xref: /freebsd/sys/kern/kern_resource.c (revision fe6060f10f634930ff71b7c50291ddc610da2475)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)kern_resource.c	8.5 (Berkeley) 1/21/94
37  */
38 
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/sysproto.h>
45 #include <sys/file.h>
46 #include <sys/kernel.h>
47 #include <sys/lock.h>
48 #include <sys/malloc.h>
49 #include <sys/mutex.h>
50 #include <sys/priv.h>
51 #include <sys/proc.h>
52 #include <sys/refcount.h>
53 #include <sys/racct.h>
54 #include <sys/resourcevar.h>
55 #include <sys/rwlock.h>
56 #include <sys/sched.h>
57 #include <sys/sx.h>
58 #include <sys/syscallsubr.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysent.h>
61 #include <sys/time.h>
62 #include <sys/umtxvar.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_param.h>
66 #include <vm/pmap.h>
67 #include <vm/vm_map.h>
68 
69 static MALLOC_DEFINE(M_PLIMIT, "plimit", "plimit structures");
70 static MALLOC_DEFINE(M_UIDINFO, "uidinfo", "uidinfo structures");
71 #define	UIHASH(uid)	(&uihashtbl[(uid) & uihash])
72 static struct rwlock uihashtbl_lock;
73 static LIST_HEAD(uihashhead, uidinfo) *uihashtbl;
74 static u_long uihash;		/* size of hash table - 1 */
75 
76 static void	calcru1(struct proc *p, struct rusage_ext *ruxp,
77 		    struct timeval *up, struct timeval *sp);
78 static int	donice(struct thread *td, struct proc *chgp, int n);
79 static struct uidinfo *uilookup(uid_t uid);
80 static void	ruxagg_ext_locked(struct rusage_ext *rux, struct thread *td);
81 
82 /*
83  * Resource controls and accounting.
84  */
85 #ifndef _SYS_SYSPROTO_H_
86 struct getpriority_args {
87 	int	which;
88 	int	who;
89 };
90 #endif
91 int
92 sys_getpriority(struct thread *td, struct getpriority_args *uap)
93 {
94 
95 	return (kern_getpriority(td, uap->which, uap->who));
96 }
97 
98 int
99 kern_getpriority(struct thread *td, int which, int who)
100 {
101 	struct proc *p;
102 	struct pgrp *pg;
103 	int error, low;
104 
105 	error = 0;
106 	low = PRIO_MAX + 1;
107 	switch (which) {
108 	case PRIO_PROCESS:
109 		if (who == 0)
110 			low = td->td_proc->p_nice;
111 		else {
112 			p = pfind(who);
113 			if (p == NULL)
114 				break;
115 			if (p_cansee(td, p) == 0)
116 				low = p->p_nice;
117 			PROC_UNLOCK(p);
118 		}
119 		break;
120 
121 	case PRIO_PGRP:
122 		sx_slock(&proctree_lock);
123 		if (who == 0) {
124 			pg = td->td_proc->p_pgrp;
125 			PGRP_LOCK(pg);
126 		} else {
127 			pg = pgfind(who);
128 			if (pg == NULL) {
129 				sx_sunlock(&proctree_lock);
130 				break;
131 			}
132 		}
133 		sx_sunlock(&proctree_lock);
134 		LIST_FOREACH(p, &pg->pg_members, p_pglist) {
135 			PROC_LOCK(p);
136 			if (p->p_state == PRS_NORMAL &&
137 			    p_cansee(td, p) == 0) {
138 				if (p->p_nice < low)
139 					low = p->p_nice;
140 			}
141 			PROC_UNLOCK(p);
142 		}
143 		PGRP_UNLOCK(pg);
144 		break;
145 
146 	case PRIO_USER:
147 		if (who == 0)
148 			who = td->td_ucred->cr_uid;
149 		sx_slock(&allproc_lock);
150 		FOREACH_PROC_IN_SYSTEM(p) {
151 			PROC_LOCK(p);
152 			if (p->p_state == PRS_NORMAL &&
153 			    p_cansee(td, p) == 0 &&
154 			    p->p_ucred->cr_uid == who) {
155 				if (p->p_nice < low)
156 					low = p->p_nice;
157 			}
158 			PROC_UNLOCK(p);
159 		}
160 		sx_sunlock(&allproc_lock);
161 		break;
162 
163 	default:
164 		error = EINVAL;
165 		break;
166 	}
167 	if (low == PRIO_MAX + 1 && error == 0)
168 		error = ESRCH;
169 	td->td_retval[0] = low;
170 	return (error);
171 }
172 
173 #ifndef _SYS_SYSPROTO_H_
174 struct setpriority_args {
175 	int	which;
176 	int	who;
177 	int	prio;
178 };
179 #endif
180 int
181 sys_setpriority(struct thread *td, struct setpriority_args *uap)
182 {
183 
184 	return (kern_setpriority(td, uap->which, uap->who, uap->prio));
185 }
186 
187 int
188 kern_setpriority(struct thread *td, int which, int who, int prio)
189 {
190 	struct proc *curp, *p;
191 	struct pgrp *pg;
192 	int found = 0, error = 0;
193 
194 	curp = td->td_proc;
195 	switch (which) {
196 	case PRIO_PROCESS:
197 		if (who == 0) {
198 			PROC_LOCK(curp);
199 			error = donice(td, curp, prio);
200 			PROC_UNLOCK(curp);
201 		} else {
202 			p = pfind(who);
203 			if (p == NULL)
204 				break;
205 			error = p_cansee(td, p);
206 			if (error == 0)
207 				error = donice(td, p, prio);
208 			PROC_UNLOCK(p);
209 		}
210 		found++;
211 		break;
212 
213 	case PRIO_PGRP:
214 		sx_slock(&proctree_lock);
215 		if (who == 0) {
216 			pg = curp->p_pgrp;
217 			PGRP_LOCK(pg);
218 		} else {
219 			pg = pgfind(who);
220 			if (pg == NULL) {
221 				sx_sunlock(&proctree_lock);
222 				break;
223 			}
224 		}
225 		sx_sunlock(&proctree_lock);
226 		LIST_FOREACH(p, &pg->pg_members, p_pglist) {
227 			PROC_LOCK(p);
228 			if (p->p_state == PRS_NORMAL &&
229 			    p_cansee(td, p) == 0) {
230 				error = donice(td, p, prio);
231 				found++;
232 			}
233 			PROC_UNLOCK(p);
234 		}
235 		PGRP_UNLOCK(pg);
236 		break;
237 
238 	case PRIO_USER:
239 		if (who == 0)
240 			who = td->td_ucred->cr_uid;
241 		sx_slock(&allproc_lock);
242 		FOREACH_PROC_IN_SYSTEM(p) {
243 			PROC_LOCK(p);
244 			if (p->p_state == PRS_NORMAL &&
245 			    p->p_ucred->cr_uid == who &&
246 			    p_cansee(td, p) == 0) {
247 				error = donice(td, p, prio);
248 				found++;
249 			}
250 			PROC_UNLOCK(p);
251 		}
252 		sx_sunlock(&allproc_lock);
253 		break;
254 
255 	default:
256 		error = EINVAL;
257 		break;
258 	}
259 	if (found == 0 && error == 0)
260 		error = ESRCH;
261 	return (error);
262 }
263 
264 /*
265  * Set "nice" for a (whole) process.
266  */
267 static int
268 donice(struct thread *td, struct proc *p, int n)
269 {
270 	int error;
271 
272 	PROC_LOCK_ASSERT(p, MA_OWNED);
273 	if ((error = p_cansched(td, p)))
274 		return (error);
275 	if (n > PRIO_MAX)
276 		n = PRIO_MAX;
277 	if (n < PRIO_MIN)
278 		n = PRIO_MIN;
279 	if (n < p->p_nice && priv_check(td, PRIV_SCHED_SETPRIORITY) != 0)
280 		return (EACCES);
281 	sched_nice(p, n);
282 	return (0);
283 }
284 
285 static int unprivileged_idprio;
286 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_idprio, CTLFLAG_RW,
287     &unprivileged_idprio, 0, "Allow non-root users to set an idle priority");
288 
289 /*
290  * Set realtime priority for LWP.
291  */
292 #ifndef _SYS_SYSPROTO_H_
293 struct rtprio_thread_args {
294 	int		function;
295 	lwpid_t		lwpid;
296 	struct rtprio	*rtp;
297 };
298 #endif
299 int
300 sys_rtprio_thread(struct thread *td, struct rtprio_thread_args *uap)
301 {
302 	struct proc *p;
303 	struct rtprio rtp;
304 	struct thread *td1;
305 	int cierror, error;
306 
307 	/* Perform copyin before acquiring locks if needed. */
308 	if (uap->function == RTP_SET)
309 		cierror = copyin(uap->rtp, &rtp, sizeof(struct rtprio));
310 	else
311 		cierror = 0;
312 
313 	if (uap->lwpid == 0 || uap->lwpid == td->td_tid) {
314 		p = td->td_proc;
315 		td1 = td;
316 		PROC_LOCK(p);
317 	} else {
318 		td1 = tdfind(uap->lwpid, -1);
319 		if (td1 == NULL)
320 			return (ESRCH);
321 		p = td1->td_proc;
322 	}
323 
324 	switch (uap->function) {
325 	case RTP_LOOKUP:
326 		if ((error = p_cansee(td, p)))
327 			break;
328 		pri_to_rtp(td1, &rtp);
329 		PROC_UNLOCK(p);
330 		return (copyout(&rtp, uap->rtp, sizeof(struct rtprio)));
331 	case RTP_SET:
332 		if ((error = p_cansched(td, p)) || (error = cierror))
333 			break;
334 
335 		/* Disallow setting rtprio in most cases if not superuser. */
336 
337 		/*
338 		 * Realtime priority has to be restricted for reasons which
339 		 * should be obvious.  However, for idleprio processes, there is
340 		 * a potential for system deadlock if an idleprio process gains
341 		 * a lock on a resource that other processes need (and the
342 		 * idleprio process can't run due to a CPU-bound normal
343 		 * process).  Fix me!  XXX
344 		 *
345 		 * This problem is not only related to idleprio process.
346 		 * A user level program can obtain a file lock and hold it
347 		 * indefinitely.  Additionally, without idleprio processes it is
348 		 * still conceivable that a program with low priority will never
349 		 * get to run.  In short, allowing this feature might make it
350 		 * easier to lock a resource indefinitely, but it is not the
351 		 * only thing that makes it possible.
352 		 */
353 		if (RTP_PRIO_BASE(rtp.type) == RTP_PRIO_REALTIME ||
354 		    (RTP_PRIO_BASE(rtp.type) == RTP_PRIO_IDLE &&
355 		    unprivileged_idprio == 0)) {
356 			error = priv_check(td, PRIV_SCHED_RTPRIO);
357 			if (error)
358 				break;
359 		}
360 		error = rtp_to_pri(&rtp, td1);
361 		break;
362 	default:
363 		error = EINVAL;
364 		break;
365 	}
366 	PROC_UNLOCK(p);
367 	return (error);
368 }
369 
370 /*
371  * Set realtime priority.
372  */
373 #ifndef _SYS_SYSPROTO_H_
374 struct rtprio_args {
375 	int		function;
376 	pid_t		pid;
377 	struct rtprio	*rtp;
378 };
379 #endif
380 int
381 sys_rtprio(struct thread *td, struct rtprio_args *uap)
382 {
383 	struct proc *p;
384 	struct thread *tdp;
385 	struct rtprio rtp;
386 	int cierror, error;
387 
388 	/* Perform copyin before acquiring locks if needed. */
389 	if (uap->function == RTP_SET)
390 		cierror = copyin(uap->rtp, &rtp, sizeof(struct rtprio));
391 	else
392 		cierror = 0;
393 
394 	if (uap->pid == 0) {
395 		p = td->td_proc;
396 		PROC_LOCK(p);
397 	} else {
398 		p = pfind(uap->pid);
399 		if (p == NULL)
400 			return (ESRCH);
401 	}
402 
403 	switch (uap->function) {
404 	case RTP_LOOKUP:
405 		if ((error = p_cansee(td, p)))
406 			break;
407 		/*
408 		 * Return OUR priority if no pid specified,
409 		 * or if one is, report the highest priority
410 		 * in the process.  There isn't much more you can do as
411 		 * there is only room to return a single priority.
412 		 * Note: specifying our own pid is not the same
413 		 * as leaving it zero.
414 		 */
415 		if (uap->pid == 0) {
416 			pri_to_rtp(td, &rtp);
417 		} else {
418 			struct rtprio rtp2;
419 
420 			rtp.type = RTP_PRIO_IDLE;
421 			rtp.prio = RTP_PRIO_MAX;
422 			FOREACH_THREAD_IN_PROC(p, tdp) {
423 				pri_to_rtp(tdp, &rtp2);
424 				if (rtp2.type <  rtp.type ||
425 				    (rtp2.type == rtp.type &&
426 				    rtp2.prio < rtp.prio)) {
427 					rtp.type = rtp2.type;
428 					rtp.prio = rtp2.prio;
429 				}
430 			}
431 		}
432 		PROC_UNLOCK(p);
433 		return (copyout(&rtp, uap->rtp, sizeof(struct rtprio)));
434 	case RTP_SET:
435 		if ((error = p_cansched(td, p)) || (error = cierror))
436 			break;
437 
438 		/*
439 		 * Disallow setting rtprio in most cases if not superuser.
440 		 * See the comment in sys_rtprio_thread about idprio
441 		 * threads holding a lock.
442 		 */
443 		if (RTP_PRIO_BASE(rtp.type) == RTP_PRIO_REALTIME ||
444 		    (RTP_PRIO_BASE(rtp.type) == RTP_PRIO_IDLE &&
445 		    !unprivileged_idprio)) {
446 			error = priv_check(td, PRIV_SCHED_RTPRIO);
447 			if (error)
448 				break;
449 		}
450 
451 		/*
452 		 * If we are setting our own priority, set just our
453 		 * thread but if we are doing another process,
454 		 * do all the threads on that process. If we
455 		 * specify our own pid we do the latter.
456 		 */
457 		if (uap->pid == 0) {
458 			error = rtp_to_pri(&rtp, td);
459 		} else {
460 			FOREACH_THREAD_IN_PROC(p, td) {
461 				if ((error = rtp_to_pri(&rtp, td)) != 0)
462 					break;
463 			}
464 		}
465 		break;
466 	default:
467 		error = EINVAL;
468 		break;
469 	}
470 	PROC_UNLOCK(p);
471 	return (error);
472 }
473 
474 int
475 rtp_to_pri(struct rtprio *rtp, struct thread *td)
476 {
477 	u_char  newpri, oldclass, oldpri;
478 
479 	switch (RTP_PRIO_BASE(rtp->type)) {
480 	case RTP_PRIO_REALTIME:
481 		if (rtp->prio > RTP_PRIO_MAX)
482 			return (EINVAL);
483 		newpri = PRI_MIN_REALTIME + rtp->prio;
484 		break;
485 	case RTP_PRIO_NORMAL:
486 		if (rtp->prio > (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE))
487 			return (EINVAL);
488 		newpri = PRI_MIN_TIMESHARE + rtp->prio;
489 		break;
490 	case RTP_PRIO_IDLE:
491 		if (rtp->prio > RTP_PRIO_MAX)
492 			return (EINVAL);
493 		newpri = PRI_MIN_IDLE + rtp->prio;
494 		break;
495 	default:
496 		return (EINVAL);
497 	}
498 
499 	thread_lock(td);
500 	oldclass = td->td_pri_class;
501 	sched_class(td, rtp->type);	/* XXX fix */
502 	oldpri = td->td_user_pri;
503 	sched_user_prio(td, newpri);
504 	if (td->td_user_pri != oldpri && (oldclass != RTP_PRIO_NORMAL ||
505 	    td->td_pri_class != RTP_PRIO_NORMAL))
506 		sched_prio(td, td->td_user_pri);
507 	if (TD_ON_UPILOCK(td) && oldpri != newpri) {
508 		critical_enter();
509 		thread_unlock(td);
510 		umtx_pi_adjust(td, oldpri);
511 		critical_exit();
512 	} else
513 		thread_unlock(td);
514 	return (0);
515 }
516 
517 void
518 pri_to_rtp(struct thread *td, struct rtprio *rtp)
519 {
520 
521 	thread_lock(td);
522 	switch (PRI_BASE(td->td_pri_class)) {
523 	case PRI_REALTIME:
524 		rtp->prio = td->td_base_user_pri - PRI_MIN_REALTIME;
525 		break;
526 	case PRI_TIMESHARE:
527 		rtp->prio = td->td_base_user_pri - PRI_MIN_TIMESHARE;
528 		break;
529 	case PRI_IDLE:
530 		rtp->prio = td->td_base_user_pri - PRI_MIN_IDLE;
531 		break;
532 	default:
533 		break;
534 	}
535 	rtp->type = td->td_pri_class;
536 	thread_unlock(td);
537 }
538 
539 #if defined(COMPAT_43)
540 #ifndef _SYS_SYSPROTO_H_
541 struct osetrlimit_args {
542 	u_int	which;
543 	struct	orlimit *rlp;
544 };
545 #endif
546 int
547 osetrlimit(struct thread *td, struct osetrlimit_args *uap)
548 {
549 	struct orlimit olim;
550 	struct rlimit lim;
551 	int error;
552 
553 	if ((error = copyin(uap->rlp, &olim, sizeof(struct orlimit))))
554 		return (error);
555 	lim.rlim_cur = olim.rlim_cur;
556 	lim.rlim_max = olim.rlim_max;
557 	error = kern_setrlimit(td, uap->which, &lim);
558 	return (error);
559 }
560 
561 #ifndef _SYS_SYSPROTO_H_
562 struct ogetrlimit_args {
563 	u_int	which;
564 	struct	orlimit *rlp;
565 };
566 #endif
567 int
568 ogetrlimit(struct thread *td, struct ogetrlimit_args *uap)
569 {
570 	struct orlimit olim;
571 	struct rlimit rl;
572 	int error;
573 
574 	if (uap->which >= RLIM_NLIMITS)
575 		return (EINVAL);
576 	lim_rlimit(td, uap->which, &rl);
577 
578 	/*
579 	 * XXX would be more correct to convert only RLIM_INFINITY to the
580 	 * old RLIM_INFINITY and fail with EOVERFLOW for other larger
581 	 * values.  Most 64->32 and 32->16 conversions, including not
582 	 * unimportant ones of uids are even more broken than what we
583 	 * do here (they blindly truncate).  We don't do this correctly
584 	 * here since we have little experience with EOVERFLOW yet.
585 	 * Elsewhere, getuid() can't fail...
586 	 */
587 	olim.rlim_cur = rl.rlim_cur > 0x7fffffff ? 0x7fffffff : rl.rlim_cur;
588 	olim.rlim_max = rl.rlim_max > 0x7fffffff ? 0x7fffffff : rl.rlim_max;
589 	error = copyout(&olim, uap->rlp, sizeof(olim));
590 	return (error);
591 }
592 #endif /* COMPAT_43 */
593 
594 #ifndef _SYS_SYSPROTO_H_
595 struct __setrlimit_args {
596 	u_int	which;
597 	struct	rlimit *rlp;
598 };
599 #endif
600 int
601 sys_setrlimit(struct thread *td, struct __setrlimit_args *uap)
602 {
603 	struct rlimit alim;
604 	int error;
605 
606 	if ((error = copyin(uap->rlp, &alim, sizeof(struct rlimit))))
607 		return (error);
608 	error = kern_setrlimit(td, uap->which, &alim);
609 	return (error);
610 }
611 
612 static void
613 lim_cb(void *arg)
614 {
615 	struct rlimit rlim;
616 	struct thread *td;
617 	struct proc *p;
618 
619 	p = arg;
620 	PROC_LOCK_ASSERT(p, MA_OWNED);
621 	/*
622 	 * Check if the process exceeds its cpu resource allocation.  If
623 	 * it reaches the max, arrange to kill the process in ast().
624 	 */
625 	if (p->p_cpulimit == RLIM_INFINITY)
626 		return;
627 	PROC_STATLOCK(p);
628 	FOREACH_THREAD_IN_PROC(p, td) {
629 		ruxagg(p, td);
630 	}
631 	PROC_STATUNLOCK(p);
632 	if (p->p_rux.rux_runtime > p->p_cpulimit * cpu_tickrate()) {
633 		lim_rlimit_proc(p, RLIMIT_CPU, &rlim);
634 		if (p->p_rux.rux_runtime >= rlim.rlim_max * cpu_tickrate()) {
635 			killproc(p, "exceeded maximum CPU limit");
636 		} else {
637 			if (p->p_cpulimit < rlim.rlim_max)
638 				p->p_cpulimit += 5;
639 			kern_psignal(p, SIGXCPU);
640 		}
641 	}
642 	if ((p->p_flag & P_WEXIT) == 0)
643 		callout_reset_sbt(&p->p_limco, SBT_1S, 0,
644 		    lim_cb, p, C_PREL(1));
645 }
646 
647 int
648 kern_setrlimit(struct thread *td, u_int which, struct rlimit *limp)
649 {
650 
651 	return (kern_proc_setrlimit(td, td->td_proc, which, limp));
652 }
653 
654 int
655 kern_proc_setrlimit(struct thread *td, struct proc *p, u_int which,
656     struct rlimit *limp)
657 {
658 	struct plimit *newlim, *oldlim;
659 	struct rlimit *alimp;
660 	struct rlimit oldssiz;
661 	int error;
662 
663 	if (which >= RLIM_NLIMITS)
664 		return (EINVAL);
665 
666 	/*
667 	 * Preserve historical bugs by treating negative limits as unsigned.
668 	 */
669 	if (limp->rlim_cur < 0)
670 		limp->rlim_cur = RLIM_INFINITY;
671 	if (limp->rlim_max < 0)
672 		limp->rlim_max = RLIM_INFINITY;
673 
674 	if (which == RLIMIT_STACK && limp->rlim_cur != RLIM_INFINITY)
675 		limp->rlim_cur += p->p_vmspace->vm_stkgap;
676 
677 	oldssiz.rlim_cur = 0;
678 	newlim = lim_alloc();
679 	PROC_LOCK(p);
680 	oldlim = p->p_limit;
681 	alimp = &oldlim->pl_rlimit[which];
682 	if (limp->rlim_cur > alimp->rlim_max ||
683 	    limp->rlim_max > alimp->rlim_max)
684 		if ((error = priv_check(td, PRIV_PROC_SETRLIMIT))) {
685 			PROC_UNLOCK(p);
686 			lim_free(newlim);
687 			return (error);
688 		}
689 	if (limp->rlim_cur > limp->rlim_max)
690 		limp->rlim_cur = limp->rlim_max;
691 	lim_copy(newlim, oldlim);
692 	alimp = &newlim->pl_rlimit[which];
693 
694 	switch (which) {
695 	case RLIMIT_CPU:
696 		if (limp->rlim_cur != RLIM_INFINITY &&
697 		    p->p_cpulimit == RLIM_INFINITY)
698 			callout_reset_sbt(&p->p_limco, SBT_1S, 0,
699 			    lim_cb, p, C_PREL(1));
700 		p->p_cpulimit = limp->rlim_cur;
701 		break;
702 	case RLIMIT_DATA:
703 		if (limp->rlim_cur > maxdsiz)
704 			limp->rlim_cur = maxdsiz;
705 		if (limp->rlim_max > maxdsiz)
706 			limp->rlim_max = maxdsiz;
707 		break;
708 
709 	case RLIMIT_STACK:
710 		if (limp->rlim_cur > maxssiz)
711 			limp->rlim_cur = maxssiz;
712 		if (limp->rlim_max > maxssiz)
713 			limp->rlim_max = maxssiz;
714 		oldssiz = *alimp;
715 		if (p->p_sysent->sv_fixlimit != NULL)
716 			p->p_sysent->sv_fixlimit(&oldssiz,
717 			    RLIMIT_STACK);
718 		break;
719 
720 	case RLIMIT_NOFILE:
721 		if (limp->rlim_cur > maxfilesperproc)
722 			limp->rlim_cur = maxfilesperproc;
723 		if (limp->rlim_max > maxfilesperproc)
724 			limp->rlim_max = maxfilesperproc;
725 		break;
726 
727 	case RLIMIT_NPROC:
728 		if (limp->rlim_cur > maxprocperuid)
729 			limp->rlim_cur = maxprocperuid;
730 		if (limp->rlim_max > maxprocperuid)
731 			limp->rlim_max = maxprocperuid;
732 		if (limp->rlim_cur < 1)
733 			limp->rlim_cur = 1;
734 		if (limp->rlim_max < 1)
735 			limp->rlim_max = 1;
736 		break;
737 	}
738 	if (p->p_sysent->sv_fixlimit != NULL)
739 		p->p_sysent->sv_fixlimit(limp, which);
740 	*alimp = *limp;
741 	p->p_limit = newlim;
742 	PROC_UPDATE_COW(p);
743 	PROC_UNLOCK(p);
744 	lim_free(oldlim);
745 
746 	if (which == RLIMIT_STACK &&
747 	    /*
748 	     * Skip calls from exec_new_vmspace(), done when stack is
749 	     * not mapped yet.
750 	     */
751 	    (td != curthread || (p->p_flag & P_INEXEC) == 0)) {
752 		/*
753 		 * Stack is allocated to the max at exec time with only
754 		 * "rlim_cur" bytes accessible.  If stack limit is going
755 		 * up make more accessible, if going down make inaccessible.
756 		 */
757 		if (limp->rlim_cur != oldssiz.rlim_cur) {
758 			vm_offset_t addr;
759 			vm_size_t size;
760 			vm_prot_t prot;
761 
762 			if (limp->rlim_cur > oldssiz.rlim_cur) {
763 				prot = p->p_sysent->sv_stackprot;
764 				size = limp->rlim_cur - oldssiz.rlim_cur;
765 				addr = p->p_sysent->sv_usrstack -
766 				    limp->rlim_cur;
767 			} else {
768 				prot = VM_PROT_NONE;
769 				size = oldssiz.rlim_cur - limp->rlim_cur;
770 				addr = p->p_sysent->sv_usrstack -
771 				    oldssiz.rlim_cur;
772 			}
773 			addr = trunc_page(addr);
774 			size = round_page(size);
775 			(void)vm_map_protect(&p->p_vmspace->vm_map,
776 			    addr, addr + size, prot, 0,
777 			    VM_MAP_PROTECT_SET_PROT);
778 		}
779 	}
780 
781 	return (0);
782 }
783 
784 #ifndef _SYS_SYSPROTO_H_
785 struct __getrlimit_args {
786 	u_int	which;
787 	struct	rlimit *rlp;
788 };
789 #endif
790 /* ARGSUSED */
791 int
792 sys_getrlimit(struct thread *td, struct __getrlimit_args *uap)
793 {
794 	struct rlimit rlim;
795 	int error;
796 
797 	if (uap->which >= RLIM_NLIMITS)
798 		return (EINVAL);
799 	lim_rlimit(td, uap->which, &rlim);
800 	error = copyout(&rlim, uap->rlp, sizeof(struct rlimit));
801 	return (error);
802 }
803 
804 /*
805  * Transform the running time and tick information for children of proc p
806  * into user and system time usage.
807  */
808 void
809 calccru(struct proc *p, struct timeval *up, struct timeval *sp)
810 {
811 
812 	PROC_LOCK_ASSERT(p, MA_OWNED);
813 	calcru1(p, &p->p_crux, up, sp);
814 }
815 
816 /*
817  * Transform the running time and tick information in proc p into user
818  * and system time usage.  If appropriate, include the current time slice
819  * on this CPU.
820  */
821 void
822 calcru(struct proc *p, struct timeval *up, struct timeval *sp)
823 {
824 	struct thread *td;
825 	uint64_t runtime, u;
826 
827 	PROC_LOCK_ASSERT(p, MA_OWNED);
828 	PROC_STATLOCK_ASSERT(p, MA_OWNED);
829 	/*
830 	 * If we are getting stats for the current process, then add in the
831 	 * stats that this thread has accumulated in its current time slice.
832 	 * We reset the thread and CPU state as if we had performed a context
833 	 * switch right here.
834 	 */
835 	td = curthread;
836 	if (td->td_proc == p) {
837 		u = cpu_ticks();
838 		runtime = u - PCPU_GET(switchtime);
839 		td->td_runtime += runtime;
840 		td->td_incruntime += runtime;
841 		PCPU_SET(switchtime, u);
842 	}
843 	/* Make sure the per-thread stats are current. */
844 	FOREACH_THREAD_IN_PROC(p, td) {
845 		if (td->td_incruntime == 0)
846 			continue;
847 		ruxagg(p, td);
848 	}
849 	calcru1(p, &p->p_rux, up, sp);
850 }
851 
852 /* Collect resource usage for a single thread. */
853 void
854 rufetchtd(struct thread *td, struct rusage *ru)
855 {
856 	struct proc *p;
857 	uint64_t runtime, u;
858 
859 	p = td->td_proc;
860 	PROC_STATLOCK_ASSERT(p, MA_OWNED);
861 	THREAD_LOCK_ASSERT(td, MA_OWNED);
862 	/*
863 	 * If we are getting stats for the current thread, then add in the
864 	 * stats that this thread has accumulated in its current time slice.
865 	 * We reset the thread and CPU state as if we had performed a context
866 	 * switch right here.
867 	 */
868 	if (td == curthread) {
869 		u = cpu_ticks();
870 		runtime = u - PCPU_GET(switchtime);
871 		td->td_runtime += runtime;
872 		td->td_incruntime += runtime;
873 		PCPU_SET(switchtime, u);
874 	}
875 	ruxagg_locked(p, td);
876 	*ru = td->td_ru;
877 	calcru1(p, &td->td_rux, &ru->ru_utime, &ru->ru_stime);
878 }
879 
880 /* XXX: the MI version is too slow to use: */
881 #ifndef __HAVE_INLINE_FLSLL
882 #define	flsll(x)	(fls((x) >> 32) != 0 ? fls((x) >> 32) + 32 : fls(x))
883 #endif
884 
885 static uint64_t
886 mul64_by_fraction(uint64_t a, uint64_t b, uint64_t c)
887 {
888 	uint64_t acc, bh, bl;
889 	int i, s, sa, sb;
890 
891 	/*
892 	 * Calculate (a * b) / c accurately enough without overflowing.  c
893 	 * must be nonzero, and its top bit must be 0.  a or b must be
894 	 * <= c, and the implementation is tuned for b <= c.
895 	 *
896 	 * The comments about times are for use in calcru1() with units of
897 	 * microseconds for 'a' and stathz ticks at 128 Hz for b and c.
898 	 *
899 	 * Let n be the number of top zero bits in c.  Each iteration
900 	 * either returns, or reduces b by right shifting it by at least n.
901 	 * The number of iterations is at most 1 + 64 / n, and the error is
902 	 * at most the number of iterations.
903 	 *
904 	 * It is very unusual to need even 2 iterations.  Previous
905 	 * implementations overflowed essentially by returning early in the
906 	 * first iteration, with n = 38 giving overflow at 105+ hours and
907 	 * n = 32 giving overlow at at 388+ days despite a more careful
908 	 * calculation.  388 days is a reasonable uptime, and the calculation
909 	 * needs to work for the uptime times the number of CPUs since 'a'
910 	 * is per-process.
911 	 */
912 	if (a >= (uint64_t)1 << 63)
913 		return (0);		/* Unsupported arg -- can't happen. */
914 	acc = 0;
915 	for (i = 0; i < 128; i++) {
916 		sa = flsll(a);
917 		sb = flsll(b);
918 		if (sa + sb <= 64)
919 			/* Up to 105 hours on first iteration. */
920 			return (acc + (a * b) / c);
921 		if (a >= c) {
922 			/*
923 			 * This reduction is based on a = q * c + r, with the
924 			 * remainder r < c.  'a' may be large to start, and
925 			 * moving bits from b into 'a' at the end of the loop
926 			 * sets the top bit of 'a', so the reduction makes
927 			 * significant progress.
928 			 */
929 			acc += (a / c) * b;
930 			a %= c;
931 			sa = flsll(a);
932 			if (sa + sb <= 64)
933 				/* Up to 388 days on first iteration. */
934 				return (acc + (a * b) / c);
935 		}
936 
937 		/*
938 		 * This step writes a * b as a * ((bh << s) + bl) =
939 		 * a * (bh << s) + a * bl = (a << s) * bh + a * bl.  The 2
940 		 * additive terms are handled separately.  Splitting in
941 		 * this way is linear except for rounding errors.
942 		 *
943 		 * s = 64 - sa is the maximum such that a << s fits in 64
944 		 * bits.  Since a < c and c has at least 1 zero top bit,
945 		 * sa < 64 and s > 0.  Thus this step makes progress by
946 		 * reducing b (it increases 'a', but taking remainders on
947 		 * the next iteration completes the reduction).
948 		 *
949 		 * Finally, the choice for s is just what is needed to keep
950 		 * a * bl from overflowing, so we don't need complications
951 		 * like a recursive call mul64_by_fraction(a, bl, c) to
952 		 * handle the second additive term.
953 		 */
954 		s = 64 - sa;
955 		bh = b >> s;
956 		bl = b - (bh << s);
957 		acc += (a * bl) / c;
958 		a <<= s;
959 		b = bh;
960 	}
961 	return (0);		/* Algorithm failure -- can't happen. */
962 }
963 
964 static void
965 calcru1(struct proc *p, struct rusage_ext *ruxp, struct timeval *up,
966     struct timeval *sp)
967 {
968 	/* {user, system, interrupt, total} {ticks, usec}: */
969 	uint64_t ut, uu, st, su, it, tt, tu;
970 
971 	ut = ruxp->rux_uticks;
972 	st = ruxp->rux_sticks;
973 	it = ruxp->rux_iticks;
974 	tt = ut + st + it;
975 	if (tt == 0) {
976 		/* Avoid divide by zero */
977 		st = 1;
978 		tt = 1;
979 	}
980 	tu = cputick2usec(ruxp->rux_runtime);
981 	if ((int64_t)tu < 0) {
982 		/* XXX: this should be an assert /phk */
983 		printf("calcru: negative runtime of %jd usec for pid %d (%s)\n",
984 		    (intmax_t)tu, p->p_pid, p->p_comm);
985 		tu = ruxp->rux_tu;
986 	}
987 
988 	/* Subdivide tu.  Avoid overflow in the multiplications. */
989 	if (__predict_true(tu <= ((uint64_t)1 << 38) && tt <= (1 << 26))) {
990 		/* Up to 76 hours when stathz is 128. */
991 		uu = (tu * ut) / tt;
992 		su = (tu * st) / tt;
993 	} else {
994 		uu = mul64_by_fraction(tu, ut, tt);
995 		su = mul64_by_fraction(tu, st, tt);
996 	}
997 
998 	if (tu >= ruxp->rux_tu) {
999 		/*
1000 		 * The normal case, time increased.
1001 		 * Enforce monotonicity of bucketed numbers.
1002 		 */
1003 		if (uu < ruxp->rux_uu)
1004 			uu = ruxp->rux_uu;
1005 		if (su < ruxp->rux_su)
1006 			su = ruxp->rux_su;
1007 	} else if (tu + 3 > ruxp->rux_tu || 101 * tu > 100 * ruxp->rux_tu) {
1008 		/*
1009 		 * When we calibrate the cputicker, it is not uncommon to
1010 		 * see the presumably fixed frequency increase slightly over
1011 		 * time as a result of thermal stabilization and NTP
1012 		 * discipline (of the reference clock).  We therefore ignore
1013 		 * a bit of backwards slop because we  expect to catch up
1014 		 * shortly.  We use a 3 microsecond limit to catch low
1015 		 * counts and a 1% limit for high counts.
1016 		 */
1017 		uu = ruxp->rux_uu;
1018 		su = ruxp->rux_su;
1019 		tu = ruxp->rux_tu;
1020 	} else { /* tu < ruxp->rux_tu */
1021 		/*
1022 		 * What happened here was likely that a laptop, which ran at
1023 		 * a reduced clock frequency at boot, kicked into high gear.
1024 		 * The wisdom of spamming this message in that case is
1025 		 * dubious, but it might also be indicative of something
1026 		 * serious, so lets keep it and hope laptops can be made
1027 		 * more truthful about their CPU speed via ACPI.
1028 		 */
1029 		printf("calcru: runtime went backwards from %ju usec "
1030 		    "to %ju usec for pid %d (%s)\n",
1031 		    (uintmax_t)ruxp->rux_tu, (uintmax_t)tu,
1032 		    p->p_pid, p->p_comm);
1033 	}
1034 
1035 	ruxp->rux_uu = uu;
1036 	ruxp->rux_su = su;
1037 	ruxp->rux_tu = tu;
1038 
1039 	up->tv_sec = uu / 1000000;
1040 	up->tv_usec = uu % 1000000;
1041 	sp->tv_sec = su / 1000000;
1042 	sp->tv_usec = su % 1000000;
1043 }
1044 
1045 #ifndef _SYS_SYSPROTO_H_
1046 struct getrusage_args {
1047 	int	who;
1048 	struct	rusage *rusage;
1049 };
1050 #endif
1051 int
1052 sys_getrusage(struct thread *td, struct getrusage_args *uap)
1053 {
1054 	struct rusage ru;
1055 	int error;
1056 
1057 	error = kern_getrusage(td, uap->who, &ru);
1058 	if (error == 0)
1059 		error = copyout(&ru, uap->rusage, sizeof(struct rusage));
1060 	return (error);
1061 }
1062 
1063 int
1064 kern_getrusage(struct thread *td, int who, struct rusage *rup)
1065 {
1066 	struct proc *p;
1067 	int error;
1068 
1069 	error = 0;
1070 	p = td->td_proc;
1071 	PROC_LOCK(p);
1072 	switch (who) {
1073 	case RUSAGE_SELF:
1074 		rufetchcalc(p, rup, &rup->ru_utime,
1075 		    &rup->ru_stime);
1076 		break;
1077 
1078 	case RUSAGE_CHILDREN:
1079 		*rup = p->p_stats->p_cru;
1080 		calccru(p, &rup->ru_utime, &rup->ru_stime);
1081 		break;
1082 
1083 	case RUSAGE_THREAD:
1084 		PROC_STATLOCK(p);
1085 		thread_lock(td);
1086 		rufetchtd(td, rup);
1087 		thread_unlock(td);
1088 		PROC_STATUNLOCK(p);
1089 		break;
1090 
1091 	default:
1092 		error = EINVAL;
1093 	}
1094 	PROC_UNLOCK(p);
1095 	return (error);
1096 }
1097 
1098 void
1099 rucollect(struct rusage *ru, struct rusage *ru2)
1100 {
1101 	long *ip, *ip2;
1102 	int i;
1103 
1104 	if (ru->ru_maxrss < ru2->ru_maxrss)
1105 		ru->ru_maxrss = ru2->ru_maxrss;
1106 	ip = &ru->ru_first;
1107 	ip2 = &ru2->ru_first;
1108 	for (i = &ru->ru_last - &ru->ru_first; i >= 0; i--)
1109 		*ip++ += *ip2++;
1110 }
1111 
1112 void
1113 ruadd(struct rusage *ru, struct rusage_ext *rux, struct rusage *ru2,
1114     struct rusage_ext *rux2)
1115 {
1116 
1117 	rux->rux_runtime += rux2->rux_runtime;
1118 	rux->rux_uticks += rux2->rux_uticks;
1119 	rux->rux_sticks += rux2->rux_sticks;
1120 	rux->rux_iticks += rux2->rux_iticks;
1121 	rux->rux_uu += rux2->rux_uu;
1122 	rux->rux_su += rux2->rux_su;
1123 	rux->rux_tu += rux2->rux_tu;
1124 	rucollect(ru, ru2);
1125 }
1126 
1127 /*
1128  * Aggregate tick counts into the proc's rusage_ext.
1129  */
1130 static void
1131 ruxagg_ext_locked(struct rusage_ext *rux, struct thread *td)
1132 {
1133 
1134 	rux->rux_runtime += td->td_incruntime;
1135 	rux->rux_uticks += td->td_uticks;
1136 	rux->rux_sticks += td->td_sticks;
1137 	rux->rux_iticks += td->td_iticks;
1138 }
1139 
1140 void
1141 ruxagg_locked(struct proc *p, struct thread *td)
1142 {
1143 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1144 	PROC_STATLOCK_ASSERT(td->td_proc, MA_OWNED);
1145 
1146 	ruxagg_ext_locked(&p->p_rux, td);
1147 	ruxagg_ext_locked(&td->td_rux, td);
1148 	td->td_incruntime = 0;
1149 	td->td_uticks = 0;
1150 	td->td_iticks = 0;
1151 	td->td_sticks = 0;
1152 }
1153 
1154 void
1155 ruxagg(struct proc *p, struct thread *td)
1156 {
1157 
1158 	thread_lock(td);
1159 	ruxagg_locked(p, td);
1160 	thread_unlock(td);
1161 }
1162 
1163 /*
1164  * Update the rusage_ext structure and fetch a valid aggregate rusage
1165  * for proc p if storage for one is supplied.
1166  */
1167 void
1168 rufetch(struct proc *p, struct rusage *ru)
1169 {
1170 	struct thread *td;
1171 
1172 	PROC_STATLOCK_ASSERT(p, MA_OWNED);
1173 
1174 	*ru = p->p_ru;
1175 	if (p->p_numthreads > 0)  {
1176 		FOREACH_THREAD_IN_PROC(p, td) {
1177 			ruxagg(p, td);
1178 			rucollect(ru, &td->td_ru);
1179 		}
1180 	}
1181 }
1182 
1183 /*
1184  * Atomically perform a rufetch and a calcru together.
1185  * Consumers, can safely assume the calcru is executed only once
1186  * rufetch is completed.
1187  */
1188 void
1189 rufetchcalc(struct proc *p, struct rusage *ru, struct timeval *up,
1190     struct timeval *sp)
1191 {
1192 
1193 	PROC_STATLOCK(p);
1194 	rufetch(p, ru);
1195 	calcru(p, up, sp);
1196 	PROC_STATUNLOCK(p);
1197 }
1198 
1199 /*
1200  * Allocate a new resource limits structure and initialize its
1201  * reference count and mutex pointer.
1202  */
1203 struct plimit *
1204 lim_alloc()
1205 {
1206 	struct plimit *limp;
1207 
1208 	limp = malloc(sizeof(struct plimit), M_PLIMIT, M_WAITOK);
1209 	refcount_init(&limp->pl_refcnt, 1);
1210 	return (limp);
1211 }
1212 
1213 struct plimit *
1214 lim_hold(struct plimit *limp)
1215 {
1216 
1217 	refcount_acquire(&limp->pl_refcnt);
1218 	return (limp);
1219 }
1220 
1221 void
1222 lim_fork(struct proc *p1, struct proc *p2)
1223 {
1224 
1225 	PROC_LOCK_ASSERT(p1, MA_OWNED);
1226 	PROC_LOCK_ASSERT(p2, MA_OWNED);
1227 
1228 	p2->p_limit = lim_hold(p1->p_limit);
1229 	callout_init_mtx(&p2->p_limco, &p2->p_mtx, 0);
1230 	if (p1->p_cpulimit != RLIM_INFINITY)
1231 		callout_reset_sbt(&p2->p_limco, SBT_1S, 0,
1232 		    lim_cb, p2, C_PREL(1));
1233 }
1234 
1235 void
1236 lim_free(struct plimit *limp)
1237 {
1238 
1239 	if (refcount_release(&limp->pl_refcnt))
1240 		free((void *)limp, M_PLIMIT);
1241 }
1242 
1243 void
1244 lim_freen(struct plimit *limp, int n)
1245 {
1246 
1247 	if (refcount_releasen(&limp->pl_refcnt, n))
1248 		free((void *)limp, M_PLIMIT);
1249 }
1250 
1251 /*
1252  * Make a copy of the plimit structure.
1253  * We share these structures copy-on-write after fork.
1254  */
1255 void
1256 lim_copy(struct plimit *dst, struct plimit *src)
1257 {
1258 
1259 	KASSERT(dst->pl_refcnt <= 1, ("lim_copy to shared limit"));
1260 	bcopy(src->pl_rlimit, dst->pl_rlimit, sizeof(src->pl_rlimit));
1261 }
1262 
1263 /*
1264  * Return the hard limit for a particular system resource.  The
1265  * which parameter specifies the index into the rlimit array.
1266  */
1267 rlim_t
1268 lim_max(struct thread *td, int which)
1269 {
1270 	struct rlimit rl;
1271 
1272 	lim_rlimit(td, which, &rl);
1273 	return (rl.rlim_max);
1274 }
1275 
1276 rlim_t
1277 lim_max_proc(struct proc *p, int which)
1278 {
1279 	struct rlimit rl;
1280 
1281 	lim_rlimit_proc(p, which, &rl);
1282 	return (rl.rlim_max);
1283 }
1284 
1285 /*
1286  * Return the current (soft) limit for a particular system resource.
1287  * The which parameter which specifies the index into the rlimit array
1288  */
1289 rlim_t
1290 (lim_cur)(struct thread *td, int which)
1291 {
1292 	struct rlimit rl;
1293 
1294 	lim_rlimit(td, which, &rl);
1295 	return (rl.rlim_cur);
1296 }
1297 
1298 rlim_t
1299 lim_cur_proc(struct proc *p, int which)
1300 {
1301 	struct rlimit rl;
1302 
1303 	lim_rlimit_proc(p, which, &rl);
1304 	return (rl.rlim_cur);
1305 }
1306 
1307 /*
1308  * Return a copy of the entire rlimit structure for the system limit
1309  * specified by 'which' in the rlimit structure pointed to by 'rlp'.
1310  */
1311 void
1312 lim_rlimit(struct thread *td, int which, struct rlimit *rlp)
1313 {
1314 	struct proc *p = td->td_proc;
1315 
1316 	MPASS(td == curthread);
1317 	KASSERT(which >= 0 && which < RLIM_NLIMITS,
1318 	    ("request for invalid resource limit"));
1319 	*rlp = td->td_limit->pl_rlimit[which];
1320 	if (p->p_sysent->sv_fixlimit != NULL)
1321 		p->p_sysent->sv_fixlimit(rlp, which);
1322 }
1323 
1324 void
1325 lim_rlimit_proc(struct proc *p, int which, struct rlimit *rlp)
1326 {
1327 
1328 	PROC_LOCK_ASSERT(p, MA_OWNED);
1329 	KASSERT(which >= 0 && which < RLIM_NLIMITS,
1330 	    ("request for invalid resource limit"));
1331 	*rlp = p->p_limit->pl_rlimit[which];
1332 	if (p->p_sysent->sv_fixlimit != NULL)
1333 		p->p_sysent->sv_fixlimit(rlp, which);
1334 }
1335 
1336 void
1337 uihashinit()
1338 {
1339 
1340 	uihashtbl = hashinit(maxproc / 16, M_UIDINFO, &uihash);
1341 	rw_init(&uihashtbl_lock, "uidinfo hash");
1342 }
1343 
1344 /*
1345  * Look up a uidinfo struct for the parameter uid.
1346  * uihashtbl_lock must be locked.
1347  * Increase refcount on uidinfo struct returned.
1348  */
1349 static struct uidinfo *
1350 uilookup(uid_t uid)
1351 {
1352 	struct uihashhead *uipp;
1353 	struct uidinfo *uip;
1354 
1355 	rw_assert(&uihashtbl_lock, RA_LOCKED);
1356 	uipp = UIHASH(uid);
1357 	LIST_FOREACH(uip, uipp, ui_hash)
1358 		if (uip->ui_uid == uid) {
1359 			uihold(uip);
1360 			break;
1361 		}
1362 
1363 	return (uip);
1364 }
1365 
1366 /*
1367  * Find or allocate a struct uidinfo for a particular uid.
1368  * Returns with uidinfo struct referenced.
1369  * uifree() should be called on a struct uidinfo when released.
1370  */
1371 struct uidinfo *
1372 uifind(uid_t uid)
1373 {
1374 	struct uidinfo *new_uip, *uip;
1375 	struct ucred *cred;
1376 
1377 	cred = curthread->td_ucred;
1378 	if (cred->cr_uidinfo->ui_uid == uid) {
1379 		uip = cred->cr_uidinfo;
1380 		uihold(uip);
1381 		return (uip);
1382 	} else if (cred->cr_ruidinfo->ui_uid == uid) {
1383 		uip = cred->cr_ruidinfo;
1384 		uihold(uip);
1385 		return (uip);
1386 	}
1387 
1388 	rw_rlock(&uihashtbl_lock);
1389 	uip = uilookup(uid);
1390 	rw_runlock(&uihashtbl_lock);
1391 	if (uip != NULL)
1392 		return (uip);
1393 
1394 	new_uip = malloc(sizeof(*new_uip), M_UIDINFO, M_WAITOK | M_ZERO);
1395 	racct_create(&new_uip->ui_racct);
1396 	refcount_init(&new_uip->ui_ref, 1);
1397 	new_uip->ui_uid = uid;
1398 
1399 	rw_wlock(&uihashtbl_lock);
1400 	/*
1401 	 * There's a chance someone created our uidinfo while we
1402 	 * were in malloc and not holding the lock, so we have to
1403 	 * make sure we don't insert a duplicate uidinfo.
1404 	 */
1405 	if ((uip = uilookup(uid)) == NULL) {
1406 		LIST_INSERT_HEAD(UIHASH(uid), new_uip, ui_hash);
1407 		rw_wunlock(&uihashtbl_lock);
1408 		uip = new_uip;
1409 	} else {
1410 		rw_wunlock(&uihashtbl_lock);
1411 		racct_destroy(&new_uip->ui_racct);
1412 		free(new_uip, M_UIDINFO);
1413 	}
1414 	return (uip);
1415 }
1416 
1417 /*
1418  * Place another refcount on a uidinfo struct.
1419  */
1420 void
1421 uihold(struct uidinfo *uip)
1422 {
1423 
1424 	refcount_acquire(&uip->ui_ref);
1425 }
1426 
1427 /*-
1428  * Since uidinfo structs have a long lifetime, we use an
1429  * opportunistic refcounting scheme to avoid locking the lookup hash
1430  * for each release.
1431  *
1432  * If the refcount hits 0, we need to free the structure,
1433  * which means we need to lock the hash.
1434  * Optimal case:
1435  *   After locking the struct and lowering the refcount, if we find
1436  *   that we don't need to free, simply unlock and return.
1437  * Suboptimal case:
1438  *   If refcount lowering results in need to free, bump the count
1439  *   back up, lose the lock and acquire the locks in the proper
1440  *   order to try again.
1441  */
1442 void
1443 uifree(struct uidinfo *uip)
1444 {
1445 
1446 	if (refcount_release_if_not_last(&uip->ui_ref))
1447 		return;
1448 
1449 	rw_wlock(&uihashtbl_lock);
1450 	if (refcount_release(&uip->ui_ref) == 0) {
1451 		rw_wunlock(&uihashtbl_lock);
1452 		return;
1453 	}
1454 
1455 	racct_destroy(&uip->ui_racct);
1456 	LIST_REMOVE(uip, ui_hash);
1457 	rw_wunlock(&uihashtbl_lock);
1458 
1459 	if (uip->ui_sbsize != 0)
1460 		printf("freeing uidinfo: uid = %d, sbsize = %ld\n",
1461 		    uip->ui_uid, uip->ui_sbsize);
1462 	if (uip->ui_proccnt != 0)
1463 		printf("freeing uidinfo: uid = %d, proccnt = %ld\n",
1464 		    uip->ui_uid, uip->ui_proccnt);
1465 	if (uip->ui_vmsize != 0)
1466 		printf("freeing uidinfo: uid = %d, swapuse = %lld\n",
1467 		    uip->ui_uid, (unsigned long long)uip->ui_vmsize);
1468 	free(uip, M_UIDINFO);
1469 }
1470 
1471 #ifdef RACCT
1472 void
1473 ui_racct_foreach(void (*callback)(struct racct *racct,
1474     void *arg2, void *arg3), void (*pre)(void), void (*post)(void),
1475     void *arg2, void *arg3)
1476 {
1477 	struct uidinfo *uip;
1478 	struct uihashhead *uih;
1479 
1480 	rw_rlock(&uihashtbl_lock);
1481 	if (pre != NULL)
1482 		(pre)();
1483 	for (uih = &uihashtbl[uihash]; uih >= uihashtbl; uih--) {
1484 		LIST_FOREACH(uip, uih, ui_hash) {
1485 			(callback)(uip->ui_racct, arg2, arg3);
1486 		}
1487 	}
1488 	if (post != NULL)
1489 		(post)();
1490 	rw_runlock(&uihashtbl_lock);
1491 }
1492 #endif
1493 
1494 static inline int
1495 chglimit(struct uidinfo *uip, long *limit, int diff, rlim_t max, const char *name)
1496 {
1497 	long new;
1498 
1499 	/* Don't allow them to exceed max, but allow subtraction. */
1500 	new = atomic_fetchadd_long(limit, (long)diff) + diff;
1501 	if (diff > 0 && max != 0) {
1502 		if (new < 0 || new > max) {
1503 			atomic_subtract_long(limit, (long)diff);
1504 			return (0);
1505 		}
1506 	} else if (new < 0)
1507 		printf("negative %s for uid = %d\n", name, uip->ui_uid);
1508 	return (1);
1509 }
1510 
1511 /*
1512  * Change the count associated with number of processes
1513  * a given user is using.  When 'max' is 0, don't enforce a limit
1514  */
1515 int
1516 chgproccnt(struct uidinfo *uip, int diff, rlim_t max)
1517 {
1518 
1519 	return (chglimit(uip, &uip->ui_proccnt, diff, max, "proccnt"));
1520 }
1521 
1522 /*
1523  * Change the total socket buffer size a user has used.
1524  */
1525 int
1526 chgsbsize(struct uidinfo *uip, u_int *hiwat, u_int to, rlim_t max)
1527 {
1528 	int diff, rv;
1529 
1530 	diff = to - *hiwat;
1531 	if (diff > 0 && max == 0) {
1532 		rv = 0;
1533 	} else {
1534 		rv = chglimit(uip, &uip->ui_sbsize, diff, max, "sbsize");
1535 		if (rv != 0)
1536 			*hiwat = to;
1537 	}
1538 	return (rv);
1539 }
1540 
1541 /*
1542  * Change the count associated with number of pseudo-terminals
1543  * a given user is using.  When 'max' is 0, don't enforce a limit
1544  */
1545 int
1546 chgptscnt(struct uidinfo *uip, int diff, rlim_t max)
1547 {
1548 
1549 	return (chglimit(uip, &uip->ui_ptscnt, diff, max, "ptscnt"));
1550 }
1551 
1552 int
1553 chgkqcnt(struct uidinfo *uip, int diff, rlim_t max)
1554 {
1555 
1556 	return (chglimit(uip, &uip->ui_kqcnt, diff, max, "kqcnt"));
1557 }
1558 
1559 int
1560 chgumtxcnt(struct uidinfo *uip, int diff, rlim_t max)
1561 {
1562 
1563 	return (chglimit(uip, &uip->ui_umtxcnt, diff, max, "umtxcnt"));
1564 }
1565