xref: /titanic_50/usr/src/lib/libc/port/threads/synch.c (revision 883492d5a933deb34cd27521e7f2756773cd27af)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/sdt.h>
30 
31 #include "lint.h"
32 #include "thr_uberdata.h"
33 
34 /*
35  * This mutex is initialized to be held by lwp#1.
36  * It is used to block a thread that has returned from a mutex_lock()
37  * of a LOCK_PRIO_INHERIT mutex with an unrecoverable error.
38  */
39 mutex_t	stall_mutex = DEFAULTMUTEX;
40 
41 static int shared_mutex_held(mutex_t *);
42 static int mutex_unlock_internal(mutex_t *, int);
43 static int mutex_queuelock_adaptive(mutex_t *);
44 static void mutex_wakeup_all(mutex_t *);
45 
46 /*
47  * Lock statistics support functions.
48  */
49 void
50 record_begin_hold(tdb_mutex_stats_t *msp)
51 {
52 	tdb_incr(msp->mutex_lock);
53 	msp->mutex_begin_hold = gethrtime();
54 }
55 
56 hrtime_t
57 record_hold_time(tdb_mutex_stats_t *msp)
58 {
59 	hrtime_t now = gethrtime();
60 
61 	if (msp->mutex_begin_hold)
62 		msp->mutex_hold_time += now - msp->mutex_begin_hold;
63 	msp->mutex_begin_hold = 0;
64 	return (now);
65 }
66 
67 /*
68  * Called once at library initialization.
69  */
70 void
71 mutex_setup(void)
72 {
73 	if (set_lock_byte(&stall_mutex.mutex_lockw))
74 		thr_panic("mutex_setup() cannot acquire stall_mutex");
75 	stall_mutex.mutex_owner = (uintptr_t)curthread;
76 }
77 
78 /*
79  * The default spin counts of 1000 and 500 are experimentally determined.
80  * On sun4u machines with any number of processors they could be raised
81  * to 10,000 but that (experimentally) makes almost no difference.
82  * The environment variables:
83  *	_THREAD_ADAPTIVE_SPIN=count
84  *	_THREAD_RELEASE_SPIN=count
85  * can be used to override and set the counts in the range [0 .. 1,000,000].
86  */
87 int	thread_adaptive_spin = 1000;
88 uint_t	thread_max_spinners = 100;
89 int	thread_release_spin = 500;
90 int	thread_queue_verify = 0;
91 static	int	ncpus;
92 
93 /*
94  * Distinguish spinning for queue locks from spinning for regular locks.
95  * The environment variable:
96  *	_THREAD_QUEUE_SPIN=count
97  * can be used to override and set the count in the range [0 .. 1,000,000].
98  * There is no release spin concept for queue locks.
99  */
100 int	thread_queue_spin = 1000;
101 
102 /*
103  * Use the otherwise-unused 'mutex_ownerpid' field of a USYNC_THREAD
104  * mutex to be a count of adaptive spins in progress.
105  */
106 #define	mutex_spinners	mutex_ownerpid
107 
108 #define	ALL_ATTRIBUTES				\
109 	(LOCK_RECURSIVE | LOCK_ERRORCHECK |	\
110 	LOCK_PRIO_INHERIT | LOCK_PRIO_PROTECT |	\
111 	LOCK_ROBUST)
112 
113 /*
114  * 'type' can be one of USYNC_THREAD, USYNC_PROCESS, or USYNC_PROCESS_ROBUST,
115  * augmented by zero or more the flags:
116  *	LOCK_RECURSIVE
117  *	LOCK_ERRORCHECK
118  *	LOCK_PRIO_INHERIT
119  *	LOCK_PRIO_PROTECT
120  *	LOCK_ROBUST
121  */
122 #pragma weak _private_mutex_init = __mutex_init
123 #pragma weak mutex_init = __mutex_init
124 #pragma weak _mutex_init = __mutex_init
125 /* ARGSUSED2 */
126 int
127 __mutex_init(mutex_t *mp, int type, void *arg)
128 {
129 	int basetype = (type & ~ALL_ATTRIBUTES);
130 	int error = 0;
131 
132 	if (basetype == USYNC_PROCESS_ROBUST) {
133 		/*
134 		 * USYNC_PROCESS_ROBUST is a deprecated historical type.
135 		 * We change it into (USYNC_PROCESS | LOCK_ROBUST) but
136 		 * retain the USYNC_PROCESS_ROBUST flag so we can return
137 		 * ELOCKUNMAPPED when necessary (only USYNC_PROCESS_ROBUST
138 		 * mutexes will ever draw ELOCKUNMAPPED).
139 		 */
140 		type |= (USYNC_PROCESS | LOCK_ROBUST);
141 		basetype = USYNC_PROCESS;
142 	}
143 
144 	if (!(basetype == USYNC_THREAD || basetype == USYNC_PROCESS) ||
145 	    (type & (LOCK_PRIO_INHERIT | LOCK_PRIO_PROTECT))
146 	    == (LOCK_PRIO_INHERIT | LOCK_PRIO_PROTECT)) {
147 		error = EINVAL;
148 	} else if (type & LOCK_ROBUST) {
149 		/*
150 		 * Callers of mutex_init() with the LOCK_ROBUST attribute
151 		 * are required to pass an initially all-zero mutex.
152 		 * Multiple calls to mutex_init() are allowed; all but
153 		 * the first return EBUSY.  A call to mutex_init() is
154 		 * allowed to make an inconsistent robust lock consistent
155 		 * (for historical usage, even though the proper interface
156 		 * for this is mutex_consistent()).  Note that we use
157 		 * atomic_or_16() to set the LOCK_INITED flag so as
158 		 * not to disturb surrounding bits (LOCK_OWNERDEAD, etc).
159 		 */
160 		extern void _atomic_or_16(volatile uint16_t *, uint16_t);
161 		if (!(mp->mutex_flag & LOCK_INITED)) {
162 			mp->mutex_type = (uint8_t)type;
163 			_atomic_or_16(&mp->mutex_flag, LOCK_INITED);
164 			mp->mutex_magic = MUTEX_MAGIC;
165 		} else if (type != mp->mutex_type ||
166 		    ((type & LOCK_PRIO_PROTECT) &&
167 		    mp->mutex_ceiling != (*(int *)arg))) {
168 			error = EINVAL;
169 		} else if (__mutex_consistent(mp) != 0) {
170 			error = EBUSY;
171 		}
172 		/* register a process robust mutex with the kernel */
173 		if (basetype == USYNC_PROCESS)
174 			register_lock(mp);
175 	} else {
176 		(void) _memset(mp, 0, sizeof (*mp));
177 		mp->mutex_type = (uint8_t)type;
178 		mp->mutex_flag = LOCK_INITED;
179 		mp->mutex_magic = MUTEX_MAGIC;
180 	}
181 
182 	if (error == 0 && (type & LOCK_PRIO_PROTECT))
183 		mp->mutex_ceiling = (uint8_t)(*(int *)arg);
184 
185 	return (error);
186 }
187 
188 /*
189  * Delete mp from list of ceil mutexes owned by curthread.
190  * Return 1 if the head of the chain was updated.
191  */
192 int
193 _ceil_mylist_del(mutex_t *mp)
194 {
195 	ulwp_t *self = curthread;
196 	mxchain_t **mcpp;
197 	mxchain_t *mcp;
198 
199 	mcpp = &self->ul_mxchain;
200 	while ((*mcpp)->mxchain_mx != mp)
201 		mcpp = &(*mcpp)->mxchain_next;
202 	mcp = *mcpp;
203 	*mcpp = mcp->mxchain_next;
204 	lfree(mcp, sizeof (*mcp));
205 	return (mcpp == &self->ul_mxchain);
206 }
207 
208 /*
209  * Add mp to head of list of ceil mutexes owned by curthread.
210  * Return ENOMEM if no memory could be allocated.
211  */
212 int
213 _ceil_mylist_add(mutex_t *mp)
214 {
215 	ulwp_t *self = curthread;
216 	mxchain_t *mcp;
217 
218 	if ((mcp = lmalloc(sizeof (*mcp))) == NULL)
219 		return (ENOMEM);
220 	mcp->mxchain_mx = mp;
221 	mcp->mxchain_next = self->ul_mxchain;
222 	self->ul_mxchain = mcp;
223 	return (0);
224 }
225 
226 /*
227  * Inherit priority from ceiling.  The inheritance impacts the effective
228  * priority, not the assigned priority.  See _thread_setschedparam_main().
229  */
230 void
231 _ceil_prio_inherit(int ceil)
232 {
233 	ulwp_t *self = curthread;
234 	struct sched_param param;
235 
236 	(void) _memset(&param, 0, sizeof (param));
237 	param.sched_priority = ceil;
238 	if (_thread_setschedparam_main(self->ul_lwpid,
239 	    self->ul_policy, &param, PRIO_INHERIT)) {
240 		/*
241 		 * Panic since unclear what error code to return.
242 		 * If we do return the error codes returned by above
243 		 * called routine, update the man page...
244 		 */
245 		thr_panic("_thread_setschedparam_main() fails");
246 	}
247 }
248 
249 /*
250  * Waive inherited ceiling priority.  Inherit from head of owned ceiling locks
251  * if holding at least one ceiling lock.  If no ceiling locks are held at this
252  * point, disinherit completely, reverting back to assigned priority.
253  */
254 void
255 _ceil_prio_waive(void)
256 {
257 	ulwp_t *self = curthread;
258 	struct sched_param param;
259 
260 	(void) _memset(&param, 0, sizeof (param));
261 	if (self->ul_mxchain == NULL) {
262 		/*
263 		 * No ceil locks held.  Zero the epri, revert back to ul_pri.
264 		 * Since thread's hash lock is not held, one cannot just
265 		 * read ul_pri here...do it in the called routine...
266 		 */
267 		param.sched_priority = self->ul_pri;	/* ignored */
268 		if (_thread_setschedparam_main(self->ul_lwpid,
269 		    self->ul_policy, &param, PRIO_DISINHERIT))
270 			thr_panic("_thread_setschedparam_main() fails");
271 	} else {
272 		/*
273 		 * Set priority to that of the mutex at the head
274 		 * of the ceilmutex chain.
275 		 */
276 		param.sched_priority =
277 		    self->ul_mxchain->mxchain_mx->mutex_ceiling;
278 		if (_thread_setschedparam_main(self->ul_lwpid,
279 		    self->ul_policy, &param, PRIO_INHERIT))
280 			thr_panic("_thread_setschedparam_main() fails");
281 	}
282 }
283 
284 /*
285  * Non-preemptive spin locks.  Used by queue_lock().
286  * No lock statistics are gathered for these locks.
287  */
288 void
289 spin_lock_set(mutex_t *mp)
290 {
291 	ulwp_t *self = curthread;
292 
293 	no_preempt(self);
294 	if (set_lock_byte(&mp->mutex_lockw) == 0) {
295 		mp->mutex_owner = (uintptr_t)self;
296 		return;
297 	}
298 	/*
299 	 * Spin for a while, attempting to acquire the lock.
300 	 */
301 	if (self->ul_spin_lock_spin != UINT_MAX)
302 		self->ul_spin_lock_spin++;
303 	if (mutex_queuelock_adaptive(mp) == 0 ||
304 	    set_lock_byte(&mp->mutex_lockw) == 0) {
305 		mp->mutex_owner = (uintptr_t)self;
306 		return;
307 	}
308 	/*
309 	 * Try harder if we were previously at a no premption level.
310 	 */
311 	if (self->ul_preempt > 1) {
312 		if (self->ul_spin_lock_spin2 != UINT_MAX)
313 			self->ul_spin_lock_spin2++;
314 		if (mutex_queuelock_adaptive(mp) == 0 ||
315 		    set_lock_byte(&mp->mutex_lockw) == 0) {
316 			mp->mutex_owner = (uintptr_t)self;
317 			return;
318 		}
319 	}
320 	/*
321 	 * Give up and block in the kernel for the mutex.
322 	 */
323 	if (self->ul_spin_lock_sleep != UINT_MAX)
324 		self->ul_spin_lock_sleep++;
325 	(void) ___lwp_mutex_timedlock(mp, NULL);
326 	mp->mutex_owner = (uintptr_t)self;
327 }
328 
329 void
330 spin_lock_clear(mutex_t *mp)
331 {
332 	ulwp_t *self = curthread;
333 
334 	mp->mutex_owner = 0;
335 	if (atomic_swap_32(&mp->mutex_lockword, 0) & WAITERMASK) {
336 		(void) ___lwp_mutex_wakeup(mp, 0);
337 		if (self->ul_spin_lock_wakeup != UINT_MAX)
338 			self->ul_spin_lock_wakeup++;
339 	}
340 	preempt(self);
341 }
342 
343 /*
344  * Allocate the sleep queue hash table.
345  */
346 void
347 queue_alloc(void)
348 {
349 	ulwp_t *self = curthread;
350 	uberdata_t *udp = self->ul_uberdata;
351 	mutex_t *mp;
352 	void *data;
353 	int i;
354 
355 	/*
356 	 * No locks are needed; we call here only when single-threaded.
357 	 */
358 	ASSERT(self == udp->ulwp_one);
359 	ASSERT(!udp->uberflags.uf_mt);
360 	if ((data = _private_mmap(NULL, 2 * QHASHSIZE * sizeof (queue_head_t),
361 	    PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, -1, (off_t)0))
362 	    == MAP_FAILED)
363 		thr_panic("cannot allocate thread queue_head table");
364 	udp->queue_head = (queue_head_t *)data;
365 	for (i = 0; i < 2 * QHASHSIZE; i++) {
366 		mp = &udp->queue_head[i].qh_lock;
367 		mp->mutex_flag = LOCK_INITED;
368 		mp->mutex_magic = MUTEX_MAGIC;
369 	}
370 }
371 
372 #if defined(THREAD_DEBUG)
373 
374 /*
375  * Debugging: verify correctness of a sleep queue.
376  */
377 void
378 QVERIFY(queue_head_t *qp)
379 {
380 	ulwp_t *self = curthread;
381 	uberdata_t *udp = self->ul_uberdata;
382 	ulwp_t *ulwp;
383 	ulwp_t *prev;
384 	uint_t index;
385 	uint32_t cnt = 0;
386 	char qtype;
387 	void *wchan;
388 
389 	ASSERT(qp >= udp->queue_head && (qp - udp->queue_head) < 2 * QHASHSIZE);
390 	ASSERT(MUTEX_OWNED(&qp->qh_lock, self));
391 	ASSERT((qp->qh_head != NULL && qp->qh_tail != NULL) ||
392 		(qp->qh_head == NULL && qp->qh_tail == NULL));
393 	if (!thread_queue_verify)
394 		return;
395 	/* real expensive stuff, only for _THREAD_QUEUE_VERIFY */
396 	qtype = ((qp - udp->queue_head) < QHASHSIZE)? MX : CV;
397 	for (prev = NULL, ulwp = qp->qh_head; ulwp != NULL;
398 	    prev = ulwp, ulwp = ulwp->ul_link, cnt++) {
399 		ASSERT(ulwp->ul_qtype == qtype);
400 		ASSERT(ulwp->ul_wchan != NULL);
401 		ASSERT(ulwp->ul_sleepq == qp);
402 		wchan = ulwp->ul_wchan;
403 		index = QUEUE_HASH(wchan, qtype);
404 		ASSERT(&udp->queue_head[index] == qp);
405 	}
406 	ASSERT(qp->qh_tail == prev);
407 	ASSERT(qp->qh_qlen == cnt);
408 }
409 
410 #else	/* THREAD_DEBUG */
411 
412 #define	QVERIFY(qp)
413 
414 #endif	/* THREAD_DEBUG */
415 
416 /*
417  * Acquire a queue head.
418  */
419 queue_head_t *
420 queue_lock(void *wchan, int qtype)
421 {
422 	uberdata_t *udp = curthread->ul_uberdata;
423 	queue_head_t *qp;
424 
425 	ASSERT(qtype == MX || qtype == CV);
426 
427 	/*
428 	 * It is possible that we could be called while still single-threaded.
429 	 * If so, we call queue_alloc() to allocate the queue_head[] array.
430 	 */
431 	if ((qp = udp->queue_head) == NULL) {
432 		queue_alloc();
433 		qp = udp->queue_head;
434 	}
435 	qp += QUEUE_HASH(wchan, qtype);
436 	spin_lock_set(&qp->qh_lock);
437 	/*
438 	 * At once per nanosecond, qh_lockcount will wrap after 512 years.
439 	 * Were we to care about this, we could peg the value at UINT64_MAX.
440 	 */
441 	qp->qh_lockcount++;
442 	QVERIFY(qp);
443 	return (qp);
444 }
445 
446 /*
447  * Release a queue head.
448  */
449 void
450 queue_unlock(queue_head_t *qp)
451 {
452 	QVERIFY(qp);
453 	spin_lock_clear(&qp->qh_lock);
454 }
455 
456 /*
457  * For rwlock queueing, we must queue writers ahead of readers of the
458  * same priority.  We do this by making writers appear to have a half
459  * point higher priority for purposes of priority comparisons below.
460  */
461 #define	CMP_PRIO(ulwp)	((real_priority(ulwp) << 1) + (ulwp)->ul_writer)
462 
463 void
464 enqueue(queue_head_t *qp, ulwp_t *ulwp, void *wchan, int qtype)
465 {
466 	ulwp_t **ulwpp;
467 	ulwp_t *next;
468 	int pri = CMP_PRIO(ulwp);
469 	int force_fifo = (qtype & FIFOQ);
470 	int do_fifo;
471 
472 	qtype &= ~FIFOQ;
473 	ASSERT(qtype == MX || qtype == CV);
474 	ASSERT(MUTEX_OWNED(&qp->qh_lock, curthread));
475 	ASSERT(ulwp->ul_sleepq != qp);
476 
477 	/*
478 	 * LIFO queue ordering is unfair and can lead to starvation,
479 	 * but it gives better performance for heavily contended locks.
480 	 * We use thread_queue_fifo (range is 0..8) to determine
481 	 * the frequency of FIFO vs LIFO queuing:
482 	 *	0 : every 256th time	(almost always LIFO)
483 	 *	1 : every 128th time
484 	 *	2 : every 64th  time
485 	 *	3 : every 32nd  time
486 	 *	4 : every 16th  time	(the default value, mostly LIFO)
487 	 *	5 : every 8th   time
488 	 *	6 : every 4th   time
489 	 *	7 : every 2nd   time
490 	 *	8 : every time		(never LIFO, always FIFO)
491 	 * Note that there is always some degree of FIFO ordering.
492 	 * This breaks live lock conditions that occur in applications
493 	 * that are written assuming (incorrectly) that threads acquire
494 	 * locks fairly, that is, in roughly round-robin order.
495 	 * In any event, the queue is maintained in priority order.
496 	 *
497 	 * If we are given the FIFOQ flag in qtype, fifo queueing is forced.
498 	 * SUSV3 requires this for semaphores.
499 	 */
500 	do_fifo = (force_fifo ||
501 		((++qp->qh_qcnt << curthread->ul_queue_fifo) & 0xff) == 0);
502 
503 	if (qp->qh_head == NULL) {
504 		/*
505 		 * The queue is empty.  LIFO/FIFO doesn't matter.
506 		 */
507 		ASSERT(qp->qh_tail == NULL);
508 		ulwpp = &qp->qh_head;
509 	} else if (do_fifo) {
510 		/*
511 		 * Enqueue after the last thread whose priority is greater
512 		 * than or equal to the priority of the thread being queued.
513 		 * Attempt first to go directly onto the tail of the queue.
514 		 */
515 		if (pri <= CMP_PRIO(qp->qh_tail))
516 			ulwpp = &qp->qh_tail->ul_link;
517 		else {
518 			for (ulwpp = &qp->qh_head; (next = *ulwpp) != NULL;
519 			    ulwpp = &next->ul_link)
520 				if (pri > CMP_PRIO(next))
521 					break;
522 		}
523 	} else {
524 		/*
525 		 * Enqueue before the first thread whose priority is less
526 		 * than or equal to the priority of the thread being queued.
527 		 * Hopefully we can go directly onto the head of the queue.
528 		 */
529 		for (ulwpp = &qp->qh_head; (next = *ulwpp) != NULL;
530 		    ulwpp = &next->ul_link)
531 			if (pri >= CMP_PRIO(next))
532 				break;
533 	}
534 	if ((ulwp->ul_link = *ulwpp) == NULL)
535 		qp->qh_tail = ulwp;
536 	*ulwpp = ulwp;
537 
538 	ulwp->ul_sleepq = qp;
539 	ulwp->ul_wchan = wchan;
540 	ulwp->ul_qtype = qtype;
541 	if (qp->qh_qmax < ++qp->qh_qlen)
542 		qp->qh_qmax = qp->qh_qlen;
543 }
544 
545 /*
546  * Return a pointer to the queue slot of the
547  * highest priority thread on the queue.
548  * On return, prevp, if not NULL, will contain a pointer
549  * to the thread's predecessor on the queue
550  */
551 static ulwp_t **
552 queue_slot(queue_head_t *qp, void *wchan, int *more, ulwp_t **prevp)
553 {
554 	ulwp_t **ulwpp;
555 	ulwp_t *ulwp;
556 	ulwp_t *prev = NULL;
557 	ulwp_t **suspp = NULL;
558 	ulwp_t *susprev;
559 
560 	ASSERT(MUTEX_OWNED(&qp->qh_lock, curthread));
561 
562 	/*
563 	 * Find a waiter on the sleep queue.
564 	 */
565 	for (ulwpp = &qp->qh_head; (ulwp = *ulwpp) != NULL;
566 	    prev = ulwp, ulwpp = &ulwp->ul_link) {
567 		if (ulwp->ul_wchan == wchan) {
568 			if (!ulwp->ul_stop)
569 				break;
570 			/*
571 			 * Try not to return a suspended thread.
572 			 * This mimics the old libthread's behavior.
573 			 */
574 			if (suspp == NULL) {
575 				suspp = ulwpp;
576 				susprev = prev;
577 			}
578 		}
579 	}
580 
581 	if (ulwp == NULL && suspp != NULL) {
582 		ulwp = *(ulwpp = suspp);
583 		prev = susprev;
584 		suspp = NULL;
585 	}
586 	if (ulwp == NULL) {
587 		if (more != NULL)
588 			*more = 0;
589 		return (NULL);
590 	}
591 
592 	if (prevp != NULL)
593 		*prevp = prev;
594 	if (more == NULL)
595 		return (ulwpp);
596 
597 	/*
598 	 * Scan the remainder of the queue for another waiter.
599 	 */
600 	if (suspp != NULL) {
601 		*more = 1;
602 		return (ulwpp);
603 	}
604 	for (ulwp = ulwp->ul_link; ulwp != NULL; ulwp = ulwp->ul_link) {
605 		if (ulwp->ul_wchan == wchan) {
606 			*more = 1;
607 			return (ulwpp);
608 		}
609 	}
610 
611 	*more = 0;
612 	return (ulwpp);
613 }
614 
615 ulwp_t *
616 queue_unlink(queue_head_t *qp, ulwp_t **ulwpp, ulwp_t *prev)
617 {
618 	ulwp_t *ulwp;
619 
620 	ulwp = *ulwpp;
621 	*ulwpp = ulwp->ul_link;
622 	ulwp->ul_link = NULL;
623 	if (qp->qh_tail == ulwp)
624 		qp->qh_tail = prev;
625 	qp->qh_qlen--;
626 	ulwp->ul_sleepq = NULL;
627 	ulwp->ul_wchan = NULL;
628 
629 	return (ulwp);
630 }
631 
632 ulwp_t *
633 dequeue(queue_head_t *qp, void *wchan, int *more)
634 {
635 	ulwp_t **ulwpp;
636 	ulwp_t *prev;
637 
638 	if ((ulwpp = queue_slot(qp, wchan, more, &prev)) == NULL)
639 		return (NULL);
640 	return (queue_unlink(qp, ulwpp, prev));
641 }
642 
643 /*
644  * Return a pointer to the highest priority thread sleeping on wchan.
645  */
646 ulwp_t *
647 queue_waiter(queue_head_t *qp, void *wchan)
648 {
649 	ulwp_t **ulwpp;
650 
651 	if ((ulwpp = queue_slot(qp, wchan, NULL, NULL)) == NULL)
652 		return (NULL);
653 	return (*ulwpp);
654 }
655 
656 uint8_t
657 dequeue_self(queue_head_t *qp, void *wchan)
658 {
659 	ulwp_t *self = curthread;
660 	ulwp_t **ulwpp;
661 	ulwp_t *ulwp;
662 	ulwp_t *prev = NULL;
663 	int found = 0;
664 	int more = 0;
665 
666 	ASSERT(MUTEX_OWNED(&qp->qh_lock, self));
667 
668 	/* find self on the sleep queue */
669 	for (ulwpp = &qp->qh_head; (ulwp = *ulwpp) != NULL;
670 	    prev = ulwp, ulwpp = &ulwp->ul_link) {
671 		if (ulwp == self) {
672 			/* dequeue ourself */
673 			ASSERT(self->ul_wchan == wchan);
674 			(void) queue_unlink(qp, ulwpp, prev);
675 			self->ul_cvmutex = NULL;
676 			self->ul_cv_wake = 0;
677 			found = 1;
678 			break;
679 		}
680 		if (ulwp->ul_wchan == wchan)
681 			more = 1;
682 	}
683 
684 	if (!found)
685 		thr_panic("dequeue_self(): curthread not found on queue");
686 
687 	if (more)
688 		return (1);
689 
690 	/* scan the remainder of the queue for another waiter */
691 	for (ulwp = *ulwpp; ulwp != NULL; ulwp = ulwp->ul_link) {
692 		if (ulwp->ul_wchan == wchan)
693 			return (1);
694 	}
695 
696 	return (0);
697 }
698 
699 /*
700  * Called from call_user_handler() and _thrp_suspend() to take
701  * ourself off of our sleep queue so we can grab locks.
702  */
703 void
704 unsleep_self(void)
705 {
706 	ulwp_t *self = curthread;
707 	queue_head_t *qp;
708 
709 	/*
710 	 * Calling enter_critical()/exit_critical() here would lead
711 	 * to recursion.  Just manipulate self->ul_critical directly.
712 	 */
713 	self->ul_critical++;
714 	while (self->ul_sleepq != NULL) {
715 		qp = queue_lock(self->ul_wchan, self->ul_qtype);
716 		/*
717 		 * We may have been moved from a CV queue to a
718 		 * mutex queue while we were attempting queue_lock().
719 		 * If so, just loop around and try again.
720 		 * dequeue_self() clears self->ul_sleepq.
721 		 */
722 		if (qp == self->ul_sleepq) {
723 			(void) dequeue_self(qp, self->ul_wchan);
724 			self->ul_writer = 0;
725 		}
726 		queue_unlock(qp);
727 	}
728 	self->ul_critical--;
729 }
730 
731 /*
732  * Common code for calling the the ___lwp_mutex_timedlock() system call.
733  * Returns with mutex_owner and mutex_ownerpid set correctly.
734  */
735 static int
736 mutex_lock_kernel(mutex_t *mp, timespec_t *tsp, tdb_mutex_stats_t *msp)
737 {
738 	ulwp_t *self = curthread;
739 	uberdata_t *udp = self->ul_uberdata;
740 	int mtype = mp->mutex_type;
741 	hrtime_t begin_sleep;
742 	int acquired;
743 	int error;
744 
745 	self->ul_sp = stkptr();
746 	self->ul_wchan = mp;
747 	if (__td_event_report(self, TD_SLEEP, udp)) {
748 		self->ul_td_evbuf.eventnum = TD_SLEEP;
749 		self->ul_td_evbuf.eventdata = mp;
750 		tdb_event(TD_SLEEP, udp);
751 	}
752 	if (msp) {
753 		tdb_incr(msp->mutex_sleep);
754 		begin_sleep = gethrtime();
755 	}
756 
757 	DTRACE_PROBE1(plockstat, mutex__block, mp);
758 
759 	for (;;) {
760 		/*
761 		 * A return value of EOWNERDEAD or ELOCKUNMAPPED
762 		 * means we successfully acquired the lock.
763 		 */
764 		if ((error = ___lwp_mutex_timedlock(mp, tsp)) != 0 &&
765 		    error != EOWNERDEAD && error != ELOCKUNMAPPED) {
766 			acquired = 0;
767 			break;
768 		}
769 
770 		if (mtype & USYNC_PROCESS) {
771 			/*
772 			 * Defend against forkall().  We may be the child,
773 			 * in which case we don't actually own the mutex.
774 			 */
775 			enter_critical(self);
776 			if (mp->mutex_ownerpid == udp->pid) {
777 				mp->mutex_owner = (uintptr_t)self;
778 				exit_critical(self);
779 				acquired = 1;
780 				break;
781 			}
782 			exit_critical(self);
783 		} else {
784 			mp->mutex_owner = (uintptr_t)self;
785 			acquired = 1;
786 			break;
787 		}
788 	}
789 	if (msp)
790 		msp->mutex_sleep_time += gethrtime() - begin_sleep;
791 	self->ul_wchan = NULL;
792 	self->ul_sp = 0;
793 
794 	if (acquired) {
795 		DTRACE_PROBE2(plockstat, mutex__blocked, mp, 1);
796 		DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
797 	} else {
798 		DTRACE_PROBE2(plockstat, mutex__blocked, mp, 0);
799 		DTRACE_PROBE2(plockstat, mutex__error, mp, error);
800 	}
801 
802 	return (error);
803 }
804 
805 /*
806  * Common code for calling the ___lwp_mutex_trylock() system call.
807  * Returns with mutex_owner and mutex_ownerpid set correctly.
808  */
809 int
810 mutex_trylock_kernel(mutex_t *mp)
811 {
812 	ulwp_t *self = curthread;
813 	uberdata_t *udp = self->ul_uberdata;
814 	int mtype = mp->mutex_type;
815 	int error;
816 	int acquired;
817 
818 	for (;;) {
819 		/*
820 		 * A return value of EOWNERDEAD or ELOCKUNMAPPED
821 		 * means we successfully acquired the lock.
822 		 */
823 		if ((error = ___lwp_mutex_trylock(mp)) != 0 &&
824 		    error != EOWNERDEAD && error != ELOCKUNMAPPED) {
825 			acquired = 0;
826 			break;
827 		}
828 
829 		if (mtype & USYNC_PROCESS) {
830 			/*
831 			 * Defend against forkall().  We may be the child,
832 			 * in which case we don't actually own the mutex.
833 			 */
834 			enter_critical(self);
835 			if (mp->mutex_ownerpid == udp->pid) {
836 				mp->mutex_owner = (uintptr_t)self;
837 				exit_critical(self);
838 				acquired = 1;
839 				break;
840 			}
841 			exit_critical(self);
842 		} else {
843 			mp->mutex_owner = (uintptr_t)self;
844 			acquired = 1;
845 			break;
846 		}
847 	}
848 
849 	if (acquired) {
850 		DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
851 	} else if (error != EBUSY) {
852 		DTRACE_PROBE2(plockstat, mutex__error, mp, error);
853 	}
854 
855 	return (error);
856 }
857 
858 volatile sc_shared_t *
859 setup_schedctl(void)
860 {
861 	ulwp_t *self = curthread;
862 	volatile sc_shared_t *scp;
863 	sc_shared_t *tmp;
864 
865 	if ((scp = self->ul_schedctl) == NULL && /* no shared state yet */
866 	    !self->ul_vfork &&			/* not a child of vfork() */
867 	    !self->ul_schedctl_called) {	/* haven't been called before */
868 		enter_critical(self);
869 		self->ul_schedctl_called = &self->ul_uberdata->uberflags;
870 		if ((tmp = __schedctl()) != (sc_shared_t *)(-1))
871 			self->ul_schedctl = scp = tmp;
872 		exit_critical(self);
873 	}
874 	/*
875 	 * Unless the call to setup_schedctl() is surrounded
876 	 * by enter_critical()/exit_critical(), the address
877 	 * we are returning could be invalid due to a forkall()
878 	 * having occurred in another thread.
879 	 */
880 	return (scp);
881 }
882 
883 /*
884  * Interfaces from libsched, incorporated into libc.
885  * libsched.so.1 is now a filter library onto libc.
886  */
887 #pragma weak schedctl_lookup = _schedctl_init
888 #pragma weak _schedctl_lookup = _schedctl_init
889 #pragma weak schedctl_init = _schedctl_init
890 schedctl_t *
891 _schedctl_init(void)
892 {
893 	volatile sc_shared_t *scp = setup_schedctl();
894 	return ((scp == NULL)? NULL : (schedctl_t *)&scp->sc_preemptctl);
895 }
896 
897 #pragma weak schedctl_exit = _schedctl_exit
898 void
899 _schedctl_exit(void)
900 {
901 }
902 
903 /*
904  * Contract private interface for java.
905  * Set up the schedctl data if it doesn't exist yet.
906  * Return a pointer to the pointer to the schedctl data.
907  */
908 volatile sc_shared_t *volatile *
909 _thr_schedctl(void)
910 {
911 	ulwp_t *self = curthread;
912 	volatile sc_shared_t *volatile *ptr;
913 
914 	if (self->ul_vfork)
915 		return (NULL);
916 	if (*(ptr = &self->ul_schedctl) == NULL)
917 		(void) setup_schedctl();
918 	return (ptr);
919 }
920 
921 /*
922  * Block signals and attempt to block preemption.
923  * no_preempt()/preempt() must be used in pairs but can be nested.
924  */
925 void
926 no_preempt(ulwp_t *self)
927 {
928 	volatile sc_shared_t *scp;
929 
930 	if (self->ul_preempt++ == 0) {
931 		enter_critical(self);
932 		if ((scp = self->ul_schedctl) != NULL ||
933 		    (scp = setup_schedctl()) != NULL) {
934 			/*
935 			 * Save the pre-existing preempt value.
936 			 */
937 			self->ul_savpreempt = scp->sc_preemptctl.sc_nopreempt;
938 			scp->sc_preemptctl.sc_nopreempt = 1;
939 		}
940 	}
941 }
942 
943 /*
944  * Undo the effects of no_preempt().
945  */
946 void
947 preempt(ulwp_t *self)
948 {
949 	volatile sc_shared_t *scp;
950 
951 	ASSERT(self->ul_preempt > 0);
952 	if (--self->ul_preempt == 0) {
953 		if ((scp = self->ul_schedctl) != NULL) {
954 			/*
955 			 * Restore the pre-existing preempt value.
956 			 */
957 			scp->sc_preemptctl.sc_nopreempt = self->ul_savpreempt;
958 			if (scp->sc_preemptctl.sc_yield &&
959 			    scp->sc_preemptctl.sc_nopreempt == 0) {
960 				lwp_yield();
961 				if (scp->sc_preemptctl.sc_yield) {
962 					/*
963 					 * Shouldn't happen.  This is either
964 					 * a race condition or the thread
965 					 * just entered the real-time class.
966 					 */
967 					lwp_yield();
968 					scp->sc_preemptctl.sc_yield = 0;
969 				}
970 			}
971 		}
972 		exit_critical(self);
973 	}
974 }
975 
976 /*
977  * If a call to preempt() would cause the current thread to yield or to
978  * take deferred actions in exit_critical(), then unpark the specified
979  * lwp so it can run while we delay.  Return the original lwpid if the
980  * unpark was not performed, else return zero.  The tests are a repeat
981  * of some of the tests in preempt(), above.  This is a statistical
982  * optimization solely for cond_sleep_queue(), below.
983  */
984 static lwpid_t
985 preempt_unpark(ulwp_t *self, lwpid_t lwpid)
986 {
987 	volatile sc_shared_t *scp = self->ul_schedctl;
988 
989 	ASSERT(self->ul_preempt == 1 && self->ul_critical > 0);
990 	if ((scp != NULL && scp->sc_preemptctl.sc_yield) ||
991 	    (self->ul_curplease && self->ul_critical == 1)) {
992 		(void) __lwp_unpark(lwpid);
993 		lwpid = 0;
994 	}
995 	return (lwpid);
996 }
997 
998 /*
999  * Spin for a while, trying to grab the lock.
1000  * If this fails, return EBUSY and let the caller deal with it.
1001  * If this succeeds, return 0 with mutex_owner set to curthread.
1002  */
1003 static int
1004 mutex_trylock_adaptive(mutex_t *mp)
1005 {
1006 	ulwp_t *self = curthread;
1007 	int error = EBUSY;
1008 	ulwp_t *ulwp;
1009 	volatile sc_shared_t *scp;
1010 	volatile uint8_t *lockp;
1011 	volatile uint64_t *ownerp;
1012 	int count;
1013 	int max;
1014 
1015 	ASSERT(!(mp->mutex_type & USYNC_PROCESS));
1016 
1017 	if (MUTEX_OWNER(mp) == self)
1018 		return (EBUSY);
1019 
1020 	/* short-cut, not definitive (see below) */
1021 	if (mp->mutex_flag & LOCK_NOTRECOVERABLE) {
1022 		ASSERT(mp->mutex_type & LOCK_ROBUST);
1023 		DTRACE_PROBE2(plockstat, mutex__error, mp, ENOTRECOVERABLE);
1024 		return (ENOTRECOVERABLE);
1025 	}
1026 
1027 	if ((max = self->ul_adaptive_spin) == 0 ||
1028 	    mp->mutex_spinners >= self->ul_max_spinners)
1029 		max = 1;	/* try at least once */
1030 
1031 	DTRACE_PROBE1(plockstat, mutex__spin, mp);
1032 
1033 	lockp = (volatile uint8_t *)&mp->mutex_lockw;
1034 	ownerp = (volatile uint64_t *)&mp->mutex_owner;
1035 	/*
1036 	 * This spin loop is unfair to lwps that have already dropped into
1037 	 * the kernel to sleep.  They will starve on a highly-contended mutex.
1038 	 * This is just too bad.  The adaptive spin algorithm is intended
1039 	 * to allow programs with highly-contended locks (that is, broken
1040 	 * programs) to execute with reasonable speed despite their contention.
1041 	 * Being fair would reduce the speed of such programs and well-written
1042 	 * programs will not suffer in any case.
1043 	 */
1044 	enter_critical(self);		/* protects ul_schedctl */
1045 	atomic_inc_32(&mp->mutex_spinners);
1046 	for (count = 1; count <= max; count++) {
1047 		if (*lockp == 0 && set_lock_byte(lockp) == 0) {
1048 			*ownerp = (uintptr_t)self;
1049 			error = 0;
1050 			break;
1051 		}
1052 		SMT_PAUSE();
1053 		/*
1054 		 * Stop spinning if the mutex owner is not running on
1055 		 * a processor; it will not drop the lock any time soon
1056 		 * and we would just be wasting time to keep spinning.
1057 		 *
1058 		 * Note that we are looking at another thread (ulwp_t)
1059 		 * without ensuring that the other thread does not exit.
1060 		 * The scheme relies on ulwp_t structures never being
1061 		 * deallocated by the library (the library employs a free
1062 		 * list of ulwp_t structs that are reused when new threads
1063 		 * are created) and on schedctl shared memory never being
1064 		 * deallocated once created via __schedctl().
1065 		 *
1066 		 * Thus, the worst that can happen when the spinning thread
1067 		 * looks at the owner's schedctl data is that it is looking
1068 		 * at some other thread's schedctl data.  This almost never
1069 		 * happens and is benign when it does.
1070 		 */
1071 		if ((ulwp = (ulwp_t *)(uintptr_t)*ownerp) != NULL &&
1072 		    ((scp = ulwp->ul_schedctl) == NULL ||
1073 		    scp->sc_state != SC_ONPROC))
1074 			break;
1075 	}
1076 	atomic_dec_32(&mp->mutex_spinners);
1077 	exit_critical(self);
1078 
1079 	if (error == 0 && (mp->mutex_flag & LOCK_NOTRECOVERABLE)) {
1080 		ASSERT(mp->mutex_type & LOCK_ROBUST);
1081 		/*
1082 		 * We shouldn't own the mutex; clear the lock.
1083 		 */
1084 		mp->mutex_owner = 0;
1085 		if (atomic_swap_32(&mp->mutex_lockword, 0) & WAITERMASK)
1086 			mutex_wakeup_all(mp);
1087 		error = ENOTRECOVERABLE;
1088 	}
1089 
1090 	if (error) {
1091 		DTRACE_PROBE2(plockstat, mutex__spun, 0, count);
1092 		if (error != EBUSY) {
1093 			DTRACE_PROBE2(plockstat, mutex__error, mp, error);
1094 		}
1095 	} else {
1096 		DTRACE_PROBE2(plockstat, mutex__spun, 1, count);
1097 		DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, count);
1098 		if (mp->mutex_flag & LOCK_OWNERDEAD) {
1099 			ASSERT(mp->mutex_type & LOCK_ROBUST);
1100 			error = EOWNERDEAD;
1101 		}
1102 	}
1103 
1104 	return (error);
1105 }
1106 
1107 /*
1108  * Same as mutex_trylock_adaptive(), except specifically for queue locks.
1109  * The owner field is not set here; the caller (spin_lock_set()) sets it.
1110  */
1111 static int
1112 mutex_queuelock_adaptive(mutex_t *mp)
1113 {
1114 	ulwp_t *ulwp;
1115 	volatile sc_shared_t *scp;
1116 	volatile uint8_t *lockp;
1117 	volatile uint64_t *ownerp;
1118 	int count = curthread->ul_queue_spin;
1119 
1120 	ASSERT(mp->mutex_type == USYNC_THREAD);
1121 
1122 	if (count == 0)
1123 		return (EBUSY);
1124 
1125 	lockp = (volatile uint8_t *)&mp->mutex_lockw;
1126 	ownerp = (volatile uint64_t *)&mp->mutex_owner;
1127 	while (--count >= 0) {
1128 		if (*lockp == 0 && set_lock_byte(lockp) == 0)
1129 			return (0);
1130 		SMT_PAUSE();
1131 		if ((ulwp = (ulwp_t *)(uintptr_t)*ownerp) != NULL &&
1132 		    ((scp = ulwp->ul_schedctl) == NULL ||
1133 		    scp->sc_state != SC_ONPROC))
1134 			break;
1135 	}
1136 
1137 	return (EBUSY);
1138 }
1139 
1140 /*
1141  * Like mutex_trylock_adaptive(), but for process-shared mutexes.
1142  * Spin for a while, trying to grab the lock.
1143  * If this fails, return EBUSY and let the caller deal with it.
1144  * If this succeeds, return 0 with mutex_owner set to curthread
1145  * and mutex_ownerpid set to the current pid.
1146  */
1147 static int
1148 mutex_trylock_process(mutex_t *mp)
1149 {
1150 	ulwp_t *self = curthread;
1151 	int error = EBUSY;
1152 	volatile uint8_t *lockp;
1153 	int count;
1154 	int max;
1155 
1156 	ASSERT(mp->mutex_type & USYNC_PROCESS);
1157 
1158 	if (shared_mutex_held(mp))
1159 		return (EBUSY);
1160 
1161 	/* short-cut, not definitive (see below) */
1162 	if (mp->mutex_flag & LOCK_NOTRECOVERABLE) {
1163 		ASSERT(mp->mutex_type & LOCK_ROBUST);
1164 		DTRACE_PROBE2(plockstat, mutex__error, mp, ENOTRECOVERABLE);
1165 		return (ENOTRECOVERABLE);
1166 	}
1167 
1168 	if (ncpus == 0)
1169 		ncpus = (int)_sysconf(_SC_NPROCESSORS_ONLN);
1170 	max = (ncpus > 1)? self->ul_adaptive_spin : 1;
1171 	if (max == 0)
1172 		max = 1;	/* try at least once */
1173 
1174 	DTRACE_PROBE1(plockstat, mutex__spin, mp);
1175 
1176 	lockp = (volatile uint8_t *)&mp->mutex_lockw;
1177 	/*
1178 	 * This is a process-shared mutex.
1179 	 * We cannot know if the owner is running on a processor.
1180 	 * We just spin and hope that it is on a processor.
1181 	 */
1182 	enter_critical(self);
1183 	for (count = 1; count <= max; count++) {
1184 		if (*lockp == 0 && set_lock_byte(lockp) == 0) {
1185 			mp->mutex_owner = (uintptr_t)self;
1186 			mp->mutex_ownerpid = self->ul_uberdata->pid;
1187 			error = 0;
1188 			break;
1189 		}
1190 		SMT_PAUSE();
1191 	}
1192 	exit_critical(self);
1193 
1194 	if (error == 0 && (mp->mutex_flag & LOCK_NOTRECOVERABLE)) {
1195 		ASSERT(mp->mutex_type & LOCK_ROBUST);
1196 		/*
1197 		 * We shouldn't own the mutex; clear the lock.
1198 		 */
1199 		mp->mutex_owner = 0;
1200 		mp->mutex_ownerpid = 0;
1201 		if (atomic_swap_32(&mp->mutex_lockword, 0) & WAITERMASK) {
1202 			no_preempt(self);
1203 			(void) ___lwp_mutex_wakeup(mp, 1);
1204 			preempt(self);
1205 		}
1206 		error = ENOTRECOVERABLE;
1207 	}
1208 
1209 	if (error) {
1210 		DTRACE_PROBE2(plockstat, mutex__spun, 0, count);
1211 		if (error != EBUSY) {
1212 			DTRACE_PROBE2(plockstat, mutex__error, mp, error);
1213 		}
1214 	} else {
1215 		DTRACE_PROBE2(plockstat, mutex__spun, 1, count);
1216 		DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, count);
1217 		if (mp->mutex_flag & (LOCK_OWNERDEAD | LOCK_UNMAPPED)) {
1218 			ASSERT(mp->mutex_type & LOCK_ROBUST);
1219 			if (mp->mutex_flag & LOCK_OWNERDEAD)
1220 				error = EOWNERDEAD;
1221 			else if (mp->mutex_type & USYNC_PROCESS_ROBUST)
1222 				error = ELOCKUNMAPPED;
1223 			else
1224 				error = EOWNERDEAD;
1225 		}
1226 	}
1227 
1228 	return (error);
1229 }
1230 
1231 /*
1232  * Mutex wakeup code for releasing a USYNC_THREAD mutex.
1233  * Returns the lwpid of the thread that was dequeued, if any.
1234  * The caller of mutex_wakeup() must call __lwp_unpark(lwpid)
1235  * to wake up the specified lwp.
1236  */
1237 static lwpid_t
1238 mutex_wakeup(mutex_t *mp)
1239 {
1240 	lwpid_t lwpid = 0;
1241 	queue_head_t *qp;
1242 	ulwp_t *ulwp;
1243 	int more;
1244 
1245 	/*
1246 	 * Dequeue a waiter from the sleep queue.  Don't touch the mutex
1247 	 * waiters bit if no one was found on the queue because the mutex
1248 	 * might have been deallocated or reallocated for another purpose.
1249 	 */
1250 	qp = queue_lock(mp, MX);
1251 	if ((ulwp = dequeue(qp, mp, &more)) != NULL) {
1252 		lwpid = ulwp->ul_lwpid;
1253 		mp->mutex_waiters = (more? 1 : 0);
1254 	}
1255 	queue_unlock(qp);
1256 	return (lwpid);
1257 }
1258 
1259 /*
1260  * Mutex wakeup code for releasing all waiters on a USYNC_THREAD mutex.
1261  */
1262 static void
1263 mutex_wakeup_all(mutex_t *mp)
1264 {
1265 	queue_head_t *qp;
1266 	int nlwpid = 0;
1267 	int maxlwps = MAXLWPS;
1268 	ulwp_t **ulwpp;
1269 	ulwp_t *ulwp;
1270 	ulwp_t *prev = NULL;
1271 	lwpid_t buffer[MAXLWPS];
1272 	lwpid_t *lwpid = buffer;
1273 
1274 	/*
1275 	 * Walk the list of waiters and prepare to wake up all of them.
1276 	 * The waiters flag has already been cleared from the mutex.
1277 	 *
1278 	 * We keep track of lwpids that are to be unparked in lwpid[].
1279 	 * __lwp_unpark_all() is called to unpark all of them after
1280 	 * they have been removed from the sleep queue and the sleep
1281 	 * queue lock has been dropped.  If we run out of space in our
1282 	 * on-stack buffer, we need to allocate more but we can't call
1283 	 * lmalloc() because we are holding a queue lock when the overflow
1284 	 * occurs and lmalloc() acquires a lock.  We can't use alloca()
1285 	 * either because the application may have allocated a small
1286 	 * stack and we don't want to overrun the stack.  So we call
1287 	 * alloc_lwpids() to allocate a bigger buffer using the mmap()
1288 	 * system call directly since that path acquires no locks.
1289 	 */
1290 	qp = queue_lock(mp, MX);
1291 	ulwpp = &qp->qh_head;
1292 	while ((ulwp = *ulwpp) != NULL) {
1293 		if (ulwp->ul_wchan != mp) {
1294 			prev = ulwp;
1295 			ulwpp = &ulwp->ul_link;
1296 		} else {
1297 			if (nlwpid == maxlwps)
1298 				lwpid = alloc_lwpids(lwpid, &nlwpid, &maxlwps);
1299 			(void) queue_unlink(qp, ulwpp, prev);
1300 			lwpid[nlwpid++] = ulwp->ul_lwpid;
1301 		}
1302 	}
1303 	mp->mutex_waiters = 0;
1304 
1305 	if (nlwpid == 0) {
1306 		queue_unlock(qp);
1307 	} else {
1308 		no_preempt(curthread);
1309 		queue_unlock(qp);
1310 		if (nlwpid == 1)
1311 			(void) __lwp_unpark(lwpid[0]);
1312 		else
1313 			(void) __lwp_unpark_all(lwpid, nlwpid);
1314 		preempt(curthread);
1315 	}
1316 
1317 	if (lwpid != buffer)
1318 		(void) _private_munmap(lwpid, maxlwps * sizeof (lwpid_t));
1319 }
1320 
1321 /*
1322  * Spin for a while, testing to see if the lock has been grabbed.
1323  * If this fails, call mutex_wakeup() to release a waiter.
1324  */
1325 static lwpid_t
1326 mutex_unlock_queue(mutex_t *mp, int release_all)
1327 {
1328 	ulwp_t *self = curthread;
1329 	uint32_t *lockw = &mp->mutex_lockword;
1330 	lwpid_t lwpid;
1331 	volatile uint8_t *lockp;
1332 	volatile uint32_t *spinp;
1333 	int count;
1334 
1335 	/*
1336 	 * We use the swap primitive to clear the lock, but we must
1337 	 * atomically retain the waiters bit for the remainder of this
1338 	 * code to work.  We first check to see if the waiters bit is
1339 	 * set and if so clear the lock by swapping in a word containing
1340 	 * only the waiters bit.  This could produce a false positive test
1341 	 * for whether there are waiters that need to be waked up, but
1342 	 * this just causes an extra call to mutex_wakeup() to do nothing.
1343 	 * The opposite case is more delicate:  If there are no waiters,
1344 	 * we swap in a zero lock byte and a zero waiters bit.  The result
1345 	 * of the swap could indicate that there really was a waiter so in
1346 	 * this case we go directly to mutex_wakeup() without performing
1347 	 * any of the adaptive code because the waiter bit has been cleared
1348 	 * and the adaptive code is unreliable in this case.
1349 	 */
1350 	if (release_all || !(*lockw & WAITERMASK)) {
1351 		mp->mutex_owner = 0;
1352 		DTRACE_PROBE2(plockstat, mutex__release, mp, 0);
1353 		if (!(atomic_swap_32(lockw, 0) & WAITERMASK))
1354 			return (0);	/* no waiters */
1355 		no_preempt(self);	/* ensure a prompt wakeup */
1356 	} else {
1357 		no_preempt(self);	/* ensure a prompt wakeup */
1358 		lockp = (volatile uint8_t *)&mp->mutex_lockw;
1359 		spinp = (volatile uint32_t *)&mp->mutex_spinners;
1360 		mp->mutex_owner = 0;
1361 		DTRACE_PROBE2(plockstat, mutex__release, mp, 0);
1362 		/* clear lock, retain waiter */
1363 		(void) atomic_swap_32(lockw, WAITER);
1364 
1365 		/*
1366 		 * We spin here fewer times than mutex_trylock_adaptive().
1367 		 * We are trying to balance two conflicting goals:
1368 		 * 1. Avoid waking up anyone if a spinning thread
1369 		 *    grabs the lock.
1370 		 * 2. Wake up a sleeping thread promptly to get on
1371 		 *    with useful work.
1372 		 * We don't spin at all if there is no acquiring spinner;
1373 		 * (mp->mutex_spinners is non-zero if there are spinners).
1374 		 */
1375 		for (count = self->ul_release_spin;
1376 		    *spinp && count > 0; count--) {
1377 			/*
1378 			 * There is a waiter that we will have to wake
1379 			 * up unless someone else grabs the lock while
1380 			 * we are busy spinning.  Like the spin loop in
1381 			 * mutex_trylock_adaptive(), this spin loop is
1382 			 * unfair to lwps that have already dropped into
1383 			 * the kernel to sleep.  They will starve on a
1384 			 * highly-contended mutex.  Too bad.
1385 			 */
1386 			if (*lockp != 0) {	/* somebody grabbed the lock */
1387 				preempt(self);
1388 				return (0);
1389 			}
1390 			SMT_PAUSE();
1391 		}
1392 
1393 		/*
1394 		 * No one grabbed the lock.
1395 		 * Wake up some lwp that is waiting for it.
1396 		 */
1397 		mp->mutex_waiters = 0;
1398 	}
1399 
1400 	if (release_all) {
1401 		mutex_wakeup_all(mp);
1402 		lwpid = 0;
1403 	} else {
1404 		lwpid = mutex_wakeup(mp);
1405 	}
1406 	if (lwpid == 0)
1407 		preempt(self);
1408 	return (lwpid);
1409 }
1410 
1411 /*
1412  * Like mutex_unlock_queue(), but for process-shared mutexes.
1413  * We tested the waiters field before calling here and it was non-zero.
1414  */
1415 static void
1416 mutex_unlock_process(mutex_t *mp, int release_all)
1417 {
1418 	ulwp_t *self = curthread;
1419 	int count;
1420 	volatile uint8_t *lockp;
1421 
1422 	/*
1423 	 * See the comments in mutex_unlock_queue(), above.
1424 	 */
1425 	if (ncpus == 0)
1426 		ncpus = (int)_sysconf(_SC_NPROCESSORS_ONLN);
1427 	count = (ncpus > 1)? self->ul_release_spin : 0;
1428 	no_preempt(self);
1429 	mp->mutex_owner = 0;
1430 	mp->mutex_ownerpid = 0;
1431 	DTRACE_PROBE2(plockstat, mutex__release, mp, 0);
1432 	if (release_all || count == 0) {
1433 		/* clear lock, test waiter */
1434 		if (!(atomic_swap_32(&mp->mutex_lockword, 0) & WAITERMASK)) {
1435 			/* no waiters now */
1436 			preempt(self);
1437 			return;
1438 		}
1439 	} else {
1440 		/* clear lock, retain waiter */
1441 		(void) atomic_swap_32(&mp->mutex_lockword, WAITER);
1442 		lockp = (volatile uint8_t *)&mp->mutex_lockw;
1443 		while (--count >= 0) {
1444 			if (*lockp != 0) {
1445 				/* somebody grabbed the lock */
1446 				preempt(self);
1447 				return;
1448 			}
1449 			SMT_PAUSE();
1450 		}
1451 		/*
1452 		 * We must clear the waiters field before going
1453 		 * to the kernel, else it could remain set forever.
1454 		 */
1455 		mp->mutex_waiters = 0;
1456 	}
1457 	(void) ___lwp_mutex_wakeup(mp, release_all);
1458 	preempt(self);
1459 }
1460 
1461 /*
1462  * Return the real priority of a thread.
1463  */
1464 int
1465 real_priority(ulwp_t *ulwp)
1466 {
1467 	if (ulwp->ul_epri == 0)
1468 		return (ulwp->ul_mappedpri? ulwp->ul_mappedpri : ulwp->ul_pri);
1469 	return (ulwp->ul_emappedpri? ulwp->ul_emappedpri : ulwp->ul_epri);
1470 }
1471 
1472 void
1473 stall(void)
1474 {
1475 	for (;;)
1476 		(void) mutex_lock_kernel(&stall_mutex, NULL, NULL);
1477 }
1478 
1479 /*
1480  * Acquire a USYNC_THREAD mutex via user-level sleep queues.
1481  * We failed set_lock_byte(&mp->mutex_lockw) before coming here.
1482  * If successful, returns with mutex_owner set correctly.
1483  */
1484 int
1485 mutex_lock_queue(ulwp_t *self, tdb_mutex_stats_t *msp, mutex_t *mp,
1486 	timespec_t *tsp)
1487 {
1488 	uberdata_t *udp = curthread->ul_uberdata;
1489 	queue_head_t *qp;
1490 	hrtime_t begin_sleep;
1491 	int error = 0;
1492 
1493 	self->ul_sp = stkptr();
1494 	if (__td_event_report(self, TD_SLEEP, udp)) {
1495 		self->ul_wchan = mp;
1496 		self->ul_td_evbuf.eventnum = TD_SLEEP;
1497 		self->ul_td_evbuf.eventdata = mp;
1498 		tdb_event(TD_SLEEP, udp);
1499 	}
1500 	if (msp) {
1501 		tdb_incr(msp->mutex_sleep);
1502 		begin_sleep = gethrtime();
1503 	}
1504 
1505 	DTRACE_PROBE1(plockstat, mutex__block, mp);
1506 
1507 	/*
1508 	 * Put ourself on the sleep queue, and while we are
1509 	 * unable to grab the lock, go park in the kernel.
1510 	 * Take ourself off the sleep queue after we acquire the lock.
1511 	 * The waiter bit can be set/cleared only while holding the queue lock.
1512 	 */
1513 	qp = queue_lock(mp, MX);
1514 	enqueue(qp, self, mp, MX);
1515 	mp->mutex_waiters = 1;
1516 	for (;;) {
1517 		if (set_lock_byte(&mp->mutex_lockw) == 0) {
1518 			mp->mutex_owner = (uintptr_t)self;
1519 			mp->mutex_waiters = dequeue_self(qp, mp);
1520 			break;
1521 		}
1522 		set_parking_flag(self, 1);
1523 		queue_unlock(qp);
1524 		/*
1525 		 * __lwp_park() will return the residual time in tsp
1526 		 * if we are unparked before the timeout expires.
1527 		 */
1528 		if ((error = __lwp_park(tsp, 0)) == EINTR)
1529 			error = 0;
1530 		set_parking_flag(self, 0);
1531 		/*
1532 		 * We could have taken a signal or suspended ourself.
1533 		 * If we did, then we removed ourself from the queue.
1534 		 * Someone else may have removed us from the queue
1535 		 * as a consequence of mutex_unlock().  We may have
1536 		 * gotten a timeout from __lwp_park().  Or we may still
1537 		 * be on the queue and this is just a spurious wakeup.
1538 		 */
1539 		qp = queue_lock(mp, MX);
1540 		if (self->ul_sleepq == NULL) {
1541 			if (error)
1542 				break;
1543 			if (set_lock_byte(&mp->mutex_lockw) == 0) {
1544 				mp->mutex_owner = (uintptr_t)self;
1545 				break;
1546 			}
1547 			enqueue(qp, self, mp, MX);
1548 			mp->mutex_waiters = 1;
1549 		}
1550 		ASSERT(self->ul_sleepq == qp &&
1551 		    self->ul_qtype == MX &&
1552 		    self->ul_wchan == mp);
1553 		if (error) {
1554 			mp->mutex_waiters = dequeue_self(qp, mp);
1555 			break;
1556 		}
1557 	}
1558 	ASSERT(self->ul_sleepq == NULL && self->ul_link == NULL &&
1559 	    self->ul_wchan == NULL);
1560 	self->ul_sp = 0;
1561 	queue_unlock(qp);
1562 
1563 	if (msp)
1564 		msp->mutex_sleep_time += gethrtime() - begin_sleep;
1565 
1566 	ASSERT(error == 0 || error == EINVAL || error == ETIME);
1567 
1568 	if (error == 0 && (mp->mutex_flag & LOCK_NOTRECOVERABLE)) {
1569 		ASSERT(mp->mutex_type & LOCK_ROBUST);
1570 		/*
1571 		 * We shouldn't own the mutex; clear the lock.
1572 		 */
1573 		mp->mutex_owner = 0;
1574 		if (atomic_swap_32(&mp->mutex_lockword, 0) & WAITERMASK)
1575 			mutex_wakeup_all(mp);
1576 		error = ENOTRECOVERABLE;
1577 	}
1578 
1579 	if (error) {
1580 		DTRACE_PROBE2(plockstat, mutex__blocked, mp, 0);
1581 		DTRACE_PROBE2(plockstat, mutex__error, mp, error);
1582 	} else {
1583 		DTRACE_PROBE2(plockstat, mutex__blocked, mp, 1);
1584 		DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
1585 		if (mp->mutex_flag & LOCK_OWNERDEAD) {
1586 			ASSERT(mp->mutex_type & LOCK_ROBUST);
1587 			error = EOWNERDEAD;
1588 		}
1589 	}
1590 
1591 	return (error);
1592 }
1593 
1594 static int
1595 mutex_recursion(mutex_t *mp, int mtype, int try)
1596 {
1597 	ASSERT(mutex_is_held(mp));
1598 	ASSERT(mtype & (LOCK_RECURSIVE|LOCK_ERRORCHECK));
1599 	ASSERT(try == MUTEX_TRY || try == MUTEX_LOCK);
1600 
1601 	if (mtype & LOCK_RECURSIVE) {
1602 		if (mp->mutex_rcount == RECURSION_MAX) {
1603 			DTRACE_PROBE2(plockstat, mutex__error, mp, EAGAIN);
1604 			return (EAGAIN);
1605 		}
1606 		mp->mutex_rcount++;
1607 		DTRACE_PROBE3(plockstat, mutex__acquire, mp, 1, 0);
1608 		return (0);
1609 	}
1610 	if (try == MUTEX_LOCK) {
1611 		DTRACE_PROBE2(plockstat, mutex__error, mp, EDEADLK);
1612 		return (EDEADLK);
1613 	}
1614 	return (EBUSY);
1615 }
1616 
1617 /*
1618  * Register this USYNC_PROCESS|LOCK_ROBUST mutex with the kernel so
1619  * it can apply LOCK_OWNERDEAD|LOCK_UNMAPPED if it becomes necessary.
1620  * We use tdb_hash_lock here and in the synch object tracking code in
1621  * the tdb_agent.c file.  There is no conflict between these two usages.
1622  */
1623 void
1624 register_lock(mutex_t *mp)
1625 {
1626 	uberdata_t *udp = curthread->ul_uberdata;
1627 	uint_t hash = LOCK_HASH(mp);
1628 	robust_t *rlp;
1629 	robust_t **rlpp;
1630 	robust_t **table;
1631 
1632 	if ((table = udp->robustlocks) == NULL) {
1633 		lmutex_lock(&udp->tdb_hash_lock);
1634 		if ((table = udp->robustlocks) == NULL) {
1635 			table = lmalloc(LOCKHASHSZ * sizeof (robust_t *));
1636 			_membar_producer();
1637 			udp->robustlocks = table;
1638 		}
1639 		lmutex_unlock(&udp->tdb_hash_lock);
1640 	}
1641 	_membar_consumer();
1642 
1643 	/*
1644 	 * First search the registered table with no locks held.
1645 	 * This is safe because the table never shrinks
1646 	 * and we can only get a false negative.
1647 	 */
1648 	for (rlp = table[hash]; rlp != NULL; rlp = rlp->robust_next) {
1649 		if (rlp->robust_lock == mp)	/* already registered */
1650 			return;
1651 	}
1652 
1653 	/*
1654 	 * The lock was not found.
1655 	 * Repeat the operation with tdb_hash_lock held.
1656 	 */
1657 	lmutex_lock(&udp->tdb_hash_lock);
1658 
1659 	for (rlpp = &table[hash];
1660 	    (rlp = *rlpp) != NULL;
1661 	    rlpp = &rlp->robust_next) {
1662 		if (rlp->robust_lock == mp) {	/* already registered */
1663 			lmutex_unlock(&udp->tdb_hash_lock);
1664 			return;
1665 		}
1666 	}
1667 
1668 	/*
1669 	 * The lock has never been registered.
1670 	 * Register it now and add it to the table.
1671 	 */
1672 	(void) ___lwp_mutex_register(mp);
1673 	rlp = lmalloc(sizeof (*rlp));
1674 	rlp->robust_lock = mp;
1675 	_membar_producer();
1676 	*rlpp = rlp;
1677 
1678 	lmutex_unlock(&udp->tdb_hash_lock);
1679 }
1680 
1681 /*
1682  * This is called in the child of fork()/forkall() to start over
1683  * with a clean slate.  (Each process must register its own locks.)
1684  * No locks are needed because all other threads are suspended or gone.
1685  */
1686 void
1687 unregister_locks(void)
1688 {
1689 	uberdata_t *udp = curthread->ul_uberdata;
1690 	uint_t hash;
1691 	robust_t **table;
1692 	robust_t *rlp;
1693 	robust_t *next;
1694 
1695 	if ((table = udp->robustlocks) != NULL) {
1696 		for (hash = 0; hash < LOCKHASHSZ; hash++) {
1697 			rlp = table[hash];
1698 			while (rlp != NULL) {
1699 				next = rlp->robust_next;
1700 				lfree(rlp, sizeof (*rlp));
1701 				rlp = next;
1702 			}
1703 		}
1704 		lfree(table, LOCKHASHSZ * sizeof (robust_t *));
1705 		udp->robustlocks = NULL;
1706 	}
1707 }
1708 
1709 /*
1710  * Returns with mutex_owner set correctly.
1711  */
1712 static int
1713 mutex_lock_internal(mutex_t *mp, timespec_t *tsp, int try)
1714 {
1715 	ulwp_t *self = curthread;
1716 	uberdata_t *udp = self->ul_uberdata;
1717 	int mtype = mp->mutex_type;
1718 	tdb_mutex_stats_t *msp = MUTEX_STATS(mp, udp);
1719 	int error = 0;
1720 	uint8_t ceil;
1721 	int myprio;
1722 
1723 	ASSERT(try == MUTEX_TRY || try == MUTEX_LOCK);
1724 
1725 	if (!self->ul_schedctl_called)
1726 		(void) setup_schedctl();
1727 
1728 	if (msp && try == MUTEX_TRY)
1729 		tdb_incr(msp->mutex_try);
1730 
1731 	if ((mtype & (LOCK_RECURSIVE|LOCK_ERRORCHECK)) && mutex_is_held(mp))
1732 		return (mutex_recursion(mp, mtype, try));
1733 
1734 	if (self->ul_error_detection && try == MUTEX_LOCK &&
1735 	    tsp == NULL && mutex_is_held(mp))
1736 		lock_error(mp, "mutex_lock", NULL, NULL);
1737 
1738 	if (mtype & LOCK_PRIO_PROTECT) {
1739 		ceil = mp->mutex_ceiling;
1740 		ASSERT(_validate_rt_prio(SCHED_FIFO, ceil) == 0);
1741 		myprio = real_priority(self);
1742 		if (myprio > ceil) {
1743 			DTRACE_PROBE2(plockstat, mutex__error, mp, EINVAL);
1744 			return (EINVAL);
1745 		}
1746 		if ((error = _ceil_mylist_add(mp)) != 0) {
1747 			DTRACE_PROBE2(plockstat, mutex__error, mp, error);
1748 			return (error);
1749 		}
1750 		if (myprio < ceil)
1751 			_ceil_prio_inherit(ceil);
1752 	}
1753 
1754 	if ((mtype & (USYNC_PROCESS | LOCK_ROBUST))
1755 	    == (USYNC_PROCESS | LOCK_ROBUST))
1756 		register_lock(mp);
1757 
1758 	if (mtype & LOCK_PRIO_INHERIT) {
1759 		/* go straight to the kernel */
1760 		if (try == MUTEX_TRY)
1761 			error = mutex_trylock_kernel(mp);
1762 		else	/* MUTEX_LOCK */
1763 			error = mutex_lock_kernel(mp, tsp, msp);
1764 		/*
1765 		 * The kernel never sets or clears the lock byte
1766 		 * for LOCK_PRIO_INHERIT mutexes.
1767 		 * Set it here for consistency.
1768 		 */
1769 		switch (error) {
1770 		case 0:
1771 			mp->mutex_lockw = LOCKSET;
1772 			break;
1773 		case EOWNERDEAD:
1774 		case ELOCKUNMAPPED:
1775 			mp->mutex_lockw = LOCKSET;
1776 			/* FALLTHROUGH */
1777 		case ENOTRECOVERABLE:
1778 			ASSERT(mtype & LOCK_ROBUST);
1779 			break;
1780 		case EDEADLK:
1781 			if (try == MUTEX_LOCK)
1782 				stall();
1783 			error = EBUSY;
1784 			break;
1785 		}
1786 	} else if (mtype & USYNC_PROCESS) {
1787 		error = mutex_trylock_process(mp);
1788 		if (error == EBUSY && try == MUTEX_LOCK)
1789 			error = mutex_lock_kernel(mp, tsp, msp);
1790 	} else  {	/* USYNC_THREAD */
1791 		error = mutex_trylock_adaptive(mp);
1792 		if (error == EBUSY && try == MUTEX_LOCK)
1793 			error = mutex_lock_queue(self, msp, mp, tsp);
1794 	}
1795 
1796 	switch (error) {
1797 	case 0:
1798 	case EOWNERDEAD:
1799 	case ELOCKUNMAPPED:
1800 		if (mtype & LOCK_ROBUST)
1801 			remember_lock(mp);
1802 		if (msp)
1803 			record_begin_hold(msp);
1804 		break;
1805 	default:
1806 		if (mtype & LOCK_PRIO_PROTECT) {
1807 			(void) _ceil_mylist_del(mp);
1808 			if (myprio < ceil)
1809 				_ceil_prio_waive();
1810 		}
1811 		if (try == MUTEX_TRY) {
1812 			if (msp)
1813 				tdb_incr(msp->mutex_try_fail);
1814 			if (__td_event_report(self, TD_LOCK_TRY, udp)) {
1815 				self->ul_td_evbuf.eventnum = TD_LOCK_TRY;
1816 				tdb_event(TD_LOCK_TRY, udp);
1817 			}
1818 		}
1819 		break;
1820 	}
1821 
1822 	return (error);
1823 }
1824 
1825 int
1826 fast_process_lock(mutex_t *mp, timespec_t *tsp, int mtype, int try)
1827 {
1828 	ulwp_t *self = curthread;
1829 	uberdata_t *udp = self->ul_uberdata;
1830 
1831 	/*
1832 	 * We know that USYNC_PROCESS is set in mtype and that
1833 	 * zero, one, or both of the flags LOCK_RECURSIVE and
1834 	 * LOCK_ERRORCHECK are set, and that no other flags are set.
1835 	 */
1836 	ASSERT((mtype & ~(USYNC_PROCESS|LOCK_RECURSIVE|LOCK_ERRORCHECK)) == 0);
1837 	enter_critical(self);
1838 	if (set_lock_byte(&mp->mutex_lockw) == 0) {
1839 		mp->mutex_owner = (uintptr_t)self;
1840 		mp->mutex_ownerpid = udp->pid;
1841 		exit_critical(self);
1842 		DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
1843 		return (0);
1844 	}
1845 	exit_critical(self);
1846 
1847 	if ((mtype & (LOCK_RECURSIVE|LOCK_ERRORCHECK)) && shared_mutex_held(mp))
1848 		return (mutex_recursion(mp, mtype, try));
1849 
1850 	/* try a little harder */
1851 	if (mutex_trylock_process(mp) == 0)
1852 		return (0);
1853 
1854 	if (try == MUTEX_LOCK)
1855 		return (mutex_lock_kernel(mp, tsp, NULL));
1856 
1857 	if (__td_event_report(self, TD_LOCK_TRY, udp)) {
1858 		self->ul_td_evbuf.eventnum = TD_LOCK_TRY;
1859 		tdb_event(TD_LOCK_TRY, udp);
1860 	}
1861 	return (EBUSY);
1862 }
1863 
1864 static int
1865 mutex_lock_impl(mutex_t *mp, timespec_t *tsp)
1866 {
1867 	ulwp_t *self = curthread;
1868 	uberdata_t *udp = self->ul_uberdata;
1869 	uberflags_t *gflags;
1870 	int mtype;
1871 
1872 	/*
1873 	 * Optimize the case of USYNC_THREAD, including
1874 	 * the LOCK_RECURSIVE and LOCK_ERRORCHECK cases,
1875 	 * no error detection, no lock statistics,
1876 	 * and the process has only a single thread.
1877 	 * (Most likely a traditional single-threaded application.)
1878 	 */
1879 	if ((((mtype = mp->mutex_type) & ~(LOCK_RECURSIVE|LOCK_ERRORCHECK)) |
1880 	    udp->uberflags.uf_all) == 0) {
1881 		/*
1882 		 * Only one thread exists so we don't need an atomic operation.
1883 		 */
1884 		if (mp->mutex_lockw == 0) {
1885 			mp->mutex_lockw = LOCKSET;
1886 			mp->mutex_owner = (uintptr_t)self;
1887 			DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
1888 			return (0);
1889 		}
1890 		if (mtype && MUTEX_OWNER(mp) == self)
1891 			return (mutex_recursion(mp, mtype, MUTEX_LOCK));
1892 		/*
1893 		 * We have reached a deadlock, probably because the
1894 		 * process is executing non-async-signal-safe code in
1895 		 * a signal handler and is attempting to acquire a lock
1896 		 * that it already owns.  This is not surprising, given
1897 		 * bad programming practices over the years that has
1898 		 * resulted in applications calling printf() and such
1899 		 * in their signal handlers.  Unless the user has told
1900 		 * us that the signal handlers are safe by setting:
1901 		 *	export _THREAD_ASYNC_SAFE=1
1902 		 * we return EDEADLK rather than actually deadlocking.
1903 		 */
1904 		if (tsp == NULL &&
1905 		    MUTEX_OWNER(mp) == self && !self->ul_async_safe) {
1906 			DTRACE_PROBE2(plockstat, mutex__error, mp, EDEADLK);
1907 			return (EDEADLK);
1908 		}
1909 	}
1910 
1911 	/*
1912 	 * Optimize the common cases of USYNC_THREAD or USYNC_PROCESS,
1913 	 * no error detection, and no lock statistics.
1914 	 * Include LOCK_RECURSIVE and LOCK_ERRORCHECK cases.
1915 	 */
1916 	if ((gflags = self->ul_schedctl_called) != NULL &&
1917 	    (gflags->uf_trs_ted |
1918 	    (mtype & ~(USYNC_PROCESS|LOCK_RECURSIVE|LOCK_ERRORCHECK))) == 0) {
1919 		if (mtype & USYNC_PROCESS)
1920 			return (fast_process_lock(mp, tsp, mtype, MUTEX_LOCK));
1921 		if (set_lock_byte(&mp->mutex_lockw) == 0) {
1922 			mp->mutex_owner = (uintptr_t)self;
1923 			DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
1924 			return (0);
1925 		}
1926 		if (mtype && MUTEX_OWNER(mp) == self)
1927 			return (mutex_recursion(mp, mtype, MUTEX_LOCK));
1928 		if (mutex_trylock_adaptive(mp) != 0)
1929 			return (mutex_lock_queue(self, NULL, mp, tsp));
1930 		return (0);
1931 	}
1932 
1933 	/* else do it the long way */
1934 	return (mutex_lock_internal(mp, tsp, MUTEX_LOCK));
1935 }
1936 
1937 #pragma weak _private_mutex_lock = __mutex_lock
1938 #pragma weak mutex_lock = __mutex_lock
1939 #pragma weak _mutex_lock = __mutex_lock
1940 #pragma weak pthread_mutex_lock = __mutex_lock
1941 #pragma weak _pthread_mutex_lock = __mutex_lock
1942 int
1943 __mutex_lock(mutex_t *mp)
1944 {
1945 	ASSERT(!curthread->ul_critical || curthread->ul_bindflags);
1946 	return (mutex_lock_impl(mp, NULL));
1947 }
1948 
1949 #pragma weak pthread_mutex_timedlock = _pthread_mutex_timedlock
1950 int
1951 _pthread_mutex_timedlock(mutex_t *mp, const timespec_t *abstime)
1952 {
1953 	timespec_t tslocal;
1954 	int error;
1955 
1956 	ASSERT(!curthread->ul_critical || curthread->ul_bindflags);
1957 	abstime_to_reltime(CLOCK_REALTIME, abstime, &tslocal);
1958 	error = mutex_lock_impl(mp, &tslocal);
1959 	if (error == ETIME)
1960 		error = ETIMEDOUT;
1961 	return (error);
1962 }
1963 
1964 #pragma weak pthread_mutex_reltimedlock_np = _pthread_mutex_reltimedlock_np
1965 int
1966 _pthread_mutex_reltimedlock_np(mutex_t *mp, const timespec_t *reltime)
1967 {
1968 	timespec_t tslocal;
1969 	int error;
1970 
1971 	ASSERT(!curthread->ul_critical || curthread->ul_bindflags);
1972 	tslocal = *reltime;
1973 	error = mutex_lock_impl(mp, &tslocal);
1974 	if (error == ETIME)
1975 		error = ETIMEDOUT;
1976 	return (error);
1977 }
1978 
1979 #pragma weak _private_mutex_trylock = __mutex_trylock
1980 #pragma weak mutex_trylock = __mutex_trylock
1981 #pragma weak _mutex_trylock = __mutex_trylock
1982 #pragma weak pthread_mutex_trylock = __mutex_trylock
1983 #pragma weak _pthread_mutex_trylock = __mutex_trylock
1984 int
1985 __mutex_trylock(mutex_t *mp)
1986 {
1987 	ulwp_t *self = curthread;
1988 	uberdata_t *udp = self->ul_uberdata;
1989 	uberflags_t *gflags;
1990 	int mtype;
1991 
1992 	ASSERT(!curthread->ul_critical || curthread->ul_bindflags);
1993 	/*
1994 	 * Optimize the case of USYNC_THREAD, including
1995 	 * the LOCK_RECURSIVE and LOCK_ERRORCHECK cases,
1996 	 * no error detection, no lock statistics,
1997 	 * and the process has only a single thread.
1998 	 * (Most likely a traditional single-threaded application.)
1999 	 */
2000 	if ((((mtype = mp->mutex_type) & ~(LOCK_RECURSIVE|LOCK_ERRORCHECK)) |
2001 	    udp->uberflags.uf_all) == 0) {
2002 		/*
2003 		 * Only one thread exists so we don't need an atomic operation.
2004 		 */
2005 		if (mp->mutex_lockw == 0) {
2006 			mp->mutex_lockw = LOCKSET;
2007 			mp->mutex_owner = (uintptr_t)self;
2008 			DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
2009 			return (0);
2010 		}
2011 		if (mtype && MUTEX_OWNER(mp) == self)
2012 			return (mutex_recursion(mp, mtype, MUTEX_TRY));
2013 		return (EBUSY);
2014 	}
2015 
2016 	/*
2017 	 * Optimize the common cases of USYNC_THREAD or USYNC_PROCESS,
2018 	 * no error detection, and no lock statistics.
2019 	 * Include LOCK_RECURSIVE and LOCK_ERRORCHECK cases.
2020 	 */
2021 	if ((gflags = self->ul_schedctl_called) != NULL &&
2022 	    (gflags->uf_trs_ted |
2023 	    (mtype & ~(USYNC_PROCESS|LOCK_RECURSIVE|LOCK_ERRORCHECK))) == 0) {
2024 		if (mtype & USYNC_PROCESS)
2025 			return (fast_process_lock(mp, NULL, mtype, MUTEX_TRY));
2026 		if (set_lock_byte(&mp->mutex_lockw) == 0) {
2027 			mp->mutex_owner = (uintptr_t)self;
2028 			DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
2029 			return (0);
2030 		}
2031 		if (mtype && MUTEX_OWNER(mp) == self)
2032 			return (mutex_recursion(mp, mtype, MUTEX_TRY));
2033 		if (mutex_trylock_adaptive(mp) != 0) {
2034 			if (__td_event_report(self, TD_LOCK_TRY, udp)) {
2035 				self->ul_td_evbuf.eventnum = TD_LOCK_TRY;
2036 				tdb_event(TD_LOCK_TRY, udp);
2037 			}
2038 			return (EBUSY);
2039 		}
2040 		return (0);
2041 	}
2042 
2043 	/* else do it the long way */
2044 	return (mutex_lock_internal(mp, NULL, MUTEX_TRY));
2045 }
2046 
2047 int
2048 mutex_unlock_internal(mutex_t *mp, int retain_robust_flags)
2049 {
2050 	ulwp_t *self = curthread;
2051 	uberdata_t *udp = self->ul_uberdata;
2052 	int mtype = mp->mutex_type;
2053 	tdb_mutex_stats_t *msp;
2054 	int error = 0;
2055 	int release_all;
2056 	lwpid_t lwpid;
2057 
2058 	if ((mtype & LOCK_ERRORCHECK) && !mutex_is_held(mp))
2059 		return (EPERM);
2060 
2061 	if (self->ul_error_detection && !mutex_is_held(mp))
2062 		lock_error(mp, "mutex_unlock", NULL, NULL);
2063 
2064 	if ((mtype & LOCK_RECURSIVE) && mp->mutex_rcount != 0) {
2065 		mp->mutex_rcount--;
2066 		DTRACE_PROBE2(plockstat, mutex__release, mp, 1);
2067 		return (0);
2068 	}
2069 
2070 	if ((msp = MUTEX_STATS(mp, udp)) != NULL)
2071 		(void) record_hold_time(msp);
2072 
2073 	if (!retain_robust_flags && !(mtype & LOCK_PRIO_INHERIT) &&
2074 	    (mp->mutex_flag & (LOCK_OWNERDEAD | LOCK_UNMAPPED))) {
2075 		ASSERT(mp->mutex_type & LOCK_ROBUST);
2076 		mp->mutex_flag &= ~(LOCK_OWNERDEAD | LOCK_UNMAPPED);
2077 		mp->mutex_flag |= LOCK_NOTRECOVERABLE;
2078 	}
2079 	release_all = ((mp->mutex_flag & LOCK_NOTRECOVERABLE) != 0);
2080 
2081 	if (mtype & LOCK_PRIO_INHERIT) {
2082 		no_preempt(self);
2083 		mp->mutex_owner = 0;
2084 		mp->mutex_ownerpid = 0;
2085 		DTRACE_PROBE2(plockstat, mutex__release, mp, 0);
2086 		mp->mutex_lockw = LOCKCLEAR;
2087 		error = ___lwp_mutex_unlock(mp);
2088 		preempt(self);
2089 	} else if (mtype & USYNC_PROCESS) {
2090 		if (mp->mutex_lockword & WAITERMASK) {
2091 			mutex_unlock_process(mp, release_all);
2092 		} else {
2093 			mp->mutex_owner = 0;
2094 			mp->mutex_ownerpid = 0;
2095 			DTRACE_PROBE2(plockstat, mutex__release, mp, 0);
2096 			if (atomic_swap_32(&mp->mutex_lockword, 0) &
2097 			    WAITERMASK) {  /* a waiter suddenly appeared */
2098 				no_preempt(self);
2099 				(void) ___lwp_mutex_wakeup(mp, release_all);
2100 				preempt(self);
2101 			}
2102 		}
2103 	} else {	/* USYNC_THREAD */
2104 		if ((lwpid = mutex_unlock_queue(mp, release_all)) != 0) {
2105 			(void) __lwp_unpark(lwpid);
2106 			preempt(self);
2107 		}
2108 	}
2109 
2110 	if (mtype & LOCK_ROBUST)
2111 		forget_lock(mp);
2112 
2113 	if ((mtype & LOCK_PRIO_PROTECT) && _ceil_mylist_del(mp))
2114 		_ceil_prio_waive();
2115 
2116 	return (error);
2117 }
2118 
2119 #pragma weak _private_mutex_unlock = __mutex_unlock
2120 #pragma weak mutex_unlock = __mutex_unlock
2121 #pragma weak _mutex_unlock = __mutex_unlock
2122 #pragma weak pthread_mutex_unlock = __mutex_unlock
2123 #pragma weak _pthread_mutex_unlock = __mutex_unlock
2124 int
2125 __mutex_unlock(mutex_t *mp)
2126 {
2127 	ulwp_t *self = curthread;
2128 	uberdata_t *udp = self->ul_uberdata;
2129 	uberflags_t *gflags;
2130 	lwpid_t lwpid;
2131 	int mtype;
2132 	short el;
2133 
2134 	/*
2135 	 * Optimize the case of USYNC_THREAD, including
2136 	 * the LOCK_RECURSIVE and LOCK_ERRORCHECK cases,
2137 	 * no error detection, no lock statistics,
2138 	 * and the process has only a single thread.
2139 	 * (Most likely a traditional single-threaded application.)
2140 	 */
2141 	if ((((mtype = mp->mutex_type) & ~(LOCK_RECURSIVE|LOCK_ERRORCHECK)) |
2142 	    udp->uberflags.uf_all) == 0) {
2143 		if (mtype) {
2144 			/*
2145 			 * At this point we know that one or both of the
2146 			 * flags LOCK_RECURSIVE or LOCK_ERRORCHECK is set.
2147 			 */
2148 			if ((mtype & LOCK_ERRORCHECK) && !MUTEX_OWNED(mp, self))
2149 				return (EPERM);
2150 			if ((mtype & LOCK_RECURSIVE) && mp->mutex_rcount != 0) {
2151 				mp->mutex_rcount--;
2152 				DTRACE_PROBE2(plockstat, mutex__release, mp, 1);
2153 				return (0);
2154 			}
2155 		}
2156 		/*
2157 		 * Only one thread exists so we don't need an atomic operation.
2158 		 * Also, there can be no waiters.
2159 		 */
2160 		mp->mutex_owner = 0;
2161 		mp->mutex_lockword = 0;
2162 		DTRACE_PROBE2(plockstat, mutex__release, mp, 0);
2163 		return (0);
2164 	}
2165 
2166 	/*
2167 	 * Optimize the common cases of USYNC_THREAD or USYNC_PROCESS,
2168 	 * no error detection, and no lock statistics.
2169 	 * Include LOCK_RECURSIVE and LOCK_ERRORCHECK cases.
2170 	 */
2171 	if ((gflags = self->ul_schedctl_called) != NULL) {
2172 		if (((el = gflags->uf_trs_ted) | mtype) == 0) {
2173 fast_unlock:
2174 			if (!(mp->mutex_lockword & WAITERMASK)) {
2175 				/* no waiter exists right now */
2176 				mp->mutex_owner = 0;
2177 				DTRACE_PROBE2(plockstat, mutex__release, mp, 0);
2178 				if (atomic_swap_32(&mp->mutex_lockword, 0) &
2179 				    WAITERMASK) {
2180 					/* a waiter suddenly appeared */
2181 					no_preempt(self);
2182 					if ((lwpid = mutex_wakeup(mp)) != 0)
2183 						(void) __lwp_unpark(lwpid);
2184 					preempt(self);
2185 				}
2186 			} else if ((lwpid = mutex_unlock_queue(mp, 0)) != 0) {
2187 				(void) __lwp_unpark(lwpid);
2188 				preempt(self);
2189 			}
2190 			return (0);
2191 		}
2192 		if (el)		/* error detection or lock statistics */
2193 			goto slow_unlock;
2194 		if ((mtype & ~(LOCK_RECURSIVE|LOCK_ERRORCHECK)) == 0) {
2195 			/*
2196 			 * At this point we know that one or both of the
2197 			 * flags LOCK_RECURSIVE or LOCK_ERRORCHECK is set.
2198 			 */
2199 			if ((mtype & LOCK_ERRORCHECK) && !MUTEX_OWNED(mp, self))
2200 				return (EPERM);
2201 			if ((mtype & LOCK_RECURSIVE) && mp->mutex_rcount != 0) {
2202 				mp->mutex_rcount--;
2203 				DTRACE_PROBE2(plockstat, mutex__release, mp, 1);
2204 				return (0);
2205 			}
2206 			goto fast_unlock;
2207 		}
2208 		if ((mtype &
2209 		    ~(USYNC_PROCESS|LOCK_RECURSIVE|LOCK_ERRORCHECK)) == 0) {
2210 			/*
2211 			 * At this point we know that zero, one, or both of the
2212 			 * flags LOCK_RECURSIVE or LOCK_ERRORCHECK is set and
2213 			 * that the USYNC_PROCESS flag is set.
2214 			 */
2215 			if ((mtype & LOCK_ERRORCHECK) && !shared_mutex_held(mp))
2216 				return (EPERM);
2217 			if ((mtype & LOCK_RECURSIVE) && mp->mutex_rcount != 0) {
2218 				mp->mutex_rcount--;
2219 				DTRACE_PROBE2(plockstat, mutex__release, mp, 1);
2220 				return (0);
2221 			}
2222 			if (mp->mutex_lockword & WAITERMASK) {
2223 				mutex_unlock_process(mp, 0);
2224 			} else {
2225 				mp->mutex_owner = 0;
2226 				mp->mutex_ownerpid = 0;
2227 				DTRACE_PROBE2(plockstat, mutex__release, mp, 0);
2228 				if (atomic_swap_32(&mp->mutex_lockword, 0) &
2229 				    WAITERMASK) {
2230 					no_preempt(self);
2231 					(void) ___lwp_mutex_wakeup(mp, 0);
2232 					preempt(self);
2233 				}
2234 			}
2235 			return (0);
2236 		}
2237 	}
2238 
2239 	/* else do it the long way */
2240 slow_unlock:
2241 	return (mutex_unlock_internal(mp, 0));
2242 }
2243 
2244 /*
2245  * Internally to the library, almost all mutex lock/unlock actions
2246  * go through these lmutex_ functions, to protect critical regions.
2247  * We replicate a bit of code from __mutex_lock() and __mutex_unlock()
2248  * to make these functions faster since we know that the mutex type
2249  * of all internal locks is USYNC_THREAD.  We also know that internal
2250  * locking can never fail, so we panic if it does.
2251  */
2252 void
2253 lmutex_lock(mutex_t *mp)
2254 {
2255 	ulwp_t *self = curthread;
2256 	uberdata_t *udp = self->ul_uberdata;
2257 
2258 	ASSERT(mp->mutex_type == USYNC_THREAD);
2259 
2260 	enter_critical(self);
2261 	/*
2262 	 * Optimize the case of no lock statistics and only a single thread.
2263 	 * (Most likely a traditional single-threaded application.)
2264 	 */
2265 	if (udp->uberflags.uf_all == 0) {
2266 		/*
2267 		 * Only one thread exists; the mutex must be free.
2268 		 */
2269 		ASSERT(mp->mutex_lockw == 0);
2270 		mp->mutex_lockw = LOCKSET;
2271 		mp->mutex_owner = (uintptr_t)self;
2272 		DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
2273 	} else {
2274 		tdb_mutex_stats_t *msp = MUTEX_STATS(mp, udp);
2275 
2276 		if (!self->ul_schedctl_called)
2277 			(void) setup_schedctl();
2278 
2279 		if (set_lock_byte(&mp->mutex_lockw) == 0) {
2280 			mp->mutex_owner = (uintptr_t)self;
2281 			DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
2282 		} else if (mutex_trylock_adaptive(mp) != 0) {
2283 			(void) mutex_lock_queue(self, msp, mp, NULL);
2284 		}
2285 
2286 		if (msp)
2287 			record_begin_hold(msp);
2288 	}
2289 }
2290 
2291 void
2292 lmutex_unlock(mutex_t *mp)
2293 {
2294 	ulwp_t *self = curthread;
2295 	uberdata_t *udp = self->ul_uberdata;
2296 
2297 	ASSERT(mp->mutex_type == USYNC_THREAD);
2298 
2299 	/*
2300 	 * Optimize the case of no lock statistics and only a single thread.
2301 	 * (Most likely a traditional single-threaded application.)
2302 	 */
2303 	if (udp->uberflags.uf_all == 0) {
2304 		/*
2305 		 * Only one thread exists so there can be no waiters.
2306 		 */
2307 		mp->mutex_owner = 0;
2308 		mp->mutex_lockword = 0;
2309 		DTRACE_PROBE2(plockstat, mutex__release, mp, 0);
2310 	} else {
2311 		tdb_mutex_stats_t *msp = MUTEX_STATS(mp, udp);
2312 		lwpid_t lwpid;
2313 
2314 		if (msp)
2315 			(void) record_hold_time(msp);
2316 		if ((lwpid = mutex_unlock_queue(mp, 0)) != 0) {
2317 			(void) __lwp_unpark(lwpid);
2318 			preempt(self);
2319 		}
2320 	}
2321 	exit_critical(self);
2322 }
2323 
2324 /*
2325  * For specialized code in libc, like the asynchronous i/o code,
2326  * the following sig_*() locking primitives are used in order
2327  * to make the code asynchronous signal safe.  Signals are
2328  * deferred while locks acquired by these functions are held.
2329  */
2330 void
2331 sig_mutex_lock(mutex_t *mp)
2332 {
2333 	sigoff(curthread);
2334 	(void) _private_mutex_lock(mp);
2335 }
2336 
2337 void
2338 sig_mutex_unlock(mutex_t *mp)
2339 {
2340 	(void) _private_mutex_unlock(mp);
2341 	sigon(curthread);
2342 }
2343 
2344 int
2345 sig_mutex_trylock(mutex_t *mp)
2346 {
2347 	int error;
2348 
2349 	sigoff(curthread);
2350 	if ((error = _private_mutex_trylock(mp)) != 0)
2351 		sigon(curthread);
2352 	return (error);
2353 }
2354 
2355 /*
2356  * sig_cond_wait() is a cancellation point.
2357  */
2358 int
2359 sig_cond_wait(cond_t *cv, mutex_t *mp)
2360 {
2361 	int error;
2362 
2363 	ASSERT(curthread->ul_sigdefer != 0);
2364 	_private_testcancel();
2365 	error = _cond_wait(cv, mp);
2366 	if (error == EINTR && curthread->ul_cursig) {
2367 		sig_mutex_unlock(mp);
2368 		/* take the deferred signal here */
2369 		sig_mutex_lock(mp);
2370 	}
2371 	_private_testcancel();
2372 	return (error);
2373 }
2374 
2375 /*
2376  * sig_cond_reltimedwait() is a cancellation point.
2377  */
2378 int
2379 sig_cond_reltimedwait(cond_t *cv, mutex_t *mp, const timespec_t *ts)
2380 {
2381 	int error;
2382 
2383 	ASSERT(curthread->ul_sigdefer != 0);
2384 	_private_testcancel();
2385 	error = _cond_reltimedwait(cv, mp, ts);
2386 	if (error == EINTR && curthread->ul_cursig) {
2387 		sig_mutex_unlock(mp);
2388 		/* take the deferred signal here */
2389 		sig_mutex_lock(mp);
2390 	}
2391 	_private_testcancel();
2392 	return (error);
2393 }
2394 
2395 static int
2396 shared_mutex_held(mutex_t *mparg)
2397 {
2398 	/*
2399 	 * The 'volatile' is necessary to make sure the compiler doesn't
2400 	 * reorder the tests of the various components of the mutex.
2401 	 * They must be tested in this order:
2402 	 *	mutex_lockw
2403 	 *	mutex_owner
2404 	 *	mutex_ownerpid
2405 	 * This relies on the fact that everywhere mutex_lockw is cleared,
2406 	 * mutex_owner and mutex_ownerpid are cleared before mutex_lockw
2407 	 * is cleared, and that everywhere mutex_lockw is set, mutex_owner
2408 	 * and mutex_ownerpid are set after mutex_lockw is set, and that
2409 	 * mutex_lockw is set or cleared with a memory barrier.
2410 	 */
2411 	volatile mutex_t *mp = (volatile mutex_t *)mparg;
2412 	ulwp_t *self = curthread;
2413 	uberdata_t *udp = self->ul_uberdata;
2414 
2415 	return (MUTEX_OWNED(mp, self) && mp->mutex_ownerpid == udp->pid);
2416 }
2417 
2418 /*
2419  * Some crufty old programs define their own version of _mutex_held()
2420  * to be simply return(1).  This breaks internal libc logic, so we
2421  * define a private version for exclusive use by libc, mutex_is_held(),
2422  * and also a new public function, __mutex_held(), to be used in new
2423  * code to circumvent these crufty old programs.
2424  */
2425 #pragma weak mutex_held = mutex_is_held
2426 #pragma weak _mutex_held = mutex_is_held
2427 #pragma weak __mutex_held = mutex_is_held
2428 int
2429 mutex_is_held(mutex_t *mparg)
2430 {
2431 	volatile mutex_t *mp = (volatile mutex_t *)mparg;
2432 
2433 	if (mparg->mutex_type & USYNC_PROCESS)
2434 		return (shared_mutex_held(mparg));
2435 	return (MUTEX_OWNED(mp, curthread));
2436 }
2437 
2438 #pragma weak _private_mutex_destroy = __mutex_destroy
2439 #pragma weak mutex_destroy = __mutex_destroy
2440 #pragma weak _mutex_destroy = __mutex_destroy
2441 #pragma weak pthread_mutex_destroy = __mutex_destroy
2442 #pragma weak _pthread_mutex_destroy = __mutex_destroy
2443 int
2444 __mutex_destroy(mutex_t *mp)
2445 {
2446 	if (mp->mutex_type & USYNC_PROCESS)
2447 		forget_lock(mp);
2448 	(void) _memset(mp, 0, sizeof (*mp));
2449 	tdb_sync_obj_deregister(mp);
2450 	return (0);
2451 }
2452 
2453 #pragma weak mutex_consistent = __mutex_consistent
2454 #pragma weak _mutex_consistent = __mutex_consistent
2455 #pragma weak pthread_mutex_consistent_np = __mutex_consistent
2456 #pragma weak _pthread_mutex_consistent_np = __mutex_consistent
2457 int
2458 __mutex_consistent(mutex_t *mp)
2459 {
2460 	/*
2461 	 * Do this only for an inconsistent, initialized robust lock
2462 	 * that we hold.  For all other cases, return EINVAL.
2463 	 */
2464 	if (mutex_is_held(mp) &&
2465 	    (mp->mutex_type & LOCK_ROBUST) &&
2466 	    (mp->mutex_flag & LOCK_INITED) &&
2467 	    (mp->mutex_flag & (LOCK_OWNERDEAD | LOCK_UNMAPPED))) {
2468 		mp->mutex_flag &= ~(LOCK_OWNERDEAD | LOCK_UNMAPPED);
2469 		mp->mutex_rcount = 0;
2470 		return (0);
2471 	}
2472 	return (EINVAL);
2473 }
2474 
2475 /*
2476  * Spin locks are separate from ordinary mutexes,
2477  * but we use the same data structure for them.
2478  */
2479 
2480 #pragma weak pthread_spin_init = _pthread_spin_init
2481 int
2482 _pthread_spin_init(pthread_spinlock_t *lock, int pshared)
2483 {
2484 	mutex_t *mp = (mutex_t *)lock;
2485 
2486 	(void) _memset(mp, 0, sizeof (*mp));
2487 	if (pshared == PTHREAD_PROCESS_SHARED)
2488 		mp->mutex_type = USYNC_PROCESS;
2489 	else
2490 		mp->mutex_type = USYNC_THREAD;
2491 	mp->mutex_flag = LOCK_INITED;
2492 	mp->mutex_magic = MUTEX_MAGIC;
2493 	return (0);
2494 }
2495 
2496 #pragma weak pthread_spin_destroy = _pthread_spin_destroy
2497 int
2498 _pthread_spin_destroy(pthread_spinlock_t *lock)
2499 {
2500 	(void) _memset(lock, 0, sizeof (*lock));
2501 	return (0);
2502 }
2503 
2504 #pragma weak pthread_spin_trylock = _pthread_spin_trylock
2505 int
2506 _pthread_spin_trylock(pthread_spinlock_t *lock)
2507 {
2508 	mutex_t *mp = (mutex_t *)lock;
2509 	ulwp_t *self = curthread;
2510 	int error = 0;
2511 
2512 	no_preempt(self);
2513 	if (set_lock_byte(&mp->mutex_lockw) != 0)
2514 		error = EBUSY;
2515 	else {
2516 		mp->mutex_owner = (uintptr_t)self;
2517 		if (mp->mutex_type == USYNC_PROCESS)
2518 			mp->mutex_ownerpid = self->ul_uberdata->pid;
2519 		DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, 0);
2520 	}
2521 	preempt(self);
2522 	return (error);
2523 }
2524 
2525 #pragma weak pthread_spin_lock = _pthread_spin_lock
2526 int
2527 _pthread_spin_lock(pthread_spinlock_t *lock)
2528 {
2529 	mutex_t *mp = (mutex_t *)lock;
2530 	ulwp_t *self = curthread;
2531 	volatile uint8_t *lockp = (volatile uint8_t *)&mp->mutex_lockw;
2532 	int count = 0;
2533 
2534 	ASSERT(!self->ul_critical || self->ul_bindflags);
2535 
2536 	DTRACE_PROBE1(plockstat, mutex__spin, mp);
2537 
2538 	/*
2539 	 * We don't care whether the owner is running on a processor.
2540 	 * We just spin because that's what this interface requires.
2541 	 */
2542 	for (;;) {
2543 		if (count < INT_MAX)
2544 			count++;
2545 		if (*lockp == 0) {	/* lock byte appears to be clear */
2546 			no_preempt(self);
2547 			if (set_lock_byte(lockp) == 0)
2548 				break;
2549 			preempt(self);
2550 		}
2551 		SMT_PAUSE();
2552 	}
2553 	mp->mutex_owner = (uintptr_t)self;
2554 	if (mp->mutex_type == USYNC_PROCESS)
2555 		mp->mutex_ownerpid = self->ul_uberdata->pid;
2556 	preempt(self);
2557 	DTRACE_PROBE2(plockstat, mutex__spun, 1, count);
2558 	DTRACE_PROBE3(plockstat, mutex__acquire, mp, 0, count);
2559 	return (0);
2560 }
2561 
2562 #pragma weak pthread_spin_unlock = _pthread_spin_unlock
2563 int
2564 _pthread_spin_unlock(pthread_spinlock_t *lock)
2565 {
2566 	mutex_t *mp = (mutex_t *)lock;
2567 	ulwp_t *self = curthread;
2568 
2569 	no_preempt(self);
2570 	mp->mutex_owner = 0;
2571 	mp->mutex_ownerpid = 0;
2572 	DTRACE_PROBE2(plockstat, mutex__release, mp, 0);
2573 	(void) atomic_swap_32(&mp->mutex_lockword, 0);
2574 	preempt(self);
2575 	return (0);
2576 }
2577 
2578 #define	INITIAL_LOCKS	8	/* initialial size of ul_heldlocks.array */
2579 
2580 /*
2581  * Find/allocate an entry for 'lock' in our array of held locks.
2582  */
2583 static mutex_t **
2584 find_lock_entry(mutex_t *lock)
2585 {
2586 	ulwp_t *self = curthread;
2587 	mutex_t **remembered = NULL;
2588 	mutex_t **lockptr;
2589 	uint_t nlocks;
2590 
2591 	if ((nlocks = self->ul_heldlockcnt) != 0)
2592 		lockptr = self->ul_heldlocks.array;
2593 	else {
2594 		nlocks = 1;
2595 		lockptr = &self->ul_heldlocks.single;
2596 	}
2597 
2598 	for (; nlocks; nlocks--, lockptr++) {
2599 		if (*lockptr == lock)
2600 			return (lockptr);
2601 		if (*lockptr == NULL && remembered == NULL)
2602 			remembered = lockptr;
2603 	}
2604 	if (remembered != NULL) {
2605 		*remembered = lock;
2606 		return (remembered);
2607 	}
2608 
2609 	/*
2610 	 * No entry available.  Allocate more space, converting
2611 	 * the single entry into an array of entries if necessary.
2612 	 */
2613 	if ((nlocks = self->ul_heldlockcnt) == 0) {
2614 		/*
2615 		 * Initial allocation of the array.
2616 		 * Convert the single entry into an array.
2617 		 */
2618 		self->ul_heldlockcnt = nlocks = INITIAL_LOCKS;
2619 		lockptr = lmalloc(nlocks * sizeof (mutex_t *));
2620 		/*
2621 		 * The single entry becomes the first entry in the array.
2622 		 */
2623 		*lockptr = self->ul_heldlocks.single;
2624 		self->ul_heldlocks.array = lockptr;
2625 		/*
2626 		 * Return the next available entry in the array.
2627 		 */
2628 		*++lockptr = lock;
2629 		return (lockptr);
2630 	}
2631 	/*
2632 	 * Reallocate the array, double the size each time.
2633 	 */
2634 	lockptr = lmalloc(nlocks * 2 * sizeof (mutex_t *));
2635 	(void) _memcpy(lockptr, self->ul_heldlocks.array,
2636 	    nlocks * sizeof (mutex_t *));
2637 	lfree(self->ul_heldlocks.array, nlocks * sizeof (mutex_t *));
2638 	self->ul_heldlocks.array = lockptr;
2639 	self->ul_heldlockcnt *= 2;
2640 	/*
2641 	 * Return the next available entry in the newly allocated array.
2642 	 */
2643 	*(lockptr += nlocks) = lock;
2644 	return (lockptr);
2645 }
2646 
2647 /*
2648  * Insert 'lock' into our list of held locks.
2649  * Currently only used for LOCK_ROBUST mutexes.
2650  */
2651 void
2652 remember_lock(mutex_t *lock)
2653 {
2654 	(void) find_lock_entry(lock);
2655 }
2656 
2657 /*
2658  * Remove 'lock' from our list of held locks.
2659  * Currently only used for LOCK_ROBUST mutexes.
2660  */
2661 void
2662 forget_lock(mutex_t *lock)
2663 {
2664 	*find_lock_entry(lock) = NULL;
2665 }
2666 
2667 /*
2668  * Free the array of held locks.
2669  */
2670 void
2671 heldlock_free(ulwp_t *ulwp)
2672 {
2673 	uint_t nlocks;
2674 
2675 	if ((nlocks = ulwp->ul_heldlockcnt) != 0)
2676 		lfree(ulwp->ul_heldlocks.array, nlocks * sizeof (mutex_t *));
2677 	ulwp->ul_heldlockcnt = 0;
2678 	ulwp->ul_heldlocks.array = NULL;
2679 }
2680 
2681 /*
2682  * Mark all held LOCK_ROBUST mutexes LOCK_OWNERDEAD.
2683  * Called from _thrp_exit() to deal with abandoned locks.
2684  */
2685 void
2686 heldlock_exit(void)
2687 {
2688 	ulwp_t *self = curthread;
2689 	mutex_t **lockptr;
2690 	uint_t nlocks;
2691 	mutex_t *mp;
2692 
2693 	if ((nlocks = self->ul_heldlockcnt) != 0)
2694 		lockptr = self->ul_heldlocks.array;
2695 	else {
2696 		nlocks = 1;
2697 		lockptr = &self->ul_heldlocks.single;
2698 	}
2699 
2700 	for (; nlocks; nlocks--, lockptr++) {
2701 		/*
2702 		 * The kernel takes care of transitioning held
2703 		 * LOCK_PRIO_INHERIT mutexes to LOCK_OWNERDEAD.
2704 		 * We avoid that case here.
2705 		 */
2706 		if ((mp = *lockptr) != NULL &&
2707 		    mutex_is_held(mp) &&
2708 		    (mp->mutex_type & (LOCK_ROBUST | LOCK_PRIO_INHERIT)) ==
2709 		    LOCK_ROBUST) {
2710 			mp->mutex_rcount = 0;
2711 			if (!(mp->mutex_flag & LOCK_UNMAPPED))
2712 				mp->mutex_flag |= LOCK_OWNERDEAD;
2713 			(void) mutex_unlock_internal(mp, 1);
2714 		}
2715 	}
2716 
2717 	heldlock_free(self);
2718 }
2719 
2720 #pragma weak cond_init = _cond_init
2721 /* ARGSUSED2 */
2722 int
2723 _cond_init(cond_t *cvp, int type, void *arg)
2724 {
2725 	if (type != USYNC_THREAD && type != USYNC_PROCESS)
2726 		return (EINVAL);
2727 	(void) _memset(cvp, 0, sizeof (*cvp));
2728 	cvp->cond_type = (uint16_t)type;
2729 	cvp->cond_magic = COND_MAGIC;
2730 	return (0);
2731 }
2732 
2733 /*
2734  * cond_sleep_queue(): utility function for cond_wait_queue().
2735  *
2736  * Go to sleep on a condvar sleep queue, expect to be waked up
2737  * by someone calling cond_signal() or cond_broadcast() or due
2738  * to receiving a UNIX signal or being cancelled, or just simply
2739  * due to a spurious wakeup (like someome calling forkall()).
2740  *
2741  * The associated mutex is *not* reacquired before returning.
2742  * That must be done by the caller of cond_sleep_queue().
2743  */
2744 static int
2745 cond_sleep_queue(cond_t *cvp, mutex_t *mp, timespec_t *tsp)
2746 {
2747 	ulwp_t *self = curthread;
2748 	queue_head_t *qp;
2749 	queue_head_t *mqp;
2750 	lwpid_t lwpid;
2751 	int signalled;
2752 	int error;
2753 	int release_all;
2754 
2755 	/*
2756 	 * Put ourself on the CV sleep queue, unlock the mutex, then
2757 	 * park ourself and unpark a candidate lwp to grab the mutex.
2758 	 * We must go onto the CV sleep queue before dropping the
2759 	 * mutex in order to guarantee atomicity of the operation.
2760 	 */
2761 	self->ul_sp = stkptr();
2762 	qp = queue_lock(cvp, CV);
2763 	enqueue(qp, self, cvp, CV);
2764 	cvp->cond_waiters_user = 1;
2765 	self->ul_cvmutex = mp;
2766 	self->ul_cv_wake = (tsp != NULL);
2767 	self->ul_signalled = 0;
2768 	if (mp->mutex_flag & LOCK_OWNERDEAD) {
2769 		mp->mutex_flag &= ~LOCK_OWNERDEAD;
2770 		mp->mutex_flag |= LOCK_NOTRECOVERABLE;
2771 	}
2772 	release_all = ((mp->mutex_flag & LOCK_NOTRECOVERABLE) != 0);
2773 	lwpid = mutex_unlock_queue(mp, release_all);
2774 	for (;;) {
2775 		set_parking_flag(self, 1);
2776 		queue_unlock(qp);
2777 		if (lwpid != 0) {
2778 			lwpid = preempt_unpark(self, lwpid);
2779 			preempt(self);
2780 		}
2781 		/*
2782 		 * We may have a deferred signal present,
2783 		 * in which case we should return EINTR.
2784 		 * Also, we may have received a SIGCANCEL; if so
2785 		 * and we are cancelable we should return EINTR.
2786 		 * We force an immediate EINTR return from
2787 		 * __lwp_park() by turning our parking flag off.
2788 		 */
2789 		if (self->ul_cursig != 0 ||
2790 		    (self->ul_cancelable && self->ul_cancel_pending))
2791 			set_parking_flag(self, 0);
2792 		/*
2793 		 * __lwp_park() will return the residual time in tsp
2794 		 * if we are unparked before the timeout expires.
2795 		 */
2796 		error = __lwp_park(tsp, lwpid);
2797 		set_parking_flag(self, 0);
2798 		lwpid = 0;	/* unpark the other lwp only once */
2799 		/*
2800 		 * We were waked up by cond_signal(), cond_broadcast(),
2801 		 * by an interrupt or timeout (EINTR or ETIME),
2802 		 * or we may just have gotten a spurious wakeup.
2803 		 */
2804 		qp = queue_lock(cvp, CV);
2805 		mqp = queue_lock(mp, MX);
2806 		if (self->ul_sleepq == NULL)
2807 			break;
2808 		/*
2809 		 * We are on either the condvar sleep queue or the
2810 		 * mutex sleep queue.  Break out of the sleep if we
2811 		 * were interrupted or we timed out (EINTR or ETIME).
2812 		 * Else this is a spurious wakeup; continue the loop.
2813 		 */
2814 		if (self->ul_sleepq == mqp) {		/* mutex queue */
2815 			if (error) {
2816 				mp->mutex_waiters = dequeue_self(mqp, mp);
2817 				break;
2818 			}
2819 			tsp = NULL;	/* no more timeout */
2820 		} else if (self->ul_sleepq == qp) {	/* condvar queue */
2821 			if (error) {
2822 				cvp->cond_waiters_user = dequeue_self(qp, cvp);
2823 				break;
2824 			}
2825 			/*
2826 			 * Else a spurious wakeup on the condvar queue.
2827 			 * __lwp_park() has already adjusted the timeout.
2828 			 */
2829 		} else {
2830 			thr_panic("cond_sleep_queue(): thread not on queue");
2831 		}
2832 		queue_unlock(mqp);
2833 	}
2834 
2835 	self->ul_sp = 0;
2836 	ASSERT(self->ul_cvmutex == NULL && self->ul_cv_wake == 0);
2837 	ASSERT(self->ul_sleepq == NULL && self->ul_link == NULL &&
2838 	    self->ul_wchan == NULL);
2839 
2840 	signalled = self->ul_signalled;
2841 	self->ul_signalled = 0;
2842 	queue_unlock(qp);
2843 	queue_unlock(mqp);
2844 
2845 	/*
2846 	 * If we were concurrently cond_signal()d and any of:
2847 	 * received a UNIX signal, were cancelled, or got a timeout,
2848 	 * then perform another cond_signal() to avoid consuming it.
2849 	 */
2850 	if (error && signalled)
2851 		(void) cond_signal_internal(cvp);
2852 
2853 	return (error);
2854 }
2855 
2856 int
2857 cond_wait_queue(cond_t *cvp, mutex_t *mp, timespec_t *tsp,
2858 	tdb_mutex_stats_t *msp)
2859 {
2860 	ulwp_t *self = curthread;
2861 	int error;
2862 	int merror;
2863 
2864 	/*
2865 	 * The old thread library was programmed to defer signals
2866 	 * while in cond_wait() so that the associated mutex would
2867 	 * be guaranteed to be held when the application signal
2868 	 * handler was invoked.
2869 	 *
2870 	 * We do not behave this way by default; the state of the
2871 	 * associated mutex in the signal handler is undefined.
2872 	 *
2873 	 * To accommodate applications that depend on the old
2874 	 * behavior, the _THREAD_COND_WAIT_DEFER environment
2875 	 * variable can be set to 1 and we will behave in the
2876 	 * old way with respect to cond_wait().
2877 	 */
2878 	if (self->ul_cond_wait_defer)
2879 		sigoff(self);
2880 
2881 	error = cond_sleep_queue(cvp, mp, tsp);
2882 
2883 	/*
2884 	 * Reacquire the mutex.
2885 	 */
2886 	if ((merror = mutex_trylock_adaptive(mp)) == EBUSY)
2887 		merror = mutex_lock_queue(self, msp, mp, NULL);
2888 	if (merror)
2889 		error = merror;
2890 	if (msp && (merror == 0 || merror == EOWNERDEAD))
2891 		record_begin_hold(msp);
2892 
2893 	/*
2894 	 * Take any deferred signal now, after we have reacquired the mutex.
2895 	 */
2896 	if (self->ul_cond_wait_defer)
2897 		sigon(self);
2898 
2899 	return (error);
2900 }
2901 
2902 /*
2903  * cond_sleep_kernel(): utility function for cond_wait_kernel().
2904  * See the comment ahead of cond_sleep_queue(), above.
2905  */
2906 static int
2907 cond_sleep_kernel(cond_t *cvp, mutex_t *mp, timespec_t *tsp)
2908 {
2909 	int mtype = mp->mutex_type;
2910 	ulwp_t *self = curthread;
2911 	int error;
2912 
2913 	if ((mtype & LOCK_PRIO_PROTECT) && _ceil_mylist_del(mp))
2914 		_ceil_prio_waive();
2915 
2916 	self->ul_sp = stkptr();
2917 	self->ul_wchan = cvp;
2918 	mp->mutex_owner = 0;
2919 	mp->mutex_ownerpid = 0;
2920 	if (mtype & LOCK_PRIO_INHERIT)
2921 		mp->mutex_lockw = LOCKCLEAR;
2922 	/*
2923 	 * ___lwp_cond_wait() returns immediately with EINTR if
2924 	 * set_parking_flag(self,0) is called on this lwp before it
2925 	 * goes to sleep in the kernel.  sigacthandler() calls this
2926 	 * when a deferred signal is noted.  This assures that we don't
2927 	 * get stuck in ___lwp_cond_wait() with all signals blocked
2928 	 * due to taking a deferred signal before going to sleep.
2929 	 */
2930 	set_parking_flag(self, 1);
2931 	if (self->ul_cursig != 0 ||
2932 	    (self->ul_cancelable && self->ul_cancel_pending))
2933 		set_parking_flag(self, 0);
2934 	error = ___lwp_cond_wait(cvp, mp, tsp, 1);
2935 	set_parking_flag(self, 0);
2936 	self->ul_sp = 0;
2937 	self->ul_wchan = NULL;
2938 	return (error);
2939 }
2940 
2941 int
2942 cond_wait_kernel(cond_t *cvp, mutex_t *mp, timespec_t *tsp)
2943 {
2944 	ulwp_t *self = curthread;
2945 	int error;
2946 	int merror;
2947 
2948 	/*
2949 	 * See the large comment in cond_wait_queue(), above.
2950 	 */
2951 	if (self->ul_cond_wait_defer)
2952 		sigoff(self);
2953 
2954 	error = cond_sleep_kernel(cvp, mp, tsp);
2955 
2956 	/*
2957 	 * Override the return code from ___lwp_cond_wait()
2958 	 * with any non-zero return code from mutex_lock().
2959 	 * This addresses robust lock failures in particular;
2960 	 * the caller must see the EOWNERDEAD or ENOTRECOVERABLE
2961 	 * errors in order to take corrective action.
2962 	 */
2963 	if ((merror = _private_mutex_lock(mp)) != 0)
2964 		error = merror;
2965 
2966 	/*
2967 	 * Take any deferred signal now, after we have reacquired the mutex.
2968 	 */
2969 	if (self->ul_cond_wait_defer)
2970 		sigon(self);
2971 
2972 	return (error);
2973 }
2974 
2975 /*
2976  * Common code for _cond_wait() and _cond_timedwait()
2977  */
2978 int
2979 cond_wait_common(cond_t *cvp, mutex_t *mp, timespec_t *tsp)
2980 {
2981 	int mtype = mp->mutex_type;
2982 	hrtime_t begin_sleep = 0;
2983 	ulwp_t *self = curthread;
2984 	uberdata_t *udp = self->ul_uberdata;
2985 	tdb_cond_stats_t *csp = COND_STATS(cvp, udp);
2986 	tdb_mutex_stats_t *msp = MUTEX_STATS(mp, udp);
2987 	uint8_t rcount;
2988 	int error = 0;
2989 
2990 	/*
2991 	 * The SUSV3 Posix spec for pthread_cond_timedwait() states:
2992 	 *	Except in the case of [ETIMEDOUT], all these error checks
2993 	 *	shall act as if they were performed immediately at the
2994 	 *	beginning of processing for the function and shall cause
2995 	 *	an error return, in effect, prior to modifying the state
2996 	 *	of the mutex specified by mutex or the condition variable
2997 	 *	specified by cond.
2998 	 * Therefore, we must return EINVAL now if the timout is invalid.
2999 	 */
3000 	if (tsp != NULL &&
3001 	    (tsp->tv_sec < 0 || (ulong_t)tsp->tv_nsec >= NANOSEC))
3002 		return (EINVAL);
3003 
3004 	if (__td_event_report(self, TD_SLEEP, udp)) {
3005 		self->ul_sp = stkptr();
3006 		self->ul_wchan = cvp;
3007 		self->ul_td_evbuf.eventnum = TD_SLEEP;
3008 		self->ul_td_evbuf.eventdata = cvp;
3009 		tdb_event(TD_SLEEP, udp);
3010 		self->ul_sp = 0;
3011 	}
3012 	if (csp) {
3013 		if (tsp)
3014 			tdb_incr(csp->cond_timedwait);
3015 		else
3016 			tdb_incr(csp->cond_wait);
3017 	}
3018 	if (msp)
3019 		begin_sleep = record_hold_time(msp);
3020 	else if (csp)
3021 		begin_sleep = gethrtime();
3022 
3023 	if (self->ul_error_detection) {
3024 		if (!mutex_is_held(mp))
3025 			lock_error(mp, "cond_wait", cvp, NULL);
3026 		if ((mtype & LOCK_RECURSIVE) && mp->mutex_rcount != 0)
3027 			lock_error(mp, "recursive mutex in cond_wait",
3028 				cvp, NULL);
3029 		if (cvp->cond_type & USYNC_PROCESS) {
3030 			if (!(mtype & USYNC_PROCESS))
3031 				lock_error(mp, "cond_wait", cvp,
3032 					"condvar process-shared, "
3033 					"mutex process-private");
3034 		} else {
3035 			if (mtype & USYNC_PROCESS)
3036 				lock_error(mp, "cond_wait", cvp,
3037 					"condvar process-private, "
3038 					"mutex process-shared");
3039 		}
3040 	}
3041 
3042 	/*
3043 	 * We deal with recursive mutexes by completely
3044 	 * dropping the lock and restoring the recursion
3045 	 * count after waking up.  This is arguably wrong,
3046 	 * but it obeys the principle of least astonishment.
3047 	 */
3048 	rcount = mp->mutex_rcount;
3049 	mp->mutex_rcount = 0;
3050 	if ((mtype &
3051 	    (USYNC_PROCESS | LOCK_PRIO_INHERIT | LOCK_PRIO_PROTECT)) |
3052 	    (cvp->cond_type & USYNC_PROCESS))
3053 		error = cond_wait_kernel(cvp, mp, tsp);
3054 	else
3055 		error = cond_wait_queue(cvp, mp, tsp, msp);
3056 	mp->mutex_rcount = rcount;
3057 
3058 	if (csp) {
3059 		hrtime_t lapse = gethrtime() - begin_sleep;
3060 		if (tsp == NULL)
3061 			csp->cond_wait_sleep_time += lapse;
3062 		else {
3063 			csp->cond_timedwait_sleep_time += lapse;
3064 			if (error == ETIME)
3065 				tdb_incr(csp->cond_timedwait_timeout);
3066 		}
3067 	}
3068 	return (error);
3069 }
3070 
3071 /*
3072  * cond_wait() is a cancellation point but _cond_wait() is not.
3073  * System libraries call the non-cancellation version.
3074  * It is expected that only applications call the cancellation version.
3075  */
3076 int
3077 _cond_wait(cond_t *cvp, mutex_t *mp)
3078 {
3079 	ulwp_t *self = curthread;
3080 	uberdata_t *udp = self->ul_uberdata;
3081 	uberflags_t *gflags;
3082 
3083 	/*
3084 	 * Optimize the common case of USYNC_THREAD plus
3085 	 * no error detection, no lock statistics, and no event tracing.
3086 	 */
3087 	if ((gflags = self->ul_schedctl_called) != NULL &&
3088 	    (cvp->cond_type | mp->mutex_type | gflags->uf_trs_ted |
3089 	    self->ul_td_events_enable |
3090 	    udp->tdb.tdb_ev_global_mask.event_bits[0]) == 0)
3091 		return (cond_wait_queue(cvp, mp, NULL, NULL));
3092 
3093 	/*
3094 	 * Else do it the long way.
3095 	 */
3096 	return (cond_wait_common(cvp, mp, NULL));
3097 }
3098 
3099 int
3100 cond_wait(cond_t *cvp, mutex_t *mp)
3101 {
3102 	int error;
3103 
3104 	_cancelon();
3105 	error = _cond_wait(cvp, mp);
3106 	if (error == EINTR)
3107 		_canceloff();
3108 	else
3109 		_canceloff_nocancel();
3110 	return (error);
3111 }
3112 
3113 #pragma weak pthread_cond_wait = _pthread_cond_wait
3114 int
3115 _pthread_cond_wait(cond_t *cvp, mutex_t *mp)
3116 {
3117 	int error;
3118 
3119 	error = cond_wait(cvp, mp);
3120 	return ((error == EINTR)? 0 : error);
3121 }
3122 
3123 /*
3124  * cond_timedwait() is a cancellation point but _cond_timedwait() is not.
3125  * System libraries call the non-cancellation version.
3126  * It is expected that only applications call the cancellation version.
3127  */
3128 int
3129 _cond_timedwait(cond_t *cvp, mutex_t *mp, const timespec_t *abstime)
3130 {
3131 	clockid_t clock_id = cvp->cond_clockid;
3132 	timespec_t reltime;
3133 	int error;
3134 
3135 	if (clock_id != CLOCK_REALTIME && clock_id != CLOCK_HIGHRES)
3136 		clock_id = CLOCK_REALTIME;
3137 	abstime_to_reltime(clock_id, abstime, &reltime);
3138 	error = cond_wait_common(cvp, mp, &reltime);
3139 	if (error == ETIME && clock_id == CLOCK_HIGHRES) {
3140 		/*
3141 		 * Don't return ETIME if we didn't really get a timeout.
3142 		 * This can happen if we return because someone resets
3143 		 * the system clock.  Just return zero in this case,
3144 		 * giving a spurious wakeup but not a timeout.
3145 		 */
3146 		if ((hrtime_t)(uint32_t)abstime->tv_sec * NANOSEC +
3147 		    abstime->tv_nsec > gethrtime())
3148 			error = 0;
3149 	}
3150 	return (error);
3151 }
3152 
3153 int
3154 cond_timedwait(cond_t *cvp, mutex_t *mp, const timespec_t *abstime)
3155 {
3156 	int error;
3157 
3158 	_cancelon();
3159 	error = _cond_timedwait(cvp, mp, abstime);
3160 	if (error == EINTR)
3161 		_canceloff();
3162 	else
3163 		_canceloff_nocancel();
3164 	return (error);
3165 }
3166 
3167 #pragma weak pthread_cond_timedwait = _pthread_cond_timedwait
3168 int
3169 _pthread_cond_timedwait(cond_t *cvp, mutex_t *mp, const timespec_t *abstime)
3170 {
3171 	int error;
3172 
3173 	error = cond_timedwait(cvp, mp, abstime);
3174 	if (error == ETIME)
3175 		error = ETIMEDOUT;
3176 	else if (error == EINTR)
3177 		error = 0;
3178 	return (error);
3179 }
3180 
3181 /*
3182  * cond_reltimedwait() is a cancellation point but _cond_reltimedwait()
3183  * is not.  System libraries call the non-cancellation version.
3184  * It is expected that only applications call the cancellation version.
3185  */
3186 int
3187 _cond_reltimedwait(cond_t *cvp, mutex_t *mp, const timespec_t *reltime)
3188 {
3189 	timespec_t tslocal = *reltime;
3190 
3191 	return (cond_wait_common(cvp, mp, &tslocal));
3192 }
3193 
3194 #pragma weak cond_reltimedwait = _cond_reltimedwait_cancel
3195 int
3196 _cond_reltimedwait_cancel(cond_t *cvp, mutex_t *mp, const timespec_t *reltime)
3197 {
3198 	int error;
3199 
3200 	_cancelon();
3201 	error = _cond_reltimedwait(cvp, mp, reltime);
3202 	if (error == EINTR)
3203 		_canceloff();
3204 	else
3205 		_canceloff_nocancel();
3206 	return (error);
3207 }
3208 
3209 #pragma weak pthread_cond_reltimedwait_np = _pthread_cond_reltimedwait_np
3210 int
3211 _pthread_cond_reltimedwait_np(cond_t *cvp, mutex_t *mp,
3212 	const timespec_t *reltime)
3213 {
3214 	int error;
3215 
3216 	error = _cond_reltimedwait_cancel(cvp, mp, reltime);
3217 	if (error == ETIME)
3218 		error = ETIMEDOUT;
3219 	else if (error == EINTR)
3220 		error = 0;
3221 	return (error);
3222 }
3223 
3224 #pragma weak pthread_cond_signal = cond_signal_internal
3225 #pragma weak _pthread_cond_signal = cond_signal_internal
3226 #pragma weak cond_signal = cond_signal_internal
3227 #pragma weak _cond_signal = cond_signal_internal
3228 int
3229 cond_signal_internal(cond_t *cvp)
3230 {
3231 	ulwp_t *self = curthread;
3232 	uberdata_t *udp = self->ul_uberdata;
3233 	tdb_cond_stats_t *csp = COND_STATS(cvp, udp);
3234 	int error = 0;
3235 	queue_head_t *qp;
3236 	mutex_t *mp;
3237 	queue_head_t *mqp;
3238 	ulwp_t **ulwpp;
3239 	ulwp_t *ulwp;
3240 	ulwp_t *prev = NULL;
3241 	ulwp_t *next;
3242 	ulwp_t **suspp = NULL;
3243 	ulwp_t *susprev;
3244 
3245 	if (csp)
3246 		tdb_incr(csp->cond_signal);
3247 
3248 	if (cvp->cond_waiters_kernel)	/* someone sleeping in the kernel? */
3249 		error = __lwp_cond_signal(cvp);
3250 
3251 	if (!cvp->cond_waiters_user)	/* no one sleeping at user-level */
3252 		return (error);
3253 
3254 	/*
3255 	 * Move someone from the condvar sleep queue to the mutex sleep
3256 	 * queue for the mutex that he will acquire on being waked up.
3257 	 * We can do this only if we own the mutex he will acquire.
3258 	 * If we do not own the mutex, or if his ul_cv_wake flag
3259 	 * is set, just dequeue and unpark him.
3260 	 */
3261 	qp = queue_lock(cvp, CV);
3262 	for (ulwpp = &qp->qh_head; (ulwp = *ulwpp) != NULL;
3263 	    prev = ulwp, ulwpp = &ulwp->ul_link) {
3264 		if (ulwp->ul_wchan == cvp) {
3265 			if (!ulwp->ul_stop)
3266 				break;
3267 			/*
3268 			 * Try not to dequeue a suspended thread.
3269 			 * This mimics the old libthread's behavior.
3270 			 */
3271 			if (suspp == NULL) {
3272 				suspp = ulwpp;
3273 				susprev = prev;
3274 			}
3275 		}
3276 	}
3277 	if (ulwp == NULL && suspp != NULL) {
3278 		ulwp = *(ulwpp = suspp);
3279 		prev = susprev;
3280 		suspp = NULL;
3281 	}
3282 	if (ulwp == NULL) {	/* no one on the sleep queue */
3283 		cvp->cond_waiters_user = 0;
3284 		queue_unlock(qp);
3285 		return (error);
3286 	}
3287 	/*
3288 	 * Scan the remainder of the CV queue for another waiter.
3289 	 */
3290 	if (suspp != NULL) {
3291 		next = *suspp;
3292 	} else {
3293 		for (next = ulwp->ul_link; next != NULL; next = next->ul_link)
3294 			if (next->ul_wchan == cvp)
3295 				break;
3296 	}
3297 	if (next == NULL)
3298 		cvp->cond_waiters_user = 0;
3299 
3300 	/*
3301 	 * Inform the thread that he was the recipient of a cond_signal().
3302 	 * This lets him deal with cond_signal() and, concurrently,
3303 	 * one or more of a cancellation, a UNIX signal, or a timeout.
3304 	 * These latter conditions must not consume a cond_signal().
3305 	 */
3306 	ulwp->ul_signalled = 1;
3307 
3308 	/*
3309 	 * Dequeue the waiter but leave his ul_sleepq non-NULL
3310 	 * while we move him to the mutex queue so that he can
3311 	 * deal properly with spurious wakeups.
3312 	 */
3313 	*ulwpp = ulwp->ul_link;
3314 	ulwp->ul_link = NULL;
3315 	if (qp->qh_tail == ulwp)
3316 		qp->qh_tail = prev;
3317 	qp->qh_qlen--;
3318 
3319 	mp = ulwp->ul_cvmutex;		/* the mutex he will acquire */
3320 	ulwp->ul_cvmutex = NULL;
3321 	ASSERT(mp != NULL);
3322 
3323 	if (ulwp->ul_cv_wake || !MUTEX_OWNED(mp, self)) {
3324 		lwpid_t lwpid = ulwp->ul_lwpid;
3325 
3326 		no_preempt(self);
3327 		ulwp->ul_sleepq = NULL;
3328 		ulwp->ul_wchan = NULL;
3329 		ulwp->ul_cv_wake = 0;
3330 		queue_unlock(qp);
3331 		(void) __lwp_unpark(lwpid);
3332 		preempt(self);
3333 	} else {
3334 		mqp = queue_lock(mp, MX);
3335 		enqueue(mqp, ulwp, mp, MX);
3336 		mp->mutex_waiters = 1;
3337 		queue_unlock(mqp);
3338 		queue_unlock(qp);
3339 	}
3340 
3341 	return (error);
3342 }
3343 
3344 /*
3345  * Utility function called by mutex_wakeup_all(), cond_broadcast(),
3346  * and rw_queue_release() to (re)allocate a big buffer to hold the
3347  * lwpids of all the threads to be set running after they are removed
3348  * from their sleep queues.  Since we are holding a queue lock, we
3349  * cannot call any function that might acquire a lock.  mmap(), munmap(),
3350  * lwp_unpark_all() are simple system calls and are safe in this regard.
3351  */
3352 lwpid_t *
3353 alloc_lwpids(lwpid_t *lwpid, int *nlwpid_ptr, int *maxlwps_ptr)
3354 {
3355 	/*
3356 	 * Allocate NEWLWPS ids on the first overflow.
3357 	 * Double the allocation each time after that.
3358 	 */
3359 	int nlwpid = *nlwpid_ptr;
3360 	int maxlwps = *maxlwps_ptr;
3361 	int first_allocation;
3362 	int newlwps;
3363 	void *vaddr;
3364 
3365 	ASSERT(nlwpid == maxlwps);
3366 
3367 	first_allocation = (maxlwps == MAXLWPS);
3368 	newlwps = first_allocation? NEWLWPS : 2 * maxlwps;
3369 	vaddr = _private_mmap(NULL, newlwps * sizeof (lwpid_t),
3370 	    PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, -1, (off_t)0);
3371 
3372 	if (vaddr == MAP_FAILED) {
3373 		/*
3374 		 * Let's hope this never happens.
3375 		 * If it does, then we have a terrible
3376 		 * thundering herd on our hands.
3377 		 */
3378 		(void) __lwp_unpark_all(lwpid, nlwpid);
3379 		*nlwpid_ptr = 0;
3380 	} else {
3381 		(void) _memcpy(vaddr, lwpid, maxlwps * sizeof (lwpid_t));
3382 		if (!first_allocation)
3383 			(void) _private_munmap(lwpid,
3384 			    maxlwps * sizeof (lwpid_t));
3385 		lwpid = vaddr;
3386 		*maxlwps_ptr = newlwps;
3387 	}
3388 
3389 	return (lwpid);
3390 }
3391 
3392 #pragma weak pthread_cond_broadcast = cond_broadcast_internal
3393 #pragma weak _pthread_cond_broadcast = cond_broadcast_internal
3394 #pragma weak cond_broadcast = cond_broadcast_internal
3395 #pragma weak _cond_broadcast = cond_broadcast_internal
3396 int
3397 cond_broadcast_internal(cond_t *cvp)
3398 {
3399 	ulwp_t *self = curthread;
3400 	uberdata_t *udp = self->ul_uberdata;
3401 	tdb_cond_stats_t *csp = COND_STATS(cvp, udp);
3402 	int error = 0;
3403 	queue_head_t *qp;
3404 	mutex_t *mp;
3405 	mutex_t *mp_cache = NULL;
3406 	queue_head_t *mqp = NULL;
3407 	ulwp_t **ulwpp;
3408 	ulwp_t *ulwp;
3409 	ulwp_t *prev = NULL;
3410 	int nlwpid = 0;
3411 	int maxlwps = MAXLWPS;
3412 	lwpid_t buffer[MAXLWPS];
3413 	lwpid_t *lwpid = buffer;
3414 
3415 	if (csp)
3416 		tdb_incr(csp->cond_broadcast);
3417 
3418 	if (cvp->cond_waiters_kernel)	/* someone sleeping in the kernel? */
3419 		error = __lwp_cond_broadcast(cvp);
3420 
3421 	if (!cvp->cond_waiters_user)	/* no one sleeping at user-level */
3422 		return (error);
3423 
3424 	/*
3425 	 * Move everyone from the condvar sleep queue to the mutex sleep
3426 	 * queue for the mutex that they will acquire on being waked up.
3427 	 * We can do this only if we own the mutex they will acquire.
3428 	 * If we do not own the mutex, or if their ul_cv_wake flag
3429 	 * is set, just dequeue and unpark them.
3430 	 *
3431 	 * We keep track of lwpids that are to be unparked in lwpid[].
3432 	 * __lwp_unpark_all() is called to unpark all of them after
3433 	 * they have been removed from the sleep queue and the sleep
3434 	 * queue lock has been dropped.  If we run out of space in our
3435 	 * on-stack buffer, we need to allocate more but we can't call
3436 	 * lmalloc() because we are holding a queue lock when the overflow
3437 	 * occurs and lmalloc() acquires a lock.  We can't use alloca()
3438 	 * either because the application may have allocated a small
3439 	 * stack and we don't want to overrun the stack.  So we call
3440 	 * alloc_lwpids() to allocate a bigger buffer using the mmap()
3441 	 * system call directly since that path acquires no locks.
3442 	 */
3443 	qp = queue_lock(cvp, CV);
3444 	cvp->cond_waiters_user = 0;
3445 	ulwpp = &qp->qh_head;
3446 	while ((ulwp = *ulwpp) != NULL) {
3447 		if (ulwp->ul_wchan != cvp) {
3448 			prev = ulwp;
3449 			ulwpp = &ulwp->ul_link;
3450 			continue;
3451 		}
3452 		*ulwpp = ulwp->ul_link;
3453 		ulwp->ul_link = NULL;
3454 		if (qp->qh_tail == ulwp)
3455 			qp->qh_tail = prev;
3456 		qp->qh_qlen--;
3457 		mp = ulwp->ul_cvmutex;		/* his mutex */
3458 		ulwp->ul_cvmutex = NULL;
3459 		ASSERT(mp != NULL);
3460 		if (ulwp->ul_cv_wake || !MUTEX_OWNED(mp, self)) {
3461 			ulwp->ul_sleepq = NULL;
3462 			ulwp->ul_wchan = NULL;
3463 			ulwp->ul_cv_wake = 0;
3464 			if (nlwpid == maxlwps)
3465 				lwpid = alloc_lwpids(lwpid, &nlwpid, &maxlwps);
3466 			lwpid[nlwpid++] = ulwp->ul_lwpid;
3467 		} else {
3468 			if (mp != mp_cache) {
3469 				mp_cache = mp;
3470 				if (mqp != NULL)
3471 					queue_unlock(mqp);
3472 				mqp = queue_lock(mp, MX);
3473 			}
3474 			enqueue(mqp, ulwp, mp, MX);
3475 			mp->mutex_waiters = 1;
3476 		}
3477 	}
3478 	if (mqp != NULL)
3479 		queue_unlock(mqp);
3480 	if (nlwpid == 0) {
3481 		queue_unlock(qp);
3482 	} else {
3483 		no_preempt(self);
3484 		queue_unlock(qp);
3485 		if (nlwpid == 1)
3486 			(void) __lwp_unpark(lwpid[0]);
3487 		else
3488 			(void) __lwp_unpark_all(lwpid, nlwpid);
3489 		preempt(self);
3490 	}
3491 	if (lwpid != buffer)
3492 		(void) _private_munmap(lwpid, maxlwps * sizeof (lwpid_t));
3493 	return (error);
3494 }
3495 
3496 #pragma weak pthread_cond_destroy = _cond_destroy
3497 #pragma weak _pthread_cond_destroy = _cond_destroy
3498 #pragma weak cond_destroy = _cond_destroy
3499 int
3500 _cond_destroy(cond_t *cvp)
3501 {
3502 	cvp->cond_magic = 0;
3503 	tdb_sync_obj_deregister(cvp);
3504 	return (0);
3505 }
3506 
3507 #if defined(THREAD_DEBUG)
3508 void
3509 assert_no_libc_locks_held(void)
3510 {
3511 	ASSERT(!curthread->ul_critical || curthread->ul_bindflags);
3512 }
3513 #endif
3514 
3515 /* protected by link_lock */
3516 uint64_t spin_lock_spin;
3517 uint64_t spin_lock_spin2;
3518 uint64_t spin_lock_sleep;
3519 uint64_t spin_lock_wakeup;
3520 
3521 /*
3522  * Record spin lock statistics.
3523  * Called by a thread exiting itself in thrp_exit().
3524  * Also called via atexit() from the thread calling
3525  * exit() to do all the other threads as well.
3526  */
3527 void
3528 record_spin_locks(ulwp_t *ulwp)
3529 {
3530 	spin_lock_spin += ulwp->ul_spin_lock_spin;
3531 	spin_lock_spin2 += ulwp->ul_spin_lock_spin2;
3532 	spin_lock_sleep += ulwp->ul_spin_lock_sleep;
3533 	spin_lock_wakeup += ulwp->ul_spin_lock_wakeup;
3534 	ulwp->ul_spin_lock_spin = 0;
3535 	ulwp->ul_spin_lock_spin2 = 0;
3536 	ulwp->ul_spin_lock_sleep = 0;
3537 	ulwp->ul_spin_lock_wakeup = 0;
3538 }
3539 
3540 /*
3541  * atexit function:  dump the queue statistics to stderr.
3542  */
3543 #if !defined(__lint)
3544 #define	fprintf	_fprintf
3545 #endif
3546 #include <stdio.h>
3547 void
3548 dump_queue_statistics(void)
3549 {
3550 	uberdata_t *udp = curthread->ul_uberdata;
3551 	queue_head_t *qp;
3552 	int qn;
3553 	uint64_t spin_lock_total = 0;
3554 
3555 	if (udp->queue_head == NULL || thread_queue_dump == 0)
3556 		return;
3557 
3558 	if (fprintf(stderr, "\n%5d mutex queues:\n", QHASHSIZE) < 0 ||
3559 	    fprintf(stderr, "queue#   lockcount    max qlen\n") < 0)
3560 		return;
3561 	for (qn = 0, qp = udp->queue_head; qn < QHASHSIZE; qn++, qp++) {
3562 		if (qp->qh_lockcount == 0)
3563 			continue;
3564 		spin_lock_total += qp->qh_lockcount;
3565 		if (fprintf(stderr, "%5d %12llu%12u\n", qn,
3566 			(u_longlong_t)qp->qh_lockcount, qp->qh_qmax) < 0)
3567 				return;
3568 	}
3569 
3570 	if (fprintf(stderr, "\n%5d condvar queues:\n", QHASHSIZE) < 0 ||
3571 	    fprintf(stderr, "queue#   lockcount    max qlen\n") < 0)
3572 		return;
3573 	for (qn = 0; qn < QHASHSIZE; qn++, qp++) {
3574 		if (qp->qh_lockcount == 0)
3575 			continue;
3576 		spin_lock_total += qp->qh_lockcount;
3577 		if (fprintf(stderr, "%5d %12llu%12u\n", qn,
3578 			(u_longlong_t)qp->qh_lockcount, qp->qh_qmax) < 0)
3579 				return;
3580 	}
3581 
3582 	(void) fprintf(stderr, "\n  spin_lock_total  = %10llu\n",
3583 		(u_longlong_t)spin_lock_total);
3584 	(void) fprintf(stderr, "  spin_lock_spin   = %10llu\n",
3585 		(u_longlong_t)spin_lock_spin);
3586 	(void) fprintf(stderr, "  spin_lock_spin2  = %10llu\n",
3587 		(u_longlong_t)spin_lock_spin2);
3588 	(void) fprintf(stderr, "  spin_lock_sleep  = %10llu\n",
3589 		(u_longlong_t)spin_lock_sleep);
3590 	(void) fprintf(stderr, "  spin_lock_wakeup = %10llu\n",
3591 		(u_longlong_t)spin_lock_wakeup);
3592 }
3593