xref: /freebsd/sys/kern/kern_resource.c (revision 180f822596ecc49d3074dcc9dfea9628aae1d48d)
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/umtx.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 		/* Only look up thread in current process */
319 		td1 = tdfind(uap->lwpid, curproc->p_pid);
320 		if (td1 == NULL)
321 			return (ESRCH);
322 		p = td1->td_proc;
323 	}
324 
325 	switch (uap->function) {
326 	case RTP_LOOKUP:
327 		if ((error = p_cansee(td, p)))
328 			break;
329 		pri_to_rtp(td1, &rtp);
330 		PROC_UNLOCK(p);
331 		return (copyout(&rtp, uap->rtp, sizeof(struct rtprio)));
332 	case RTP_SET:
333 		if ((error = p_cansched(td, p)) || (error = cierror))
334 			break;
335 
336 		/* Disallow setting rtprio in most cases if not superuser. */
337 
338 		/*
339 		 * Realtime priority has to be restricted for reasons which
340 		 * should be obvious.  However, for idleprio processes, there is
341 		 * a potential for system deadlock if an idleprio process gains
342 		 * a lock on a resource that other processes need (and the
343 		 * idleprio process can't run due to a CPU-bound normal
344 		 * process).  Fix me!  XXX
345 		 *
346 		 * This problem is not only related to idleprio process.
347 		 * A user level program can obtain a file lock and hold it
348 		 * indefinitely.  Additionally, without idleprio processes it is
349 		 * still conceivable that a program with low priority will never
350 		 * get to run.  In short, allowing this feature might make it
351 		 * easier to lock a resource indefinitely, but it is not the
352 		 * only thing that makes it possible.
353 		 */
354 		if (RTP_PRIO_BASE(rtp.type) == RTP_PRIO_REALTIME ||
355 		    (RTP_PRIO_BASE(rtp.type) == RTP_PRIO_IDLE &&
356 		    unprivileged_idprio == 0)) {
357 			error = priv_check(td, PRIV_SCHED_RTPRIO);
358 			if (error)
359 				break;
360 		}
361 		error = rtp_to_pri(&rtp, td1);
362 		break;
363 	default:
364 		error = EINVAL;
365 		break;
366 	}
367 	PROC_UNLOCK(p);
368 	return (error);
369 }
370 
371 /*
372  * Set realtime priority.
373  */
374 #ifndef _SYS_SYSPROTO_H_
375 struct rtprio_args {
376 	int		function;
377 	pid_t		pid;
378 	struct rtprio	*rtp;
379 };
380 #endif
381 int
382 sys_rtprio(struct thread *td, struct rtprio_args *uap)
383 {
384 	struct proc *p;
385 	struct thread *tdp;
386 	struct rtprio rtp;
387 	int cierror, error;
388 
389 	/* Perform copyin before acquiring locks if needed. */
390 	if (uap->function == RTP_SET)
391 		cierror = copyin(uap->rtp, &rtp, sizeof(struct rtprio));
392 	else
393 		cierror = 0;
394 
395 	if (uap->pid == 0) {
396 		p = td->td_proc;
397 		PROC_LOCK(p);
398 	} else {
399 		p = pfind(uap->pid);
400 		if (p == NULL)
401 			return (ESRCH);
402 	}
403 
404 	switch (uap->function) {
405 	case RTP_LOOKUP:
406 		if ((error = p_cansee(td, p)))
407 			break;
408 		/*
409 		 * Return OUR priority if no pid specified,
410 		 * or if one is, report the highest priority
411 		 * in the process.  There isn't much more you can do as
412 		 * there is only room to return a single priority.
413 		 * Note: specifying our own pid is not the same
414 		 * as leaving it zero.
415 		 */
416 		if (uap->pid == 0) {
417 			pri_to_rtp(td, &rtp);
418 		} else {
419 			struct rtprio rtp2;
420 
421 			rtp.type = RTP_PRIO_IDLE;
422 			rtp.prio = RTP_PRIO_MAX;
423 			FOREACH_THREAD_IN_PROC(p, tdp) {
424 				pri_to_rtp(tdp, &rtp2);
425 				if (rtp2.type <  rtp.type ||
426 				    (rtp2.type == rtp.type &&
427 				    rtp2.prio < rtp.prio)) {
428 					rtp.type = rtp2.type;
429 					rtp.prio = rtp2.prio;
430 				}
431 			}
432 		}
433 		PROC_UNLOCK(p);
434 		return (copyout(&rtp, uap->rtp, sizeof(struct rtprio)));
435 	case RTP_SET:
436 		if ((error = p_cansched(td, p)) || (error = cierror))
437 			break;
438 
439 		/*
440 		 * Disallow setting rtprio in most cases if not superuser.
441 		 * See the comment in sys_rtprio_thread about idprio
442 		 * threads holding a lock.
443 		 */
444 		if (RTP_PRIO_BASE(rtp.type) == RTP_PRIO_REALTIME ||
445 		    (RTP_PRIO_BASE(rtp.type) == RTP_PRIO_IDLE &&
446 		    !unprivileged_idprio)) {
447 			error = priv_check(td, PRIV_SCHED_RTPRIO);
448 			if (error)
449 				break;
450 		}
451 
452 		/*
453 		 * If we are setting our own priority, set just our
454 		 * thread but if we are doing another process,
455 		 * do all the threads on that process. If we
456 		 * specify our own pid we do the latter.
457 		 */
458 		if (uap->pid == 0) {
459 			error = rtp_to_pri(&rtp, td);
460 		} else {
461 			FOREACH_THREAD_IN_PROC(p, td) {
462 				if ((error = rtp_to_pri(&rtp, td)) != 0)
463 					break;
464 			}
465 		}
466 		break;
467 	default:
468 		error = EINVAL;
469 		break;
470 	}
471 	PROC_UNLOCK(p);
472 	return (error);
473 }
474 
475 int
476 rtp_to_pri(struct rtprio *rtp, struct thread *td)
477 {
478 	u_char  newpri, oldclass, oldpri;
479 
480 	switch (RTP_PRIO_BASE(rtp->type)) {
481 	case RTP_PRIO_REALTIME:
482 		if (rtp->prio > RTP_PRIO_MAX)
483 			return (EINVAL);
484 		newpri = PRI_MIN_REALTIME + rtp->prio;
485 		break;
486 	case RTP_PRIO_NORMAL:
487 		if (rtp->prio > (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE))
488 			return (EINVAL);
489 		newpri = PRI_MIN_TIMESHARE + rtp->prio;
490 		break;
491 	case RTP_PRIO_IDLE:
492 		if (rtp->prio > RTP_PRIO_MAX)
493 			return (EINVAL);
494 		newpri = PRI_MIN_IDLE + rtp->prio;
495 		break;
496 	default:
497 		return (EINVAL);
498 	}
499 
500 	thread_lock(td);
501 	oldclass = td->td_pri_class;
502 	sched_class(td, rtp->type);	/* XXX fix */
503 	oldpri = td->td_user_pri;
504 	sched_user_prio(td, newpri);
505 	if (td->td_user_pri != oldpri && (oldclass != RTP_PRIO_NORMAL ||
506 	    td->td_pri_class != RTP_PRIO_NORMAL))
507 		sched_prio(td, td->td_user_pri);
508 	if (TD_ON_UPILOCK(td) && oldpri != newpri) {
509 		critical_enter();
510 		thread_unlock(td);
511 		umtx_pi_adjust(td, oldpri);
512 		critical_exit();
513 	} else
514 		thread_unlock(td);
515 	return (0);
516 }
517 
518 void
519 pri_to_rtp(struct thread *td, struct rtprio *rtp)
520 {
521 
522 	thread_lock(td);
523 	switch (PRI_BASE(td->td_pri_class)) {
524 	case PRI_REALTIME:
525 		rtp->prio = td->td_base_user_pri - PRI_MIN_REALTIME;
526 		break;
527 	case PRI_TIMESHARE:
528 		rtp->prio = td->td_base_user_pri - PRI_MIN_TIMESHARE;
529 		break;
530 	case PRI_IDLE:
531 		rtp->prio = td->td_base_user_pri - PRI_MIN_IDLE;
532 		break;
533 	default:
534 		break;
535 	}
536 	rtp->type = td->td_pri_class;
537 	thread_unlock(td);
538 }
539 
540 #if defined(COMPAT_43)
541 #ifndef _SYS_SYSPROTO_H_
542 struct osetrlimit_args {
543 	u_int	which;
544 	struct	orlimit *rlp;
545 };
546 #endif
547 int
548 osetrlimit(struct thread *td, struct osetrlimit_args *uap)
549 {
550 	struct orlimit olim;
551 	struct rlimit lim;
552 	int error;
553 
554 	if ((error = copyin(uap->rlp, &olim, sizeof(struct orlimit))))
555 		return (error);
556 	lim.rlim_cur = olim.rlim_cur;
557 	lim.rlim_max = olim.rlim_max;
558 	error = kern_setrlimit(td, uap->which, &lim);
559 	return (error);
560 }
561 
562 #ifndef _SYS_SYSPROTO_H_
563 struct ogetrlimit_args {
564 	u_int	which;
565 	struct	orlimit *rlp;
566 };
567 #endif
568 int
569 ogetrlimit(struct thread *td, struct ogetrlimit_args *uap)
570 {
571 	struct orlimit olim;
572 	struct rlimit rl;
573 	int error;
574 
575 	if (uap->which >= RLIM_NLIMITS)
576 		return (EINVAL);
577 	lim_rlimit(td, uap->which, &rl);
578 
579 	/*
580 	 * XXX would be more correct to convert only RLIM_INFINITY to the
581 	 * old RLIM_INFINITY and fail with EOVERFLOW for other larger
582 	 * values.  Most 64->32 and 32->16 conversions, including not
583 	 * unimportant ones of uids are even more broken than what we
584 	 * do here (they blindly truncate).  We don't do this correctly
585 	 * here since we have little experience with EOVERFLOW yet.
586 	 * Elsewhere, getuid() can't fail...
587 	 */
588 	olim.rlim_cur = rl.rlim_cur > 0x7fffffff ? 0x7fffffff : rl.rlim_cur;
589 	olim.rlim_max = rl.rlim_max > 0x7fffffff ? 0x7fffffff : rl.rlim_max;
590 	error = copyout(&olim, uap->rlp, sizeof(olim));
591 	return (error);
592 }
593 #endif /* COMPAT_43 */
594 
595 #ifndef _SYS_SYSPROTO_H_
596 struct __setrlimit_args {
597 	u_int	which;
598 	struct	rlimit *rlp;
599 };
600 #endif
601 int
602 sys_setrlimit(struct thread *td, struct __setrlimit_args *uap)
603 {
604 	struct rlimit alim;
605 	int error;
606 
607 	if ((error = copyin(uap->rlp, &alim, sizeof(struct rlimit))))
608 		return (error);
609 	error = kern_setrlimit(td, uap->which, &alim);
610 	return (error);
611 }
612 
613 static void
614 lim_cb(void *arg)
615 {
616 	struct rlimit rlim;
617 	struct thread *td;
618 	struct proc *p;
619 
620 	p = arg;
621 	PROC_LOCK_ASSERT(p, MA_OWNED);
622 	/*
623 	 * Check if the process exceeds its cpu resource allocation.  If
624 	 * it reaches the max, arrange to kill the process in ast().
625 	 */
626 	if (p->p_cpulimit == RLIM_INFINITY)
627 		return;
628 	PROC_STATLOCK(p);
629 	FOREACH_THREAD_IN_PROC(p, td) {
630 		ruxagg(p, td);
631 	}
632 	PROC_STATUNLOCK(p);
633 	if (p->p_rux.rux_runtime > p->p_cpulimit * cpu_tickrate()) {
634 		lim_rlimit_proc(p, RLIMIT_CPU, &rlim);
635 		if (p->p_rux.rux_runtime >= rlim.rlim_max * cpu_tickrate()) {
636 			killproc(p, "exceeded maximum CPU limit");
637 		} else {
638 			if (p->p_cpulimit < rlim.rlim_max)
639 				p->p_cpulimit += 5;
640 			kern_psignal(p, SIGXCPU);
641 		}
642 	}
643 	if ((p->p_flag & P_WEXIT) == 0)
644 		callout_reset_sbt(&p->p_limco, SBT_1S, 0,
645 		    lim_cb, p, C_PREL(1));
646 }
647 
648 int
649 kern_setrlimit(struct thread *td, u_int which, struct rlimit *limp)
650 {
651 
652 	return (kern_proc_setrlimit(td, td->td_proc, which, limp));
653 }
654 
655 int
656 kern_proc_setrlimit(struct thread *td, struct proc *p, u_int which,
657     struct rlimit *limp)
658 {
659 	struct plimit *newlim, *oldlim;
660 	struct rlimit *alimp;
661 	struct rlimit oldssiz;
662 	int error;
663 
664 	if (which >= RLIM_NLIMITS)
665 		return (EINVAL);
666 
667 	/*
668 	 * Preserve historical bugs by treating negative limits as unsigned.
669 	 */
670 	if (limp->rlim_cur < 0)
671 		limp->rlim_cur = RLIM_INFINITY;
672 	if (limp->rlim_max < 0)
673 		limp->rlim_max = RLIM_INFINITY;
674 
675 	oldssiz.rlim_cur = 0;
676 	newlim = lim_alloc();
677 	PROC_LOCK(p);
678 	oldlim = p->p_limit;
679 	alimp = &oldlim->pl_rlimit[which];
680 	if (limp->rlim_cur > alimp->rlim_max ||
681 	    limp->rlim_max > alimp->rlim_max)
682 		if ((error = priv_check(td, PRIV_PROC_SETRLIMIT))) {
683 			PROC_UNLOCK(p);
684 			lim_free(newlim);
685 			return (error);
686 		}
687 	if (limp->rlim_cur > limp->rlim_max)
688 		limp->rlim_cur = limp->rlim_max;
689 	lim_copy(newlim, oldlim);
690 	alimp = &newlim->pl_rlimit[which];
691 
692 	switch (which) {
693 	case RLIMIT_CPU:
694 		if (limp->rlim_cur != RLIM_INFINITY &&
695 		    p->p_cpulimit == RLIM_INFINITY)
696 			callout_reset_sbt(&p->p_limco, SBT_1S, 0,
697 			    lim_cb, p, C_PREL(1));
698 		p->p_cpulimit = limp->rlim_cur;
699 		break;
700 	case RLIMIT_DATA:
701 		if (limp->rlim_cur > maxdsiz)
702 			limp->rlim_cur = maxdsiz;
703 		if (limp->rlim_max > maxdsiz)
704 			limp->rlim_max = maxdsiz;
705 		break;
706 
707 	case RLIMIT_STACK:
708 		if (limp->rlim_cur > maxssiz)
709 			limp->rlim_cur = maxssiz;
710 		if (limp->rlim_max > maxssiz)
711 			limp->rlim_max = maxssiz;
712 		oldssiz = *alimp;
713 		if (p->p_sysent->sv_fixlimit != NULL)
714 			p->p_sysent->sv_fixlimit(&oldssiz,
715 			    RLIMIT_STACK);
716 		break;
717 
718 	case RLIMIT_NOFILE:
719 		if (limp->rlim_cur > maxfilesperproc)
720 			limp->rlim_cur = maxfilesperproc;
721 		if (limp->rlim_max > maxfilesperproc)
722 			limp->rlim_max = maxfilesperproc;
723 		break;
724 
725 	case RLIMIT_NPROC:
726 		if (limp->rlim_cur > maxprocperuid)
727 			limp->rlim_cur = maxprocperuid;
728 		if (limp->rlim_max > maxprocperuid)
729 			limp->rlim_max = maxprocperuid;
730 		if (limp->rlim_cur < 1)
731 			limp->rlim_cur = 1;
732 		if (limp->rlim_max < 1)
733 			limp->rlim_max = 1;
734 		break;
735 	}
736 	if (p->p_sysent->sv_fixlimit != NULL)
737 		p->p_sysent->sv_fixlimit(limp, which);
738 	*alimp = *limp;
739 	p->p_limit = newlim;
740 	PROC_UPDATE_COW(p);
741 	PROC_UNLOCK(p);
742 	lim_free(oldlim);
743 
744 	if (which == RLIMIT_STACK &&
745 	    /*
746 	     * Skip calls from exec_new_vmspace(), done when stack is
747 	     * not mapped yet.
748 	     */
749 	    (td != curthread || (p->p_flag & P_INEXEC) == 0)) {
750 		/*
751 		 * Stack is allocated to the max at exec time with only
752 		 * "rlim_cur" bytes accessible.  If stack limit is going
753 		 * up make more accessible, if going down make inaccessible.
754 		 */
755 		if (limp->rlim_cur != oldssiz.rlim_cur) {
756 			vm_offset_t addr;
757 			vm_size_t size;
758 			vm_prot_t prot;
759 
760 			if (limp->rlim_cur > oldssiz.rlim_cur) {
761 				prot = p->p_sysent->sv_stackprot;
762 				size = limp->rlim_cur - oldssiz.rlim_cur;
763 				addr = p->p_sysent->sv_usrstack -
764 				    limp->rlim_cur;
765 			} else {
766 				prot = VM_PROT_NONE;
767 				size = oldssiz.rlim_cur - limp->rlim_cur;
768 				addr = p->p_sysent->sv_usrstack -
769 				    oldssiz.rlim_cur;
770 			}
771 			addr = trunc_page(addr);
772 			size = round_page(size);
773 			(void)vm_map_protect(&p->p_vmspace->vm_map,
774 			    addr, addr + size, prot, FALSE);
775 		}
776 	}
777 
778 	return (0);
779 }
780 
781 #ifndef _SYS_SYSPROTO_H_
782 struct __getrlimit_args {
783 	u_int	which;
784 	struct	rlimit *rlp;
785 };
786 #endif
787 /* ARGSUSED */
788 int
789 sys_getrlimit(struct thread *td, struct __getrlimit_args *uap)
790 {
791 	struct rlimit rlim;
792 	int error;
793 
794 	if (uap->which >= RLIM_NLIMITS)
795 		return (EINVAL);
796 	lim_rlimit(td, uap->which, &rlim);
797 	error = copyout(&rlim, uap->rlp, sizeof(struct rlimit));
798 	return (error);
799 }
800 
801 /*
802  * Transform the running time and tick information for children of proc p
803  * into user and system time usage.
804  */
805 void
806 calccru(struct proc *p, struct timeval *up, struct timeval *sp)
807 {
808 
809 	PROC_LOCK_ASSERT(p, MA_OWNED);
810 	calcru1(p, &p->p_crux, up, sp);
811 }
812 
813 /*
814  * Transform the running time and tick information in proc p into user
815  * and system time usage.  If appropriate, include the current time slice
816  * on this CPU.
817  */
818 void
819 calcru(struct proc *p, struct timeval *up, struct timeval *sp)
820 {
821 	struct thread *td;
822 	uint64_t runtime, u;
823 
824 	PROC_LOCK_ASSERT(p, MA_OWNED);
825 	PROC_STATLOCK_ASSERT(p, MA_OWNED);
826 	/*
827 	 * If we are getting stats for the current process, then add in the
828 	 * stats that this thread has accumulated in its current time slice.
829 	 * We reset the thread and CPU state as if we had performed a context
830 	 * switch right here.
831 	 */
832 	td = curthread;
833 	if (td->td_proc == p) {
834 		u = cpu_ticks();
835 		runtime = u - PCPU_GET(switchtime);
836 		td->td_runtime += runtime;
837 		td->td_incruntime += runtime;
838 		PCPU_SET(switchtime, u);
839 	}
840 	/* Make sure the per-thread stats are current. */
841 	FOREACH_THREAD_IN_PROC(p, td) {
842 		if (td->td_incruntime == 0)
843 			continue;
844 		ruxagg(p, td);
845 	}
846 	calcru1(p, &p->p_rux, up, sp);
847 }
848 
849 /* Collect resource usage for a single thread. */
850 void
851 rufetchtd(struct thread *td, struct rusage *ru)
852 {
853 	struct proc *p;
854 	uint64_t runtime, u;
855 
856 	p = td->td_proc;
857 	PROC_STATLOCK_ASSERT(p, MA_OWNED);
858 	THREAD_LOCK_ASSERT(td, MA_OWNED);
859 	/*
860 	 * If we are getting stats for the current thread, then add in the
861 	 * stats that this thread has accumulated in its current time slice.
862 	 * We reset the thread and CPU state as if we had performed a context
863 	 * switch right here.
864 	 */
865 	if (td == curthread) {
866 		u = cpu_ticks();
867 		runtime = u - PCPU_GET(switchtime);
868 		td->td_runtime += runtime;
869 		td->td_incruntime += runtime;
870 		PCPU_SET(switchtime, u);
871 	}
872 	ruxagg_locked(p, td);
873 	*ru = td->td_ru;
874 	calcru1(p, &td->td_rux, &ru->ru_utime, &ru->ru_stime);
875 }
876 
877 /* XXX: the MI version is too slow to use: */
878 #ifndef __HAVE_INLINE_FLSLL
879 #define	flsll(x)	(fls((x) >> 32) != 0 ? fls((x) >> 32) + 32 : fls(x))
880 #endif
881 
882 static uint64_t
883 mul64_by_fraction(uint64_t a, uint64_t b, uint64_t c)
884 {
885 	uint64_t acc, bh, bl;
886 	int i, s, sa, sb;
887 
888 	/*
889 	 * Calculate (a * b) / c accurately enough without overflowing.  c
890 	 * must be nonzero, and its top bit must be 0.  a or b must be
891 	 * <= c, and the implementation is tuned for b <= c.
892 	 *
893 	 * The comments about times are for use in calcru1() with units of
894 	 * microseconds for 'a' and stathz ticks at 128 Hz for b and c.
895 	 *
896 	 * Let n be the number of top zero bits in c.  Each iteration
897 	 * either returns, or reduces b by right shifting it by at least n.
898 	 * The number of iterations is at most 1 + 64 / n, and the error is
899 	 * at most the number of iterations.
900 	 *
901 	 * It is very unusual to need even 2 iterations.  Previous
902 	 * implementations overflowed essentially by returning early in the
903 	 * first iteration, with n = 38 giving overflow at 105+ hours and
904 	 * n = 32 giving overlow at at 388+ days despite a more careful
905 	 * calculation.  388 days is a reasonable uptime, and the calculation
906 	 * needs to work for the uptime times the number of CPUs since 'a'
907 	 * is per-process.
908 	 */
909 	if (a >= (uint64_t)1 << 63)
910 		return (0);		/* Unsupported arg -- can't happen. */
911 	acc = 0;
912 	for (i = 0; i < 128; i++) {
913 		sa = flsll(a);
914 		sb = flsll(b);
915 		if (sa + sb <= 64)
916 			/* Up to 105 hours on first iteration. */
917 			return (acc + (a * b) / c);
918 		if (a >= c) {
919 			/*
920 			 * This reduction is based on a = q * c + r, with the
921 			 * remainder r < c.  'a' may be large to start, and
922 			 * moving bits from b into 'a' at the end of the loop
923 			 * sets the top bit of 'a', so the reduction makes
924 			 * significant progress.
925 			 */
926 			acc += (a / c) * b;
927 			a %= c;
928 			sa = flsll(a);
929 			if (sa + sb <= 64)
930 				/* Up to 388 days on first iteration. */
931 				return (acc + (a * b) / c);
932 		}
933 
934 		/*
935 		 * This step writes a * b as a * ((bh << s) + bl) =
936 		 * a * (bh << s) + a * bl = (a << s) * bh + a * bl.  The 2
937 		 * additive terms are handled separately.  Splitting in
938 		 * this way is linear except for rounding errors.
939 		 *
940 		 * s = 64 - sa is the maximum such that a << s fits in 64
941 		 * bits.  Since a < c and c has at least 1 zero top bit,
942 		 * sa < 64 and s > 0.  Thus this step makes progress by
943 		 * reducing b (it increases 'a', but taking remainders on
944 		 * the next iteration completes the reduction).
945 		 *
946 		 * Finally, the choice for s is just what is needed to keep
947 		 * a * bl from overflowing, so we don't need complications
948 		 * like a recursive call mul64_by_fraction(a, bl, c) to
949 		 * handle the second additive term.
950 		 */
951 		s = 64 - sa;
952 		bh = b >> s;
953 		bl = b - (bh << s);
954 		acc += (a * bl) / c;
955 		a <<= s;
956 		b = bh;
957 	}
958 	return (0);		/* Algorithm failure -- can't happen. */
959 }
960 
961 static void
962 calcru1(struct proc *p, struct rusage_ext *ruxp, struct timeval *up,
963     struct timeval *sp)
964 {
965 	/* {user, system, interrupt, total} {ticks, usec}: */
966 	uint64_t ut, uu, st, su, it, tt, tu;
967 
968 	ut = ruxp->rux_uticks;
969 	st = ruxp->rux_sticks;
970 	it = ruxp->rux_iticks;
971 	tt = ut + st + it;
972 	if (tt == 0) {
973 		/* Avoid divide by zero */
974 		st = 1;
975 		tt = 1;
976 	}
977 	tu = cputick2usec(ruxp->rux_runtime);
978 	if ((int64_t)tu < 0) {
979 		/* XXX: this should be an assert /phk */
980 		printf("calcru: negative runtime of %jd usec for pid %d (%s)\n",
981 		    (intmax_t)tu, p->p_pid, p->p_comm);
982 		tu = ruxp->rux_tu;
983 	}
984 
985 	/* Subdivide tu.  Avoid overflow in the multiplications. */
986 	if (__predict_true(tu <= ((uint64_t)1 << 38) && tt <= (1 << 26))) {
987 		/* Up to 76 hours when stathz is 128. */
988 		uu = (tu * ut) / tt;
989 		su = (tu * st) / tt;
990 	} else {
991 		uu = mul64_by_fraction(tu, ut, tt);
992 		su = mul64_by_fraction(tu, st, tt);
993 	}
994 
995 	if (tu >= ruxp->rux_tu) {
996 		/*
997 		 * The normal case, time increased.
998 		 * Enforce monotonicity of bucketed numbers.
999 		 */
1000 		if (uu < ruxp->rux_uu)
1001 			uu = ruxp->rux_uu;
1002 		if (su < ruxp->rux_su)
1003 			su = ruxp->rux_su;
1004 	} else if (tu + 3 > ruxp->rux_tu || 101 * tu > 100 * ruxp->rux_tu) {
1005 		/*
1006 		 * When we calibrate the cputicker, it is not uncommon to
1007 		 * see the presumably fixed frequency increase slightly over
1008 		 * time as a result of thermal stabilization and NTP
1009 		 * discipline (of the reference clock).  We therefore ignore
1010 		 * a bit of backwards slop because we  expect to catch up
1011 		 * shortly.  We use a 3 microsecond limit to catch low
1012 		 * counts and a 1% limit for high counts.
1013 		 */
1014 		uu = ruxp->rux_uu;
1015 		su = ruxp->rux_su;
1016 		tu = ruxp->rux_tu;
1017 	} else { /* tu < ruxp->rux_tu */
1018 		/*
1019 		 * What happened here was likely that a laptop, which ran at
1020 		 * a reduced clock frequency at boot, kicked into high gear.
1021 		 * The wisdom of spamming this message in that case is
1022 		 * dubious, but it might also be indicative of something
1023 		 * serious, so lets keep it and hope laptops can be made
1024 		 * more truthful about their CPU speed via ACPI.
1025 		 */
1026 		printf("calcru: runtime went backwards from %ju usec "
1027 		    "to %ju usec for pid %d (%s)\n",
1028 		    (uintmax_t)ruxp->rux_tu, (uintmax_t)tu,
1029 		    p->p_pid, p->p_comm);
1030 	}
1031 
1032 	ruxp->rux_uu = uu;
1033 	ruxp->rux_su = su;
1034 	ruxp->rux_tu = tu;
1035 
1036 	up->tv_sec = uu / 1000000;
1037 	up->tv_usec = uu % 1000000;
1038 	sp->tv_sec = su / 1000000;
1039 	sp->tv_usec = su % 1000000;
1040 }
1041 
1042 #ifndef _SYS_SYSPROTO_H_
1043 struct getrusage_args {
1044 	int	who;
1045 	struct	rusage *rusage;
1046 };
1047 #endif
1048 int
1049 sys_getrusage(struct thread *td, struct getrusage_args *uap)
1050 {
1051 	struct rusage ru;
1052 	int error;
1053 
1054 	error = kern_getrusage(td, uap->who, &ru);
1055 	if (error == 0)
1056 		error = copyout(&ru, uap->rusage, sizeof(struct rusage));
1057 	return (error);
1058 }
1059 
1060 int
1061 kern_getrusage(struct thread *td, int who, struct rusage *rup)
1062 {
1063 	struct proc *p;
1064 	int error;
1065 
1066 	error = 0;
1067 	p = td->td_proc;
1068 	PROC_LOCK(p);
1069 	switch (who) {
1070 	case RUSAGE_SELF:
1071 		rufetchcalc(p, rup, &rup->ru_utime,
1072 		    &rup->ru_stime);
1073 		break;
1074 
1075 	case RUSAGE_CHILDREN:
1076 		*rup = p->p_stats->p_cru;
1077 		calccru(p, &rup->ru_utime, &rup->ru_stime);
1078 		break;
1079 
1080 	case RUSAGE_THREAD:
1081 		PROC_STATLOCK(p);
1082 		thread_lock(td);
1083 		rufetchtd(td, rup);
1084 		thread_unlock(td);
1085 		PROC_STATUNLOCK(p);
1086 		break;
1087 
1088 	default:
1089 		error = EINVAL;
1090 	}
1091 	PROC_UNLOCK(p);
1092 	return (error);
1093 }
1094 
1095 void
1096 rucollect(struct rusage *ru, struct rusage *ru2)
1097 {
1098 	long *ip, *ip2;
1099 	int i;
1100 
1101 	if (ru->ru_maxrss < ru2->ru_maxrss)
1102 		ru->ru_maxrss = ru2->ru_maxrss;
1103 	ip = &ru->ru_first;
1104 	ip2 = &ru2->ru_first;
1105 	for (i = &ru->ru_last - &ru->ru_first; i >= 0; i--)
1106 		*ip++ += *ip2++;
1107 }
1108 
1109 void
1110 ruadd(struct rusage *ru, struct rusage_ext *rux, struct rusage *ru2,
1111     struct rusage_ext *rux2)
1112 {
1113 
1114 	rux->rux_runtime += rux2->rux_runtime;
1115 	rux->rux_uticks += rux2->rux_uticks;
1116 	rux->rux_sticks += rux2->rux_sticks;
1117 	rux->rux_iticks += rux2->rux_iticks;
1118 	rux->rux_uu += rux2->rux_uu;
1119 	rux->rux_su += rux2->rux_su;
1120 	rux->rux_tu += rux2->rux_tu;
1121 	rucollect(ru, ru2);
1122 }
1123 
1124 /*
1125  * Aggregate tick counts into the proc's rusage_ext.
1126  */
1127 static void
1128 ruxagg_ext_locked(struct rusage_ext *rux, struct thread *td)
1129 {
1130 
1131 	rux->rux_runtime += td->td_incruntime;
1132 	rux->rux_uticks += td->td_uticks;
1133 	rux->rux_sticks += td->td_sticks;
1134 	rux->rux_iticks += td->td_iticks;
1135 }
1136 
1137 void
1138 ruxagg_locked(struct proc *p, struct thread *td)
1139 {
1140 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1141 	PROC_STATLOCK_ASSERT(td->td_proc, MA_OWNED);
1142 
1143 	ruxagg_ext_locked(&p->p_rux, td);
1144 	ruxagg_ext_locked(&td->td_rux, td);
1145 	td->td_incruntime = 0;
1146 	td->td_uticks = 0;
1147 	td->td_iticks = 0;
1148 	td->td_sticks = 0;
1149 }
1150 
1151 void
1152 ruxagg(struct proc *p, struct thread *td)
1153 {
1154 
1155 	thread_lock(td);
1156 	ruxagg_locked(p, td);
1157 	thread_unlock(td);
1158 }
1159 
1160 /*
1161  * Update the rusage_ext structure and fetch a valid aggregate rusage
1162  * for proc p if storage for one is supplied.
1163  */
1164 void
1165 rufetch(struct proc *p, struct rusage *ru)
1166 {
1167 	struct thread *td;
1168 
1169 	PROC_STATLOCK_ASSERT(p, MA_OWNED);
1170 
1171 	*ru = p->p_ru;
1172 	if (p->p_numthreads > 0)  {
1173 		FOREACH_THREAD_IN_PROC(p, td) {
1174 			ruxagg(p, td);
1175 			rucollect(ru, &td->td_ru);
1176 		}
1177 	}
1178 }
1179 
1180 /*
1181  * Atomically perform a rufetch and a calcru together.
1182  * Consumers, can safely assume the calcru is executed only once
1183  * rufetch is completed.
1184  */
1185 void
1186 rufetchcalc(struct proc *p, struct rusage *ru, struct timeval *up,
1187     struct timeval *sp)
1188 {
1189 
1190 	PROC_STATLOCK(p);
1191 	rufetch(p, ru);
1192 	calcru(p, up, sp);
1193 	PROC_STATUNLOCK(p);
1194 }
1195 
1196 /*
1197  * Allocate a new resource limits structure and initialize its
1198  * reference count and mutex pointer.
1199  */
1200 struct plimit *
1201 lim_alloc()
1202 {
1203 	struct plimit *limp;
1204 
1205 	limp = malloc(sizeof(struct plimit), M_PLIMIT, M_WAITOK);
1206 	refcount_init(&limp->pl_refcnt, 1);
1207 	return (limp);
1208 }
1209 
1210 struct plimit *
1211 lim_hold(struct plimit *limp)
1212 {
1213 
1214 	refcount_acquire(&limp->pl_refcnt);
1215 	return (limp);
1216 }
1217 
1218 void
1219 lim_fork(struct proc *p1, struct proc *p2)
1220 {
1221 
1222 	PROC_LOCK_ASSERT(p1, MA_OWNED);
1223 	PROC_LOCK_ASSERT(p2, MA_OWNED);
1224 
1225 	p2->p_limit = lim_hold(p1->p_limit);
1226 	callout_init_mtx(&p2->p_limco, &p2->p_mtx, 0);
1227 	if (p1->p_cpulimit != RLIM_INFINITY)
1228 		callout_reset_sbt(&p2->p_limco, SBT_1S, 0,
1229 		    lim_cb, p2, C_PREL(1));
1230 }
1231 
1232 void
1233 lim_free(struct plimit *limp)
1234 {
1235 
1236 	if (refcount_release(&limp->pl_refcnt))
1237 		free((void *)limp, M_PLIMIT);
1238 }
1239 
1240 /*
1241  * Make a copy of the plimit structure.
1242  * We share these structures copy-on-write after fork.
1243  */
1244 void
1245 lim_copy(struct plimit *dst, struct plimit *src)
1246 {
1247 
1248 	KASSERT(dst->pl_refcnt <= 1, ("lim_copy to shared limit"));
1249 	bcopy(src->pl_rlimit, dst->pl_rlimit, sizeof(src->pl_rlimit));
1250 }
1251 
1252 /*
1253  * Return the hard limit for a particular system resource.  The
1254  * which parameter specifies the index into the rlimit array.
1255  */
1256 rlim_t
1257 lim_max(struct thread *td, int which)
1258 {
1259 	struct rlimit rl;
1260 
1261 	lim_rlimit(td, which, &rl);
1262 	return (rl.rlim_max);
1263 }
1264 
1265 rlim_t
1266 lim_max_proc(struct proc *p, int which)
1267 {
1268 	struct rlimit rl;
1269 
1270 	lim_rlimit_proc(p, which, &rl);
1271 	return (rl.rlim_max);
1272 }
1273 
1274 /*
1275  * Return the current (soft) limit for a particular system resource.
1276  * The which parameter which specifies the index into the rlimit array
1277  */
1278 rlim_t
1279 (lim_cur)(struct thread *td, int which)
1280 {
1281 	struct rlimit rl;
1282 
1283 	lim_rlimit(td, which, &rl);
1284 	return (rl.rlim_cur);
1285 }
1286 
1287 rlim_t
1288 lim_cur_proc(struct proc *p, int which)
1289 {
1290 	struct rlimit rl;
1291 
1292 	lim_rlimit_proc(p, which, &rl);
1293 	return (rl.rlim_cur);
1294 }
1295 
1296 /*
1297  * Return a copy of the entire rlimit structure for the system limit
1298  * specified by 'which' in the rlimit structure pointed to by 'rlp'.
1299  */
1300 void
1301 lim_rlimit(struct thread *td, int which, struct rlimit *rlp)
1302 {
1303 	struct proc *p = td->td_proc;
1304 
1305 	MPASS(td == curthread);
1306 	KASSERT(which >= 0 && which < RLIM_NLIMITS,
1307 	    ("request for invalid resource limit"));
1308 	*rlp = td->td_limit->pl_rlimit[which];
1309 	if (p->p_sysent->sv_fixlimit != NULL)
1310 		p->p_sysent->sv_fixlimit(rlp, which);
1311 }
1312 
1313 void
1314 lim_rlimit_proc(struct proc *p, int which, struct rlimit *rlp)
1315 {
1316 
1317 	PROC_LOCK_ASSERT(p, MA_OWNED);
1318 	KASSERT(which >= 0 && which < RLIM_NLIMITS,
1319 	    ("request for invalid resource limit"));
1320 	*rlp = p->p_limit->pl_rlimit[which];
1321 	if (p->p_sysent->sv_fixlimit != NULL)
1322 		p->p_sysent->sv_fixlimit(rlp, which);
1323 }
1324 
1325 void
1326 uihashinit()
1327 {
1328 
1329 	uihashtbl = hashinit(maxproc / 16, M_UIDINFO, &uihash);
1330 	rw_init(&uihashtbl_lock, "uidinfo hash");
1331 }
1332 
1333 /*
1334  * Look up a uidinfo struct for the parameter uid.
1335  * uihashtbl_lock must be locked.
1336  * Increase refcount on uidinfo struct returned.
1337  */
1338 static struct uidinfo *
1339 uilookup(uid_t uid)
1340 {
1341 	struct uihashhead *uipp;
1342 	struct uidinfo *uip;
1343 
1344 	rw_assert(&uihashtbl_lock, RA_LOCKED);
1345 	uipp = UIHASH(uid);
1346 	LIST_FOREACH(uip, uipp, ui_hash)
1347 		if (uip->ui_uid == uid) {
1348 			uihold(uip);
1349 			break;
1350 		}
1351 
1352 	return (uip);
1353 }
1354 
1355 /*
1356  * Find or allocate a struct uidinfo for a particular uid.
1357  * Returns with uidinfo struct referenced.
1358  * uifree() should be called on a struct uidinfo when released.
1359  */
1360 struct uidinfo *
1361 uifind(uid_t uid)
1362 {
1363 	struct uidinfo *new_uip, *uip;
1364 	struct ucred *cred;
1365 
1366 	cred = curthread->td_ucred;
1367 	if (cred->cr_uidinfo->ui_uid == uid) {
1368 		uip = cred->cr_uidinfo;
1369 		uihold(uip);
1370 		return (uip);
1371 	} else if (cred->cr_ruidinfo->ui_uid == uid) {
1372 		uip = cred->cr_ruidinfo;
1373 		uihold(uip);
1374 		return (uip);
1375 	}
1376 
1377 	rw_rlock(&uihashtbl_lock);
1378 	uip = uilookup(uid);
1379 	rw_runlock(&uihashtbl_lock);
1380 	if (uip != NULL)
1381 		return (uip);
1382 
1383 	new_uip = malloc(sizeof(*new_uip), M_UIDINFO, M_WAITOK | M_ZERO);
1384 	racct_create(&new_uip->ui_racct);
1385 	refcount_init(&new_uip->ui_ref, 1);
1386 	new_uip->ui_uid = uid;
1387 
1388 	rw_wlock(&uihashtbl_lock);
1389 	/*
1390 	 * There's a chance someone created our uidinfo while we
1391 	 * were in malloc and not holding the lock, so we have to
1392 	 * make sure we don't insert a duplicate uidinfo.
1393 	 */
1394 	if ((uip = uilookup(uid)) == NULL) {
1395 		LIST_INSERT_HEAD(UIHASH(uid), new_uip, ui_hash);
1396 		rw_wunlock(&uihashtbl_lock);
1397 		uip = new_uip;
1398 	} else {
1399 		rw_wunlock(&uihashtbl_lock);
1400 		racct_destroy(&new_uip->ui_racct);
1401 		free(new_uip, M_UIDINFO);
1402 	}
1403 	return (uip);
1404 }
1405 
1406 /*
1407  * Place another refcount on a uidinfo struct.
1408  */
1409 void
1410 uihold(struct uidinfo *uip)
1411 {
1412 
1413 	refcount_acquire(&uip->ui_ref);
1414 }
1415 
1416 /*-
1417  * Since uidinfo structs have a long lifetime, we use an
1418  * opportunistic refcounting scheme to avoid locking the lookup hash
1419  * for each release.
1420  *
1421  * If the refcount hits 0, we need to free the structure,
1422  * which means we need to lock the hash.
1423  * Optimal case:
1424  *   After locking the struct and lowering the refcount, if we find
1425  *   that we don't need to free, simply unlock and return.
1426  * Suboptimal case:
1427  *   If refcount lowering results in need to free, bump the count
1428  *   back up, lose the lock and acquire the locks in the proper
1429  *   order to try again.
1430  */
1431 void
1432 uifree(struct uidinfo *uip)
1433 {
1434 
1435 	if (refcount_release_if_not_last(&uip->ui_ref))
1436 		return;
1437 
1438 	rw_wlock(&uihashtbl_lock);
1439 	if (refcount_release(&uip->ui_ref) == 0) {
1440 		rw_wunlock(&uihashtbl_lock);
1441 		return;
1442 	}
1443 
1444 	racct_destroy(&uip->ui_racct);
1445 	LIST_REMOVE(uip, ui_hash);
1446 	rw_wunlock(&uihashtbl_lock);
1447 
1448 	if (uip->ui_sbsize != 0)
1449 		printf("freeing uidinfo: uid = %d, sbsize = %ld\n",
1450 		    uip->ui_uid, uip->ui_sbsize);
1451 	if (uip->ui_proccnt != 0)
1452 		printf("freeing uidinfo: uid = %d, proccnt = %ld\n",
1453 		    uip->ui_uid, uip->ui_proccnt);
1454 	if (uip->ui_vmsize != 0)
1455 		printf("freeing uidinfo: uid = %d, swapuse = %lld\n",
1456 		    uip->ui_uid, (unsigned long long)uip->ui_vmsize);
1457 	free(uip, M_UIDINFO);
1458 }
1459 
1460 #ifdef RACCT
1461 void
1462 ui_racct_foreach(void (*callback)(struct racct *racct,
1463     void *arg2, void *arg3), void (*pre)(void), void (*post)(void),
1464     void *arg2, void *arg3)
1465 {
1466 	struct uidinfo *uip;
1467 	struct uihashhead *uih;
1468 
1469 	rw_rlock(&uihashtbl_lock);
1470 	if (pre != NULL)
1471 		(pre)();
1472 	for (uih = &uihashtbl[uihash]; uih >= uihashtbl; uih--) {
1473 		LIST_FOREACH(uip, uih, ui_hash) {
1474 			(callback)(uip->ui_racct, arg2, arg3);
1475 		}
1476 	}
1477 	if (post != NULL)
1478 		(post)();
1479 	rw_runlock(&uihashtbl_lock);
1480 }
1481 #endif
1482 
1483 static inline int
1484 chglimit(struct uidinfo *uip, long *limit, int diff, rlim_t max, const char *name)
1485 {
1486 	long new;
1487 
1488 	/* Don't allow them to exceed max, but allow subtraction. */
1489 	new = atomic_fetchadd_long(limit, (long)diff) + diff;
1490 	if (diff > 0 && max != 0) {
1491 		if (new < 0 || new > max) {
1492 			atomic_subtract_long(limit, (long)diff);
1493 			return (0);
1494 		}
1495 	} else if (new < 0)
1496 		printf("negative %s for uid = %d\n", name, uip->ui_uid);
1497 	return (1);
1498 }
1499 
1500 /*
1501  * Change the count associated with number of processes
1502  * a given user is using.  When 'max' is 0, don't enforce a limit
1503  */
1504 int
1505 chgproccnt(struct uidinfo *uip, int diff, rlim_t max)
1506 {
1507 
1508 	return (chglimit(uip, &uip->ui_proccnt, diff, max, "proccnt"));
1509 }
1510 
1511 /*
1512  * Change the total socket buffer size a user has used.
1513  */
1514 int
1515 chgsbsize(struct uidinfo *uip, u_int *hiwat, u_int to, rlim_t max)
1516 {
1517 	int diff, rv;
1518 
1519 	diff = to - *hiwat;
1520 	if (diff > 0 && max == 0) {
1521 		rv = 0;
1522 	} else {
1523 		rv = chglimit(uip, &uip->ui_sbsize, diff, max, "sbsize");
1524 		if (rv != 0)
1525 			*hiwat = to;
1526 	}
1527 	return (rv);
1528 }
1529 
1530 /*
1531  * Change the count associated with number of pseudo-terminals
1532  * a given user is using.  When 'max' is 0, don't enforce a limit
1533  */
1534 int
1535 chgptscnt(struct uidinfo *uip, int diff, rlim_t max)
1536 {
1537 
1538 	return (chglimit(uip, &uip->ui_ptscnt, diff, max, "ptscnt"));
1539 }
1540 
1541 int
1542 chgkqcnt(struct uidinfo *uip, int diff, rlim_t max)
1543 {
1544 
1545 	return (chglimit(uip, &uip->ui_kqcnt, diff, max, "kqcnt"));
1546 }
1547 
1548 int
1549 chgumtxcnt(struct uidinfo *uip, int diff, rlim_t max)
1550 {
1551 
1552 	return (chglimit(uip, &uip->ui_umtxcnt, diff, max, "umtxcnt"));
1553 }
1554