xref: /freebsd/sys/kern/kern_cpuset.c (revision d06955f9bdb1416d9196043ed781f9b36dae9adc)
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("$FreeBSD$");
35 
36 #include "opt_ddb.h"
37 
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/sysproto.h>
41 #include <sys/jail.h>
42 #include <sys/kernel.h>
43 #include <sys/lock.h>
44 #include <sys/malloc.h>
45 #include <sys/mutex.h>
46 #include <sys/priv.h>
47 #include <sys/proc.h>
48 #include <sys/refcount.h>
49 #include <sys/sched.h>
50 #include <sys/smp.h>
51 #include <sys/syscallsubr.h>
52 #include <sys/capsicum.h>
53 #include <sys/cpuset.h>
54 #include <sys/sx.h>
55 #include <sys/queue.h>
56 #include <sys/libkern.h>
57 #include <sys/limits.h>
58 #include <sys/bus.h>
59 #include <sys/interrupt.h>
60 
61 #include <vm/uma.h>
62 #include <vm/vm.h>
63 #include <vm/vm_page.h>
64 #include <vm/vm_param.h>
65 
66 #ifdef DDB
67 #include <ddb/ddb.h>
68 #endif /* DDB */
69 
70 /*
71  * cpusets provide a mechanism for creating and manipulating sets of
72  * processors for the purpose of constraining the scheduling of threads to
73  * specific processors.
74  *
75  * Each process belongs to an identified set, by default this is set 1.  Each
76  * thread may further restrict the cpus it may run on to a subset of this
77  * named set.  This creates an anonymous set which other threads and processes
78  * may not join by number.
79  *
80  * The named set is referred to herein as the 'base' set to avoid ambiguity.
81  * This set is usually a child of a 'root' set while the anonymous set may
82  * simply be referred to as a mask.  In the syscall api these are referred to
83  * as the ROOT, CPUSET, and MASK levels where CPUSET is called 'base' here.
84  *
85  * Threads inherit their set from their creator whether it be anonymous or
86  * not.  This means that anonymous sets are immutable because they may be
87  * shared.  To modify an anonymous set a new set is created with the desired
88  * mask and the same parent as the existing anonymous set.  This gives the
89  * illusion of each thread having a private mask.
90  *
91  * Via the syscall apis a user may ask to retrieve or modify the root, base,
92  * or mask that is discovered via a pid, tid, or setid.  Modifying a set
93  * modifies all numbered and anonymous child sets to comply with the new mask.
94  * Modifying a pid or tid's mask applies only to that tid but must still
95  * exist within the assigned parent set.
96  *
97  * A thread may not be assigned to a group separate from other threads in
98  * the process.  This is to remove ambiguity when the setid is queried with
99  * a pid argument.  There is no other technical limitation.
100  *
101  * This somewhat complex arrangement is intended to make it easy for
102  * applications to query available processors and bind their threads to
103  * specific processors while also allowing administrators to dynamically
104  * reprovision by changing sets which apply to groups of processes.
105  *
106  * A simple application should not concern itself with sets at all and
107  * rather apply masks to its own threads via CPU_WHICH_TID and a -1 id
108  * meaning 'curthread'.  It may query available cpus for that tid with a
109  * getaffinity call using (CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, ...).
110  */
111 static uma_zone_t cpuset_zone;
112 static struct mtx cpuset_lock;
113 static struct setlist cpuset_ids;
114 static struct unrhdr *cpuset_unr;
115 static struct cpuset *cpuset_zero, *cpuset_default;
116 
117 /* Return the size of cpuset_t at the kernel level */
118 SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD | CTLFLAG_CAPRD,
119     SYSCTL_NULL_INT_PTR, sizeof(cpuset_t), "sizeof(cpuset_t)");
120 
121 cpuset_t *cpuset_root;
122 cpuset_t cpuset_domain[MAXMEMDOM];
123 
124 /*
125  * Acquire a reference to a cpuset, all pointers must be tracked with refs.
126  */
127 struct cpuset *
128 cpuset_ref(struct cpuset *set)
129 {
130 
131 	refcount_acquire(&set->cs_ref);
132 	return (set);
133 }
134 
135 /*
136  * Walks up the tree from 'set' to find the root.  Returns the root
137  * referenced.
138  */
139 static struct cpuset *
140 cpuset_refroot(struct cpuset *set)
141 {
142 
143 	for (; set->cs_parent != NULL; set = set->cs_parent)
144 		if (set->cs_flags & CPU_SET_ROOT)
145 			break;
146 	cpuset_ref(set);
147 
148 	return (set);
149 }
150 
151 /*
152  * Find the first non-anonymous set starting from 'set'.  Returns this set
153  * referenced.  May return the passed in set with an extra ref if it is
154  * not anonymous.
155  */
156 static struct cpuset *
157 cpuset_refbase(struct cpuset *set)
158 {
159 
160 	if (set->cs_id == CPUSET_INVALID)
161 		set = set->cs_parent;
162 	cpuset_ref(set);
163 
164 	return (set);
165 }
166 
167 /*
168  * Release a reference in a context where it is safe to allocate.
169  */
170 void
171 cpuset_rel(struct cpuset *set)
172 {
173 	cpusetid_t id;
174 
175 	if (refcount_release(&set->cs_ref) == 0)
176 		return;
177 	mtx_lock_spin(&cpuset_lock);
178 	LIST_REMOVE(set, cs_siblings);
179 	id = set->cs_id;
180 	if (id != CPUSET_INVALID)
181 		LIST_REMOVE(set, cs_link);
182 	mtx_unlock_spin(&cpuset_lock);
183 	cpuset_rel(set->cs_parent);
184 	uma_zfree(cpuset_zone, set);
185 	if (id != CPUSET_INVALID)
186 		free_unr(cpuset_unr, id);
187 }
188 
189 /*
190  * Deferred release must be used when in a context that is not safe to
191  * allocate/free.  This places any unreferenced sets on the list 'head'.
192  */
193 static void
194 cpuset_rel_defer(struct setlist *head, struct cpuset *set)
195 {
196 
197 	if (refcount_release(&set->cs_ref) == 0)
198 		return;
199 	mtx_lock_spin(&cpuset_lock);
200 	LIST_REMOVE(set, cs_siblings);
201 	if (set->cs_id != CPUSET_INVALID)
202 		LIST_REMOVE(set, cs_link);
203 	LIST_INSERT_HEAD(head, set, cs_link);
204 	mtx_unlock_spin(&cpuset_lock);
205 }
206 
207 /*
208  * Complete a deferred release.  Removes the set from the list provided to
209  * cpuset_rel_defer.
210  */
211 static void
212 cpuset_rel_complete(struct cpuset *set)
213 {
214 	LIST_REMOVE(set, cs_link);
215 	cpuset_rel(set->cs_parent);
216 	uma_zfree(cpuset_zone, set);
217 }
218 
219 /*
220  * Find a set based on an id.  Returns it with a ref.
221  */
222 static struct cpuset *
223 cpuset_lookup(cpusetid_t setid, struct thread *td)
224 {
225 	struct cpuset *set;
226 
227 	if (setid == CPUSET_INVALID)
228 		return (NULL);
229 	mtx_lock_spin(&cpuset_lock);
230 	LIST_FOREACH(set, &cpuset_ids, cs_link)
231 		if (set->cs_id == setid)
232 			break;
233 	if (set)
234 		cpuset_ref(set);
235 	mtx_unlock_spin(&cpuset_lock);
236 
237 	KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__));
238 	if (set != NULL && jailed(td->td_ucred)) {
239 		struct cpuset *jset, *tset;
240 
241 		jset = td->td_ucred->cr_prison->pr_cpuset;
242 		for (tset = set; tset != NULL; tset = tset->cs_parent)
243 			if (tset == jset)
244 				break;
245 		if (tset == NULL) {
246 			cpuset_rel(set);
247 			set = NULL;
248 		}
249 	}
250 
251 	return (set);
252 }
253 
254 /*
255  * Create a set in the space provided in 'set' with the provided parameters.
256  * The set is returned with a single ref.  May return EDEADLK if the set
257  * will have no valid cpu based on restrictions from the parent.
258  */
259 static int
260 _cpuset_create(struct cpuset *set, struct cpuset *parent, const cpuset_t *mask,
261     cpusetid_t id)
262 {
263 
264 	if (!CPU_OVERLAP(&parent->cs_mask, mask))
265 		return (EDEADLK);
266 	CPU_COPY(mask, &set->cs_mask);
267 	LIST_INIT(&set->cs_children);
268 	refcount_init(&set->cs_ref, 1);
269 	set->cs_flags = 0;
270 	mtx_lock_spin(&cpuset_lock);
271 	CPU_AND(&set->cs_mask, &parent->cs_mask);
272 	set->cs_id = id;
273 	set->cs_parent = cpuset_ref(parent);
274 	LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings);
275 	if (set->cs_id != CPUSET_INVALID)
276 		LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
277 	mtx_unlock_spin(&cpuset_lock);
278 
279 	return (0);
280 }
281 
282 /*
283  * Create a new non-anonymous set with the requested parent and mask.  May
284  * return failures if the mask is invalid or a new number can not be
285  * allocated.
286  */
287 static int
288 cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask)
289 {
290 	struct cpuset *set;
291 	cpusetid_t id;
292 	int error;
293 
294 	id = alloc_unr(cpuset_unr);
295 	if (id == -1)
296 		return (ENFILE);
297 	*setp = set = uma_zalloc(cpuset_zone, M_WAITOK);
298 	error = _cpuset_create(set, parent, mask, id);
299 	if (error == 0)
300 		return (0);
301 	free_unr(cpuset_unr, id);
302 	uma_zfree(cpuset_zone, set);
303 
304 	return (error);
305 }
306 
307 /*
308  * Recursively check for errors that would occur from applying mask to
309  * the tree of sets starting at 'set'.  Checks for sets that would become
310  * empty as well as RDONLY flags.
311  */
312 static int
313 cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int check_mask)
314 {
315 	struct cpuset *nset;
316 	cpuset_t newmask;
317 	int error;
318 
319 	mtx_assert(&cpuset_lock, MA_OWNED);
320 	if (set->cs_flags & CPU_SET_RDONLY)
321 		return (EPERM);
322 	if (check_mask) {
323 		if (!CPU_OVERLAP(&set->cs_mask, mask))
324 			return (EDEADLK);
325 		CPU_COPY(&set->cs_mask, &newmask);
326 		CPU_AND(&newmask, mask);
327 	} else
328 		CPU_COPY(mask, &newmask);
329 	error = 0;
330 	LIST_FOREACH(nset, &set->cs_children, cs_siblings)
331 		if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0)
332 			break;
333 	return (error);
334 }
335 
336 /*
337  * Applies the mask 'mask' without checking for empty sets or permissions.
338  */
339 static void
340 cpuset_update(struct cpuset *set, cpuset_t *mask)
341 {
342 	struct cpuset *nset;
343 
344 	mtx_assert(&cpuset_lock, MA_OWNED);
345 	CPU_AND(&set->cs_mask, mask);
346 	LIST_FOREACH(nset, &set->cs_children, cs_siblings)
347 		cpuset_update(nset, &set->cs_mask);
348 
349 	return;
350 }
351 
352 /*
353  * Modify the set 'set' to use a copy of the mask provided.  Apply this new
354  * mask to restrict all children in the tree.  Checks for validity before
355  * applying the changes.
356  */
357 static int
358 cpuset_modify(struct cpuset *set, cpuset_t *mask)
359 {
360 	struct cpuset *root;
361 	int error;
362 
363 	error = priv_check(curthread, PRIV_SCHED_CPUSET);
364 	if (error)
365 		return (error);
366 	/*
367 	 * In case we are called from within the jail
368 	 * we do not allow modifying the dedicated root
369 	 * cpuset of the jail but may still allow to
370 	 * change child sets.
371 	 */
372 	if (jailed(curthread->td_ucred) &&
373 	    set->cs_flags & CPU_SET_ROOT)
374 		return (EPERM);
375 	/*
376 	 * Verify that we have access to this set of
377 	 * cpus.
378 	 */
379 	root = set->cs_parent;
380 	if (root && !CPU_SUBSET(&root->cs_mask, mask))
381 		return (EINVAL);
382 	mtx_lock_spin(&cpuset_lock);
383 	error = cpuset_testupdate(set, mask, 0);
384 	if (error)
385 		goto out;
386 	CPU_COPY(mask, &set->cs_mask);
387 	cpuset_update(set, mask);
388 out:
389 	mtx_unlock_spin(&cpuset_lock);
390 
391 	return (error);
392 }
393 
394 /*
395  * Resolve the 'which' parameter of several cpuset apis.
396  *
397  * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid.  Also
398  * checks for permission via p_cansched().
399  *
400  * For WHICH_SET returns a valid set with a new reference.
401  *
402  * -1 may be supplied for any argument to mean the current proc/thread or
403  * the base set of the current thread.  May fail with ESRCH/EPERM.
404  */
405 int
406 cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp,
407     struct cpuset **setp)
408 {
409 	struct cpuset *set;
410 	struct thread *td;
411 	struct proc *p;
412 	int error;
413 
414 	*pp = p = NULL;
415 	*tdp = td = NULL;
416 	*setp = set = NULL;
417 	switch (which) {
418 	case CPU_WHICH_PID:
419 		if (id == -1) {
420 			PROC_LOCK(curproc);
421 			p = curproc;
422 			break;
423 		}
424 		if ((p = pfind(id)) == NULL)
425 			return (ESRCH);
426 		break;
427 	case CPU_WHICH_TID:
428 		if (id == -1) {
429 			PROC_LOCK(curproc);
430 			p = curproc;
431 			td = curthread;
432 			break;
433 		}
434 		td = tdfind(id, -1);
435 		if (td == NULL)
436 			return (ESRCH);
437 		p = td->td_proc;
438 		break;
439 	case CPU_WHICH_CPUSET:
440 		if (id == -1) {
441 			thread_lock(curthread);
442 			set = cpuset_refbase(curthread->td_cpuset);
443 			thread_unlock(curthread);
444 		} else
445 			set = cpuset_lookup(id, curthread);
446 		if (set) {
447 			*setp = set;
448 			return (0);
449 		}
450 		return (ESRCH);
451 	case CPU_WHICH_JAIL:
452 	{
453 		/* Find `set' for prison with given id. */
454 		struct prison *pr;
455 
456 		sx_slock(&allprison_lock);
457 		pr = prison_find_child(curthread->td_ucred->cr_prison, id);
458 		sx_sunlock(&allprison_lock);
459 		if (pr == NULL)
460 			return (ESRCH);
461 		cpuset_ref(pr->pr_cpuset);
462 		*setp = pr->pr_cpuset;
463 		mtx_unlock(&pr->pr_mtx);
464 		return (0);
465 	}
466 	case CPU_WHICH_IRQ:
467 	case CPU_WHICH_DOMAIN:
468 		return (0);
469 	default:
470 		return (EINVAL);
471 	}
472 	error = p_cansched(curthread, p);
473 	if (error) {
474 		PROC_UNLOCK(p);
475 		return (error);
476 	}
477 	if (td == NULL)
478 		td = FIRST_THREAD_IN_PROC(p);
479 	*pp = p;
480 	*tdp = td;
481 	return (0);
482 }
483 
484 /*
485  * Create an anonymous set with the provided mask in the space provided by
486  * 'fset'.  If the passed in set is anonymous we use its parent otherwise
487  * the new set is a child of 'set'.
488  */
489 static int
490 cpuset_shadow(struct cpuset *set, struct cpuset *fset, const cpuset_t *mask)
491 {
492 	struct cpuset *parent;
493 
494 	if (set->cs_id == CPUSET_INVALID)
495 		parent = set->cs_parent;
496 	else
497 		parent = set;
498 	if (!CPU_SUBSET(&parent->cs_mask, mask))
499 		return (EDEADLK);
500 	return (_cpuset_create(fset, parent, mask, CPUSET_INVALID));
501 }
502 
503 /*
504  * Handle two cases for replacing the base set or mask of an entire process.
505  *
506  * 1) Set is non-null and mask is null.  This reparents all anonymous sets
507  *    to the provided set and replaces all non-anonymous td_cpusets with the
508  *    provided set.
509  * 2) Mask is non-null and set is null.  This replaces or creates anonymous
510  *    sets for every thread with the existing base as a parent.
511  *
512  * This is overly complicated because we can't allocate while holding a
513  * spinlock and spinlocks must be held while changing and examining thread
514  * state.
515  */
516 static int
517 cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask)
518 {
519 	struct setlist freelist;
520 	struct setlist droplist;
521 	struct cpuset *tdset;
522 	struct cpuset *nset;
523 	struct thread *td;
524 	struct proc *p;
525 	int threads;
526 	int nfree;
527 	int error;
528 
529 	/*
530 	 * The algorithm requires two passes due to locking considerations.
531 	 *
532 	 * 1) Lookup the process and acquire the locks in the required order.
533 	 * 2) If enough cpusets have not been allocated release the locks and
534 	 *    allocate them.  Loop.
535 	 */
536 	LIST_INIT(&freelist);
537 	LIST_INIT(&droplist);
538 	nfree = 0;
539 	for (;;) {
540 		error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset);
541 		if (error)
542 			goto out;
543 		if (nfree >= p->p_numthreads)
544 			break;
545 		threads = p->p_numthreads;
546 		PROC_UNLOCK(p);
547 		for (; nfree < threads; nfree++) {
548 			nset = uma_zalloc(cpuset_zone, M_WAITOK);
549 			LIST_INSERT_HEAD(&freelist, nset, cs_link);
550 		}
551 	}
552 	PROC_LOCK_ASSERT(p, MA_OWNED);
553 	/*
554 	 * Now that the appropriate locks are held and we have enough cpusets,
555 	 * make sure the operation will succeed before applying changes.  The
556 	 * proc lock prevents td_cpuset from changing between calls.
557 	 */
558 	error = 0;
559 	FOREACH_THREAD_IN_PROC(p, td) {
560 		thread_lock(td);
561 		tdset = td->td_cpuset;
562 		/*
563 		 * Verify that a new mask doesn't specify cpus outside of
564 		 * the set the thread is a member of.
565 		 */
566 		if (mask) {
567 			if (tdset->cs_id == CPUSET_INVALID)
568 				tdset = tdset->cs_parent;
569 			if (!CPU_SUBSET(&tdset->cs_mask, mask))
570 				error = EDEADLK;
571 		/*
572 		 * Verify that a new set won't leave an existing thread
573 		 * mask without a cpu to run on.  It can, however, restrict
574 		 * the set.
575 		 */
576 		} else if (tdset->cs_id == CPUSET_INVALID) {
577 			if (!CPU_OVERLAP(&set->cs_mask, &tdset->cs_mask))
578 				error = EDEADLK;
579 		}
580 		thread_unlock(td);
581 		if (error)
582 			goto unlock_out;
583 	}
584 	/*
585 	 * Replace each thread's cpuset while using deferred release.  We
586 	 * must do this because the thread lock must be held while operating
587 	 * on the thread and this limits the type of operations allowed.
588 	 */
589 	FOREACH_THREAD_IN_PROC(p, td) {
590 		thread_lock(td);
591 		/*
592 		 * If we presently have an anonymous set or are applying a
593 		 * mask we must create an anonymous shadow set.  That is
594 		 * either parented to our existing base or the supplied set.
595 		 *
596 		 * If we have a base set with no anonymous shadow we simply
597 		 * replace it outright.
598 		 */
599 		tdset = td->td_cpuset;
600 		if (tdset->cs_id == CPUSET_INVALID || mask) {
601 			nset = LIST_FIRST(&freelist);
602 			LIST_REMOVE(nset, cs_link);
603 			if (mask)
604 				error = cpuset_shadow(tdset, nset, mask);
605 			else
606 				error = _cpuset_create(nset, set,
607 				    &tdset->cs_mask, CPUSET_INVALID);
608 			if (error) {
609 				LIST_INSERT_HEAD(&freelist, nset, cs_link);
610 				thread_unlock(td);
611 				break;
612 			}
613 		} else
614 			nset = cpuset_ref(set);
615 		cpuset_rel_defer(&droplist, tdset);
616 		td->td_cpuset = nset;
617 		sched_affinity(td);
618 		thread_unlock(td);
619 	}
620 unlock_out:
621 	PROC_UNLOCK(p);
622 out:
623 	while ((nset = LIST_FIRST(&droplist)) != NULL)
624 		cpuset_rel_complete(nset);
625 	while ((nset = LIST_FIRST(&freelist)) != NULL) {
626 		LIST_REMOVE(nset, cs_link);
627 		uma_zfree(cpuset_zone, nset);
628 	}
629 	return (error);
630 }
631 
632 /*
633  * Return a string representing a valid layout for a cpuset_t object.
634  * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
635  */
636 char *
637 cpusetobj_strprint(char *buf, const cpuset_t *set)
638 {
639 	char *tbuf;
640 	size_t i, bytesp, bufsiz;
641 
642 	tbuf = buf;
643 	bytesp = 0;
644 	bufsiz = CPUSETBUFSIZ;
645 
646 	for (i = 0; i < (_NCPUWORDS - 1); i++) {
647 		bytesp = snprintf(tbuf, bufsiz, "%lx,", set->__bits[i]);
648 		bufsiz -= bytesp;
649 		tbuf += bytesp;
650 	}
651 	snprintf(tbuf, bufsiz, "%lx", set->__bits[_NCPUWORDS - 1]);
652 	return (buf);
653 }
654 
655 /*
656  * Build a valid cpuset_t object from a string representation.
657  * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
658  */
659 int
660 cpusetobj_strscan(cpuset_t *set, const char *buf)
661 {
662 	u_int nwords;
663 	int i, ret;
664 
665 	if (strlen(buf) > CPUSETBUFSIZ - 1)
666 		return (-1);
667 
668 	/* Allow to pass a shorter version of the mask when necessary. */
669 	nwords = 1;
670 	for (i = 0; buf[i] != '\0'; i++)
671 		if (buf[i] == ',')
672 			nwords++;
673 	if (nwords > _NCPUWORDS)
674 		return (-1);
675 
676 	CPU_ZERO(set);
677 	for (i = 0; i < (nwords - 1); i++) {
678 		ret = sscanf(buf, "%lx,", &set->__bits[i]);
679 		if (ret == 0 || ret == -1)
680 			return (-1);
681 		buf = strstr(buf, ",");
682 		if (buf == NULL)
683 			return (-1);
684 		buf++;
685 	}
686 	ret = sscanf(buf, "%lx", &set->__bits[nwords - 1]);
687 	if (ret == 0 || ret == -1)
688 		return (-1);
689 	return (0);
690 }
691 
692 /*
693  * Apply an anonymous mask to a single thread.
694  */
695 int
696 cpuset_setthread(lwpid_t id, cpuset_t *mask)
697 {
698 	struct cpuset *nset;
699 	struct cpuset *set;
700 	struct thread *td;
701 	struct proc *p;
702 	int error;
703 
704 	nset = uma_zalloc(cpuset_zone, M_WAITOK);
705 	error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set);
706 	if (error)
707 		goto out;
708 	set = NULL;
709 	thread_lock(td);
710 	error = cpuset_shadow(td->td_cpuset, nset, mask);
711 	if (error == 0) {
712 		set = td->td_cpuset;
713 		td->td_cpuset = nset;
714 		sched_affinity(td);
715 		nset = NULL;
716 	}
717 	thread_unlock(td);
718 	PROC_UNLOCK(p);
719 	if (set)
720 		cpuset_rel(set);
721 out:
722 	if (nset)
723 		uma_zfree(cpuset_zone, nset);
724 	return (error);
725 }
726 
727 /*
728  * Apply new cpumask to the ithread.
729  */
730 int
731 cpuset_setithread(lwpid_t id, int cpu)
732 {
733 	struct cpuset *nset, *rset;
734 	struct cpuset *parent, *old_set;
735 	struct thread *td;
736 	struct proc *p;
737 	cpusetid_t cs_id;
738 	cpuset_t mask;
739 	int error;
740 
741 	nset = uma_zalloc(cpuset_zone, M_WAITOK);
742 	rset = uma_zalloc(cpuset_zone, M_WAITOK);
743 	cs_id = CPUSET_INVALID;
744 
745 	CPU_ZERO(&mask);
746 	if (cpu == NOCPU)
747 		CPU_COPY(cpuset_root, &mask);
748 	else
749 		CPU_SET(cpu, &mask);
750 
751 	error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &old_set);
752 	if (error != 0 || ((cs_id = alloc_unr(cpuset_unr)) == CPUSET_INVALID))
753 		goto out;
754 
755 	/* cpuset_which() returns with PROC_LOCK held. */
756 	old_set = td->td_cpuset;
757 
758 	if (cpu == NOCPU) {
759 
760 		/*
761 		 * roll back to default set. We're not using cpuset_shadow()
762 		 * here because we can fail CPU_SUBSET() check. This can happen
763 		 * if default set does not contain all CPUs.
764 		 */
765 		error = _cpuset_create(nset, cpuset_default, &mask,
766 		    CPUSET_INVALID);
767 
768 		goto applyset;
769 	}
770 
771 	if (old_set->cs_id == 1 || (old_set->cs_id == CPUSET_INVALID &&
772 	    old_set->cs_parent->cs_id == 1)) {
773 
774 		/*
775 		 * Current set is either default (1) or
776 		 * shadowed version of default set.
777 		 *
778 		 * Allocate new root set to be able to shadow it
779 		 * with any mask.
780 		 */
781 		error = _cpuset_create(rset, cpuset_zero,
782 		    &cpuset_zero->cs_mask, cs_id);
783 		if (error != 0) {
784 			PROC_UNLOCK(p);
785 			goto out;
786 		}
787 		rset->cs_flags |= CPU_SET_ROOT;
788 		parent = rset;
789 		rset = NULL;
790 		cs_id = CPUSET_INVALID;
791 	} else {
792 		/* Assume existing set was already allocated by previous call */
793 		parent = old_set;
794 		old_set = NULL;
795 	}
796 
797 	error = cpuset_shadow(parent, nset, &mask);
798 applyset:
799 	if (error == 0) {
800 		thread_lock(td);
801 		td->td_cpuset = nset;
802 		sched_affinity(td);
803 		thread_unlock(td);
804 		nset = NULL;
805 	} else
806 		old_set = NULL;
807 	PROC_UNLOCK(p);
808 	if (old_set != NULL)
809 		cpuset_rel(old_set);
810 out:
811 	if (nset != NULL)
812 		uma_zfree(cpuset_zone, nset);
813 	if (rset != NULL)
814 		uma_zfree(cpuset_zone, rset);
815 	if (cs_id != CPUSET_INVALID)
816 		free_unr(cpuset_unr, cs_id);
817 	return (error);
818 }
819 
820 
821 /*
822  * Creates system-wide cpusets and the cpuset for thread0 including two
823  * sets:
824  *
825  * 0 - The root set which should represent all valid processors in the
826  *     system.  It is initially created with a mask of all processors
827  *     because we don't know what processors are valid until cpuset_init()
828  *     runs.  This set is immutable.
829  * 1 - The default set which all processes are a member of until changed.
830  *     This allows an administrator to move all threads off of given cpus to
831  *     dedicate them to high priority tasks or save power etc.
832  */
833 struct cpuset *
834 cpuset_thread0(void)
835 {
836 	struct cpuset *set;
837 	int error, i;
838 
839 	cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL,
840 	    NULL, NULL, UMA_ALIGN_PTR, 0);
841 	mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE);
842 
843 	/*
844 	 * Create the root system set for the whole machine.  Doesn't use
845 	 * cpuset_create() due to NULL parent.
846 	 */
847 	set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
848 	CPU_FILL(&set->cs_mask);
849 	LIST_INIT(&set->cs_children);
850 	LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
851 	set->cs_ref = 1;
852 	set->cs_flags = CPU_SET_ROOT;
853 	cpuset_zero = set;
854 	cpuset_root = &set->cs_mask;
855 
856 	/*
857 	 * Now derive a default, modifiable set from that to give out.
858 	 */
859 	set = uma_zalloc(cpuset_zone, M_WAITOK);
860 	error = _cpuset_create(set, cpuset_zero, &cpuset_zero->cs_mask, 1);
861 	KASSERT(error == 0, ("Error creating default set: %d\n", error));
862 	cpuset_default = set;
863 
864 	/*
865 	 * Initialize the unit allocator. 0 and 1 are allocated above.
866 	 */
867 	cpuset_unr = new_unrhdr(2, INT_MAX, NULL);
868 
869 	/*
870 	 * If MD code has not initialized per-domain cpusets, place all
871 	 * CPUs in domain 0.
872 	 */
873 	for (i = 0; i < MAXMEMDOM; i++)
874 		if (!CPU_EMPTY(&cpuset_domain[i]))
875 			goto domains_set;
876 	CPU_COPY(&all_cpus, &cpuset_domain[0]);
877 domains_set:
878 
879 	return (set);
880 }
881 
882 /*
883  * Create a cpuset, which would be cpuset_create() but
884  * mark the new 'set' as root.
885  *
886  * We are not going to reparent the td to it.  Use cpuset_setproc_update_set()
887  * for that.
888  *
889  * In case of no error, returns the set in *setp locked with a reference.
890  */
891 int
892 cpuset_create_root(struct prison *pr, struct cpuset **setp)
893 {
894 	struct cpuset *set;
895 	int error;
896 
897 	KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__));
898 	KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__));
899 
900 	error = cpuset_create(setp, pr->pr_cpuset, &pr->pr_cpuset->cs_mask);
901 	if (error)
902 		return (error);
903 
904 	KASSERT(*setp != NULL, ("[%s:%d] cpuset_create returned invalid data",
905 	    __func__, __LINE__));
906 
907 	/* Mark the set as root. */
908 	set = *setp;
909 	set->cs_flags |= CPU_SET_ROOT;
910 
911 	return (0);
912 }
913 
914 int
915 cpuset_setproc_update_set(struct proc *p, struct cpuset *set)
916 {
917 	int error;
918 
919 	KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__));
920 	KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__));
921 
922 	cpuset_ref(set);
923 	error = cpuset_setproc(p->p_pid, set, NULL);
924 	if (error)
925 		return (error);
926 	cpuset_rel(set);
927 	return (0);
928 }
929 
930 /*
931  * This is called once the final set of system cpus is known.  Modifies
932  * the root set and all children and mark the root read-only.
933  */
934 static void
935 cpuset_init(void *arg)
936 {
937 	cpuset_t mask;
938 
939 	mask = all_cpus;
940 	if (cpuset_modify(cpuset_zero, &mask))
941 		panic("Can't set initial cpuset mask.\n");
942 	cpuset_zero->cs_flags |= CPU_SET_RDONLY;
943 }
944 SYSINIT(cpuset, SI_SUB_SMP, SI_ORDER_ANY, cpuset_init, NULL);
945 
946 #ifndef _SYS_SYSPROTO_H_
947 struct cpuset_args {
948 	cpusetid_t	*setid;
949 };
950 #endif
951 int
952 sys_cpuset(struct thread *td, struct cpuset_args *uap)
953 {
954 	struct cpuset *root;
955 	struct cpuset *set;
956 	int error;
957 
958 	thread_lock(td);
959 	root = cpuset_refroot(td->td_cpuset);
960 	thread_unlock(td);
961 	error = cpuset_create(&set, root, &root->cs_mask);
962 	cpuset_rel(root);
963 	if (error)
964 		return (error);
965 	error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id));
966 	if (error == 0)
967 		error = cpuset_setproc(-1, set, NULL);
968 	cpuset_rel(set);
969 	return (error);
970 }
971 
972 #ifndef _SYS_SYSPROTO_H_
973 struct cpuset_setid_args {
974 	cpuwhich_t	which;
975 	id_t		id;
976 	cpusetid_t	setid;
977 };
978 #endif
979 int
980 sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap)
981 {
982 
983 	return (kern_cpuset_setid(td, uap->which, uap->id, uap->setid));
984 }
985 
986 int
987 kern_cpuset_setid(struct thread *td, cpuwhich_t which,
988     id_t id, cpusetid_t setid)
989 {
990 	struct cpuset *set;
991 	int error;
992 
993 	/*
994 	 * Presently we only support per-process sets.
995 	 */
996 	if (which != CPU_WHICH_PID)
997 		return (EINVAL);
998 	set = cpuset_lookup(setid, td);
999 	if (set == NULL)
1000 		return (ESRCH);
1001 	error = cpuset_setproc(id, set, NULL);
1002 	cpuset_rel(set);
1003 	return (error);
1004 }
1005 
1006 #ifndef _SYS_SYSPROTO_H_
1007 struct cpuset_getid_args {
1008 	cpulevel_t	level;
1009 	cpuwhich_t	which;
1010 	id_t		id;
1011 	cpusetid_t	*setid;
1012 };
1013 #endif
1014 int
1015 sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap)
1016 {
1017 
1018 	return (kern_cpuset_getid(td, uap->level, uap->which, uap->id,
1019 	    uap->setid));
1020 }
1021 
1022 int
1023 kern_cpuset_getid(struct thread *td, cpulevel_t level, cpuwhich_t which,
1024     id_t id, cpusetid_t *setid)
1025 {
1026 	struct cpuset *nset;
1027 	struct cpuset *set;
1028 	struct thread *ttd;
1029 	struct proc *p;
1030 	cpusetid_t tmpid;
1031 	int error;
1032 
1033 	if (level == CPU_LEVEL_WHICH && which != CPU_WHICH_CPUSET)
1034 		return (EINVAL);
1035 	error = cpuset_which(which, id, &p, &ttd, &set);
1036 	if (error)
1037 		return (error);
1038 	switch (which) {
1039 	case CPU_WHICH_TID:
1040 	case CPU_WHICH_PID:
1041 		thread_lock(ttd);
1042 		set = cpuset_refbase(ttd->td_cpuset);
1043 		thread_unlock(ttd);
1044 		PROC_UNLOCK(p);
1045 		break;
1046 	case CPU_WHICH_CPUSET:
1047 	case CPU_WHICH_JAIL:
1048 		break;
1049 	case CPU_WHICH_IRQ:
1050 	case CPU_WHICH_DOMAIN:
1051 		return (EINVAL);
1052 	}
1053 	switch (level) {
1054 	case CPU_LEVEL_ROOT:
1055 		nset = cpuset_refroot(set);
1056 		cpuset_rel(set);
1057 		set = nset;
1058 		break;
1059 	case CPU_LEVEL_CPUSET:
1060 		break;
1061 	case CPU_LEVEL_WHICH:
1062 		break;
1063 	}
1064 	tmpid = set->cs_id;
1065 	cpuset_rel(set);
1066 	if (error == 0)
1067 		error = copyout(&tmpid, setid, sizeof(tmpid));
1068 
1069 	return (error);
1070 }
1071 
1072 #ifndef _SYS_SYSPROTO_H_
1073 struct cpuset_getaffinity_args {
1074 	cpulevel_t	level;
1075 	cpuwhich_t	which;
1076 	id_t		id;
1077 	size_t		cpusetsize;
1078 	cpuset_t	*mask;
1079 };
1080 #endif
1081 int
1082 sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap)
1083 {
1084 
1085 	return (kern_cpuset_getaffinity(td, uap->level, uap->which,
1086 	    uap->id, uap->cpusetsize, uap->mask));
1087 }
1088 
1089 int
1090 kern_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1091     id_t id, size_t cpusetsize, cpuset_t *maskp)
1092 {
1093 	struct thread *ttd;
1094 	struct cpuset *nset;
1095 	struct cpuset *set;
1096 	struct proc *p;
1097 	cpuset_t *mask;
1098 	int error;
1099 	size_t size;
1100 
1101 	if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1102 		return (ERANGE);
1103 	/* In Capability mode, you can only get your own CPU set. */
1104 	if (IN_CAPABILITY_MODE(td)) {
1105 	    if (level != CPU_LEVEL_WHICH)
1106 		return (ECAPMODE);
1107 	    if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1108 		return (ECAPMODE);
1109 	    if (id != -1)
1110 		return (ECAPMODE);
1111 	}
1112 	size = cpusetsize;
1113 	mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO);
1114 	error = cpuset_which(which, id, &p, &ttd, &set);
1115 	if (error)
1116 		goto out;
1117 	switch (level) {
1118 	case CPU_LEVEL_ROOT:
1119 	case CPU_LEVEL_CPUSET:
1120 		switch (which) {
1121 		case CPU_WHICH_TID:
1122 		case CPU_WHICH_PID:
1123 			thread_lock(ttd);
1124 			set = cpuset_ref(ttd->td_cpuset);
1125 			thread_unlock(ttd);
1126 			break;
1127 		case CPU_WHICH_CPUSET:
1128 		case CPU_WHICH_JAIL:
1129 			break;
1130 		case CPU_WHICH_IRQ:
1131 		case CPU_WHICH_INTRHANDLER:
1132 		case CPU_WHICH_ITHREAD:
1133 		case CPU_WHICH_DOMAIN:
1134 			error = EINVAL;
1135 			goto out;
1136 		}
1137 		if (level == CPU_LEVEL_ROOT)
1138 			nset = cpuset_refroot(set);
1139 		else
1140 			nset = cpuset_refbase(set);
1141 		CPU_COPY(&nset->cs_mask, mask);
1142 		cpuset_rel(nset);
1143 		break;
1144 	case CPU_LEVEL_WHICH:
1145 		switch (which) {
1146 		case CPU_WHICH_TID:
1147 			thread_lock(ttd);
1148 			CPU_COPY(&ttd->td_cpuset->cs_mask, mask);
1149 			thread_unlock(ttd);
1150 			break;
1151 		case CPU_WHICH_PID:
1152 			FOREACH_THREAD_IN_PROC(p, ttd) {
1153 				thread_lock(ttd);
1154 				CPU_OR(mask, &ttd->td_cpuset->cs_mask);
1155 				thread_unlock(ttd);
1156 			}
1157 			break;
1158 		case CPU_WHICH_CPUSET:
1159 		case CPU_WHICH_JAIL:
1160 			CPU_COPY(&set->cs_mask, mask);
1161 			break;
1162 		case CPU_WHICH_IRQ:
1163 		case CPU_WHICH_INTRHANDLER:
1164 		case CPU_WHICH_ITHREAD:
1165 			error = intr_getaffinity(id, which, mask);
1166 			break;
1167 		case CPU_WHICH_DOMAIN:
1168 			if (id < 0 || id >= MAXMEMDOM)
1169 				error = ESRCH;
1170 			else
1171 				CPU_COPY(&cpuset_domain[id], mask);
1172 			break;
1173 		}
1174 		break;
1175 	default:
1176 		error = EINVAL;
1177 		break;
1178 	}
1179 	if (set)
1180 		cpuset_rel(set);
1181 	if (p)
1182 		PROC_UNLOCK(p);
1183 	if (error == 0)
1184 		error = copyout(mask, maskp, size);
1185 out:
1186 	free(mask, M_TEMP);
1187 	return (error);
1188 }
1189 
1190 #ifndef _SYS_SYSPROTO_H_
1191 struct cpuset_setaffinity_args {
1192 	cpulevel_t	level;
1193 	cpuwhich_t	which;
1194 	id_t		id;
1195 	size_t		cpusetsize;
1196 	const cpuset_t	*mask;
1197 };
1198 #endif
1199 int
1200 sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap)
1201 {
1202 
1203 	return (kern_cpuset_setaffinity(td, uap->level, uap->which,
1204 	    uap->id, uap->cpusetsize, uap->mask));
1205 }
1206 
1207 int
1208 kern_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which,
1209     id_t id, size_t cpusetsize, const cpuset_t *maskp)
1210 {
1211 	struct cpuset *nset;
1212 	struct cpuset *set;
1213 	struct thread *ttd;
1214 	struct proc *p;
1215 	cpuset_t *mask;
1216 	int error;
1217 
1218 	if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY)
1219 		return (ERANGE);
1220 	/* In Capability mode, you can only set your own CPU set. */
1221 	if (IN_CAPABILITY_MODE(td)) {
1222 	    if (level != CPU_LEVEL_WHICH)
1223 		return (ECAPMODE);
1224 	    if (which != CPU_WHICH_TID && which != CPU_WHICH_PID)
1225 		return (ECAPMODE);
1226 	    if (id != -1)
1227 		return (ECAPMODE);
1228 	}
1229 	mask = malloc(cpusetsize, M_TEMP, M_WAITOK | M_ZERO);
1230 	error = copyin(maskp, mask, cpusetsize);
1231 	if (error)
1232 		goto out;
1233 	/*
1234 	 * Verify that no high bits are set.
1235 	 */
1236 	if (cpusetsize > sizeof(cpuset_t)) {
1237 		char *end;
1238 		char *cp;
1239 
1240 		end = cp = (char *)&mask->__bits;
1241 		end += cpusetsize;
1242 		cp += sizeof(cpuset_t);
1243 		while (cp != end)
1244 			if (*cp++ != 0) {
1245 				error = EINVAL;
1246 				goto out;
1247 			}
1248 
1249 	}
1250 	switch (level) {
1251 	case CPU_LEVEL_ROOT:
1252 	case CPU_LEVEL_CPUSET:
1253 		error = cpuset_which(which, id, &p, &ttd, &set);
1254 		if (error)
1255 			break;
1256 		switch (which) {
1257 		case CPU_WHICH_TID:
1258 		case CPU_WHICH_PID:
1259 			thread_lock(ttd);
1260 			set = cpuset_ref(ttd->td_cpuset);
1261 			thread_unlock(ttd);
1262 			PROC_UNLOCK(p);
1263 			break;
1264 		case CPU_WHICH_CPUSET:
1265 		case CPU_WHICH_JAIL:
1266 			break;
1267 		case CPU_WHICH_IRQ:
1268 		case CPU_WHICH_INTRHANDLER:
1269 		case CPU_WHICH_ITHREAD:
1270 		case CPU_WHICH_DOMAIN:
1271 			error = EINVAL;
1272 			goto out;
1273 		}
1274 		if (level == CPU_LEVEL_ROOT)
1275 			nset = cpuset_refroot(set);
1276 		else
1277 			nset = cpuset_refbase(set);
1278 		error = cpuset_modify(nset, mask);
1279 		cpuset_rel(nset);
1280 		cpuset_rel(set);
1281 		break;
1282 	case CPU_LEVEL_WHICH:
1283 		switch (which) {
1284 		case CPU_WHICH_TID:
1285 			error = cpuset_setthread(id, mask);
1286 			break;
1287 		case CPU_WHICH_PID:
1288 			error = cpuset_setproc(id, NULL, mask);
1289 			break;
1290 		case CPU_WHICH_CPUSET:
1291 		case CPU_WHICH_JAIL:
1292 			error = cpuset_which(which, id, &p, &ttd, &set);
1293 			if (error == 0) {
1294 				error = cpuset_modify(set, mask);
1295 				cpuset_rel(set);
1296 			}
1297 			break;
1298 		case CPU_WHICH_IRQ:
1299 		case CPU_WHICH_INTRHANDLER:
1300 		case CPU_WHICH_ITHREAD:
1301 			error = intr_setaffinity(id, which, mask);
1302 			break;
1303 		default:
1304 			error = EINVAL;
1305 			break;
1306 		}
1307 		break;
1308 	default:
1309 		error = EINVAL;
1310 		break;
1311 	}
1312 out:
1313 	free(mask, M_TEMP);
1314 	return (error);
1315 }
1316 
1317 #ifdef DDB
1318 void
1319 ddb_display_cpuset(const cpuset_t *set)
1320 {
1321 	int cpu, once;
1322 
1323 	for (once = 0, cpu = 0; cpu < CPU_SETSIZE; cpu++) {
1324 		if (CPU_ISSET(cpu, set)) {
1325 			if (once == 0) {
1326 				db_printf("%d", cpu);
1327 				once = 1;
1328 			} else
1329 				db_printf(",%d", cpu);
1330 		}
1331 	}
1332 	if (once == 0)
1333 		db_printf("<none>");
1334 }
1335 
1336 DB_SHOW_COMMAND(cpusets, db_show_cpusets)
1337 {
1338 	struct cpuset *set;
1339 
1340 	LIST_FOREACH(set, &cpuset_ids, cs_link) {
1341 		db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n",
1342 		    set, set->cs_id, set->cs_ref, set->cs_flags,
1343 		    (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0);
1344 		db_printf("  mask=");
1345 		ddb_display_cpuset(&set->cs_mask);
1346 		db_printf("\n");
1347 		if (db_pager_quit)
1348 			break;
1349 	}
1350 }
1351 #endif /* DDB */
1352