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_firsttouch;
123 struct domainset __read_mostly domainset_fixed[MAXMEMDOM];
124 struct domainset __read_mostly domainset_interleave;
125 struct domainset __read_mostly domainset_prefer[MAXMEMDOM];
126 struct domainset __read_mostly domainset_roundrobin;
128 static uma_zone_t cpuset_zone;
129 static uma_zone_t domainset_zone;
130 static struct mtx cpuset_lock;
131 static struct setlist cpuset_ids;
132 static struct domainlist cpuset_domains;
133 static struct unrhdr *cpuset_unr;
134 static struct cpuset *cpuset_zero, *cpuset_default, *cpuset_kernel;
135 static struct domainset *domainset0, *domainset2;
137 /* Return the size of cpuset_t at the kernel level */
138 SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD | CTLFLAG_CAPRD,
139 SYSCTL_NULL_INT_PTR, sizeof(cpuset_t), "sizeof(cpuset_t)");
141 cpuset_t *cpuset_root;
142 cpuset_t cpuset_domain[MAXMEMDOM];
144 static int domainset_valid(const struct domainset *, const struct domainset *);
147 * Find the first non-anonymous set starting from 'set'.
149 static struct cpuset *
150 cpuset_getbase(struct cpuset *set)
153 if (set->cs_id == CPUSET_INVALID)
154 set = set->cs_parent;
159 * Walks up the tree from 'set' to find the root.
161 static struct cpuset *
162 cpuset_getroot(struct cpuset *set)
165 while ((set->cs_flags & CPU_SET_ROOT) == 0 && set->cs_parent != NULL)
166 set = set->cs_parent;
171 * Acquire a reference to a cpuset, all pointers must be tracked with refs.
174 cpuset_ref(struct cpuset *set)
177 refcount_acquire(&set->cs_ref);
182 * Walks up the tree from 'set' to find the root. Returns the root
185 static struct cpuset *
186 cpuset_refroot(struct cpuset *set)
189 return (cpuset_ref(cpuset_getroot(set)));
193 * Find the first non-anonymous set starting from 'set'. Returns this set
194 * referenced. May return the passed in set with an extra ref if it is
197 static struct cpuset *
198 cpuset_refbase(struct cpuset *set)
201 return (cpuset_ref(cpuset_getbase(set)));
205 * Release a reference in a context where it is safe to allocate.
208 cpuset_rel(struct cpuset *set)
212 if (refcount_release_if_not_last(&set->cs_ref))
214 mtx_lock_spin(&cpuset_lock);
215 if (!refcount_release(&set->cs_ref)) {
216 mtx_unlock_spin(&cpuset_lock);
219 LIST_REMOVE(set, cs_siblings);
221 if (id != CPUSET_INVALID)
222 LIST_REMOVE(set, cs_link);
223 mtx_unlock_spin(&cpuset_lock);
224 cpuset_rel(set->cs_parent);
225 uma_zfree(cpuset_zone, set);
226 if (id != CPUSET_INVALID)
227 free_unr(cpuset_unr, id);
231 * Deferred release must be used when in a context that is not safe to
232 * allocate/free. This places any unreferenced sets on the list 'head'.
235 cpuset_rel_defer(struct setlist *head, struct cpuset *set)
238 if (refcount_release_if_not_last(&set->cs_ref))
240 mtx_lock_spin(&cpuset_lock);
241 if (!refcount_release(&set->cs_ref)) {
242 mtx_unlock_spin(&cpuset_lock);
245 LIST_REMOVE(set, cs_siblings);
246 if (set->cs_id != CPUSET_INVALID)
247 LIST_REMOVE(set, cs_link);
248 LIST_INSERT_HEAD(head, set, cs_link);
249 mtx_unlock_spin(&cpuset_lock);
253 * Complete a deferred release. Removes the set from the list provided to
257 cpuset_rel_complete(struct cpuset *set)
262 LIST_REMOVE(set, cs_link);
263 cpuset_rel(set->cs_parent);
264 uma_zfree(cpuset_zone, set);
265 if (id != CPUSET_INVALID)
266 free_unr(cpuset_unr, id);
270 * Find a set based on an id. Returns it with a ref.
272 static struct cpuset *
273 cpuset_lookup(cpusetid_t setid, struct thread *td)
277 if (setid == CPUSET_INVALID)
279 mtx_lock_spin(&cpuset_lock);
280 LIST_FOREACH(set, &cpuset_ids, cs_link)
281 if (set->cs_id == setid)
285 mtx_unlock_spin(&cpuset_lock);
287 KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__));
288 if (set != NULL && jailed(td->td_ucred)) {
289 struct cpuset *jset, *tset;
291 jset = td->td_ucred->cr_prison->pr_cpuset;
292 for (tset = set; tset != NULL; tset = tset->cs_parent)
305 * Initialize a set in the space provided in 'set' with the provided parameters.
306 * The set is returned with a single ref. May return EDEADLK if the set
307 * will have no valid cpu based on restrictions from the parent.
310 cpuset_init(struct cpuset *set, struct cpuset *parent,
311 const cpuset_t *mask, struct domainset *domain, cpusetid_t id)
315 domain = parent->cs_domain;
317 mask = &parent->cs_mask;
318 if (!CPU_OVERLAP(&parent->cs_mask, mask))
320 /* The domain must be prepared ahead of time. */
321 if (!domainset_valid(parent->cs_domain, domain))
323 CPU_COPY(mask, &set->cs_mask);
324 LIST_INIT(&set->cs_children);
325 refcount_init(&set->cs_ref, 1);
327 mtx_lock_spin(&cpuset_lock);
328 set->cs_domain = domain;
329 CPU_AND(&set->cs_mask, &parent->cs_mask);
331 set->cs_parent = cpuset_ref(parent);
332 LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings);
333 if (set->cs_id != CPUSET_INVALID)
334 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
335 mtx_unlock_spin(&cpuset_lock);
341 * Create a new non-anonymous set with the requested parent and mask. May
342 * return failures if the mask is invalid or a new number can not be
345 * If *setp is not NULL, then it will be used as-is. The caller must take
346 * into account that *setp will be inserted at the head of cpuset_ids and
347 * plan any potentially conflicting cs_link usage accordingly.
350 cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask)
357 id = alloc_unr(cpuset_unr);
360 dofree = (*setp == NULL);
364 *setp = set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
365 error = cpuset_init(set, parent, mask, NULL, id);
368 free_unr(cpuset_unr, id);
370 uma_zfree(cpuset_zone, set);
376 cpuset_freelist_add(struct setlist *list, int count)
381 for (i = 0; i < count; i++) {
382 set = uma_zalloc(cpuset_zone, M_ZERO | M_WAITOK);
383 LIST_INSERT_HEAD(list, set, cs_link);
388 cpuset_freelist_init(struct setlist *list, int count)
392 cpuset_freelist_add(list, count);
396 cpuset_freelist_free(struct setlist *list)
400 while ((set = LIST_FIRST(list)) != NULL) {
401 LIST_REMOVE(set, cs_link);
402 uma_zfree(cpuset_zone, set);
407 domainset_freelist_add(struct domainlist *list, int count)
409 struct domainset *set;
412 for (i = 0; i < count; i++) {
413 set = uma_zalloc(domainset_zone, M_ZERO | M_WAITOK);
414 LIST_INSERT_HEAD(list, set, ds_link);
419 domainset_freelist_init(struct domainlist *list, int count)
423 domainset_freelist_add(list, count);
427 domainset_freelist_free(struct domainlist *list)
429 struct domainset *set;
431 while ((set = LIST_FIRST(list)) != NULL) {
432 LIST_REMOVE(set, ds_link);
433 uma_zfree(domainset_zone, set);
437 /* Copy a domainset preserving mask and policy. */
439 domainset_copy(const struct domainset *from, struct domainset *to)
442 DOMAINSET_COPY(&from->ds_mask, &to->ds_mask);
443 to->ds_policy = from->ds_policy;
444 to->ds_prefer = from->ds_prefer;
447 /* Return 1 if mask and policy are equal, otherwise 0. */
449 domainset_equal(const struct domainset *one, const struct domainset *two)
452 return (DOMAINSET_CMP(&one->ds_mask, &two->ds_mask) == 0 &&
453 one->ds_policy == two->ds_policy &&
454 one->ds_prefer == two->ds_prefer);
457 /* Return 1 if child is a valid subset of parent. */
459 domainset_valid(const struct domainset *parent, const struct domainset *child)
461 if (child->ds_policy != DOMAINSET_POLICY_PREFER)
462 return (DOMAINSET_SUBSET(&parent->ds_mask, &child->ds_mask));
463 return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask));
467 domainset_restrict(const struct domainset *parent,
468 const struct domainset *child)
470 if (child->ds_policy != DOMAINSET_POLICY_PREFER)
471 return (DOMAINSET_OVERLAP(&parent->ds_mask, &child->ds_mask));
472 return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask));
476 * Lookup or create a domainset. The key is provided in ds_mask and
477 * ds_policy. If the domainset does not yet exist the storage in
478 * 'domain' is used to insert. Otherwise this storage is freed to the
479 * domainset_zone and the existing domainset is returned.
481 static struct domainset *
482 _domainset_create(struct domainset *domain, struct domainlist *freelist)
484 struct domainset *ndomain;
487 KASSERT(domain->ds_cnt <= vm_ndomains,
488 ("invalid domain count in domainset %p", domain));
489 KASSERT(domain->ds_policy != DOMAINSET_POLICY_PREFER ||
490 domain->ds_prefer < vm_ndomains,
491 ("invalid preferred domain in domains %p", domain));
493 mtx_lock_spin(&cpuset_lock);
494 LIST_FOREACH(ndomain, &cpuset_domains, ds_link)
495 if (domainset_equal(ndomain, domain))
498 * If the domain does not yet exist we insert it and initialize
499 * various iteration helpers which are not part of the key.
501 if (ndomain == NULL) {
502 LIST_INSERT_HEAD(&cpuset_domains, domain, ds_link);
503 domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask);
504 for (i = 0, j = 0; i < DOMAINSET_FLS(&domain->ds_mask); i++)
505 if (DOMAINSET_ISSET(i, &domain->ds_mask))
506 domain->ds_order[j++] = i;
508 mtx_unlock_spin(&cpuset_lock);
511 if (freelist != NULL)
512 LIST_INSERT_HEAD(freelist, domain, ds_link);
514 uma_zfree(domainset_zone, domain);
520 * Are any of the domains in the mask empty? If so, silently
521 * remove them and update the domainset accordingly. If only empty
522 * domains are present, we must return failure.
525 domainset_empty_vm(struct domainset *domain)
530 DOMAINSET_ZERO(&empty);
531 for (i = 0; i < vm_ndomains; i++)
532 if (VM_DOMAIN_EMPTY(i))
533 DOMAINSET_SET(i, &empty);
534 if (DOMAINSET_SUBSET(&empty, &domain->ds_mask))
537 /* Remove empty domains from the set and recompute. */
538 DOMAINSET_ANDNOT(&domain->ds_mask, &empty);
539 domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask);
540 for (i = j = 0; i < DOMAINSET_FLS(&domain->ds_mask); i++)
541 if (DOMAINSET_ISSET(i, &domain->ds_mask))
542 domain->ds_order[j++] = i;
544 /* Convert a PREFER policy referencing an empty domain to RR. */
545 if (domain->ds_policy == DOMAINSET_POLICY_PREFER &&
546 DOMAINSET_ISSET(domain->ds_prefer, &empty)) {
547 domain->ds_policy = DOMAINSET_POLICY_ROUNDROBIN;
548 domain->ds_prefer = -1;
555 * Create or lookup a domainset based on the key held in 'domain'.
558 domainset_create(const struct domainset *domain)
560 struct domainset *ndomain;
563 * Validate the policy. It must specify a useable policy number with
564 * only valid domains. Preferred must include the preferred domain
567 if (domain->ds_policy <= DOMAINSET_POLICY_INVALID ||
568 domain->ds_policy > DOMAINSET_POLICY_MAX)
570 if (domain->ds_policy == DOMAINSET_POLICY_PREFER &&
571 !DOMAINSET_ISSET(domain->ds_prefer, &domain->ds_mask))
573 if (!DOMAINSET_SUBSET(&domainset0->ds_mask, &domain->ds_mask))
575 ndomain = uma_zalloc(domainset_zone, M_WAITOK | M_ZERO);
576 domainset_copy(domain, ndomain);
577 return _domainset_create(ndomain, NULL);
581 * Update thread domainset pointers.
584 domainset_notify(void)
589 sx_slock(&allproc_lock);
590 FOREACH_PROC_IN_SYSTEM(p) {
592 if (p->p_state == PRS_NEW) {
596 FOREACH_THREAD_IN_PROC(p, td) {
598 td->td_domain.dr_policy = td->td_cpuset->cs_domain;
603 sx_sunlock(&allproc_lock);
604 kernel_object->domain.dr_policy = cpuset_kernel->cs_domain;
608 * Create a new set that is a subset of a parent.
610 static struct domainset *
611 domainset_shadow(const struct domainset *pdomain,
612 const struct domainset *domain, struct domainlist *freelist)
614 struct domainset *ndomain;
616 ndomain = LIST_FIRST(freelist);
617 LIST_REMOVE(ndomain, ds_link);
620 * Initialize the key from the request.
622 domainset_copy(domain, ndomain);
625 * Restrict the key by the parent.
627 DOMAINSET_AND(&ndomain->ds_mask, &pdomain->ds_mask);
629 return _domainset_create(ndomain, freelist);
633 * Recursively check for errors that would occur from applying mask to
634 * the tree of sets starting at 'set'. Checks for sets that would become
635 * empty as well as RDONLY flags.
638 cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int augment_mask)
644 mtx_assert(&cpuset_lock, MA_OWNED);
645 if (set->cs_flags & CPU_SET_RDONLY)
648 CPU_COPY(&set->cs_mask, &newmask);
649 CPU_AND(&newmask, mask);
651 CPU_COPY(mask, &newmask);
653 if (CPU_EMPTY(&newmask))
656 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
657 if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0)
663 * Applies the mask 'mask' without checking for empty sets or permissions.
666 cpuset_update(struct cpuset *set, cpuset_t *mask)
670 mtx_assert(&cpuset_lock, MA_OWNED);
671 CPU_AND(&set->cs_mask, mask);
672 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
673 cpuset_update(nset, &set->cs_mask);
679 * Modify the set 'set' to use a copy of the mask provided. Apply this new
680 * mask to restrict all children in the tree. Checks for validity before
681 * applying the changes.
684 cpuset_modify(struct cpuset *set, cpuset_t *mask)
689 error = priv_check(curthread, PRIV_SCHED_CPUSET);
693 * In case we are called from within the jail,
694 * we do not allow modifying the dedicated root
695 * cpuset of the jail but may still allow to
696 * change child sets, including subordinate jails'
699 if ((set->cs_flags & CPU_SET_ROOT) != 0 &&
700 jailed(curthread->td_ucred) &&
701 set == curthread->td_ucred->cr_prison->pr_cpuset)
704 * Verify that we have access to this set of
707 if ((set->cs_flags & (CPU_SET_ROOT | CPU_SET_RDONLY)) == CPU_SET_ROOT) {
708 KASSERT(set->cs_parent != NULL,
709 ("jail.cpuset=%d is not a proper child of parent jail's root.",
713 * cpuset_getroot() cannot work here due to how top-level jail
714 * roots are constructed. Top-level jails are parented to
715 * thread0's cpuset (i.e. cpuset 1) rather than the system root.
717 root = set->cs_parent;
719 root = cpuset_getroot(set);
721 mtx_lock_spin(&cpuset_lock);
722 if (root && !CPU_SUBSET(&root->cs_mask, mask)) {
726 error = cpuset_testupdate(set, mask, 0);
729 CPU_COPY(mask, &set->cs_mask);
730 cpuset_update(set, mask);
732 mtx_unlock_spin(&cpuset_lock);
738 * Recursively check for errors that would occur from applying mask to
739 * the tree of sets starting at 'set'. Checks for sets that would become
740 * empty as well as RDONLY flags.
743 cpuset_testupdate_domain(struct cpuset *set, struct domainset *dset,
744 struct domainset *orig, int *count, int augment_mask __unused)
747 struct domainset *domain;
748 struct domainset newset;
751 mtx_assert(&cpuset_lock, MA_OWNED);
752 if (set->cs_flags & CPU_SET_RDONLY)
754 domain = set->cs_domain;
755 domainset_copy(domain, &newset);
756 if (!domainset_equal(domain, orig)) {
757 if (!domainset_restrict(domain, dset))
759 DOMAINSET_AND(&newset.ds_mask, &dset->ds_mask);
760 /* Count the number of domains that are changing. */
764 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
765 if ((error = cpuset_testupdate_domain(nset, &newset, domain,
772 * Applies the mask 'mask' without checking for empty sets or permissions.
775 cpuset_update_domain(struct cpuset *set, struct domainset *domain,
776 struct domainset *orig, struct domainlist *domains)
780 mtx_assert(&cpuset_lock, MA_OWNED);
782 * If this domainset has changed from the parent we must calculate
783 * a new set. Otherwise it simply inherits from the parent. When
784 * we inherit from the parent we get a new mask and policy. If the
785 * set is modified from the parent we keep the policy and only
788 if (set->cs_domain != orig) {
789 orig = set->cs_domain;
790 set->cs_domain = domainset_shadow(domain, orig, domains);
792 set->cs_domain = domain;
793 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
794 cpuset_update_domain(nset, set->cs_domain, orig, domains);
800 * Modify the set 'set' to use a copy the domainset provided. Apply this new
801 * mask to restrict all children in the tree. Checks for validity before
802 * applying the changes.
805 cpuset_modify_domain(struct cpuset *set, struct domainset *domain)
807 struct domainlist domains;
808 struct domainset temp;
809 struct domainset *dset;
811 int ndomains, needed;
814 error = priv_check(curthread, PRIV_SCHED_CPUSET);
818 * In case we are called from within the jail
819 * we do not allow modifying the dedicated root
820 * cpuset of the jail but may still allow to
823 if (jailed(curthread->td_ucred) &&
824 set->cs_flags & CPU_SET_ROOT)
826 domainset_freelist_init(&domains, 0);
827 domain = domainset_create(domain);
830 mtx_lock_spin(&cpuset_lock);
832 root = cpuset_getroot(set);
833 dset = root->cs_domain;
835 * Verify that we have access to this set of domains.
837 if (!domainset_valid(dset, domain)) {
842 * If applying prefer we keep the current set as the fallback.
844 if (domain->ds_policy == DOMAINSET_POLICY_PREFER)
845 DOMAINSET_COPY(&set->cs_domain->ds_mask,
848 * Determine whether we can apply this set of domains and
849 * how many new domain structures it will require.
851 domainset_copy(domain, &temp);
853 error = cpuset_testupdate_domain(set, &temp, set->cs_domain,
857 if (ndomains >= needed)
860 /* Dropping the lock; we'll need to re-evaluate again. */
861 mtx_unlock_spin(&cpuset_lock);
862 domainset_freelist_add(&domains, needed - ndomains);
864 mtx_lock_spin(&cpuset_lock);
866 dset = set->cs_domain;
867 cpuset_update_domain(set, domain, dset, &domains);
869 mtx_unlock_spin(&cpuset_lock);
870 domainset_freelist_free(&domains);
878 * Resolve the 'which' parameter of several cpuset apis.
880 * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid. Also
881 * checks for permission via p_cansched().
883 * For WHICH_SET returns a valid set with a new reference.
885 * -1 may be supplied for any argument to mean the current proc/thread or
886 * the base set of the current thread. May fail with ESRCH/EPERM.
889 cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp,
890 struct cpuset **setp)
907 if ((p = pfind(id)) == NULL)
922 case CPU_WHICH_CPUSET:
924 thread_lock(curthread);
925 set = cpuset_refbase(curthread->td_cpuset);
926 thread_unlock(curthread);
928 set = cpuset_lookup(id, curthread);
936 /* Find `set' for prison with given id. */
939 sx_slock(&allprison_lock);
940 pr = prison_find_child(curthread->td_ucred->cr_prison, id);
941 sx_sunlock(&allprison_lock);
944 cpuset_ref(pr->pr_cpuset);
945 *setp = pr->pr_cpuset;
946 mtx_unlock(&pr->pr_mtx);
950 case CPU_WHICH_DOMAIN:
955 error = p_cansched(curthread, p);
961 td = FIRST_THREAD_IN_PROC(p);
968 cpuset_testshadow(struct cpuset *set, const cpuset_t *mask,
969 const struct domainset *domain)
971 struct cpuset *parent;
972 struct domainset *dset;
974 parent = cpuset_getbase(set);
976 * If we are restricting a cpu mask it must be a subset of the
977 * parent or invalid CPUs have been specified.
979 if (mask != NULL && !CPU_SUBSET(&parent->cs_mask, mask))
983 * If we are restricting a domain mask it must be a subset of the
984 * parent or invalid domains have been specified.
986 dset = parent->cs_domain;
987 if (domain != NULL && !domainset_valid(dset, domain))
994 * Create an anonymous set with the provided mask in the space provided by
995 * 'nset'. If the passed in set is anonymous we use its parent otherwise
996 * the new set is a child of 'set'.
999 cpuset_shadow(struct cpuset *set, struct cpuset **nsetp,
1000 const cpuset_t *mask, const struct domainset *domain,
1001 struct setlist *cpusets, struct domainlist *domains)
1003 struct cpuset *parent;
1004 struct cpuset *nset;
1005 struct domainset *dset;
1006 struct domainset *d;
1009 error = cpuset_testshadow(set, mask, domain);
1013 parent = cpuset_getbase(set);
1014 dset = parent->cs_domain;
1016 mask = &set->cs_mask;
1018 d = domainset_shadow(dset, domain, domains);
1021 nset = LIST_FIRST(cpusets);
1022 error = cpuset_init(nset, parent, mask, d, CPUSET_INVALID);
1024 LIST_REMOVE(nset, cs_link);
1030 static struct cpuset *
1031 cpuset_update_thread(struct thread *td, struct cpuset *nset)
1033 struct cpuset *tdset;
1035 tdset = td->td_cpuset;
1036 td->td_cpuset = nset;
1037 td->td_domain.dr_policy = nset->cs_domain;
1044 cpuset_setproc_test_maskthread(struct cpuset *tdset, cpuset_t *mask,
1045 struct domainset *domain)
1047 struct cpuset *parent;
1049 parent = cpuset_getbase(tdset);
1051 mask = &tdset->cs_mask;
1053 domain = tdset->cs_domain;
1054 return cpuset_testshadow(parent, mask, domain);
1058 cpuset_setproc_maskthread(struct cpuset *tdset, cpuset_t *mask,
1059 struct domainset *domain, struct cpuset **nsetp,
1060 struct setlist *freelist, struct domainlist *domainlist)
1062 struct cpuset *parent;
1064 parent = cpuset_getbase(tdset);
1066 mask = &tdset->cs_mask;
1068 domain = tdset->cs_domain;
1069 return cpuset_shadow(parent, nsetp, mask, domain, freelist,
1074 cpuset_setproc_setthread_mask(struct cpuset *tdset, struct cpuset *set,
1075 cpuset_t *mask, struct domainset *domain)
1077 struct cpuset *parent;
1079 parent = cpuset_getbase(tdset);
1082 * If the thread restricted its mask then apply that same
1083 * restriction to the new set, otherwise take it wholesale.
1085 if (CPU_CMP(&tdset->cs_mask, &parent->cs_mask) != 0) {
1086 CPU_COPY(&tdset->cs_mask, mask);
1087 CPU_AND(mask, &set->cs_mask);
1089 CPU_COPY(&set->cs_mask, mask);
1092 * If the thread restricted the domain then we apply the
1093 * restriction to the new set but retain the policy.
1095 if (tdset->cs_domain != parent->cs_domain) {
1096 domainset_copy(tdset->cs_domain, domain);
1097 DOMAINSET_AND(&domain->ds_mask, &set->cs_domain->ds_mask);
1099 domainset_copy(set->cs_domain, domain);
1101 if (CPU_EMPTY(mask) || DOMAINSET_EMPTY(&domain->ds_mask))
1108 cpuset_setproc_test_setthread(struct cpuset *tdset, struct cpuset *set)
1110 struct domainset domain;
1113 if (tdset->cs_id != CPUSET_INVALID)
1115 return cpuset_setproc_setthread_mask(tdset, set, &mask, &domain);
1119 cpuset_setproc_setthread(struct cpuset *tdset, struct cpuset *set,
1120 struct cpuset **nsetp, struct setlist *freelist,
1121 struct domainlist *domainlist)
1123 struct domainset domain;
1128 * If we're replacing on a thread that has not constrained the
1129 * original set we can simply accept the new set.
1131 if (tdset->cs_id != CPUSET_INVALID) {
1132 *nsetp = cpuset_ref(set);
1135 error = cpuset_setproc_setthread_mask(tdset, set, &mask, &domain);
1139 return cpuset_shadow(set, nsetp, &mask, &domain, freelist,
1144 cpuset_setproc_newbase(struct thread *td, struct cpuset *set,
1145 struct cpuset *nroot, struct cpuset **nsetp,
1146 struct setlist *cpusets, struct domainlist *domainlist)
1148 struct domainset ndomain;
1150 struct cpuset *pbase;
1153 pbase = cpuset_getbase(td->td_cpuset);
1155 /* Copy process mask, then further apply the new root mask. */
1156 CPU_COPY(&pbase->cs_mask, &nmask);
1157 CPU_AND(&nmask, &nroot->cs_mask);
1159 domainset_copy(pbase->cs_domain, &ndomain);
1160 DOMAINSET_AND(&ndomain.ds_mask, &set->cs_domain->ds_mask);
1162 /* Policy is too restrictive, will not work. */
1163 if (CPU_EMPTY(&nmask) || DOMAINSET_EMPTY(&ndomain.ds_mask))
1167 * Remove pbase from the freelist in advance, it'll be pushed to
1168 * cpuset_ids on success. We assume here that cpuset_create() will not
1169 * touch pbase on failure, and we just enqueue it back to the freelist
1170 * to remain in a consistent state.
1172 pbase = LIST_FIRST(cpusets);
1173 LIST_REMOVE(pbase, cs_link);
1174 error = cpuset_create(&pbase, set, &nmask);
1176 LIST_INSERT_HEAD(cpusets, pbase, cs_link);
1180 /* Duplicates some work from above... oh well. */
1181 pbase->cs_domain = domainset_shadow(set->cs_domain, &ndomain,
1188 * Handle four cases for updating an entire process.
1190 * 1) Set is non-null and the process is not rebasing onto a new root. This
1191 * reparents all anonymous sets to the provided set and replaces all
1192 * non-anonymous td_cpusets with the provided set.
1193 * 2) Set is non-null and the process is rebasing onto a new root. This
1194 * creates a new base set if the process previously had its own base set,
1195 * then reparents all anonymous sets either to that set or the provided set
1196 * if one was not created. Non-anonymous sets are similarly replaced.
1197 * 3) Mask is non-null. This replaces or creates anonymous sets for every
1198 * thread with the existing base as a parent.
1199 * 4) domain is non-null. This creates anonymous sets for every thread
1200 * and replaces the domain set.
1202 * This is overly complicated because we can't allocate while holding a
1203 * spinlock and spinlocks must be held while changing and examining thread
1207 cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask,
1208 struct domainset *domain, bool rebase)
1210 struct setlist freelist;
1211 struct setlist droplist;
1212 struct domainlist domainlist;
1213 struct cpuset *base, *nset, *nroot, *tdroot;
1221 * The algorithm requires two passes due to locking considerations.
1223 * 1) Lookup the process and acquire the locks in the required order.
1224 * 2) If enough cpusets have not been allocated release the locks and
1225 * allocate them. Loop.
1227 cpuset_freelist_init(&freelist, 1);
1228 domainset_freelist_init(&domainlist, 1);
1230 LIST_INIT(&droplist);
1235 nroot = cpuset_getroot(set);
1237 error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset);
1240 tdroot = cpuset_getroot(td->td_cpuset);
1241 needed = p->p_numthreads;
1242 if (set != NULL && rebase && tdroot != nroot)
1244 if (nfree >= needed)
1247 if (nfree < needed) {
1248 cpuset_freelist_add(&freelist, needed - nfree);
1249 domainset_freelist_add(&domainlist, needed - nfree);
1253 PROC_LOCK_ASSERT(p, MA_OWNED);
1256 * If we're changing roots and the root set is what has been specified
1257 * as the parent, then we'll check if the process was previously using
1258 * the root set and, if it wasn't, create a new base with the process's
1259 * mask applied to it.
1261 * If the new root is incompatible with the existing mask, then we allow
1262 * the process to take on the new root if and only if they have
1263 * privilege to widen their mask anyways. Unprivileged processes get
1264 * rejected with EDEADLK.
1266 if (set != NULL && rebase && nroot != tdroot) {
1267 cpusetid_t base_id, root_id;
1269 root_id = td->td_ucred->cr_prison->pr_cpuset->cs_id;
1270 base_id = cpuset_getbase(td->td_cpuset)->cs_id;
1272 if (base_id != root_id) {
1273 error = cpuset_setproc_newbase(td, set, nroot, &base,
1274 &freelist, &domainlist);
1275 if (error == EDEADLK &&
1276 priv_check(td, PRIV_SCHED_CPUSET) == 0)
1284 * Now that the appropriate locks are held and we have enough cpusets,
1285 * make sure the operation will succeed before applying changes. The
1286 * proc lock prevents td_cpuset from changing between calls.
1289 FOREACH_THREAD_IN_PROC(p, td) {
1292 error = cpuset_setproc_test_setthread(td->td_cpuset,
1295 error = cpuset_setproc_test_maskthread(td->td_cpuset,
1302 * Replace each thread's cpuset while using deferred release. We
1303 * must do this because the thread lock must be held while operating
1304 * on the thread and this limits the type of operations allowed.
1306 FOREACH_THREAD_IN_PROC(p, td) {
1309 error = cpuset_setproc_setthread(td->td_cpuset, base,
1310 &nset, &freelist, &domainlist);
1312 error = cpuset_setproc_maskthread(td->td_cpuset, mask,
1313 domain, &nset, &freelist, &domainlist);
1318 cpuset_rel_defer(&droplist, cpuset_update_thread(td, nset));
1324 if (base != NULL && base != set)
1326 while ((nset = LIST_FIRST(&droplist)) != NULL)
1327 cpuset_rel_complete(nset);
1328 cpuset_freelist_free(&freelist);
1329 domainset_freelist_free(&domainlist);
1334 bitset_strprint(char *buf, size_t bufsiz, const struct bitset *set, int setlen)
1342 for (i = 0; i < __bitset_words(setlen); i++) {
1351 if (bufsiz < sizeof(__STRING(ULONG_MAX)))
1353 bytes = snprintf(p, bufsiz, "%lx", set->__bits[i]);
1361 bitset_strscan(struct bitset *set, int setlen, const char *buf)
1366 BIT_ZERO(setlen, set);
1368 for (i = 0; i < __bitset_words(setlen); i++) {
1373 ret = sscanf(p, "%lx", &set->__bits[i]);
1374 if (ret == 0 || ret == -1)
1376 while (isxdigit(*p))
1383 * Return a string representing a valid layout for a cpuset_t object.
1384 * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
1387 cpusetobj_strprint(char *buf, const cpuset_t *set)
1390 bitset_strprint(buf, CPUSETBUFSIZ, (const struct bitset *)set,
1396 * Build a valid cpuset_t object from a string representation.
1397 * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
1400 cpusetobj_strscan(cpuset_t *set, const char *buf)
1404 if (strlen(buf) > CPUSETBUFSIZ - 1)
1407 p = buf[bitset_strscan((struct bitset *)set, CPU_SETSIZE, buf)];
1415 * Handle a domainset specifier in the sysctl tree. A poiner to a pointer to
1416 * a domainset is in arg1. If the user specifies a valid domainset the
1417 * pointer is updated.
1420 * hex mask word 0,hex mask word 1,...:decimal policy:decimal preferred
1423 sysctl_handle_domainset(SYSCTL_HANDLER_ARGS)
1425 char buf[DOMAINSETBUFSIZ];
1426 struct domainset *dset;
1427 struct domainset key;
1428 int policy, prefer, error;
1431 dset = *(struct domainset **)arg1;
1435 p = buf + bitset_strprint(buf, DOMAINSETBUFSIZ,
1436 (const struct bitset *)&dset->ds_mask, DOMAINSET_SETSIZE);
1437 sprintf(p, ":%d:%d", dset->ds_policy, dset->ds_prefer);
1439 sprintf(buf, "<NULL>");
1440 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
1441 if (error != 0 || req->newptr == NULL)
1445 * Read in and validate the string.
1447 memset(&key, 0, sizeof(key));
1448 p = &buf[bitset_strscan((struct bitset *)&key.ds_mask,
1449 DOMAINSET_SETSIZE, buf)];
1452 if (sscanf(p, ":%d:%d", &policy, &prefer) != 2)
1454 key.ds_policy = policy;
1455 key.ds_prefer = prefer;
1457 /* Domainset_create() validates the policy.*/
1458 dset = domainset_create(&key);
1461 *(struct domainset **)arg1 = dset;
1467 * Apply an anonymous mask or a domain to a single thread.
1470 _cpuset_setthread(lwpid_t id, cpuset_t *mask, struct domainset *domain)
1472 struct setlist cpusets;
1473 struct domainlist domainlist;
1474 struct cpuset *nset;
1480 cpuset_freelist_init(&cpusets, 1);
1481 domainset_freelist_init(&domainlist, domain != NULL);
1482 error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set);
1487 error = cpuset_shadow(td->td_cpuset, &nset, mask, domain,
1488 &cpusets, &domainlist);
1490 set = cpuset_update_thread(td, nset);
1496 cpuset_freelist_free(&cpusets);
1497 domainset_freelist_free(&domainlist);
1502 * Apply an anonymous mask to a single thread.
1505 cpuset_setthread(lwpid_t id, cpuset_t *mask)
1508 return _cpuset_setthread(id, mask, NULL);
1512 * Apply new cpumask to the ithread.
1515 cpuset_setithread(lwpid_t id, int cpu)
1521 CPU_COPY(cpuset_root, &mask);
1523 CPU_SET(cpu, &mask);
1524 return _cpuset_setthread(id, &mask, NULL);
1528 * Initialize static domainsets after NUMA information is available. This is
1529 * called before memory allocators are initialized.
1532 domainset_init(void)
1534 struct domainset *dset;
1537 dset = &domainset_firsttouch;
1538 DOMAINSET_COPY(&all_domains, &dset->ds_mask);
1539 dset->ds_policy = DOMAINSET_POLICY_FIRSTTOUCH;
1540 dset->ds_prefer = -1;
1541 _domainset_create(dset, NULL);
1543 dset = &domainset_interleave;
1544 DOMAINSET_COPY(&all_domains, &dset->ds_mask);
1545 dset->ds_policy = DOMAINSET_POLICY_INTERLEAVE;
1546 dset->ds_prefer = -1;
1547 _domainset_create(dset, NULL);
1549 dset = &domainset_roundrobin;
1550 DOMAINSET_COPY(&all_domains, &dset->ds_mask);
1551 dset->ds_policy = DOMAINSET_POLICY_ROUNDROBIN;
1552 dset->ds_prefer = -1;
1553 _domainset_create(dset, NULL);
1555 for (i = 0; i < vm_ndomains; i++) {
1556 dset = &domainset_fixed[i];
1557 DOMAINSET_ZERO(&dset->ds_mask);
1558 DOMAINSET_SET(i, &dset->ds_mask);
1559 dset->ds_policy = DOMAINSET_POLICY_ROUNDROBIN;
1560 _domainset_create(dset, NULL);
1562 dset = &domainset_prefer[i];
1563 DOMAINSET_COPY(&all_domains, &dset->ds_mask);
1564 dset->ds_policy = DOMAINSET_POLICY_PREFER;
1565 dset->ds_prefer = i;
1566 _domainset_create(dset, NULL);
1571 * Define the domainsets for cpuset 0, 1 and cpuset 2.
1574 domainset_zero(void)
1576 struct domainset *dset, *tmp;
1578 mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE);
1580 domainset0 = &domainset_firsttouch;
1581 curthread->td_domain.dr_policy = domainset0;
1583 domainset2 = &domainset_interleave;
1584 kernel_object->domain.dr_policy = domainset2;
1586 /* Remove empty domains from the global policies. */
1587 LIST_FOREACH_SAFE(dset, &cpuset_domains, ds_link, tmp)
1588 if (domainset_empty_vm(dset))
1589 LIST_REMOVE(dset, ds_link);
1593 * Creates system-wide cpusets and the cpuset for thread0 including three
1596 * 0 - The root set which should represent all valid processors in the
1597 * system. This set is immutable.
1598 * 1 - The default set which all processes are a member of until changed.
1599 * This allows an administrator to move all threads off of given cpus to
1600 * dedicate them to high priority tasks or save power etc.
1601 * 2 - The kernel set which allows restriction and policy to be applied only
1602 * to kernel threads and the kernel_object.
1605 cpuset_thread0(void)
1611 cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL,
1612 NULL, NULL, UMA_ALIGN_CACHE, 0);
1613 domainset_zone = uma_zcreate("domainset", sizeof(struct domainset),
1614 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
1617 * Create the root system set (0) for the whole machine. Doesn't use
1618 * cpuset_create() due to NULL parent.
1620 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1621 CPU_COPY(&all_cpus, &set->cs_mask);
1622 LIST_INIT(&set->cs_children);
1623 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
1624 refcount_init(&set->cs_ref, 1);
1625 set->cs_flags = CPU_SET_ROOT | CPU_SET_RDONLY;
1626 set->cs_domain = domainset0;
1628 cpuset_root = &set->cs_mask;
1631 * Now derive a default (1), modifiable set from that to give out.
1633 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1634 error = cpuset_init(set, cpuset_zero, NULL, NULL, 1);
1635 KASSERT(error == 0, ("Error creating default set: %d\n", error));
1636 cpuset_default = set;
1638 * Create the kernel set (2).
1640 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1641 error = cpuset_init(set, cpuset_zero, NULL, NULL, 2);
1642 KASSERT(error == 0, ("Error creating kernel set: %d\n", error));
1643 set->cs_domain = domainset2;
1644 cpuset_kernel = set;
1647 * Initialize the unit allocator. 0 and 1 are allocated above.
1649 cpuset_unr = new_unrhdr(3, INT_MAX, NULL);
1652 * If MD code has not initialized per-domain cpusets, place all
1655 for (i = 0; i < MAXMEMDOM; i++)
1656 if (!CPU_EMPTY(&cpuset_domain[i]))
1658 CPU_COPY(&all_cpus, &cpuset_domain[0]);
1661 return (cpuset_default);
1665 cpuset_kernthread(struct thread *td)
1670 set = td->td_cpuset;
1671 td->td_cpuset = cpuset_ref(cpuset_kernel);
1677 * Create a cpuset, which would be cpuset_create() but
1678 * mark the new 'set' as root.
1680 * We are not going to reparent the td to it. Use cpuset_setproc_update_set()
1683 * In case of no error, returns the set in *setp locked with a reference.
1686 cpuset_create_root(struct prison *pr, struct cpuset **setp)
1691 KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__));
1692 KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__));
1695 error = cpuset_create(&set, pr->pr_cpuset, &pr->pr_cpuset->cs_mask);
1699 KASSERT(set != NULL, ("[%s:%d] cpuset_create returned invalid data",
1700 __func__, __LINE__));
1702 /* Mark the set as root. */
1703 set->cs_flags |= CPU_SET_ROOT;
1710 cpuset_setproc_update_set(struct proc *p, struct cpuset *set)
1714 KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__));
1715 KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__));
1718 error = cpuset_setproc(p->p_pid, set, NULL, NULL, true);
1726 * In Capability mode, the only accesses that are permitted are to the current
1727 * thread and process' CPU and domain sets.
1730 cpuset_check_capabilities(struct thread *td, cpulevel_t level, cpuwhich_t which,
1733 if (IN_CAPABILITY_MODE(td)) {
1734 if (level != CPU_LEVEL_WHICH)
1736 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1739 !(which == CPU_WHICH_TID && id == td->td_tid) &&
1740 !(which == CPU_WHICH_PID && id == td->td_proc->p_pid))
1746 #ifndef _SYS_SYSPROTO_H_
1747 struct cpuset_args {
1752 sys_cpuset(struct thread *td, struct cpuset_args *uap)
1754 struct cpuset *root;
1759 root = cpuset_refroot(td->td_cpuset);
1762 error = cpuset_create(&set, root, &root->cs_mask);
1766 error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id));
1768 error = cpuset_setproc(-1, set, NULL, NULL, false);
1773 #ifndef _SYS_SYSPROTO_H_
1774 struct cpuset_setid_args {
1781 sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap)
1784 return (kern_cpuset_setid(td, uap->which, uap->id, uap->setid));
1788 kern_cpuset_setid(struct thread *td, cpuwhich_t which,
1789 id_t id, cpusetid_t setid)
1795 * Presently we only support per-process sets.
1797 if (which != CPU_WHICH_PID)
1799 set = cpuset_lookup(setid, td);
1802 error = cpuset_setproc(id, set, NULL, NULL, false);
1807 #ifndef _SYS_SYSPROTO_H_
1808 struct cpuset_getid_args {
1816 sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap)
1819 return (kern_cpuset_getid(td, uap->level, uap->which, uap->id,
1824 kern_cpuset_getid(struct thread *td, cpulevel_t level, cpuwhich_t which,
1825 id_t id, cpusetid_t *setid)
1827 struct cpuset *nset;
1834 if (level == CPU_LEVEL_WHICH && which != CPU_WHICH_CPUSET)
1836 error = cpuset_which(which, id, &p, &ttd, &set);
1843 set = cpuset_refbase(ttd->td_cpuset);
1847 case CPU_WHICH_CPUSET:
1848 case CPU_WHICH_JAIL:
1851 case CPU_WHICH_DOMAIN:
1855 case CPU_LEVEL_ROOT:
1856 nset = cpuset_refroot(set);
1860 case CPU_LEVEL_CPUSET:
1862 case CPU_LEVEL_WHICH:
1868 error = copyout(&tmpid, setid, sizeof(tmpid));
1873 #ifndef _SYS_SYSPROTO_H_
1874 struct cpuset_getaffinity_args {
1883 sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap)
1886 return (kern_cpuset_getaffinity(td, uap->level, uap->which,
1887 uap->id, uap->cpusetsize, uap->mask));
1891 kern_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1892 id_t id, size_t cpusetsize, cpuset_t *maskp)
1895 struct cpuset *nset;
1902 if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1904 error = cpuset_check_capabilities(td, level, which, id);
1908 mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO);
1909 error = cpuset_which(which, id, &p, &ttd, &set);
1913 case CPU_LEVEL_ROOT:
1914 case CPU_LEVEL_CPUSET:
1919 set = cpuset_ref(ttd->td_cpuset);
1922 case CPU_WHICH_CPUSET:
1923 case CPU_WHICH_JAIL:
1926 case CPU_WHICH_INTRHANDLER:
1927 case CPU_WHICH_ITHREAD:
1928 case CPU_WHICH_DOMAIN:
1932 if (level == CPU_LEVEL_ROOT)
1933 nset = cpuset_refroot(set);
1935 nset = cpuset_refbase(set);
1936 CPU_COPY(&nset->cs_mask, mask);
1939 case CPU_LEVEL_WHICH:
1943 CPU_COPY(&ttd->td_cpuset->cs_mask, mask);
1947 FOREACH_THREAD_IN_PROC(p, ttd) {
1949 CPU_OR(mask, &ttd->td_cpuset->cs_mask);
1953 case CPU_WHICH_CPUSET:
1954 case CPU_WHICH_JAIL:
1955 CPU_COPY(&set->cs_mask, mask);
1958 case CPU_WHICH_INTRHANDLER:
1959 case CPU_WHICH_ITHREAD:
1960 error = intr_getaffinity(id, which, mask);
1962 case CPU_WHICH_DOMAIN:
1963 if (id < 0 || id >= MAXMEMDOM)
1966 CPU_COPY(&cpuset_domain[id], mask);
1979 error = copyout(mask, maskp, size);
1985 #ifndef _SYS_SYSPROTO_H_
1986 struct cpuset_setaffinity_args {
1991 const cpuset_t *mask;
1995 sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap)
1998 return (kern_cpuset_setaffinity(td, uap->level, uap->which,
1999 uap->id, uap->cpusetsize, uap->mask));
2003 kern_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
2004 id_t id, size_t cpusetsize, const cpuset_t *maskp)
2006 struct cpuset *nset;
2013 if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
2015 error = cpuset_check_capabilities(td, level, which, id);
2018 mask = malloc(cpusetsize, M_TEMP, M_WAITOK | M_ZERO);
2019 error = copyin(maskp, mask, cpusetsize);
2023 * Verify that no high bits are set.
2025 if (cpusetsize > sizeof(cpuset_t)) {
2029 end = cp = (char *)&mask->__bits;
2031 cp += sizeof(cpuset_t);
2038 if (CPU_EMPTY(mask)) {
2043 case CPU_LEVEL_ROOT:
2044 case CPU_LEVEL_CPUSET:
2045 error = cpuset_which(which, id, &p, &ttd, &set);
2052 set = cpuset_ref(ttd->td_cpuset);
2056 case CPU_WHICH_CPUSET:
2057 case CPU_WHICH_JAIL:
2060 case CPU_WHICH_INTRHANDLER:
2061 case CPU_WHICH_ITHREAD:
2062 case CPU_WHICH_DOMAIN:
2066 if (level == CPU_LEVEL_ROOT)
2067 nset = cpuset_refroot(set);
2069 nset = cpuset_refbase(set);
2070 error = cpuset_modify(nset, mask);
2074 case CPU_LEVEL_WHICH:
2077 error = cpuset_setthread(id, mask);
2080 error = cpuset_setproc(id, NULL, mask, NULL, false);
2082 case CPU_WHICH_CPUSET:
2083 case CPU_WHICH_JAIL:
2084 error = cpuset_which(which, id, &p, &ttd, &set);
2086 error = cpuset_modify(set, mask);
2091 case CPU_WHICH_INTRHANDLER:
2092 case CPU_WHICH_ITHREAD:
2093 error = intr_setaffinity(id, which, mask);
2109 #ifndef _SYS_SYSPROTO_H_
2110 struct cpuset_getdomain_args {
2114 size_t domainsetsize;
2120 sys_cpuset_getdomain(struct thread *td, struct cpuset_getdomain_args *uap)
2123 return (kern_cpuset_getdomain(td, uap->level, uap->which,
2124 uap->id, uap->domainsetsize, uap->mask, uap->policy));
2128 kern_cpuset_getdomain(struct thread *td, cpulevel_t level, cpuwhich_t which,
2129 id_t id, size_t domainsetsize, domainset_t *maskp, int *policyp)
2131 struct domainset outset;
2133 struct cpuset *nset;
2135 struct domainset *dset;
2140 if (domainsetsize < sizeof(domainset_t) ||
2141 domainsetsize > DOMAINSET_MAXSIZE / NBBY)
2143 error = cpuset_check_capabilities(td, level, which, id);
2146 mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO);
2147 bzero(&outset, sizeof(outset));
2148 error = cpuset_which(which, id, &p, &ttd, &set);
2152 case CPU_LEVEL_ROOT:
2153 case CPU_LEVEL_CPUSET:
2158 set = cpuset_ref(ttd->td_cpuset);
2161 case CPU_WHICH_CPUSET:
2162 case CPU_WHICH_JAIL:
2165 case CPU_WHICH_INTRHANDLER:
2166 case CPU_WHICH_ITHREAD:
2167 case CPU_WHICH_DOMAIN:
2171 if (level == CPU_LEVEL_ROOT)
2172 nset = cpuset_refroot(set);
2174 nset = cpuset_refbase(set);
2175 domainset_copy(nset->cs_domain, &outset);
2178 case CPU_LEVEL_WHICH:
2182 domainset_copy(ttd->td_cpuset->cs_domain, &outset);
2186 FOREACH_THREAD_IN_PROC(p, ttd) {
2188 dset = ttd->td_cpuset->cs_domain;
2189 /* Show all domains in the proc. */
2190 DOMAINSET_OR(&outset.ds_mask, &dset->ds_mask);
2191 /* Last policy wins. */
2192 outset.ds_policy = dset->ds_policy;
2193 outset.ds_prefer = dset->ds_prefer;
2197 case CPU_WHICH_CPUSET:
2198 case CPU_WHICH_JAIL:
2199 domainset_copy(set->cs_domain, &outset);
2202 case CPU_WHICH_INTRHANDLER:
2203 case CPU_WHICH_ITHREAD:
2204 case CPU_WHICH_DOMAIN:
2218 * Translate prefer into a set containing only the preferred domain,
2219 * not the entire fallback set.
2221 if (outset.ds_policy == DOMAINSET_POLICY_PREFER) {
2222 DOMAINSET_ZERO(&outset.ds_mask);
2223 DOMAINSET_SET(outset.ds_prefer, &outset.ds_mask);
2225 DOMAINSET_COPY(&outset.ds_mask, mask);
2227 error = copyout(mask, maskp, domainsetsize);
2229 if (suword32(policyp, outset.ds_policy) != 0)
2236 #ifndef _SYS_SYSPROTO_H_
2237 struct cpuset_setdomain_args {
2241 size_t domainsetsize;
2247 sys_cpuset_setdomain(struct thread *td, struct cpuset_setdomain_args *uap)
2250 return (kern_cpuset_setdomain(td, uap->level, uap->which,
2251 uap->id, uap->domainsetsize, uap->mask, uap->policy));
2255 kern_cpuset_setdomain(struct thread *td, cpulevel_t level, cpuwhich_t which,
2256 id_t id, size_t domainsetsize, const domainset_t *maskp, int policy)
2258 struct cpuset *nset;
2262 struct domainset domain;
2266 if (domainsetsize < sizeof(domainset_t) ||
2267 domainsetsize > DOMAINSET_MAXSIZE / NBBY)
2269 if (policy <= DOMAINSET_POLICY_INVALID ||
2270 policy > DOMAINSET_POLICY_MAX)
2272 error = cpuset_check_capabilities(td, level, which, id);
2275 memset(&domain, 0, sizeof(domain));
2276 mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO);
2277 error = copyin(maskp, mask, domainsetsize);
2281 * Verify that no high bits are set.
2283 if (domainsetsize > sizeof(domainset_t)) {
2287 end = cp = (char *)&mask->__bits;
2288 end += domainsetsize;
2289 cp += sizeof(domainset_t);
2296 if (DOMAINSET_EMPTY(mask)) {
2300 DOMAINSET_COPY(mask, &domain.ds_mask);
2301 domain.ds_policy = policy;
2304 * Sanitize the provided mask.
2306 if (!DOMAINSET_SUBSET(&all_domains, &domain.ds_mask)) {
2311 /* Translate preferred policy into a mask and fallback. */
2312 if (policy == DOMAINSET_POLICY_PREFER) {
2313 /* Only support a single preferred domain. */
2314 if (DOMAINSET_COUNT(&domain.ds_mask) != 1) {
2318 domain.ds_prefer = DOMAINSET_FFS(&domain.ds_mask) - 1;
2319 /* This will be constrained by domainset_shadow(). */
2320 DOMAINSET_COPY(&all_domains, &domain.ds_mask);
2324 * When given an impossible policy, fall back to interleaving
2325 * across all domains.
2327 if (domainset_empty_vm(&domain))
2328 domainset_copy(domainset2, &domain);
2331 case CPU_LEVEL_ROOT:
2332 case CPU_LEVEL_CPUSET:
2333 error = cpuset_which(which, id, &p, &ttd, &set);
2340 set = cpuset_ref(ttd->td_cpuset);
2344 case CPU_WHICH_CPUSET:
2345 case CPU_WHICH_JAIL:
2348 case CPU_WHICH_INTRHANDLER:
2349 case CPU_WHICH_ITHREAD:
2350 case CPU_WHICH_DOMAIN:
2354 if (level == CPU_LEVEL_ROOT)
2355 nset = cpuset_refroot(set);
2357 nset = cpuset_refbase(set);
2358 error = cpuset_modify_domain(nset, &domain);
2362 case CPU_LEVEL_WHICH:
2365 error = _cpuset_setthread(id, NULL, &domain);
2368 error = cpuset_setproc(id, NULL, NULL, &domain, false);
2370 case CPU_WHICH_CPUSET:
2371 case CPU_WHICH_JAIL:
2372 error = cpuset_which(which, id, &p, &ttd, &set);
2374 error = cpuset_modify_domain(set, &domain);
2379 case CPU_WHICH_INTRHANDLER:
2380 case CPU_WHICH_ITHREAD:
2398 ddb_display_bitset(const struct bitset *set, int size)
2402 for (once = 0, bit = 0; bit < size; bit++) {
2403 if (CPU_ISSET(bit, set)) {
2405 db_printf("%d", bit);
2408 db_printf(",%d", bit);
2412 db_printf("<none>");
2416 ddb_display_cpuset(const cpuset_t *set)
2418 ddb_display_bitset((const struct bitset *)set, CPU_SETSIZE);
2422 ddb_display_domainset(const domainset_t *set)
2424 ddb_display_bitset((const struct bitset *)set, DOMAINSET_SETSIZE);
2427 DB_SHOW_COMMAND(cpusets, db_show_cpusets)
2431 LIST_FOREACH(set, &cpuset_ids, cs_link) {
2432 db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n",
2433 set, set->cs_id, refcount_load(&set->cs_ref), set->cs_flags,
2434 (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0);
2435 db_printf(" cpu mask=");
2436 ddb_display_cpuset(&set->cs_mask);
2438 db_printf(" domain policy %d prefer %d mask=",
2439 set->cs_domain->ds_policy, set->cs_domain->ds_prefer);
2440 ddb_display_domainset(&set->cs_domain->ds_mask);
2447 DB_SHOW_COMMAND(domainsets, db_show_domainsets)
2449 struct domainset *set;
2451 LIST_FOREACH(set, &cpuset_domains, ds_link) {
2452 db_printf("set=%p policy %d prefer %d cnt %d\n",
2453 set, set->ds_policy, set->ds_prefer, set->ds_cnt);
2454 db_printf(" mask =");
2455 ddb_display_domainset(&set->ds_mask);