xref: /openbsd-src/sys/kern/kern_synch.c (revision edddd55a3e8317b9e3fe2f3551a9d21da4cc8e58)
1 /*	$OpenBSD: kern_synch.c,v 1.204 2024/05/22 09:24:11 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 	/* exiting processes are not allowed to catch signals */
346 	if (p->p_flag & P_WEXIT)
347 		CLR(prio, PCATCH);
348 
349 	SCHED_LOCK(s);
350 
351 	TRACEPOINT(sched, sleep, NULL);
352 
353 	p->p_wchan = ident;
354 	p->p_wmesg = wmesg;
355 	p->p_slptime = 0;
356 	p->p_slppri = prio & PRIMASK;
357 	atomic_setbits_int(&p->p_flag, P_WSLEEP);
358 	TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_runq);
359 	if (prio & PCATCH)
360 		atomic_setbits_int(&p->p_flag, P_SINTR);
361 	p->p_stat = SSLEEP;
362 
363 	SCHED_UNLOCK(s);
364 }
365 
366 int
367 sleep_finish(int timo, int do_sleep)
368 {
369 	struct proc *p = curproc;
370 	int s, catch, error = 0, error1 = 0;
371 
372 	catch = p->p_flag & P_SINTR;
373 
374 	if (timo != 0) {
375 		KASSERT((p->p_flag & P_TIMEOUT) == 0);
376 		timeout_add(&p->p_sleep_to, timo);
377 	}
378 
379 	if (catch != 0) {
380 		/*
381 		 * We put ourselves on the sleep queue and start our
382 		 * timeout before calling sleep_signal_check(), as we could
383 		 * stop there, and a wakeup or a SIGCONT (or both) could
384 		 * occur while we were stopped.  A SIGCONT would cause
385 		 * us to be marked as SSLEEP without resuming us, thus
386 		 * we must be ready for sleep when sleep_signal_check() is
387 		 * called.
388 		 */
389 		if ((error = sleep_signal_check()) != 0) {
390 			catch = 0;
391 			do_sleep = 0;
392 		}
393 	}
394 
395 	SCHED_LOCK(s);
396 	/*
397 	 * If the wakeup happens while going to sleep, p->p_wchan
398 	 * will be NULL. In that case unwind immediately but still
399 	 * check for possible signals and timeouts.
400 	 */
401 	if (p->p_wchan == NULL)
402 		do_sleep = 0;
403 	atomic_clearbits_int(&p->p_flag, P_WSLEEP);
404 
405 	if (do_sleep) {
406 		KASSERT(p->p_stat == SSLEEP || p->p_stat == SSTOP);
407 		p->p_ru.ru_nvcsw++;
408 		mi_switch();
409 	} else {
410 		KASSERT(p->p_stat == SONPROC || p->p_stat == SSLEEP ||
411 		    p->p_stat == SSTOP);
412 		unsleep(p);
413 		p->p_stat = SONPROC;
414 	}
415 
416 #ifdef DIAGNOSTIC
417 	if (p->p_stat != SONPROC)
418 		panic("sleep_finish !SONPROC");
419 #endif
420 
421 	p->p_cpu->ci_schedstate.spc_curpriority = p->p_usrpri;
422 	SCHED_UNLOCK(s);
423 
424 	/*
425 	 * Even though this belongs to the signal handling part of sleep,
426 	 * we need to clear it before the ktrace.
427 	 */
428 	atomic_clearbits_int(&p->p_flag, P_SINTR);
429 
430 	if (timo != 0) {
431 		if (p->p_flag & P_TIMEOUT) {
432 			error1 = EWOULDBLOCK;
433 		} else {
434 			/* This can sleep. It must not use timeouts. */
435 			timeout_del_barrier(&p->p_sleep_to);
436 		}
437 		atomic_clearbits_int(&p->p_flag, P_TIMEOUT);
438 	}
439 
440 	/* Check if thread was woken up because of a unwind or signal */
441 	if (catch != 0)
442 		error = sleep_signal_check();
443 
444 	/* Signal errors are higher priority than timeouts. */
445 	if (error == 0 && error1 != 0)
446 		error = error1;
447 
448 	return error;
449 }
450 
451 /*
452  * Check and handle signals and suspensions around a sleep cycle.
453  */
454 int
455 sleep_signal_check(void)
456 {
457 	struct proc *p = curproc;
458 	struct sigctx ctx;
459 	int err, sig;
460 
461 	if ((err = single_thread_check(p, 1)) != 0)
462 		return err;
463 	if ((sig = cursig(p, &ctx)) != 0) {
464 		if (ctx.sig_intr)
465 			return EINTR;
466 		else
467 			return ERESTART;
468 	}
469 	return 0;
470 }
471 
472 int
473 wakeup_proc(struct proc *p, int flags)
474 {
475 	int awakened = 0;
476 
477 	SCHED_ASSERT_LOCKED();
478 
479 	if (p->p_wchan != NULL) {
480 		awakened = 1;
481 		if (flags)
482 			atomic_setbits_int(&p->p_flag, flags);
483 #ifdef DIAGNOSTIC
484 		if (p->p_stat != SSLEEP && p->p_stat != SSTOP)
485 			panic("thread %d p_stat is %d", p->p_tid, p->p_stat);
486 #endif
487 		unsleep(p);
488 		if (p->p_stat == SSLEEP)
489 			setrunnable(p);
490 	}
491 
492 	return awakened;
493 }
494 
495 
496 /*
497  * Implement timeout for tsleep.
498  * If process hasn't been awakened (wchan non-zero),
499  * set timeout flag and undo the sleep.  If proc
500  * is stopped, just unsleep so it will remain stopped.
501  */
502 void
503 endtsleep(void *arg)
504 {
505 	struct proc *p = arg;
506 	int s;
507 
508 	SCHED_LOCK(s);
509 	wakeup_proc(p, P_TIMEOUT);
510 	SCHED_UNLOCK(s);
511 }
512 
513 /*
514  * Remove a process from its wait queue
515  */
516 void
517 unsleep(struct proc *p)
518 {
519 	SCHED_ASSERT_LOCKED();
520 
521 	if (p->p_wchan != NULL) {
522 		TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_runq);
523 		p->p_wchan = NULL;
524 		p->p_wmesg = NULL;
525 		TRACEPOINT(sched, unsleep, p->p_tid + THREAD_PID_OFFSET,
526 		    p->p_p->ps_pid);
527 	}
528 }
529 
530 /*
531  * Make a number of processes sleeping on the specified identifier runnable.
532  */
533 void
534 wakeup_n(const volatile void *ident, int n)
535 {
536 	struct slpque *qp, wakeq;
537 	struct proc *p;
538 	struct proc *pnext;
539 	int s;
540 
541 	TAILQ_INIT(&wakeq);
542 
543 	SCHED_LOCK(s);
544 	qp = &slpque[LOOKUP(ident)];
545 	for (p = TAILQ_FIRST(qp); p != NULL && n != 0; p = pnext) {
546 		pnext = TAILQ_NEXT(p, p_runq);
547 #ifdef DIAGNOSTIC
548 		if (p->p_stat != SSLEEP && p->p_stat != SSTOP)
549 			panic("thread %d p_stat is %d", p->p_tid, p->p_stat);
550 #endif
551 		KASSERT(p->p_wchan != NULL);
552 		if (p->p_wchan == ident) {
553 			TAILQ_REMOVE(qp, p, p_runq);
554 			p->p_wchan = NULL;
555 			p->p_wmesg = NULL;
556 			TAILQ_INSERT_TAIL(&wakeq, p, p_runq);
557 			--n;
558 		}
559 	}
560 	while ((p = TAILQ_FIRST(&wakeq))) {
561 		TAILQ_REMOVE(&wakeq, p, p_runq);
562 		TRACEPOINT(sched, unsleep, p->p_tid + THREAD_PID_OFFSET,
563 		    p->p_p->ps_pid);
564 		if (p->p_stat == SSLEEP)
565 			setrunnable(p);
566 	}
567 	SCHED_UNLOCK(s);
568 }
569 
570 /*
571  * Make all processes sleeping on the specified identifier runnable.
572  */
573 void
574 wakeup(const volatile void *chan)
575 {
576 	wakeup_n(chan, -1);
577 }
578 
579 int
580 sys_sched_yield(struct proc *p, void *v, register_t *retval)
581 {
582 	struct proc *q;
583 	uint8_t newprio;
584 	int s;
585 
586 	/*
587 	 * If one of the threads of a multi-threaded process called
588 	 * sched_yield(2), drop its priority to ensure its siblings
589 	 * can make some progress.
590 	 */
591 	mtx_enter(&p->p_p->ps_mtx);
592 	newprio = p->p_usrpri;
593 	TAILQ_FOREACH(q, &p->p_p->ps_threads, p_thr_link)
594 		newprio = max(newprio, q->p_runpri);
595 	mtx_leave(&p->p_p->ps_mtx);
596 
597 	SCHED_LOCK(s);
598 	setrunqueue(p->p_cpu, p, newprio);
599 	p->p_ru.ru_nvcsw++;
600 	mi_switch();
601 	SCHED_UNLOCK(s);
602 
603 	return (0);
604 }
605 
606 int
607 thrsleep_unlock(void *lock)
608 {
609 	static _atomic_lock_t unlocked = _ATOMIC_LOCK_UNLOCKED;
610 	_atomic_lock_t *atomiclock = lock;
611 
612 	if (!lock)
613 		return 0;
614 
615 	return copyout(&unlocked, atomiclock, sizeof(unlocked));
616 }
617 
618 struct tslpentry {
619 	TAILQ_ENTRY(tslpentry)	tslp_link;
620 	long			tslp_ident;
621 };
622 
623 /* thrsleep queue shared between processes */
624 static struct tslpqueue thrsleep_queue = TAILQ_HEAD_INITIALIZER(thrsleep_queue);
625 static struct rwlock thrsleep_lock = RWLOCK_INITIALIZER("thrsleeplk");
626 
627 int
628 thrsleep(struct proc *p, struct sys___thrsleep_args *v)
629 {
630 	struct sys___thrsleep_args /* {
631 		syscallarg(const volatile void *) ident;
632 		syscallarg(clockid_t) clock_id;
633 		syscallarg(const struct timespec *) tp;
634 		syscallarg(void *) lock;
635 		syscallarg(const int *) abort;
636 	} */ *uap = v;
637 	long ident = (long)SCARG(uap, ident);
638 	struct tslpentry entry;
639 	struct tslpqueue *queue;
640 	struct rwlock *qlock;
641 	struct timespec *tsp = (struct timespec *)SCARG(uap, tp);
642 	void *lock = SCARG(uap, lock);
643 	uint64_t nsecs = INFSLP;
644 	int abort = 0, error;
645 	clockid_t clock_id = SCARG(uap, clock_id);
646 
647 	if (ident == 0)
648 		return (EINVAL);
649 	if (tsp != NULL) {
650 		struct timespec now;
651 
652 		if ((error = clock_gettime(p, clock_id, &now)))
653 			return (error);
654 #ifdef KTRACE
655 		if (KTRPOINT(p, KTR_STRUCT))
656 			ktrabstimespec(p, tsp);
657 #endif
658 
659 		if (timespeccmp(tsp, &now, <=)) {
660 			/* already passed: still do the unlock */
661 			if ((error = thrsleep_unlock(lock)))
662 				return (error);
663 			return (EWOULDBLOCK);
664 		}
665 
666 		timespecsub(tsp, &now, tsp);
667 		nsecs = MIN(TIMESPEC_TO_NSEC(tsp), MAXTSLP);
668 	}
669 
670 	if (ident == -1) {
671 		queue = &thrsleep_queue;
672 		qlock = &thrsleep_lock;
673 	} else {
674 		queue = &p->p_p->ps_tslpqueue;
675 		qlock = &p->p_p->ps_lock;
676 	}
677 
678 	/* Interlock with wakeup. */
679 	entry.tslp_ident = ident;
680 	rw_enter_write(qlock);
681 	TAILQ_INSERT_TAIL(queue, &entry, tslp_link);
682 	rw_exit_write(qlock);
683 
684 	error = thrsleep_unlock(lock);
685 
686 	if (error == 0 && SCARG(uap, abort) != NULL)
687 		error = copyin(SCARG(uap, abort), &abort, sizeof(abort));
688 
689 	rw_enter_write(qlock);
690 	if (error != 0)
691 		goto out;
692 	if (abort != 0) {
693 		error = EINTR;
694 		goto out;
695 	}
696 	if (entry.tslp_ident != 0) {
697 		error = rwsleep_nsec(&entry, qlock, PWAIT|PCATCH, "thrsleep",
698 		    nsecs);
699 	}
700 
701 out:
702 	if (entry.tslp_ident != 0)
703 		TAILQ_REMOVE(queue, &entry, tslp_link);
704 	rw_exit_write(qlock);
705 
706 	if (error == ERESTART)
707 		error = ECANCELED;
708 
709 	return (error);
710 
711 }
712 
713 int
714 sys___thrsleep(struct proc *p, void *v, register_t *retval)
715 {
716 	struct sys___thrsleep_args /* {
717 		syscallarg(const volatile void *) ident;
718 		syscallarg(clockid_t) clock_id;
719 		syscallarg(struct timespec *) tp;
720 		syscallarg(void *) lock;
721 		syscallarg(const int *) abort;
722 	} */ *uap = v;
723 	struct timespec ts;
724 	int error;
725 
726 	if (SCARG(uap, tp) != NULL) {
727 		if ((error = copyin(SCARG(uap, tp), &ts, sizeof(ts)))) {
728 			*retval = error;
729 			return 0;
730 		}
731 		if (!timespecisvalid(&ts)) {
732 			*retval = EINVAL;
733 			return 0;
734 		}
735 		SCARG(uap, tp) = &ts;
736 	}
737 
738 	*retval = thrsleep(p, uap);
739 	return 0;
740 }
741 
742 int
743 sys___thrwakeup(struct proc *p, void *v, register_t *retval)
744 {
745 	struct sys___thrwakeup_args /* {
746 		syscallarg(const volatile void *) ident;
747 		syscallarg(int) n;
748 	} */ *uap = v;
749 	struct tslpentry *entry, *tmp;
750 	struct tslpqueue *queue;
751 	struct rwlock *qlock;
752 	long ident = (long)SCARG(uap, ident);
753 	int n = SCARG(uap, n);
754 	int found = 0;
755 
756 	if (ident == 0)
757 		*retval = EINVAL;
758 	else {
759 		if (ident == -1) {
760 			queue = &thrsleep_queue;
761 			qlock = &thrsleep_lock;
762 			/*
763 			 * Wake up all waiters with ident -1. This is needed
764 			 * because ident -1 can be shared by multiple userspace
765 			 * lock state machines concurrently. The implementation
766 			 * has no way to direct the wakeup to a particular
767 			 * state machine.
768 			 */
769 			n = 0;
770 		} else {
771 			queue = &p->p_p->ps_tslpqueue;
772 			qlock = &p->p_p->ps_lock;
773 		}
774 
775 		rw_enter_write(qlock);
776 		TAILQ_FOREACH_SAFE(entry, queue, tslp_link, tmp) {
777 			if (entry->tslp_ident == ident) {
778 				TAILQ_REMOVE(queue, entry, tslp_link);
779 				entry->tslp_ident = 0;
780 				wakeup_one(entry);
781 				if (++found == n)
782 					break;
783 			}
784 		}
785 		rw_exit_write(qlock);
786 
787 		if (ident == -1)
788 			*retval = 0;
789 		else
790 			*retval = found ? 0 : ESRCH;
791 	}
792 
793 	return (0);
794 }
795 
796 void
797 refcnt_init(struct refcnt *r)
798 {
799 	refcnt_init_trace(r, 0);
800 }
801 
802 void
803 refcnt_init_trace(struct refcnt *r, int idx)
804 {
805 	r->r_traceidx = idx;
806 	atomic_store_int(&r->r_refs, 1);
807 	TRACEINDEX(refcnt, r->r_traceidx, r, 0, +1);
808 }
809 
810 void
811 refcnt_take(struct refcnt *r)
812 {
813 	u_int refs;
814 
815 	refs = atomic_inc_int_nv(&r->r_refs);
816 	KASSERT(refs != 0);
817 	TRACEINDEX(refcnt, r->r_traceidx, r, refs - 1, +1);
818 	(void)refs;
819 }
820 
821 int
822 refcnt_rele(struct refcnt *r)
823 {
824 	u_int refs;
825 
826 	membar_exit_before_atomic();
827 	refs = atomic_dec_int_nv(&r->r_refs);
828 	KASSERT(refs != ~0);
829 	TRACEINDEX(refcnt, r->r_traceidx, r, refs + 1, -1);
830 	if (refs == 0) {
831 		membar_enter_after_atomic();
832 		return (1);
833 	}
834 	return (0);
835 }
836 
837 void
838 refcnt_rele_wake(struct refcnt *r)
839 {
840 	if (refcnt_rele(r))
841 		wakeup_one(r);
842 }
843 
844 void
845 refcnt_finalize(struct refcnt *r, const char *wmesg)
846 {
847 	u_int refs;
848 
849 	membar_exit_before_atomic();
850 	refs = atomic_dec_int_nv(&r->r_refs);
851 	KASSERT(refs != ~0);
852 	TRACEINDEX(refcnt, r->r_traceidx, r, refs + 1, -1);
853 	while (refs) {
854 		sleep_setup(r, PWAIT, wmesg);
855 		refs = atomic_load_int(&r->r_refs);
856 		sleep_finish(0, refs);
857 	}
858 	TRACEINDEX(refcnt, r->r_traceidx, r, refs, 0);
859 	/* Order subsequent loads and stores after refs == 0 load. */
860 	membar_sync();
861 }
862 
863 int
864 refcnt_shared(struct refcnt *r)
865 {
866 	u_int refs;
867 
868 	refs = atomic_load_int(&r->r_refs);
869 	TRACEINDEX(refcnt, r->r_traceidx, r, refs, 0);
870 	return (refs > 1);
871 }
872 
873 unsigned int
874 refcnt_read(struct refcnt *r)
875 {
876 	u_int refs;
877 
878 	refs = atomic_load_int(&r->r_refs);
879 	TRACEINDEX(refcnt, r->r_traceidx, r, refs, 0);
880 	return (refs);
881 }
882 
883 void
884 cond_init(struct cond *c)
885 {
886 	atomic_store_int(&c->c_wait, 1);
887 }
888 
889 void
890 cond_signal(struct cond *c)
891 {
892 	atomic_store_int(&c->c_wait, 0);
893 
894 	wakeup_one(c);
895 }
896 
897 void
898 cond_wait(struct cond *c, const char *wmesg)
899 {
900 	unsigned int wait;
901 
902 	wait = atomic_load_int(&c->c_wait);
903 	while (wait) {
904 		sleep_setup(c, PWAIT, wmesg);
905 		wait = atomic_load_int(&c->c_wait);
906 		sleep_finish(0, wait);
907 	}
908 }
909