2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2008, Jeffrey Roberson <jeff@freebsd.org>
7 * Copyright (c) 2008 Nokia Corporation
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice unmodified, this list of conditions, and the following
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
21 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
22 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
23 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
24 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
25 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
26 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
27 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
28 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
29 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/sysctl.h>
41 #include <sys/ctype.h>
42 #include <sys/sysproto.h>
44 #include <sys/kernel.h>
46 #include <sys/malloc.h>
47 #include <sys/mutex.h>
50 #include <sys/refcount.h>
51 #include <sys/sched.h>
53 #include <sys/syscallsubr.h>
54 #include <sys/capsicum.h>
55 #include <sys/cpuset.h>
56 #include <sys/domainset.h>
58 #include <sys/queue.h>
59 #include <sys/libkern.h>
60 #include <sys/limits.h>
62 #include <sys/interrupt.h>
63 #include <sys/vmmeter.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_pageout.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_param.h>
72 #include <vm/vm_phys.h>
73 #include <vm/vm_pagequeue.h>
80 * cpusets provide a mechanism for creating and manipulating sets of
81 * processors for the purpose of constraining the scheduling of threads to
82 * specific processors.
84 * Each process belongs to an identified set, by default this is set 1. Each
85 * thread may further restrict the cpus it may run on to a subset of this
86 * named set. This creates an anonymous set which other threads and processes
87 * may not join by number.
89 * The named set is referred to herein as the 'base' set to avoid ambiguity.
90 * This set is usually a child of a 'root' set while the anonymous set may
91 * simply be referred to as a mask. In the syscall api these are referred to
92 * as the ROOT, CPUSET, and MASK levels where CPUSET is called 'base' here.
94 * Threads inherit their set from their creator whether it be anonymous or
95 * not. This means that anonymous sets are immutable because they may be
96 * shared. To modify an anonymous set a new set is created with the desired
97 * mask and the same parent as the existing anonymous set. This gives the
98 * illusion of each thread having a private mask.
100 * Via the syscall apis a user may ask to retrieve or modify the root, base,
101 * or mask that is discovered via a pid, tid, or setid. Modifying a set
102 * modifies all numbered and anonymous child sets to comply with the new mask.
103 * Modifying a pid or tid's mask applies only to that tid but must still
104 * exist within the assigned parent set.
106 * A thread may not be assigned to a group separate from other threads in
107 * the process. This is to remove ambiguity when the setid is queried with
108 * a pid argument. There is no other technical limitation.
110 * This somewhat complex arrangement is intended to make it easy for
111 * applications to query available processors and bind their threads to
112 * specific processors while also allowing administrators to dynamically
113 * reprovision by changing sets which apply to groups of processes.
115 * A simple application should not concern itself with sets at all and
116 * rather apply masks to its own threads via CPU_WHICH_TID and a -1 id
117 * meaning 'curthread'. It may query available cpus for that tid with a
118 * getaffinity call using (CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, ...).
121 LIST_HEAD(domainlist, domainset);
123 static uma_zone_t cpuset_zone;
124 static uma_zone_t domainset_zone;
125 static struct mtx cpuset_lock;
126 static struct setlist cpuset_ids;
127 static struct domainlist cpuset_domains;
128 static struct unrhdr *cpuset_unr;
129 static struct cpuset *cpuset_zero, *cpuset_default, *cpuset_kernel;
130 static struct domainset domainset0, domainset2;
132 /* Return the size of cpuset_t at the kernel level */
133 SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD | CTLFLAG_CAPRD,
134 SYSCTL_NULL_INT_PTR, sizeof(cpuset_t), "sizeof(cpuset_t)");
136 cpuset_t *cpuset_root;
137 cpuset_t cpuset_domain[MAXMEMDOM];
139 static int domainset_valid(const struct domainset *, const struct domainset *);
142 * Find the first non-anonymous set starting from 'set'.
144 static struct cpuset *
145 cpuset_getbase(struct cpuset *set)
148 if (set->cs_id == CPUSET_INVALID)
149 set = set->cs_parent;
154 * Walks up the tree from 'set' to find the root.
156 static struct cpuset *
157 cpuset_getroot(struct cpuset *set)
160 while ((set->cs_flags & CPU_SET_ROOT) == 0 && set->cs_parent != NULL)
161 set = set->cs_parent;
166 * Acquire a reference to a cpuset, all pointers must be tracked with refs.
169 cpuset_ref(struct cpuset *set)
172 refcount_acquire(&set->cs_ref);
177 * Walks up the tree from 'set' to find the root. Returns the root
180 static struct cpuset *
181 cpuset_refroot(struct cpuset *set)
184 return (cpuset_ref(cpuset_getroot(set)));
188 * Find the first non-anonymous set starting from 'set'. Returns this set
189 * referenced. May return the passed in set with an extra ref if it is
192 static struct cpuset *
193 cpuset_refbase(struct cpuset *set)
196 return (cpuset_ref(cpuset_getbase(set)));
200 * Release a reference in a context where it is safe to allocate.
203 cpuset_rel(struct cpuset *set)
207 if (refcount_release(&set->cs_ref) == 0)
209 mtx_lock_spin(&cpuset_lock);
210 LIST_REMOVE(set, cs_siblings);
212 if (id != CPUSET_INVALID)
213 LIST_REMOVE(set, cs_link);
214 mtx_unlock_spin(&cpuset_lock);
215 cpuset_rel(set->cs_parent);
216 uma_zfree(cpuset_zone, set);
217 if (id != CPUSET_INVALID)
218 free_unr(cpuset_unr, id);
222 * Deferred release must be used when in a context that is not safe to
223 * allocate/free. This places any unreferenced sets on the list 'head'.
226 cpuset_rel_defer(struct setlist *head, struct cpuset *set)
229 if (refcount_release(&set->cs_ref) == 0)
231 mtx_lock_spin(&cpuset_lock);
232 LIST_REMOVE(set, cs_siblings);
233 if (set->cs_id != CPUSET_INVALID)
234 LIST_REMOVE(set, cs_link);
235 LIST_INSERT_HEAD(head, set, cs_link);
236 mtx_unlock_spin(&cpuset_lock);
240 * Complete a deferred release. Removes the set from the list provided to
244 cpuset_rel_complete(struct cpuset *set)
246 LIST_REMOVE(set, cs_link);
247 cpuset_rel(set->cs_parent);
248 uma_zfree(cpuset_zone, set);
252 * Find a set based on an id. Returns it with a ref.
254 static struct cpuset *
255 cpuset_lookup(cpusetid_t setid, struct thread *td)
259 if (setid == CPUSET_INVALID)
261 mtx_lock_spin(&cpuset_lock);
262 LIST_FOREACH(set, &cpuset_ids, cs_link)
263 if (set->cs_id == setid)
267 mtx_unlock_spin(&cpuset_lock);
269 KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__));
270 if (set != NULL && jailed(td->td_ucred)) {
271 struct cpuset *jset, *tset;
273 jset = td->td_ucred->cr_prison->pr_cpuset;
274 for (tset = set; tset != NULL; tset = tset->cs_parent)
287 * Create a set in the space provided in 'set' with the provided parameters.
288 * The set is returned with a single ref. May return EDEADLK if the set
289 * will have no valid cpu based on restrictions from the parent.
292 _cpuset_create(struct cpuset *set, struct cpuset *parent,
293 const cpuset_t *mask, struct domainset *domain, cpusetid_t id)
297 domain = parent->cs_domain;
299 mask = &parent->cs_mask;
300 if (!CPU_OVERLAP(&parent->cs_mask, mask))
302 /* The domain must be prepared ahead of time. */
303 if (!domainset_valid(parent->cs_domain, domain))
305 CPU_COPY(mask, &set->cs_mask);
306 LIST_INIT(&set->cs_children);
307 refcount_init(&set->cs_ref, 1);
309 mtx_lock_spin(&cpuset_lock);
310 set->cs_domain = domain;
311 CPU_AND(&set->cs_mask, &parent->cs_mask);
313 set->cs_parent = cpuset_ref(parent);
314 LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings);
315 if (set->cs_id != CPUSET_INVALID)
316 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
317 mtx_unlock_spin(&cpuset_lock);
323 * Create a new non-anonymous set with the requested parent and mask. May
324 * return failures if the mask is invalid or a new number can not be
328 cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask)
334 id = alloc_unr(cpuset_unr);
337 *setp = set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
338 error = _cpuset_create(set, parent, mask, NULL, id);
341 free_unr(cpuset_unr, id);
342 uma_zfree(cpuset_zone, set);
348 cpuset_freelist_add(struct setlist *list, int count)
353 for (i = 0; i < count; i++) {
354 set = uma_zalloc(cpuset_zone, M_ZERO | M_WAITOK);
355 LIST_INSERT_HEAD(list, set, cs_link);
360 cpuset_freelist_init(struct setlist *list, int count)
364 cpuset_freelist_add(list, count);
368 cpuset_freelist_free(struct setlist *list)
372 while ((set = LIST_FIRST(list)) != NULL) {
373 LIST_REMOVE(set, cs_link);
374 uma_zfree(cpuset_zone, set);
379 domainset_freelist_add(struct domainlist *list, int count)
381 struct domainset *set;
384 for (i = 0; i < count; i++) {
385 set = uma_zalloc(domainset_zone, M_ZERO | M_WAITOK);
386 LIST_INSERT_HEAD(list, set, ds_link);
391 domainset_freelist_init(struct domainlist *list, int count)
395 domainset_freelist_add(list, count);
399 domainset_freelist_free(struct domainlist *list)
401 struct domainset *set;
403 while ((set = LIST_FIRST(list)) != NULL) {
404 LIST_REMOVE(set, ds_link);
405 uma_zfree(domainset_zone, set);
409 /* Copy a domainset preserving mask and policy. */
411 domainset_copy(const struct domainset *from, struct domainset *to)
414 DOMAINSET_COPY(&from->ds_mask, &to->ds_mask);
415 to->ds_policy = from->ds_policy;
416 to->ds_prefer = from->ds_prefer;
419 /* Return 1 if mask and policy are equal, otherwise 0. */
421 domainset_equal(const struct domainset *one, const struct domainset *two)
424 return (DOMAINSET_CMP(&one->ds_mask, &two->ds_mask) == 0 &&
425 one->ds_policy == two->ds_policy &&
426 one->ds_prefer == two->ds_prefer);
429 /* Return 1 if child is a valid subset of parent. */
431 domainset_valid(const struct domainset *parent, const struct domainset *child)
433 if (child->ds_policy != DOMAINSET_POLICY_PREFER)
434 return (DOMAINSET_SUBSET(&parent->ds_mask, &child->ds_mask));
435 return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask));
439 domainset_restrict(const struct domainset *parent,
440 const struct domainset *child)
442 if (child->ds_policy != DOMAINSET_POLICY_PREFER)
443 return (DOMAINSET_OVERLAP(&parent->ds_mask, &child->ds_mask));
444 return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask));
448 * Lookup or create a domainset. The key is provided in ds_mask and
449 * ds_policy. If the domainset does not yet exist the storage in
450 * 'domain' is used to insert. Otherwise this storage is freed to the
451 * domainset_zone and the existing domainset is returned.
453 static struct domainset *
454 _domainset_create(struct domainset *domain, struct domainlist *freelist)
456 struct domainset *ndomain;
459 mtx_lock_spin(&cpuset_lock);
460 LIST_FOREACH(ndomain, &cpuset_domains, ds_link)
461 if (domainset_equal(ndomain, domain))
464 * If the domain does not yet exist we insert it and initialize
465 * various iteration helpers which are not part of the key.
467 if (ndomain == NULL) {
468 LIST_INSERT_HEAD(&cpuset_domains, domain, ds_link);
469 domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask);
470 max = DOMAINSET_FLS(&domain->ds_mask) + 1;
471 for (i = 0, j = 0; i < max; i++)
472 if (DOMAINSET_ISSET(i, &domain->ds_mask))
473 domain->ds_order[j++] = i;
475 mtx_unlock_spin(&cpuset_lock);
478 if (freelist != NULL)
479 LIST_INSERT_HEAD(freelist, domain, ds_link);
481 uma_zfree(domainset_zone, domain);
487 * Are any of the domains in the mask empty? If so, silently
488 * remove them. If only empty domains are present, we must
492 domainset_empty_vm(struct domainset *domain)
496 max = DOMAINSET_FLS(&domain->ds_mask) + 1;
497 for (i = 0; i < max; i++) {
498 if (DOMAINSET_ISSET(i, &domain->ds_mask) &&
500 DOMAINSET_CLR(i, &domain->ds_mask);
503 return (DOMAINSET_EMPTY(&domain->ds_mask));
507 * Create or lookup a domainset based on the key held in 'domain'.
510 domainset_create(const struct domainset *domain)
512 struct domainset *ndomain;
515 * Validate the policy. It must specify a useable policy number with
516 * only valid domains. Preferred must include the preferred domain
519 if (domain->ds_policy <= DOMAINSET_POLICY_INVALID ||
520 domain->ds_policy > DOMAINSET_POLICY_MAX)
522 if (domain->ds_policy == DOMAINSET_POLICY_PREFER &&
523 !DOMAINSET_ISSET(domain->ds_prefer, &domain->ds_mask))
525 if (!DOMAINSET_SUBSET(&domainset0.ds_mask, &domain->ds_mask))
527 ndomain = uma_zalloc(domainset_zone, M_WAITOK | M_ZERO);
528 domainset_copy(domain, ndomain);
529 return _domainset_create(ndomain, NULL);
533 * Update thread domainset pointers.
536 domainset_notify(void)
541 sx_slock(&allproc_lock);
542 FOREACH_PROC_IN_SYSTEM(p) {
544 if (p->p_state == PRS_NEW) {
548 FOREACH_THREAD_IN_PROC(p, td) {
550 td->td_domain.dr_policy = td->td_cpuset->cs_domain;
555 sx_sunlock(&allproc_lock);
556 kernel_object->domain.dr_policy = cpuset_kernel->cs_domain;
560 * Create a new set that is a subset of a parent.
562 static struct domainset *
563 domainset_shadow(const struct domainset *pdomain,
564 const struct domainset *domain, struct domainlist *freelist)
566 struct domainset *ndomain;
568 ndomain = LIST_FIRST(freelist);
569 LIST_REMOVE(ndomain, ds_link);
572 * Initialize the key from the request.
574 domainset_copy(domain, ndomain);
577 * Restrict the key by the parent.
579 DOMAINSET_AND(&ndomain->ds_mask, &pdomain->ds_mask);
581 return _domainset_create(ndomain, freelist);
585 * Recursively check for errors that would occur from applying mask to
586 * the tree of sets starting at 'set'. Checks for sets that would become
587 * empty as well as RDONLY flags.
590 cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int check_mask)
596 mtx_assert(&cpuset_lock, MA_OWNED);
597 if (set->cs_flags & CPU_SET_RDONLY)
600 if (!CPU_OVERLAP(&set->cs_mask, mask))
602 CPU_COPY(&set->cs_mask, &newmask);
603 CPU_AND(&newmask, mask);
605 CPU_COPY(mask, &newmask);
607 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
608 if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0)
614 * Applies the mask 'mask' without checking for empty sets or permissions.
617 cpuset_update(struct cpuset *set, cpuset_t *mask)
621 mtx_assert(&cpuset_lock, MA_OWNED);
622 CPU_AND(&set->cs_mask, mask);
623 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
624 cpuset_update(nset, &set->cs_mask);
630 * Modify the set 'set' to use a copy of the mask provided. Apply this new
631 * mask to restrict all children in the tree. Checks for validity before
632 * applying the changes.
635 cpuset_modify(struct cpuset *set, cpuset_t *mask)
640 error = priv_check(curthread, PRIV_SCHED_CPUSET);
644 * In case we are called from within the jail
645 * we do not allow modifying the dedicated root
646 * cpuset of the jail but may still allow to
649 if (jailed(curthread->td_ucred) &&
650 set->cs_flags & CPU_SET_ROOT)
653 * Verify that we have access to this set of
656 root = cpuset_getroot(set);
657 mtx_lock_spin(&cpuset_lock);
658 if (root && !CPU_SUBSET(&root->cs_mask, mask)) {
662 error = cpuset_testupdate(set, mask, 0);
665 CPU_COPY(mask, &set->cs_mask);
666 cpuset_update(set, mask);
668 mtx_unlock_spin(&cpuset_lock);
674 * Recursively check for errors that would occur from applying mask to
675 * the tree of sets starting at 'set'. Checks for sets that would become
676 * empty as well as RDONLY flags.
679 cpuset_testupdate_domain(struct cpuset *set, struct domainset *dset,
680 struct domainset *orig, int *count, int check_mask)
683 struct domainset *domain;
684 struct domainset newset;
687 mtx_assert(&cpuset_lock, MA_OWNED);
688 if (set->cs_flags & CPU_SET_RDONLY)
690 domain = set->cs_domain;
691 domainset_copy(domain, &newset);
692 if (!domainset_equal(domain, orig)) {
693 if (!domainset_restrict(domain, dset))
695 DOMAINSET_AND(&newset.ds_mask, &dset->ds_mask);
696 /* Count the number of domains that are changing. */
700 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
701 if ((error = cpuset_testupdate_domain(nset, &newset, domain,
708 * Applies the mask 'mask' without checking for empty sets or permissions.
711 cpuset_update_domain(struct cpuset *set, struct domainset *domain,
712 struct domainset *orig, struct domainlist *domains)
716 mtx_assert(&cpuset_lock, MA_OWNED);
718 * If this domainset has changed from the parent we must calculate
719 * a new set. Otherwise it simply inherits from the parent. When
720 * we inherit from the parent we get a new mask and policy. If the
721 * set is modified from the parent we keep the policy and only
724 if (set->cs_domain != orig) {
725 orig = set->cs_domain;
726 set->cs_domain = domainset_shadow(domain, orig, domains);
728 set->cs_domain = domain;
729 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
730 cpuset_update_domain(nset, set->cs_domain, orig, domains);
736 * Modify the set 'set' to use a copy the domainset provided. Apply this new
737 * mask to restrict all children in the tree. Checks for validity before
738 * applying the changes.
741 cpuset_modify_domain(struct cpuset *set, struct domainset *domain)
743 struct domainlist domains;
744 struct domainset temp;
745 struct domainset *dset;
747 int ndomains, needed;
750 error = priv_check(curthread, PRIV_SCHED_CPUSET);
754 * In case we are called from within the jail
755 * we do not allow modifying the dedicated root
756 * cpuset of the jail but may still allow to
759 if (jailed(curthread->td_ucred) &&
760 set->cs_flags & CPU_SET_ROOT)
762 domainset_freelist_init(&domains, 0);
763 domain = domainset_create(domain);
764 ndomains = needed = 0;
766 if (ndomains < needed) {
767 domainset_freelist_add(&domains, needed - ndomains);
770 root = cpuset_getroot(set);
771 mtx_lock_spin(&cpuset_lock);
772 dset = root->cs_domain;
774 * Verify that we have access to this set of domains.
776 if (root && !domainset_valid(dset, domain)) {
781 * If applying prefer we keep the current set as the fallback.
783 if (domain->ds_policy == DOMAINSET_POLICY_PREFER)
784 DOMAINSET_COPY(&set->cs_domain->ds_mask,
787 * Determine whether we can apply this set of domains and
788 * how many new domain structures it will require.
790 domainset_copy(domain, &temp);
792 error = cpuset_testupdate_domain(set, &temp, set->cs_domain,
796 } while (ndomains < needed);
797 dset = set->cs_domain;
798 cpuset_update_domain(set, domain, dset, &domains);
800 mtx_unlock_spin(&cpuset_lock);
801 domainset_freelist_free(&domains);
809 * Resolve the 'which' parameter of several cpuset apis.
811 * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid. Also
812 * checks for permission via p_cansched().
814 * For WHICH_SET returns a valid set with a new reference.
816 * -1 may be supplied for any argument to mean the current proc/thread or
817 * the base set of the current thread. May fail with ESRCH/EPERM.
820 cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp,
821 struct cpuset **setp)
838 if ((p = pfind(id)) == NULL)
853 case CPU_WHICH_CPUSET:
855 thread_lock(curthread);
856 set = cpuset_refbase(curthread->td_cpuset);
857 thread_unlock(curthread);
859 set = cpuset_lookup(id, curthread);
867 /* Find `set' for prison with given id. */
870 sx_slock(&allprison_lock);
871 pr = prison_find_child(curthread->td_ucred->cr_prison, id);
872 sx_sunlock(&allprison_lock);
875 cpuset_ref(pr->pr_cpuset);
876 *setp = pr->pr_cpuset;
877 mtx_unlock(&pr->pr_mtx);
881 case CPU_WHICH_DOMAIN:
886 error = p_cansched(curthread, p);
892 td = FIRST_THREAD_IN_PROC(p);
899 cpuset_testshadow(struct cpuset *set, const cpuset_t *mask,
900 const struct domainset *domain)
902 struct cpuset *parent;
903 struct domainset *dset;
905 parent = cpuset_getbase(set);
907 * If we are restricting a cpu mask it must be a subset of the
908 * parent or invalid CPUs have been specified.
910 if (mask != NULL && !CPU_SUBSET(&parent->cs_mask, mask))
914 * If we are restricting a domain mask it must be a subset of the
915 * parent or invalid domains have been specified.
917 dset = parent->cs_domain;
918 if (domain != NULL && !domainset_valid(dset, domain))
925 * Create an anonymous set with the provided mask in the space provided by
926 * 'nset'. If the passed in set is anonymous we use its parent otherwise
927 * the new set is a child of 'set'.
930 cpuset_shadow(struct cpuset *set, struct cpuset **nsetp,
931 const cpuset_t *mask, const struct domainset *domain,
932 struct setlist *cpusets, struct domainlist *domains)
934 struct cpuset *parent;
936 struct domainset *dset;
940 error = cpuset_testshadow(set, mask, domain);
944 parent = cpuset_getbase(set);
945 dset = parent->cs_domain;
947 mask = &set->cs_mask;
949 d = domainset_shadow(dset, domain, domains);
952 nset = LIST_FIRST(cpusets);
953 error = _cpuset_create(nset, parent, mask, d, CPUSET_INVALID);
955 LIST_REMOVE(nset, cs_link);
961 static struct cpuset *
962 cpuset_update_thread(struct thread *td, struct cpuset *nset)
964 struct cpuset *tdset;
966 tdset = td->td_cpuset;
967 td->td_cpuset = nset;
968 td->td_domain.dr_policy = nset->cs_domain;
975 cpuset_setproc_test_maskthread(struct cpuset *tdset, cpuset_t *mask,
976 struct domainset *domain)
978 struct cpuset *parent;
980 parent = cpuset_getbase(tdset);
982 mask = &tdset->cs_mask;
984 domain = tdset->cs_domain;
985 return cpuset_testshadow(parent, mask, domain);
989 cpuset_setproc_maskthread(struct cpuset *tdset, cpuset_t *mask,
990 struct domainset *domain, struct cpuset **nsetp,
991 struct setlist *freelist, struct domainlist *domainlist)
993 struct cpuset *parent;
995 parent = cpuset_getbase(tdset);
997 mask = &tdset->cs_mask;
999 domain = tdset->cs_domain;
1000 return cpuset_shadow(parent, nsetp, mask, domain, freelist,
1005 cpuset_setproc_setthread_mask(struct cpuset *tdset, struct cpuset *set,
1006 cpuset_t *mask, struct domainset *domain)
1008 struct cpuset *parent;
1010 parent = cpuset_getbase(tdset);
1013 * If the thread restricted its mask then apply that same
1014 * restriction to the new set, otherwise take it wholesale.
1016 if (CPU_CMP(&tdset->cs_mask, &parent->cs_mask) != 0) {
1017 CPU_COPY(&tdset->cs_mask, mask);
1018 CPU_AND(mask, &set->cs_mask);
1020 CPU_COPY(&set->cs_mask, mask);
1023 * If the thread restricted the domain then we apply the
1024 * restriction to the new set but retain the policy.
1026 if (tdset->cs_domain != parent->cs_domain) {
1027 domainset_copy(tdset->cs_domain, domain);
1028 DOMAINSET_AND(&domain->ds_mask, &set->cs_domain->ds_mask);
1030 domainset_copy(set->cs_domain, domain);
1032 if (CPU_EMPTY(mask) || DOMAINSET_EMPTY(&domain->ds_mask))
1039 cpuset_setproc_test_setthread(struct cpuset *tdset, struct cpuset *set)
1041 struct domainset domain;
1044 if (tdset->cs_id != CPUSET_INVALID)
1046 return cpuset_setproc_setthread_mask(tdset, set, &mask, &domain);
1050 cpuset_setproc_setthread(struct cpuset *tdset, struct cpuset *set,
1051 struct cpuset **nsetp, struct setlist *freelist,
1052 struct domainlist *domainlist)
1054 struct domainset domain;
1059 * If we're replacing on a thread that has not constrained the
1060 * original set we can simply accept the new set.
1062 if (tdset->cs_id != CPUSET_INVALID) {
1063 *nsetp = cpuset_ref(set);
1066 error = cpuset_setproc_setthread_mask(tdset, set, &mask, &domain);
1070 return cpuset_shadow(tdset, nsetp, &mask, &domain, freelist,
1075 * Handle three cases for updating an entire process.
1077 * 1) Set is non-null. This reparents all anonymous sets to the provided
1078 * set and replaces all non-anonymous td_cpusets with the provided set.
1079 * 2) Mask is non-null. This replaces or creates anonymous sets for every
1080 * thread with the existing base as a parent.
1081 * 3) domain is non-null. This creates anonymous sets for every thread
1082 * and replaces the domain set.
1084 * This is overly complicated because we can't allocate while holding a
1085 * spinlock and spinlocks must be held while changing and examining thread
1089 cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask,
1090 struct domainset *domain)
1092 struct setlist freelist;
1093 struct setlist droplist;
1094 struct domainlist domainlist;
1095 struct cpuset *nset;
1103 * The algorithm requires two passes due to locking considerations.
1105 * 1) Lookup the process and acquire the locks in the required order.
1106 * 2) If enough cpusets have not been allocated release the locks and
1107 * allocate them. Loop.
1109 cpuset_freelist_init(&freelist, 1);
1110 domainset_freelist_init(&domainlist, 1);
1112 LIST_INIT(&droplist);
1115 error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset);
1118 if (nfree >= p->p_numthreads)
1120 threads = p->p_numthreads;
1122 if (nfree < threads) {
1123 cpuset_freelist_add(&freelist, threads - nfree);
1124 domainset_freelist_add(&domainlist, threads - nfree);
1128 PROC_LOCK_ASSERT(p, MA_OWNED);
1130 * Now that the appropriate locks are held and we have enough cpusets,
1131 * make sure the operation will succeed before applying changes. The
1132 * proc lock prevents td_cpuset from changing between calls.
1135 FOREACH_THREAD_IN_PROC(p, td) {
1138 error = cpuset_setproc_test_setthread(td->td_cpuset,
1141 error = cpuset_setproc_test_maskthread(td->td_cpuset,
1148 * Replace each thread's cpuset while using deferred release. We
1149 * must do this because the thread lock must be held while operating
1150 * on the thread and this limits the type of operations allowed.
1152 FOREACH_THREAD_IN_PROC(p, td) {
1155 error = cpuset_setproc_setthread(td->td_cpuset, set,
1156 &nset, &freelist, &domainlist);
1158 error = cpuset_setproc_maskthread(td->td_cpuset, mask,
1159 domain, &nset, &freelist, &domainlist);
1164 cpuset_rel_defer(&droplist, cpuset_update_thread(td, nset));
1170 while ((nset = LIST_FIRST(&droplist)) != NULL)
1171 cpuset_rel_complete(nset);
1172 cpuset_freelist_free(&freelist);
1173 domainset_freelist_free(&domainlist);
1178 bitset_strprint(char *buf, size_t bufsiz, const struct bitset *set, int setlen)
1186 for (i = 0; i < __bitset_words(setlen); i++) {
1195 if (bufsiz < sizeof(__STRING(ULONG_MAX)))
1197 bytes = snprintf(p, bufsiz, "%lx", set->__bits[i]);
1205 bitset_strscan(struct bitset *set, int setlen, const char *buf)
1210 BIT_ZERO(setlen, set);
1212 for (i = 0; i < __bitset_words(setlen); i++) {
1217 ret = sscanf(p, "%lx", &set->__bits[i]);
1218 if (ret == 0 || ret == -1)
1220 while (isxdigit(*p))
1227 * Return a string representing a valid layout for a cpuset_t object.
1228 * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
1231 cpusetobj_strprint(char *buf, const cpuset_t *set)
1234 bitset_strprint(buf, CPUSETBUFSIZ, (const struct bitset *)set,
1240 * Build a valid cpuset_t object from a string representation.
1241 * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
1244 cpusetobj_strscan(cpuset_t *set, const char *buf)
1248 if (strlen(buf) > CPUSETBUFSIZ - 1)
1251 p = buf[bitset_strscan((struct bitset *)set, CPU_SETSIZE, buf)];
1259 * Handle a domainset specifier in the sysctl tree. A poiner to a pointer to
1260 * a domainset is in arg1. If the user specifies a valid domainset the
1261 * pointer is updated.
1264 * hex mask word 0,hex mask word 1,...:decimal policy:decimal preferred
1267 sysctl_handle_domainset(SYSCTL_HANDLER_ARGS)
1269 char buf[DOMAINSETBUFSIZ];
1270 struct domainset *dset;
1271 struct domainset key;
1272 int policy, prefer, error;
1275 dset = *(struct domainset **)arg1;
1279 p = buf + bitset_strprint(buf, DOMAINSETBUFSIZ,
1280 (const struct bitset *)&dset->ds_mask, DOMAINSET_SETSIZE);
1281 sprintf(p, ":%d:%d", dset->ds_policy, dset->ds_prefer);
1283 sprintf(buf, "<NULL>");
1284 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
1285 if (error != 0 || req->newptr == NULL)
1289 * Read in and validate the string.
1291 memset(&key, 0, sizeof(key));
1292 p = &buf[bitset_strscan((struct bitset *)&key.ds_mask,
1293 DOMAINSET_SETSIZE, buf)];
1296 if (sscanf(p, ":%d:%d", &policy, &prefer) != 2)
1298 key.ds_policy = policy;
1299 key.ds_prefer = prefer;
1301 /* Domainset_create() validates the policy.*/
1302 dset = domainset_create(&key);
1305 *(struct domainset **)arg1 = dset;
1311 * Apply an anonymous mask or a domain to a single thread.
1314 _cpuset_setthread(lwpid_t id, cpuset_t *mask, struct domainset *domain)
1316 struct setlist cpusets;
1317 struct domainlist domainlist;
1318 struct cpuset *nset;
1324 cpuset_freelist_init(&cpusets, 1);
1325 domainset_freelist_init(&domainlist, domain != NULL);
1326 error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set);
1331 error = cpuset_shadow(td->td_cpuset, &nset, mask, domain,
1332 &cpusets, &domainlist);
1334 set = cpuset_update_thread(td, nset);
1340 cpuset_freelist_free(&cpusets);
1341 domainset_freelist_free(&domainlist);
1346 * Apply an anonymous mask to a single thread.
1349 cpuset_setthread(lwpid_t id, cpuset_t *mask)
1352 return _cpuset_setthread(id, mask, NULL);
1356 * Apply new cpumask to the ithread.
1359 cpuset_setithread(lwpid_t id, int cpu)
1365 CPU_COPY(cpuset_root, &mask);
1367 CPU_SET(cpu, &mask);
1368 return _cpuset_setthread(id, &mask, NULL);
1372 * Create the domainset for cpuset 0, 1 and cpuset 2.
1375 domainset_zero(void)
1377 struct domainset *dset;
1380 mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE);
1383 DOMAINSET_ZERO(&dset->ds_mask);
1384 for (i = 0; i < vm_ndomains; i++)
1385 DOMAINSET_SET(i, &dset->ds_mask);
1386 dset->ds_policy = DOMAINSET_POLICY_FIRSTTOUCH;
1387 dset->ds_prefer = -1;
1388 (void)domainset_empty_vm(dset);
1389 curthread->td_domain.dr_policy = _domainset_create(dset, NULL);
1391 domainset_copy(dset, &domainset2);
1392 domainset2.ds_policy = DOMAINSET_POLICY_INTERLEAVE;
1393 kernel_object->domain.dr_policy = _domainset_create(&domainset2, NULL);
1397 * Creates system-wide cpusets and the cpuset for thread0 including three
1400 * 0 - The root set which should represent all valid processors in the
1401 * system. It is initially created with a mask of all processors
1402 * because we don't know what processors are valid until cpuset_init()
1403 * runs. This set is immutable.
1404 * 1 - The default set which all processes are a member of until changed.
1405 * This allows an administrator to move all threads off of given cpus to
1406 * dedicate them to high priority tasks or save power etc.
1407 * 2 - The kernel set which allows restriction and policy to be applied only
1408 * to kernel threads and the kernel_object.
1411 cpuset_thread0(void)
1417 cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL,
1418 NULL, NULL, UMA_ALIGN_CACHE, 0);
1419 domainset_zone = uma_zcreate("domainset", sizeof(struct domainset),
1420 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
1423 * Create the root system set (0) for the whole machine. Doesn't use
1424 * cpuset_create() due to NULL parent.
1426 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1427 CPU_COPY(&all_cpus, &set->cs_mask);
1428 LIST_INIT(&set->cs_children);
1429 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
1431 set->cs_flags = CPU_SET_ROOT | CPU_SET_RDONLY;
1432 set->cs_domain = &domainset0;
1434 cpuset_root = &set->cs_mask;
1437 * Now derive a default (1), modifiable set from that to give out.
1439 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1440 error = _cpuset_create(set, cpuset_zero, NULL, NULL, 1);
1441 KASSERT(error == 0, ("Error creating default set: %d\n", error));
1442 cpuset_default = set;
1444 * Create the kernel set (2).
1446 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1447 error = _cpuset_create(set, cpuset_zero, NULL, NULL, 2);
1448 KASSERT(error == 0, ("Error creating kernel set: %d\n", error));
1449 set->cs_domain = &domainset2;
1450 cpuset_kernel = set;
1453 * Initialize the unit allocator. 0 and 1 are allocated above.
1455 cpuset_unr = new_unrhdr(2, INT_MAX, NULL);
1458 * If MD code has not initialized per-domain cpusets, place all
1461 for (i = 0; i < MAXMEMDOM; i++)
1462 if (!CPU_EMPTY(&cpuset_domain[i]))
1464 CPU_COPY(&all_cpus, &cpuset_domain[0]);
1467 return (cpuset_default);
1471 cpuset_kernthread(struct thread *td)
1476 set = td->td_cpuset;
1477 td->td_cpuset = cpuset_ref(cpuset_kernel);
1483 * Create a cpuset, which would be cpuset_create() but
1484 * mark the new 'set' as root.
1486 * We are not going to reparent the td to it. Use cpuset_setproc_update_set()
1489 * In case of no error, returns the set in *setp locked with a reference.
1492 cpuset_create_root(struct prison *pr, struct cpuset **setp)
1497 KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__));
1498 KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__));
1500 error = cpuset_create(setp, pr->pr_cpuset, &pr->pr_cpuset->cs_mask);
1504 KASSERT(*setp != NULL, ("[%s:%d] cpuset_create returned invalid data",
1505 __func__, __LINE__));
1507 /* Mark the set as root. */
1509 set->cs_flags |= CPU_SET_ROOT;
1515 cpuset_setproc_update_set(struct proc *p, struct cpuset *set)
1519 KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__));
1520 KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__));
1523 error = cpuset_setproc(p->p_pid, set, NULL, NULL);
1530 #ifndef _SYS_SYSPROTO_H_
1531 struct cpuset_args {
1536 sys_cpuset(struct thread *td, struct cpuset_args *uap)
1538 struct cpuset *root;
1543 root = cpuset_refroot(td->td_cpuset);
1545 error = cpuset_create(&set, root, &root->cs_mask);
1549 error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id));
1551 error = cpuset_setproc(-1, set, NULL, NULL);
1556 #ifndef _SYS_SYSPROTO_H_
1557 struct cpuset_setid_args {
1564 sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap)
1567 return (kern_cpuset_setid(td, uap->which, uap->id, uap->setid));
1571 kern_cpuset_setid(struct thread *td, cpuwhich_t which,
1572 id_t id, cpusetid_t setid)
1578 * Presently we only support per-process sets.
1580 if (which != CPU_WHICH_PID)
1582 set = cpuset_lookup(setid, td);
1585 error = cpuset_setproc(id, set, NULL, NULL);
1590 #ifndef _SYS_SYSPROTO_H_
1591 struct cpuset_getid_args {
1599 sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap)
1602 return (kern_cpuset_getid(td, uap->level, uap->which, uap->id,
1607 kern_cpuset_getid(struct thread *td, cpulevel_t level, cpuwhich_t which,
1608 id_t id, cpusetid_t *setid)
1610 struct cpuset *nset;
1617 if (level == CPU_LEVEL_WHICH && which != CPU_WHICH_CPUSET)
1619 error = cpuset_which(which, id, &p, &ttd, &set);
1626 set = cpuset_refbase(ttd->td_cpuset);
1630 case CPU_WHICH_CPUSET:
1631 case CPU_WHICH_JAIL:
1634 case CPU_WHICH_DOMAIN:
1638 case CPU_LEVEL_ROOT:
1639 nset = cpuset_refroot(set);
1643 case CPU_LEVEL_CPUSET:
1645 case CPU_LEVEL_WHICH:
1651 error = copyout(&tmpid, setid, sizeof(tmpid));
1656 #ifndef _SYS_SYSPROTO_H_
1657 struct cpuset_getaffinity_args {
1666 sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap)
1669 return (kern_cpuset_getaffinity(td, uap->level, uap->which,
1670 uap->id, uap->cpusetsize, uap->mask));
1674 kern_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1675 id_t id, size_t cpusetsize, cpuset_t *maskp)
1678 struct cpuset *nset;
1685 if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1687 /* In Capability mode, you can only get your own CPU set. */
1688 if (IN_CAPABILITY_MODE(td)) {
1689 if (level != CPU_LEVEL_WHICH)
1691 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1697 mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO);
1698 error = cpuset_which(which, id, &p, &ttd, &set);
1702 case CPU_LEVEL_ROOT:
1703 case CPU_LEVEL_CPUSET:
1708 set = cpuset_ref(ttd->td_cpuset);
1711 case CPU_WHICH_CPUSET:
1712 case CPU_WHICH_JAIL:
1715 case CPU_WHICH_INTRHANDLER:
1716 case CPU_WHICH_ITHREAD:
1717 case CPU_WHICH_DOMAIN:
1721 if (level == CPU_LEVEL_ROOT)
1722 nset = cpuset_refroot(set);
1724 nset = cpuset_refbase(set);
1725 CPU_COPY(&nset->cs_mask, mask);
1728 case CPU_LEVEL_WHICH:
1732 CPU_COPY(&ttd->td_cpuset->cs_mask, mask);
1736 FOREACH_THREAD_IN_PROC(p, ttd) {
1738 CPU_OR(mask, &ttd->td_cpuset->cs_mask);
1742 case CPU_WHICH_CPUSET:
1743 case CPU_WHICH_JAIL:
1744 CPU_COPY(&set->cs_mask, mask);
1747 case CPU_WHICH_INTRHANDLER:
1748 case CPU_WHICH_ITHREAD:
1749 error = intr_getaffinity(id, which, mask);
1751 case CPU_WHICH_DOMAIN:
1752 if (id < 0 || id >= MAXMEMDOM)
1755 CPU_COPY(&cpuset_domain[id], mask);
1768 error = copyout(mask, maskp, size);
1774 #ifndef _SYS_SYSPROTO_H_
1775 struct cpuset_setaffinity_args {
1780 const cpuset_t *mask;
1784 sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap)
1787 return (kern_cpuset_setaffinity(td, uap->level, uap->which,
1788 uap->id, uap->cpusetsize, uap->mask));
1792 kern_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1793 id_t id, size_t cpusetsize, const cpuset_t *maskp)
1795 struct cpuset *nset;
1802 if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1804 /* In Capability mode, you can only set your own CPU set. */
1805 if (IN_CAPABILITY_MODE(td)) {
1806 if (level != CPU_LEVEL_WHICH)
1808 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1813 mask = malloc(cpusetsize, M_TEMP, M_WAITOK | M_ZERO);
1814 error = copyin(maskp, mask, cpusetsize);
1818 * Verify that no high bits are set.
1820 if (cpusetsize > sizeof(cpuset_t)) {
1824 end = cp = (char *)&mask->__bits;
1826 cp += sizeof(cpuset_t);
1835 case CPU_LEVEL_ROOT:
1836 case CPU_LEVEL_CPUSET:
1837 error = cpuset_which(which, id, &p, &ttd, &set);
1844 set = cpuset_ref(ttd->td_cpuset);
1848 case CPU_WHICH_CPUSET:
1849 case CPU_WHICH_JAIL:
1852 case CPU_WHICH_INTRHANDLER:
1853 case CPU_WHICH_ITHREAD:
1854 case CPU_WHICH_DOMAIN:
1858 if (level == CPU_LEVEL_ROOT)
1859 nset = cpuset_refroot(set);
1861 nset = cpuset_refbase(set);
1862 error = cpuset_modify(nset, mask);
1866 case CPU_LEVEL_WHICH:
1869 error = cpuset_setthread(id, mask);
1872 error = cpuset_setproc(id, NULL, mask, NULL);
1874 case CPU_WHICH_CPUSET:
1875 case CPU_WHICH_JAIL:
1876 error = cpuset_which(which, id, &p, &ttd, &set);
1878 error = cpuset_modify(set, mask);
1883 case CPU_WHICH_INTRHANDLER:
1884 case CPU_WHICH_ITHREAD:
1885 error = intr_setaffinity(id, which, mask);
1901 #ifndef _SYS_SYSPROTO_H_
1902 struct cpuset_getdomain_args {
1906 size_t domainsetsize;
1912 sys_cpuset_getdomain(struct thread *td, struct cpuset_getdomain_args *uap)
1915 return (kern_cpuset_getdomain(td, uap->level, uap->which,
1916 uap->id, uap->domainsetsize, uap->mask, uap->policy));
1920 kern_cpuset_getdomain(struct thread *td, cpulevel_t level, cpuwhich_t which,
1921 id_t id, size_t domainsetsize, domainset_t *maskp, int *policyp)
1923 struct domainset outset;
1925 struct cpuset *nset;
1927 struct domainset *dset;
1932 if (domainsetsize < sizeof(domainset_t) ||
1933 domainsetsize > DOMAINSET_MAXSIZE / NBBY)
1935 /* In Capability mode, you can only get your own domain set. */
1936 if (IN_CAPABILITY_MODE(td)) {
1937 if (level != CPU_LEVEL_WHICH)
1939 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1944 mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO);
1945 bzero(&outset, sizeof(outset));
1946 error = cpuset_which(which, id, &p, &ttd, &set);
1950 case CPU_LEVEL_ROOT:
1951 case CPU_LEVEL_CPUSET:
1956 set = cpuset_ref(ttd->td_cpuset);
1959 case CPU_WHICH_CPUSET:
1960 case CPU_WHICH_JAIL:
1963 case CPU_WHICH_INTRHANDLER:
1964 case CPU_WHICH_ITHREAD:
1965 case CPU_WHICH_DOMAIN:
1969 if (level == CPU_LEVEL_ROOT)
1970 nset = cpuset_refroot(set);
1972 nset = cpuset_refbase(set);
1973 domainset_copy(nset->cs_domain, &outset);
1976 case CPU_LEVEL_WHICH:
1980 domainset_copy(ttd->td_cpuset->cs_domain, &outset);
1984 FOREACH_THREAD_IN_PROC(p, ttd) {
1986 dset = ttd->td_cpuset->cs_domain;
1987 /* Show all domains in the proc. */
1988 DOMAINSET_OR(&outset.ds_mask, &dset->ds_mask);
1989 /* Last policy wins. */
1990 outset.ds_policy = dset->ds_policy;
1991 outset.ds_prefer = dset->ds_prefer;
1995 case CPU_WHICH_CPUSET:
1996 case CPU_WHICH_JAIL:
1997 domainset_copy(set->cs_domain, &outset);
2000 case CPU_WHICH_INTRHANDLER:
2001 case CPU_WHICH_ITHREAD:
2002 case CPU_WHICH_DOMAIN:
2016 * Translate prefer into a set containing only the preferred domain,
2017 * not the entire fallback set.
2019 if (outset.ds_policy == DOMAINSET_POLICY_PREFER) {
2020 DOMAINSET_ZERO(&outset.ds_mask);
2021 DOMAINSET_SET(outset.ds_prefer, &outset.ds_mask);
2023 DOMAINSET_COPY(&outset.ds_mask, mask);
2025 error = copyout(mask, maskp, domainsetsize);
2027 if (suword32(policyp, outset.ds_policy) != 0)
2034 #ifndef _SYS_SYSPROTO_H_
2035 struct cpuset_setdomain_args {
2039 size_t domainsetsize;
2045 sys_cpuset_setdomain(struct thread *td, struct cpuset_setdomain_args *uap)
2048 return (kern_cpuset_setdomain(td, uap->level, uap->which,
2049 uap->id, uap->domainsetsize, uap->mask, uap->policy));
2053 kern_cpuset_setdomain(struct thread *td, cpulevel_t level, cpuwhich_t which,
2054 id_t id, size_t domainsetsize, const domainset_t *maskp, int policy)
2056 struct cpuset *nset;
2060 struct domainset domain;
2064 if (domainsetsize < sizeof(domainset_t) ||
2065 domainsetsize > DOMAINSET_MAXSIZE / NBBY)
2067 if (policy <= DOMAINSET_POLICY_INVALID ||
2068 policy > DOMAINSET_POLICY_MAX)
2070 /* In Capability mode, you can only set your own CPU set. */
2071 if (IN_CAPABILITY_MODE(td)) {
2072 if (level != CPU_LEVEL_WHICH)
2074 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
2079 memset(&domain, 0, sizeof(domain));
2080 mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO);
2081 error = copyin(maskp, mask, domainsetsize);
2085 * Verify that no high bits are set.
2087 if (domainsetsize > sizeof(domainset_t)) {
2091 end = cp = (char *)&mask->__bits;
2092 end += domainsetsize;
2093 cp += sizeof(domainset_t);
2101 DOMAINSET_COPY(mask, &domain.ds_mask);
2102 domain.ds_policy = policy;
2104 /* Translate preferred policy into a mask and fallback. */
2105 if (policy == DOMAINSET_POLICY_PREFER) {
2106 /* Only support a single preferred domain. */
2107 if (DOMAINSET_COUNT(&domain.ds_mask) != 1) {
2111 domain.ds_prefer = DOMAINSET_FFS(&domain.ds_mask) - 1;
2112 /* This will be constrained by domainset_shadow(). */
2113 DOMAINSET_FILL(&domain.ds_mask);
2117 * When given an impossible policy, fall back to interleaving
2118 * across all domains
2120 if (domainset_empty_vm(&domain))
2121 domainset_copy(&domainset2, &domain);
2124 case CPU_LEVEL_ROOT:
2125 case CPU_LEVEL_CPUSET:
2126 error = cpuset_which(which, id, &p, &ttd, &set);
2133 set = cpuset_ref(ttd->td_cpuset);
2137 case CPU_WHICH_CPUSET:
2138 case CPU_WHICH_JAIL:
2141 case CPU_WHICH_INTRHANDLER:
2142 case CPU_WHICH_ITHREAD:
2143 case CPU_WHICH_DOMAIN:
2147 if (level == CPU_LEVEL_ROOT)
2148 nset = cpuset_refroot(set);
2150 nset = cpuset_refbase(set);
2151 error = cpuset_modify_domain(nset, &domain);
2155 case CPU_LEVEL_WHICH:
2158 error = _cpuset_setthread(id, NULL, &domain);
2161 error = cpuset_setproc(id, NULL, NULL, &domain);
2163 case CPU_WHICH_CPUSET:
2164 case CPU_WHICH_JAIL:
2165 error = cpuset_which(which, id, &p, &ttd, &set);
2167 error = cpuset_modify_domain(set, &domain);
2172 case CPU_WHICH_INTRHANDLER:
2173 case CPU_WHICH_ITHREAD:
2191 ddb_display_bitset(const struct bitset *set, int size)
2195 for (once = 0, bit = 0; bit < size; bit++) {
2196 if (CPU_ISSET(bit, set)) {
2198 db_printf("%d", bit);
2201 db_printf(",%d", bit);
2205 db_printf("<none>");
2209 ddb_display_cpuset(const cpuset_t *set)
2211 ddb_display_bitset((const struct bitset *)set, CPU_SETSIZE);
2215 ddb_display_domainset(const domainset_t *set)
2217 ddb_display_bitset((const struct bitset *)set, DOMAINSET_SETSIZE);
2220 DB_SHOW_COMMAND(cpusets, db_show_cpusets)
2224 LIST_FOREACH(set, &cpuset_ids, cs_link) {
2225 db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n",
2226 set, set->cs_id, set->cs_ref, set->cs_flags,
2227 (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0);
2228 db_printf(" cpu mask=");
2229 ddb_display_cpuset(&set->cs_mask);
2231 db_printf(" domain policy %d prefer %d mask=",
2232 set->cs_domain->ds_policy, set->cs_domain->ds_prefer);
2233 ddb_display_domainset(&set->cs_domain->ds_mask);
2240 DB_SHOW_COMMAND(domainsets, db_show_domainsets)
2242 struct domainset *set;
2244 LIST_FOREACH(set, &cpuset_domains, ds_link) {
2245 db_printf("set=%p policy %d prefer %d cnt %d\n",
2246 set, set->ds_policy, set->ds_prefer, set->ds_cnt);
2247 db_printf(" mask =");
2248 ddb_display_domainset(&set->ds_mask);