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("$FreeBSD$");
37 #include "opt_ktrace.h"
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/sysctl.h>
42 #include <sys/ctype.h>
43 #include <sys/sysproto.h>
45 #include <sys/kernel.h>
47 #include <sys/malloc.h>
48 #include <sys/mutex.h>
51 #include <sys/refcount.h>
52 #include <sys/sched.h>
54 #include <sys/syscallsubr.h>
55 #include <sys/sysent.h>
56 #include <sys/capsicum.h>
57 #include <sys/cpuset.h>
58 #include <sys/domainset.h>
60 #include <sys/queue.h>
61 #include <sys/libkern.h>
62 #include <sys/limits.h>
64 #include <sys/interrupt.h>
65 #include <sys/vmmeter.h>
66 #include <sys/ktrace.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_page.h>
72 #include <vm/vm_pageout.h>
73 #include <vm/vm_extern.h>
74 #include <vm/vm_param.h>
75 #include <vm/vm_phys.h>
76 #include <vm/vm_pagequeue.h>
83 * cpusets provide a mechanism for creating and manipulating sets of
84 * processors for the purpose of constraining the scheduling of threads to
85 * specific processors.
87 * Each process belongs to an identified set, by default this is set 1. Each
88 * thread may further restrict the cpus it may run on to a subset of this
89 * named set. This creates an anonymous set which other threads and processes
90 * may not join by number.
92 * The named set is referred to herein as the 'base' set to avoid ambiguity.
93 * This set is usually a child of a 'root' set while the anonymous set may
94 * simply be referred to as a mask. In the syscall api these are referred to
95 * as the ROOT, CPUSET, and MASK levels where CPUSET is called 'base' here.
97 * Threads inherit their set from their creator whether it be anonymous or
98 * not. This means that anonymous sets are immutable because they may be
99 * shared. To modify an anonymous set a new set is created with the desired
100 * mask and the same parent as the existing anonymous set. This gives the
101 * illusion of each thread having a private mask.
103 * Via the syscall apis a user may ask to retrieve or modify the root, base,
104 * or mask that is discovered via a pid, tid, or setid. Modifying a set
105 * modifies all numbered and anonymous child sets to comply with the new mask.
106 * Modifying a pid or tid's mask applies only to that tid but must still
107 * exist within the assigned parent set.
109 * A thread may not be assigned to a group separate from other threads in
110 * the process. This is to remove ambiguity when the setid is queried with
111 * a pid argument. There is no other technical limitation.
113 * This somewhat complex arrangement is intended to make it easy for
114 * applications to query available processors and bind their threads to
115 * specific processors while also allowing administrators to dynamically
116 * reprovision by changing sets which apply to groups of processes.
118 * A simple application should not concern itself with sets at all and
119 * rather apply masks to its own threads via CPU_WHICH_TID and a -1 id
120 * meaning 'curthread'. It may query available cpus for that tid with a
121 * getaffinity call using (CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, ...).
124 LIST_HEAD(domainlist, domainset);
125 struct domainset __read_mostly domainset_firsttouch;
126 struct domainset __read_mostly domainset_fixed[MAXMEMDOM];
127 struct domainset __read_mostly domainset_interleave;
128 struct domainset __read_mostly domainset_prefer[MAXMEMDOM];
129 struct domainset __read_mostly domainset_roundrobin;
131 static uma_zone_t cpuset_zone;
132 static uma_zone_t domainset_zone;
133 static struct mtx cpuset_lock;
134 static struct setlist cpuset_ids;
135 static struct domainlist cpuset_domains;
136 static struct unrhdr *cpuset_unr;
137 static struct cpuset *cpuset_zero, *cpuset_default, *cpuset_kernel;
138 static struct domainset *domainset0, *domainset2;
140 /* Return the size of cpuset_t at the kernel level */
141 SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD | CTLFLAG_CAPRD,
142 SYSCTL_NULL_INT_PTR, sizeof(cpuset_t), "sizeof(cpuset_t)");
144 cpuset_t *cpuset_root;
145 cpuset_t cpuset_domain[MAXMEMDOM];
147 static int domainset_valid(const struct domainset *, const struct domainset *);
150 * Find the first non-anonymous set starting from 'set'.
152 static struct cpuset *
153 cpuset_getbase(struct cpuset *set)
156 if (set->cs_id == CPUSET_INVALID)
157 set = set->cs_parent;
162 * Walks up the tree from 'set' to find the root.
164 static struct cpuset *
165 cpuset_getroot(struct cpuset *set)
168 while ((set->cs_flags & CPU_SET_ROOT) == 0 && set->cs_parent != NULL)
169 set = set->cs_parent;
174 * Acquire a reference to a cpuset, all pointers must be tracked with refs.
177 cpuset_ref(struct cpuset *set)
180 refcount_acquire(&set->cs_ref);
185 * Walks up the tree from 'set' to find the root. Returns the root
188 static struct cpuset *
189 cpuset_refroot(struct cpuset *set)
192 return (cpuset_ref(cpuset_getroot(set)));
196 * Find the first non-anonymous set starting from 'set'. Returns this set
197 * referenced. May return the passed in set with an extra ref if it is
200 static struct cpuset *
201 cpuset_refbase(struct cpuset *set)
204 return (cpuset_ref(cpuset_getbase(set)));
208 * Release a reference in a context where it is safe to allocate.
211 cpuset_rel(struct cpuset *set)
215 if (refcount_release_if_not_last(&set->cs_ref))
217 mtx_lock_spin(&cpuset_lock);
218 if (!refcount_release(&set->cs_ref)) {
219 mtx_unlock_spin(&cpuset_lock);
222 LIST_REMOVE(set, cs_siblings);
224 if (id != CPUSET_INVALID)
225 LIST_REMOVE(set, cs_link);
226 mtx_unlock_spin(&cpuset_lock);
227 cpuset_rel(set->cs_parent);
228 uma_zfree(cpuset_zone, set);
229 if (id != CPUSET_INVALID)
230 free_unr(cpuset_unr, id);
234 * Deferred release must be used when in a context that is not safe to
235 * allocate/free. This places any unreferenced sets on the list 'head'.
238 cpuset_rel_defer(struct setlist *head, struct cpuset *set)
241 if (refcount_release_if_not_last(&set->cs_ref))
243 mtx_lock_spin(&cpuset_lock);
244 if (!refcount_release(&set->cs_ref)) {
245 mtx_unlock_spin(&cpuset_lock);
248 LIST_REMOVE(set, cs_siblings);
249 if (set->cs_id != CPUSET_INVALID)
250 LIST_REMOVE(set, cs_link);
251 LIST_INSERT_HEAD(head, set, cs_link);
252 mtx_unlock_spin(&cpuset_lock);
256 * Complete a deferred release. Removes the set from the list provided to
260 cpuset_rel_complete(struct cpuset *set)
265 LIST_REMOVE(set, cs_link);
266 cpuset_rel(set->cs_parent);
267 uma_zfree(cpuset_zone, set);
268 if (id != CPUSET_INVALID)
269 free_unr(cpuset_unr, id);
273 * Find a set based on an id. Returns it with a ref.
275 static struct cpuset *
276 cpuset_lookup(cpusetid_t setid, struct thread *td)
280 if (setid == CPUSET_INVALID)
282 mtx_lock_spin(&cpuset_lock);
283 LIST_FOREACH(set, &cpuset_ids, cs_link)
284 if (set->cs_id == setid)
288 mtx_unlock_spin(&cpuset_lock);
290 KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__));
291 if (set != NULL && jailed(td->td_ucred)) {
292 struct cpuset *jset, *tset;
294 jset = td->td_ucred->cr_prison->pr_cpuset;
295 for (tset = set; tset != NULL; tset = tset->cs_parent)
308 * Initialize a set in the space provided in 'set' with the provided parameters.
309 * The set is returned with a single ref. May return EDEADLK if the set
310 * will have no valid cpu based on restrictions from the parent.
313 cpuset_init(struct cpuset *set, struct cpuset *parent,
314 const cpuset_t *mask, struct domainset *domain, cpusetid_t id)
318 domain = parent->cs_domain;
320 mask = &parent->cs_mask;
321 if (!CPU_OVERLAP(&parent->cs_mask, mask))
323 /* The domain must be prepared ahead of time. */
324 if (!domainset_valid(parent->cs_domain, domain))
326 CPU_COPY(mask, &set->cs_mask);
327 LIST_INIT(&set->cs_children);
328 refcount_init(&set->cs_ref, 1);
330 mtx_lock_spin(&cpuset_lock);
331 set->cs_domain = domain;
332 CPU_AND(&set->cs_mask, &set->cs_mask, &parent->cs_mask);
334 set->cs_parent = cpuset_ref(parent);
335 LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings);
336 if (set->cs_id != CPUSET_INVALID)
337 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
338 mtx_unlock_spin(&cpuset_lock);
344 * Create a new non-anonymous set with the requested parent and mask. May
345 * return failures if the mask is invalid or a new number can not be
348 * If *setp is not NULL, then it will be used as-is. The caller must take
349 * into account that *setp will be inserted at the head of cpuset_ids and
350 * plan any potentially conflicting cs_link usage accordingly.
353 cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask)
360 id = alloc_unr(cpuset_unr);
363 dofree = (*setp == NULL);
367 *setp = set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
368 error = cpuset_init(set, parent, mask, NULL, id);
371 free_unr(cpuset_unr, id);
373 uma_zfree(cpuset_zone, set);
379 cpuset_freelist_add(struct setlist *list, int count)
384 for (i = 0; i < count; i++) {
385 set = uma_zalloc(cpuset_zone, M_ZERO | M_WAITOK);
386 LIST_INSERT_HEAD(list, set, cs_link);
391 cpuset_freelist_init(struct setlist *list, int count)
395 cpuset_freelist_add(list, count);
399 cpuset_freelist_free(struct setlist *list)
403 while ((set = LIST_FIRST(list)) != NULL) {
404 LIST_REMOVE(set, cs_link);
405 uma_zfree(cpuset_zone, set);
410 domainset_freelist_add(struct domainlist *list, int count)
412 struct domainset *set;
415 for (i = 0; i < count; i++) {
416 set = uma_zalloc(domainset_zone, M_ZERO | M_WAITOK);
417 LIST_INSERT_HEAD(list, set, ds_link);
422 domainset_freelist_init(struct domainlist *list, int count)
426 domainset_freelist_add(list, count);
430 domainset_freelist_free(struct domainlist *list)
432 struct domainset *set;
434 while ((set = LIST_FIRST(list)) != NULL) {
435 LIST_REMOVE(set, ds_link);
436 uma_zfree(domainset_zone, set);
440 /* Copy a domainset preserving mask and policy. */
442 domainset_copy(const struct domainset *from, struct domainset *to)
445 DOMAINSET_COPY(&from->ds_mask, &to->ds_mask);
446 to->ds_policy = from->ds_policy;
447 to->ds_prefer = from->ds_prefer;
450 /* Return 1 if mask and policy are equal, otherwise 0. */
452 domainset_equal(const struct domainset *one, const struct domainset *two)
455 return (DOMAINSET_CMP(&one->ds_mask, &two->ds_mask) == 0 &&
456 one->ds_policy == two->ds_policy &&
457 one->ds_prefer == two->ds_prefer);
460 /* Return 1 if child is a valid subset of parent. */
462 domainset_valid(const struct domainset *parent, const struct domainset *child)
464 if (child->ds_policy != DOMAINSET_POLICY_PREFER)
465 return (DOMAINSET_SUBSET(&parent->ds_mask, &child->ds_mask));
466 return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask));
470 domainset_restrict(const struct domainset *parent,
471 const struct domainset *child)
473 if (child->ds_policy != DOMAINSET_POLICY_PREFER)
474 return (DOMAINSET_OVERLAP(&parent->ds_mask, &child->ds_mask));
475 return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask));
479 * Lookup or create a domainset. The key is provided in ds_mask and
480 * ds_policy. If the domainset does not yet exist the storage in
481 * 'domain' is used to insert. Otherwise this storage is freed to the
482 * domainset_zone and the existing domainset is returned.
484 static struct domainset *
485 _domainset_create(struct domainset *domain, struct domainlist *freelist)
487 struct domainset *ndomain;
490 KASSERT(domain->ds_cnt <= vm_ndomains,
491 ("invalid domain count in domainset %p", domain));
492 KASSERT(domain->ds_policy != DOMAINSET_POLICY_PREFER ||
493 domain->ds_prefer < vm_ndomains,
494 ("invalid preferred domain in domains %p", domain));
496 mtx_lock_spin(&cpuset_lock);
497 LIST_FOREACH(ndomain, &cpuset_domains, ds_link)
498 if (domainset_equal(ndomain, domain))
501 * If the domain does not yet exist we insert it and initialize
502 * various iteration helpers which are not part of the key.
504 if (ndomain == NULL) {
505 LIST_INSERT_HEAD(&cpuset_domains, domain, ds_link);
506 domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask);
507 for (i = 0, j = 0; i < DOMAINSET_FLS(&domain->ds_mask); i++)
508 if (DOMAINSET_ISSET(i, &domain->ds_mask))
509 domain->ds_order[j++] = i;
511 mtx_unlock_spin(&cpuset_lock);
514 if (freelist != NULL)
515 LIST_INSERT_HEAD(freelist, domain, ds_link);
517 uma_zfree(domainset_zone, domain);
523 * Are any of the domains in the mask empty? If so, silently
524 * remove them and update the domainset accordingly. If only empty
525 * domains are present, we must return failure.
528 domainset_empty_vm(struct domainset *domain)
533 DOMAINSET_ZERO(&empty);
534 for (i = 0; i < vm_ndomains; i++)
535 if (VM_DOMAIN_EMPTY(i))
536 DOMAINSET_SET(i, &empty);
537 if (DOMAINSET_SUBSET(&empty, &domain->ds_mask))
540 /* Remove empty domains from the set and recompute. */
541 DOMAINSET_ANDNOT(&domain->ds_mask, &empty);
542 domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask);
543 for (i = j = 0; i < DOMAINSET_FLS(&domain->ds_mask); i++)
544 if (DOMAINSET_ISSET(i, &domain->ds_mask))
545 domain->ds_order[j++] = i;
547 /* Convert a PREFER policy referencing an empty domain to RR. */
548 if (domain->ds_policy == DOMAINSET_POLICY_PREFER &&
549 DOMAINSET_ISSET(domain->ds_prefer, &empty)) {
550 domain->ds_policy = DOMAINSET_POLICY_ROUNDROBIN;
551 domain->ds_prefer = -1;
558 * Create or lookup a domainset based on the key held in 'domain'.
561 domainset_create(const struct domainset *domain)
563 struct domainset *ndomain;
566 * Validate the policy. It must specify a useable policy number with
567 * only valid domains. Preferred must include the preferred domain
570 if (domain->ds_policy <= DOMAINSET_POLICY_INVALID ||
571 domain->ds_policy > DOMAINSET_POLICY_MAX)
573 if (domain->ds_policy == DOMAINSET_POLICY_PREFER &&
574 !DOMAINSET_ISSET(domain->ds_prefer, &domain->ds_mask))
576 if (!DOMAINSET_SUBSET(&domainset0->ds_mask, &domain->ds_mask))
578 ndomain = uma_zalloc(domainset_zone, M_WAITOK | M_ZERO);
579 domainset_copy(domain, ndomain);
580 return _domainset_create(ndomain, NULL);
584 * Update thread domainset pointers.
587 domainset_notify(void)
592 sx_slock(&allproc_lock);
593 FOREACH_PROC_IN_SYSTEM(p) {
595 if (p->p_state == PRS_NEW) {
599 FOREACH_THREAD_IN_PROC(p, td) {
601 td->td_domain.dr_policy = td->td_cpuset->cs_domain;
606 sx_sunlock(&allproc_lock);
607 kernel_object->domain.dr_policy = cpuset_kernel->cs_domain;
611 * Create a new set that is a subset of a parent.
613 static struct domainset *
614 domainset_shadow(const struct domainset *pdomain,
615 const struct domainset *domain, struct domainlist *freelist)
617 struct domainset *ndomain;
619 ndomain = LIST_FIRST(freelist);
620 LIST_REMOVE(ndomain, ds_link);
623 * Initialize the key from the request.
625 domainset_copy(domain, ndomain);
628 * Restrict the key by the parent.
630 DOMAINSET_AND(&ndomain->ds_mask, &pdomain->ds_mask);
632 return _domainset_create(ndomain, freelist);
636 * Recursively check for errors that would occur from applying mask to
637 * the tree of sets starting at 'set'. Checks for sets that would become
638 * empty as well as RDONLY flags.
641 cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int augment_mask)
647 mtx_assert(&cpuset_lock, MA_OWNED);
648 if (set->cs_flags & CPU_SET_RDONLY)
651 CPU_AND(&newmask, &set->cs_mask, mask);
653 CPU_COPY(mask, &newmask);
655 if (CPU_EMPTY(&newmask))
658 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
659 if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0)
665 * Applies the mask 'mask' without checking for empty sets or permissions.
668 cpuset_update(struct cpuset *set, cpuset_t *mask)
672 mtx_assert(&cpuset_lock, MA_OWNED);
673 CPU_AND(&set->cs_mask, &set->cs_mask, mask);
674 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
675 cpuset_update(nset, &set->cs_mask);
681 * Modify the set 'set' to use a copy of the mask provided. Apply this new
682 * mask to restrict all children in the tree. Checks for validity before
683 * applying the changes.
686 cpuset_modify(struct cpuset *set, cpuset_t *mask)
691 error = priv_check(curthread, PRIV_SCHED_CPUSET);
695 * In case we are called from within the jail,
696 * we do not allow modifying the dedicated root
697 * cpuset of the jail but may still allow to
698 * change child sets, including subordinate jails'
701 if ((set->cs_flags & CPU_SET_ROOT) != 0 &&
702 jailed(curthread->td_ucred) &&
703 set == curthread->td_ucred->cr_prison->pr_cpuset)
706 * Verify that we have access to this set of
709 if ((set->cs_flags & (CPU_SET_ROOT | CPU_SET_RDONLY)) == CPU_SET_ROOT) {
710 KASSERT(set->cs_parent != NULL,
711 ("jail.cpuset=%d is not a proper child of parent jail's root.",
715 * cpuset_getroot() cannot work here due to how top-level jail
716 * roots are constructed. Top-level jails are parented to
717 * thread0's cpuset (i.e. cpuset 1) rather than the system root.
719 root = set->cs_parent;
721 root = cpuset_getroot(set);
723 mtx_lock_spin(&cpuset_lock);
724 if (root && !CPU_SUBSET(&root->cs_mask, mask)) {
728 error = cpuset_testupdate(set, mask, 0);
731 CPU_COPY(mask, &set->cs_mask);
732 cpuset_update(set, mask);
734 mtx_unlock_spin(&cpuset_lock);
740 * Recursively check for errors that would occur from applying mask to
741 * the tree of sets starting at 'set'. Checks for sets that would become
742 * empty as well as RDONLY flags.
745 cpuset_testupdate_domain(struct cpuset *set, struct domainset *dset,
746 struct domainset *orig, int *count, int augment_mask __unused)
749 struct domainset *domain;
750 struct domainset newset;
753 mtx_assert(&cpuset_lock, MA_OWNED);
754 if (set->cs_flags & CPU_SET_RDONLY)
756 domain = set->cs_domain;
757 domainset_copy(domain, &newset);
758 if (!domainset_equal(domain, orig)) {
759 if (!domainset_restrict(domain, dset))
761 DOMAINSET_AND(&newset.ds_mask, &dset->ds_mask);
762 /* Count the number of domains that are changing. */
766 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
767 if ((error = cpuset_testupdate_domain(nset, &newset, domain,
774 * Applies the mask 'mask' without checking for empty sets or permissions.
777 cpuset_update_domain(struct cpuset *set, struct domainset *domain,
778 struct domainset *orig, struct domainlist *domains)
782 mtx_assert(&cpuset_lock, MA_OWNED);
784 * If this domainset has changed from the parent we must calculate
785 * a new set. Otherwise it simply inherits from the parent. When
786 * we inherit from the parent we get a new mask and policy. If the
787 * set is modified from the parent we keep the policy and only
790 if (set->cs_domain != orig) {
791 orig = set->cs_domain;
792 set->cs_domain = domainset_shadow(domain, orig, domains);
794 set->cs_domain = domain;
795 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
796 cpuset_update_domain(nset, set->cs_domain, orig, domains);
802 * Modify the set 'set' to use a copy the domainset provided. Apply this new
803 * mask to restrict all children in the tree. Checks for validity before
804 * applying the changes.
807 cpuset_modify_domain(struct cpuset *set, struct domainset *domain)
809 struct domainlist domains;
810 struct domainset temp;
811 struct domainset *dset;
813 int ndomains, needed;
816 error = priv_check(curthread, PRIV_SCHED_CPUSET);
820 * In case we are called from within the jail
821 * we do not allow modifying the dedicated root
822 * cpuset of the jail but may still allow to
825 if (jailed(curthread->td_ucred) &&
826 set->cs_flags & CPU_SET_ROOT)
828 domainset_freelist_init(&domains, 0);
829 domain = domainset_create(domain);
832 mtx_lock_spin(&cpuset_lock);
834 root = cpuset_getroot(set);
835 dset = root->cs_domain;
837 * Verify that we have access to this set of domains.
839 if (!domainset_valid(dset, domain)) {
844 * If applying prefer we keep the current set as the fallback.
846 if (domain->ds_policy == DOMAINSET_POLICY_PREFER)
847 DOMAINSET_COPY(&set->cs_domain->ds_mask,
850 * Determine whether we can apply this set of domains and
851 * how many new domain structures it will require.
853 domainset_copy(domain, &temp);
855 error = cpuset_testupdate_domain(set, &temp, set->cs_domain,
859 if (ndomains >= needed)
862 /* Dropping the lock; we'll need to re-evaluate again. */
863 mtx_unlock_spin(&cpuset_lock);
864 domainset_freelist_add(&domains, needed - ndomains);
866 mtx_lock_spin(&cpuset_lock);
868 dset = set->cs_domain;
869 cpuset_update_domain(set, domain, dset, &domains);
871 mtx_unlock_spin(&cpuset_lock);
872 domainset_freelist_free(&domains);
880 * Resolve the 'which' parameter of several cpuset apis.
882 * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid. Also
883 * checks for permission via p_cansched().
885 * For WHICH_SET returns a valid set with a new reference.
887 * -1 may be supplied for any argument to mean the current proc/thread or
888 * the base set of the current thread. May fail with ESRCH/EPERM.
891 cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp,
892 struct cpuset **setp)
909 if ((p = pfind(id)) == NULL)
924 case CPU_WHICH_CPUSET:
926 thread_lock(curthread);
927 set = cpuset_refbase(curthread->td_cpuset);
928 thread_unlock(curthread);
930 set = cpuset_lookup(id, curthread);
938 /* Find `set' for prison with given id. */
941 sx_slock(&allprison_lock);
942 pr = prison_find_child(curthread->td_ucred->cr_prison, id);
943 sx_sunlock(&allprison_lock);
946 cpuset_ref(pr->pr_cpuset);
947 *setp = pr->pr_cpuset;
948 mtx_unlock(&pr->pr_mtx);
952 case CPU_WHICH_DOMAIN:
957 error = p_cansched(curthread, p);
963 td = FIRST_THREAD_IN_PROC(p);
970 cpuset_testshadow(struct cpuset *set, const cpuset_t *mask,
971 const struct domainset *domain)
973 struct cpuset *parent;
974 struct domainset *dset;
976 parent = cpuset_getbase(set);
978 * If we are restricting a cpu mask it must be a subset of the
979 * parent or invalid CPUs have been specified.
981 if (mask != NULL && !CPU_SUBSET(&parent->cs_mask, mask))
985 * If we are restricting a domain mask it must be a subset of the
986 * parent or invalid domains have been specified.
988 dset = parent->cs_domain;
989 if (domain != NULL && !domainset_valid(dset, domain))
996 * Create an anonymous set with the provided mask in the space provided by
997 * 'nset'. If the passed in set is anonymous we use its parent otherwise
998 * the new set is a child of 'set'.
1001 cpuset_shadow(struct cpuset *set, struct cpuset **nsetp,
1002 const cpuset_t *mask, const struct domainset *domain,
1003 struct setlist *cpusets, struct domainlist *domains)
1005 struct cpuset *parent;
1006 struct cpuset *nset;
1007 struct domainset *dset;
1008 struct domainset *d;
1011 error = cpuset_testshadow(set, mask, domain);
1015 parent = cpuset_getbase(set);
1016 dset = parent->cs_domain;
1018 mask = &set->cs_mask;
1020 d = domainset_shadow(dset, domain, domains);
1023 nset = LIST_FIRST(cpusets);
1024 error = cpuset_init(nset, parent, mask, d, CPUSET_INVALID);
1026 LIST_REMOVE(nset, cs_link);
1032 static struct cpuset *
1033 cpuset_update_thread(struct thread *td, struct cpuset *nset)
1035 struct cpuset *tdset;
1037 tdset = td->td_cpuset;
1038 td->td_cpuset = nset;
1039 td->td_domain.dr_policy = nset->cs_domain;
1046 cpuset_setproc_test_maskthread(struct cpuset *tdset, cpuset_t *mask,
1047 struct domainset *domain)
1049 struct cpuset *parent;
1051 parent = cpuset_getbase(tdset);
1053 mask = &tdset->cs_mask;
1055 domain = tdset->cs_domain;
1056 return cpuset_testshadow(parent, mask, domain);
1060 cpuset_setproc_maskthread(struct cpuset *tdset, cpuset_t *mask,
1061 struct domainset *domain, struct cpuset **nsetp,
1062 struct setlist *freelist, struct domainlist *domainlist)
1064 struct cpuset *parent;
1066 parent = cpuset_getbase(tdset);
1068 mask = &tdset->cs_mask;
1070 domain = tdset->cs_domain;
1071 return cpuset_shadow(parent, nsetp, mask, domain, freelist,
1076 cpuset_setproc_setthread_mask(struct cpuset *tdset, struct cpuset *set,
1077 cpuset_t *mask, struct domainset *domain)
1079 struct cpuset *parent;
1081 parent = cpuset_getbase(tdset);
1084 * If the thread restricted its mask then apply that same
1085 * restriction to the new set, otherwise take it wholesale.
1087 if (CPU_CMP(&tdset->cs_mask, &parent->cs_mask) != 0) {
1088 CPU_AND(mask, &tdset->cs_mask, &set->cs_mask);
1090 CPU_COPY(&set->cs_mask, mask);
1093 * If the thread restricted the domain then we apply the
1094 * restriction to the new set but retain the policy.
1096 if (tdset->cs_domain != parent->cs_domain) {
1097 domainset_copy(tdset->cs_domain, domain);
1098 DOMAINSET_AND(&domain->ds_mask, &set->cs_domain->ds_mask);
1100 domainset_copy(set->cs_domain, domain);
1102 if (CPU_EMPTY(mask) || DOMAINSET_EMPTY(&domain->ds_mask))
1109 cpuset_setproc_test_setthread(struct cpuset *tdset, struct cpuset *set)
1111 struct domainset domain;
1114 if (tdset->cs_id != CPUSET_INVALID)
1116 return cpuset_setproc_setthread_mask(tdset, set, &mask, &domain);
1120 cpuset_setproc_setthread(struct cpuset *tdset, struct cpuset *set,
1121 struct cpuset **nsetp, struct setlist *freelist,
1122 struct domainlist *domainlist)
1124 struct domainset domain;
1129 * If we're replacing on a thread that has not constrained the
1130 * original set we can simply accept the new set.
1132 if (tdset->cs_id != CPUSET_INVALID) {
1133 *nsetp = cpuset_ref(set);
1136 error = cpuset_setproc_setthread_mask(tdset, set, &mask, &domain);
1140 return cpuset_shadow(set, nsetp, &mask, &domain, freelist,
1145 cpuset_setproc_newbase(struct thread *td, struct cpuset *set,
1146 struct cpuset *nroot, struct cpuset **nsetp,
1147 struct setlist *cpusets, struct domainlist *domainlist)
1149 struct domainset ndomain;
1151 struct cpuset *pbase;
1154 pbase = cpuset_getbase(td->td_cpuset);
1156 /* Copy process mask, then further apply the new root mask. */
1157 CPU_AND(&nmask, &pbase->cs_mask, &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 #if defined(__powerpc__)
1748 * TODO: At least powerpc64 and powerpc64le kernels panic with
1749 * exception 0x480 (instruction segment exception) when copyin/copyout,
1750 * are set as a function pointer in cpuset_copy_cb struct and called by
1751 * an external module (like pfsync). Tip: copyin/copyout have an ifunc
1752 * resolver function.
1754 * Bisect of LLVM shows that the behavior changed on LLVM 10.0 with
1755 * https://reviews.llvm.org/rGdc06b0bc9ad055d06535462d91bfc2a744b2f589
1757 * This is a hack/workaround while problem is being discussed with LLVM
1761 cpuset_copyin(const void *uaddr, void *kaddr, size_t len)
1763 return(copyin(uaddr, kaddr, len));
1767 cpuset_copyout(const void *kaddr, void *uaddr, size_t len)
1769 return(copyout(kaddr, uaddr, len));
1772 static const struct cpuset_copy_cb copy_set = {
1773 .cpuset_copyin = cpuset_copyin,
1774 .cpuset_copyout = cpuset_copyout
1777 static const struct cpuset_copy_cb copy_set = {
1778 .cpuset_copyin = copyin,
1779 .cpuset_copyout = copyout
1783 #ifndef _SYS_SYSPROTO_H_
1784 struct cpuset_args {
1789 sys_cpuset(struct thread *td, struct cpuset_args *uap)
1791 struct cpuset *root;
1796 root = cpuset_refroot(td->td_cpuset);
1799 error = cpuset_create(&set, root, &root->cs_mask);
1803 error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id));
1805 error = cpuset_setproc(-1, set, NULL, NULL, false);
1810 #ifndef _SYS_SYSPROTO_H_
1811 struct cpuset_setid_args {
1818 sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap)
1821 return (kern_cpuset_setid(td, uap->which, uap->id, uap->setid));
1825 kern_cpuset_setid(struct thread *td, cpuwhich_t which,
1826 id_t id, cpusetid_t setid)
1832 * Presently we only support per-process sets.
1834 if (which != CPU_WHICH_PID)
1836 set = cpuset_lookup(setid, td);
1839 error = cpuset_setproc(id, set, NULL, NULL, false);
1844 #ifndef _SYS_SYSPROTO_H_
1845 struct cpuset_getid_args {
1853 sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap)
1856 return (kern_cpuset_getid(td, uap->level, uap->which, uap->id,
1861 kern_cpuset_getid(struct thread *td, cpulevel_t level, cpuwhich_t which,
1862 id_t id, cpusetid_t *setid)
1864 struct cpuset *nset;
1871 if (level == CPU_LEVEL_WHICH && which != CPU_WHICH_CPUSET)
1873 error = cpuset_which(which, id, &p, &ttd, &set);
1880 set = cpuset_refbase(ttd->td_cpuset);
1884 case CPU_WHICH_CPUSET:
1885 case CPU_WHICH_JAIL:
1888 case CPU_WHICH_DOMAIN:
1892 case CPU_LEVEL_ROOT:
1893 nset = cpuset_refroot(set);
1897 case CPU_LEVEL_CPUSET:
1899 case CPU_LEVEL_WHICH:
1905 error = copyout(&tmpid, setid, sizeof(tmpid));
1910 #ifndef _SYS_SYSPROTO_H_
1911 struct cpuset_getaffinity_args {
1920 sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap)
1923 return (user_cpuset_getaffinity(td, uap->level, uap->which,
1924 uap->id, uap->cpusetsize, uap->mask, ©_set));
1928 kern_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1929 id_t id, size_t cpusetsize, cpuset_t *mask)
1932 struct cpuset *nset;
1937 error = cpuset_check_capabilities(td, level, which, id);
1940 error = cpuset_which(which, id, &p, &ttd, &set);
1944 case CPU_LEVEL_ROOT:
1945 case CPU_LEVEL_CPUSET:
1950 set = cpuset_ref(ttd->td_cpuset);
1953 case CPU_WHICH_CPUSET:
1954 case CPU_WHICH_JAIL:
1957 case CPU_WHICH_INTRHANDLER:
1958 case CPU_WHICH_ITHREAD:
1959 case CPU_WHICH_DOMAIN:
1962 if (level == CPU_LEVEL_ROOT)
1963 nset = cpuset_refroot(set);
1965 nset = cpuset_refbase(set);
1966 CPU_COPY(&nset->cs_mask, mask);
1969 case CPU_LEVEL_WHICH:
1973 CPU_COPY(&ttd->td_cpuset->cs_mask, mask);
1977 FOREACH_THREAD_IN_PROC(p, ttd) {
1979 CPU_OR(mask, mask, &ttd->td_cpuset->cs_mask);
1983 case CPU_WHICH_CPUSET:
1984 case CPU_WHICH_JAIL:
1985 CPU_COPY(&set->cs_mask, mask);
1988 case CPU_WHICH_INTRHANDLER:
1989 case CPU_WHICH_ITHREAD:
1990 error = intr_getaffinity(id, which, mask);
1992 case CPU_WHICH_DOMAIN:
1993 if (id < 0 || id >= MAXMEMDOM)
1996 CPU_COPY(&cpuset_domain[id], mask);
2009 if (cpusetsize < howmany(CPU_FLS(mask), NBBY))
2012 if (KTRPOINT(td, KTR_STRUCT))
2013 ktrcpuset(mask, cpusetsize);
2020 user_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
2021 id_t id, size_t cpusetsize, cpuset_t *maskp, const struct cpuset_copy_cb *cb)
2027 mask = malloc(sizeof(cpuset_t), M_TEMP, M_WAITOK | M_ZERO);
2028 size = min(cpusetsize, sizeof(cpuset_t));
2029 error = kern_cpuset_getaffinity(td, level, which, id, size, mask);
2031 error = cb->cpuset_copyout(mask, maskp, size);
2034 if (cpusetsize > size) {
2039 end = cp = (char *)&maskp->__bits;
2057 #ifndef _SYS_SYSPROTO_H_
2058 struct cpuset_setaffinity_args {
2063 const cpuset_t *mask;
2067 sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap)
2070 return (user_cpuset_setaffinity(td, uap->level, uap->which,
2071 uap->id, uap->cpusetsize, uap->mask, ©_set));
2075 kern_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
2076 id_t id, cpuset_t *mask)
2078 struct cpuset *nset;
2085 if (KTRPOINT(td, KTR_STRUCT))
2086 ktrcpuset(mask, sizeof(cpuset_t));
2088 error = cpuset_check_capabilities(td, level, which, id);
2091 if (CPU_EMPTY(mask))
2094 case CPU_LEVEL_ROOT:
2095 case CPU_LEVEL_CPUSET:
2096 error = cpuset_which(which, id, &p, &ttd, &set);
2103 set = cpuset_ref(ttd->td_cpuset);
2107 case CPU_WHICH_CPUSET:
2108 case CPU_WHICH_JAIL:
2111 case CPU_WHICH_INTRHANDLER:
2112 case CPU_WHICH_ITHREAD:
2113 case CPU_WHICH_DOMAIN:
2116 if (level == CPU_LEVEL_ROOT)
2117 nset = cpuset_refroot(set);
2119 nset = cpuset_refbase(set);
2120 error = cpuset_modify(nset, mask);
2124 case CPU_LEVEL_WHICH:
2127 error = cpuset_setthread(id, mask);
2130 error = cpuset_setproc(id, NULL, mask, NULL, false);
2132 case CPU_WHICH_CPUSET:
2133 case CPU_WHICH_JAIL:
2134 error = cpuset_which(which, id, &p, &ttd, &set);
2136 error = cpuset_modify(set, mask);
2141 case CPU_WHICH_INTRHANDLER:
2142 case CPU_WHICH_ITHREAD:
2143 error = intr_setaffinity(id, which, mask);
2158 user_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
2159 id_t id, size_t cpusetsize, const cpuset_t *maskp, const struct cpuset_copy_cb *cb)
2165 size = min(cpusetsize, sizeof(cpuset_t));
2166 mask = malloc(sizeof(cpuset_t), M_TEMP, M_WAITOK | M_ZERO);
2167 error = cb->cpuset_copyin(maskp, mask, size);
2171 * Verify that no high bits are set.
2173 if (cpusetsize > sizeof(cpuset_t)) {
2174 const char *end, *cp;
2176 end = cp = (const char *)&maskp->__bits;
2178 cp += sizeof(cpuset_t);
2193 error = kern_cpuset_setaffinity(td, level, which, id, mask);
2200 #ifndef _SYS_SYSPROTO_H_
2201 struct cpuset_getdomain_args {
2205 size_t domainsetsize;
2211 sys_cpuset_getdomain(struct thread *td, struct cpuset_getdomain_args *uap)
2214 return (kern_cpuset_getdomain(td, uap->level, uap->which,
2215 uap->id, uap->domainsetsize, uap->mask, uap->policy, ©_set));
2219 kern_cpuset_getdomain(struct thread *td, cpulevel_t level, cpuwhich_t which,
2220 id_t id, size_t domainsetsize, domainset_t *maskp, int *policyp,
2221 const struct cpuset_copy_cb *cb)
2223 struct domainset outset;
2225 struct cpuset *nset;
2227 struct domainset *dset;
2232 if (domainsetsize < sizeof(domainset_t) ||
2233 domainsetsize > DOMAINSET_MAXSIZE / NBBY)
2235 error = cpuset_check_capabilities(td, level, which, id);
2238 mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO);
2239 bzero(&outset, sizeof(outset));
2240 error = cpuset_which(which, id, &p, &ttd, &set);
2244 case CPU_LEVEL_ROOT:
2245 case CPU_LEVEL_CPUSET:
2250 set = cpuset_ref(ttd->td_cpuset);
2253 case CPU_WHICH_CPUSET:
2254 case CPU_WHICH_JAIL:
2257 case CPU_WHICH_INTRHANDLER:
2258 case CPU_WHICH_ITHREAD:
2259 case CPU_WHICH_DOMAIN:
2263 if (level == CPU_LEVEL_ROOT)
2264 nset = cpuset_refroot(set);
2266 nset = cpuset_refbase(set);
2267 domainset_copy(nset->cs_domain, &outset);
2270 case CPU_LEVEL_WHICH:
2274 domainset_copy(ttd->td_cpuset->cs_domain, &outset);
2278 FOREACH_THREAD_IN_PROC(p, ttd) {
2280 dset = ttd->td_cpuset->cs_domain;
2281 /* Show all domains in the proc. */
2282 DOMAINSET_OR(&outset.ds_mask, &dset->ds_mask);
2283 /* Last policy wins. */
2284 outset.ds_policy = dset->ds_policy;
2285 outset.ds_prefer = dset->ds_prefer;
2289 case CPU_WHICH_CPUSET:
2290 case CPU_WHICH_JAIL:
2291 domainset_copy(set->cs_domain, &outset);
2294 case CPU_WHICH_INTRHANDLER:
2295 case CPU_WHICH_ITHREAD:
2296 case CPU_WHICH_DOMAIN:
2310 * Translate prefer into a set containing only the preferred domain,
2311 * not the entire fallback set.
2313 if (outset.ds_policy == DOMAINSET_POLICY_PREFER) {
2314 DOMAINSET_ZERO(&outset.ds_mask);
2315 DOMAINSET_SET(outset.ds_prefer, &outset.ds_mask);
2317 DOMAINSET_COPY(&outset.ds_mask, mask);
2319 error = cb->cpuset_copyout(mask, maskp, domainsetsize);
2321 if (suword32(policyp, outset.ds_policy) != 0)
2328 #ifndef _SYS_SYSPROTO_H_
2329 struct cpuset_setdomain_args {
2333 size_t domainsetsize;
2339 sys_cpuset_setdomain(struct thread *td, struct cpuset_setdomain_args *uap)
2342 return (kern_cpuset_setdomain(td, uap->level, uap->which,
2343 uap->id, uap->domainsetsize, uap->mask, uap->policy, ©_set));
2347 kern_cpuset_setdomain(struct thread *td, cpulevel_t level, cpuwhich_t which,
2348 id_t id, size_t domainsetsize, const domainset_t *maskp, int policy,
2349 const struct cpuset_copy_cb *cb)
2351 struct cpuset *nset;
2355 struct domainset domain;
2359 if (domainsetsize < sizeof(domainset_t) ||
2360 domainsetsize > DOMAINSET_MAXSIZE / NBBY)
2362 if (policy <= DOMAINSET_POLICY_INVALID ||
2363 policy > DOMAINSET_POLICY_MAX)
2365 error = cpuset_check_capabilities(td, level, which, id);
2368 memset(&domain, 0, sizeof(domain));
2369 mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO);
2370 error = cb->cpuset_copyin(maskp, mask, domainsetsize);
2374 * Verify that no high bits are set.
2376 if (domainsetsize > sizeof(domainset_t)) {
2380 end = cp = (char *)&mask->__bits;
2381 end += domainsetsize;
2382 cp += sizeof(domainset_t);
2389 if (DOMAINSET_EMPTY(mask)) {
2393 DOMAINSET_COPY(mask, &domain.ds_mask);
2394 domain.ds_policy = policy;
2397 * Sanitize the provided mask.
2399 if (!DOMAINSET_SUBSET(&all_domains, &domain.ds_mask)) {
2404 /* Translate preferred policy into a mask and fallback. */
2405 if (policy == DOMAINSET_POLICY_PREFER) {
2406 /* Only support a single preferred domain. */
2407 if (DOMAINSET_COUNT(&domain.ds_mask) != 1) {
2411 domain.ds_prefer = DOMAINSET_FFS(&domain.ds_mask) - 1;
2412 /* This will be constrained by domainset_shadow(). */
2413 DOMAINSET_COPY(&all_domains, &domain.ds_mask);
2417 * When given an impossible policy, fall back to interleaving
2418 * across all domains.
2420 if (domainset_empty_vm(&domain))
2421 domainset_copy(domainset2, &domain);
2424 case CPU_LEVEL_ROOT:
2425 case CPU_LEVEL_CPUSET:
2426 error = cpuset_which(which, id, &p, &ttd, &set);
2433 set = cpuset_ref(ttd->td_cpuset);
2437 case CPU_WHICH_CPUSET:
2438 case CPU_WHICH_JAIL:
2441 case CPU_WHICH_INTRHANDLER:
2442 case CPU_WHICH_ITHREAD:
2443 case CPU_WHICH_DOMAIN:
2447 if (level == CPU_LEVEL_ROOT)
2448 nset = cpuset_refroot(set);
2450 nset = cpuset_refbase(set);
2451 error = cpuset_modify_domain(nset, &domain);
2455 case CPU_LEVEL_WHICH:
2458 error = _cpuset_setthread(id, NULL, &domain);
2461 error = cpuset_setproc(id, NULL, NULL, &domain, false);
2463 case CPU_WHICH_CPUSET:
2464 case CPU_WHICH_JAIL:
2465 error = cpuset_which(which, id, &p, &ttd, &set);
2467 error = cpuset_modify_domain(set, &domain);
2472 case CPU_WHICH_INTRHANDLER:
2473 case CPU_WHICH_ITHREAD:
2491 ddb_display_bitset(const struct bitset *set, int size)
2495 for (once = 0, bit = 0; bit < size; bit++) {
2496 if (CPU_ISSET(bit, set)) {
2498 db_printf("%d", bit);
2501 db_printf(",%d", bit);
2505 db_printf("<none>");
2509 ddb_display_cpuset(const cpuset_t *set)
2511 ddb_display_bitset((const struct bitset *)set, CPU_SETSIZE);
2515 ddb_display_domainset(const domainset_t *set)
2517 ddb_display_bitset((const struct bitset *)set, DOMAINSET_SETSIZE);
2520 DB_SHOW_COMMAND(cpusets, db_show_cpusets)
2524 LIST_FOREACH(set, &cpuset_ids, cs_link) {
2525 db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n",
2526 set, set->cs_id, refcount_load(&set->cs_ref), set->cs_flags,
2527 (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0);
2528 db_printf(" cpu mask=");
2529 ddb_display_cpuset(&set->cs_mask);
2531 db_printf(" domain policy %d prefer %d mask=",
2532 set->cs_domain->ds_policy, set->cs_domain->ds_prefer);
2533 ddb_display_domainset(&set->cs_domain->ds_mask);
2540 DB_SHOW_COMMAND(domainsets, db_show_domainsets)
2542 struct domainset *set;
2544 LIST_FOREACH(set, &cpuset_domains, ds_link) {
2545 db_printf("set=%p policy %d prefer %d cnt %d\n",
2546 set, set->ds_policy, set->ds_prefer, set->ds_cnt);
2547 db_printf(" mask =");
2548 ddb_display_domainset(&set->ds_mask);