xref: /freebsd/sys/kern/kern_cpuset.c (revision 6472ac3d8a86336899b6cfb789a4cd9897e3fab5)
1 /*-
2  * Copyright (c) 2008,  Jeffrey Roberson <jeff@freebsd.org>
3  * All rights reserved.
4  *
5  * Copyright (c) 2008 Nokia Corporation
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice unmodified, this list of conditions, and the following
13  *    disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
27  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28  *
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include "opt_ddb.h"
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/sysproto.h>
39 #include <sys/jail.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/mutex.h>
44 #include <sys/priv.h>
45 #include <sys/proc.h>
46 #include <sys/refcount.h>
47 #include <sys/sched.h>
48 #include <sys/smp.h>
49 #include <sys/syscallsubr.h>
50 #include <sys/cpuset.h>
51 #include <sys/sx.h>
52 #include <sys/queue.h>
53 #include <sys/libkern.h>
54 #include <sys/limits.h>
55 #include <sys/bus.h>
56 #include <sys/interrupt.h>
57 
58 #include <vm/uma.h>
59 
60 #ifdef DDB
61 #include <ddb/ddb.h>
62 #endif /* DDB */
63 
64 /*
65  * cpusets provide a mechanism for creating and manipulating sets of
66  * processors for the purpose of constraining the scheduling of threads to
67  * specific processors.
68  *
69  * Each process belongs to an identified set, by default this is set 1.  Each
70  * thread may further restrict the cpus it may run on to a subset of this
71  * named set.  This creates an anonymous set which other threads and processes
72  * may not join by number.
73  *
74  * The named set is referred to herein as the 'base' set to avoid ambiguity.
75  * This set is usually a child of a 'root' set while the anonymous set may
76  * simply be referred to as a mask.  In the syscall api these are referred to
77  * as the ROOT, CPUSET, and MASK levels where CPUSET is called 'base' here.
78  *
79  * Threads inherit their set from their creator whether it be anonymous or
80  * not.  This means that anonymous sets are immutable because they may be
81  * shared.  To modify an anonymous set a new set is created with the desired
82  * mask and the same parent as the existing anonymous set.  This gives the
83  * illusion of each thread having a private mask.
84  *
85  * Via the syscall apis a user may ask to retrieve or modify the root, base,
86  * or mask that is discovered via a pid, tid, or setid.  Modifying a set
87  * modifies all numbered and anonymous child sets to comply with the new mask.
88  * Modifying a pid or tid's mask applies only to that tid but must still
89  * exist within the assigned parent set.
90  *
91  * A thread may not be assigned to a group separate from other threads in
92  * the process.  This is to remove ambiguity when the setid is queried with
93  * a pid argument.  There is no other technical limitation.
94  *
95  * This somewhat complex arrangement is intended to make it easy for
96  * applications to query available processors and bind their threads to
97  * specific processors while also allowing administrators to dynamically
98  * reprovision by changing sets which apply to groups of processes.
99  *
100  * A simple application should not concern itself with sets at all and
101  * rather apply masks to its own threads via CPU_WHICH_TID and a -1 id
102  * meaning 'curthread'.  It may query available cpus for that tid with a
103  * getaffinity call using (CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, ...).
104  */
105 static uma_zone_t cpuset_zone;
106 static struct mtx cpuset_lock;
107 static struct setlist cpuset_ids;
108 static struct unrhdr *cpuset_unr;
109 static struct cpuset *cpuset_zero;
110 
111 /* Return the size of cpuset_t at the kernel level */
112 SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD,
113 	0, sizeof(cpuset_t), "sizeof(cpuset_t)");
114 
115 cpuset_t *cpuset_root;
116 
117 /*
118  * Acquire a reference to a cpuset, all pointers must be tracked with refs.
119  */
120 struct cpuset *
121 cpuset_ref(struct cpuset *set)
122 {
123 
124 	refcount_acquire(&set->cs_ref);
125 	return (set);
126 }
127 
128 /*
129  * Walks up the tree from 'set' to find the root.  Returns the root
130  * referenced.
131  */
132 static struct cpuset *
133 cpuset_refroot(struct cpuset *set)
134 {
135 
136 	for (; set->cs_parent != NULL; set = set->cs_parent)
137 		if (set->cs_flags & CPU_SET_ROOT)
138 			break;
139 	cpuset_ref(set);
140 
141 	return (set);
142 }
143 
144 /*
145  * Find the first non-anonymous set starting from 'set'.  Returns this set
146  * referenced.  May return the passed in set with an extra ref if it is
147  * not anonymous.
148  */
149 static struct cpuset *
150 cpuset_refbase(struct cpuset *set)
151 {
152 
153 	if (set->cs_id == CPUSET_INVALID)
154 		set = set->cs_parent;
155 	cpuset_ref(set);
156 
157 	return (set);
158 }
159 
160 /*
161  * Release a reference in a context where it is safe to allocate.
162  */
163 void
164 cpuset_rel(struct cpuset *set)
165 {
166 	cpusetid_t id;
167 
168 	if (refcount_release(&set->cs_ref) == 0)
169 		return;
170 	mtx_lock_spin(&cpuset_lock);
171 	LIST_REMOVE(set, cs_siblings);
172 	id = set->cs_id;
173 	if (id != CPUSET_INVALID)
174 		LIST_REMOVE(set, cs_link);
175 	mtx_unlock_spin(&cpuset_lock);
176 	cpuset_rel(set->cs_parent);
177 	uma_zfree(cpuset_zone, set);
178 	if (id != CPUSET_INVALID)
179 		free_unr(cpuset_unr, id);
180 }
181 
182 /*
183  * Deferred release must be used when in a context that is not safe to
184  * allocate/free.  This places any unreferenced sets on the list 'head'.
185  */
186 static void
187 cpuset_rel_defer(struct setlist *head, struct cpuset *set)
188 {
189 
190 	if (refcount_release(&set->cs_ref) == 0)
191 		return;
192 	mtx_lock_spin(&cpuset_lock);
193 	LIST_REMOVE(set, cs_siblings);
194 	if (set->cs_id != CPUSET_INVALID)
195 		LIST_REMOVE(set, cs_link);
196 	LIST_INSERT_HEAD(head, set, cs_link);
197 	mtx_unlock_spin(&cpuset_lock);
198 }
199 
200 /*
201  * Complete a deferred release.  Removes the set from the list provided to
202  * cpuset_rel_defer.
203  */
204 static void
205 cpuset_rel_complete(struct cpuset *set)
206 {
207 	LIST_REMOVE(set, cs_link);
208 	cpuset_rel(set->cs_parent);
209 	uma_zfree(cpuset_zone, set);
210 }
211 
212 /*
213  * Find a set based on an id.  Returns it with a ref.
214  */
215 static struct cpuset *
216 cpuset_lookup(cpusetid_t setid, struct thread *td)
217 {
218 	struct cpuset *set;
219 
220 	if (setid == CPUSET_INVALID)
221 		return (NULL);
222 	mtx_lock_spin(&cpuset_lock);
223 	LIST_FOREACH(set, &cpuset_ids, cs_link)
224 		if (set->cs_id == setid)
225 			break;
226 	if (set)
227 		cpuset_ref(set);
228 	mtx_unlock_spin(&cpuset_lock);
229 
230 	KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__));
231 	if (set != NULL && jailed(td->td_ucred)) {
232 		struct cpuset *jset, *tset;
233 
234 		jset = td->td_ucred->cr_prison->pr_cpuset;
235 		for (tset = set; tset != NULL; tset = tset->cs_parent)
236 			if (tset == jset)
237 				break;
238 		if (tset == NULL) {
239 			cpuset_rel(set);
240 			set = NULL;
241 		}
242 	}
243 
244 	return (set);
245 }
246 
247 /*
248  * Create a set in the space provided in 'set' with the provided parameters.
249  * The set is returned with a single ref.  May return EDEADLK if the set
250  * will have no valid cpu based on restrictions from the parent.
251  */
252 static int
253 _cpuset_create(struct cpuset *set, struct cpuset *parent, const cpuset_t *mask,
254     cpusetid_t id)
255 {
256 
257 	if (!CPU_OVERLAP(&parent->cs_mask, mask))
258 		return (EDEADLK);
259 	CPU_COPY(mask, &set->cs_mask);
260 	LIST_INIT(&set->cs_children);
261 	refcount_init(&set->cs_ref, 1);
262 	set->cs_flags = 0;
263 	mtx_lock_spin(&cpuset_lock);
264 	CPU_AND(&set->cs_mask, &parent->cs_mask);
265 	set->cs_id = id;
266 	set->cs_parent = cpuset_ref(parent);
267 	LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings);
268 	if (set->cs_id != CPUSET_INVALID)
269 		LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
270 	mtx_unlock_spin(&cpuset_lock);
271 
272 	return (0);
273 }
274 
275 /*
276  * Create a new non-anonymous set with the requested parent and mask.  May
277  * return failures if the mask is invalid or a new number can not be
278  * allocated.
279  */
280 static int
281 cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask)
282 {
283 	struct cpuset *set;
284 	cpusetid_t id;
285 	int error;
286 
287 	id = alloc_unr(cpuset_unr);
288 	if (id == -1)
289 		return (ENFILE);
290 	*setp = set = uma_zalloc(cpuset_zone, M_WAITOK);
291 	error = _cpuset_create(set, parent, mask, id);
292 	if (error == 0)
293 		return (0);
294 	free_unr(cpuset_unr, id);
295 	uma_zfree(cpuset_zone, set);
296 
297 	return (error);
298 }
299 
300 /*
301  * Recursively check for errors that would occur from applying mask to
302  * the tree of sets starting at 'set'.  Checks for sets that would become
303  * empty as well as RDONLY flags.
304  */
305 static int
306 cpuset_testupdate(struct cpuset *set, cpuset_t *mask)
307 {
308 	struct cpuset *nset;
309 	cpuset_t newmask;
310 	int error;
311 
312 	mtx_assert(&cpuset_lock, MA_OWNED);
313 	if (set->cs_flags & CPU_SET_RDONLY)
314 		return (EPERM);
315 	if (!CPU_OVERLAP(&set->cs_mask, mask))
316 		return (EDEADLK);
317 	CPU_COPY(&set->cs_mask, &newmask);
318 	CPU_AND(&newmask, mask);
319 	error = 0;
320 	LIST_FOREACH(nset, &set->cs_children, cs_siblings)
321 		if ((error = cpuset_testupdate(nset, &newmask)) != 0)
322 			break;
323 	return (error);
324 }
325 
326 /*
327  * Applies the mask 'mask' without checking for empty sets or permissions.
328  */
329 static void
330 cpuset_update(struct cpuset *set, cpuset_t *mask)
331 {
332 	struct cpuset *nset;
333 
334 	mtx_assert(&cpuset_lock, MA_OWNED);
335 	CPU_AND(&set->cs_mask, mask);
336 	LIST_FOREACH(nset, &set->cs_children, cs_siblings)
337 		cpuset_update(nset, &set->cs_mask);
338 
339 	return;
340 }
341 
342 /*
343  * Modify the set 'set' to use a copy of the mask provided.  Apply this new
344  * mask to restrict all children in the tree.  Checks for validity before
345  * applying the changes.
346  */
347 static int
348 cpuset_modify(struct cpuset *set, cpuset_t *mask)
349 {
350 	struct cpuset *root;
351 	int error;
352 
353 	error = priv_check(curthread, PRIV_SCHED_CPUSET);
354 	if (error)
355 		return (error);
356 	/*
357 	 * In case we are called from within the jail
358 	 * we do not allow modifying the dedicated root
359 	 * cpuset of the jail but may still allow to
360 	 * change child sets.
361 	 */
362 	if (jailed(curthread->td_ucred) &&
363 	    set->cs_flags & CPU_SET_ROOT)
364 		return (EPERM);
365 	/*
366 	 * Verify that we have access to this set of
367 	 * cpus.
368 	 */
369 	root = set->cs_parent;
370 	if (root && !CPU_SUBSET(&root->cs_mask, mask))
371 		return (EINVAL);
372 	mtx_lock_spin(&cpuset_lock);
373 	error = cpuset_testupdate(set, mask);
374 	if (error)
375 		goto out;
376 	cpuset_update(set, mask);
377 	CPU_COPY(mask, &set->cs_mask);
378 out:
379 	mtx_unlock_spin(&cpuset_lock);
380 
381 	return (error);
382 }
383 
384 /*
385  * Resolve the 'which' parameter of several cpuset apis.
386  *
387  * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid.  Also
388  * checks for permission via p_cansched().
389  *
390  * For WHICH_SET returns a valid set with a new reference.
391  *
392  * -1 may be supplied for any argument to mean the current proc/thread or
393  * the base set of the current thread.  May fail with ESRCH/EPERM.
394  */
395 static int
396 cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp,
397     struct cpuset **setp)
398 {
399 	struct cpuset *set;
400 	struct thread *td;
401 	struct proc *p;
402 	int error;
403 
404 	*pp = p = NULL;
405 	*tdp = td = NULL;
406 	*setp = set = NULL;
407 	switch (which) {
408 	case CPU_WHICH_PID:
409 		if (id == -1) {
410 			PROC_LOCK(curproc);
411 			p = curproc;
412 			break;
413 		}
414 		if ((p = pfind(id)) == NULL)
415 			return (ESRCH);
416 		break;
417 	case CPU_WHICH_TID:
418 		if (id == -1) {
419 			PROC_LOCK(curproc);
420 			p = curproc;
421 			td = curthread;
422 			break;
423 		}
424 		td = tdfind(id, -1);
425 		if (td == NULL)
426 			return (ESRCH);
427 		p = td->td_proc;
428 		break;
429 	case CPU_WHICH_CPUSET:
430 		if (id == -1) {
431 			thread_lock(curthread);
432 			set = cpuset_refbase(curthread->td_cpuset);
433 			thread_unlock(curthread);
434 		} else
435 			set = cpuset_lookup(id, curthread);
436 		if (set) {
437 			*setp = set;
438 			return (0);
439 		}
440 		return (ESRCH);
441 	case CPU_WHICH_JAIL:
442 	{
443 		/* Find `set' for prison with given id. */
444 		struct prison *pr;
445 
446 		sx_slock(&allprison_lock);
447 		pr = prison_find_child(curthread->td_ucred->cr_prison, id);
448 		sx_sunlock(&allprison_lock);
449 		if (pr == NULL)
450 			return (ESRCH);
451 		cpuset_ref(pr->pr_cpuset);
452 		*setp = pr->pr_cpuset;
453 		mtx_unlock(&pr->pr_mtx);
454 		return (0);
455 	}
456 	case CPU_WHICH_IRQ:
457 		return (0);
458 	default:
459 		return (EINVAL);
460 	}
461 	error = p_cansched(curthread, p);
462 	if (error) {
463 		PROC_UNLOCK(p);
464 		return (error);
465 	}
466 	if (td == NULL)
467 		td = FIRST_THREAD_IN_PROC(p);
468 	*pp = p;
469 	*tdp = td;
470 	return (0);
471 }
472 
473 /*
474  * Create an anonymous set with the provided mask in the space provided by
475  * 'fset'.  If the passed in set is anonymous we use its parent otherwise
476  * the new set is a child of 'set'.
477  */
478 static int
479 cpuset_shadow(struct cpuset *set, struct cpuset *fset, const cpuset_t *mask)
480 {
481 	struct cpuset *parent;
482 
483 	if (set->cs_id == CPUSET_INVALID)
484 		parent = set->cs_parent;
485 	else
486 		parent = set;
487 	if (!CPU_SUBSET(&parent->cs_mask, mask))
488 		return (EDEADLK);
489 	return (_cpuset_create(fset, parent, mask, CPUSET_INVALID));
490 }
491 
492 /*
493  * Handle two cases for replacing the base set or mask of an entire process.
494  *
495  * 1) Set is non-null and mask is null.  This reparents all anonymous sets
496  *    to the provided set and replaces all non-anonymous td_cpusets with the
497  *    provided set.
498  * 2) Mask is non-null and set is null.  This replaces or creates anonymous
499  *    sets for every thread with the existing base as a parent.
500  *
501  * This is overly complicated because we can't allocate while holding a
502  * spinlock and spinlocks must be held while changing and examining thread
503  * state.
504  */
505 static int
506 cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask)
507 {
508 	struct setlist freelist;
509 	struct setlist droplist;
510 	struct cpuset *tdset;
511 	struct cpuset *nset;
512 	struct thread *td;
513 	struct proc *p;
514 	int threads;
515 	int nfree;
516 	int error;
517 	/*
518 	 * The algorithm requires two passes due to locking considerations.
519 	 *
520 	 * 1) Lookup the process and acquire the locks in the required order.
521 	 * 2) If enough cpusets have not been allocated release the locks and
522 	 *    allocate them.  Loop.
523 	 */
524 	LIST_INIT(&freelist);
525 	LIST_INIT(&droplist);
526 	nfree = 0;
527 	for (;;) {
528 		error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset);
529 		if (error)
530 			goto out;
531 		if (nfree >= p->p_numthreads)
532 			break;
533 		threads = p->p_numthreads;
534 		PROC_UNLOCK(p);
535 		for (; nfree < threads; nfree++) {
536 			nset = uma_zalloc(cpuset_zone, M_WAITOK);
537 			LIST_INSERT_HEAD(&freelist, nset, cs_link);
538 		}
539 	}
540 	PROC_LOCK_ASSERT(p, MA_OWNED);
541 	/*
542 	 * Now that the appropriate locks are held and we have enough cpusets,
543 	 * make sure the operation will succeed before applying changes.  The
544 	 * proc lock prevents td_cpuset from changing between calls.
545 	 */
546 	error = 0;
547 	FOREACH_THREAD_IN_PROC(p, td) {
548 		thread_lock(td);
549 		tdset = td->td_cpuset;
550 		/*
551 		 * Verify that a new mask doesn't specify cpus outside of
552 		 * the set the thread is a member of.
553 		 */
554 		if (mask) {
555 			if (tdset->cs_id == CPUSET_INVALID)
556 				tdset = tdset->cs_parent;
557 			if (!CPU_SUBSET(&tdset->cs_mask, mask))
558 				error = EDEADLK;
559 		/*
560 		 * Verify that a new set won't leave an existing thread
561 		 * mask without a cpu to run on.  It can, however, restrict
562 		 * the set.
563 		 */
564 		} else if (tdset->cs_id == CPUSET_INVALID) {
565 			if (!CPU_OVERLAP(&set->cs_mask, &tdset->cs_mask))
566 				error = EDEADLK;
567 		}
568 		thread_unlock(td);
569 		if (error)
570 			goto unlock_out;
571 	}
572 	/*
573 	 * Replace each thread's cpuset while using deferred release.  We
574 	 * must do this because the thread lock must be held while operating
575 	 * on the thread and this limits the type of operations allowed.
576 	 */
577 	FOREACH_THREAD_IN_PROC(p, td) {
578 		thread_lock(td);
579 		/*
580 		 * If we presently have an anonymous set or are applying a
581 		 * mask we must create an anonymous shadow set.  That is
582 		 * either parented to our existing base or the supplied set.
583 		 *
584 		 * If we have a base set with no anonymous shadow we simply
585 		 * replace it outright.
586 		 */
587 		tdset = td->td_cpuset;
588 		if (tdset->cs_id == CPUSET_INVALID || mask) {
589 			nset = LIST_FIRST(&freelist);
590 			LIST_REMOVE(nset, cs_link);
591 			if (mask)
592 				error = cpuset_shadow(tdset, nset, mask);
593 			else
594 				error = _cpuset_create(nset, set,
595 				    &tdset->cs_mask, CPUSET_INVALID);
596 			if (error) {
597 				LIST_INSERT_HEAD(&freelist, nset, cs_link);
598 				thread_unlock(td);
599 				break;
600 			}
601 		} else
602 			nset = cpuset_ref(set);
603 		cpuset_rel_defer(&droplist, tdset);
604 		td->td_cpuset = nset;
605 		sched_affinity(td);
606 		thread_unlock(td);
607 	}
608 unlock_out:
609 	PROC_UNLOCK(p);
610 out:
611 	while ((nset = LIST_FIRST(&droplist)) != NULL)
612 		cpuset_rel_complete(nset);
613 	while ((nset = LIST_FIRST(&freelist)) != NULL) {
614 		LIST_REMOVE(nset, cs_link);
615 		uma_zfree(cpuset_zone, nset);
616 	}
617 	return (error);
618 }
619 
620 /*
621  * Calculate the ffs() of the cpuset.
622  */
623 int
624 cpusetobj_ffs(const cpuset_t *set)
625 {
626 	size_t i;
627 	int cbit;
628 
629 	cbit = 0;
630 	for (i = 0; i < _NCPUWORDS; i++) {
631 		if (set->__bits[i] != 0) {
632 			cbit = ffsl(set->__bits[i]);
633 			cbit += i * _NCPUBITS;
634 			break;
635 		}
636 	}
637 	return (cbit);
638 }
639 
640 /*
641  * Return a string representing a valid layout for a cpuset_t object.
642  * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
643  */
644 char *
645 cpusetobj_strprint(char *buf, const cpuset_t *set)
646 {
647 	char *tbuf;
648 	size_t i, bytesp, bufsiz;
649 
650 	tbuf = buf;
651 	bytesp = 0;
652 	bufsiz = CPUSETBUFSIZ;
653 
654 	for (i = _NCPUWORDS - 1; i > 0; i--) {
655 		bytesp = snprintf(tbuf, bufsiz, "%lx, ", set->__bits[i]);
656 		bufsiz -= bytesp;
657 		tbuf += bytesp;
658 	}
659 	snprintf(tbuf, bufsiz, "%lx", set->__bits[0]);
660 	return (buf);
661 }
662 
663 /*
664  * Build a valid cpuset_t object from a string representation.
665  * It expects an incoming buffer at least sized as CPUSETBUFSIZ.
666  */
667 int
668 cpusetobj_strscan(cpuset_t *set, const char *buf)
669 {
670 	u_int nwords;
671 	int i, ret;
672 
673 	if (strlen(buf) > CPUSETBUFSIZ - 1)
674 		return (-1);
675 
676 	/* Allow to pass a shorter version of the mask when necessary. */
677 	nwords = 1;
678 	for (i = 0; buf[i] != '\0'; i++)
679 		if (buf[i] == ',')
680 			nwords++;
681 	if (nwords > _NCPUWORDS)
682 		return (-1);
683 
684 	CPU_ZERO(set);
685 	for (i = nwords - 1; i > 0; i--) {
686 		ret = sscanf(buf, "%lx, ", &set->__bits[i]);
687 		if (ret == 0 || ret == -1)
688 			return (-1);
689 		buf = strstr(buf, " ");
690 		if (buf == NULL)
691 			return (-1);
692 		buf++;
693 	}
694 	ret = sscanf(buf, "%lx", &set->__bits[0]);
695 	if (ret == 0 || ret == -1)
696 		return (-1);
697 	return (0);
698 }
699 
700 /*
701  * Apply an anonymous mask to a single thread.
702  */
703 int
704 cpuset_setthread(lwpid_t id, cpuset_t *mask)
705 {
706 	struct cpuset *nset;
707 	struct cpuset *set;
708 	struct thread *td;
709 	struct proc *p;
710 	int error;
711 
712 	nset = uma_zalloc(cpuset_zone, M_WAITOK);
713 	error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set);
714 	if (error)
715 		goto out;
716 	set = NULL;
717 	thread_lock(td);
718 	error = cpuset_shadow(td->td_cpuset, nset, mask);
719 	if (error == 0) {
720 		set = td->td_cpuset;
721 		td->td_cpuset = nset;
722 		sched_affinity(td);
723 		nset = NULL;
724 	}
725 	thread_unlock(td);
726 	PROC_UNLOCK(p);
727 	if (set)
728 		cpuset_rel(set);
729 out:
730 	if (nset)
731 		uma_zfree(cpuset_zone, nset);
732 	return (error);
733 }
734 
735 /*
736  * Creates the cpuset for thread0.  We make two sets:
737  *
738  * 0 - The root set which should represent all valid processors in the
739  *     system.  It is initially created with a mask of all processors
740  *     because we don't know what processors are valid until cpuset_init()
741  *     runs.  This set is immutable.
742  * 1 - The default set which all processes are a member of until changed.
743  *     This allows an administrator to move all threads off of given cpus to
744  *     dedicate them to high priority tasks or save power etc.
745  */
746 struct cpuset *
747 cpuset_thread0(void)
748 {
749 	struct cpuset *set;
750 	int error;
751 
752 	cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL,
753 	    NULL, NULL, UMA_ALIGN_PTR, 0);
754 	mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE);
755 	/*
756 	 * Create the root system set for the whole machine.  Doesn't use
757 	 * cpuset_create() due to NULL parent.
758 	 */
759 	set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
760 	CPU_FILL(&set->cs_mask);
761 	LIST_INIT(&set->cs_children);
762 	LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
763 	set->cs_ref = 1;
764 	set->cs_flags = CPU_SET_ROOT;
765 	cpuset_zero = set;
766 	cpuset_root = &set->cs_mask;
767 	/*
768 	 * Now derive a default, modifiable set from that to give out.
769 	 */
770 	set = uma_zalloc(cpuset_zone, M_WAITOK);
771 	error = _cpuset_create(set, cpuset_zero, &cpuset_zero->cs_mask, 1);
772 	KASSERT(error == 0, ("Error creating default set: %d\n", error));
773 	/*
774 	 * Initialize the unit allocator. 0 and 1 are allocated above.
775 	 */
776 	cpuset_unr = new_unrhdr(2, INT_MAX, NULL);
777 
778 	return (set);
779 }
780 
781 /*
782  * Create a cpuset, which would be cpuset_create() but
783  * mark the new 'set' as root.
784  *
785  * We are not going to reparent the td to it.  Use cpuset_setproc_update_set()
786  * for that.
787  *
788  * In case of no error, returns the set in *setp locked with a reference.
789  */
790 int
791 cpuset_create_root(struct prison *pr, struct cpuset **setp)
792 {
793 	struct cpuset *set;
794 	int error;
795 
796 	KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__));
797 	KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__));
798 
799 	error = cpuset_create(setp, pr->pr_cpuset, &pr->pr_cpuset->cs_mask);
800 	if (error)
801 		return (error);
802 
803 	KASSERT(*setp != NULL, ("[%s:%d] cpuset_create returned invalid data",
804 	    __func__, __LINE__));
805 
806 	/* Mark the set as root. */
807 	set = *setp;
808 	set->cs_flags |= CPU_SET_ROOT;
809 
810 	return (0);
811 }
812 
813 int
814 cpuset_setproc_update_set(struct proc *p, struct cpuset *set)
815 {
816 	int error;
817 
818 	KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__));
819 	KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__));
820 
821 	cpuset_ref(set);
822 	error = cpuset_setproc(p->p_pid, set, NULL);
823 	if (error)
824 		return (error);
825 	cpuset_rel(set);
826 	return (0);
827 }
828 
829 /*
830  * This is called once the final set of system cpus is known.  Modifies
831  * the root set and all children and mark the root read-only.
832  */
833 static void
834 cpuset_init(void *arg)
835 {
836 	cpuset_t mask;
837 
838 	mask = all_cpus;
839 	if (cpuset_modify(cpuset_zero, &mask))
840 		panic("Can't set initial cpuset mask.\n");
841 	cpuset_zero->cs_flags |= CPU_SET_RDONLY;
842 }
843 SYSINIT(cpuset, SI_SUB_SMP, SI_ORDER_ANY, cpuset_init, NULL);
844 
845 #ifndef _SYS_SYSPROTO_H_
846 struct cpuset_args {
847 	cpusetid_t	*setid;
848 };
849 #endif
850 int
851 sys_cpuset(struct thread *td, struct cpuset_args *uap)
852 {
853 	struct cpuset *root;
854 	struct cpuset *set;
855 	int error;
856 
857 	thread_lock(td);
858 	root = cpuset_refroot(td->td_cpuset);
859 	thread_unlock(td);
860 	error = cpuset_create(&set, root, &root->cs_mask);
861 	cpuset_rel(root);
862 	if (error)
863 		return (error);
864 	error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id));
865 	if (error == 0)
866 		error = cpuset_setproc(-1, set, NULL);
867 	cpuset_rel(set);
868 	return (error);
869 }
870 
871 #ifndef _SYS_SYSPROTO_H_
872 struct cpuset_setid_args {
873 	cpuwhich_t	which;
874 	id_t		id;
875 	cpusetid_t	setid;
876 };
877 #endif
878 int
879 sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap)
880 {
881 	struct cpuset *set;
882 	int error;
883 
884 	/*
885 	 * Presently we only support per-process sets.
886 	 */
887 	if (uap->which != CPU_WHICH_PID)
888 		return (EINVAL);
889 	set = cpuset_lookup(uap->setid, td);
890 	if (set == NULL)
891 		return (ESRCH);
892 	error = cpuset_setproc(uap->id, set, NULL);
893 	cpuset_rel(set);
894 	return (error);
895 }
896 
897 #ifndef _SYS_SYSPROTO_H_
898 struct cpuset_getid_args {
899 	cpulevel_t	level;
900 	cpuwhich_t	which;
901 	id_t		id;
902 	cpusetid_t	*setid;
903 #endif
904 int
905 sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap)
906 {
907 	struct cpuset *nset;
908 	struct cpuset *set;
909 	struct thread *ttd;
910 	struct proc *p;
911 	cpusetid_t id;
912 	int error;
913 
914 	if (uap->level == CPU_LEVEL_WHICH && uap->which != CPU_WHICH_CPUSET)
915 		return (EINVAL);
916 	error = cpuset_which(uap->which, uap->id, &p, &ttd, &set);
917 	if (error)
918 		return (error);
919 	switch (uap->which) {
920 	case CPU_WHICH_TID:
921 	case CPU_WHICH_PID:
922 		thread_lock(ttd);
923 		set = cpuset_refbase(ttd->td_cpuset);
924 		thread_unlock(ttd);
925 		PROC_UNLOCK(p);
926 		break;
927 	case CPU_WHICH_CPUSET:
928 	case CPU_WHICH_JAIL:
929 		break;
930 	case CPU_WHICH_IRQ:
931 		return (EINVAL);
932 	}
933 	switch (uap->level) {
934 	case CPU_LEVEL_ROOT:
935 		nset = cpuset_refroot(set);
936 		cpuset_rel(set);
937 		set = nset;
938 		break;
939 	case CPU_LEVEL_CPUSET:
940 		break;
941 	case CPU_LEVEL_WHICH:
942 		break;
943 	}
944 	id = set->cs_id;
945 	cpuset_rel(set);
946 	if (error == 0)
947 		error = copyout(&id, uap->setid, sizeof(id));
948 
949 	return (error);
950 }
951 
952 #ifndef _SYS_SYSPROTO_H_
953 struct cpuset_getaffinity_args {
954 	cpulevel_t	level;
955 	cpuwhich_t	which;
956 	id_t		id;
957 	size_t		cpusetsize;
958 	cpuset_t	*mask;
959 };
960 #endif
961 int
962 sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap)
963 {
964 	struct thread *ttd;
965 	struct cpuset *nset;
966 	struct cpuset *set;
967 	struct proc *p;
968 	cpuset_t *mask;
969 	int error;
970 	size_t size;
971 
972 	if (uap->cpusetsize < sizeof(cpuset_t) ||
973 	    uap->cpusetsize > CPU_MAXSIZE / NBBY)
974 		return (ERANGE);
975 	size = uap->cpusetsize;
976 	mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO);
977 	error = cpuset_which(uap->which, uap->id, &p, &ttd, &set);
978 	if (error)
979 		goto out;
980 	switch (uap->level) {
981 	case CPU_LEVEL_ROOT:
982 	case CPU_LEVEL_CPUSET:
983 		switch (uap->which) {
984 		case CPU_WHICH_TID:
985 		case CPU_WHICH_PID:
986 			thread_lock(ttd);
987 			set = cpuset_ref(ttd->td_cpuset);
988 			thread_unlock(ttd);
989 			break;
990 		case CPU_WHICH_CPUSET:
991 		case CPU_WHICH_JAIL:
992 			break;
993 		case CPU_WHICH_IRQ:
994 			error = EINVAL;
995 			goto out;
996 		}
997 		if (uap->level == CPU_LEVEL_ROOT)
998 			nset = cpuset_refroot(set);
999 		else
1000 			nset = cpuset_refbase(set);
1001 		CPU_COPY(&nset->cs_mask, mask);
1002 		cpuset_rel(nset);
1003 		break;
1004 	case CPU_LEVEL_WHICH:
1005 		switch (uap->which) {
1006 		case CPU_WHICH_TID:
1007 			thread_lock(ttd);
1008 			CPU_COPY(&ttd->td_cpuset->cs_mask, mask);
1009 			thread_unlock(ttd);
1010 			break;
1011 		case CPU_WHICH_PID:
1012 			FOREACH_THREAD_IN_PROC(p, ttd) {
1013 				thread_lock(ttd);
1014 				CPU_OR(mask, &ttd->td_cpuset->cs_mask);
1015 				thread_unlock(ttd);
1016 			}
1017 			break;
1018 		case CPU_WHICH_CPUSET:
1019 		case CPU_WHICH_JAIL:
1020 			CPU_COPY(&set->cs_mask, mask);
1021 			break;
1022 		case CPU_WHICH_IRQ:
1023 			error = intr_getaffinity(uap->id, mask);
1024 			break;
1025 		}
1026 		break;
1027 	default:
1028 		error = EINVAL;
1029 		break;
1030 	}
1031 	if (set)
1032 		cpuset_rel(set);
1033 	if (p)
1034 		PROC_UNLOCK(p);
1035 	if (error == 0)
1036 		error = copyout(mask, uap->mask, size);
1037 out:
1038 	free(mask, M_TEMP);
1039 	return (error);
1040 }
1041 
1042 #ifndef _SYS_SYSPROTO_H_
1043 struct cpuset_setaffinity_args {
1044 	cpulevel_t	level;
1045 	cpuwhich_t	which;
1046 	id_t		id;
1047 	size_t		cpusetsize;
1048 	const cpuset_t	*mask;
1049 };
1050 #endif
1051 int
1052 sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap)
1053 {
1054 	struct cpuset *nset;
1055 	struct cpuset *set;
1056 	struct thread *ttd;
1057 	struct proc *p;
1058 	cpuset_t *mask;
1059 	int error;
1060 
1061 	if (uap->cpusetsize < sizeof(cpuset_t) ||
1062 	    uap->cpusetsize > CPU_MAXSIZE / NBBY)
1063 		return (ERANGE);
1064 	mask = malloc(uap->cpusetsize, M_TEMP, M_WAITOK | M_ZERO);
1065 	error = copyin(uap->mask, mask, uap->cpusetsize);
1066 	if (error)
1067 		goto out;
1068 	/*
1069 	 * Verify that no high bits are set.
1070 	 */
1071 	if (uap->cpusetsize > sizeof(cpuset_t)) {
1072 		char *end;
1073 		char *cp;
1074 
1075 		end = cp = (char *)&mask->__bits;
1076 		end += uap->cpusetsize;
1077 		cp += sizeof(cpuset_t);
1078 		while (cp != end)
1079 			if (*cp++ != 0) {
1080 				error = EINVAL;
1081 				goto out;
1082 			}
1083 
1084 	}
1085 	switch (uap->level) {
1086 	case CPU_LEVEL_ROOT:
1087 	case CPU_LEVEL_CPUSET:
1088 		error = cpuset_which(uap->which, uap->id, &p, &ttd, &set);
1089 		if (error)
1090 			break;
1091 		switch (uap->which) {
1092 		case CPU_WHICH_TID:
1093 		case CPU_WHICH_PID:
1094 			thread_lock(ttd);
1095 			set = cpuset_ref(ttd->td_cpuset);
1096 			thread_unlock(ttd);
1097 			PROC_UNLOCK(p);
1098 			break;
1099 		case CPU_WHICH_CPUSET:
1100 		case CPU_WHICH_JAIL:
1101 			break;
1102 		case CPU_WHICH_IRQ:
1103 			error = EINVAL;
1104 			goto out;
1105 		}
1106 		if (uap->level == CPU_LEVEL_ROOT)
1107 			nset = cpuset_refroot(set);
1108 		else
1109 			nset = cpuset_refbase(set);
1110 		error = cpuset_modify(nset, mask);
1111 		cpuset_rel(nset);
1112 		cpuset_rel(set);
1113 		break;
1114 	case CPU_LEVEL_WHICH:
1115 		switch (uap->which) {
1116 		case CPU_WHICH_TID:
1117 			error = cpuset_setthread(uap->id, mask);
1118 			break;
1119 		case CPU_WHICH_PID:
1120 			error = cpuset_setproc(uap->id, NULL, mask);
1121 			break;
1122 		case CPU_WHICH_CPUSET:
1123 		case CPU_WHICH_JAIL:
1124 			error = cpuset_which(uap->which, uap->id, &p,
1125 			    &ttd, &set);
1126 			if (error == 0) {
1127 				error = cpuset_modify(set, mask);
1128 				cpuset_rel(set);
1129 			}
1130 			break;
1131 		case CPU_WHICH_IRQ:
1132 			error = intr_setaffinity(uap->id, mask);
1133 			break;
1134 		default:
1135 			error = EINVAL;
1136 			break;
1137 		}
1138 		break;
1139 	default:
1140 		error = EINVAL;
1141 		break;
1142 	}
1143 out:
1144 	free(mask, M_TEMP);
1145 	return (error);
1146 }
1147 
1148 #ifdef DDB
1149 DB_SHOW_COMMAND(cpusets, db_show_cpusets)
1150 {
1151 	struct cpuset *set;
1152 	int cpu, once;
1153 
1154 	LIST_FOREACH(set, &cpuset_ids, cs_link) {
1155 		db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n",
1156 		    set, set->cs_id, set->cs_ref, set->cs_flags,
1157 		    (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0);
1158 		db_printf("  mask=");
1159 		for (once = 0, cpu = 0; cpu < CPU_SETSIZE; cpu++) {
1160 			if (CPU_ISSET(cpu, &set->cs_mask)) {
1161 				if (once == 0) {
1162 					db_printf("%d", cpu);
1163 					once = 1;
1164 				} else
1165 					db_printf(",%d", cpu);
1166 			}
1167 		}
1168 		db_printf("\n");
1169 		if (db_pager_quit)
1170 			break;
1171 	}
1172 }
1173 #endif /* DDB */
1174