1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2008, Jeffrey Roberson <jeff@freebsd.org>
5 * All rights reserved.
6 *
7 * Copyright (c) 2008 Nokia Corporation
8 * All rights reserved.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice unmodified, this list of conditions, and the following
15 * disclaimer.
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.
19 *
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.
30 *
31 */
32
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD: releng/12.0/sys/kern/kern_cpuset.c 340401 2018-11-13 18:21:47Z markj $");
35
36 #include "opt_ddb.h"
37
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>
43 #include <sys/jail.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/mutex.h>
48 #include <sys/priv.h>
49 #include <sys/proc.h>
50 #include <sys/refcount.h>
51 #include <sys/sched.h>
52 #include <sys/smp.h>
53 #include <sys/syscallsubr.h>
54 #include <sys/capsicum.h>
55 #include <sys/cpuset.h>
56 #include <sys/domainset.h>
57 #include <sys/sx.h>
58 #include <sys/queue.h>
59 #include <sys/libkern.h>
60 #include <sys/limits.h>
61 #include <sys/bus.h>
62 #include <sys/interrupt.h>
63 #include <sys/vmmeter.h>
64
65 #include <vm/uma.h>
66 #include <vm/vm.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>
74
75 #ifdef DDB
76 #include <ddb/ddb.h>
77 #endif /* DDB */
78
79 /*
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.
83 *
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.
88 *
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.
93 *
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.
99 *
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.
105 *
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.
109 *
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.
114 *
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, ...).
119 */
120
121 LIST_HEAD(domainlist, domainset);
122 struct domainset __read_mostly domainset_fixed[MAXMEMDOM];
123 struct domainset __read_mostly domainset_prefer[MAXMEMDOM];
124 struct domainset __read_mostly domainset_roundrobin;
125
126 static uma_zone_t cpuset_zone;
127 static uma_zone_t domainset_zone;
128 static struct mtx cpuset_lock;
129 static struct setlist cpuset_ids;
130 static struct domainlist cpuset_domains;
131 static struct unrhdr *cpuset_unr;
132 static struct cpuset *cpuset_zero, *cpuset_default, *cpuset_kernel;
133 static struct domainset domainset0, domainset2;
134
135 /* Return the size of cpuset_t at the kernel level */
136 SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD | CTLFLAG_CAPRD,
137 SYSCTL_NULL_INT_PTR, sizeof(cpuset_t), "sizeof(cpuset_t)");
138
139 cpuset_t *cpuset_root;
140 cpuset_t cpuset_domain[MAXMEMDOM];
141
142 static int domainset_valid(const struct domainset *, const struct domainset *);
143
144 /*
145 * Find the first non-anonymous set starting from 'set'.
146 */
147 static struct cpuset *
148 cpuset_getbase(struct cpuset *set)
149 {
150
151 if (set->cs_id == CPUSET_INVALID)
152 set = set->cs_parent;
153 return (set);
154 }
155
156 /*
157 * Walks up the tree from 'set' to find the root.
158 */
159 static struct cpuset *
160 cpuset_getroot(struct cpuset *set)
161 {
162
163 while ((set->cs_flags & CPU_SET_ROOT) == 0 && set->cs_parent != NULL)
164 set = set->cs_parent;
165 return (set);
166 }
167
168 /*
169 * Acquire a reference to a cpuset, all pointers must be tracked with refs.
170 */
171 struct cpuset *
172 cpuset_ref(struct cpuset *set)
173 {
174
175 refcount_acquire(&set->cs_ref);
176 return (set);
177 }
178
179 /*
180 * Walks up the tree from 'set' to find the root. Returns the root
181 * referenced.
182 */
183 static struct cpuset *
184 cpuset_refroot(struct cpuset *set)
185 {
186
187 return (cpuset_ref(cpuset_getroot(set)));
188 }
189
190 /*
191 * Find the first non-anonymous set starting from 'set'. Returns this set
192 * referenced. May return the passed in set with an extra ref if it is
193 * not anonymous.
194 */
195 static struct cpuset *
196 cpuset_refbase(struct cpuset *set)
197 {
198
199 return (cpuset_ref(cpuset_getbase(set)));
200 }
201
202 /*
203 * Release a reference in a context where it is safe to allocate.
204 */
205 void
206 cpuset_rel(struct cpuset *set)
207 {
208 cpusetid_t id;
209
210 if (refcount_release(&set->cs_ref) == 0)
211 return;
212 mtx_lock_spin(&cpuset_lock);
213 LIST_REMOVE(set, cs_siblings);
214 id = set->cs_id;
215 if (id != CPUSET_INVALID)
216 LIST_REMOVE(set, cs_link);
217 mtx_unlock_spin(&cpuset_lock);
218 cpuset_rel(set->cs_parent);
219 uma_zfree(cpuset_zone, set);
220 if (id != CPUSET_INVALID)
221 free_unr(cpuset_unr, id);
222 }
223
224 /*
225 * Deferred release must be used when in a context that is not safe to
226 * allocate/free. This places any unreferenced sets on the list 'head'.
227 */
228 static void
229 cpuset_rel_defer(struct setlist *head, struct cpuset *set)
230 {
231
232 if (refcount_release(&set->cs_ref) == 0)
233 return;
234 mtx_lock_spin(&cpuset_lock);
235 LIST_REMOVE(set, cs_siblings);
236 if (set->cs_id != CPUSET_INVALID)
237 LIST_REMOVE(set, cs_link);
238 LIST_INSERT_HEAD(head, set, cs_link);
239 mtx_unlock_spin(&cpuset_lock);
240 }
241
242 /*
243 * Complete a deferred release. Removes the set from the list provided to
244 * cpuset_rel_defer.
245 */
246 static void
247 cpuset_rel_complete(struct cpuset *set)
248 {
249 LIST_REMOVE(set, cs_link);
250 cpuset_rel(set->cs_parent);
251 uma_zfree(cpuset_zone, set);
252 }
253
254 /*
255 * Find a set based on an id. Returns it with a ref.
256 */
257 static struct cpuset *
258 cpuset_lookup(cpusetid_t setid, struct thread *td)
259 {
260 struct cpuset *set;
261
262 if (setid == CPUSET_INVALID)
263 return (NULL);
264 mtx_lock_spin(&cpuset_lock);
265 LIST_FOREACH(set, &cpuset_ids, cs_link)
266 if (set->cs_id == setid)
267 break;
268 if (set)
269 cpuset_ref(set);
270 mtx_unlock_spin(&cpuset_lock);
271
272 KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__));
273 if (set != NULL && jailed(td->td_ucred)) {
274 struct cpuset *jset, *tset;
275
276 jset = td->td_ucred->cr_prison->pr_cpuset;
277 for (tset = set; tset != NULL; tset = tset->cs_parent)
278 if (tset == jset)
279 break;
280 if (tset == NULL) {
281 cpuset_rel(set);
282 set = NULL;
283 }
284 }
285
286 return (set);
287 }
288
289 /*
290 * Create a set in the space provided in 'set' with the provided parameters.
291 * The set is returned with a single ref. May return EDEADLK if the set
292 * will have no valid cpu based on restrictions from the parent.
293 */
294 static int
295 _cpuset_create(struct cpuset *set, struct cpuset *parent,
296 const cpuset_t *mask, struct domainset *domain, cpusetid_t id)
297 {
298
299 if (domain == NULL)
300 domain = parent->cs_domain;
301 if (mask == NULL)
302 mask = &parent->cs_mask;
303 if (!CPU_OVERLAP(&parent->cs_mask, mask))
304 return (EDEADLK);
305 /* The domain must be prepared ahead of time. */
306 if (!domainset_valid(parent->cs_domain, domain))
307 return (EDEADLK);
308 CPU_COPY(mask, &set->cs_mask);
309 LIST_INIT(&set->cs_children);
310 refcount_init(&set->cs_ref, 1);
311 set->cs_flags = 0;
312 mtx_lock_spin(&cpuset_lock);
313 set->cs_domain = domain;
314 CPU_AND(&set->cs_mask, &parent->cs_mask);
315 set->cs_id = id;
316 set->cs_parent = cpuset_ref(parent);
317 LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings);
318 if (set->cs_id != CPUSET_INVALID)
319 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
320 mtx_unlock_spin(&cpuset_lock);
321
322 return (0);
323 }
324
325 /*
326 * Create a new non-anonymous set with the requested parent and mask. May
327 * return failures if the mask is invalid or a new number can not be
328 * allocated.
329 */
330 static int
331 cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask)
332 {
333 struct cpuset *set;
334 cpusetid_t id;
335 int error;
336
337 id = alloc_unr(cpuset_unr);
338 if (id == -1)
339 return (ENFILE);
340 *setp = set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
341 error = _cpuset_create(set, parent, mask, NULL, id);
342 if (error == 0)
343 return (0);
344 free_unr(cpuset_unr, id);
345 uma_zfree(cpuset_zone, set);
346
347 return (error);
348 }
349
350 static void
351 cpuset_freelist_add(struct setlist *list, int count)
352 {
353 struct cpuset *set;
354 int i;
355
356 for (i = 0; i < count; i++) {
357 set = uma_zalloc(cpuset_zone, M_ZERO | M_WAITOK);
358 LIST_INSERT_HEAD(list, set, cs_link);
359 }
360 }
361
362 static void
363 cpuset_freelist_init(struct setlist *list, int count)
364 {
365
366 LIST_INIT(list);
367 cpuset_freelist_add(list, count);
368 }
369
370 static void
371 cpuset_freelist_free(struct setlist *list)
372 {
373 struct cpuset *set;
374
375 while ((set = LIST_FIRST(list)) != NULL) {
376 LIST_REMOVE(set, cs_link);
377 uma_zfree(cpuset_zone, set);
378 }
379 }
380
381 static void
382 domainset_freelist_add(struct domainlist *list, int count)
383 {
384 struct domainset *set;
385 int i;
386
387 for (i = 0; i < count; i++) {
388 set = uma_zalloc(domainset_zone, M_ZERO | M_WAITOK);
389 LIST_INSERT_HEAD(list, set, ds_link);
390 }
391 }
392
393 static void
394 domainset_freelist_init(struct domainlist *list, int count)
395 {
396
397 LIST_INIT(list);
398 domainset_freelist_add(list, count);
399 }
400
401 static void
402 domainset_freelist_free(struct domainlist *list)
403 {
404 struct domainset *set;
405
406 while ((set = LIST_FIRST(list)) != NULL) {
407 LIST_REMOVE(set, ds_link);
408 uma_zfree(domainset_zone, set);
409 }
410 }
411
412 /* Copy a domainset preserving mask and policy. */
413 static void
414 domainset_copy(const struct domainset *from, struct domainset *to)
415 {
416
417 DOMAINSET_COPY(&from->ds_mask, &to->ds_mask);
418 to->ds_policy = from->ds_policy;
419 to->ds_prefer = from->ds_prefer;
420 }
421
422 /* Return 1 if mask and policy are equal, otherwise 0. */
423 static int
424 domainset_equal(const struct domainset *one, const struct domainset *two)
425 {
426
427 return (DOMAINSET_CMP(&one->ds_mask, &two->ds_mask) == 0 &&
428 one->ds_policy == two->ds_policy &&
429 one->ds_prefer == two->ds_prefer);
430 }
431
432 /* Return 1 if child is a valid subset of parent. */
433 static int
434 domainset_valid(const struct domainset *parent, const struct domainset *child)
435 {
436 if (child->ds_policy != DOMAINSET_POLICY_PREFER)
437 return (DOMAINSET_SUBSET(&parent->ds_mask, &child->ds_mask));
438 return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask));
439 }
440
441 static int
442 domainset_restrict(const struct domainset *parent,
443 const struct domainset *child)
444 {
445 if (child->ds_policy != DOMAINSET_POLICY_PREFER)
446 return (DOMAINSET_OVERLAP(&parent->ds_mask, &child->ds_mask));
447 return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask));
448 }
449
450 /*
451 * Lookup or create a domainset. The key is provided in ds_mask and
452 * ds_policy. If the domainset does not yet exist the storage in
453 * 'domain' is used to insert. Otherwise this storage is freed to the
454 * domainset_zone and the existing domainset is returned.
455 */
456 static struct domainset *
457 _domainset_create(struct domainset *domain, struct domainlist *freelist)
458 {
459 struct domainset *ndomain;
460 int i, j, max;
461
462 KASSERT(domain->ds_cnt <= vm_ndomains,
463 ("invalid domain count in domainset %p", domain));
464 KASSERT(domain->ds_policy != DOMAINSET_POLICY_PREFER ||
465 domain->ds_prefer < vm_ndomains,
466 ("invalid preferred domain in domains %p", domain));
467
468 mtx_lock_spin(&cpuset_lock);
469 LIST_FOREACH(ndomain, &cpuset_domains, ds_link)
470 if (domainset_equal(ndomain, domain))
471 break;
472 /*
473 * If the domain does not yet exist we insert it and initialize
474 * various iteration helpers which are not part of the key.
475 */
476 if (ndomain == NULL) {
477 LIST_INSERT_HEAD(&cpuset_domains, domain, ds_link);
478 domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask);
479 max = DOMAINSET_FLS(&domain->ds_mask) + 1;
480 for (i = 0, j = 0; i < max; i++)
481 if (DOMAINSET_ISSET(i, &domain->ds_mask))
482 domain->ds_order[j++] = i;
483 }
484 mtx_unlock_spin(&cpuset_lock);
485 if (ndomain == NULL)
486 return (domain);
487 if (freelist != NULL)
488 LIST_INSERT_HEAD(freelist, domain, ds_link);
489 else
490 uma_zfree(domainset_zone, domain);
491 return (ndomain);
492
493 }
494
495 /*
496 * Are any of the domains in the mask empty? If so, silently
497 * remove them and update the domainset accordingly. If only empty
498 * domains are present, we must return failure.
499 */
500 static bool
501 domainset_empty_vm(struct domainset *domain)
502 {
503 int i, j, max;
504
505 max = DOMAINSET_FLS(&domain->ds_mask) + 1;
506 for (i = 0; i < max; i++)
507 if (DOMAINSET_ISSET(i, &domain->ds_mask) && VM_DOMAIN_EMPTY(i))
508 DOMAINSET_CLR(i, &domain->ds_mask);
509 domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask);
510 max = DOMAINSET_FLS(&domain->ds_mask) + 1;
511 for (i = j = 0; i < max; i++) {
512 if (DOMAINSET_ISSET(i, &domain->ds_mask))
513 domain->ds_order[j++] = i;
514 else if (domain->ds_policy == DOMAINSET_POLICY_PREFER &&
515 domain->ds_prefer == i && domain->ds_cnt > 1) {
516 domain->ds_policy = DOMAINSET_POLICY_ROUNDROBIN;
517 domain->ds_prefer = -1;
518 }
519 }
520
521 return (DOMAINSET_EMPTY(&domain->ds_mask));
522 }
523
524 /*
525 * Create or lookup a domainset based on the key held in 'domain'.
526 */
527 struct domainset *
528 domainset_create(const struct domainset *domain)
529 {
530 struct domainset *ndomain;
531
532 /*
533 * Validate the policy. It must specify a useable policy number with
534 * only valid domains. Preferred must include the preferred domain
535 * in the mask.
536 */
537 if (domain->ds_policy <= DOMAINSET_POLICY_INVALID ||
538 domain->ds_policy > DOMAINSET_POLICY_MAX)
539 return (NULL);
540 if (domain->ds_policy == DOMAINSET_POLICY_PREFER &&
541 !DOMAINSET_ISSET(domain->ds_prefer, &domain->ds_mask))
542 return (NULL);
543 if (!DOMAINSET_SUBSET(&domainset0.ds_mask, &domain->ds_mask))
544 return (NULL);
545 ndomain = uma_zalloc(domainset_zone, M_WAITOK | M_ZERO);
546 domainset_copy(domain, ndomain);
547 return _domainset_create(ndomain, NULL);
548 }
549
550 /*
551 * Update thread domainset pointers.
552 */
553 static void
554 domainset_notify(void)
555 {
556 struct thread *td;
557 struct proc *p;
558
559 sx_slock(&allproc_lock);
560 FOREACH_PROC_IN_SYSTEM(p) {
561 PROC_LOCK(p);
562 if (p->p_state == PRS_NEW) {
563 PROC_UNLOCK(p);
564 continue;
565 }
566 FOREACH_THREAD_IN_PROC(p, td) {
567 thread_lock(td);
568 td->td_domain.dr_policy = td->td_cpuset->cs_domain;
569 thread_unlock(td);
570 }
571 PROC_UNLOCK(p);
572 }
573 sx_sunlock(&allproc_lock);
574 kernel_object->domain.dr_policy = cpuset_kernel->cs_domain;
575 }
576
577 /*
578 * Create a new set that is a subset of a parent.
579 */
580 static struct domainset *
581 domainset_shadow(const struct domainset *pdomain,
582 const struct domainset *domain, struct domainlist *freelist)
583 {
584 struct domainset *ndomain;
585
586 ndomain = LIST_FIRST(freelist);
587 LIST_REMOVE(ndomain, ds_link);
588
589 /*
590 * Initialize the key from the request.
591 */
592 domainset_copy(domain, ndomain);
593
594 /*
595 * Restrict the key by the parent.
596 */
597 DOMAINSET_AND(&ndomain->ds_mask, &pdomain->ds_mask);
598
599 return _domainset_create(ndomain, freelist);
600 }
601
602 /*
603 * Recursively check for errors that would occur from applying mask to
604 * the tree of sets starting at 'set'. Checks for sets that would become
605 * empty as well as RDONLY flags.
606 */
607 static int
608 cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int check_mask)
609 {
610 struct cpuset *nset;
611 cpuset_t newmask;
612 int error;
613
614 mtx_assert(&cpuset_lock, MA_OWNED);
615 if (set->cs_flags & CPU_SET_RDONLY)
616 return (EPERM);
617 if (check_mask) {
618 if (!CPU_OVERLAP(&set->cs_mask, mask))
619 return (EDEADLK);
620 CPU_COPY(&set->cs_mask, &newmask);
621 CPU_AND(&newmask, mask);
622 } else
623 CPU_COPY(mask, &newmask);
624 error = 0;
625 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
626 if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0)
627 break;
628 return (error);
629 }
630
631 /*
632 * Applies the mask 'mask' without checking for empty sets or permissions.
633 */
634 static void
635 cpuset_update(struct cpuset *set, cpuset_t *mask)
636 {
637 struct cpuset *nset;
638
639 mtx_assert(&cpuset_lock, MA_OWNED);
640 CPU_AND(&set->cs_mask, mask);
641 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
642 cpuset_update(nset, &set->cs_mask);
643
644 return;
645 }
646
647 /*
648 * Modify the set 'set' to use a copy of the mask provided. Apply this new
649 * mask to restrict all children in the tree. Checks for validity before
650 * applying the changes.
651 */
652 static int
653 cpuset_modify(struct cpuset *set, cpuset_t *mask)
654 {
655 struct cpuset *root;
656 int error;
657
658 error = priv_check(curthread, PRIV_SCHED_CPUSET);
659 if (error)
660 return (error);
661 /*
662 * In case we are called from within the jail
663 * we do not allow modifying the dedicated root
664 * cpuset of the jail but may still allow to
665 * change child sets.
666 */
667 if (jailed(curthread->td_ucred) &&
668 set->cs_flags & CPU_SET_ROOT)
669 return (EPERM);
670 /*
671 * Verify that we have access to this set of
672 * cpus.
673 */
674 root = cpuset_getroot(set);
675 mtx_lock_spin(&cpuset_lock);
676 if (root && !CPU_SUBSET(&root->cs_mask, mask)) {
677 error = EINVAL;
678 goto out;
679 }
680 error = cpuset_testupdate(set, mask, 0);
681 if (error)
682 goto out;
683 CPU_COPY(mask, &set->cs_mask);
684 cpuset_update(set, mask);
685 out:
686 mtx_unlock_spin(&cpuset_lock);
687
688 return (error);
689 }
690
691 /*
692 * Recursively check for errors that would occur from applying mask to
693 * the tree of sets starting at 'set'. Checks for sets that would become
694 * empty as well as RDONLY flags.
695 */
696 static int
697 cpuset_testupdate_domain(struct cpuset *set, struct domainset *dset,
698 struct domainset *orig, int *count, int check_mask)
699 {
700 struct cpuset *nset;
701 struct domainset *domain;
702 struct domainset newset;
703 int error;
704
705 mtx_assert(&cpuset_lock, MA_OWNED);
706 if (set->cs_flags & CPU_SET_RDONLY)
707 return (EPERM);
708 domain = set->cs_domain;
709 domainset_copy(domain, &newset);
710 if (!domainset_equal(domain, orig)) {
711 if (!domainset_restrict(domain, dset))
712 return (EDEADLK);
713 DOMAINSET_AND(&newset.ds_mask, &dset->ds_mask);
714 /* Count the number of domains that are changing. */
715 (*count)++;
716 }
717 error = 0;
718 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
719 if ((error = cpuset_testupdate_domain(nset, &newset, domain,
720 count, 1)) != 0)
721 break;
722 return (error);
723 }
724
725 /*
726 * Applies the mask 'mask' without checking for empty sets or permissions.
727 */
728 static void
729 cpuset_update_domain(struct cpuset *set, struct domainset *domain,
730 struct domainset *orig, struct domainlist *domains)
731 {
732 struct cpuset *nset;
733
734 mtx_assert(&cpuset_lock, MA_OWNED);
735 /*
736 * If this domainset has changed from the parent we must calculate
737 * a new set. Otherwise it simply inherits from the parent. When
738 * we inherit from the parent we get a new mask and policy. If the
739 * set is modified from the parent we keep the policy and only
740 * update the mask.
741 */
742 if (set->cs_domain != orig) {
743 orig = set->cs_domain;
744 set->cs_domain = domainset_shadow(domain, orig, domains);
745 } else
746 set->cs_domain = domain;
747 LIST_FOREACH(nset, &set->cs_children, cs_siblings)
748 cpuset_update_domain(nset, set->cs_domain, orig, domains);
749
750 return;
751 }
752
753 /*
754 * Modify the set 'set' to use a copy the domainset provided. Apply this new
755 * mask to restrict all children in the tree. Checks for validity before
756 * applying the changes.
757 */
758 static int
759 cpuset_modify_domain(struct cpuset *set, struct domainset *domain)
760 {
761 struct domainlist domains;
762 struct domainset temp;
763 struct domainset *dset;
764 struct cpuset *root;
765 int ndomains, needed;
766 int error;
767
768 error = priv_check(curthread, PRIV_SCHED_CPUSET);
769 if (error)
770 return (error);
771 /*
772 * In case we are called from within the jail
773 * we do not allow modifying the dedicated root
774 * cpuset of the jail but may still allow to
775 * change child sets.
776 */
777 if (jailed(curthread->td_ucred) &&
778 set->cs_flags & CPU_SET_ROOT)
779 return (EPERM);
780 domainset_freelist_init(&domains, 0);
781 domain = domainset_create(domain);
782 ndomains = needed = 0;
783 do {
784 if (ndomains < needed) {
785 domainset_freelist_add(&domains, needed - ndomains);
786 ndomains = needed;
787 }
788 root = cpuset_getroot(set);
789 mtx_lock_spin(&cpuset_lock);
790 dset = root->cs_domain;
791 /*
792 * Verify that we have access to this set of domains.
793 */
794 if (root && !domainset_valid(dset, domain)) {
795 error = EINVAL;
796 goto out;
797 }
798 /*
799 * If applying prefer we keep the current set as the fallback.
800 */
801 if (domain->ds_policy == DOMAINSET_POLICY_PREFER)
802 DOMAINSET_COPY(&set->cs_domain->ds_mask,
803 &domain->ds_mask);
804 /*
805 * Determine whether we can apply this set of domains and
806 * how many new domain structures it will require.
807 */
808 domainset_copy(domain, &temp);
809 needed = 0;
810 error = cpuset_testupdate_domain(set, &temp, set->cs_domain,
811 &needed, 0);
812 if (error)
813 goto out;
814 } while (ndomains < needed);
815 dset = set->cs_domain;
816 cpuset_update_domain(set, domain, dset, &domains);
817 out:
818 mtx_unlock_spin(&cpuset_lock);
819 domainset_freelist_free(&domains);
820 if (error == 0)
821 domainset_notify();
822
823 return (error);
824 }
825
826 /*
827 * Resolve the 'which' parameter of several cpuset apis.
828 *
829 * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid. Also
830 * checks for permission via p_cansched().
831 *
832 * For WHICH_SET returns a valid set with a new reference.
833 *
834 * -1 may be supplied for any argument to mean the current proc/thread or
835 * the base set of the current thread. May fail with ESRCH/EPERM.
836 */
837 int
838 cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp,
839 struct cpuset **setp)
840 {
841 struct cpuset *set;
842 struct thread *td;
843 struct proc *p;
844 int error;
845
846 *pp = p = NULL;
847 *tdp = td = NULL;
848 *setp = set = NULL;
849 switch (which) {
850 case CPU_WHICH_PID:
851 if (id == -1) {
852 PROC_LOCK(curproc);
853 p = curproc;
854 break;
855 }
856 if ((p = pfind(id)) == NULL)
857 return (ESRCH);
858 break;
859 case CPU_WHICH_TID:
860 if (id == -1) {
861 PROC_LOCK(curproc);
862 p = curproc;
863 td = curthread;
864 break;
865 }
866 td = tdfind(id, -1);
867 if (td == NULL)
868 return (ESRCH);
869 p = td->td_proc;
870 break;
871 case CPU_WHICH_CPUSET:
872 if (id == -1) {
873 thread_lock(curthread);
874 set = cpuset_refbase(curthread->td_cpuset);
875 thread_unlock(curthread);
876 } else
877 set = cpuset_lookup(id, curthread);
878 if (set) {
879 *setp = set;
880 return (0);
881 }
882 return (ESRCH);
883 case CPU_WHICH_JAIL:
884 {
885 /* Find `set' for prison with given id. */
886 struct prison *pr;
887
888 sx_slock(&allprison_lock);
889 pr = prison_find_child(curthread->td_ucred->cr_prison, id);
890 sx_sunlock(&allprison_lock);
891 if (pr == NULL)
892 return (ESRCH);
893 cpuset_ref(pr->pr_cpuset);
894 *setp = pr->pr_cpuset;
895 mtx_unlock(&pr->pr_mtx);
896 return (0);
897 }
898 case CPU_WHICH_IRQ:
899 case CPU_WHICH_DOMAIN:
900 return (0);
901 default:
902 return (EINVAL);
903 }
904 error = p_cansched(curthread, p);
905 if (error) {
906 PROC_UNLOCK(p);
907 return (error);
908 }
909 if (td == NULL)
910 td = FIRST_THREAD_IN_PROC(p);
911 *pp = p;
912 *tdp = td;
913 return (0);
914 }
915
916 static int
917 cpuset_testshadow(struct cpuset *set, const cpuset_t *mask,
918 const struct domainset *domain)
919 {
920 struct cpuset *parent;
921 struct domainset *dset;
922
923 parent = cpuset_getbase(set);
924 /*
925 * If we are restricting a cpu mask it must be a subset of the
926 * parent or invalid CPUs have been specified.
927 */
928 if (mask != NULL && !CPU_SUBSET(&parent->cs_mask, mask))
929 return (EINVAL);
930
931 /*
932 * If we are restricting a domain mask it must be a subset of the
933 * parent or invalid domains have been specified.
934 */
935 dset = parent->cs_domain;
936 if (domain != NULL && !domainset_valid(dset, domain))
937 return (EINVAL);
938
939 return (0);
940 }
941
942 /*
943 * Create an anonymous set with the provided mask in the space provided by
944 * 'nset'. If the passed in set is anonymous we use its parent otherwise
945 * the new set is a child of 'set'.
946 */
947 static int
948 cpuset_shadow(struct cpuset *set, struct cpuset **nsetp,
949 const cpuset_t *mask, const struct domainset *domain,
950 struct setlist *cpusets, struct domainlist *domains)
951 {
952 struct cpuset *parent;
953 struct cpuset *nset;
954 struct domainset *dset;
955 struct domainset *d;
956 int error;
957
958 error = cpuset_testshadow(set, mask, domain);
959 if (error)
960 return (error);
961
962 parent = cpuset_getbase(set);
963 dset = parent->cs_domain;
964 if (mask == NULL)
965 mask = &set->cs_mask;
966 if (domain != NULL)
967 d = domainset_shadow(dset, domain, domains);
968 else
969 d = set->cs_domain;
970 nset = LIST_FIRST(cpusets);
971 error = _cpuset_create(nset, parent, mask, d, CPUSET_INVALID);
972 if (error == 0) {
973 LIST_REMOVE(nset, cs_link);
974 *nsetp = nset;
975 }
976 return (error);
977 }
978
979 static struct cpuset *
980 cpuset_update_thread(struct thread *td, struct cpuset *nset)
981 {
982 struct cpuset *tdset;
983
984 tdset = td->td_cpuset;
985 td->td_cpuset = nset;
986 td->td_domain.dr_policy = nset->cs_domain;
987 sched_affinity(td);
988
989 return (tdset);
990 }
991
992 static int
993 cpuset_setproc_test_maskthread(struct cpuset *tdset, cpuset_t *mask,
994 struct domainset *domain)
995 {
996 struct cpuset *parent;
997
998 parent = cpuset_getbase(tdset);
999 if (mask == NULL)
1000 mask = &tdset->cs_mask;
1001 if (domain == NULL)
1002 domain = tdset->cs_domain;
1003 return cpuset_testshadow(parent, mask, domain);
1004 }
1005
1006 static int
1007 cpuset_setproc_maskthread(struct cpuset *tdset, cpuset_t *mask,
1008 struct domainset *domain, struct cpuset **nsetp,
1009 struct setlist *freelist, struct domainlist *domainlist)
1010 {
1011 struct cpuset *parent;
1012
1013 parent = cpuset_getbase(tdset);
1014 if (mask == NULL)
1015 mask = &tdset->cs_mask;
1016 if (domain == NULL)
1017 domain = tdset->cs_domain;
1018 return cpuset_shadow(parent, nsetp, mask, domain, freelist,
1019 domainlist);
1020 }
1021
1022 static int
1023 cpuset_setproc_setthread_mask(struct cpuset *tdset, struct cpuset *set,
1024 cpuset_t *mask, struct domainset *domain)
1025 {
1026 struct cpuset *parent;
1027
1028 parent = cpuset_getbase(tdset);
1029
1030 /*
1031 * If the thread restricted its mask then apply that same
1032 * restriction to the new set, otherwise take it wholesale.
1033 */
1034 if (CPU_CMP(&tdset->cs_mask, &parent->cs_mask) != 0) {
1035 CPU_COPY(&tdset->cs_mask, mask);
1036 CPU_AND(mask, &set->cs_mask);
1037 } else
1038 CPU_COPY(&set->cs_mask, mask);
1039
1040 /*
1041 * If the thread restricted the domain then we apply the
1042 * restriction to the new set but retain the policy.
1043 */
1044 if (tdset->cs_domain != parent->cs_domain) {
1045 domainset_copy(tdset->cs_domain, domain);
1046 DOMAINSET_AND(&domain->ds_mask, &set->cs_domain->ds_mask);
1047 } else
1048 domainset_copy(set->cs_domain, domain);
1049
1050 if (CPU_EMPTY(mask) || DOMAINSET_EMPTY(&domain->ds_mask))
1051 return (EDEADLK);
1052
1053 return (0);
1054 }
1055
1056 static int
1057 cpuset_setproc_test_setthread(struct cpuset *tdset, struct cpuset *set)
1058 {
1059 struct domainset domain;
1060 cpuset_t mask;
1061
1062 if (tdset->cs_id != CPUSET_INVALID)
1063 return (0);
1064 return cpuset_setproc_setthread_mask(tdset, set, &mask, &domain);
1065 }
1066
1067 static int
1068 cpuset_setproc_setthread(struct cpuset *tdset, struct cpuset *set,
1069 struct cpuset **nsetp, struct setlist *freelist,
1070 struct domainlist *domainlist)
1071 {
1072 struct domainset domain;
1073 cpuset_t mask;
1074 int error;
1075
1076 /*
1077 * If we're replacing on a thread that has not constrained the
1078 * original set we can simply accept the new set.
1079 */
1080 if (tdset->cs_id != CPUSET_INVALID) {
1081 *nsetp = cpuset_ref(set);
1082 return (0);
1083 }
1084 error = cpuset_setproc_setthread_mask(tdset, set, &mask, &domain);
1085 if (error)
1086 return (error);
1087
1088 return cpuset_shadow(tdset, nsetp, &mask, &domain, freelist,
1089 domainlist);
1090 }
1091
1092 /*
1093 * Handle three cases for updating an entire process.
1094 *
1095 * 1) Set is non-null. This reparents all anonymous sets to the provided
1096 * set and replaces all non-anonymous td_cpusets with the provided set.
1097 * 2) Mask is non-null. This replaces or creates anonymous sets for every
1098 * thread with the existing base as a parent.
1099 * 3) domain is non-null. This creates anonymous sets for every thread
1100 * and replaces the domain set.
1101 *
1102 * This is overly complicated because we can't allocate while holding a
1103 * spinlock and spinlocks must be held while changing and examining thread
1104 * state.
1105 */
1106 static int
1107 cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask,
1108 struct domainset *domain)
1109 {
1110 struct setlist freelist;
1111 struct setlist droplist;
1112 struct domainlist domainlist;
1113 struct cpuset *nset;
1114 struct thread *td;
1115 struct proc *p;
1116 int threads;
1117 int nfree;
1118 int error;
1119
1120 /*
1121 * The algorithm requires two passes due to locking considerations.
1122 *
1123 * 1) Lookup the process and acquire the locks in the required order.
1124 * 2) If enough cpusets have not been allocated release the locks and
1125 * allocate them. Loop.
1126 */
1127 cpuset_freelist_init(&freelist, 1);
1128 domainset_freelist_init(&domainlist, 1);
1129 nfree = 1;
1130 LIST_INIT(&droplist);
1131 nfree = 0;
1132 for (;;) {
1133 error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset);
1134 if (error)
1135 goto out;
1136 if (nfree >= p->p_numthreads)
1137 break;
1138 threads = p->p_numthreads;
1139 PROC_UNLOCK(p);
1140 if (nfree < threads) {
1141 cpuset_freelist_add(&freelist, threads - nfree);
1142 domainset_freelist_add(&domainlist, threads - nfree);
1143 nfree = threads;
1144 }
1145 }
1146 PROC_LOCK_ASSERT(p, MA_OWNED);
1147 /*
1148 * Now that the appropriate locks are held and we have enough cpusets,
1149 * make sure the operation will succeed before applying changes. The
1150 * proc lock prevents td_cpuset from changing between calls.
1151 */
1152 error = 0;
1153 FOREACH_THREAD_IN_PROC(p, td) {
1154 thread_lock(td);
1155 if (set != NULL)
1156 error = cpuset_setproc_test_setthread(td->td_cpuset,
1157 set);
1158 else
1159 error = cpuset_setproc_test_maskthread(td->td_cpuset,
1160 mask, domain);
1161 thread_unlock(td);
1162 if (error)
1163 goto unlock_out;
1164 }
1165 /*
1166 * Replace each thread's cpuset while using deferred release. We
1167 * must do this because the thread lock must be held while operating
1168 * on the thread and this limits the type of operations allowed.
1169 */
1170 FOREACH_THREAD_IN_PROC(p, td) {
1171 thread_lock(td);
1172 if (set != NULL)
1173 error = cpuset_setproc_setthread(td->td_cpuset, set,
1174 &nset, &freelist, &domainlist);
1175 else
1176 error = cpuset_setproc_maskthread(td->td_cpuset, mask,
1177 domain, &nset, &freelist, &domainlist);
1178 if (error) {
1179 thread_unlock(td);
1180 break;
1181 }
1182 cpuset_rel_defer(&droplist, cpuset_update_thread(td, nset));
1183 thread_unlock(td);
1184 }
1185 unlock_out:
1186 PROC_UNLOCK(p);
1187 out:
1188 while ((nset = LIST_FIRST(&droplist)) != NULL)
1189 cpuset_rel_complete(nset);
1190 cpuset_freelist_free(&freelist);
1191 domainset_freelist_free(&domainlist);
1192 return (error);
1193 }
1194
1195 static int
1196 bitset_strprint(char *buf, size_t bufsiz, const struct bitset *set, int setlen)
1197 {
1198 size_t bytes;
1199 int i, once;
1200 char *p;
1201
1202 once = 0;
1203 p = buf;
1204 for (i = 0; i < __bitset_words(setlen); i++) {
1205 if (once != 0) {
1206 if (bufsiz < 1)
1207 return (0);
1208 *p = ',';
1209 p++;
1210 bufsiz--;
1211 } else
1212 once = 1;
1213 if (bufsiz < sizeof(__STRING(ULONG_MAX)))
1214 return (0);
1215 bytes = snprintf(p, bufsiz, "%lx", set->__bits[i]);
1216 p += bytes;
1217 bufsiz -= bytes;
1218 }
1219 return (p - buf);
1220 }
1221
1222 static int
1223 bitset_strscan(struct bitset *set, int setlen, const char *buf)
1224 {
1225 int i, ret;
1226 const char *p;
1227
1228 BIT_ZERO(setlen, set);
1229 p = buf;
1230 for (i = 0; i < __bitset_words(setlen); i++) {
1231 if (*p == ',') {
1232 p++;
1233 continue;
1234 }
1235 ret = sscanf(p, "%lx", &set->__bits[i]);
1236 if (ret == 0 || ret == -1)
1237 break;
1238 while (isxdigit(*p))
1239 p++;
1240 }
1241 return (p - buf);
1242 }
1243
1244 /*
1245 * Return a string representing a valid layout for a cpuset_t object.
1246 * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
1247 */
1248 char *
1249 cpusetobj_strprint(char *buf, const cpuset_t *set)
1250 {
1251
1252 bitset_strprint(buf, CPUSETBUFSIZ, (const struct bitset *)set,
1253 CPU_SETSIZE);
1254 return (buf);
1255 }
1256
1257 /*
1258 * Build a valid cpuset_t object from a string representation.
1259 * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
1260 */
1261 int
1262 cpusetobj_strscan(cpuset_t *set, const char *buf)
1263 {
1264 char p;
1265
1266 if (strlen(buf) > CPUSETBUFSIZ - 1)
1267 return (-1);
1268
1269 p = buf[bitset_strscan((struct bitset *)set, CPU_SETSIZE, buf)];
1270 if (p != '\0')
1271 return (-1);
1272
1273 return (0);
1274 }
1275
1276 /*
1277 * Handle a domainset specifier in the sysctl tree. A poiner to a pointer to
1278 * a domainset is in arg1. If the user specifies a valid domainset the
1279 * pointer is updated.
1280 *
1281 * Format is:
1282 * hex mask word 0,hex mask word 1,...:decimal policy:decimal preferred
1283 */
1284 int
1285 sysctl_handle_domainset(SYSCTL_HANDLER_ARGS)
1286 {
1287 char buf[DOMAINSETBUFSIZ];
1288 struct domainset *dset;
1289 struct domainset key;
1290 int policy, prefer, error;
1291 char *p;
1292
1293 dset = *(struct domainset **)arg1;
1294 error = 0;
1295
1296 if (dset != NULL) {
1297 p = buf + bitset_strprint(buf, DOMAINSETBUFSIZ,
1298 (const struct bitset *)&dset->ds_mask, DOMAINSET_SETSIZE);
1299 sprintf(p, ":%d:%d", dset->ds_policy, dset->ds_prefer);
1300 } else
1301 sprintf(buf, "<NULL>");
1302 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
1303 if (error != 0 || req->newptr == NULL)
1304 return (error);
1305
1306 /*
1307 * Read in and validate the string.
1308 */
1309 memset(&key, 0, sizeof(key));
1310 p = &buf[bitset_strscan((struct bitset *)&key.ds_mask,
1311 DOMAINSET_SETSIZE, buf)];
1312 if (p == buf)
1313 return (EINVAL);
1314 if (sscanf(p, ":%d:%d", &policy, &prefer) != 2)
1315 return (EINVAL);
1316 key.ds_policy = policy;
1317 key.ds_prefer = prefer;
1318
1319 /* Domainset_create() validates the policy.*/
1320 dset = domainset_create(&key);
1321 if (dset == NULL)
1322 return (EINVAL);
1323 *(struct domainset **)arg1 = dset;
1324
1325 return (error);
1326 }
1327
1328 /*
1329 * Apply an anonymous mask or a domain to a single thread.
1330 */
1331 static int
1332 _cpuset_setthread(lwpid_t id, cpuset_t *mask, struct domainset *domain)
1333 {
1334 struct setlist cpusets;
1335 struct domainlist domainlist;
1336 struct cpuset *nset;
1337 struct cpuset *set;
1338 struct thread *td;
1339 struct proc *p;
1340 int error;
1341
1342 cpuset_freelist_init(&cpusets, 1);
1343 domainset_freelist_init(&domainlist, domain != NULL);
1344 error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set);
1345 if (error)
1346 goto out;
1347 set = NULL;
1348 thread_lock(td);
1349 error = cpuset_shadow(td->td_cpuset, &nset, mask, domain,
1350 &cpusets, &domainlist);
1351 if (error == 0)
1352 set = cpuset_update_thread(td, nset);
1353 thread_unlock(td);
1354 PROC_UNLOCK(p);
1355 if (set)
1356 cpuset_rel(set);
1357 out:
1358 cpuset_freelist_free(&cpusets);
1359 domainset_freelist_free(&domainlist);
1360 return (error);
1361 }
1362
1363 /*
1364 * Apply an anonymous mask to a single thread.
1365 */
1366 int
1367 cpuset_setthread(lwpid_t id, cpuset_t *mask)
1368 {
1369
1370 return _cpuset_setthread(id, mask, NULL);
1371 }
1372
1373 /*
1374 * Apply new cpumask to the ithread.
1375 */
1376 int
1377 cpuset_setithread(lwpid_t id, int cpu)
1378 {
1379 cpuset_t mask;
1380
1381 CPU_ZERO(&mask);
1382 if (cpu == NOCPU)
1383 CPU_COPY(cpuset_root, &mask);
1384 else
1385 CPU_SET(cpu, &mask);
1386 return _cpuset_setthread(id, &mask, NULL);
1387 }
1388
1389 /*
1390 * Initialize static domainsets after NUMA information is available. This is
1391 * called before memory allocators are initialized.
1392 */
1393 void
1394 domainset_init(void)
1395 {
1396 struct domainset *dset;
1397 int i;
1398
1399 dset = &domainset_roundrobin;
1400 DOMAINSET_COPY(&all_domains, &dset->ds_mask);
1401 dset->ds_policy = DOMAINSET_POLICY_ROUNDROBIN;
1402 dset->ds_prefer = -1;
1403 _domainset_create(dset, NULL);
1404
1405 for (i = 0; i < vm_ndomains; i++) {
1406 dset = &domainset_fixed[i];
1407 DOMAINSET_ZERO(&dset->ds_mask);
1408 DOMAINSET_SET(i, &dset->ds_mask);
1409 dset->ds_policy = DOMAINSET_POLICY_ROUNDROBIN;
1410 _domainset_create(dset, NULL);
1411
1412 dset = &domainset_prefer[i];
1413 DOMAINSET_COPY(&all_domains, &dset->ds_mask);
1414 dset->ds_policy = DOMAINSET_POLICY_PREFER;
1415 dset->ds_prefer = i;
1416 _domainset_create(dset, NULL);
1417 }
1418 }
1419
1420 /*
1421 * Create the domainset for cpuset 0, 1 and cpuset 2.
1422 */
1423 void
1424 domainset_zero(void)
1425 {
1426 struct domainset *dset, *tmp;
1427
1428 mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE);
1429
1430 dset = &domainset0;
1431 DOMAINSET_COPY(&all_domains, &dset->ds_mask);
1432 dset->ds_policy = DOMAINSET_POLICY_FIRSTTOUCH;
1433 dset->ds_prefer = -1;
1434 curthread->td_domain.dr_policy = _domainset_create(dset, NULL);
1435
1436 domainset_copy(dset, &domainset2);
1437 domainset2.ds_policy = DOMAINSET_POLICY_INTERLEAVE;
1438 kernel_object->domain.dr_policy = _domainset_create(&domainset2, NULL);
1439
1440 /* Remove empty domains from the global policies. */
1441 LIST_FOREACH_SAFE(dset, &cpuset_domains, ds_link, tmp)
1442 if (domainset_empty_vm(dset))
1443 LIST_REMOVE(dset, ds_link);
1444 }
1445
1446 /*
1447 * Creates system-wide cpusets and the cpuset for thread0 including three
1448 * sets:
1449 *
1450 * 0 - The root set which should represent all valid processors in the
1451 * system. It is initially created with a mask of all processors
1452 * because we don't know what processors are valid until cpuset_init()
1453 * runs. This set is immutable.
1454 * 1 - The default set which all processes are a member of until changed.
1455 * This allows an administrator to move all threads off of given cpus to
1456 * dedicate them to high priority tasks or save power etc.
1457 * 2 - The kernel set which allows restriction and policy to be applied only
1458 * to kernel threads and the kernel_object.
1459 */
1460 struct cpuset *
1461 cpuset_thread0(void)
1462 {
1463 struct cpuset *set;
1464 int i;
1465 int error __unused;
1466
1467 cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL,
1468 NULL, NULL, UMA_ALIGN_CACHE, 0);
1469 domainset_zone = uma_zcreate("domainset", sizeof(struct domainset),
1470 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
1471
1472 /*
1473 * Create the root system set (0) for the whole machine. Doesn't use
1474 * cpuset_create() due to NULL parent.
1475 */
1476 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1477 CPU_COPY(&all_cpus, &set->cs_mask);
1478 LIST_INIT(&set->cs_children);
1479 LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
1480 set->cs_ref = 1;
1481 set->cs_flags = CPU_SET_ROOT | CPU_SET_RDONLY;
1482 set->cs_domain = &domainset0;
1483 cpuset_zero = set;
1484 cpuset_root = &set->cs_mask;
1485
1486 /*
1487 * Now derive a default (1), modifiable set from that to give out.
1488 */
1489 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1490 error = _cpuset_create(set, cpuset_zero, NULL, NULL, 1);
1491 KASSERT(error == 0, ("Error creating default set: %d\n", error));
1492 cpuset_default = set;
1493 /*
1494 * Create the kernel set (2).
1495 */
1496 set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
1497 error = _cpuset_create(set, cpuset_zero, NULL, NULL, 2);
1498 KASSERT(error == 0, ("Error creating kernel set: %d\n", error));
1499 set->cs_domain = &domainset2;
1500 cpuset_kernel = set;
1501
1502 /*
1503 * Initialize the unit allocator. 0 and 1 are allocated above.
1504 */
1505 cpuset_unr = new_unrhdr(2, INT_MAX, NULL);
1506
1507 /*
1508 * If MD code has not initialized per-domain cpusets, place all
1509 * CPUs in domain 0.
1510 */
1511 for (i = 0; i < MAXMEMDOM; i++)
1512 if (!CPU_EMPTY(&cpuset_domain[i]))
1513 goto domains_set;
1514 CPU_COPY(&all_cpus, &cpuset_domain[0]);
1515 domains_set:
1516
1517 return (cpuset_default);
1518 }
1519
1520 void
1521 cpuset_kernthread(struct thread *td)
1522 {
1523 struct cpuset *set;
1524
1525 thread_lock(td);
1526 set = td->td_cpuset;
1527 td->td_cpuset = cpuset_ref(cpuset_kernel);
1528 thread_unlock(td);
1529 cpuset_rel(set);
1530 }
1531
1532 /*
1533 * Create a cpuset, which would be cpuset_create() but
1534 * mark the new 'set' as root.
1535 *
1536 * We are not going to reparent the td to it. Use cpuset_setproc_update_set()
1537 * for that.
1538 *
1539 * In case of no error, returns the set in *setp locked with a reference.
1540 */
1541 int
1542 cpuset_create_root(struct prison *pr, struct cpuset **setp)
1543 {
1544 struct cpuset *set;
1545 int error;
1546
1547 KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__));
1548 KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__));
1549
1550 error = cpuset_create(setp, pr->pr_cpuset, &pr->pr_cpuset->cs_mask);
1551 if (error)
1552 return (error);
1553
1554 KASSERT(*setp != NULL, ("[%s:%d] cpuset_create returned invalid data",
1555 __func__, __LINE__));
1556
1557 /* Mark the set as root. */
1558 set = *setp;
1559 set->cs_flags |= CPU_SET_ROOT;
1560
1561 return (0);
1562 }
1563
1564 int
1565 cpuset_setproc_update_set(struct proc *p, struct cpuset *set)
1566 {
1567 int error;
1568
1569 KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__));
1570 KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__));
1571
1572 cpuset_ref(set);
1573 error = cpuset_setproc(p->p_pid, set, NULL, NULL);
1574 if (error)
1575 return (error);
1576 cpuset_rel(set);
1577 return (0);
1578 }
1579
1580 #ifndef _SYS_SYSPROTO_H_
1581 struct cpuset_args {
1582 cpusetid_t *setid;
1583 };
1584 #endif
1585 int
1586 sys_cpuset(struct thread *td, struct cpuset_args *uap)
1587 {
1588 struct cpuset *root;
1589 struct cpuset *set;
1590 int error;
1591
1592 thread_lock(td);
1593 root = cpuset_refroot(td->td_cpuset);
1594 thread_unlock(td);
1595 error = cpuset_create(&set, root, &root->cs_mask);
1596 cpuset_rel(root);
1597 if (error)
1598 return (error);
1599 error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id));
1600 if (error == 0)
1601 error = cpuset_setproc(-1, set, NULL, NULL);
1602 cpuset_rel(set);
1603 return (error);
1604 }
1605
1606 #ifndef _SYS_SYSPROTO_H_
1607 struct cpuset_setid_args {
1608 cpuwhich_t which;
1609 id_t id;
1610 cpusetid_t setid;
1611 };
1612 #endif
1613 int
1614 sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap)
1615 {
1616
1617 return (kern_cpuset_setid(td, uap->which, uap->id, uap->setid));
1618 }
1619
1620 int
1621 kern_cpuset_setid(struct thread *td, cpuwhich_t which,
1622 id_t id, cpusetid_t setid)
1623 {
1624 struct cpuset *set;
1625 int error;
1626
1627 /*
1628 * Presently we only support per-process sets.
1629 */
1630 if (which != CPU_WHICH_PID)
1631 return (EINVAL);
1632 set = cpuset_lookup(setid, td);
1633 if (set == NULL)
1634 return (ESRCH);
1635 error = cpuset_setproc(id, set, NULL, NULL);
1636 cpuset_rel(set);
1637 return (error);
1638 }
1639
1640 #ifndef _SYS_SYSPROTO_H_
1641 struct cpuset_getid_args {
1642 cpulevel_t level;
1643 cpuwhich_t which;
1644 id_t id;
1645 cpusetid_t *setid;
1646 };
1647 #endif
1648 int
1649 sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap)
1650 {
1651
1652 return (kern_cpuset_getid(td, uap->level, uap->which, uap->id,
1653 uap->setid));
1654 }
1655
1656 int
1657 kern_cpuset_getid(struct thread *td, cpulevel_t level, cpuwhich_t which,
1658 id_t id, cpusetid_t *setid)
1659 {
1660 struct cpuset *nset;
1661 struct cpuset *set;
1662 struct thread *ttd;
1663 struct proc *p;
1664 cpusetid_t tmpid;
1665 int error;
1666
1667 if (level == CPU_LEVEL_WHICH && which != CPU_WHICH_CPUSET)
1668 return (EINVAL);
1669 error = cpuset_which(which, id, &p, &ttd, &set);
1670 if (error)
1671 return (error);
1672 switch (which) {
1673 case CPU_WHICH_TID:
1674 case CPU_WHICH_PID:
1675 thread_lock(ttd);
1676 set = cpuset_refbase(ttd->td_cpuset);
1677 thread_unlock(ttd);
1678 PROC_UNLOCK(p);
1679 break;
1680 case CPU_WHICH_CPUSET:
1681 case CPU_WHICH_JAIL:
1682 break;
1683 case CPU_WHICH_IRQ:
1684 case CPU_WHICH_DOMAIN:
1685 return (EINVAL);
1686 }
1687 switch (level) {
1688 case CPU_LEVEL_ROOT:
1689 nset = cpuset_refroot(set);
1690 cpuset_rel(set);
1691 set = nset;
1692 break;
1693 case CPU_LEVEL_CPUSET:
1694 break;
1695 case CPU_LEVEL_WHICH:
1696 break;
1697 }
1698 tmpid = set->cs_id;
1699 cpuset_rel(set);
1700 if (error == 0)
1701 error = copyout(&tmpid, setid, sizeof(tmpid));
1702
1703 return (error);
1704 }
1705
1706 #ifndef _SYS_SYSPROTO_H_
1707 struct cpuset_getaffinity_args {
1708 cpulevel_t level;
1709 cpuwhich_t which;
1710 id_t id;
1711 size_t cpusetsize;
1712 cpuset_t *mask;
1713 };
1714 #endif
1715 int
1716 sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap)
1717 {
1718
1719 return (kern_cpuset_getaffinity(td, uap->level, uap->which,
1720 uap->id, uap->cpusetsize, uap->mask));
1721 }
1722
1723 int
1724 kern_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1725 id_t id, size_t cpusetsize, cpuset_t *maskp)
1726 {
1727 struct thread *ttd;
1728 struct cpuset *nset;
1729 struct cpuset *set;
1730 struct proc *p;
1731 cpuset_t *mask;
1732 int error;
1733 size_t size;
1734
1735 if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1736 return (ERANGE);
1737 /* In Capability mode, you can only get your own CPU set. */
1738 if (IN_CAPABILITY_MODE(td)) {
1739 if (level != CPU_LEVEL_WHICH)
1740 return (ECAPMODE);
1741 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1742 return (ECAPMODE);
1743 if (id != -1)
1744 return (ECAPMODE);
1745 }
1746 size = cpusetsize;
1747 mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO);
1748 error = cpuset_which(which, id, &p, &ttd, &set);
1749 if (error)
1750 goto out;
1751 switch (level) {
1752 case CPU_LEVEL_ROOT:
1753 case CPU_LEVEL_CPUSET:
1754 switch (which) {
1755 case CPU_WHICH_TID:
1756 case CPU_WHICH_PID:
1757 thread_lock(ttd);
1758 set = cpuset_ref(ttd->td_cpuset);
1759 thread_unlock(ttd);
1760 break;
1761 case CPU_WHICH_CPUSET:
1762 case CPU_WHICH_JAIL:
1763 break;
1764 case CPU_WHICH_IRQ:
1765 case CPU_WHICH_INTRHANDLER:
1766 case CPU_WHICH_ITHREAD:
1767 case CPU_WHICH_DOMAIN:
1768 error = EINVAL;
1769 goto out;
1770 }
1771 if (level == CPU_LEVEL_ROOT)
1772 nset = cpuset_refroot(set);
1773 else
1774 nset = cpuset_refbase(set);
1775 CPU_COPY(&nset->cs_mask, mask);
1776 cpuset_rel(nset);
1777 break;
1778 case CPU_LEVEL_WHICH:
1779 switch (which) {
1780 case CPU_WHICH_TID:
1781 thread_lock(ttd);
1782 CPU_COPY(&ttd->td_cpuset->cs_mask, mask);
1783 thread_unlock(ttd);
1784 break;
1785 case CPU_WHICH_PID:
1786 FOREACH_THREAD_IN_PROC(p, ttd) {
1787 thread_lock(ttd);
1788 CPU_OR(mask, &ttd->td_cpuset->cs_mask);
1789 thread_unlock(ttd);
1790 }
1791 break;
1792 case CPU_WHICH_CPUSET:
1793 case CPU_WHICH_JAIL:
1794 CPU_COPY(&set->cs_mask, mask);
1795 break;
1796 case CPU_WHICH_IRQ:
1797 case CPU_WHICH_INTRHANDLER:
1798 case CPU_WHICH_ITHREAD:
1799 error = intr_getaffinity(id, which, mask);
1800 break;
1801 case CPU_WHICH_DOMAIN:
1802 if (id < 0 || id >= MAXMEMDOM)
1803 error = ESRCH;
1804 else
1805 CPU_COPY(&cpuset_domain[id], mask);
1806 break;
1807 }
1808 break;
1809 default:
1810 error = EINVAL;
1811 break;
1812 }
1813 if (set)
1814 cpuset_rel(set);
1815 if (p)
1816 PROC_UNLOCK(p);
1817 if (error == 0)
1818 error = copyout(mask, maskp, size);
1819 out:
1820 free(mask, M_TEMP);
1821 return (error);
1822 }
1823
1824 #ifndef _SYS_SYSPROTO_H_
1825 struct cpuset_setaffinity_args {
1826 cpulevel_t level;
1827 cpuwhich_t which;
1828 id_t id;
1829 size_t cpusetsize;
1830 const cpuset_t *mask;
1831 };
1832 #endif
1833 int
1834 sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap)
1835 {
1836
1837 return (kern_cpuset_setaffinity(td, uap->level, uap->which,
1838 uap->id, uap->cpusetsize, uap->mask));
1839 }
1840
1841 int
1842 kern_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1843 id_t id, size_t cpusetsize, const cpuset_t *maskp)
1844 {
1845 struct cpuset *nset;
1846 struct cpuset *set;
1847 struct thread *ttd;
1848 struct proc *p;
1849 cpuset_t *mask;
1850 int error;
1851
1852 if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1853 return (ERANGE);
1854 /* In Capability mode, you can only set your own CPU set. */
1855 if (IN_CAPABILITY_MODE(td)) {
1856 if (level != CPU_LEVEL_WHICH)
1857 return (ECAPMODE);
1858 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1859 return (ECAPMODE);
1860 if (id != -1)
1861 return (ECAPMODE);
1862 }
1863 mask = malloc(cpusetsize, M_TEMP, M_WAITOK | M_ZERO);
1864 error = copyin(maskp, mask, cpusetsize);
1865 if (error)
1866 goto out;
1867 /*
1868 * Verify that no high bits are set.
1869 */
1870 if (cpusetsize > sizeof(cpuset_t)) {
1871 char *end;
1872 char *cp;
1873
1874 end = cp = (char *)&mask->__bits;
1875 end += cpusetsize;
1876 cp += sizeof(cpuset_t);
1877 while (cp != end)
1878 if (*cp++ != 0) {
1879 error = EINVAL;
1880 goto out;
1881 }
1882
1883 }
1884 switch (level) {
1885 case CPU_LEVEL_ROOT:
1886 case CPU_LEVEL_CPUSET:
1887 error = cpuset_which(which, id, &p, &ttd, &set);
1888 if (error)
1889 break;
1890 switch (which) {
1891 case CPU_WHICH_TID:
1892 case CPU_WHICH_PID:
1893 thread_lock(ttd);
1894 set = cpuset_ref(ttd->td_cpuset);
1895 thread_unlock(ttd);
1896 PROC_UNLOCK(p);
1897 break;
1898 case CPU_WHICH_CPUSET:
1899 case CPU_WHICH_JAIL:
1900 break;
1901 case CPU_WHICH_IRQ:
1902 case CPU_WHICH_INTRHANDLER:
1903 case CPU_WHICH_ITHREAD:
1904 case CPU_WHICH_DOMAIN:
1905 error = EINVAL;
1906 goto out;
1907 }
1908 if (level == CPU_LEVEL_ROOT)
1909 nset = cpuset_refroot(set);
1910 else
1911 nset = cpuset_refbase(set);
1912 error = cpuset_modify(nset, mask);
1913 cpuset_rel(nset);
1914 cpuset_rel(set);
1915 break;
1916 case CPU_LEVEL_WHICH:
1917 switch (which) {
1918 case CPU_WHICH_TID:
1919 error = cpuset_setthread(id, mask);
1920 break;
1921 case CPU_WHICH_PID:
1922 error = cpuset_setproc(id, NULL, mask, NULL);
1923 break;
1924 case CPU_WHICH_CPUSET:
1925 case CPU_WHICH_JAIL:
1926 error = cpuset_which(which, id, &p, &ttd, &set);
1927 if (error == 0) {
1928 error = cpuset_modify(set, mask);
1929 cpuset_rel(set);
1930 }
1931 break;
1932 case CPU_WHICH_IRQ:
1933 case CPU_WHICH_INTRHANDLER:
1934 case CPU_WHICH_ITHREAD:
1935 error = intr_setaffinity(id, which, mask);
1936 break;
1937 default:
1938 error = EINVAL;
1939 break;
1940 }
1941 break;
1942 default:
1943 error = EINVAL;
1944 break;
1945 }
1946 out:
1947 free(mask, M_TEMP);
1948 return (error);
1949 }
1950
1951 #ifndef _SYS_SYSPROTO_H_
1952 struct cpuset_getdomain_args {
1953 cpulevel_t level;
1954 cpuwhich_t which;
1955 id_t id;
1956 size_t domainsetsize;
1957 domainset_t *mask;
1958 int *policy;
1959 };
1960 #endif
1961 int
1962 sys_cpuset_getdomain(struct thread *td, struct cpuset_getdomain_args *uap)
1963 {
1964
1965 return (kern_cpuset_getdomain(td, uap->level, uap->which,
1966 uap->id, uap->domainsetsize, uap->mask, uap->policy));
1967 }
1968
1969 int
1970 kern_cpuset_getdomain(struct thread *td, cpulevel_t level, cpuwhich_t which,
1971 id_t id, size_t domainsetsize, domainset_t *maskp, int *policyp)
1972 {
1973 struct domainset outset;
1974 struct thread *ttd;
1975 struct cpuset *nset;
1976 struct cpuset *set;
1977 struct domainset *dset;
1978 struct proc *p;
1979 domainset_t *mask;
1980 int error;
1981
1982 if (domainsetsize < sizeof(domainset_t) ||
1983 domainsetsize > DOMAINSET_MAXSIZE / NBBY)
1984 return (ERANGE);
1985 /* In Capability mode, you can only get your own domain set. */
1986 if (IN_CAPABILITY_MODE(td)) {
1987 if (level != CPU_LEVEL_WHICH)
1988 return (ECAPMODE);
1989 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1990 return (ECAPMODE);
1991 if (id != -1)
1992 return (ECAPMODE);
1993 }
1994 mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO);
1995 bzero(&outset, sizeof(outset));
1996 error = cpuset_which(which, id, &p, &ttd, &set);
1997 if (error)
1998 goto out;
1999 switch (level) {
2000 case CPU_LEVEL_ROOT:
2001 case CPU_LEVEL_CPUSET:
2002 switch (which) {
2003 case CPU_WHICH_TID:
2004 case CPU_WHICH_PID:
2005 thread_lock(ttd);
2006 set = cpuset_ref(ttd->td_cpuset);
2007 thread_unlock(ttd);
2008 break;
2009 case CPU_WHICH_CPUSET:
2010 case CPU_WHICH_JAIL:
2011 break;
2012 case CPU_WHICH_IRQ:
2013 case CPU_WHICH_INTRHANDLER:
2014 case CPU_WHICH_ITHREAD:
2015 case CPU_WHICH_DOMAIN:
2016 error = EINVAL;
2017 goto out;
2018 }
2019 if (level == CPU_LEVEL_ROOT)
2020 nset = cpuset_refroot(set);
2021 else
2022 nset = cpuset_refbase(set);
2023 domainset_copy(nset->cs_domain, &outset);
2024 cpuset_rel(nset);
2025 break;
2026 case CPU_LEVEL_WHICH:
2027 switch (which) {
2028 case CPU_WHICH_TID:
2029 thread_lock(ttd);
2030 domainset_copy(ttd->td_cpuset->cs_domain, &outset);
2031 thread_unlock(ttd);
2032 break;
2033 case CPU_WHICH_PID:
2034 FOREACH_THREAD_IN_PROC(p, ttd) {
2035 thread_lock(ttd);
2036 dset = ttd->td_cpuset->cs_domain;
2037 /* Show all domains in the proc. */
2038 DOMAINSET_OR(&outset.ds_mask, &dset->ds_mask);
2039 /* Last policy wins. */
2040 outset.ds_policy = dset->ds_policy;
2041 outset.ds_prefer = dset->ds_prefer;
2042 thread_unlock(ttd);
2043 }
2044 break;
2045 case CPU_WHICH_CPUSET:
2046 case CPU_WHICH_JAIL:
2047 domainset_copy(set->cs_domain, &outset);
2048 break;
2049 case CPU_WHICH_IRQ:
2050 case CPU_WHICH_INTRHANDLER:
2051 case CPU_WHICH_ITHREAD:
2052 case CPU_WHICH_DOMAIN:
2053 error = EINVAL;
2054 break;
2055 }
2056 break;
2057 default:
2058 error = EINVAL;
2059 break;
2060 }
2061 if (set)
2062 cpuset_rel(set);
2063 if (p)
2064 PROC_UNLOCK(p);
2065 /*
2066 * Translate prefer into a set containing only the preferred domain,
2067 * not the entire fallback set.
2068 */
2069 if (outset.ds_policy == DOMAINSET_POLICY_PREFER) {
2070 DOMAINSET_ZERO(&outset.ds_mask);
2071 DOMAINSET_SET(outset.ds_prefer, &outset.ds_mask);
2072 }
2073 DOMAINSET_COPY(&outset.ds_mask, mask);
2074 if (error == 0)
2075 error = copyout(mask, maskp, domainsetsize);
2076 if (error == 0)
2077 if (suword32(policyp, outset.ds_policy) != 0)
2078 error = EFAULT;
2079 out:
2080 free(mask, M_TEMP);
2081 return (error);
2082 }
2083
2084 #ifndef _SYS_SYSPROTO_H_
2085 struct cpuset_setdomain_args {
2086 cpulevel_t level;
2087 cpuwhich_t which;
2088 id_t id;
2089 size_t domainsetsize;
2090 domainset_t *mask;
2091 int policy;
2092 };
2093 #endif
2094 int
2095 sys_cpuset_setdomain(struct thread *td, struct cpuset_setdomain_args *uap)
2096 {
2097
2098 return (kern_cpuset_setdomain(td, uap->level, uap->which,
2099 uap->id, uap->domainsetsize, uap->mask, uap->policy));
2100 }
2101
2102 int
2103 kern_cpuset_setdomain(struct thread *td, cpulevel_t level, cpuwhich_t which,
2104 id_t id, size_t domainsetsize, const domainset_t *maskp, int policy)
2105 {
2106 struct cpuset *nset;
2107 struct cpuset *set;
2108 struct thread *ttd;
2109 struct proc *p;
2110 struct domainset domain;
2111 domainset_t *mask;
2112 int error;
2113
2114 if (domainsetsize < sizeof(domainset_t) ||
2115 domainsetsize > DOMAINSET_MAXSIZE / NBBY)
2116 return (ERANGE);
2117 if (policy <= DOMAINSET_POLICY_INVALID ||
2118 policy > DOMAINSET_POLICY_MAX)
2119 return (EINVAL);
2120 /* In Capability mode, you can only set your own CPU set. */
2121 if (IN_CAPABILITY_MODE(td)) {
2122 if (level != CPU_LEVEL_WHICH)
2123 return (ECAPMODE);
2124 if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
2125 return (ECAPMODE);
2126 if (id != -1)
2127 return (ECAPMODE);
2128 }
2129 memset(&domain, 0, sizeof(domain));
2130 mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO);
2131 error = copyin(maskp, mask, domainsetsize);
2132 if (error)
2133 goto out;
2134 /*
2135 * Verify that no high bits are set.
2136 */
2137 if (domainsetsize > sizeof(domainset_t)) {
2138 char *end;
2139 char *cp;
2140
2141 end = cp = (char *)&mask->__bits;
2142 end += domainsetsize;
2143 cp += sizeof(domainset_t);
2144 while (cp != end)
2145 if (*cp++ != 0) {
2146 error = EINVAL;
2147 goto out;
2148 }
2149
2150 }
2151 DOMAINSET_COPY(mask, &domain.ds_mask);
2152 domain.ds_policy = policy;
2153
2154 /* Translate preferred policy into a mask and fallback. */
2155 if (policy == DOMAINSET_POLICY_PREFER) {
2156 /* Only support a single preferred domain. */
2157 if (DOMAINSET_COUNT(&domain.ds_mask) != 1) {
2158 error = EINVAL;
2159 goto out;
2160 }
2161 domain.ds_prefer = DOMAINSET_FFS(&domain.ds_mask) - 1;
2162 /* This will be constrained by domainset_shadow(). */
2163 DOMAINSET_FILL(&domain.ds_mask);
2164 }
2165
2166 /*
2167 * When given an impossible policy, fall back to interleaving
2168 * across all domains
2169 */
2170 if (domainset_empty_vm(&domain))
2171 domainset_copy(&domainset2, &domain);
2172
2173 switch (level) {
2174 case CPU_LEVEL_ROOT:
2175 case CPU_LEVEL_CPUSET:
2176 error = cpuset_which(which, id, &p, &ttd, &set);
2177 if (error)
2178 break;
2179 switch (which) {
2180 case CPU_WHICH_TID:
2181 case CPU_WHICH_PID:
2182 thread_lock(ttd);
2183 set = cpuset_ref(ttd->td_cpuset);
2184 thread_unlock(ttd);
2185 PROC_UNLOCK(p);
2186 break;
2187 case CPU_WHICH_CPUSET:
2188 case CPU_WHICH_JAIL:
2189 break;
2190 case CPU_WHICH_IRQ:
2191 case CPU_WHICH_INTRHANDLER:
2192 case CPU_WHICH_ITHREAD:
2193 case CPU_WHICH_DOMAIN:
2194 error = EINVAL;
2195 goto out;
2196 }
2197 if (level == CPU_LEVEL_ROOT)
2198 nset = cpuset_refroot(set);
2199 else
2200 nset = cpuset_refbase(set);
2201 error = cpuset_modify_domain(nset, &domain);
2202 cpuset_rel(nset);
2203 cpuset_rel(set);
2204 break;
2205 case CPU_LEVEL_WHICH:
2206 switch (which) {
2207 case CPU_WHICH_TID:
2208 error = _cpuset_setthread(id, NULL, &domain);
2209 break;
2210 case CPU_WHICH_PID:
2211 error = cpuset_setproc(id, NULL, NULL, &domain);
2212 break;
2213 case CPU_WHICH_CPUSET:
2214 case CPU_WHICH_JAIL:
2215 error = cpuset_which(which, id, &p, &ttd, &set);
2216 if (error == 0) {
2217 error = cpuset_modify_domain(set, &domain);
2218 cpuset_rel(set);
2219 }
2220 break;
2221 case CPU_WHICH_IRQ:
2222 case CPU_WHICH_INTRHANDLER:
2223 case CPU_WHICH_ITHREAD:
2224 default:
2225 error = EINVAL;
2226 break;
2227 }
2228 break;
2229 default:
2230 error = EINVAL;
2231 break;
2232 }
2233 out:
2234 free(mask, M_TEMP);
2235 return (error);
2236 }
2237
2238 #ifdef DDB
2239
2240 static void
2241 ddb_display_bitset(const struct bitset *set, int size)
2242 {
2243 int bit, once;
2244
2245 for (once = 0, bit = 0; bit < size; bit++) {
2246 if (CPU_ISSET(bit, set)) {
2247 if (once == 0) {
2248 db_printf("%d", bit);
2249 once = 1;
2250 } else
2251 db_printf(",%d", bit);
2252 }
2253 }
2254 if (once == 0)
2255 db_printf("<none>");
2256 }
2257
2258 void
2259 ddb_display_cpuset(const cpuset_t *set)
2260 {
2261 ddb_display_bitset((const struct bitset *)set, CPU_SETSIZE);
2262 }
2263
2264 static void
2265 ddb_display_domainset(const domainset_t *set)
2266 {
2267 ddb_display_bitset((const struct bitset *)set, DOMAINSET_SETSIZE);
2268 }
2269
2270 DB_SHOW_COMMAND(cpusets, db_show_cpusets)
2271 {
2272 struct cpuset *set;
2273
2274 LIST_FOREACH(set, &cpuset_ids, cs_link) {
2275 db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n",
2276 set, set->cs_id, set->cs_ref, set->cs_flags,
2277 (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0);
2278 db_printf(" cpu mask=");
2279 ddb_display_cpuset(&set->cs_mask);
2280 db_printf("\n");
2281 db_printf(" domain policy %d prefer %d mask=",
2282 set->cs_domain->ds_policy, set->cs_domain->ds_prefer);
2283 ddb_display_domainset(&set->cs_domain->ds_mask);
2284 db_printf("\n");
2285 if (db_pager_quit)
2286 break;
2287 }
2288 }
2289
2290 DB_SHOW_COMMAND(domainsets, db_show_domainsets)
2291 {
2292 struct domainset *set;
2293
2294 LIST_FOREACH(set, &cpuset_domains, ds_link) {
2295 db_printf("set=%p policy %d prefer %d cnt %d\n",
2296 set, set->ds_policy, set->ds_prefer, set->ds_cnt);
2297 db_printf(" mask =");
2298 ddb_display_domainset(&set->ds_mask);
2299 db_printf("\n");
2300 }
2301 }
2302 #endif /* DDB */
Cache object: 2cbcb5f8ea12fa7c3838c7e92fc2fcbb
|