 /* No match - a current sched domain not in new doms_new[] */
1875 detach_destroy_domains(doms_cur
[i
]);
1883 doms_new
= &fallback_doms
;
1884 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
1887 /* Build new domains: */
1888 for (i
= 0; i
< ndoms_new
; i
++) {
1889 for (j
= 0; j
< n
&& !new_topology
; j
++) {
1890 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
1891 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
1894 /* No match - add a new doms_new */
1895 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
1900 /* Remember the new sched domains: */
1901 if (doms_cur
!= &fallback_doms
)
1902 free_sched_domains(doms_cur
, ndoms_cur
);
1905 doms_cur
= doms_new
;
1906 dattr_cur
= dattr_new
;
1907 ndoms_cur
= ndoms_new
;
1909 register_sched_domain_sysctl();
1911 mutex_unlock(&sched_domains_mutex
);