xref: /freebsd/sys/kern/kern_cpuset.c (revision a9148abd9da5db2f1c682fb17bed791845fc41c9)
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/kernel.h>
40 #include <sys/lock.h>
41 #include <sys/malloc.h>
42 #include <sys/mutex.h>
43 #include <sys/priv.h>
44 #include <sys/proc.h>
45 #include <sys/refcount.h>
46 #include <sys/sched.h>
47 #include <sys/smp.h>
48 #include <sys/syscallsubr.h>
49 #include <sys/cpuset.h>
50 #include <sys/sx.h>
51 #include <sys/refcount.h>
52 #include <sys/queue.h>
53 #include <sys/limits.h>
54 #include <sys/bus.h>
55 #include <sys/interrupt.h>
56 
57 #include <vm/uma.h>
58 
59 #ifdef DDB
60 #include <ddb/ddb.h>
61 #endif /* DDB */
62 
63 /*
64  * cpusets provide a mechanism for creating and manipulating sets of
65  * processors for the purpose of constraining the scheduling of threads to
66  * specific processors.
67  *
68  * Each process belongs to an identified set, by default this is set 1.  Each
69  * thread may further restrict the cpus it may run on to a subset of this
70  * named set.  This creates an anonymous set which other threads and processes
71  * may not join by number.
72  *
73  * The named set is referred to herein as the 'base' set to avoid ambiguity.
74  * This set is usually a child of a 'root' set while the anonymous set may
75  * simply be referred to as a mask.  In the syscall api these are referred to
76  * as the ROOT, CPUSET, and MASK levels where CPUSET is called 'base' here.
77  *
78  * Threads inherit their set from their creator whether it be anonymous or
79  * not.  This means that anonymous sets are immutable because they may be
80  * shared.  To modify an anonymous set a new set is created with the desired
81  * mask and the same parent as the existing anonymous set.  This gives the
82  * illusion of each thread having a private mask.A
83  *
84  * Via the syscall apis a user may ask to retrieve or modify the root, base,
85  * or mask that is discovered via a pid, tid, or setid.  Modifying a set
86  * modifies all numbered and anonymous child sets to comply with the new mask.
87  * Modifying a pid or tid's mask applies only to that tid but must still
88  * exist within the assigned parent set.
89  *
90  * A thread may not be assigned to a a group seperate from other threads in
91  * the process.  This is to remove ambiguity when the setid is queried with
92  * a pid argument.  There is no other technical limitation.
93  *
94  * This somewhat complex arrangement is intended to make it easy for
95  * applications to query available processors and bind their threads to
96  * specific processors while also allowing administrators to dynamically
97  * reprovision by changing sets which apply to groups of processes.
98  *
99  * A simple application should not concern itself with sets at all and
100  * rather apply masks to its own threads via CPU_WHICH_TID and a -1 id
101  * meaning 'curthread'.  It may query availble cpus for that tid with a
102  * getaffinity call using (CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, ...).
103  */
104 static uma_zone_t cpuset_zone;
105 static struct mtx cpuset_lock;
106 static struct setlist cpuset_ids;
107 static struct unrhdr *cpuset_unr;
108 static struct cpuset *cpuset_zero;
109 
110 cpuset_t *cpuset_root;
111 
112 /*
113  * Acquire a reference to a cpuset, all pointers must be tracked with refs.
114  */
115 struct cpuset *
116 cpuset_ref(struct cpuset *set)
117 {
118 
119 	refcount_acquire(&set->cs_ref);
120 	return (set);
121 }
122 
123 /*
124  * Walks up the tree from 'set' to find the root.  Returns the root
125  * referenced.
126  */
127 static struct cpuset *
128 cpuset_refroot(struct cpuset *set)
129 {
130 
131 	for (; set->cs_parent != NULL; set = set->cs_parent)
132 		if (set->cs_flags & CPU_SET_ROOT)
133 			break;
134 	cpuset_ref(set);
135 
136 	return (set);
137 }
138 
139 /*
140  * Find the first non-anonymous set starting from 'set'.  Returns this set
141  * referenced.  May return the passed in set with an extra ref if it is
142  * not anonymous.
143  */
144 static struct cpuset *
145 cpuset_refbase(struct cpuset *set)
146 {
147 
148 	if (set->cs_id == CPUSET_INVALID)
149 		set = set->cs_parent;
150 	cpuset_ref(set);
151 
152 	return (set);
153 }
154 
155 /*
156  * Release a reference in a context where it is safe to allocte.
157  */
158 void
159 cpuset_rel(struct cpuset *set)
160 {
161 	cpusetid_t id;
162 
163 	if (refcount_release(&set->cs_ref) == 0)
164 		return;
165 	mtx_lock_spin(&cpuset_lock);
166 	LIST_REMOVE(set, cs_siblings);
167 	id = set->cs_id;
168 	if (id != CPUSET_INVALID)
169 		LIST_REMOVE(set, cs_link);
170 	mtx_unlock_spin(&cpuset_lock);
171 	cpuset_rel(set->cs_parent);
172 	uma_zfree(cpuset_zone, set);
173 	if (id != CPUSET_INVALID)
174 		free_unr(cpuset_unr, id);
175 }
176 
177 /*
178  * Deferred release must be used when in a context that is not safe to
179  * allocate/free.  This places any unreferenced sets on the list 'head'.
180  */
181 static void
182 cpuset_rel_defer(struct setlist *head, struct cpuset *set)
183 {
184 
185 	if (refcount_release(&set->cs_ref) == 0)
186 		return;
187 	mtx_lock_spin(&cpuset_lock);
188 	LIST_REMOVE(set, cs_siblings);
189 	if (set->cs_id != CPUSET_INVALID)
190 		LIST_REMOVE(set, cs_link);
191 	LIST_INSERT_HEAD(head, set, cs_link);
192 	mtx_unlock_spin(&cpuset_lock);
193 }
194 
195 /*
196  * Complete a deferred release.  Removes the set from the list provided to
197  * cpuset_rel_defer.
198  */
199 static void
200 cpuset_rel_complete(struct cpuset *set)
201 {
202 	LIST_REMOVE(set, cs_link);
203 	cpuset_rel(set->cs_parent);
204 	uma_zfree(cpuset_zone, set);
205 }
206 
207 /*
208  * Find a set based on an id.  Returns it with a ref.
209  */
210 static struct cpuset *
211 cpuset_lookup(cpusetid_t setid)
212 {
213 	struct cpuset *set;
214 
215 	if (setid == CPUSET_INVALID)
216 		return (NULL);
217 	mtx_lock_spin(&cpuset_lock);
218 	LIST_FOREACH(set, &cpuset_ids, cs_link)
219 		if (set->cs_id == setid)
220 			break;
221 	if (set)
222 		cpuset_ref(set);
223 	mtx_unlock_spin(&cpuset_lock);
224 	return (set);
225 }
226 
227 /*
228  * Create a set in the space provided in 'set' with the provided parameters.
229  * The set is returned with a single ref.  May return EDEADLK if the set
230  * will have no valid cpu based on restrictions from the parent.
231  */
232 static int
233 _cpuset_create(struct cpuset *set, struct cpuset *parent, cpuset_t *mask,
234     cpusetid_t id)
235 {
236 
237 	if (!CPU_OVERLAP(&parent->cs_mask, mask))
238 		return (EDEADLK);
239 	CPU_COPY(mask, &set->cs_mask);
240 	LIST_INIT(&set->cs_children);
241 	refcount_init(&set->cs_ref, 1);
242 	set->cs_flags = 0;
243 	mtx_lock_spin(&cpuset_lock);
244 	CPU_AND(mask, &parent->cs_mask);
245 	set->cs_id = id;
246 	set->cs_parent = cpuset_ref(parent);
247 	LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings);
248 	if (set->cs_id != CPUSET_INVALID)
249 		LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
250 	mtx_unlock_spin(&cpuset_lock);
251 
252 	return (0);
253 }
254 
255 /*
256  * Create a new non-anonymous set with the requested parent and mask.  May
257  * return failures if the mask is invalid or a new number can not be
258  * allocated.
259  */
260 static int
261 cpuset_create(struct cpuset **setp, struct cpuset *parent, cpuset_t *mask)
262 {
263 	struct cpuset *set;
264 	cpusetid_t id;
265 	int error;
266 
267 	id = alloc_unr(cpuset_unr);
268 	if (id == -1)
269 		return (ENFILE);
270 	*setp = set = uma_zalloc(cpuset_zone, M_WAITOK);
271 	error = _cpuset_create(set, parent, mask, id);
272 	if (error == 0)
273 		return (0);
274 	free_unr(cpuset_unr, id);
275 	uma_zfree(cpuset_zone, set);
276 
277 	return (error);
278 }
279 
280 /*
281  * Recursively check for errors that would occur from applying mask to
282  * the tree of sets starting at 'set'.  Checks for sets that would become
283  * empty as well as RDONLY flags.
284  */
285 static int
286 cpuset_testupdate(struct cpuset *set, cpuset_t *mask)
287 {
288 	struct cpuset *nset;
289 	cpuset_t newmask;
290 	int error;
291 
292 	mtx_assert(&cpuset_lock, MA_OWNED);
293 	if (set->cs_flags & CPU_SET_RDONLY)
294 		return (EPERM);
295 	if (!CPU_OVERLAP(&set->cs_mask, mask))
296 		return (EDEADLK);
297 	CPU_COPY(&set->cs_mask, &newmask);
298 	CPU_AND(&newmask, mask);
299 	error = 0;
300 	LIST_FOREACH(nset, &set->cs_children, cs_siblings)
301 		if ((error = cpuset_testupdate(nset, &newmask)) != 0)
302 			break;
303 	return (error);
304 }
305 
306 /*
307  * Applies the mask 'mask' without checking for empty sets or permissions.
308  */
309 static void
310 cpuset_update(struct cpuset *set, cpuset_t *mask)
311 {
312 	struct cpuset *nset;
313 
314 	mtx_assert(&cpuset_lock, MA_OWNED);
315 	CPU_AND(&set->cs_mask, mask);
316 	LIST_FOREACH(nset, &set->cs_children, cs_siblings)
317 		cpuset_update(nset, &set->cs_mask);
318 
319 	return;
320 }
321 
322 /*
323  * Modify the set 'set' to use a copy of the mask provided.  Apply this new
324  * mask to restrict all children in the tree.  Checks for validity before
325  * applying the changes.
326  */
327 static int
328 cpuset_modify(struct cpuset *set, cpuset_t *mask)
329 {
330 	struct cpuset *root;
331 	int error;
332 
333 	error = priv_check(curthread, PRIV_SCHED_CPUSET);
334 	if (error)
335 		return (error);
336 	/*
337 	 * Verify that we have access to this set of
338 	 * cpus.
339 	 */
340 	root = set->cs_parent;
341 	if (root && !CPU_SUBSET(&root->cs_mask, mask))
342 		return (EINVAL);
343 	mtx_lock_spin(&cpuset_lock);
344 	error = cpuset_testupdate(set, mask);
345 	if (error)
346 		goto out;
347 	cpuset_update(set, mask);
348 	CPU_COPY(mask, &set->cs_mask);
349 out:
350 	mtx_unlock_spin(&cpuset_lock);
351 
352 	return (error);
353 }
354 
355 /*
356  * Resolve the 'which' parameter of several cpuset apis.
357  *
358  * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid.  Also
359  * checks for permission via p_cansched().
360  *
361  * For WHICH_SET returns a valid set with a new reference.
362  *
363  * -1 may be supplied for any argument to mean the current proc/thread or
364  * the base set of the current thread.  May fail with ESRCH/EPERM.
365  */
366 static int
367 cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp,
368     struct cpuset **setp)
369 {
370 	struct cpuset *set;
371 	struct thread *td;
372 	struct proc *p;
373 	int error;
374 
375 	*pp = p = NULL;
376 	*tdp = td = NULL;
377 	*setp = set = NULL;
378 	switch (which) {
379 	case CPU_WHICH_PID:
380 		if (id == -1) {
381 			PROC_LOCK(curproc);
382 			p = curproc;
383 			break;
384 		}
385 		if ((p = pfind(id)) == NULL)
386 			return (ESRCH);
387 		break;
388 	case CPU_WHICH_TID:
389 		if (id == -1) {
390 			PROC_LOCK(curproc);
391 			p = curproc;
392 			td = curthread;
393 			break;
394 		}
395 		sx_slock(&allproc_lock);
396 		FOREACH_PROC_IN_SYSTEM(p) {
397 			PROC_LOCK(p);
398 			FOREACH_THREAD_IN_PROC(p, td)
399 				if (td->td_tid == id)
400 					break;
401 			if (td != NULL)
402 				break;
403 			PROC_UNLOCK(p);
404 		}
405 		sx_sunlock(&allproc_lock);
406 		if (td == NULL)
407 			return (ESRCH);
408 		break;
409 	case CPU_WHICH_CPUSET:
410 		if (id == -1) {
411 			thread_lock(curthread);
412 			set = cpuset_refbase(curthread->td_cpuset);
413 			thread_unlock(curthread);
414 		} else
415 			set = cpuset_lookup(id);
416 		if (set) {
417 			*setp = set;
418 			return (0);
419 		}
420 		return (ESRCH);
421 	case CPU_WHICH_IRQ:
422 		return (0);
423 	default:
424 		return (EINVAL);
425 	}
426 	error = p_cansched(curthread, p);
427 	if (error) {
428 		PROC_UNLOCK(p);
429 		return (error);
430 	}
431 	if (td == NULL)
432 		td = FIRST_THREAD_IN_PROC(p);
433 	*pp = p;
434 	*tdp = td;
435 	return (0);
436 }
437 
438 /*
439  * Create an anonymous set with the provided mask in the space provided by
440  * 'fset'.  If the passed in set is anonymous we use its parent otherwise
441  * the new set is a child of 'set'.
442  */
443 static int
444 cpuset_shadow(struct cpuset *set, struct cpuset *fset, cpuset_t *mask)
445 {
446 	struct cpuset *parent;
447 
448 	if (set->cs_id == CPUSET_INVALID)
449 		parent = set->cs_parent;
450 	else
451 		parent = set;
452 	if (!CPU_SUBSET(&parent->cs_mask, mask))
453 		return (EDEADLK);
454 	return (_cpuset_create(fset, parent, mask, CPUSET_INVALID));
455 }
456 
457 /*
458  * Handle two cases for replacing the base set or mask of an entire process.
459  *
460  * 1) Set is non-null and mask is null.  This reparents all anonymous sets
461  *    to the provided set and replaces all non-anonymous td_cpusets with the
462  *    provided set.
463  * 2) Mask is non-null and set is null.  This replaces or creates anonymous
464  *    sets for every thread with the existing base as a parent.
465  *
466  * This is overly complicated because we can't allocate while holding a
467  * spinlock and spinlocks must be held while changing and examining thread
468  * state.
469  */
470 static int
471 cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask)
472 {
473 	struct setlist freelist;
474 	struct setlist droplist;
475 	struct cpuset *tdset;
476 	struct cpuset *nset;
477 	struct thread *td;
478 	struct proc *p;
479 	int threads;
480 	int nfree;
481 	int error;
482 	/*
483 	 * The algorithm requires two passes due to locking considerations.
484 	 *
485 	 * 1) Lookup the process and acquire the locks in the required order.
486 	 * 2) If enough cpusets have not been allocated release the locks and
487 	 *    allocate them.  Loop.
488 	 */
489 	LIST_INIT(&freelist);
490 	LIST_INIT(&droplist);
491 	nfree = 0;
492 	for (;;) {
493 		error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset);
494 		if (error)
495 			goto out;
496 		if (nfree >= p->p_numthreads)
497 			break;
498 		threads = p->p_numthreads;
499 		PROC_UNLOCK(p);
500 		for (; nfree < threads; nfree++) {
501 			nset = uma_zalloc(cpuset_zone, M_WAITOK);
502 			LIST_INSERT_HEAD(&freelist, nset, cs_link);
503 		}
504 	}
505 	PROC_LOCK_ASSERT(p, MA_OWNED);
506 	/*
507 	 * Now that the appropriate locks are held and we have enough cpusets,
508 	 * make sure the operation will succeed before applying changes.  The
509 	 * proc lock prevents td_cpuset from changing between calls.
510 	 */
511 	error = 0;
512 	FOREACH_THREAD_IN_PROC(p, td) {
513 		thread_lock(td);
514 		tdset = td->td_cpuset;
515 		/*
516 		 * Verify that a new mask doesn't specify cpus outside of
517 		 * the set the thread is a member of.
518 		 */
519 		if (mask) {
520 			if (tdset->cs_id == CPUSET_INVALID)
521 				tdset = tdset->cs_parent;
522 			if (!CPU_SUBSET(&tdset->cs_mask, mask))
523 				error = EDEADLK;
524 		/*
525 		 * Verify that a new set won't leave an existing thread
526 		 * mask without a cpu to run on.  It can, however, restrict
527 		 * the set.
528 		 */
529 		} else if (tdset->cs_id == CPUSET_INVALID) {
530 			if (!CPU_OVERLAP(&set->cs_mask, &tdset->cs_mask))
531 				error = EDEADLK;
532 		}
533 		thread_unlock(td);
534 		if (error)
535 			goto unlock_out;
536 	}
537 	/*
538 	 * Replace each thread's cpuset while using deferred release.  We
539 	 * must do this because the thread lock must be held while operating
540 	 * on the thread and this limits the type of operations allowed.
541 	 */
542 	FOREACH_THREAD_IN_PROC(p, td) {
543 		thread_lock(td);
544 		/*
545 		 * If we presently have an anonymous set or are applying a
546 		 * mask we must create an anonymous shadow set.  That is
547 		 * either parented to our existing base or the supplied set.
548 		 *
549 		 * If we have a base set with no anonymous shadow we simply
550 		 * replace it outright.
551 		 */
552 		tdset = td->td_cpuset;
553 		if (tdset->cs_id == CPUSET_INVALID || mask) {
554 			nset = LIST_FIRST(&freelist);
555 			LIST_REMOVE(nset, cs_link);
556 			if (mask)
557 				error = cpuset_shadow(tdset, nset, mask);
558 			else
559 				error = _cpuset_create(nset, set,
560 				    &tdset->cs_mask, CPUSET_INVALID);
561 			if (error) {
562 				LIST_INSERT_HEAD(&freelist, nset, cs_link);
563 				thread_unlock(td);
564 				break;
565 			}
566 		} else
567 			nset = cpuset_ref(set);
568 		cpuset_rel_defer(&droplist, tdset);
569 		td->td_cpuset = nset;
570 		sched_affinity(td);
571 		thread_unlock(td);
572 	}
573 unlock_out:
574 	PROC_UNLOCK(p);
575 out:
576 	while ((nset = LIST_FIRST(&droplist)) != NULL)
577 		cpuset_rel_complete(nset);
578 	while ((nset = LIST_FIRST(&freelist)) != NULL) {
579 		LIST_REMOVE(nset, cs_link);
580 		uma_zfree(cpuset_zone, nset);
581 	}
582 	return (error);
583 }
584 
585 /*
586  * Apply an anonymous mask to a single thread.
587  */
588 int
589 cpuset_setthread(lwpid_t id, cpuset_t *mask)
590 {
591 	struct cpuset *nset;
592 	struct cpuset *set;
593 	struct thread *td;
594 	struct proc *p;
595 	int error;
596 
597 	nset = uma_zalloc(cpuset_zone, M_WAITOK);
598 	error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set);
599 	if (error)
600 		goto out;
601 	set = NULL;
602 	thread_lock(td);
603 	error = cpuset_shadow(td->td_cpuset, nset, mask);
604 	if (error == 0) {
605 		set = td->td_cpuset;
606 		td->td_cpuset = nset;
607 		sched_affinity(td);
608 		nset = NULL;
609 	}
610 	thread_unlock(td);
611 	PROC_UNLOCK(p);
612 	if (set)
613 		cpuset_rel(set);
614 out:
615 	if (nset)
616 		uma_zfree(cpuset_zone, nset);
617 	return (error);
618 }
619 
620 /*
621  * Creates the cpuset for thread0.  We make two sets:
622  *
623  * 0 - The root set which should represent all valid processors in the
624  *     system.  It is initially created with a mask of all processors
625  *     because we don't know what processors are valid until cpuset_init()
626  *     runs.  This set is immutable.
627  * 1 - The default set which all processes are a member of until changed.
628  *     This allows an administrator to move all threads off of given cpus to
629  *     dedicate them to high priority tasks or save power etc.
630  */
631 struct cpuset *
632 cpuset_thread0(void)
633 {
634 	struct cpuset *set;
635 	int error;
636 
637 	cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL,
638 	    NULL, NULL, UMA_ALIGN_PTR, 0);
639 	mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE);
640 	/*
641 	 * Create the root system set for the whole machine.  Doesn't use
642 	 * cpuset_create() due to NULL parent.
643 	 */
644 	set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO);
645 	set->cs_mask.__bits[0] = -1;
646 	LIST_INIT(&set->cs_children);
647 	LIST_INSERT_HEAD(&cpuset_ids, set, cs_link);
648 	set->cs_ref = 1;
649 	set->cs_flags = CPU_SET_ROOT;
650 	cpuset_zero = set;
651 	cpuset_root = &set->cs_mask;
652 	/*
653 	 * Now derive a default, modifiable set from that to give out.
654 	 */
655 	set = uma_zalloc(cpuset_zone, M_WAITOK);
656 	error = _cpuset_create(set, cpuset_zero, &cpuset_zero->cs_mask, 1);
657 	KASSERT(error == 0, ("Error creating default set: %d\n", error));
658 	/*
659 	 * Initialize the unit allocator. 0 and 1 are allocated above.
660 	 */
661 	cpuset_unr = new_unrhdr(2, INT_MAX, NULL);
662 
663 	return (set);
664 }
665 
666 /*
667  * This is called once the final set of system cpus is known.  Modifies
668  * the root set and all children and mark the root readonly.
669  */
670 static void
671 cpuset_init(void *arg)
672 {
673 	cpuset_t mask;
674 
675 	CPU_ZERO(&mask);
676 #ifdef SMP
677 	mask.__bits[0] = all_cpus;
678 #else
679 	mask.__bits[0] = 1;
680 #endif
681 	if (cpuset_modify(cpuset_zero, &mask))
682 		panic("Can't set initial cpuset mask.\n");
683 	cpuset_zero->cs_flags |= CPU_SET_RDONLY;
684 }
685 SYSINIT(cpuset, SI_SUB_SMP, SI_ORDER_ANY, cpuset_init, NULL);
686 
687 #ifndef _SYS_SYSPROTO_H_
688 struct cpuset_args {
689 	cpusetid_t	*setid;
690 };
691 #endif
692 int
693 cpuset(struct thread *td, struct cpuset_args *uap)
694 {
695 	struct cpuset *root;
696 	struct cpuset *set;
697 	int error;
698 
699 	thread_lock(td);
700 	root = cpuset_refroot(td->td_cpuset);
701 	thread_unlock(td);
702 	error = cpuset_create(&set, root, &root->cs_mask);
703 	cpuset_rel(root);
704 	if (error)
705 		return (error);
706 	error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id));
707 	if (error == 0)
708 		error = cpuset_setproc(-1, set, NULL);
709 	cpuset_rel(set);
710 	return (error);
711 }
712 
713 #ifndef _SYS_SYSPROTO_H_
714 struct cpuset_setid_args {
715 	cpuwhich_t	which;
716 	id_t		id;
717 	cpusetid_t	setid;
718 };
719 #endif
720 int
721 cpuset_setid(struct thread *td, struct cpuset_setid_args *uap)
722 {
723 	struct cpuset *set;
724 	int error;
725 
726 	/*
727 	 * Presently we only support per-process sets.
728 	 */
729 	if (uap->which != CPU_WHICH_PID)
730 		return (EINVAL);
731 	set = cpuset_lookup(uap->setid);
732 	if (set == NULL)
733 		return (ESRCH);
734 	error = cpuset_setproc(uap->id, set, NULL);
735 	cpuset_rel(set);
736 	return (error);
737 }
738 
739 #ifndef _SYS_SYSPROTO_H_
740 struct cpuset_getid_args {
741 	cpulevel_t	level;
742 	cpuwhich_t	which;
743 	id_t		id;
744 	cpusetid_t	*setid;
745 #endif
746 int
747 cpuset_getid(struct thread *td, struct cpuset_getid_args *uap)
748 {
749 	struct cpuset *nset;
750 	struct cpuset *set;
751 	struct thread *ttd;
752 	struct proc *p;
753 	cpusetid_t id;
754 	int error;
755 
756 	if (uap->level == CPU_LEVEL_WHICH && uap->which != CPU_WHICH_CPUSET)
757 		return (EINVAL);
758 	error = cpuset_which(uap->which, uap->id, &p, &ttd, &set);
759 	if (error)
760 		return (error);
761 	switch (uap->which) {
762 	case CPU_WHICH_TID:
763 	case CPU_WHICH_PID:
764 		thread_lock(ttd);
765 		set = cpuset_refbase(ttd->td_cpuset);
766 		thread_unlock(ttd);
767 		PROC_UNLOCK(p);
768 		break;
769 	case CPU_WHICH_CPUSET:
770 		break;
771 	case CPU_WHICH_IRQ:
772 		return (EINVAL);
773 	}
774 	switch (uap->level) {
775 	case CPU_LEVEL_ROOT:
776 		nset = cpuset_refroot(set);
777 		cpuset_rel(set);
778 		set = nset;
779 		break;
780 	case CPU_LEVEL_CPUSET:
781 		break;
782 	case CPU_LEVEL_WHICH:
783 		break;
784 	}
785 	id = set->cs_id;
786 	cpuset_rel(set);
787 	if (error == 0)
788 		error = copyout(&id, uap->setid, sizeof(id));
789 
790 	return (error);
791 }
792 
793 #ifndef _SYS_SYSPROTO_H_
794 struct cpuset_getaffinity_args {
795 	cpulevel_t	level;
796 	cpuwhich_t	which;
797 	id_t		id;
798 	size_t		cpusetsize;
799 	cpuset_t	*mask;
800 };
801 #endif
802 int
803 cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap)
804 {
805 	struct thread *ttd;
806 	struct cpuset *nset;
807 	struct cpuset *set;
808 	struct proc *p;
809 	cpuset_t *mask;
810 	int error;
811 	size_t size;
812 
813 	if (uap->cpusetsize < sizeof(cpuset_t) ||
814 	    uap->cpusetsize > CPU_MAXSIZE / NBBY)
815 		return (ERANGE);
816 	size = uap->cpusetsize;
817 	mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO);
818 	error = cpuset_which(uap->which, uap->id, &p, &ttd, &set);
819 	if (error)
820 		goto out;
821 	switch (uap->level) {
822 	case CPU_LEVEL_ROOT:
823 	case CPU_LEVEL_CPUSET:
824 		switch (uap->which) {
825 		case CPU_WHICH_TID:
826 		case CPU_WHICH_PID:
827 			thread_lock(ttd);
828 			set = cpuset_ref(ttd->td_cpuset);
829 			thread_unlock(ttd);
830 			break;
831 		case CPU_WHICH_CPUSET:
832 			break;
833 		case CPU_WHICH_IRQ:
834 			error = EINVAL;
835 			goto out;
836 		}
837 		if (uap->level == CPU_LEVEL_ROOT)
838 			nset = cpuset_refroot(set);
839 		else
840 			nset = cpuset_refbase(set);
841 		CPU_COPY(&nset->cs_mask, mask);
842 		cpuset_rel(nset);
843 		break;
844 	case CPU_LEVEL_WHICH:
845 		switch (uap->which) {
846 		case CPU_WHICH_TID:
847 			thread_lock(ttd);
848 			CPU_COPY(&ttd->td_cpuset->cs_mask, mask);
849 			thread_unlock(ttd);
850 			break;
851 		case CPU_WHICH_PID:
852 			FOREACH_THREAD_IN_PROC(p, ttd) {
853 				thread_lock(ttd);
854 				CPU_OR(mask, &ttd->td_cpuset->cs_mask);
855 				thread_unlock(ttd);
856 			}
857 			break;
858 		case CPU_WHICH_CPUSET:
859 			CPU_COPY(&set->cs_mask, mask);
860 			break;
861 		case CPU_WHICH_IRQ:
862 			error = intr_getaffinity(uap->id, mask);
863 			break;
864 		}
865 		break;
866 	default:
867 		error = EINVAL;
868 		break;
869 	}
870 	if (set)
871 		cpuset_rel(set);
872 	if (p)
873 		PROC_UNLOCK(p);
874 	if (error == 0)
875 		error = copyout(mask, uap->mask, size);
876 out:
877 	free(mask, M_TEMP);
878 	return (error);
879 }
880 
881 #ifndef _SYS_SYSPROTO_H_
882 struct cpuset_setaffinity_args {
883 	cpulevel_t	level;
884 	cpuwhich_t	which;
885 	id_t		id;
886 	size_t		cpusetsize;
887 	const cpuset_t	*mask;
888 };
889 #endif
890 int
891 cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap)
892 {
893 	struct cpuset *nset;
894 	struct cpuset *set;
895 	struct thread *ttd;
896 	struct proc *p;
897 	cpuset_t *mask;
898 	int error;
899 
900 	if (uap->cpusetsize < sizeof(cpuset_t) ||
901 	    uap->cpusetsize > CPU_MAXSIZE / NBBY)
902 		return (ERANGE);
903 	mask = malloc(uap->cpusetsize, M_TEMP, M_WAITOK | M_ZERO);
904 	error = copyin(uap->mask, mask, uap->cpusetsize);
905 	if (error)
906 		goto out;
907 	/*
908 	 * Verify that no high bits are set.
909 	 */
910 	if (uap->cpusetsize > sizeof(cpuset_t)) {
911 		char *end;
912 		char *cp;
913 
914 		end = cp = (char *)&mask->__bits;
915 		end += uap->cpusetsize;
916 		cp += sizeof(cpuset_t);
917 		while (cp != end)
918 			if (*cp++ != 0) {
919 				error = EINVAL;
920 				goto out;
921 			}
922 
923 	}
924 	switch (uap->level) {
925 	case CPU_LEVEL_ROOT:
926 	case CPU_LEVEL_CPUSET:
927 		error = cpuset_which(uap->which, uap->id, &p, &ttd, &set);
928 		if (error)
929 			break;
930 		switch (uap->which) {
931 		case CPU_WHICH_TID:
932 		case CPU_WHICH_PID:
933 			thread_lock(ttd);
934 			set = cpuset_ref(ttd->td_cpuset);
935 			thread_unlock(ttd);
936 			PROC_UNLOCK(p);
937 			break;
938 		case CPU_WHICH_CPUSET:
939 			break;
940 		case CPU_WHICH_IRQ:
941 			error = EINVAL;
942 			goto out;
943 		}
944 		if (uap->level == CPU_LEVEL_ROOT)
945 			nset = cpuset_refroot(set);
946 		else
947 			nset = cpuset_refbase(set);
948 		error = cpuset_modify(nset, mask);
949 		cpuset_rel(nset);
950 		cpuset_rel(set);
951 		break;
952 	case CPU_LEVEL_WHICH:
953 		switch (uap->which) {
954 		case CPU_WHICH_TID:
955 			error = cpuset_setthread(uap->id, mask);
956 			break;
957 		case CPU_WHICH_PID:
958 			error = cpuset_setproc(uap->id, NULL, mask);
959 			break;
960 		case CPU_WHICH_CPUSET:
961 			error = cpuset_which(CPU_WHICH_CPUSET, uap->id, &p,
962 			    &ttd, &set);
963 			if (error == 0) {
964 				error = cpuset_modify(set, mask);
965 				cpuset_rel(set);
966 			}
967 			break;
968 		case CPU_WHICH_IRQ:
969 			error = intr_setaffinity(uap->id, mask);
970 			break;
971 		default:
972 			error = EINVAL;
973 			break;
974 		}
975 		break;
976 	default:
977 		error = EINVAL;
978 		break;
979 	}
980 out:
981 	free(mask, M_TEMP);
982 	return (error);
983 }
984 
985 #ifdef DDB
986 DB_SHOW_COMMAND(cpusets, db_show_cpusets)
987 {
988 	struct cpuset *set;
989 	int cpu, once;
990 
991 	LIST_FOREACH(set, &cpuset_ids, cs_link) {
992 		db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n",
993 		    set, set->cs_id, set->cs_ref, set->cs_flags,
994 		    (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0);
995 		db_printf("  mask=");
996 		for (once = 0, cpu = 0; cpu < CPU_SETSIZE; cpu++) {
997 			if (CPU_ISSET(cpu, &set->cs_mask)) {
998 				if (once == 0) {
999 					db_printf("%d", cpu);
1000 					once = 1;
1001 				} else
1002 					db_printf(",%d", cpu);
1003 			}
1004 		}
1005 		db_printf("\n");
1006 		if (db_pager_quit)
1007 			break;
1008 	}
1009 }
1010 #endif /* DDB */
1011