xref: /openbsd-src/sys/kern/kern_synch.c (revision 13095e6dd0735888515c168c112f29db29ff12a2)
1 /*	$OpenBSD: kern_synch.c,v 1.199 2023/09/08 09:06:31 claudio Exp $	*/
2 /*	$NetBSD: kern_synch.c,v 1.37 1996/04/22 01:38:37 christos Exp $	*/
3 
4 /*
5  * Copyright (c) 1982, 1986, 1990, 1991, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  * (c) UNIX System Laboratories, Inc.
8  * All or some portions of this file are derived from material licensed
9  * to the University of California by American Telephone and Telegraph
10  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
11  * the permission of UNIX System Laboratories, Inc.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. Neither the name of the University nor the names of its contributors
22  *    may be used to endorse or promote products derived from this software
23  *    without specific prior written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
29  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  *
37  *	@(#)kern_synch.c	8.6 (Berkeley) 1/21/94
38  */
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/proc.h>
43 #include <sys/kernel.h>
44 #include <sys/signalvar.h>
45 #include <sys/sched.h>
46 #include <sys/timeout.h>
47 #include <sys/mount.h>
48 #include <sys/syscallargs.h>
49 #include <sys/refcnt.h>
50 #include <sys/atomic.h>
51 #include <sys/tracepoint.h>
52 
53 #include <ddb/db_output.h>
54 
55 #include <machine/spinlock.h>
56 
57 #ifdef DIAGNOSTIC
58 #include <sys/syslog.h>
59 #endif
60 
61 #ifdef KTRACE
62 #include <sys/ktrace.h>
63 #endif
64 
65 int	sleep_signal_check(void);
66 int	thrsleep(struct proc *, struct sys___thrsleep_args *);
67 int	thrsleep_unlock(void *);
68 
69 /*
70  * We're only looking at 7 bits of the address; everything is
71  * aligned to 4, lots of things are aligned to greater powers
72  * of 2.  Shift right by 8, i.e. drop the bottom 256 worth.
73  */
74 #define TABLESIZE	128
75 #define LOOKUP(x)	(((long)(x) >> 8) & (TABLESIZE - 1))
76 TAILQ_HEAD(slpque,proc) slpque[TABLESIZE];
77 
78 void
79 sleep_queue_init(void)
80 {
81 	int i;
82 
83 	for (i = 0; i < TABLESIZE; i++)
84 		TAILQ_INIT(&slpque[i]);
85 }
86 
87 /*
88  * Global sleep channel for threads that do not want to
89  * receive wakeup(9) broadcasts.
90  */
91 int nowake;
92 
93 /*
94  * During autoconfiguration or after a panic, a sleep will simply
95  * lower the priority briefly to allow interrupts, then return.
96  * The priority to be used (safepri) is machine-dependent, thus this
97  * value is initialized and maintained in the machine-dependent layers.
98  * This priority will typically be 0, or the lowest priority
99  * that is safe for use on the interrupt stack; it can be made
100  * higher to block network software interrupts after panics.
101  */
102 extern int safepri;
103 
104 /*
105  * General sleep call.  Suspends the current process until a wakeup is
106  * performed on the specified identifier.  The process will then be made
107  * runnable with the specified priority.  Sleeps at most timo/hz seconds
108  * (0 means no timeout).  If pri includes PCATCH flag, signals are checked
109  * before and after sleeping, else signals are not checked.  Returns 0 if
110  * awakened, EWOULDBLOCK if the timeout expires.  If PCATCH is set and a
111  * signal needs to be delivered, ERESTART is returned if the current system
112  * call should be restarted if possible, and EINTR is returned if the system
113  * call should be interrupted by the signal (return EINTR).
114  */
115 int
116 tsleep(const volatile void *ident, int priority, const char *wmesg, int timo)
117 {
118 #ifdef MULTIPROCESSOR
119 	int hold_count;
120 #endif
121 
122 	KASSERT((priority & ~(PRIMASK | PCATCH)) == 0);
123 	KASSERT(ident != &nowake || ISSET(priority, PCATCH) || timo != 0);
124 
125 #ifdef MULTIPROCESSOR
126 	KASSERT(ident == &nowake || timo || _kernel_lock_held());
127 #endif
128 
129 #ifdef DDB
130 	if (cold == 2)
131 		db_stack_dump();
132 #endif
133 	if (cold || panicstr) {
134 		int s;
135 		/*
136 		 * After a panic, or during autoconfiguration,
137 		 * just give interrupts a chance, then just return;
138 		 * don't run any other procs or panic below,
139 		 * in case this is the idle process and already asleep.
140 		 */
141 		s = splhigh();
142 		splx(safepri);
143 #ifdef MULTIPROCESSOR
144 		if (_kernel_lock_held()) {
145 			hold_count = __mp_release_all(&kernel_lock);
146 			__mp_acquire_count(&kernel_lock, hold_count);
147 		}
148 #endif
149 		splx(s);
150 		return (0);
151 	}
152 
153 	sleep_setup(ident, priority, wmesg);
154 	return sleep_finish(timo, 1);
155 }
156 
157 int
158 tsleep_nsec(const volatile void *ident, int priority, const char *wmesg,
159     uint64_t nsecs)
160 {
161 	uint64_t to_ticks;
162 
163 	if (nsecs == INFSLP)
164 		return tsleep(ident, priority, wmesg, 0);
165 #ifdef DIAGNOSTIC
166 	if (nsecs == 0) {
167 		log(LOG_WARNING,
168 		    "%s: %s[%d]: %s: trying to sleep zero nanoseconds\n",
169 		    __func__, curproc->p_p->ps_comm, curproc->p_p->ps_pid,
170 		    wmesg);
171 	}
172 #endif
173 	/*
174 	 * We want to sleep at least nsecs nanoseconds worth of ticks.
175 	 *
176 	 *  - Clamp nsecs to prevent arithmetic overflow.
177 	 *
178 	 *  - Round nsecs up to account for any nanoseconds that do not
179 	 *    divide evenly into tick_nsec, otherwise we'll lose them to
180 	 *    integer division in the next step.  We add (tick_nsec - 1)
181 	 *    to keep from introducing a spurious tick if there are no
182 	 *    such nanoseconds, i.e. nsecs % tick_nsec == 0.
183 	 *
184 	 *  - Divide the rounded value to a count of ticks.  We divide
185 	 *    by (tick_nsec + 1) to discard the extra tick introduced if,
186 	 *    before rounding, nsecs % tick_nsec == 1.
187 	 *
188 	 *  - Finally, add a tick to the result.  We need to wait out
189 	 *    the current tick before we can begin counting our interval,
190 	 *    as we do not know how much time has elapsed since the
191 	 *    current tick began.
192 	 */
193 	nsecs = MIN(nsecs, UINT64_MAX - tick_nsec);
194 	to_ticks = (nsecs + tick_nsec - 1) / (tick_nsec + 1) + 1;
195 	if (to_ticks > INT_MAX)
196 		to_ticks = INT_MAX;
197 	return tsleep(ident, priority, wmesg, (int)to_ticks);
198 }
199 
200 /*
201  * Same as tsleep, but if we have a mutex provided, then once we've
202  * entered the sleep queue we drop the mutex. After sleeping we re-lock.
203  */
204 int
205 msleep(const volatile void *ident, struct mutex *mtx, int priority,
206     const char *wmesg, int timo)
207 {
208 	int error, spl;
209 #ifdef MULTIPROCESSOR
210 	int hold_count;
211 #endif
212 
213 	KASSERT((priority & ~(PRIMASK | PCATCH | PNORELOCK)) == 0);
214 	KASSERT(ident != &nowake || ISSET(priority, PCATCH) || timo != 0);
215 	KASSERT(mtx != NULL);
216 
217 #ifdef DDB
218 	if (cold == 2)
219 		db_stack_dump();
220 #endif
221 	if (cold || panicstr) {
222 		/*
223 		 * After a panic, or during autoconfiguration,
224 		 * just give interrupts a chance, then just return;
225 		 * don't run any other procs or panic below,
226 		 * in case this is the idle process and already asleep.
227 		 */
228 		spl = MUTEX_OLDIPL(mtx);
229 		MUTEX_OLDIPL(mtx) = safepri;
230 		mtx_leave(mtx);
231 #ifdef MULTIPROCESSOR
232 		if (_kernel_lock_held()) {
233 			hold_count = __mp_release_all(&kernel_lock);
234 			__mp_acquire_count(&kernel_lock, hold_count);
235 		}
236 #endif
237 		if ((priority & PNORELOCK) == 0) {
238 			mtx_enter(mtx);
239 			MUTEX_OLDIPL(mtx) = spl;
240 		} else
241 			splx(spl);
242 		return (0);
243 	}
244 
245 	sleep_setup(ident, priority, wmesg);
246 
247 	mtx_leave(mtx);
248 	/* signal may stop the process, release mutex before that */
249 	error = sleep_finish(timo, 1);
250 
251 	if ((priority & PNORELOCK) == 0)
252 		mtx_enter(mtx);
253 
254 	return error;
255 }
256 
257 int
258 msleep_nsec(const volatile void *ident, struct mutex *mtx, int priority,
259     const char *wmesg, uint64_t nsecs)
260 {
261 	uint64_t to_ticks;
262 
263 	if (nsecs == INFSLP)
264 		return msleep(ident, mtx, priority, wmesg, 0);
265 #ifdef DIAGNOSTIC
266 	if (nsecs == 0) {
267 		log(LOG_WARNING,
268 		    "%s: %s[%d]: %s: trying to sleep zero nanoseconds\n",
269 		    __func__, curproc->p_p->ps_comm, curproc->p_p->ps_pid,
270 		    wmesg);
271 	}
272 #endif
273 	nsecs = MIN(nsecs, UINT64_MAX - tick_nsec);
274 	to_ticks = (nsecs + tick_nsec - 1) / (tick_nsec + 1) + 1;
275 	if (to_ticks > INT_MAX)
276 		to_ticks = INT_MAX;
277 	return msleep(ident, mtx, priority, wmesg, (int)to_ticks);
278 }
279 
280 /*
281  * Same as tsleep, but if we have a rwlock provided, then once we've
282  * entered the sleep queue we drop the it. After sleeping we re-lock.
283  */
284 int
285 rwsleep(const volatile void *ident, struct rwlock *rwl, int priority,
286     const char *wmesg, int timo)
287 {
288 	int error, status;
289 
290 	KASSERT((priority & ~(PRIMASK | PCATCH | PNORELOCK)) == 0);
291 	KASSERT(ident != &nowake || ISSET(priority, PCATCH) || timo != 0);
292 	KASSERT(ident != rwl);
293 	rw_assert_anylock(rwl);
294 	status = rw_status(rwl);
295 
296 	sleep_setup(ident, priority, wmesg);
297 
298 	rw_exit(rwl);
299 	/* signal may stop the process, release rwlock before that */
300 	error = sleep_finish(timo, 1);
301 
302 	if ((priority & PNORELOCK) == 0)
303 		rw_enter(rwl, status);
304 
305 	return error;
306 }
307 
308 int
309 rwsleep_nsec(const volatile void *ident, struct rwlock *rwl, int priority,
310     const char *wmesg, uint64_t nsecs)
311 {
312 	uint64_t to_ticks;
313 
314 	if (nsecs == INFSLP)
315 		return rwsleep(ident, rwl, priority, wmesg, 0);
316 #ifdef DIAGNOSTIC
317 	if (nsecs == 0) {
318 		log(LOG_WARNING,
319 		    "%s: %s[%d]: %s: trying to sleep zero nanoseconds\n",
320 		    __func__, curproc->p_p->ps_comm, curproc->p_p->ps_pid,
321 		    wmesg);
322 	}
323 #endif
324 	nsecs = MIN(nsecs, UINT64_MAX - tick_nsec);
325 	to_ticks = (nsecs + tick_nsec - 1) / (tick_nsec + 1) + 1;
326 	if (to_ticks > INT_MAX)
327 		to_ticks = INT_MAX;
328 	return 	rwsleep(ident, rwl, priority, wmesg, (int)to_ticks);
329 }
330 
331 void
332 sleep_setup(const volatile void *ident, int prio, const char *wmesg)
333 {
334 	struct proc *p = curproc;
335 	int s;
336 
337 #ifdef DIAGNOSTIC
338 	if (p->p_flag & P_CANTSLEEP)
339 		panic("sleep: %s failed insomnia", p->p_p->ps_comm);
340 	if (ident == NULL)
341 		panic("tsleep: no ident");
342 	if (p->p_stat != SONPROC)
343 		panic("tsleep: not SONPROC");
344 #endif
345 
346 	SCHED_LOCK(s);
347 
348 	TRACEPOINT(sched, sleep, NULL);
349 
350 	p->p_wchan = ident;
351 	p->p_wmesg = wmesg;
352 	p->p_slptime = 0;
353 	p->p_slppri = prio & PRIMASK;
354 	atomic_setbits_int(&p->p_flag, P_WSLEEP);
355 	TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_runq);
356 	if (prio & PCATCH)
357 		atomic_setbits_int(&p->p_flag, P_SINTR);
358 	p->p_stat = SSLEEP;
359 
360 	SCHED_UNLOCK(s);
361 }
362 
363 int
364 sleep_finish(int timo, int do_sleep)
365 {
366 	struct proc *p = curproc;
367 	int s, catch, error = 0, error1 = 0;
368 
369 	catch = p->p_flag & P_SINTR;
370 
371 	if (timo != 0) {
372 		KASSERT((p->p_flag & P_TIMEOUT) == 0);
373 		timeout_add(&p->p_sleep_to, timo);
374 	}
375 
376 	if (catch != 0) {
377 		/*
378 		 * We put ourselves on the sleep queue and start our
379 		 * timeout before calling sleep_signal_check(), as we could
380 		 * stop there, and a wakeup or a SIGCONT (or both) could
381 		 * occur while we were stopped.  A SIGCONT would cause
382 		 * us to be marked as SSLEEP without resuming us, thus
383 		 * we must be ready for sleep when sleep_signal_check() is
384 		 * called.
385 		 */
386 		if ((error = sleep_signal_check()) != 0) {
387 			catch = 0;
388 			do_sleep = 0;
389 		}
390 	}
391 
392 	SCHED_LOCK(s);
393 	/*
394 	 * If the wakeup happens while going to sleep, p->p_wchan
395 	 * will be NULL. In that case unwind immediately but still
396 	 * check for possible signals and timeouts.
397 	 */
398 	if (p->p_wchan == NULL)
399 		do_sleep = 0;
400 	atomic_clearbits_int(&p->p_flag, P_WSLEEP);
401 
402 	if (do_sleep) {
403 		KASSERT(p->p_stat == SSLEEP || p->p_stat == SSTOP);
404 		p->p_ru.ru_nvcsw++;
405 		mi_switch();
406 	} else {
407 		KASSERT(p->p_stat == SONPROC || p->p_stat == SSLEEP ||
408 		    p->p_stat == SSTOP);
409 		unsleep(p);
410 		p->p_stat = SONPROC;
411 	}
412 
413 #ifdef DIAGNOSTIC
414 	if (p->p_stat != SONPROC)
415 		panic("sleep_finish !SONPROC");
416 #endif
417 
418 	p->p_cpu->ci_schedstate.spc_curpriority = p->p_usrpri;
419 	SCHED_UNLOCK(s);
420 
421 	/*
422 	 * Even though this belongs to the signal handling part of sleep,
423 	 * we need to clear it before the ktrace.
424 	 */
425 	atomic_clearbits_int(&p->p_flag, P_SINTR);
426 
427 	if (timo != 0) {
428 		if (p->p_flag & P_TIMEOUT) {
429 			error1 = EWOULDBLOCK;
430 		} else {
431 			/* This can sleep. It must not use timeouts. */
432 			timeout_del_barrier(&p->p_sleep_to);
433 		}
434 		atomic_clearbits_int(&p->p_flag, P_TIMEOUT);
435 	}
436 
437 	/* Check if thread was woken up because of a unwind or signal */
438 	if (catch != 0)
439 		error = sleep_signal_check();
440 
441 	/* Signal errors are higher priority than timeouts. */
442 	if (error == 0 && error1 != 0)
443 		error = error1;
444 
445 	return error;
446 }
447 
448 /*
449  * Check and handle signals and suspensions around a sleep cycle.
450  */
451 int
452 sleep_signal_check(void)
453 {
454 	struct proc *p = curproc;
455 	struct sigctx ctx;
456 	int err, sig;
457 
458 	if ((err = single_thread_check(p, 1)) != 0)
459 		return err;
460 	if ((sig = cursig(p, &ctx)) != 0) {
461 		if (ctx.sig_intr)
462 			return EINTR;
463 		else
464 			return ERESTART;
465 	}
466 	return 0;
467 }
468 
469 int
470 wakeup_proc(struct proc *p, const volatile void *chan, int flags)
471 {
472 	int awakened = 0;
473 
474 	SCHED_ASSERT_LOCKED();
475 
476 	if (p->p_wchan != NULL &&
477 	   ((chan == NULL) || (p->p_wchan == chan))) {
478 		awakened = 1;
479 		if (flags)
480 			atomic_setbits_int(&p->p_flag, flags);
481 		if (p->p_stat == SSLEEP)
482 			setrunnable(p);
483 		else if (p->p_stat == SSTOP)
484 			unsleep(p);
485 #ifdef DIAGNOSTIC
486 		else
487 			panic("wakeup: p_stat is %d", (int)p->p_stat);
488 #endif
489 	}
490 
491 	return awakened;
492 }
493 
494 
495 /*
496  * Implement timeout for tsleep.
497  * If process hasn't been awakened (wchan non-zero),
498  * set timeout flag and undo the sleep.  If proc
499  * is stopped, just unsleep so it will remain stopped.
500  */
501 void
502 endtsleep(void *arg)
503 {
504 	struct proc *p = arg;
505 	int s;
506 
507 	SCHED_LOCK(s);
508 	wakeup_proc(p, NULL, P_TIMEOUT);
509 	SCHED_UNLOCK(s);
510 }
511 
512 /*
513  * Remove a process from its wait queue
514  */
515 void
516 unsleep(struct proc *p)
517 {
518 	SCHED_ASSERT_LOCKED();
519 
520 	if (p->p_wchan != NULL) {
521 		TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_runq);
522 		p->p_wchan = NULL;
523 		TRACEPOINT(sched, unsleep, p->p_tid + THREAD_PID_OFFSET,
524 		    p->p_p->ps_pid);
525 	}
526 }
527 
528 /*
529  * Make a number of processes sleeping on the specified identifier runnable.
530  */
531 void
532 wakeup_n(const volatile void *ident, int n)
533 {
534 	struct slpque *qp;
535 	struct proc *p;
536 	struct proc *pnext;
537 	int s;
538 
539 	SCHED_LOCK(s);
540 	qp = &slpque[LOOKUP(ident)];
541 	for (p = TAILQ_FIRST(qp); p != NULL && n != 0; p = pnext) {
542 		pnext = TAILQ_NEXT(p, p_runq);
543 #ifdef DIAGNOSTIC
544 		if (p->p_stat != SSLEEP && p->p_stat != SSTOP)
545 			panic("wakeup: p_stat is %d", (int)p->p_stat);
546 #endif
547 		if (wakeup_proc(p, ident, 0))
548 			--n;
549 	}
550 	SCHED_UNLOCK(s);
551 }
552 
553 /*
554  * Make all processes sleeping on the specified identifier runnable.
555  */
556 void
557 wakeup(const volatile void *chan)
558 {
559 	wakeup_n(chan, -1);
560 }
561 
562 int
563 sys_sched_yield(struct proc *p, void *v, register_t *retval)
564 {
565 	struct proc *q;
566 	uint8_t newprio;
567 	int s;
568 
569 	/*
570 	 * If one of the threads of a multi-threaded process called
571 	 * sched_yield(2), drop its priority to ensure its siblings
572 	 * can make some progress.
573 	 */
574 	mtx_enter(&p->p_p->ps_mtx);
575 	newprio = p->p_usrpri;
576 	TAILQ_FOREACH(q, &p->p_p->ps_threads, p_thr_link)
577 		newprio = max(newprio, q->p_runpri);
578 	mtx_leave(&p->p_p->ps_mtx);
579 
580 	SCHED_LOCK(s);
581 	setrunqueue(p->p_cpu, p, newprio);
582 	p->p_ru.ru_nvcsw++;
583 	mi_switch();
584 	SCHED_UNLOCK(s);
585 
586 	return (0);
587 }
588 
589 int
590 thrsleep_unlock(void *lock)
591 {
592 	static _atomic_lock_t unlocked = _ATOMIC_LOCK_UNLOCKED;
593 	_atomic_lock_t *atomiclock = lock;
594 
595 	if (!lock)
596 		return 0;
597 
598 	return copyout(&unlocked, atomiclock, sizeof(unlocked));
599 }
600 
601 struct tslpentry {
602 	TAILQ_ENTRY(tslpentry)	tslp_link;
603 	long			tslp_ident;
604 };
605 
606 /* thrsleep queue shared between processes */
607 static struct tslpqueue thrsleep_queue = TAILQ_HEAD_INITIALIZER(thrsleep_queue);
608 static struct rwlock thrsleep_lock = RWLOCK_INITIALIZER("thrsleeplk");
609 
610 int
611 thrsleep(struct proc *p, struct sys___thrsleep_args *v)
612 {
613 	struct sys___thrsleep_args /* {
614 		syscallarg(const volatile void *) ident;
615 		syscallarg(clockid_t) clock_id;
616 		syscallarg(const struct timespec *) tp;
617 		syscallarg(void *) lock;
618 		syscallarg(const int *) abort;
619 	} */ *uap = v;
620 	long ident = (long)SCARG(uap, ident);
621 	struct tslpentry entry;
622 	struct tslpqueue *queue;
623 	struct rwlock *qlock;
624 	struct timespec *tsp = (struct timespec *)SCARG(uap, tp);
625 	void *lock = SCARG(uap, lock);
626 	uint64_t nsecs = INFSLP;
627 	int abort = 0, error;
628 	clockid_t clock_id = SCARG(uap, clock_id);
629 
630 	if (ident == 0)
631 		return (EINVAL);
632 	if (tsp != NULL) {
633 		struct timespec now;
634 
635 		if ((error = clock_gettime(p, clock_id, &now)))
636 			return (error);
637 #ifdef KTRACE
638 		if (KTRPOINT(p, KTR_STRUCT))
639 			ktrabstimespec(p, tsp);
640 #endif
641 
642 		if (timespeccmp(tsp, &now, <=)) {
643 			/* already passed: still do the unlock */
644 			if ((error = thrsleep_unlock(lock)))
645 				return (error);
646 			return (EWOULDBLOCK);
647 		}
648 
649 		timespecsub(tsp, &now, tsp);
650 		nsecs = MIN(TIMESPEC_TO_NSEC(tsp), MAXTSLP);
651 	}
652 
653 	if (ident == -1) {
654 		queue = &thrsleep_queue;
655 		qlock = &thrsleep_lock;
656 	} else {
657 		queue = &p->p_p->ps_tslpqueue;
658 		qlock = &p->p_p->ps_lock;
659 	}
660 
661 	/* Interlock with wakeup. */
662 	entry.tslp_ident = ident;
663 	rw_enter_write(qlock);
664 	TAILQ_INSERT_TAIL(queue, &entry, tslp_link);
665 	rw_exit_write(qlock);
666 
667 	error = thrsleep_unlock(lock);
668 
669 	if (error == 0 && SCARG(uap, abort) != NULL)
670 		error = copyin(SCARG(uap, abort), &abort, sizeof(abort));
671 
672 	rw_enter_write(qlock);
673 	if (error != 0)
674 		goto out;
675 	if (abort != 0) {
676 		error = EINTR;
677 		goto out;
678 	}
679 	if (entry.tslp_ident != 0) {
680 		error = rwsleep_nsec(&entry, qlock, PWAIT|PCATCH, "thrsleep",
681 		    nsecs);
682 	}
683 
684 out:
685 	if (entry.tslp_ident != 0)
686 		TAILQ_REMOVE(queue, &entry, tslp_link);
687 	rw_exit_write(qlock);
688 
689 	if (error == ERESTART)
690 		error = ECANCELED;
691 
692 	return (error);
693 
694 }
695 
696 int
697 sys___thrsleep(struct proc *p, void *v, register_t *retval)
698 {
699 	struct sys___thrsleep_args /* {
700 		syscallarg(const volatile void *) ident;
701 		syscallarg(clockid_t) clock_id;
702 		syscallarg(struct timespec *) tp;
703 		syscallarg(void *) lock;
704 		syscallarg(const int *) abort;
705 	} */ *uap = v;
706 	struct timespec ts;
707 	int error;
708 
709 	if (SCARG(uap, tp) != NULL) {
710 		if ((error = copyin(SCARG(uap, tp), &ts, sizeof(ts)))) {
711 			*retval = error;
712 			return 0;
713 		}
714 		if (!timespecisvalid(&ts)) {
715 			*retval = EINVAL;
716 			return 0;
717 		}
718 		SCARG(uap, tp) = &ts;
719 	}
720 
721 	*retval = thrsleep(p, uap);
722 	return 0;
723 }
724 
725 int
726 sys___thrwakeup(struct proc *p, void *v, register_t *retval)
727 {
728 	struct sys___thrwakeup_args /* {
729 		syscallarg(const volatile void *) ident;
730 		syscallarg(int) n;
731 	} */ *uap = v;
732 	struct tslpentry *entry, *tmp;
733 	struct tslpqueue *queue;
734 	struct rwlock *qlock;
735 	long ident = (long)SCARG(uap, ident);
736 	int n = SCARG(uap, n);
737 	int found = 0;
738 
739 	if (ident == 0)
740 		*retval = EINVAL;
741 	else {
742 		if (ident == -1) {
743 			queue = &thrsleep_queue;
744 			qlock = &thrsleep_lock;
745 			/*
746 			 * Wake up all waiters with ident -1. This is needed
747 			 * because ident -1 can be shared by multiple userspace
748 			 * lock state machines concurrently. The implementation
749 			 * has no way to direct the wakeup to a particular
750 			 * state machine.
751 			 */
752 			n = 0;
753 		} else {
754 			queue = &p->p_p->ps_tslpqueue;
755 			qlock = &p->p_p->ps_lock;
756 		}
757 
758 		rw_enter_write(qlock);
759 		TAILQ_FOREACH_SAFE(entry, queue, tslp_link, tmp) {
760 			if (entry->tslp_ident == ident) {
761 				TAILQ_REMOVE(queue, entry, tslp_link);
762 				entry->tslp_ident = 0;
763 				wakeup_one(entry);
764 				if (++found == n)
765 					break;
766 			}
767 		}
768 		rw_exit_write(qlock);
769 
770 		if (ident == -1)
771 			*retval = 0;
772 		else
773 			*retval = found ? 0 : ESRCH;
774 	}
775 
776 	return (0);
777 }
778 
779 void
780 refcnt_init(struct refcnt *r)
781 {
782 	refcnt_init_trace(r, 0);
783 }
784 
785 void
786 refcnt_init_trace(struct refcnt *r, int idx)
787 {
788 	r->r_traceidx = idx;
789 	atomic_store_int(&r->r_refs, 1);
790 	TRACEINDEX(refcnt, r->r_traceidx, r, 0, +1);
791 }
792 
793 void
794 refcnt_take(struct refcnt *r)
795 {
796 	u_int refs;
797 
798 	refs = atomic_inc_int_nv(&r->r_refs);
799 	KASSERT(refs != 0);
800 	TRACEINDEX(refcnt, r->r_traceidx, r, refs - 1, +1);
801 	(void)refs;
802 }
803 
804 int
805 refcnt_rele(struct refcnt *r)
806 {
807 	u_int refs;
808 
809 	membar_exit_before_atomic();
810 	refs = atomic_dec_int_nv(&r->r_refs);
811 	KASSERT(refs != ~0);
812 	TRACEINDEX(refcnt, r->r_traceidx, r, refs + 1, -1);
813 	if (refs == 0) {
814 		membar_enter_after_atomic();
815 		return (1);
816 	}
817 	return (0);
818 }
819 
820 void
821 refcnt_rele_wake(struct refcnt *r)
822 {
823 	if (refcnt_rele(r))
824 		wakeup_one(r);
825 }
826 
827 void
828 refcnt_finalize(struct refcnt *r, const char *wmesg)
829 {
830 	u_int refs;
831 
832 	membar_exit_before_atomic();
833 	refs = atomic_dec_int_nv(&r->r_refs);
834 	KASSERT(refs != ~0);
835 	TRACEINDEX(refcnt, r->r_traceidx, r, refs + 1, -1);
836 	while (refs) {
837 		sleep_setup(r, PWAIT, wmesg);
838 		refs = atomic_load_int(&r->r_refs);
839 		sleep_finish(0, refs);
840 	}
841 	TRACEINDEX(refcnt, r->r_traceidx, r, refs, 0);
842 	/* Order subsequent loads and stores after refs == 0 load. */
843 	membar_sync();
844 }
845 
846 int
847 refcnt_shared(struct refcnt *r)
848 {
849 	u_int refs;
850 
851 	refs = atomic_load_int(&r->r_refs);
852 	TRACEINDEX(refcnt, r->r_traceidx, r, refs, 0);
853 	return (refs > 1);
854 }
855 
856 unsigned int
857 refcnt_read(struct refcnt *r)
858 {
859 	u_int refs;
860 
861 	refs = atomic_load_int(&r->r_refs);
862 	TRACEINDEX(refcnt, r->r_traceidx, r, refs, 0);
863 	return (refs);
864 }
865 
866 void
867 cond_init(struct cond *c)
868 {
869 	atomic_store_int(&c->c_wait, 1);
870 }
871 
872 void
873 cond_signal(struct cond *c)
874 {
875 	atomic_store_int(&c->c_wait, 0);
876 
877 	wakeup_one(c);
878 }
879 
880 void
881 cond_wait(struct cond *c, const char *wmesg)
882 {
883 	unsigned int wait;
884 
885 	wait = atomic_load_int(&c->c_wait);
886 	while (wait) {
887 		sleep_setup(c, PWAIT, wmesg);
888 		wait = atomic_load_int(&c->c_wait);
889 		sleep_finish(0, wait);
890 	}
891 }
892