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);
122 struct domainset __read_mostly domainset_fixed[MAXMEMDOM];
123 struct domainset __read_mostly domainset_prefer[MAXMEMDOM];
124 struct domainset __read_mostly domainset_roundrobin;
126 static uma_zone_t cpuset_zone;
127 static uma_zone_t domainset_zone;
128 static struct mtx cpuset_lock;
129 static struct setlist cpuset_ids;
130 static struct domainlist cpuset_domains;
131 static struct unrhdr *cpuset_unr;
132 static struct cpuset *cpuset_zero, *cpuset_default, *cpuset_kernel;
133 static struct domainset domainset0, domainset2;
135 /* Return the size of cpuset_t at the kernel level */
136 SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD | CTLFLAG_CAPRD,
137 SYSCTL_NULL_INT_PTR, sizeof(cpuset_t), "sizeof(cpuset_t)");
139 cpuset_t *cpuset_root;
140 cpuset_t cpuset_domain[MAXMEMDOM];
142 static int domainset_valid(const struct domainset *, const struct domainset *);
145 * Find the first non-anonymous set starting from 'set'.
147 static struct cpuset *
148 cpuset_getbase(struct cpuset *set)
151 if (set->cs_id == CPUSET_INVALID)
152 set = set->cs_parent;
157 * Walks up the tree from 'set' to find the root.
159 static struct cpuset *
160 cpuset_getroot(struct cpuset *set)
163 while ((set->cs_flags & CPU_SET_ROOT) == 0 && set->cs_parent != NULL)
164 set = set->cs_parent;
169 * Acquire a reference to a cpuset, all pointers must be tracked with refs.
172 cpuset_ref(struct cpuset *set)
175 refcount_acquire(&set->cs_ref);
180 * Walks up the tree from 'set' to find the root. Returns the root
183 static struct cpuset *
184 cpuset_refroot(struct cpuset *set)
187 return (cpuset_ref(cpuset_getroot(set)));
191 * Find the first non-anonymous set starting from 'set'. Returns this set
192 * referenced. May return the passed in set with an extra ref if it is
195 static struct cpuset *
196 cpuset_refbase(struct cpuset *set)
199 return (cpuset_ref(cpuset_getbase(set)));
203 * Release a reference in a context where it is safe to allocate.
206 cpuset_rel(struct cpuset *set)
210 if (refcount_release_if_not_last(&set->cs_ref))
212 mtx_lock_spin(&cpuset_lock);
213 if (!refcount_release(&set->cs_ref)) {
214 mtx_unlock_spin(&cpuset_lock);
217 LIST_REMOVE(set, cs_siblings);
219 if (id != CPUSET_INVALID)
220 LIST_REMOVE(set, cs_link);
221 mtx_unlock_spin(&cpuset_lock);
222 cpuset_rel(set->cs_parent);
223 uma_zfree(cpuset_zone, set);
224 if (id != CPUSET_INVALID)
225 free_unr(cpuset_unr, id);
229 * Deferred release must be used when in a context that is not safe to
230 * allocate/free. This places any unreferenced sets on the list 'head'.
233 cpuset_rel_defer(struct setlist *head, struct cpuset *set)
236 if (refcount_release_if_not_last(&set->cs_ref))
238 mtx_lock_spin(&cpuset_lock);
239 if (!refcount_release(&set->cs_ref)) {
240 mtx_unlock_spin(&cpuset_lock);
243 LIST_REMOVE(set, cs_siblings);
244 if (set->cs_id != CPUSET_INVALID)
245 LIST_REMOVE(set, cs_link);
246 LIST_INSERT_HEAD(head, set, cs_link);
247 mtx_unlock_spin(&cpuset_lock);
251 * Complete a deferred release. Removes the set from the list provided to
255 cpuset_rel_complete(struct cpuset *set)
260 LIST_REMOVE(set, cs_link);
261 cpuset_rel(set->cs_parent);
262 uma_zfree(cpuset_zone, set);
263 if (id != CPUSET_INVALID)
264 free_unr(cpuset_unr, id);
268 * Find a set based on an id. Returns it with a ref.
270 static struct cpuset *
271 cpuset_lookup(cpusetid_t setid, struct thread *td)
275 if (setid == CPUSET_INVALID)
277 mtx_lock_spin(&cpuset_lock);
278 LIST_FOREACH(set, &cpuset_ids, cs_link)
279 if (set->cs_id == setid)
283 mtx_unlock_spin(&cpuset_lock);
285 KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__));
286 if (set != NULL && jailed(td->td_ucred)) {
287 struct cpuset *jset, *tset;
289 jset = td->td_ucred->cr_prison->pr_cpuset;
290 for (tset = set; tset != NULL; tset = tset->cs_parent)
303 * Initialize a set in the space provided in 'set' with the provided parameters.
304 * The set is returned with a single ref. May return EDEADLK if the set
305 * will have no valid cpu based on restrictions from the parent.
308 cpuset_init(struct cpuset *set, struct cpuset *parent,
309 const cpuset_t *mask, struct domainset *domain, cpusetid_t id)
313 domain = parent->cs_domain;
315 mask = &parent->cs_mask;
316 if (!CPU_OVERLAP(&parent->cs_mask, mask))
318 /* The domain must be prepared ahead of time. */
319 if (!domainset_valid(parent->cs_domain, domain))
321 CPU_COPY(mask, &set->cs_mask);
322 LIST_INIT(&set->cs_children);
323 refcount_init(&set->cs_ref, 1);
325 mtx_lock_spin(&cpuset_lock);
326 set->cs_domain = domain;
327 CPU_AND(&set->cs_mask, &parent->cs_mask);
329 set->cs_parent = cpuset_ref(parent);
330 LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings);
331 if (set->cs_id != CPUSET_INVALID)
332 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
333 mtx_unlock_spin(&cpuset_lock);
339 * Create a new non-anonymous set with the requested parent and mask. May
340 * return failures if the mask is invalid or a new number can not be
343 * If *setp is not NULL, then it will be used as-is. The caller must take
344 * into account that *setp will be inserted at the head of cpuset_ids and
345 * plan any potentially conflicting cs_link usage accordingly.
348 cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask)
355 id = alloc_unr(cpuset_unr);
358 dofree = (*setp == NULL);
362 *setp = set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
363 error = cpuset_init(set, parent, mask, NULL, id);
366 free_unr(cpuset_unr, id);
368 uma_zfree(cpuset_zone, set);
374 cpuset_freelist_add(struct setlist *list, int count)
379 for (i = 0; i < count; i++) {
380 set = uma_zalloc(cpuset_zone, M_ZERO | M_WAITOK);
381 LIST_INSERT_HEAD(list, set, cs_link);
386 cpuset_freelist_init(struct setlist *list, int count)
390 cpuset_freelist_add(list, count);
394 cpuset_freelist_free(struct setlist *list)
398 while ((set = LIST_FIRST(list)) != NULL) {
399 LIST_REMOVE(set, cs_link);
400 uma_zfree(cpuset_zone, set);
405 domainset_freelist_add(struct domainlist *list, int count)
407 struct domainset *set;
410 for (i = 0; i < count; i++) {
411 set = uma_zalloc(domainset_zone, M_ZERO | M_WAITOK);
412 LIST_INSERT_HEAD(list, set, ds_link);
417 domainset_freelist_init(struct domainlist *list, int count)
421 domainset_freelist_add(list, count);
425 domainset_freelist_free(struct domainlist *list)
427 struct domainset *set;
429 while ((set = LIST_FIRST(list)) != NULL) {
430 LIST_REMOVE(set, ds_link);
431 uma_zfree(domainset_zone, set);
435 /* Copy a domainset preserving mask and policy. */
437 domainset_copy(const struct domainset *from, struct domainset *to)
440 DOMAINSET_COPY(&from->ds_mask, &to->ds_mask);
441 to->ds_policy = from->ds_policy;
442 to->ds_prefer = from->ds_prefer;
445 /* Return 1 if mask and policy are equal, otherwise 0. */
447 domainset_equal(const struct domainset *one, const struct domainset *two)
450 return (DOMAINSET_CMP(&one->ds_mask, &two->ds_mask) == 0 &&
451 one->ds_policy == two->ds_policy &&
452 one->ds_prefer == two->ds_prefer);
455 /* Return 1 if child is a valid subset of parent. */
457 domainset_valid(const struct domainset *parent, const struct domainset *child)
459 if (child->ds_policy != DOMAINSET_POLICY_PREFER)
460 return (DOMAINSET_SUBSET(&parent->ds_mask, &child->ds_mask));
461 return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask));
465 domainset_restrict(const struct domainset *parent,
466 const struct domainset *child)
468 if (child->ds_policy != DOMAINSET_POLICY_PREFER)
469 return (DOMAINSET_OVERLAP(&parent->ds_mask, &child->ds_mask));
470 return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask));
474 * Lookup or create a domainset. The key is provided in ds_mask and
475 * ds_policy. If the domainset does not yet exist the storage in
476 * 'domain' is used to insert. Otherwise this storage is freed to the
477 * domainset_zone and the existing domainset is returned.
479 static struct domainset *
480 _domainset_create(struct domainset *domain, struct domainlist *freelist)
482 struct domainset *ndomain;
485 KASSERT(domain->ds_cnt <= vm_ndomains,
486 ("invalid domain count in domainset %p", domain));
487 KASSERT(domain->ds_policy != DOMAINSET_POLICY_PREFER ||
488 domain->ds_prefer < vm_ndomains,
489 ("invalid preferred domain in domains %p", domain));
491 mtx_lock_spin(&cpuset_lock);
492 LIST_FOREACH(ndomain, &cpuset_domains, ds_link)
493 if (domainset_equal(ndomain, domain))
496 * If the domain does not yet exist we insert it and initialize
497 * various iteration helpers which are not part of the key.
499 if (ndomain == NULL) {
500 LIST_INSERT_HEAD(&cpuset_domains, domain, ds_link);
501 domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask);
502 for (i = 0, j = 0; i < DOMAINSET_FLS(&domain->ds_mask); i++)
503 if (DOMAINSET_ISSET(i, &domain->ds_mask))
504 domain->ds_order[j++] = i;
506 mtx_unlock_spin(&cpuset_lock);
509 if (freelist != NULL)
510 LIST_INSERT_HEAD(freelist, domain, ds_link);
512 uma_zfree(domainset_zone, domain);
518 * Are any of the domains in the mask empty? If so, silently
519 * remove them and update the domainset accordingly. If only empty
520 * domains are present, we must return failure.
523 domainset_empty_vm(struct domainset *domain)
528 DOMAINSET_ZERO(&empty);
529 for (i = 0; i < vm_ndomains; i++)
530 if (VM_DOMAIN_EMPTY(i))
531 DOMAINSET_SET(i, &empty);
532 if (DOMAINSET_SUBSET(&empty, &domain->ds_mask))
535 /* Remove empty domains from the set and recompute. */
536 DOMAINSET_ANDNOT(&domain->ds_mask, &empty);
537 domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask);
538 for (i = j = 0; i < DOMAINSET_FLS(&domain->ds_mask); i++)
539 if (DOMAINSET_ISSET(i, &domain->ds_mask))
540 domain->ds_order[j++] = i;
542 /* Convert a PREFER policy referencing an empty domain to RR. */
543 if (domain->ds_policy == DOMAINSET_POLICY_PREFER &&
544 DOMAINSET_ISSET(domain->ds_prefer, &empty)) {
545 domain->ds_policy = DOMAINSET_POLICY_ROUNDROBIN;
546 domain->ds_prefer = -1;
553 * Create or lookup a domainset based on the key held in 'domain'.
556 domainset_create(const struct domainset *domain)
558 struct domainset *ndomain;
561 * Validate the policy. It must specify a useable policy number with
562 * only valid domains. Preferred must include the preferred domain
565 if (domain->ds_policy <= DOMAINSET_POLICY_INVALID ||
566 domain->ds_policy > DOMAINSET_POLICY_MAX)
568 if (domain->ds_policy == DOMAINSET_POLICY_PREFER &&
569 !DOMAINSET_ISSET(domain->ds_prefer, &domain->ds_mask))
571 if (!DOMAINSET_SUBSET(&domainset0.ds_mask, &domain->ds_mask))
573 ndomain = uma_zalloc(domainset_zone, M_WAITOK | M_ZERO);
574 domainset_copy(domain, ndomain);
575 return _domainset_create(ndomain, NULL);
579 * Update thread domainset pointers.
582 domainset_notify(void)
587 sx_slock(&allproc_lock);
588 FOREACH_PROC_IN_SYSTEM(p) {
590 if (p->p_state == PRS_NEW) {
594 FOREACH_THREAD_IN_PROC(p, td) {
596 td->td_domain.dr_policy = td->td_cpuset->cs_domain;
601 sx_sunlock(&allproc_lock);
602 kernel_object->domain.dr_policy = cpuset_kernel->cs_domain;
606 * Create a new set that is a subset of a parent.
608 static struct domainset *
609 domainset_shadow(const struct domainset *pdomain,
610 const struct domainset *domain, struct domainlist *freelist)
612 struct domainset *ndomain;
614 ndomain = LIST_FIRST(freelist);
615 LIST_REMOVE(ndomain, ds_link);
618 * Initialize the key from the request.
620 domainset_copy(domain, ndomain);
623 * Restrict the key by the parent.
625 DOMAINSET_AND(&ndomain->ds_mask, &pdomain->ds_mask);
627 return _domainset_create(ndomain, freelist);
631 * Recursively check for errors that would occur from applying mask to
632 * the tree of sets starting at 'set'. Checks for sets that would become
633 * empty as well as RDONLY flags.
636 cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int augment_mask)
642 mtx_assert(&cpuset_lock, MA_OWNED);
643 if (set->cs_flags & CPU_SET_RDONLY)
646 CPU_COPY(&set->cs_mask, &newmask);
647 CPU_AND(&newmask, mask);
649 CPU_COPY(mask, &newmask);
651 if (CPU_EMPTY(&newmask))
654 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
655 if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0)
661 * Applies the mask 'mask' without checking for empty sets or permissions.
664 cpuset_update(struct cpuset *set, cpuset_t *mask)
668 mtx_assert(&cpuset_lock, MA_OWNED);
669 CPU_AND(&set->cs_mask, mask);
670 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
671 cpuset_update(nset, &set->cs_mask);
677 * Modify the set 'set' to use a copy of the mask provided. Apply this new
678 * mask to restrict all children in the tree. Checks for validity before
679 * applying the changes.
682 cpuset_modify(struct cpuset *set, cpuset_t *mask)
687 error = priv_check(curthread, PRIV_SCHED_CPUSET);
691 * In case we are called from within the jail
692 * we do not allow modifying the dedicated root
693 * cpuset of the jail but may still allow to
696 if (jailed(curthread->td_ucred) &&
697 set->cs_flags & CPU_SET_ROOT)
700 * Verify that we have access to this set of
703 root = cpuset_getroot(set);
704 mtx_lock_spin(&cpuset_lock);
705 if (root && !CPU_SUBSET(&root->cs_mask, mask)) {
709 error = cpuset_testupdate(set, mask, 0);
712 CPU_COPY(mask, &set->cs_mask);
713 cpuset_update(set, mask);
715 mtx_unlock_spin(&cpuset_lock);
721 * Recursively check for errors that would occur from applying mask to
722 * the tree of sets starting at 'set'. Checks for sets that would become
723 * empty as well as RDONLY flags.
726 cpuset_testupdate_domain(struct cpuset *set, struct domainset *dset,
727 struct domainset *orig, int *count, int augment_mask __unused)
730 struct domainset *domain;
731 struct domainset newset;
734 mtx_assert(&cpuset_lock, MA_OWNED);
735 if (set->cs_flags & CPU_SET_RDONLY)
737 domain = set->cs_domain;
738 domainset_copy(domain, &newset);
739 if (!domainset_equal(domain, orig)) {
740 if (!domainset_restrict(domain, dset))
742 DOMAINSET_AND(&newset.ds_mask, &dset->ds_mask);
743 /* Count the number of domains that are changing. */
747 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
748 if ((error = cpuset_testupdate_domain(nset, &newset, domain,
755 * Applies the mask 'mask' without checking for empty sets or permissions.
758 cpuset_update_domain(struct cpuset *set, struct domainset *domain,
759 struct domainset *orig, struct domainlist *domains)
763 mtx_assert(&cpuset_lock, MA_OWNED);
765 * If this domainset has changed from the parent we must calculate
766 * a new set. Otherwise it simply inherits from the parent. When
767 * we inherit from the parent we get a new mask and policy. If the
768 * set is modified from the parent we keep the policy and only
771 if (set->cs_domain != orig) {
772 orig = set->cs_domain;
773 set->cs_domain = domainset_shadow(domain, orig, domains);
775 set->cs_domain = domain;
776 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
777 cpuset_update_domain(nset, set->cs_domain, orig, domains);
783 * Modify the set 'set' to use a copy the domainset provided. Apply this new
784 * mask to restrict all children in the tree. Checks for validity before
785 * applying the changes.
788 cpuset_modify_domain(struct cpuset *set, struct domainset *domain)
790 struct domainlist domains;
791 struct domainset temp;
792 struct domainset *dset;
794 int ndomains, needed;
797 error = priv_check(curthread, PRIV_SCHED_CPUSET);
801 * In case we are called from within the jail
802 * we do not allow modifying the dedicated root
803 * cpuset of the jail but may still allow to
806 if (jailed(curthread->td_ucred) &&
807 set->cs_flags & CPU_SET_ROOT)
809 domainset_freelist_init(&domains, 0);
810 domain = domainset_create(domain);
813 mtx_lock_spin(&cpuset_lock);
815 root = cpuset_getroot(set);
816 dset = root->cs_domain;
818 * Verify that we have access to this set of domains.
820 if (!domainset_valid(dset, domain)) {
825 * If applying prefer we keep the current set as the fallback.
827 if (domain->ds_policy == DOMAINSET_POLICY_PREFER)
828 DOMAINSET_COPY(&set->cs_domain->ds_mask,
831 * Determine whether we can apply this set of domains and
832 * how many new domain structures it will require.
834 domainset_copy(domain, &temp);
836 error = cpuset_testupdate_domain(set, &temp, set->cs_domain,
840 if (ndomains >= needed)
843 /* Dropping the lock; we'll need to re-evaluate again. */
844 mtx_unlock_spin(&cpuset_lock);
845 domainset_freelist_add(&domains, needed - ndomains);
847 mtx_lock_spin(&cpuset_lock);
849 dset = set->cs_domain;
850 cpuset_update_domain(set, domain, dset, &domains);
852 mtx_unlock_spin(&cpuset_lock);
853 domainset_freelist_free(&domains);
861 * Resolve the 'which' parameter of several cpuset apis.
863 * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid. Also
864 * checks for permission via p_cansched().
866 * For WHICH_SET returns a valid set with a new reference.
868 * -1 may be supplied for any argument to mean the current proc/thread or
869 * the base set of the current thread. May fail with ESRCH/EPERM.
872 cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp,
873 struct cpuset **setp)
890 if ((p = pfind(id)) == NULL)
905 case CPU_WHICH_CPUSET:
907 thread_lock(curthread);
908 set = cpuset_refbase(curthread->td_cpuset);
909 thread_unlock(curthread);
911 set = cpuset_lookup(id, curthread);
919 /* Find `set' for prison with given id. */
922 sx_slock(&allprison_lock);
923 pr = prison_find_child(curthread->td_ucred->cr_prison, id);
924 sx_sunlock(&allprison_lock);
927 cpuset_ref(pr->pr_cpuset);
928 *setp = pr->pr_cpuset;
929 mtx_unlock(&pr->pr_mtx);
933 case CPU_WHICH_DOMAIN:
938 error = p_cansched(curthread, p);
944 td = FIRST_THREAD_IN_PROC(p);
951 cpuset_testshadow(struct cpuset *set, const cpuset_t *mask,
952 const struct domainset *domain)
954 struct cpuset *parent;
955 struct domainset *dset;
957 parent = cpuset_getbase(set);
959 * If we are restricting a cpu mask it must be a subset of the
960 * parent or invalid CPUs have been specified.
962 if (mask != NULL && !CPU_SUBSET(&parent->cs_mask, mask))
966 * If we are restricting a domain mask it must be a subset of the
967 * parent or invalid domains have been specified.
969 dset = parent->cs_domain;
970 if (domain != NULL && !domainset_valid(dset, domain))
977 * Create an anonymous set with the provided mask in the space provided by
978 * 'nset'. If the passed in set is anonymous we use its parent otherwise
979 * the new set is a child of 'set'.
982 cpuset_shadow(struct cpuset *set, struct cpuset **nsetp,
983 const cpuset_t *mask, const struct domainset *domain,
984 struct setlist *cpusets, struct domainlist *domains)
986 struct cpuset *parent;
988 struct domainset *dset;
992 error = cpuset_testshadow(set, mask, domain);
996 parent = cpuset_getbase(set);
997 dset = parent->cs_domain;
999 mask = &set->cs_mask;
1001 d = domainset_shadow(dset, domain, domains);
1004 nset = LIST_FIRST(cpusets);
1005 error = cpuset_init(nset, parent, mask, d, CPUSET_INVALID);
1007 LIST_REMOVE(nset, cs_link);
1013 static struct cpuset *
1014 cpuset_update_thread(struct thread *td, struct cpuset *nset)
1016 struct cpuset *tdset;
1018 tdset = td->td_cpuset;
1019 td->td_cpuset = nset;
1020 td->td_domain.dr_policy = nset->cs_domain;
1027 cpuset_setproc_test_maskthread(struct cpuset *tdset, cpuset_t *mask,
1028 struct domainset *domain)
1030 struct cpuset *parent;
1032 parent = cpuset_getbase(tdset);
1034 mask = &tdset->cs_mask;
1036 domain = tdset->cs_domain;
1037 return cpuset_testshadow(parent, mask, domain);
1041 cpuset_setproc_maskthread(struct cpuset *tdset, cpuset_t *mask,
1042 struct domainset *domain, struct cpuset **nsetp,
1043 struct setlist *freelist, struct domainlist *domainlist)
1045 struct cpuset *parent;
1047 parent = cpuset_getbase(tdset);
1049 mask = &tdset->cs_mask;
1051 domain = tdset->cs_domain;
1052 return cpuset_shadow(parent, nsetp, mask, domain, freelist,
1057 cpuset_setproc_setthread_mask(struct cpuset *tdset, struct cpuset *set,
1058 cpuset_t *mask, struct domainset *domain)
1060 struct cpuset *parent;
1062 parent = cpuset_getbase(tdset);
1065 * If the thread restricted its mask then apply that same
1066 * restriction to the new set, otherwise take it wholesale.
1068 if (CPU_CMP(&tdset->cs_mask, &parent->cs_mask) != 0) {
1069 CPU_COPY(&tdset->cs_mask, mask);
1070 CPU_AND(mask, &set->cs_mask);
1072 CPU_COPY(&set->cs_mask, mask);
1075 * If the thread restricted the domain then we apply the
1076 * restriction to the new set but retain the policy.
1078 if (tdset->cs_domain != parent->cs_domain) {
1079 domainset_copy(tdset->cs_domain, domain);
1080 DOMAINSET_AND(&domain->ds_mask, &set->cs_domain->ds_mask);
1082 domainset_copy(set->cs_domain, domain);
1084 if (CPU_EMPTY(mask) || DOMAINSET_EMPTY(&domain->ds_mask))
1091 cpuset_setproc_test_setthread(struct cpuset *tdset, struct cpuset *set)
1093 struct domainset domain;
1096 if (tdset->cs_id != CPUSET_INVALID)
1098 return cpuset_setproc_setthread_mask(tdset, set, &mask, &domain);
1102 cpuset_setproc_setthread(struct cpuset *tdset, struct cpuset *set,
1103 struct cpuset **nsetp, struct setlist *freelist,
1104 struct domainlist *domainlist)
1106 struct domainset domain;
1111 * If we're replacing on a thread that has not constrained the
1112 * original set we can simply accept the new set.
1114 if (tdset->cs_id != CPUSET_INVALID) {
1115 *nsetp = cpuset_ref(set);
1118 error = cpuset_setproc_setthread_mask(tdset, set, &mask, &domain);
1122 return cpuset_shadow(set, nsetp, &mask, &domain, freelist,
1127 cpuset_setproc_newbase(struct thread *td, struct cpuset *set,
1128 struct cpuset *nroot, struct cpuset **nsetp,
1129 struct setlist *cpusets, struct domainlist *domainlist)
1131 struct domainset ndomain;
1133 struct cpuset *pbase;
1136 pbase = cpuset_getbase(td->td_cpuset);
1138 /* Copy process mask, then further apply the new root mask. */
1139 CPU_COPY(&pbase->cs_mask, &nmask);
1140 CPU_AND(&nmask, &nroot->cs_mask);
1142 domainset_copy(pbase->cs_domain, &ndomain);
1143 DOMAINSET_AND(&ndomain.ds_mask, &set->cs_domain->ds_mask);
1145 /* Policy is too restrictive, will not work. */
1146 if (CPU_EMPTY(&nmask) || DOMAINSET_EMPTY(&ndomain.ds_mask))
1150 * Remove pbase from the freelist in advance, it'll be pushed to
1151 * cpuset_ids on success. We assume here that cpuset_create() will not
1152 * touch pbase on failure, and we just enqueue it back to the freelist
1153 * to remain in a consistent state.
1155 pbase = LIST_FIRST(cpusets);
1156 LIST_REMOVE(pbase, cs_link);
1157 error = cpuset_create(&pbase, set, &nmask);
1159 LIST_INSERT_HEAD(cpusets, pbase, cs_link);
1163 /* Duplicates some work from above... oh well. */
1164 pbase->cs_domain = domainset_shadow(set->cs_domain, &ndomain,
1171 * Handle four cases for updating an entire process.
1173 * 1) Set is non-null and the process is not rebasing onto a new root. This
1174 * reparents all anonymous sets to the provided set and replaces all
1175 * non-anonymous td_cpusets with the provided set.
1176 * 2) Set is non-null and the process is rebasing onto a new root. This
1177 * creates a new base set if the process previously had its own base set,
1178 * then reparents all anonymous sets either to that set or the provided set
1179 * if one was not created. Non-anonymous sets are similarly replaced.
1180 * 3) Mask is non-null. This replaces or creates anonymous sets for every
1181 * thread with the existing base as a parent.
1182 * 4) domain is non-null. This creates anonymous sets for every thread
1183 * and replaces the domain set.
1185 * This is overly complicated because we can't allocate while holding a
1186 * spinlock and spinlocks must be held while changing and examining thread
1190 cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask,
1191 struct domainset *domain, bool rebase)
1193 struct setlist freelist;
1194 struct setlist droplist;
1195 struct domainlist domainlist;
1196 struct cpuset *base, *nset, *nroot, *tdroot;
1204 * The algorithm requires two passes due to locking considerations.
1206 * 1) Lookup the process and acquire the locks in the required order.
1207 * 2) If enough cpusets have not been allocated release the locks and
1208 * allocate them. Loop.
1210 cpuset_freelist_init(&freelist, 1);
1211 domainset_freelist_init(&domainlist, 1);
1213 LIST_INIT(&droplist);
1218 nroot = cpuset_getroot(set);
1220 error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset);
1223 tdroot = cpuset_getroot(td->td_cpuset);
1224 needed = p->p_numthreads;
1225 if (set != NULL && rebase && tdroot != nroot)
1227 if (nfree >= needed)
1230 if (nfree < needed) {
1231 cpuset_freelist_add(&freelist, needed - nfree);
1232 domainset_freelist_add(&domainlist, needed - nfree);
1236 PROC_LOCK_ASSERT(p, MA_OWNED);
1239 * If we're changing roots and the root set is what has been specified
1240 * as the parent, then we'll check if the process was previously using
1241 * the root set and, if it wasn't, create a new base with the process's
1242 * mask applied to it.
1244 if (set != NULL && rebase && nroot != tdroot) {
1245 cpusetid_t base_id, root_id;
1247 root_id = td->td_ucred->cr_prison->pr_cpuset->cs_id;
1248 base_id = cpuset_getbase(td->td_cpuset)->cs_id;
1250 if (base_id != root_id) {
1251 error = cpuset_setproc_newbase(td, set, nroot, &base,
1252 &freelist, &domainlist);
1259 * Now that the appropriate locks are held and we have enough cpusets,
1260 * make sure the operation will succeed before applying changes. The
1261 * proc lock prevents td_cpuset from changing between calls.
1264 FOREACH_THREAD_IN_PROC(p, td) {
1267 error = cpuset_setproc_test_setthread(td->td_cpuset,
1270 error = cpuset_setproc_test_maskthread(td->td_cpuset,
1277 * Replace each thread's cpuset while using deferred release. We
1278 * must do this because the thread lock must be held while operating
1279 * on the thread and this limits the type of operations allowed.
1281 FOREACH_THREAD_IN_PROC(p, td) {
1284 error = cpuset_setproc_setthread(td->td_cpuset, base,
1285 &nset, &freelist, &domainlist);
1287 error = cpuset_setproc_maskthread(td->td_cpuset, mask,
1288 domain, &nset, &freelist, &domainlist);
1293 cpuset_rel_defer(&droplist, cpuset_update_thread(td, nset));
1299 if (base != NULL && base != set)
1301 while ((nset = LIST_FIRST(&droplist)) != NULL)
1302 cpuset_rel_complete(nset);
1303 cpuset_freelist_free(&freelist);
1304 domainset_freelist_free(&domainlist);
1309 bitset_strprint(char *buf, size_t bufsiz, const struct bitset *set, int setlen)
1317 for (i = 0; i < __bitset_words(setlen); i++) {
1326 if (bufsiz < sizeof(__STRING(ULONG_MAX)))
1328 bytes = snprintf(p, bufsiz, "%lx", set->__bits[i]);
1336 bitset_strscan(struct bitset *set, int setlen, const char *buf)
1341 BIT_ZERO(setlen, set);
1343 for (i = 0; i < __bitset_words(setlen); i++) {
1348 ret = sscanf(p, "%lx", &set->__bits[i]);
1349 if (ret == 0 || ret == -1)
1351 while (isxdigit(*p))
1358 * Return a string representing a valid layout for a cpuset_t object.
1359 * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
1362 cpusetobj_strprint(char *buf, const cpuset_t *set)
1365 bitset_strprint(buf, CPUSETBUFSIZ, (const struct bitset *)set,
1371 * Build a valid cpuset_t object from a string representation.
1372 * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
1375 cpusetobj_strscan(cpuset_t *set, const char *buf)
1379 if (strlen(buf) > CPUSETBUFSIZ - 1)
1382 p = buf[bitset_strscan((struct bitset *)set, CPU_SETSIZE, buf)];
1390 * Handle a domainset specifier in the sysctl tree. A poiner to a pointer to
1391 * a domainset is in arg1. If the user specifies a valid domainset the
1392 * pointer is updated.
1395 * hex mask word 0,hex mask word 1,...:decimal policy:decimal preferred
1398 sysctl_handle_domainset(SYSCTL_HANDLER_ARGS)
1400 char buf[DOMAINSETBUFSIZ];
1401 struct domainset *dset;
1402 struct domainset key;
1403 int policy, prefer, error;
1406 dset = *(struct domainset **)arg1;
1410 p = buf + bitset_strprint(buf, DOMAINSETBUFSIZ,
1411 (const struct bitset *)&dset->ds_mask, DOMAINSET_SETSIZE);
1412 sprintf(p, ":%d:%d", dset->ds_policy, dset->ds_prefer);
1414 sprintf(buf, "<NULL>");
1415 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
1416 if (error != 0 || req->newptr == NULL)
1420 * Read in and validate the string.
1422 memset(&key, 0, sizeof(key));
1423 p = &buf[bitset_strscan((struct bitset *)&key.ds_mask,
1424 DOMAINSET_SETSIZE, buf)];
1427 if (sscanf(p, ":%d:%d", &policy, &prefer) != 2)
1429 key.ds_policy = policy;
1430 key.ds_prefer = prefer;
1432 /* Domainset_create() validates the policy.*/
1433 dset = domainset_create(&key);
1436 *(struct domainset **)arg1 = dset;
1442 * Apply an anonymous mask or a domain to a single thread.
1445 _cpuset_setthread(lwpid_t id, cpuset_t *mask, struct domainset *domain)
1447 struct setlist cpusets;
1448 struct domainlist domainlist;
1449 struct cpuset *nset;
1455 cpuset_freelist_init(&cpusets, 1);
1456 domainset_freelist_init(&domainlist, domain != NULL);
1457 error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set);
1462 error = cpuset_shadow(td->td_cpuset, &nset, mask, domain,
1463 &cpusets, &domainlist);
1465 set = cpuset_update_thread(td, nset);
1471 cpuset_freelist_free(&cpusets);
1472 domainset_freelist_free(&domainlist);
1477 * Apply an anonymous mask to a single thread.
1480 cpuset_setthread(lwpid_t id, cpuset_t *mask)
1483 return _cpuset_setthread(id, mask, NULL);
1487 * Apply new cpumask to the ithread.
1490 cpuset_setithread(lwpid_t id, int cpu)
1496 CPU_COPY(cpuset_root, &mask);
1498 CPU_SET(cpu, &mask);
1499 return _cpuset_setthread(id, &mask, NULL);
1503 * Initialize static domainsets after NUMA information is available. This is
1504 * called before memory allocators are initialized.
1507 domainset_init(void)
1509 struct domainset *dset;
1512 dset = &domainset_roundrobin;
1513 DOMAINSET_COPY(&all_domains, &dset->ds_mask);
1514 dset->ds_policy = DOMAINSET_POLICY_ROUNDROBIN;
1515 dset->ds_prefer = -1;
1516 _domainset_create(dset, NULL);
1518 for (i = 0; i < vm_ndomains; i++) {
1519 dset = &domainset_fixed[i];
1520 DOMAINSET_ZERO(&dset->ds_mask);
1521 DOMAINSET_SET(i, &dset->ds_mask);
1522 dset->ds_policy = DOMAINSET_POLICY_ROUNDROBIN;
1523 _domainset_create(dset, NULL);
1525 dset = &domainset_prefer[i];
1526 DOMAINSET_COPY(&all_domains, &dset->ds_mask);
1527 dset->ds_policy = DOMAINSET_POLICY_PREFER;
1528 dset->ds_prefer = i;
1529 _domainset_create(dset, NULL);
1534 * Create the domainset for cpuset 0, 1 and cpuset 2.
1537 domainset_zero(void)
1539 struct domainset *dset, *tmp;
1541 mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE);
1544 DOMAINSET_COPY(&all_domains, &dset->ds_mask);
1545 dset->ds_policy = DOMAINSET_POLICY_FIRSTTOUCH;
1546 dset->ds_prefer = -1;
1547 curthread->td_domain.dr_policy = _domainset_create(dset, NULL);
1549 domainset_copy(dset, &domainset2);
1550 domainset2.ds_policy = DOMAINSET_POLICY_INTERLEAVE;
1551 kernel_object->domain.dr_policy = _domainset_create(&domainset2, NULL);
1553 /* Remove empty domains from the global policies. */
1554 LIST_FOREACH_SAFE(dset, &cpuset_domains, ds_link, tmp)
1555 if (domainset_empty_vm(dset))
1556 LIST_REMOVE(dset, ds_link);
1560 * Creates system-wide cpusets and the cpuset for thread0 including three
1563 * 0 - The root set which should represent all valid processors in the
1564 * system. This set is immutable.
1565 * 1 - The default set which all processes are a member of until changed.
1566 * This allows an administrator to move all threads off of given cpus to
1567 * dedicate them to high priority tasks or save power etc.
1568 * 2 - The kernel set which allows restriction and policy to be applied only
1569 * to kernel threads and the kernel_object.
1572 cpuset_thread0(void)
1578 cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL,
1579 NULL, NULL, UMA_ALIGN_CACHE, 0);
1580 domainset_zone = uma_zcreate("domainset", sizeof(struct domainset),
1581 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
1584 * Create the root system set (0) for the whole machine. Doesn't use
1585 * cpuset_create() due to NULL parent.
1587 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1588 CPU_COPY(&all_cpus, &set->cs_mask);
1589 LIST_INIT(&set->cs_children);
1590 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
1591 refcount_init(&set->cs_ref, 1);
1592 set->cs_flags = CPU_SET_ROOT | CPU_SET_RDONLY;
1593 set->cs_domain = &domainset0;
1595 cpuset_root = &set->cs_mask;
1598 * Now derive a default (1), modifiable set from that to give out.
1600 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1601 error = cpuset_init(set, cpuset_zero, NULL, NULL, 1);
1602 KASSERT(error == 0, ("Error creating default set: %d\n", error));
1603 cpuset_default = set;
1605 * Create the kernel set (2).
1607 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1608 error = cpuset_init(set, cpuset_zero, NULL, NULL, 2);
1609 KASSERT(error == 0, ("Error creating kernel set: %d\n", error));
1610 set->cs_domain = &domainset2;
1611 cpuset_kernel = set;
1614 * Initialize the unit allocator. 0 and 1 are allocated above.
1616 cpuset_unr = new_unrhdr(3, INT_MAX, NULL);
1619 * If MD code has not initialized per-domain cpusets, place all
1622 for (i = 0; i < MAXMEMDOM; i++)
1623 if (!CPU_EMPTY(&cpuset_domain[i]))
1625 CPU_COPY(&all_cpus, &cpuset_domain[0]);
1628 return (cpuset_default);
1632 cpuset_kernthread(struct thread *td)
1637 set = td->td_cpuset;
1638 td->td_cpuset = cpuset_ref(cpuset_kernel);
1644 * Create a cpuset, which would be cpuset_create() but
1645 * mark the new 'set' as root.
1647 * We are not going to reparent the td to it. Use cpuset_setproc_update_set()
1650 * In case of no error, returns the set in *setp locked with a reference.
1653 cpuset_create_root(struct prison *pr, struct cpuset **setp)
1658 KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__));
1659 KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__));
1662 error = cpuset_create(&set, pr->pr_cpuset, &pr->pr_cpuset->cs_mask);
1666 KASSERT(set != NULL, ("[%s:%d] cpuset_create returned invalid data",
1667 __func__, __LINE__));
1669 /* Mark the set as root. */
1670 set->cs_flags |= CPU_SET_ROOT;
1677 cpuset_setproc_update_set(struct proc *p, struct cpuset *set)
1681 KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__));
1682 KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__));
1685 error = cpuset_setproc(p->p_pid, set, NULL, NULL, true);
1693 * In Capability mode, the only accesses that are permitted are to the current
1694 * thread and process' CPU and domain sets.
1697 cpuset_check_capabilities(struct thread *td, cpulevel_t level, cpuwhich_t which,
1700 if (IN_CAPABILITY_MODE(td)) {
1701 if (level != CPU_LEVEL_WHICH)
1703 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1706 !(which == CPU_WHICH_TID && id == td->td_tid) &&
1707 !(which == CPU_WHICH_PID && id == td->td_proc->p_pid))
1713 #ifndef _SYS_SYSPROTO_H_
1714 struct cpuset_args {
1719 sys_cpuset(struct thread *td, struct cpuset_args *uap)
1721 struct cpuset *root;
1726 root = cpuset_refroot(td->td_cpuset);
1729 error = cpuset_create(&set, root, &root->cs_mask);
1733 error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id));
1735 error = cpuset_setproc(-1, set, NULL, NULL, false);
1740 #ifndef _SYS_SYSPROTO_H_
1741 struct cpuset_setid_args {
1748 sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap)
1751 return (kern_cpuset_setid(td, uap->which, uap->id, uap->setid));
1755 kern_cpuset_setid(struct thread *td, cpuwhich_t which,
1756 id_t id, cpusetid_t setid)
1762 * Presently we only support per-process sets.
1764 if (which != CPU_WHICH_PID)
1766 set = cpuset_lookup(setid, td);
1769 error = cpuset_setproc(id, set, NULL, NULL, false);
1774 #ifndef _SYS_SYSPROTO_H_
1775 struct cpuset_getid_args {
1783 sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap)
1786 return (kern_cpuset_getid(td, uap->level, uap->which, uap->id,
1791 kern_cpuset_getid(struct thread *td, cpulevel_t level, cpuwhich_t which,
1792 id_t id, cpusetid_t *setid)
1794 struct cpuset *nset;
1801 if (level == CPU_LEVEL_WHICH && which != CPU_WHICH_CPUSET)
1803 error = cpuset_which(which, id, &p, &ttd, &set);
1810 set = cpuset_refbase(ttd->td_cpuset);
1814 case CPU_WHICH_CPUSET:
1815 case CPU_WHICH_JAIL:
1818 case CPU_WHICH_DOMAIN:
1822 case CPU_LEVEL_ROOT:
1823 nset = cpuset_refroot(set);
1827 case CPU_LEVEL_CPUSET:
1829 case CPU_LEVEL_WHICH:
1835 error = copyout(&tmpid, setid, sizeof(tmpid));
1840 #ifndef _SYS_SYSPROTO_H_
1841 struct cpuset_getaffinity_args {
1850 sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap)
1853 return (kern_cpuset_getaffinity(td, uap->level, uap->which,
1854 uap->id, uap->cpusetsize, uap->mask));
1858 kern_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1859 id_t id, size_t cpusetsize, cpuset_t *maskp)
1862 struct cpuset *nset;
1869 if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1871 error = cpuset_check_capabilities(td, level, which, id);
1875 mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO);
1876 error = cpuset_which(which, id, &p, &ttd, &set);
1880 case CPU_LEVEL_ROOT:
1881 case CPU_LEVEL_CPUSET:
1886 set = cpuset_ref(ttd->td_cpuset);
1889 case CPU_WHICH_CPUSET:
1890 case CPU_WHICH_JAIL:
1893 case CPU_WHICH_INTRHANDLER:
1894 case CPU_WHICH_ITHREAD:
1895 case CPU_WHICH_DOMAIN:
1899 if (level == CPU_LEVEL_ROOT)
1900 nset = cpuset_refroot(set);
1902 nset = cpuset_refbase(set);
1903 CPU_COPY(&nset->cs_mask, mask);
1906 case CPU_LEVEL_WHICH:
1910 CPU_COPY(&ttd->td_cpuset->cs_mask, mask);
1914 FOREACH_THREAD_IN_PROC(p, ttd) {
1916 CPU_OR(mask, &ttd->td_cpuset->cs_mask);
1920 case CPU_WHICH_CPUSET:
1921 case CPU_WHICH_JAIL:
1922 CPU_COPY(&set->cs_mask, mask);
1925 case CPU_WHICH_INTRHANDLER:
1926 case CPU_WHICH_ITHREAD:
1927 error = intr_getaffinity(id, which, mask);
1929 case CPU_WHICH_DOMAIN:
1930 if (id < 0 || id >= MAXMEMDOM)
1933 CPU_COPY(&cpuset_domain[id], mask);
1946 error = copyout(mask, maskp, size);
1952 #ifndef _SYS_SYSPROTO_H_
1953 struct cpuset_setaffinity_args {
1958 const cpuset_t *mask;
1962 sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap)
1965 return (kern_cpuset_setaffinity(td, uap->level, uap->which,
1966 uap->id, uap->cpusetsize, uap->mask));
1970 kern_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1971 id_t id, size_t cpusetsize, const cpuset_t *maskp)
1973 struct cpuset *nset;
1980 if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1982 error = cpuset_check_capabilities(td, level, which, id);
1985 mask = malloc(cpusetsize, M_TEMP, M_WAITOK | M_ZERO);
1986 error = copyin(maskp, mask, cpusetsize);
1990 * Verify that no high bits are set.
1992 if (cpusetsize > sizeof(cpuset_t)) {
1996 end = cp = (char *)&mask->__bits;
1998 cp += sizeof(cpuset_t);
2005 if (CPU_EMPTY(mask)) {
2010 case CPU_LEVEL_ROOT:
2011 case CPU_LEVEL_CPUSET:
2012 error = cpuset_which(which, id, &p, &ttd, &set);
2019 set = cpuset_ref(ttd->td_cpuset);
2023 case CPU_WHICH_CPUSET:
2024 case CPU_WHICH_JAIL:
2027 case CPU_WHICH_INTRHANDLER:
2028 case CPU_WHICH_ITHREAD:
2029 case CPU_WHICH_DOMAIN:
2033 if (level == CPU_LEVEL_ROOT)
2034 nset = cpuset_refroot(set);
2036 nset = cpuset_refbase(set);
2037 error = cpuset_modify(nset, mask);
2041 case CPU_LEVEL_WHICH:
2044 error = cpuset_setthread(id, mask);
2047 error = cpuset_setproc(id, NULL, mask, NULL, false);
2049 case CPU_WHICH_CPUSET:
2050 case CPU_WHICH_JAIL:
2051 error = cpuset_which(which, id, &p, &ttd, &set);
2053 error = cpuset_modify(set, mask);
2058 case CPU_WHICH_INTRHANDLER:
2059 case CPU_WHICH_ITHREAD:
2060 error = intr_setaffinity(id, which, mask);
2076 #ifndef _SYS_SYSPROTO_H_
2077 struct cpuset_getdomain_args {
2081 size_t domainsetsize;
2087 sys_cpuset_getdomain(struct thread *td, struct cpuset_getdomain_args *uap)
2090 return (kern_cpuset_getdomain(td, uap->level, uap->which,
2091 uap->id, uap->domainsetsize, uap->mask, uap->policy));
2095 kern_cpuset_getdomain(struct thread *td, cpulevel_t level, cpuwhich_t which,
2096 id_t id, size_t domainsetsize, domainset_t *maskp, int *policyp)
2098 struct domainset outset;
2100 struct cpuset *nset;
2102 struct domainset *dset;
2107 if (domainsetsize < sizeof(domainset_t) ||
2108 domainsetsize > DOMAINSET_MAXSIZE / NBBY)
2110 error = cpuset_check_capabilities(td, level, which, id);
2113 mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO);
2114 bzero(&outset, sizeof(outset));
2115 error = cpuset_which(which, id, &p, &ttd, &set);
2119 case CPU_LEVEL_ROOT:
2120 case CPU_LEVEL_CPUSET:
2125 set = cpuset_ref(ttd->td_cpuset);
2128 case CPU_WHICH_CPUSET:
2129 case CPU_WHICH_JAIL:
2132 case CPU_WHICH_INTRHANDLER:
2133 case CPU_WHICH_ITHREAD:
2134 case CPU_WHICH_DOMAIN:
2138 if (level == CPU_LEVEL_ROOT)
2139 nset = cpuset_refroot(set);
2141 nset = cpuset_refbase(set);
2142 domainset_copy(nset->cs_domain, &outset);
2145 case CPU_LEVEL_WHICH:
2149 domainset_copy(ttd->td_cpuset->cs_domain, &outset);
2153 FOREACH_THREAD_IN_PROC(p, ttd) {
2155 dset = ttd->td_cpuset->cs_domain;
2156 /* Show all domains in the proc. */
2157 DOMAINSET_OR(&outset.ds_mask, &dset->ds_mask);
2158 /* Last policy wins. */
2159 outset.ds_policy = dset->ds_policy;
2160 outset.ds_prefer = dset->ds_prefer;
2164 case CPU_WHICH_CPUSET:
2165 case CPU_WHICH_JAIL:
2166 domainset_copy(set->cs_domain, &outset);
2169 case CPU_WHICH_INTRHANDLER:
2170 case CPU_WHICH_ITHREAD:
2171 case CPU_WHICH_DOMAIN:
2185 * Translate prefer into a set containing only the preferred domain,
2186 * not the entire fallback set.
2188 if (outset.ds_policy == DOMAINSET_POLICY_PREFER) {
2189 DOMAINSET_ZERO(&outset.ds_mask);
2190 DOMAINSET_SET(outset.ds_prefer, &outset.ds_mask);
2192 DOMAINSET_COPY(&outset.ds_mask, mask);
2194 error = copyout(mask, maskp, domainsetsize);
2196 if (suword32(policyp, outset.ds_policy) != 0)
2203 #ifndef _SYS_SYSPROTO_H_
2204 struct cpuset_setdomain_args {
2208 size_t domainsetsize;
2214 sys_cpuset_setdomain(struct thread *td, struct cpuset_setdomain_args *uap)
2217 return (kern_cpuset_setdomain(td, uap->level, uap->which,
2218 uap->id, uap->domainsetsize, uap->mask, uap->policy));
2222 kern_cpuset_setdomain(struct thread *td, cpulevel_t level, cpuwhich_t which,
2223 id_t id, size_t domainsetsize, const domainset_t *maskp, int policy)
2225 struct cpuset *nset;
2229 struct domainset domain;
2233 if (domainsetsize < sizeof(domainset_t) ||
2234 domainsetsize > DOMAINSET_MAXSIZE / NBBY)
2236 if (policy <= DOMAINSET_POLICY_INVALID ||
2237 policy > DOMAINSET_POLICY_MAX)
2239 error = cpuset_check_capabilities(td, level, which, id);
2242 memset(&domain, 0, sizeof(domain));
2243 mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO);
2244 error = copyin(maskp, mask, domainsetsize);
2248 * Verify that no high bits are set.
2250 if (domainsetsize > sizeof(domainset_t)) {
2254 end = cp = (char *)&mask->__bits;
2255 end += domainsetsize;
2256 cp += sizeof(domainset_t);
2263 if (DOMAINSET_EMPTY(mask)) {
2267 DOMAINSET_COPY(mask, &domain.ds_mask);
2268 domain.ds_policy = policy;
2271 * Sanitize the provided mask.
2273 if (!DOMAINSET_SUBSET(&all_domains, &domain.ds_mask)) {
2278 /* Translate preferred policy into a mask and fallback. */
2279 if (policy == DOMAINSET_POLICY_PREFER) {
2280 /* Only support a single preferred domain. */
2281 if (DOMAINSET_COUNT(&domain.ds_mask) != 1) {
2285 domain.ds_prefer = DOMAINSET_FFS(&domain.ds_mask) - 1;
2286 /* This will be constrained by domainset_shadow(). */
2287 DOMAINSET_COPY(&all_domains, &domain.ds_mask);
2291 * When given an impossible policy, fall back to interleaving
2292 * across all domains.
2294 if (domainset_empty_vm(&domain))
2295 domainset_copy(&domainset2, &domain);
2298 case CPU_LEVEL_ROOT:
2299 case CPU_LEVEL_CPUSET:
2300 error = cpuset_which(which, id, &p, &ttd, &set);
2307 set = cpuset_ref(ttd->td_cpuset);
2311 case CPU_WHICH_CPUSET:
2312 case CPU_WHICH_JAIL:
2315 case CPU_WHICH_INTRHANDLER:
2316 case CPU_WHICH_ITHREAD:
2317 case CPU_WHICH_DOMAIN:
2321 if (level == CPU_LEVEL_ROOT)
2322 nset = cpuset_refroot(set);
2324 nset = cpuset_refbase(set);
2325 error = cpuset_modify_domain(nset, &domain);
2329 case CPU_LEVEL_WHICH:
2332 error = _cpuset_setthread(id, NULL, &domain);
2335 error = cpuset_setproc(id, NULL, NULL, &domain, false);
2337 case CPU_WHICH_CPUSET:
2338 case CPU_WHICH_JAIL:
2339 error = cpuset_which(which, id, &p, &ttd, &set);
2341 error = cpuset_modify_domain(set, &domain);
2346 case CPU_WHICH_INTRHANDLER:
2347 case CPU_WHICH_ITHREAD:
2365 ddb_display_bitset(const struct bitset *set, int size)
2369 for (once = 0, bit = 0; bit < size; bit++) {
2370 if (CPU_ISSET(bit, set)) {
2372 db_printf("%d", bit);
2375 db_printf(",%d", bit);
2379 db_printf("<none>");
2383 ddb_display_cpuset(const cpuset_t *set)
2385 ddb_display_bitset((const struct bitset *)set, CPU_SETSIZE);
2389 ddb_display_domainset(const domainset_t *set)
2391 ddb_display_bitset((const struct bitset *)set, DOMAINSET_SETSIZE);
2394 DB_SHOW_COMMAND(cpusets, db_show_cpusets)
2398 LIST_FOREACH(set, &cpuset_ids, cs_link) {
2399 db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n",
2400 set, set->cs_id, refcount_load(&set->cs_ref), set->cs_flags,
2401 (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0);
2402 db_printf(" cpu mask=");
2403 ddb_display_cpuset(&set->cs_mask);
2405 db_printf(" domain policy %d prefer %d mask=",
2406 set->cs_domain->ds_policy, set->cs_domain->ds_prefer);
2407 ddb_display_domainset(&set->cs_domain->ds_mask);
2414 DB_SHOW_COMMAND(domainsets, db_show_domainsets)
2416 struct domainset *set;
2418 LIST_FOREACH(set, &cpuset_domains, ds_link) {
2419 db_printf("set=%p policy %d prefer %d cnt %d\n",
2420 set, set->ds_policy, set->ds_prefer, set->ds_cnt);
2421 db_printf(" mask =");
2422 ddb_display_domainset(&set->ds_mask);