xref: /freebsd/sys/kern/kern_umtx.c (revision f5463265955b829775bbb32e1fd0bc11dafc36ce)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2015, 2016 The FreeBSD Foundation
5  * Copyright (c) 2004, David Xu <davidxu@freebsd.org>
6  * Copyright (c) 2002, Jeffrey Roberson <jeff@freebsd.org>
7  * All rights reserved.
8  *
9  * Portions of this software were developed by Konstantin Belousov
10  * under sponsorship from the FreeBSD Foundation.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice unmodified, this list of conditions, and the following
17  *    disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
23  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
24  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
25  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
27  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
31  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include <sys/cdefs.h>
35 #include "opt_umtx_profiling.h"
36 
37 #include <sys/param.h>
38 #include <sys/kernel.h>
39 #include <sys/fcntl.h>
40 #include <sys/file.h>
41 #include <sys/filedesc.h>
42 #include <sys/limits.h>
43 #include <sys/lock.h>
44 #include <sys/malloc.h>
45 #include <sys/mman.h>
46 #include <sys/mutex.h>
47 #include <sys/priv.h>
48 #include <sys/proc.h>
49 #include <sys/resource.h>
50 #include <sys/resourcevar.h>
51 #include <sys/rwlock.h>
52 #include <sys/sbuf.h>
53 #include <sys/sched.h>
54 #include <sys/smp.h>
55 #include <sys/sysctl.h>
56 #include <sys/systm.h>
57 #include <sys/sysproto.h>
58 #include <sys/syscallsubr.h>
59 #include <sys/taskqueue.h>
60 #include <sys/time.h>
61 #include <sys/eventhandler.h>
62 #include <sys/umtx.h>
63 #include <sys/umtxvar.h>
64 
65 #include <security/mac/mac_framework.h>
66 
67 #include <vm/vm.h>
68 #include <vm/vm_param.h>
69 #include <vm/pmap.h>
70 #include <vm/vm_map.h>
71 #include <vm/vm_object.h>
72 
73 #include <machine/atomic.h>
74 #include <machine/cpu.h>
75 
76 #include <compat/freebsd32/freebsd32.h>
77 #ifdef COMPAT_FREEBSD32
78 #include <compat/freebsd32/freebsd32_proto.h>
79 #endif
80 
81 #define _UMUTEX_TRY		1
82 #define _UMUTEX_WAIT		2
83 
84 #ifdef UMTX_PROFILING
85 #define	UPROF_PERC_BIGGER(w, f, sw, sf)					\
86 	(((w) > (sw)) || ((w) == (sw) && (f) > (sf)))
87 #endif
88 
89 #define	UMTXQ_LOCKED_ASSERT(uc)		mtx_assert(&(uc)->uc_lock, MA_OWNED)
90 #ifdef INVARIANTS
91 #define	UMTXQ_ASSERT_LOCKED_BUSY(key) do {				\
92 	struct umtxq_chain *uc;						\
93 									\
94 	uc = umtxq_getchain(key);					\
95 	mtx_assert(&uc->uc_lock, MA_OWNED);				\
96 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));		\
97 } while (0)
98 #else
99 #define	UMTXQ_ASSERT_LOCKED_BUSY(key) do {} while (0)
100 #endif
101 
102 /*
103  * Don't propagate time-sharing priority, there is a security reason,
104  * a user can simply introduce PI-mutex, let thread A lock the mutex,
105  * and let another thread B block on the mutex, because B is
106  * sleeping, its priority will be boosted, this causes A's priority to
107  * be boosted via priority propagating too and will never be lowered even
108  * if it is using 100%CPU, this is unfair to other processes.
109  */
110 
111 #define UPRI(td)	(((td)->td_user_pri >= PRI_MIN_TIMESHARE &&\
112 			  (td)->td_user_pri <= PRI_MAX_TIMESHARE) ?\
113 			 PRI_MAX_TIMESHARE : (td)->td_user_pri)
114 
115 #define	GOLDEN_RATIO_PRIME	2654404609U
116 #ifndef	UMTX_CHAINS
117 #define	UMTX_CHAINS		512
118 #endif
119 #define	UMTX_SHIFTS		(__WORD_BIT - 9)
120 
121 #define	GET_SHARE(flags)	\
122     (((flags) & USYNC_PROCESS_SHARED) == 0 ? THREAD_SHARE : PROCESS_SHARE)
123 
124 #define BUSY_SPINS		200
125 
126 struct umtx_copyops {
127 	int	(*copyin_timeout)(const void *uaddr, struct timespec *tsp);
128 	int	(*copyin_umtx_time)(const void *uaddr, size_t size,
129 	    struct _umtx_time *tp);
130 	int	(*copyin_robust_lists)(const void *uaddr, size_t size,
131 	    struct umtx_robust_lists_params *rbp);
132 	int	(*copyout_timeout)(void *uaddr, size_t size,
133 	    struct timespec *tsp);
134 	const size_t	timespec_sz;
135 	const size_t	umtx_time_sz;
136 	const bool	compat32;
137 };
138 
139 _Static_assert(sizeof(struct umutex) == sizeof(struct umutex32), "umutex32");
140 _Static_assert(__offsetof(struct umutex, m_spare[0]) ==
141     __offsetof(struct umutex32, m_spare[0]), "m_spare32");
142 
143 int umtx_shm_vnobj_persistent = 0;
144 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_vnode_persistent, CTLFLAG_RWTUN,
145     &umtx_shm_vnobj_persistent, 0,
146     "False forces destruction of umtx attached to file, on last close");
147 static int umtx_max_rb = 1000;
148 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_max_robust, CTLFLAG_RWTUN,
149     &umtx_max_rb, 0,
150     "Maximum number of robust mutexes allowed for each thread");
151 
152 static uma_zone_t		umtx_pi_zone;
153 static struct umtxq_chain	umtxq_chains[2][UMTX_CHAINS];
154 static MALLOC_DEFINE(M_UMTX, "umtx", "UMTX queue memory");
155 static int			umtx_pi_allocated;
156 
157 static SYSCTL_NODE(_debug, OID_AUTO, umtx, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
158     "umtx debug");
159 SYSCTL_INT(_debug_umtx, OID_AUTO, umtx_pi_allocated, CTLFLAG_RD,
160     &umtx_pi_allocated, 0, "Allocated umtx_pi");
161 static int umtx_verbose_rb = 1;
162 SYSCTL_INT(_debug_umtx, OID_AUTO, robust_faults_verbose, CTLFLAG_RWTUN,
163     &umtx_verbose_rb, 0,
164     "");
165 
166 #ifdef UMTX_PROFILING
167 static long max_length;
168 SYSCTL_LONG(_debug_umtx, OID_AUTO, max_length, CTLFLAG_RD, &max_length, 0, "max_length");
169 static SYSCTL_NODE(_debug_umtx, OID_AUTO, chains, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
170     "umtx chain stats");
171 #endif
172 
173 static inline void umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
174     const struct _umtx_time *umtxtime);
175 
176 static void umtx_shm_init(void);
177 static void umtxq_sysinit(void *);
178 static void umtxq_hash(struct umtx_key *key);
179 static int do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags,
180     bool rb);
181 static void umtx_thread_cleanup(struct thread *td);
182 SYSINIT(umtx, SI_SUB_EVENTHANDLER+1, SI_ORDER_MIDDLE, umtxq_sysinit, NULL);
183 
184 #define umtxq_signal(key, nwake)	umtxq_signal_queue((key), (nwake), UMTX_SHARED_QUEUE)
185 
186 static struct mtx umtx_lock;
187 
188 #ifdef UMTX_PROFILING
189 static void
190 umtx_init_profiling(void)
191 {
192 	struct sysctl_oid *chain_oid;
193 	char chain_name[10];
194 	int i;
195 
196 	for (i = 0; i < UMTX_CHAINS; ++i) {
197 		snprintf(chain_name, sizeof(chain_name), "%d", i);
198 		chain_oid = SYSCTL_ADD_NODE(NULL,
199 		    SYSCTL_STATIC_CHILDREN(_debug_umtx_chains), OID_AUTO,
200 		    chain_name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
201 		    "umtx hash stats");
202 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
203 		    "max_length0", CTLFLAG_RD, &umtxq_chains[0][i].max_length, 0, NULL);
204 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
205 		    "max_length1", CTLFLAG_RD, &umtxq_chains[1][i].max_length, 0, NULL);
206 	}
207 }
208 
209 static int
210 sysctl_debug_umtx_chains_peaks(SYSCTL_HANDLER_ARGS)
211 {
212 	char buf[512];
213 	struct sbuf sb;
214 	struct umtxq_chain *uc;
215 	u_int fract, i, j, tot, whole;
216 	u_int sf0, sf1, sf2, sf3, sf4;
217 	u_int si0, si1, si2, si3, si4;
218 	u_int sw0, sw1, sw2, sw3, sw4;
219 
220 	sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
221 	for (i = 0; i < 2; i++) {
222 		tot = 0;
223 		for (j = 0; j < UMTX_CHAINS; ++j) {
224 			uc = &umtxq_chains[i][j];
225 			mtx_lock(&uc->uc_lock);
226 			tot += uc->max_length;
227 			mtx_unlock(&uc->uc_lock);
228 		}
229 		if (tot == 0)
230 			sbuf_printf(&sb, "%u) Empty ", i);
231 		else {
232 			sf0 = sf1 = sf2 = sf3 = sf4 = 0;
233 			si0 = si1 = si2 = si3 = si4 = 0;
234 			sw0 = sw1 = sw2 = sw3 = sw4 = 0;
235 			for (j = 0; j < UMTX_CHAINS; j++) {
236 				uc = &umtxq_chains[i][j];
237 				mtx_lock(&uc->uc_lock);
238 				whole = uc->max_length * 100;
239 				mtx_unlock(&uc->uc_lock);
240 				fract = (whole % tot) * 100;
241 				if (UPROF_PERC_BIGGER(whole, fract, sw0, sf0)) {
242 					sf0 = fract;
243 					si0 = j;
244 					sw0 = whole;
245 				} else if (UPROF_PERC_BIGGER(whole, fract, sw1,
246 				    sf1)) {
247 					sf1 = fract;
248 					si1 = j;
249 					sw1 = whole;
250 				} else if (UPROF_PERC_BIGGER(whole, fract, sw2,
251 				    sf2)) {
252 					sf2 = fract;
253 					si2 = j;
254 					sw2 = whole;
255 				} else if (UPROF_PERC_BIGGER(whole, fract, sw3,
256 				    sf3)) {
257 					sf3 = fract;
258 					si3 = j;
259 					sw3 = whole;
260 				} else if (UPROF_PERC_BIGGER(whole, fract, sw4,
261 				    sf4)) {
262 					sf4 = fract;
263 					si4 = j;
264 					sw4 = whole;
265 				}
266 			}
267 			sbuf_printf(&sb, "queue %u:\n", i);
268 			sbuf_printf(&sb, "1st: %u.%u%% idx: %u\n", sw0 / tot,
269 			    sf0 / tot, si0);
270 			sbuf_printf(&sb, "2nd: %u.%u%% idx: %u\n", sw1 / tot,
271 			    sf1 / tot, si1);
272 			sbuf_printf(&sb, "3rd: %u.%u%% idx: %u\n", sw2 / tot,
273 			    sf2 / tot, si2);
274 			sbuf_printf(&sb, "4th: %u.%u%% idx: %u\n", sw3 / tot,
275 			    sf3 / tot, si3);
276 			sbuf_printf(&sb, "5th: %u.%u%% idx: %u\n", sw4 / tot,
277 			    sf4 / tot, si4);
278 		}
279 	}
280 	sbuf_trim(&sb);
281 	sbuf_finish(&sb);
282 	sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
283 	sbuf_delete(&sb);
284 	return (0);
285 }
286 
287 static int
288 sysctl_debug_umtx_chains_clear(SYSCTL_HANDLER_ARGS)
289 {
290 	struct umtxq_chain *uc;
291 	u_int i, j;
292 	int clear, error;
293 
294 	clear = 0;
295 	error = sysctl_handle_int(oidp, &clear, 0, req);
296 	if (error != 0 || req->newptr == NULL)
297 		return (error);
298 
299 	if (clear != 0) {
300 		for (i = 0; i < 2; ++i) {
301 			for (j = 0; j < UMTX_CHAINS; ++j) {
302 				uc = &umtxq_chains[i][j];
303 				mtx_lock(&uc->uc_lock);
304 				uc->length = 0;
305 				uc->max_length = 0;
306 				mtx_unlock(&uc->uc_lock);
307 			}
308 		}
309 	}
310 	return (0);
311 }
312 
313 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, clear,
314     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
315     sysctl_debug_umtx_chains_clear, "I",
316     "Clear umtx chains statistics");
317 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, peaks,
318     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
319     sysctl_debug_umtx_chains_peaks, "A",
320     "Highest peaks in chains max length");
321 #endif
322 
323 static void
324 umtxq_sysinit(void *arg __unused)
325 {
326 	int i, j;
327 
328 	umtx_pi_zone = uma_zcreate("umtx pi", sizeof(struct umtx_pi),
329 		NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
330 	for (i = 0; i < 2; ++i) {
331 		for (j = 0; j < UMTX_CHAINS; ++j) {
332 			mtx_init(&umtxq_chains[i][j].uc_lock, "umtxql", NULL,
333 				 MTX_DEF | MTX_DUPOK);
334 			LIST_INIT(&umtxq_chains[i][j].uc_queue[0]);
335 			LIST_INIT(&umtxq_chains[i][j].uc_queue[1]);
336 			LIST_INIT(&umtxq_chains[i][j].uc_spare_queue);
337 			TAILQ_INIT(&umtxq_chains[i][j].uc_pi_list);
338 			umtxq_chains[i][j].uc_busy = 0;
339 			umtxq_chains[i][j].uc_waiters = 0;
340 #ifdef UMTX_PROFILING
341 			umtxq_chains[i][j].length = 0;
342 			umtxq_chains[i][j].max_length = 0;
343 #endif
344 		}
345 	}
346 #ifdef UMTX_PROFILING
347 	umtx_init_profiling();
348 #endif
349 	mtx_init(&umtx_lock, "umtx lock", NULL, MTX_DEF);
350 	umtx_shm_init();
351 }
352 
353 struct umtx_q *
354 umtxq_alloc(void)
355 {
356 	struct umtx_q *uq;
357 
358 	uq = malloc(sizeof(struct umtx_q), M_UMTX, M_WAITOK | M_ZERO);
359 	uq->uq_spare_queue = malloc(sizeof(struct umtxq_queue), M_UMTX,
360 	    M_WAITOK | M_ZERO);
361 	TAILQ_INIT(&uq->uq_spare_queue->head);
362 	TAILQ_INIT(&uq->uq_pi_contested);
363 	uq->uq_inherited_pri = PRI_MAX;
364 	return (uq);
365 }
366 
367 void
368 umtxq_free(struct umtx_q *uq)
369 {
370 
371 	MPASS(uq->uq_spare_queue != NULL);
372 	free(uq->uq_spare_queue, M_UMTX);
373 	free(uq, M_UMTX);
374 }
375 
376 static inline void
377 umtxq_hash(struct umtx_key *key)
378 {
379 	unsigned n;
380 
381 	n = (uintptr_t)key->info.both.a + key->info.both.b;
382 	key->hash = ((n * GOLDEN_RATIO_PRIME) >> UMTX_SHIFTS) % UMTX_CHAINS;
383 }
384 
385 struct umtxq_chain *
386 umtxq_getchain(struct umtx_key *key)
387 {
388 
389 	if (key->type <= TYPE_SEM)
390 		return (&umtxq_chains[1][key->hash]);
391 	return (&umtxq_chains[0][key->hash]);
392 }
393 
394 /*
395  * Set chain to busy state when following operation
396  * may be blocked (kernel mutex can not be used).
397  */
398 void
399 umtxq_busy(struct umtx_key *key)
400 {
401 	struct umtxq_chain *uc;
402 
403 	uc = umtxq_getchain(key);
404 	mtx_assert(&uc->uc_lock, MA_OWNED);
405 	if (uc->uc_busy) {
406 #ifdef SMP
407 		if (smp_cpus > 1) {
408 			int count = BUSY_SPINS;
409 			if (count > 0) {
410 				umtxq_unlock(key);
411 				while (uc->uc_busy && --count > 0)
412 					cpu_spinwait();
413 				umtxq_lock(key);
414 			}
415 		}
416 #endif
417 		while (uc->uc_busy) {
418 			uc->uc_waiters++;
419 			msleep(uc, &uc->uc_lock, 0, "umtxqb", 0);
420 			uc->uc_waiters--;
421 		}
422 	}
423 	uc->uc_busy = 1;
424 }
425 
426 /*
427  * Unbusy a chain.
428  */
429 void
430 umtxq_unbusy(struct umtx_key *key)
431 {
432 	struct umtxq_chain *uc;
433 
434 	uc = umtxq_getchain(key);
435 	mtx_assert(&uc->uc_lock, MA_OWNED);
436 	KASSERT(uc->uc_busy != 0, ("not busy"));
437 	uc->uc_busy = 0;
438 	if (uc->uc_waiters)
439 		wakeup_one(uc);
440 }
441 
442 void
443 umtxq_unbusy_unlocked(struct umtx_key *key)
444 {
445 
446 	umtxq_lock(key);
447 	umtxq_unbusy(key);
448 	umtxq_unlock(key);
449 }
450 
451 static struct umtxq_queue *
452 umtxq_queue_lookup(struct umtx_key *key, int q)
453 {
454 	struct umtxq_queue *uh;
455 	struct umtxq_chain *uc;
456 
457 	uc = umtxq_getchain(key);
458 	UMTXQ_LOCKED_ASSERT(uc);
459 	LIST_FOREACH(uh, &uc->uc_queue[q], link) {
460 		if (umtx_key_match(&uh->key, key))
461 			return (uh);
462 	}
463 
464 	return (NULL);
465 }
466 
467 void
468 umtxq_insert_queue(struct umtx_q *uq, int q)
469 {
470 	struct umtxq_queue *uh;
471 	struct umtxq_chain *uc;
472 
473 	uc = umtxq_getchain(&uq->uq_key);
474 	UMTXQ_LOCKED_ASSERT(uc);
475 	KASSERT((uq->uq_flags & UQF_UMTXQ) == 0, ("umtx_q is already on queue"));
476 	uh = umtxq_queue_lookup(&uq->uq_key, q);
477 	if (uh != NULL) {
478 		LIST_INSERT_HEAD(&uc->uc_spare_queue, uq->uq_spare_queue, link);
479 	} else {
480 		uh = uq->uq_spare_queue;
481 		uh->key = uq->uq_key;
482 		LIST_INSERT_HEAD(&uc->uc_queue[q], uh, link);
483 #ifdef UMTX_PROFILING
484 		uc->length++;
485 		if (uc->length > uc->max_length) {
486 			uc->max_length = uc->length;
487 			if (uc->max_length > max_length)
488 				max_length = uc->max_length;
489 		}
490 #endif
491 	}
492 	uq->uq_spare_queue = NULL;
493 
494 	TAILQ_INSERT_TAIL(&uh->head, uq, uq_link);
495 	uh->length++;
496 	uq->uq_flags |= UQF_UMTXQ;
497 	uq->uq_cur_queue = uh;
498 	return;
499 }
500 
501 void
502 umtxq_remove_queue(struct umtx_q *uq, int q)
503 {
504 	struct umtxq_chain *uc;
505 	struct umtxq_queue *uh;
506 
507 	uc = umtxq_getchain(&uq->uq_key);
508 	UMTXQ_LOCKED_ASSERT(uc);
509 	if (uq->uq_flags & UQF_UMTXQ) {
510 		uh = uq->uq_cur_queue;
511 		TAILQ_REMOVE(&uh->head, uq, uq_link);
512 		uh->length--;
513 		uq->uq_flags &= ~UQF_UMTXQ;
514 		if (TAILQ_EMPTY(&uh->head)) {
515 			KASSERT(uh->length == 0,
516 			    ("inconsistent umtxq_queue length"));
517 #ifdef UMTX_PROFILING
518 			uc->length--;
519 #endif
520 			LIST_REMOVE(uh, link);
521 		} else {
522 			uh = LIST_FIRST(&uc->uc_spare_queue);
523 			KASSERT(uh != NULL, ("uc_spare_queue is empty"));
524 			LIST_REMOVE(uh, link);
525 		}
526 		uq->uq_spare_queue = uh;
527 		uq->uq_cur_queue = NULL;
528 	}
529 }
530 
531 /*
532  * Check if there are multiple waiters
533  */
534 int
535 umtxq_count(struct umtx_key *key)
536 {
537 	struct umtxq_queue *uh;
538 
539 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
540 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
541 	if (uh != NULL)
542 		return (uh->length);
543 	return (0);
544 }
545 
546 /*
547  * Check if there are multiple PI waiters and returns first
548  * waiter.
549  */
550 static int
551 umtxq_count_pi(struct umtx_key *key, struct umtx_q **first)
552 {
553 	struct umtxq_queue *uh;
554 
555 	*first = NULL;
556 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
557 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
558 	if (uh != NULL) {
559 		*first = TAILQ_FIRST(&uh->head);
560 		return (uh->length);
561 	}
562 	return (0);
563 }
564 
565 /*
566  * Wake up threads waiting on an userland object by a bit mask.
567  */
568 int
569 umtxq_signal_mask(struct umtx_key *key, int n_wake, u_int bitset)
570 {
571 	struct umtxq_queue *uh;
572 	struct umtx_q *uq, *uq_temp;
573 	int ret;
574 
575 	ret = 0;
576 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
577 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
578 	if (uh == NULL)
579 		return (0);
580 	TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
581 		if ((uq->uq_bitset & bitset) == 0)
582 			continue;
583 		umtxq_remove_queue(uq, UMTX_SHARED_QUEUE);
584 		wakeup_one(uq);
585 		if (++ret >= n_wake)
586 			break;
587 	}
588 	return (ret);
589 }
590 
591 /*
592  * Wake up threads waiting on an userland object.
593  */
594 
595 static int
596 umtxq_signal_queue(struct umtx_key *key, int n_wake, int q)
597 {
598 	struct umtxq_queue *uh;
599 	struct umtx_q *uq;
600 	int ret;
601 
602 	ret = 0;
603 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
604 	uh = umtxq_queue_lookup(key, q);
605 	if (uh != NULL) {
606 		while ((uq = TAILQ_FIRST(&uh->head)) != NULL) {
607 			umtxq_remove_queue(uq, q);
608 			wakeup(uq);
609 			if (++ret >= n_wake)
610 				return (ret);
611 		}
612 	}
613 	return (ret);
614 }
615 
616 /*
617  * Wake up specified thread.
618  */
619 static inline void
620 umtxq_signal_thread(struct umtx_q *uq)
621 {
622 
623 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
624 	umtxq_remove(uq);
625 	wakeup(uq);
626 }
627 
628 /*
629  * Wake up a maximum of n_wake threads that are waiting on an userland
630  * object identified by key. The remaining threads are removed from queue
631  * identified by key and added to the queue identified by key2 (requeued).
632  * The n_requeue specifies an upper limit on the number of threads that
633  * are requeued to the second queue.
634  */
635 int
636 umtxq_requeue(struct umtx_key *key, int n_wake, struct umtx_key *key2,
637     int n_requeue)
638 {
639 	struct umtxq_queue *uh;
640 	struct umtx_q *uq, *uq_temp;
641 	int ret;
642 
643 	ret = 0;
644 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
645 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key2));
646 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
647 	if (uh == NULL)
648 		return (0);
649 	TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
650 		if (++ret <= n_wake) {
651 			umtxq_remove(uq);
652 			wakeup_one(uq);
653 		} else {
654 			umtxq_remove(uq);
655 			uq->uq_key = *key2;
656 			umtxq_insert(uq);
657 			if (ret - n_wake == n_requeue)
658 				break;
659 		}
660 	}
661 	return (ret);
662 }
663 
664 static inline int
665 tstohz(const struct timespec *tsp)
666 {
667 	struct timeval tv;
668 
669 	TIMESPEC_TO_TIMEVAL(&tv, tsp);
670 	return tvtohz(&tv);
671 }
672 
673 void
674 umtx_abs_timeout_init(struct umtx_abs_timeout *timo, int clockid,
675     int absolute, const struct timespec *timeout)
676 {
677 
678 	timo->clockid = clockid;
679 	if (!absolute) {
680 		timo->is_abs_real = false;
681 		kern_clock_gettime(curthread, timo->clockid, &timo->cur);
682 		timespecadd(&timo->cur, timeout, &timo->end);
683 	} else {
684 		timo->end = *timeout;
685 		timo->is_abs_real = clockid == CLOCK_REALTIME ||
686 		    clockid == CLOCK_REALTIME_FAST ||
687 		    clockid == CLOCK_REALTIME_PRECISE ||
688 		    clockid == CLOCK_SECOND;
689 	}
690 }
691 
692 static void
693 umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
694     const struct _umtx_time *umtxtime)
695 {
696 
697 	umtx_abs_timeout_init(timo, umtxtime->_clockid,
698 	    (umtxtime->_flags & UMTX_ABSTIME) != 0, &umtxtime->_timeout);
699 }
700 
701 static void
702 umtx_abs_timeout_enforce_min(sbintime_t *sbt)
703 {
704 	sbintime_t when, mint;
705 
706 	mint = curproc->p_umtx_min_timeout;
707 	if (__predict_false(mint != 0)) {
708 		when = sbinuptime() + mint;
709 		if (*sbt < when)
710 			*sbt = when;
711 	}
712 }
713 
714 static int
715 umtx_abs_timeout_getsbt(struct umtx_abs_timeout *timo, sbintime_t *sbt,
716     int *flags)
717 {
718 	struct bintime bt, bbt;
719 	struct timespec tts;
720 	sbintime_t rem;
721 
722 	switch (timo->clockid) {
723 
724 	/* Clocks that can be converted into absolute time. */
725 	case CLOCK_REALTIME:
726 	case CLOCK_REALTIME_PRECISE:
727 	case CLOCK_REALTIME_FAST:
728 	case CLOCK_MONOTONIC:
729 	case CLOCK_MONOTONIC_PRECISE:
730 	case CLOCK_MONOTONIC_FAST:
731 	case CLOCK_UPTIME:
732 	case CLOCK_UPTIME_PRECISE:
733 	case CLOCK_UPTIME_FAST:
734 	case CLOCK_SECOND:
735 		timespec2bintime(&timo->end, &bt);
736 		switch (timo->clockid) {
737 		case CLOCK_REALTIME:
738 		case CLOCK_REALTIME_PRECISE:
739 		case CLOCK_REALTIME_FAST:
740 		case CLOCK_SECOND:
741 			getboottimebin(&bbt);
742 			bintime_sub(&bt, &bbt);
743 			break;
744 		}
745 		if (bt.sec < 0)
746 			return (ETIMEDOUT);
747 		if (bt.sec >= (SBT_MAX >> 32)) {
748 			*sbt = 0;
749 			*flags = 0;
750 			return (0);
751 		}
752 		*sbt = bttosbt(bt);
753 		umtx_abs_timeout_enforce_min(sbt);
754 
755 		/*
756 		 * Check if the absolute time should be aligned to
757 		 * avoid firing multiple timer events in non-periodic
758 		 * timer mode.
759 		 */
760 		switch (timo->clockid) {
761 		case CLOCK_REALTIME_FAST:
762 		case CLOCK_MONOTONIC_FAST:
763 		case CLOCK_UPTIME_FAST:
764 			rem = *sbt % tc_tick_sbt;
765 			if (__predict_true(rem != 0))
766 				*sbt += tc_tick_sbt - rem;
767 			break;
768 		case CLOCK_SECOND:
769 			rem = *sbt % SBT_1S;
770 			if (__predict_true(rem != 0))
771 				*sbt += SBT_1S - rem;
772 			break;
773 		}
774 		*flags = C_ABSOLUTE;
775 		return (0);
776 
777 	/* Clocks that has to be periodically polled. */
778 	case CLOCK_VIRTUAL:
779 	case CLOCK_PROF:
780 	case CLOCK_THREAD_CPUTIME_ID:
781 	case CLOCK_PROCESS_CPUTIME_ID:
782 	default:
783 		kern_clock_gettime(curthread, timo->clockid, &timo->cur);
784 		if (timespeccmp(&timo->end, &timo->cur, <=))
785 			return (ETIMEDOUT);
786 		timespecsub(&timo->end, &timo->cur, &tts);
787 		*sbt = tick_sbt * tstohz(&tts);
788 		*flags = C_HARDCLOCK;
789 		return (0);
790 	}
791 }
792 
793 static uint32_t
794 umtx_unlock_val(uint32_t flags, bool rb)
795 {
796 
797 	if (rb)
798 		return (UMUTEX_RB_OWNERDEAD);
799 	else if ((flags & UMUTEX_NONCONSISTENT) != 0)
800 		return (UMUTEX_RB_NOTRECOV);
801 	else
802 		return (UMUTEX_UNOWNED);
803 
804 }
805 
806 /*
807  * Put thread into sleep state, before sleeping, check if
808  * thread was removed from umtx queue.
809  */
810 int
811 umtxq_sleep(struct umtx_q *uq, const char *wmesg,
812     struct umtx_abs_timeout *timo)
813 {
814 	struct umtxq_chain *uc;
815 	sbintime_t sbt = 0;
816 	int error, flags = 0;
817 
818 	uc = umtxq_getchain(&uq->uq_key);
819 	UMTXQ_LOCKED_ASSERT(uc);
820 	for (;;) {
821 		if (!(uq->uq_flags & UQF_UMTXQ)) {
822 			error = 0;
823 			break;
824 		}
825 		if (timo != NULL) {
826 			if (timo->is_abs_real)
827 				curthread->td_rtcgen =
828 				    atomic_load_acq_int(&rtc_generation);
829 			error = umtx_abs_timeout_getsbt(timo, &sbt, &flags);
830 			if (error != 0)
831 				break;
832 		}
833 		error = msleep_sbt(uq, &uc->uc_lock, PCATCH | PDROP, wmesg,
834 		    sbt, 0, flags);
835 		uc = umtxq_getchain(&uq->uq_key);
836 		mtx_lock(&uc->uc_lock);
837 		if (error == EINTR || error == ERESTART)
838 			break;
839 		if (error == EWOULDBLOCK && (flags & C_ABSOLUTE) != 0) {
840 			error = ETIMEDOUT;
841 			break;
842 		}
843 	}
844 
845 	curthread->td_rtcgen = 0;
846 	return (error);
847 }
848 
849 /*
850  * Convert userspace address into unique logical address.
851  */
852 int
853 umtx_key_get(const void *addr, int type, int share, struct umtx_key *key)
854 {
855 	struct thread *td = curthread;
856 	vm_map_t map;
857 	vm_map_entry_t entry;
858 	vm_pindex_t pindex;
859 	vm_prot_t prot;
860 	boolean_t wired;
861 
862 	key->type = type;
863 	if (share == THREAD_SHARE) {
864 		key->shared = 0;
865 		key->info.private.vs = td->td_proc->p_vmspace;
866 		key->info.private.addr = (uintptr_t)addr;
867 	} else {
868 		MPASS(share == PROCESS_SHARE || share == AUTO_SHARE);
869 		map = &td->td_proc->p_vmspace->vm_map;
870 		if (vm_map_lookup(&map, (vm_offset_t)addr, VM_PROT_WRITE,
871 		    &entry, &key->info.shared.object, &pindex, &prot,
872 		    &wired) != KERN_SUCCESS) {
873 			return (EFAULT);
874 		}
875 
876 		if ((share == PROCESS_SHARE) ||
877 		    (share == AUTO_SHARE &&
878 		     VM_INHERIT_SHARE == entry->inheritance)) {
879 			key->shared = 1;
880 			key->info.shared.offset = (vm_offset_t)addr -
881 			    entry->start + entry->offset;
882 			vm_object_reference(key->info.shared.object);
883 		} else {
884 			key->shared = 0;
885 			key->info.private.vs = td->td_proc->p_vmspace;
886 			key->info.private.addr = (uintptr_t)addr;
887 		}
888 		vm_map_lookup_done(map, entry);
889 	}
890 
891 	umtxq_hash(key);
892 	return (0);
893 }
894 
895 /*
896  * Release key.
897  */
898 void
899 umtx_key_release(struct umtx_key *key)
900 {
901 	if (key->shared)
902 		vm_object_deallocate(key->info.shared.object);
903 }
904 
905 #ifdef COMPAT_FREEBSD10
906 /*
907  * Lock a umtx object.
908  */
909 static int
910 do_lock_umtx(struct thread *td, struct umtx *umtx, u_long id,
911     const struct timespec *timeout)
912 {
913 	struct umtx_abs_timeout timo;
914 	struct umtx_q *uq;
915 	u_long owner;
916 	u_long old;
917 	int error = 0;
918 
919 	uq = td->td_umtxq;
920 	if (timeout != NULL)
921 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
922 
923 	/*
924 	 * Care must be exercised when dealing with umtx structure. It
925 	 * can fault on any access.
926 	 */
927 	for (;;) {
928 		/*
929 		 * Try the uncontested case.  This should be done in userland.
930 		 */
931 		owner = casuword(&umtx->u_owner, UMTX_UNOWNED, id);
932 
933 		/* The acquire succeeded. */
934 		if (owner == UMTX_UNOWNED)
935 			return (0);
936 
937 		/* The address was invalid. */
938 		if (owner == -1)
939 			return (EFAULT);
940 
941 		/* If no one owns it but it is contested try to acquire it. */
942 		if (owner == UMTX_CONTESTED) {
943 			owner = casuword(&umtx->u_owner,
944 			    UMTX_CONTESTED, id | UMTX_CONTESTED);
945 
946 			if (owner == UMTX_CONTESTED)
947 				return (0);
948 
949 			/* The address was invalid. */
950 			if (owner == -1)
951 				return (EFAULT);
952 
953 			error = thread_check_susp(td, false);
954 			if (error != 0)
955 				break;
956 
957 			/* If this failed the lock has changed, restart. */
958 			continue;
959 		}
960 
961 		/*
962 		 * If we caught a signal, we have retried and now
963 		 * exit immediately.
964 		 */
965 		if (error != 0)
966 			break;
967 
968 		if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK,
969 			AUTO_SHARE, &uq->uq_key)) != 0)
970 			return (error);
971 
972 		umtxq_lock(&uq->uq_key);
973 		umtxq_busy(&uq->uq_key);
974 		umtxq_insert(uq);
975 		umtxq_unbusy(&uq->uq_key);
976 		umtxq_unlock(&uq->uq_key);
977 
978 		/*
979 		 * Set the contested bit so that a release in user space
980 		 * knows to use the system call for unlock.  If this fails
981 		 * either some one else has acquired the lock or it has been
982 		 * released.
983 		 */
984 		old = casuword(&umtx->u_owner, owner, owner | UMTX_CONTESTED);
985 
986 		/* The address was invalid. */
987 		if (old == -1) {
988 			umtxq_lock(&uq->uq_key);
989 			umtxq_remove(uq);
990 			umtxq_unlock(&uq->uq_key);
991 			umtx_key_release(&uq->uq_key);
992 			return (EFAULT);
993 		}
994 
995 		/*
996 		 * We set the contested bit, sleep. Otherwise the lock changed
997 		 * and we need to retry or we lost a race to the thread
998 		 * unlocking the umtx.
999 		 */
1000 		umtxq_lock(&uq->uq_key);
1001 		if (old == owner)
1002 			error = umtxq_sleep(uq, "umtx", timeout == NULL ? NULL :
1003 			    &timo);
1004 		umtxq_remove(uq);
1005 		umtxq_unlock(&uq->uq_key);
1006 		umtx_key_release(&uq->uq_key);
1007 
1008 		if (error == 0)
1009 			error = thread_check_susp(td, false);
1010 	}
1011 
1012 	if (timeout == NULL) {
1013 		/* Mutex locking is restarted if it is interrupted. */
1014 		if (error == EINTR)
1015 			error = ERESTART;
1016 	} else {
1017 		/* Timed-locking is not restarted. */
1018 		if (error == ERESTART)
1019 			error = EINTR;
1020 	}
1021 	return (error);
1022 }
1023 
1024 /*
1025  * Unlock a umtx object.
1026  */
1027 static int
1028 do_unlock_umtx(struct thread *td, struct umtx *umtx, u_long id)
1029 {
1030 	struct umtx_key key;
1031 	u_long owner;
1032 	u_long old;
1033 	int error;
1034 	int count;
1035 
1036 	/*
1037 	 * Make sure we own this mtx.
1038 	 */
1039 	owner = fuword(__DEVOLATILE(u_long *, &umtx->u_owner));
1040 	if (owner == -1)
1041 		return (EFAULT);
1042 
1043 	if ((owner & ~UMTX_CONTESTED) != id)
1044 		return (EPERM);
1045 
1046 	/* This should be done in userland */
1047 	if ((owner & UMTX_CONTESTED) == 0) {
1048 		old = casuword(&umtx->u_owner, owner, UMTX_UNOWNED);
1049 		if (old == -1)
1050 			return (EFAULT);
1051 		if (old == owner)
1052 			return (0);
1053 		owner = old;
1054 	}
1055 
1056 	/* We should only ever be in here for contested locks */
1057 	if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1058 	    &key)) != 0)
1059 		return (error);
1060 
1061 	umtxq_lock(&key);
1062 	umtxq_busy(&key);
1063 	count = umtxq_count(&key);
1064 	umtxq_unlock(&key);
1065 
1066 	/*
1067 	 * When unlocking the umtx, it must be marked as unowned if
1068 	 * there is zero or one thread only waiting for it.
1069 	 * Otherwise, it must be marked as contested.
1070 	 */
1071 	old = casuword(&umtx->u_owner, owner,
1072 	    count <= 1 ? UMTX_UNOWNED : UMTX_CONTESTED);
1073 	umtxq_lock(&key);
1074 	umtxq_signal(&key,1);
1075 	umtxq_unbusy(&key);
1076 	umtxq_unlock(&key);
1077 	umtx_key_release(&key);
1078 	if (old == -1)
1079 		return (EFAULT);
1080 	if (old != owner)
1081 		return (EINVAL);
1082 	return (0);
1083 }
1084 
1085 #ifdef COMPAT_FREEBSD32
1086 
1087 /*
1088  * Lock a umtx object.
1089  */
1090 static int
1091 do_lock_umtx32(struct thread *td, uint32_t *m, uint32_t id,
1092 	const struct timespec *timeout)
1093 {
1094 	struct umtx_abs_timeout timo;
1095 	struct umtx_q *uq;
1096 	uint32_t owner;
1097 	uint32_t old;
1098 	int error = 0;
1099 
1100 	uq = td->td_umtxq;
1101 
1102 	if (timeout != NULL)
1103 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
1104 
1105 	/*
1106 	 * Care must be exercised when dealing with umtx structure. It
1107 	 * can fault on any access.
1108 	 */
1109 	for (;;) {
1110 		/*
1111 		 * Try the uncontested case.  This should be done in userland.
1112 		 */
1113 		owner = casuword32(m, UMUTEX_UNOWNED, id);
1114 
1115 		/* The acquire succeeded. */
1116 		if (owner == UMUTEX_UNOWNED)
1117 			return (0);
1118 
1119 		/* The address was invalid. */
1120 		if (owner == -1)
1121 			return (EFAULT);
1122 
1123 		/* If no one owns it but it is contested try to acquire it. */
1124 		if (owner == UMUTEX_CONTESTED) {
1125 			owner = casuword32(m,
1126 			    UMUTEX_CONTESTED, id | UMUTEX_CONTESTED);
1127 			if (owner == UMUTEX_CONTESTED)
1128 				return (0);
1129 
1130 			/* The address was invalid. */
1131 			if (owner == -1)
1132 				return (EFAULT);
1133 
1134 			error = thread_check_susp(td, false);
1135 			if (error != 0)
1136 				break;
1137 
1138 			/* If this failed the lock has changed, restart. */
1139 			continue;
1140 		}
1141 
1142 		/*
1143 		 * If we caught a signal, we have retried and now
1144 		 * exit immediately.
1145 		 */
1146 		if (error != 0)
1147 			return (error);
1148 
1149 		if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK,
1150 			AUTO_SHARE, &uq->uq_key)) != 0)
1151 			return (error);
1152 
1153 		umtxq_lock(&uq->uq_key);
1154 		umtxq_busy(&uq->uq_key);
1155 		umtxq_insert(uq);
1156 		umtxq_unbusy(&uq->uq_key);
1157 		umtxq_unlock(&uq->uq_key);
1158 
1159 		/*
1160 		 * Set the contested bit so that a release in user space
1161 		 * knows to use the system call for unlock.  If this fails
1162 		 * either some one else has acquired the lock or it has been
1163 		 * released.
1164 		 */
1165 		old = casuword32(m, owner, owner | UMUTEX_CONTESTED);
1166 
1167 		/* The address was invalid. */
1168 		if (old == -1) {
1169 			umtxq_lock(&uq->uq_key);
1170 			umtxq_remove(uq);
1171 			umtxq_unlock(&uq->uq_key);
1172 			umtx_key_release(&uq->uq_key);
1173 			return (EFAULT);
1174 		}
1175 
1176 		/*
1177 		 * We set the contested bit, sleep. Otherwise the lock changed
1178 		 * and we need to retry or we lost a race to the thread
1179 		 * unlocking the umtx.
1180 		 */
1181 		umtxq_lock(&uq->uq_key);
1182 		if (old == owner)
1183 			error = umtxq_sleep(uq, "umtx", timeout == NULL ?
1184 			    NULL : &timo);
1185 		umtxq_remove(uq);
1186 		umtxq_unlock(&uq->uq_key);
1187 		umtx_key_release(&uq->uq_key);
1188 
1189 		if (error == 0)
1190 			error = thread_check_susp(td, false);
1191 	}
1192 
1193 	if (timeout == NULL) {
1194 		/* Mutex locking is restarted if it is interrupted. */
1195 		if (error == EINTR)
1196 			error = ERESTART;
1197 	} else {
1198 		/* Timed-locking is not restarted. */
1199 		if (error == ERESTART)
1200 			error = EINTR;
1201 	}
1202 	return (error);
1203 }
1204 
1205 /*
1206  * Unlock a umtx object.
1207  */
1208 static int
1209 do_unlock_umtx32(struct thread *td, uint32_t *m, uint32_t id)
1210 {
1211 	struct umtx_key key;
1212 	uint32_t owner;
1213 	uint32_t old;
1214 	int error;
1215 	int count;
1216 
1217 	/*
1218 	 * Make sure we own this mtx.
1219 	 */
1220 	owner = fuword32(m);
1221 	if (owner == -1)
1222 		return (EFAULT);
1223 
1224 	if ((owner & ~UMUTEX_CONTESTED) != id)
1225 		return (EPERM);
1226 
1227 	/* This should be done in userland */
1228 	if ((owner & UMUTEX_CONTESTED) == 0) {
1229 		old = casuword32(m, owner, UMUTEX_UNOWNED);
1230 		if (old == -1)
1231 			return (EFAULT);
1232 		if (old == owner)
1233 			return (0);
1234 		owner = old;
1235 	}
1236 
1237 	/* We should only ever be in here for contested locks */
1238 	if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1239 		&key)) != 0)
1240 		return (error);
1241 
1242 	umtxq_lock(&key);
1243 	umtxq_busy(&key);
1244 	count = umtxq_count(&key);
1245 	umtxq_unlock(&key);
1246 
1247 	/*
1248 	 * When unlocking the umtx, it must be marked as unowned if
1249 	 * there is zero or one thread only waiting for it.
1250 	 * Otherwise, it must be marked as contested.
1251 	 */
1252 	old = casuword32(m, owner,
1253 		count <= 1 ? UMUTEX_UNOWNED : UMUTEX_CONTESTED);
1254 	umtxq_lock(&key);
1255 	umtxq_signal(&key,1);
1256 	umtxq_unbusy(&key);
1257 	umtxq_unlock(&key);
1258 	umtx_key_release(&key);
1259 	if (old == -1)
1260 		return (EFAULT);
1261 	if (old != owner)
1262 		return (EINVAL);
1263 	return (0);
1264 }
1265 #endif	/* COMPAT_FREEBSD32 */
1266 #endif	/* COMPAT_FREEBSD10 */
1267 
1268 /*
1269  * Fetch and compare value, sleep on the address if value is not changed.
1270  */
1271 static int
1272 do_wait(struct thread *td, void *addr, u_long id,
1273     struct _umtx_time *timeout, int compat32, int is_private)
1274 {
1275 	struct umtx_abs_timeout timo;
1276 	struct umtx_q *uq;
1277 	u_long tmp;
1278 	uint32_t tmp32;
1279 	int error = 0;
1280 
1281 	uq = td->td_umtxq;
1282 	if ((error = umtx_key_get(addr, TYPE_SIMPLE_WAIT,
1283 	    is_private ? THREAD_SHARE : AUTO_SHARE, &uq->uq_key)) != 0)
1284 		return (error);
1285 
1286 	if (timeout != NULL)
1287 		umtx_abs_timeout_init2(&timo, timeout);
1288 
1289 	umtxq_lock(&uq->uq_key);
1290 	umtxq_insert(uq);
1291 	umtxq_unlock(&uq->uq_key);
1292 	if (compat32 == 0) {
1293 		error = fueword(addr, &tmp);
1294 		if (error != 0)
1295 			error = EFAULT;
1296 	} else {
1297 		error = fueword32(addr, &tmp32);
1298 		if (error == 0)
1299 			tmp = tmp32;
1300 		else
1301 			error = EFAULT;
1302 	}
1303 	umtxq_lock(&uq->uq_key);
1304 	if (error == 0) {
1305 		if (tmp == id)
1306 			error = umtxq_sleep(uq, "uwait", timeout == NULL ?
1307 			    NULL : &timo);
1308 		if ((uq->uq_flags & UQF_UMTXQ) == 0)
1309 			error = 0;
1310 		else
1311 			umtxq_remove(uq);
1312 	} else if ((uq->uq_flags & UQF_UMTXQ) != 0) {
1313 		umtxq_remove(uq);
1314 	}
1315 	umtxq_unlock(&uq->uq_key);
1316 	umtx_key_release(&uq->uq_key);
1317 	if (error == ERESTART)
1318 		error = EINTR;
1319 	return (error);
1320 }
1321 
1322 /*
1323  * Wake up threads sleeping on the specified address.
1324  */
1325 int
1326 kern_umtx_wake(struct thread *td, void *uaddr, int n_wake, int is_private)
1327 {
1328 	struct umtx_key key;
1329 	int ret;
1330 
1331 	if ((ret = umtx_key_get(uaddr, TYPE_SIMPLE_WAIT,
1332 	    is_private ? THREAD_SHARE : AUTO_SHARE, &key)) != 0)
1333 		return (ret);
1334 	umtxq_lock(&key);
1335 	umtxq_signal(&key, n_wake);
1336 	umtxq_unlock(&key);
1337 	umtx_key_release(&key);
1338 	return (0);
1339 }
1340 
1341 /*
1342  * Lock PTHREAD_PRIO_NONE protocol POSIX mutex.
1343  */
1344 static int
1345 do_lock_normal(struct thread *td, struct umutex *m, uint32_t flags,
1346     struct _umtx_time *timeout, int mode)
1347 {
1348 	struct umtx_abs_timeout timo;
1349 	struct umtx_q *uq;
1350 	uint32_t owner, old, id;
1351 	int error, rv;
1352 
1353 	id = td->td_tid;
1354 	uq = td->td_umtxq;
1355 	error = 0;
1356 	if (timeout != NULL)
1357 		umtx_abs_timeout_init2(&timo, timeout);
1358 
1359 	/*
1360 	 * Care must be exercised when dealing with umtx structure. It
1361 	 * can fault on any access.
1362 	 */
1363 	for (;;) {
1364 		rv = fueword32(&m->m_owner, &owner);
1365 		if (rv == -1)
1366 			return (EFAULT);
1367 		if (mode == _UMUTEX_WAIT) {
1368 			if (owner == UMUTEX_UNOWNED ||
1369 			    owner == UMUTEX_CONTESTED ||
1370 			    owner == UMUTEX_RB_OWNERDEAD ||
1371 			    owner == UMUTEX_RB_NOTRECOV)
1372 				return (0);
1373 		} else {
1374 			/*
1375 			 * Robust mutex terminated.  Kernel duty is to
1376 			 * return EOWNERDEAD to the userspace.  The
1377 			 * umutex.m_flags UMUTEX_NONCONSISTENT is set
1378 			 * by the common userspace code.
1379 			 */
1380 			if (owner == UMUTEX_RB_OWNERDEAD) {
1381 				rv = casueword32(&m->m_owner,
1382 				    UMUTEX_RB_OWNERDEAD, &owner,
1383 				    id | UMUTEX_CONTESTED);
1384 				if (rv == -1)
1385 					return (EFAULT);
1386 				if (rv == 0) {
1387 					MPASS(owner == UMUTEX_RB_OWNERDEAD);
1388 					return (EOWNERDEAD); /* success */
1389 				}
1390 				MPASS(rv == 1);
1391 				rv = thread_check_susp(td, false);
1392 				if (rv != 0)
1393 					return (rv);
1394 				continue;
1395 			}
1396 			if (owner == UMUTEX_RB_NOTRECOV)
1397 				return (ENOTRECOVERABLE);
1398 
1399 			/*
1400 			 * Try the uncontested case.  This should be
1401 			 * done in userland.
1402 			 */
1403 			rv = casueword32(&m->m_owner, UMUTEX_UNOWNED,
1404 			    &owner, id);
1405 			/* The address was invalid. */
1406 			if (rv == -1)
1407 				return (EFAULT);
1408 
1409 			/* The acquire succeeded. */
1410 			if (rv == 0) {
1411 				MPASS(owner == UMUTEX_UNOWNED);
1412 				return (0);
1413 			}
1414 
1415 			/*
1416 			 * If no one owns it but it is contested try
1417 			 * to acquire it.
1418 			 */
1419 			MPASS(rv == 1);
1420 			if (owner == UMUTEX_CONTESTED) {
1421 				rv = casueword32(&m->m_owner,
1422 				    UMUTEX_CONTESTED, &owner,
1423 				    id | UMUTEX_CONTESTED);
1424 				/* The address was invalid. */
1425 				if (rv == -1)
1426 					return (EFAULT);
1427 				if (rv == 0) {
1428 					MPASS(owner == UMUTEX_CONTESTED);
1429 					return (0);
1430 				}
1431 				if (rv == 1) {
1432 					rv = thread_check_susp(td, false);
1433 					if (rv != 0)
1434 						return (rv);
1435 				}
1436 
1437 				/*
1438 				 * If this failed the lock has
1439 				 * changed, restart.
1440 				 */
1441 				continue;
1442 			}
1443 
1444 			/* rv == 1 but not contested, likely store failure */
1445 			rv = thread_check_susp(td, false);
1446 			if (rv != 0)
1447 				return (rv);
1448 		}
1449 
1450 		if (mode == _UMUTEX_TRY)
1451 			return (EBUSY);
1452 
1453 		/*
1454 		 * If we caught a signal, we have retried and now
1455 		 * exit immediately.
1456 		 */
1457 		if (error != 0)
1458 			return (error);
1459 
1460 		if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX,
1461 		    GET_SHARE(flags), &uq->uq_key)) != 0)
1462 			return (error);
1463 
1464 		umtxq_lock(&uq->uq_key);
1465 		umtxq_busy(&uq->uq_key);
1466 		umtxq_insert(uq);
1467 		umtxq_unlock(&uq->uq_key);
1468 
1469 		/*
1470 		 * Set the contested bit so that a release in user space
1471 		 * knows to use the system call for unlock.  If this fails
1472 		 * either some one else has acquired the lock or it has been
1473 		 * released.
1474 		 */
1475 		rv = casueword32(&m->m_owner, owner, &old,
1476 		    owner | UMUTEX_CONTESTED);
1477 
1478 		/* The address was invalid or casueword failed to store. */
1479 		if (rv == -1 || rv == 1) {
1480 			umtxq_lock(&uq->uq_key);
1481 			umtxq_remove(uq);
1482 			umtxq_unbusy(&uq->uq_key);
1483 			umtxq_unlock(&uq->uq_key);
1484 			umtx_key_release(&uq->uq_key);
1485 			if (rv == -1)
1486 				return (EFAULT);
1487 			if (rv == 1) {
1488 				rv = thread_check_susp(td, false);
1489 				if (rv != 0)
1490 					return (rv);
1491 			}
1492 			continue;
1493 		}
1494 
1495 		/*
1496 		 * We set the contested bit, sleep. Otherwise the lock changed
1497 		 * and we need to retry or we lost a race to the thread
1498 		 * unlocking the umtx.
1499 		 */
1500 		umtxq_lock(&uq->uq_key);
1501 		umtxq_unbusy(&uq->uq_key);
1502 		MPASS(old == owner);
1503 		error = umtxq_sleep(uq, "umtxn", timeout == NULL ?
1504 		    NULL : &timo);
1505 		umtxq_remove(uq);
1506 		umtxq_unlock(&uq->uq_key);
1507 		umtx_key_release(&uq->uq_key);
1508 
1509 		if (error == 0)
1510 			error = thread_check_susp(td, false);
1511 	}
1512 
1513 	return (0);
1514 }
1515 
1516 /*
1517  * Unlock PTHREAD_PRIO_NONE protocol POSIX mutex.
1518  */
1519 static int
1520 do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1521 {
1522 	struct umtx_key key;
1523 	uint32_t owner, old, id, newlock;
1524 	int error, count;
1525 
1526 	id = td->td_tid;
1527 
1528 again:
1529 	/*
1530 	 * Make sure we own this mtx.
1531 	 */
1532 	error = fueword32(&m->m_owner, &owner);
1533 	if (error == -1)
1534 		return (EFAULT);
1535 
1536 	if ((owner & ~UMUTEX_CONTESTED) != id)
1537 		return (EPERM);
1538 
1539 	newlock = umtx_unlock_val(flags, rb);
1540 	if ((owner & UMUTEX_CONTESTED) == 0) {
1541 		error = casueword32(&m->m_owner, owner, &old, newlock);
1542 		if (error == -1)
1543 			return (EFAULT);
1544 		if (error == 1) {
1545 			error = thread_check_susp(td, false);
1546 			if (error != 0)
1547 				return (error);
1548 			goto again;
1549 		}
1550 		MPASS(old == owner);
1551 		return (0);
1552 	}
1553 
1554 	/* We should only ever be in here for contested locks */
1555 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1556 	    &key)) != 0)
1557 		return (error);
1558 
1559 	umtxq_lock(&key);
1560 	umtxq_busy(&key);
1561 	count = umtxq_count(&key);
1562 	umtxq_unlock(&key);
1563 
1564 	/*
1565 	 * When unlocking the umtx, it must be marked as unowned if
1566 	 * there is zero or one thread only waiting for it.
1567 	 * Otherwise, it must be marked as contested.
1568 	 */
1569 	if (count > 1)
1570 		newlock |= UMUTEX_CONTESTED;
1571 	error = casueword32(&m->m_owner, owner, &old, newlock);
1572 	umtxq_lock(&key);
1573 	umtxq_signal(&key, 1);
1574 	umtxq_unbusy(&key);
1575 	umtxq_unlock(&key);
1576 	umtx_key_release(&key);
1577 	if (error == -1)
1578 		return (EFAULT);
1579 	if (error == 1) {
1580 		if (old != owner)
1581 			return (EINVAL);
1582 		error = thread_check_susp(td, false);
1583 		if (error != 0)
1584 			return (error);
1585 		goto again;
1586 	}
1587 	return (0);
1588 }
1589 
1590 /*
1591  * Check if the mutex is available and wake up a waiter,
1592  * only for simple mutex.
1593  */
1594 static int
1595 do_wake_umutex(struct thread *td, struct umutex *m)
1596 {
1597 	struct umtx_key key;
1598 	uint32_t owner;
1599 	uint32_t flags;
1600 	int error;
1601 	int count;
1602 
1603 again:
1604 	error = fueword32(&m->m_owner, &owner);
1605 	if (error == -1)
1606 		return (EFAULT);
1607 
1608 	if ((owner & ~UMUTEX_CONTESTED) != 0 && owner != UMUTEX_RB_OWNERDEAD &&
1609 	    owner != UMUTEX_RB_NOTRECOV)
1610 		return (0);
1611 
1612 	error = fueword32(&m->m_flags, &flags);
1613 	if (error == -1)
1614 		return (EFAULT);
1615 
1616 	/* We should only ever be in here for contested locks */
1617 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1618 	    &key)) != 0)
1619 		return (error);
1620 
1621 	umtxq_lock(&key);
1622 	umtxq_busy(&key);
1623 	count = umtxq_count(&key);
1624 	umtxq_unlock(&key);
1625 
1626 	if (count <= 1 && owner != UMUTEX_RB_OWNERDEAD &&
1627 	    owner != UMUTEX_RB_NOTRECOV) {
1628 		error = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
1629 		    UMUTEX_UNOWNED);
1630 		if (error == -1) {
1631 			error = EFAULT;
1632 		} else if (error == 1) {
1633 			umtxq_lock(&key);
1634 			umtxq_unbusy(&key);
1635 			umtxq_unlock(&key);
1636 			umtx_key_release(&key);
1637 			error = thread_check_susp(td, false);
1638 			if (error != 0)
1639 				return (error);
1640 			goto again;
1641 		}
1642 	}
1643 
1644 	umtxq_lock(&key);
1645 	if (error == 0 && count != 0) {
1646 		MPASS((owner & ~UMUTEX_CONTESTED) == 0 ||
1647 		    owner == UMUTEX_RB_OWNERDEAD ||
1648 		    owner == UMUTEX_RB_NOTRECOV);
1649 		umtxq_signal(&key, 1);
1650 	}
1651 	umtxq_unbusy(&key);
1652 	umtxq_unlock(&key);
1653 	umtx_key_release(&key);
1654 	return (error);
1655 }
1656 
1657 /*
1658  * Check if the mutex has waiters and tries to fix contention bit.
1659  */
1660 static int
1661 do_wake2_umutex(struct thread *td, struct umutex *m, uint32_t flags)
1662 {
1663 	struct umtx_key key;
1664 	uint32_t owner, old;
1665 	int type;
1666 	int error;
1667 	int count;
1668 
1669 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT |
1670 	    UMUTEX_ROBUST)) {
1671 	case 0:
1672 	case UMUTEX_ROBUST:
1673 		type = TYPE_NORMAL_UMUTEX;
1674 		break;
1675 	case UMUTEX_PRIO_INHERIT:
1676 		type = TYPE_PI_UMUTEX;
1677 		break;
1678 	case (UMUTEX_PRIO_INHERIT | UMUTEX_ROBUST):
1679 		type = TYPE_PI_ROBUST_UMUTEX;
1680 		break;
1681 	case UMUTEX_PRIO_PROTECT:
1682 		type = TYPE_PP_UMUTEX;
1683 		break;
1684 	case (UMUTEX_PRIO_PROTECT | UMUTEX_ROBUST):
1685 		type = TYPE_PP_ROBUST_UMUTEX;
1686 		break;
1687 	default:
1688 		return (EINVAL);
1689 	}
1690 	if ((error = umtx_key_get(m, type, GET_SHARE(flags), &key)) != 0)
1691 		return (error);
1692 
1693 	owner = 0;
1694 	umtxq_lock(&key);
1695 	umtxq_busy(&key);
1696 	count = umtxq_count(&key);
1697 	umtxq_unlock(&key);
1698 
1699 	error = fueword32(&m->m_owner, &owner);
1700 	if (error == -1)
1701 		error = EFAULT;
1702 
1703 	/*
1704 	 * Only repair contention bit if there is a waiter, this means
1705 	 * the mutex is still being referenced by userland code,
1706 	 * otherwise don't update any memory.
1707 	 */
1708 	while (error == 0 && (owner & UMUTEX_CONTESTED) == 0 &&
1709 	    (count > 1 || (count == 1 && (owner & ~UMUTEX_CONTESTED) != 0))) {
1710 		error = casueword32(&m->m_owner, owner, &old,
1711 		    owner | UMUTEX_CONTESTED);
1712 		if (error == -1) {
1713 			error = EFAULT;
1714 			break;
1715 		}
1716 		if (error == 0) {
1717 			MPASS(old == owner);
1718 			break;
1719 		}
1720 		owner = old;
1721 		error = thread_check_susp(td, false);
1722 	}
1723 
1724 	umtxq_lock(&key);
1725 	if (error == EFAULT) {
1726 		umtxq_signal(&key, INT_MAX);
1727 	} else if (count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1728 	    owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1729 		umtxq_signal(&key, 1);
1730 	umtxq_unbusy(&key);
1731 	umtxq_unlock(&key);
1732 	umtx_key_release(&key);
1733 	return (error);
1734 }
1735 
1736 struct umtx_pi *
1737 umtx_pi_alloc(int flags)
1738 {
1739 	struct umtx_pi *pi;
1740 
1741 	pi = uma_zalloc(umtx_pi_zone, M_ZERO | flags);
1742 	TAILQ_INIT(&pi->pi_blocked);
1743 	atomic_add_int(&umtx_pi_allocated, 1);
1744 	return (pi);
1745 }
1746 
1747 void
1748 umtx_pi_free(struct umtx_pi *pi)
1749 {
1750 	uma_zfree(umtx_pi_zone, pi);
1751 	atomic_add_int(&umtx_pi_allocated, -1);
1752 }
1753 
1754 /*
1755  * Adjust the thread's position on a pi_state after its priority has been
1756  * changed.
1757  */
1758 static int
1759 umtx_pi_adjust_thread(struct umtx_pi *pi, struct thread *td)
1760 {
1761 	struct umtx_q *uq, *uq1, *uq2;
1762 	struct thread *td1;
1763 
1764 	mtx_assert(&umtx_lock, MA_OWNED);
1765 	if (pi == NULL)
1766 		return (0);
1767 
1768 	uq = td->td_umtxq;
1769 
1770 	/*
1771 	 * Check if the thread needs to be moved on the blocked chain.
1772 	 * It needs to be moved if either its priority is lower than
1773 	 * the previous thread or higher than the next thread.
1774 	 */
1775 	uq1 = TAILQ_PREV(uq, umtxq_head, uq_lockq);
1776 	uq2 = TAILQ_NEXT(uq, uq_lockq);
1777 	if ((uq1 != NULL && UPRI(td) < UPRI(uq1->uq_thread)) ||
1778 	    (uq2 != NULL && UPRI(td) > UPRI(uq2->uq_thread))) {
1779 		/*
1780 		 * Remove thread from blocked chain and determine where
1781 		 * it should be moved to.
1782 		 */
1783 		TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1784 		TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1785 			td1 = uq1->uq_thread;
1786 			MPASS(td1->td_proc->p_magic == P_MAGIC);
1787 			if (UPRI(td1) > UPRI(td))
1788 				break;
1789 		}
1790 
1791 		if (uq1 == NULL)
1792 			TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1793 		else
1794 			TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1795 	}
1796 	return (1);
1797 }
1798 
1799 static struct umtx_pi *
1800 umtx_pi_next(struct umtx_pi *pi)
1801 {
1802 	struct umtx_q *uq_owner;
1803 
1804 	if (pi->pi_owner == NULL)
1805 		return (NULL);
1806 	uq_owner = pi->pi_owner->td_umtxq;
1807 	if (uq_owner == NULL)
1808 		return (NULL);
1809 	return (uq_owner->uq_pi_blocked);
1810 }
1811 
1812 /*
1813  * Floyd's Cycle-Finding Algorithm.
1814  */
1815 static bool
1816 umtx_pi_check_loop(struct umtx_pi *pi)
1817 {
1818 	struct umtx_pi *pi1;	/* fast iterator */
1819 
1820 	mtx_assert(&umtx_lock, MA_OWNED);
1821 	if (pi == NULL)
1822 		return (false);
1823 	pi1 = pi;
1824 	for (;;) {
1825 		pi = umtx_pi_next(pi);
1826 		if (pi == NULL)
1827 			break;
1828 		pi1 = umtx_pi_next(pi1);
1829 		if (pi1 == NULL)
1830 			break;
1831 		pi1 = umtx_pi_next(pi1);
1832 		if (pi1 == NULL)
1833 			break;
1834 		if (pi == pi1)
1835 			return (true);
1836 	}
1837 	return (false);
1838 }
1839 
1840 /*
1841  * Propagate priority when a thread is blocked on POSIX
1842  * PI mutex.
1843  */
1844 static void
1845 umtx_propagate_priority(struct thread *td)
1846 {
1847 	struct umtx_q *uq;
1848 	struct umtx_pi *pi;
1849 	int pri;
1850 
1851 	mtx_assert(&umtx_lock, MA_OWNED);
1852 	pri = UPRI(td);
1853 	uq = td->td_umtxq;
1854 	pi = uq->uq_pi_blocked;
1855 	if (pi == NULL)
1856 		return;
1857 	if (umtx_pi_check_loop(pi))
1858 		return;
1859 
1860 	for (;;) {
1861 		td = pi->pi_owner;
1862 		if (td == NULL || td == curthread)
1863 			return;
1864 
1865 		MPASS(td->td_proc != NULL);
1866 		MPASS(td->td_proc->p_magic == P_MAGIC);
1867 
1868 		thread_lock(td);
1869 		if (td->td_lend_user_pri > pri)
1870 			sched_lend_user_prio(td, pri);
1871 		else {
1872 			thread_unlock(td);
1873 			break;
1874 		}
1875 		thread_unlock(td);
1876 
1877 		/*
1878 		 * Pick up the lock that td is blocked on.
1879 		 */
1880 		uq = td->td_umtxq;
1881 		pi = uq->uq_pi_blocked;
1882 		if (pi == NULL)
1883 			break;
1884 		/* Resort td on the list if needed. */
1885 		umtx_pi_adjust_thread(pi, td);
1886 	}
1887 }
1888 
1889 /*
1890  * Unpropagate priority for a PI mutex when a thread blocked on
1891  * it is interrupted by signal or resumed by others.
1892  */
1893 static void
1894 umtx_repropagate_priority(struct umtx_pi *pi)
1895 {
1896 	struct umtx_q *uq, *uq_owner;
1897 	struct umtx_pi *pi2;
1898 	int pri;
1899 
1900 	mtx_assert(&umtx_lock, MA_OWNED);
1901 
1902 	if (umtx_pi_check_loop(pi))
1903 		return;
1904 	while (pi != NULL && pi->pi_owner != NULL) {
1905 		pri = PRI_MAX;
1906 		uq_owner = pi->pi_owner->td_umtxq;
1907 
1908 		TAILQ_FOREACH(pi2, &uq_owner->uq_pi_contested, pi_link) {
1909 			uq = TAILQ_FIRST(&pi2->pi_blocked);
1910 			if (uq != NULL) {
1911 				if (pri > UPRI(uq->uq_thread))
1912 					pri = UPRI(uq->uq_thread);
1913 			}
1914 		}
1915 
1916 		if (pri > uq_owner->uq_inherited_pri)
1917 			pri = uq_owner->uq_inherited_pri;
1918 		thread_lock(pi->pi_owner);
1919 		sched_lend_user_prio(pi->pi_owner, pri);
1920 		thread_unlock(pi->pi_owner);
1921 		if ((pi = uq_owner->uq_pi_blocked) != NULL)
1922 			umtx_pi_adjust_thread(pi, uq_owner->uq_thread);
1923 	}
1924 }
1925 
1926 /*
1927  * Insert a PI mutex into owned list.
1928  */
1929 static void
1930 umtx_pi_setowner(struct umtx_pi *pi, struct thread *owner)
1931 {
1932 	struct umtx_q *uq_owner;
1933 
1934 	uq_owner = owner->td_umtxq;
1935 	mtx_assert(&umtx_lock, MA_OWNED);
1936 	MPASS(pi->pi_owner == NULL);
1937 	pi->pi_owner = owner;
1938 	TAILQ_INSERT_TAIL(&uq_owner->uq_pi_contested, pi, pi_link);
1939 }
1940 
1941 /*
1942  * Disown a PI mutex, and remove it from the owned list.
1943  */
1944 static void
1945 umtx_pi_disown(struct umtx_pi *pi)
1946 {
1947 
1948 	mtx_assert(&umtx_lock, MA_OWNED);
1949 	TAILQ_REMOVE(&pi->pi_owner->td_umtxq->uq_pi_contested, pi, pi_link);
1950 	pi->pi_owner = NULL;
1951 }
1952 
1953 /*
1954  * Claim ownership of a PI mutex.
1955  */
1956 int
1957 umtx_pi_claim(struct umtx_pi *pi, struct thread *owner)
1958 {
1959 	struct umtx_q *uq;
1960 	int pri;
1961 
1962 	mtx_lock(&umtx_lock);
1963 	if (pi->pi_owner == owner) {
1964 		mtx_unlock(&umtx_lock);
1965 		return (0);
1966 	}
1967 
1968 	if (pi->pi_owner != NULL) {
1969 		/*
1970 		 * userland may have already messed the mutex, sigh.
1971 		 */
1972 		mtx_unlock(&umtx_lock);
1973 		return (EPERM);
1974 	}
1975 	umtx_pi_setowner(pi, owner);
1976 	uq = TAILQ_FIRST(&pi->pi_blocked);
1977 	if (uq != NULL) {
1978 		pri = UPRI(uq->uq_thread);
1979 		thread_lock(owner);
1980 		if (pri < UPRI(owner))
1981 			sched_lend_user_prio(owner, pri);
1982 		thread_unlock(owner);
1983 	}
1984 	mtx_unlock(&umtx_lock);
1985 	return (0);
1986 }
1987 
1988 /*
1989  * Adjust a thread's order position in its blocked PI mutex,
1990  * this may result new priority propagating process.
1991  */
1992 void
1993 umtx_pi_adjust(struct thread *td, u_char oldpri)
1994 {
1995 	struct umtx_q *uq;
1996 	struct umtx_pi *pi;
1997 
1998 	uq = td->td_umtxq;
1999 	mtx_lock(&umtx_lock);
2000 	/*
2001 	 * Pick up the lock that td is blocked on.
2002 	 */
2003 	pi = uq->uq_pi_blocked;
2004 	if (pi != NULL) {
2005 		umtx_pi_adjust_thread(pi, td);
2006 		umtx_repropagate_priority(pi);
2007 	}
2008 	mtx_unlock(&umtx_lock);
2009 }
2010 
2011 /*
2012  * Sleep on a PI mutex.
2013  */
2014 int
2015 umtxq_sleep_pi(struct umtx_q *uq, struct umtx_pi *pi, uint32_t owner,
2016     const char *wmesg, struct umtx_abs_timeout *timo, bool shared)
2017 {
2018 	struct thread *td, *td1;
2019 	struct umtx_q *uq1;
2020 	int error, pri;
2021 #ifdef INVARIANTS
2022 	struct umtxq_chain *uc;
2023 
2024 	uc = umtxq_getchain(&pi->pi_key);
2025 #endif
2026 	error = 0;
2027 	td = uq->uq_thread;
2028 	KASSERT(td == curthread, ("inconsistent uq_thread"));
2029 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
2030 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));
2031 	umtxq_insert(uq);
2032 	mtx_lock(&umtx_lock);
2033 	if (pi->pi_owner == NULL) {
2034 		mtx_unlock(&umtx_lock);
2035 		td1 = tdfind(owner, shared ? -1 : td->td_proc->p_pid);
2036 		mtx_lock(&umtx_lock);
2037 		if (td1 != NULL) {
2038 			if (pi->pi_owner == NULL)
2039 				umtx_pi_setowner(pi, td1);
2040 			PROC_UNLOCK(td1->td_proc);
2041 		}
2042 	}
2043 
2044 	TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
2045 		pri = UPRI(uq1->uq_thread);
2046 		if (pri > UPRI(td))
2047 			break;
2048 	}
2049 
2050 	if (uq1 != NULL)
2051 		TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
2052 	else
2053 		TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
2054 
2055 	uq->uq_pi_blocked = pi;
2056 	thread_lock(td);
2057 	td->td_flags |= TDF_UPIBLOCKED;
2058 	thread_unlock(td);
2059 	umtx_propagate_priority(td);
2060 	mtx_unlock(&umtx_lock);
2061 	umtxq_unbusy(&uq->uq_key);
2062 
2063 	error = umtxq_sleep(uq, wmesg, timo);
2064 	umtxq_remove(uq);
2065 
2066 	mtx_lock(&umtx_lock);
2067 	uq->uq_pi_blocked = NULL;
2068 	thread_lock(td);
2069 	td->td_flags &= ~TDF_UPIBLOCKED;
2070 	thread_unlock(td);
2071 	TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
2072 	umtx_repropagate_priority(pi);
2073 	mtx_unlock(&umtx_lock);
2074 	umtxq_unlock(&uq->uq_key);
2075 
2076 	return (error);
2077 }
2078 
2079 /*
2080  * Add reference count for a PI mutex.
2081  */
2082 void
2083 umtx_pi_ref(struct umtx_pi *pi)
2084 {
2085 
2086 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&pi->pi_key));
2087 	pi->pi_refcount++;
2088 }
2089 
2090 /*
2091  * Decrease reference count for a PI mutex, if the counter
2092  * is decreased to zero, its memory space is freed.
2093  */
2094 void
2095 umtx_pi_unref(struct umtx_pi *pi)
2096 {
2097 	struct umtxq_chain *uc;
2098 
2099 	uc = umtxq_getchain(&pi->pi_key);
2100 	UMTXQ_LOCKED_ASSERT(uc);
2101 	KASSERT(pi->pi_refcount > 0, ("invalid reference count"));
2102 	if (--pi->pi_refcount == 0) {
2103 		mtx_lock(&umtx_lock);
2104 		if (pi->pi_owner != NULL)
2105 			umtx_pi_disown(pi);
2106 		KASSERT(TAILQ_EMPTY(&pi->pi_blocked),
2107 			("blocked queue not empty"));
2108 		mtx_unlock(&umtx_lock);
2109 		TAILQ_REMOVE(&uc->uc_pi_list, pi, pi_hashlink);
2110 		umtx_pi_free(pi);
2111 	}
2112 }
2113 
2114 /*
2115  * Find a PI mutex in hash table.
2116  */
2117 struct umtx_pi *
2118 umtx_pi_lookup(struct umtx_key *key)
2119 {
2120 	struct umtxq_chain *uc;
2121 	struct umtx_pi *pi;
2122 
2123 	uc = umtxq_getchain(key);
2124 	UMTXQ_LOCKED_ASSERT(uc);
2125 
2126 	TAILQ_FOREACH(pi, &uc->uc_pi_list, pi_hashlink) {
2127 		if (umtx_key_match(&pi->pi_key, key)) {
2128 			return (pi);
2129 		}
2130 	}
2131 	return (NULL);
2132 }
2133 
2134 /*
2135  * Insert a PI mutex into hash table.
2136  */
2137 void
2138 umtx_pi_insert(struct umtx_pi *pi)
2139 {
2140 	struct umtxq_chain *uc;
2141 
2142 	uc = umtxq_getchain(&pi->pi_key);
2143 	UMTXQ_LOCKED_ASSERT(uc);
2144 	TAILQ_INSERT_TAIL(&uc->uc_pi_list, pi, pi_hashlink);
2145 }
2146 
2147 /*
2148  * Drop a PI mutex and wakeup a top waiter.
2149  */
2150 int
2151 umtx_pi_drop(struct thread *td, struct umtx_key *key, bool rb, int *count)
2152 {
2153 	struct umtx_q *uq_first, *uq_first2, *uq_me;
2154 	struct umtx_pi *pi, *pi2;
2155 	int pri;
2156 
2157 	UMTXQ_ASSERT_LOCKED_BUSY(key);
2158 	*count = umtxq_count_pi(key, &uq_first);
2159 	if (uq_first != NULL) {
2160 		mtx_lock(&umtx_lock);
2161 		pi = uq_first->uq_pi_blocked;
2162 		KASSERT(pi != NULL, ("pi == NULL?"));
2163 		if (pi->pi_owner != td && !(rb && pi->pi_owner == NULL)) {
2164 			mtx_unlock(&umtx_lock);
2165 			/* userland messed the mutex */
2166 			return (EPERM);
2167 		}
2168 		uq_me = td->td_umtxq;
2169 		if (pi->pi_owner == td)
2170 			umtx_pi_disown(pi);
2171 		/* get highest priority thread which is still sleeping. */
2172 		uq_first = TAILQ_FIRST(&pi->pi_blocked);
2173 		while (uq_first != NULL &&
2174 		    (uq_first->uq_flags & UQF_UMTXQ) == 0) {
2175 			uq_first = TAILQ_NEXT(uq_first, uq_lockq);
2176 		}
2177 		pri = PRI_MAX;
2178 		TAILQ_FOREACH(pi2, &uq_me->uq_pi_contested, pi_link) {
2179 			uq_first2 = TAILQ_FIRST(&pi2->pi_blocked);
2180 			if (uq_first2 != NULL) {
2181 				if (pri > UPRI(uq_first2->uq_thread))
2182 					pri = UPRI(uq_first2->uq_thread);
2183 			}
2184 		}
2185 		thread_lock(td);
2186 		sched_lend_user_prio(td, pri);
2187 		thread_unlock(td);
2188 		mtx_unlock(&umtx_lock);
2189 		if (uq_first)
2190 			umtxq_signal_thread(uq_first);
2191 	} else {
2192 		pi = umtx_pi_lookup(key);
2193 		/*
2194 		 * A umtx_pi can exist if a signal or timeout removed the
2195 		 * last waiter from the umtxq, but there is still
2196 		 * a thread in do_lock_pi() holding the umtx_pi.
2197 		 */
2198 		if (pi != NULL) {
2199 			/*
2200 			 * The umtx_pi can be unowned, such as when a thread
2201 			 * has just entered do_lock_pi(), allocated the
2202 			 * umtx_pi, and unlocked the umtxq.
2203 			 * If the current thread owns it, it must disown it.
2204 			 */
2205 			mtx_lock(&umtx_lock);
2206 			if (pi->pi_owner == td)
2207 				umtx_pi_disown(pi);
2208 			mtx_unlock(&umtx_lock);
2209 		}
2210 	}
2211 	return (0);
2212 }
2213 
2214 /*
2215  * Lock a PI mutex.
2216  */
2217 static int
2218 do_lock_pi(struct thread *td, struct umutex *m, uint32_t flags,
2219     struct _umtx_time *timeout, int try)
2220 {
2221 	struct umtx_abs_timeout timo;
2222 	struct umtx_q *uq;
2223 	struct umtx_pi *pi, *new_pi;
2224 	uint32_t id, old_owner, owner, old;
2225 	int error, rv;
2226 
2227 	id = td->td_tid;
2228 	uq = td->td_umtxq;
2229 
2230 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2231 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2232 	    &uq->uq_key)) != 0)
2233 		return (error);
2234 
2235 	if (timeout != NULL)
2236 		umtx_abs_timeout_init2(&timo, timeout);
2237 
2238 	umtxq_lock(&uq->uq_key);
2239 	pi = umtx_pi_lookup(&uq->uq_key);
2240 	if (pi == NULL) {
2241 		new_pi = umtx_pi_alloc(M_NOWAIT);
2242 		if (new_pi == NULL) {
2243 			umtxq_unlock(&uq->uq_key);
2244 			new_pi = umtx_pi_alloc(M_WAITOK);
2245 			umtxq_lock(&uq->uq_key);
2246 			pi = umtx_pi_lookup(&uq->uq_key);
2247 			if (pi != NULL) {
2248 				umtx_pi_free(new_pi);
2249 				new_pi = NULL;
2250 			}
2251 		}
2252 		if (new_pi != NULL) {
2253 			new_pi->pi_key = uq->uq_key;
2254 			umtx_pi_insert(new_pi);
2255 			pi = new_pi;
2256 		}
2257 	}
2258 	umtx_pi_ref(pi);
2259 	umtxq_unlock(&uq->uq_key);
2260 
2261 	/*
2262 	 * Care must be exercised when dealing with umtx structure.  It
2263 	 * can fault on any access.
2264 	 */
2265 	for (;;) {
2266 		/*
2267 		 * Try the uncontested case.  This should be done in userland.
2268 		 */
2269 		rv = casueword32(&m->m_owner, UMUTEX_UNOWNED, &owner, id);
2270 		/* The address was invalid. */
2271 		if (rv == -1) {
2272 			error = EFAULT;
2273 			break;
2274 		}
2275 		/* The acquire succeeded. */
2276 		if (rv == 0) {
2277 			MPASS(owner == UMUTEX_UNOWNED);
2278 			error = 0;
2279 			break;
2280 		}
2281 
2282 		if (owner == UMUTEX_RB_NOTRECOV) {
2283 			error = ENOTRECOVERABLE;
2284 			break;
2285 		}
2286 
2287 		/*
2288 		 * Nobody owns it, but the acquire failed. This can happen
2289 		 * with ll/sc atomics.
2290 		 */
2291 		if (owner == UMUTEX_UNOWNED) {
2292 			error = thread_check_susp(td, true);
2293 			if (error != 0)
2294 				break;
2295 			continue;
2296 		}
2297 
2298 		/*
2299 		 * Avoid overwriting a possible error from sleep due
2300 		 * to the pending signal with suspension check result.
2301 		 */
2302 		if (error == 0) {
2303 			error = thread_check_susp(td, true);
2304 			if (error != 0)
2305 				break;
2306 		}
2307 
2308 		/* If no one owns it but it is contested try to acquire it. */
2309 		if (owner == UMUTEX_CONTESTED || owner == UMUTEX_RB_OWNERDEAD) {
2310 			old_owner = owner;
2311 			rv = casueword32(&m->m_owner, owner, &owner,
2312 			    id | UMUTEX_CONTESTED);
2313 			/* The address was invalid. */
2314 			if (rv == -1) {
2315 				error = EFAULT;
2316 				break;
2317 			}
2318 			if (rv == 1) {
2319 				if (error == 0) {
2320 					error = thread_check_susp(td, true);
2321 					if (error != 0)
2322 						break;
2323 				}
2324 
2325 				/*
2326 				 * If this failed the lock could
2327 				 * changed, restart.
2328 				 */
2329 				continue;
2330 			}
2331 
2332 			MPASS(rv == 0);
2333 			MPASS(owner == old_owner);
2334 			umtxq_lock(&uq->uq_key);
2335 			umtxq_busy(&uq->uq_key);
2336 			error = umtx_pi_claim(pi, td);
2337 			umtxq_unbusy(&uq->uq_key);
2338 			umtxq_unlock(&uq->uq_key);
2339 			if (error != 0) {
2340 				/*
2341 				 * Since we're going to return an
2342 				 * error, restore the m_owner to its
2343 				 * previous, unowned state to avoid
2344 				 * compounding the problem.
2345 				 */
2346 				(void)casuword32(&m->m_owner,
2347 				    id | UMUTEX_CONTESTED, old_owner);
2348 			}
2349 			if (error == 0 && old_owner == UMUTEX_RB_OWNERDEAD)
2350 				error = EOWNERDEAD;
2351 			break;
2352 		}
2353 
2354 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2355 			error = EDEADLK;
2356 			break;
2357 		}
2358 
2359 		if (try != 0) {
2360 			error = EBUSY;
2361 			break;
2362 		}
2363 
2364 		/*
2365 		 * If we caught a signal, we have retried and now
2366 		 * exit immediately.
2367 		 */
2368 		if (error != 0)
2369 			break;
2370 
2371 		umtxq_lock(&uq->uq_key);
2372 		umtxq_busy(&uq->uq_key);
2373 		umtxq_unlock(&uq->uq_key);
2374 
2375 		/*
2376 		 * Set the contested bit so that a release in user space
2377 		 * knows to use the system call for unlock.  If this fails
2378 		 * either some one else has acquired the lock or it has been
2379 		 * released.
2380 		 */
2381 		rv = casueword32(&m->m_owner, owner, &old, owner |
2382 		    UMUTEX_CONTESTED);
2383 
2384 		/* The address was invalid. */
2385 		if (rv == -1) {
2386 			umtxq_unbusy_unlocked(&uq->uq_key);
2387 			error = EFAULT;
2388 			break;
2389 		}
2390 		if (rv == 1) {
2391 			umtxq_unbusy_unlocked(&uq->uq_key);
2392 			error = thread_check_susp(td, true);
2393 			if (error != 0)
2394 				break;
2395 
2396 			/*
2397 			 * The lock changed and we need to retry or we
2398 			 * lost a race to the thread unlocking the
2399 			 * umtx.  Note that the UMUTEX_RB_OWNERDEAD
2400 			 * value for owner is impossible there.
2401 			 */
2402 			continue;
2403 		}
2404 
2405 		umtxq_lock(&uq->uq_key);
2406 
2407 		/* We set the contested bit, sleep. */
2408 		MPASS(old == owner);
2409 		error = umtxq_sleep_pi(uq, pi, owner & ~UMUTEX_CONTESTED,
2410 		    "umtxpi", timeout == NULL ? NULL : &timo,
2411 		    (flags & USYNC_PROCESS_SHARED) != 0);
2412 		if (error != 0)
2413 			continue;
2414 
2415 		error = thread_check_susp(td, false);
2416 		if (error != 0)
2417 			break;
2418 	}
2419 
2420 	umtxq_lock(&uq->uq_key);
2421 	umtx_pi_unref(pi);
2422 	umtxq_unlock(&uq->uq_key);
2423 
2424 	umtx_key_release(&uq->uq_key);
2425 	return (error);
2426 }
2427 
2428 /*
2429  * Unlock a PI mutex.
2430  */
2431 static int
2432 do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2433 {
2434 	struct umtx_key key;
2435 	uint32_t id, new_owner, old, owner;
2436 	int count, error;
2437 
2438 	id = td->td_tid;
2439 
2440 usrloop:
2441 	/*
2442 	 * Make sure we own this mtx.
2443 	 */
2444 	error = fueword32(&m->m_owner, &owner);
2445 	if (error == -1)
2446 		return (EFAULT);
2447 
2448 	if ((owner & ~UMUTEX_CONTESTED) != id)
2449 		return (EPERM);
2450 
2451 	new_owner = umtx_unlock_val(flags, rb);
2452 
2453 	/* This should be done in userland */
2454 	if ((owner & UMUTEX_CONTESTED) == 0) {
2455 		error = casueword32(&m->m_owner, owner, &old, new_owner);
2456 		if (error == -1)
2457 			return (EFAULT);
2458 		if (error == 1) {
2459 			error = thread_check_susp(td, true);
2460 			if (error != 0)
2461 				return (error);
2462 			goto usrloop;
2463 		}
2464 		if (old == owner)
2465 			return (0);
2466 		owner = old;
2467 	}
2468 
2469 	/* We should only ever be in here for contested locks */
2470 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2471 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2472 	    &key)) != 0)
2473 		return (error);
2474 
2475 	umtxq_lock(&key);
2476 	umtxq_busy(&key);
2477 	error = umtx_pi_drop(td, &key, rb, &count);
2478 	if (error != 0) {
2479 		umtxq_unbusy(&key);
2480 		umtxq_unlock(&key);
2481 		umtx_key_release(&key);
2482 		/* userland messed the mutex */
2483 		return (error);
2484 	}
2485 	umtxq_unlock(&key);
2486 
2487 	/*
2488 	 * When unlocking the umtx, it must be marked as unowned if
2489 	 * there is zero or one thread only waiting for it.
2490 	 * Otherwise, it must be marked as contested.
2491 	 */
2492 
2493 	if (count > 1)
2494 		new_owner |= UMUTEX_CONTESTED;
2495 again:
2496 	error = casueword32(&m->m_owner, owner, &old, new_owner);
2497 	if (error == 1) {
2498 		error = thread_check_susp(td, false);
2499 		if (error == 0)
2500 			goto again;
2501 	}
2502 	umtxq_unbusy_unlocked(&key);
2503 	umtx_key_release(&key);
2504 	if (error == -1)
2505 		return (EFAULT);
2506 	if (error == 0 && old != owner)
2507 		return (EINVAL);
2508 	return (error);
2509 }
2510 
2511 /*
2512  * Lock a PP mutex.
2513  */
2514 static int
2515 do_lock_pp(struct thread *td, struct umutex *m, uint32_t flags,
2516     struct _umtx_time *timeout, int try)
2517 {
2518 	struct umtx_abs_timeout timo;
2519 	struct umtx_q *uq, *uq2;
2520 	struct umtx_pi *pi;
2521 	uint32_t ceiling;
2522 	uint32_t owner, id;
2523 	int error, pri, old_inherited_pri, su, rv;
2524 
2525 	id = td->td_tid;
2526 	uq = td->td_umtxq;
2527 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2528 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2529 	    &uq->uq_key)) != 0)
2530 		return (error);
2531 
2532 	if (timeout != NULL)
2533 		umtx_abs_timeout_init2(&timo, timeout);
2534 
2535 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2536 	for (;;) {
2537 		old_inherited_pri = uq->uq_inherited_pri;
2538 		umtxq_lock(&uq->uq_key);
2539 		umtxq_busy(&uq->uq_key);
2540 		umtxq_unlock(&uq->uq_key);
2541 
2542 		rv = fueword32(&m->m_ceilings[0], &ceiling);
2543 		if (rv == -1) {
2544 			error = EFAULT;
2545 			goto out;
2546 		}
2547 		ceiling = RTP_PRIO_MAX - ceiling;
2548 		if (ceiling > RTP_PRIO_MAX) {
2549 			error = EINVAL;
2550 			goto out;
2551 		}
2552 
2553 		mtx_lock(&umtx_lock);
2554 		if (UPRI(td) < PRI_MIN_REALTIME + ceiling) {
2555 			mtx_unlock(&umtx_lock);
2556 			error = EINVAL;
2557 			goto out;
2558 		}
2559 		if (su && PRI_MIN_REALTIME + ceiling < uq->uq_inherited_pri) {
2560 			uq->uq_inherited_pri = PRI_MIN_REALTIME + ceiling;
2561 			thread_lock(td);
2562 			if (uq->uq_inherited_pri < UPRI(td))
2563 				sched_lend_user_prio(td, uq->uq_inherited_pri);
2564 			thread_unlock(td);
2565 		}
2566 		mtx_unlock(&umtx_lock);
2567 
2568 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2569 		    id | UMUTEX_CONTESTED);
2570 		/* The address was invalid. */
2571 		if (rv == -1) {
2572 			error = EFAULT;
2573 			break;
2574 		}
2575 		if (rv == 0) {
2576 			MPASS(owner == UMUTEX_CONTESTED);
2577 			error = 0;
2578 			break;
2579 		}
2580 		/* rv == 1 */
2581 		if (owner == UMUTEX_RB_OWNERDEAD) {
2582 			rv = casueword32(&m->m_owner, UMUTEX_RB_OWNERDEAD,
2583 			    &owner, id | UMUTEX_CONTESTED);
2584 			if (rv == -1) {
2585 				error = EFAULT;
2586 				break;
2587 			}
2588 			if (rv == 0) {
2589 				MPASS(owner == UMUTEX_RB_OWNERDEAD);
2590 				error = EOWNERDEAD; /* success */
2591 				break;
2592 			}
2593 
2594 			/*
2595 			 *  rv == 1, only check for suspension if we
2596 			 *  did not already catched a signal.  If we
2597 			 *  get an error from the check, the same
2598 			 *  condition is checked by the umtxq_sleep()
2599 			 *  call below, so we should obliterate the
2600 			 *  error to not skip the last loop iteration.
2601 			 */
2602 			if (error == 0) {
2603 				error = thread_check_susp(td, false);
2604 				if (error == 0) {
2605 					if (try != 0)
2606 						error = EBUSY;
2607 					else
2608 						continue;
2609 				}
2610 				error = 0;
2611 			}
2612 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2613 			error = ENOTRECOVERABLE;
2614 		}
2615 
2616 		if (try != 0)
2617 			error = EBUSY;
2618 
2619 		/*
2620 		 * If we caught a signal, we have retried and now
2621 		 * exit immediately.
2622 		 */
2623 		if (error != 0)
2624 			break;
2625 
2626 		umtxq_lock(&uq->uq_key);
2627 		umtxq_insert(uq);
2628 		umtxq_unbusy(&uq->uq_key);
2629 		error = umtxq_sleep(uq, "umtxpp", timeout == NULL ?
2630 		    NULL : &timo);
2631 		umtxq_remove(uq);
2632 		umtxq_unlock(&uq->uq_key);
2633 
2634 		mtx_lock(&umtx_lock);
2635 		uq->uq_inherited_pri = old_inherited_pri;
2636 		pri = PRI_MAX;
2637 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2638 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2639 			if (uq2 != NULL) {
2640 				if (pri > UPRI(uq2->uq_thread))
2641 					pri = UPRI(uq2->uq_thread);
2642 			}
2643 		}
2644 		if (pri > uq->uq_inherited_pri)
2645 			pri = uq->uq_inherited_pri;
2646 		thread_lock(td);
2647 		sched_lend_user_prio(td, pri);
2648 		thread_unlock(td);
2649 		mtx_unlock(&umtx_lock);
2650 	}
2651 
2652 	if (error != 0 && error != EOWNERDEAD) {
2653 		mtx_lock(&umtx_lock);
2654 		uq->uq_inherited_pri = old_inherited_pri;
2655 		pri = PRI_MAX;
2656 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2657 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2658 			if (uq2 != NULL) {
2659 				if (pri > UPRI(uq2->uq_thread))
2660 					pri = UPRI(uq2->uq_thread);
2661 			}
2662 		}
2663 		if (pri > uq->uq_inherited_pri)
2664 			pri = uq->uq_inherited_pri;
2665 		thread_lock(td);
2666 		sched_lend_user_prio(td, pri);
2667 		thread_unlock(td);
2668 		mtx_unlock(&umtx_lock);
2669 	}
2670 
2671 out:
2672 	umtxq_unbusy_unlocked(&uq->uq_key);
2673 	umtx_key_release(&uq->uq_key);
2674 	return (error);
2675 }
2676 
2677 /*
2678  * Unlock a PP mutex.
2679  */
2680 static int
2681 do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2682 {
2683 	struct umtx_key key;
2684 	struct umtx_q *uq, *uq2;
2685 	struct umtx_pi *pi;
2686 	uint32_t id, owner, rceiling;
2687 	int error, pri, new_inherited_pri, su;
2688 
2689 	id = td->td_tid;
2690 	uq = td->td_umtxq;
2691 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2692 
2693 	/*
2694 	 * Make sure we own this mtx.
2695 	 */
2696 	error = fueword32(&m->m_owner, &owner);
2697 	if (error == -1)
2698 		return (EFAULT);
2699 
2700 	if ((owner & ~UMUTEX_CONTESTED) != id)
2701 		return (EPERM);
2702 
2703 	error = copyin(&m->m_ceilings[1], &rceiling, sizeof(uint32_t));
2704 	if (error != 0)
2705 		return (error);
2706 
2707 	if (rceiling == -1)
2708 		new_inherited_pri = PRI_MAX;
2709 	else {
2710 		rceiling = RTP_PRIO_MAX - rceiling;
2711 		if (rceiling > RTP_PRIO_MAX)
2712 			return (EINVAL);
2713 		new_inherited_pri = PRI_MIN_REALTIME + rceiling;
2714 	}
2715 
2716 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2717 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2718 	    &key)) != 0)
2719 		return (error);
2720 	umtxq_lock(&key);
2721 	umtxq_busy(&key);
2722 	umtxq_unlock(&key);
2723 	/*
2724 	 * For priority protected mutex, always set unlocked state
2725 	 * to UMUTEX_CONTESTED, so that userland always enters kernel
2726 	 * to lock the mutex, it is necessary because thread priority
2727 	 * has to be adjusted for such mutex.
2728 	 */
2729 	error = suword32(&m->m_owner, umtx_unlock_val(flags, rb) |
2730 	    UMUTEX_CONTESTED);
2731 
2732 	umtxq_lock(&key);
2733 	if (error == 0)
2734 		umtxq_signal(&key, 1);
2735 	umtxq_unbusy(&key);
2736 	umtxq_unlock(&key);
2737 
2738 	if (error == -1)
2739 		error = EFAULT;
2740 	else {
2741 		mtx_lock(&umtx_lock);
2742 		if (su != 0)
2743 			uq->uq_inherited_pri = new_inherited_pri;
2744 		pri = PRI_MAX;
2745 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2746 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2747 			if (uq2 != NULL) {
2748 				if (pri > UPRI(uq2->uq_thread))
2749 					pri = UPRI(uq2->uq_thread);
2750 			}
2751 		}
2752 		if (pri > uq->uq_inherited_pri)
2753 			pri = uq->uq_inherited_pri;
2754 		thread_lock(td);
2755 		sched_lend_user_prio(td, pri);
2756 		thread_unlock(td);
2757 		mtx_unlock(&umtx_lock);
2758 	}
2759 	umtx_key_release(&key);
2760 	return (error);
2761 }
2762 
2763 static int
2764 do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
2765     uint32_t *old_ceiling)
2766 {
2767 	struct umtx_q *uq;
2768 	uint32_t flags, id, owner, save_ceiling;
2769 	int error, rv, rv1;
2770 
2771 	error = fueword32(&m->m_flags, &flags);
2772 	if (error == -1)
2773 		return (EFAULT);
2774 	if ((flags & UMUTEX_PRIO_PROTECT) == 0)
2775 		return (EINVAL);
2776 	if (ceiling > RTP_PRIO_MAX)
2777 		return (EINVAL);
2778 	id = td->td_tid;
2779 	uq = td->td_umtxq;
2780 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2781 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2782 	    &uq->uq_key)) != 0)
2783 		return (error);
2784 	for (;;) {
2785 		umtxq_lock(&uq->uq_key);
2786 		umtxq_busy(&uq->uq_key);
2787 		umtxq_unlock(&uq->uq_key);
2788 
2789 		rv = fueword32(&m->m_ceilings[0], &save_ceiling);
2790 		if (rv == -1) {
2791 			error = EFAULT;
2792 			break;
2793 		}
2794 
2795 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2796 		    id | UMUTEX_CONTESTED);
2797 		if (rv == -1) {
2798 			error = EFAULT;
2799 			break;
2800 		}
2801 
2802 		if (rv == 0) {
2803 			MPASS(owner == UMUTEX_CONTESTED);
2804 			rv = suword32(&m->m_ceilings[0], ceiling);
2805 			rv1 = suword32(&m->m_owner, UMUTEX_CONTESTED);
2806 			error = (rv == 0 && rv1 == 0) ? 0: EFAULT;
2807 			break;
2808 		}
2809 
2810 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2811 			rv = suword32(&m->m_ceilings[0], ceiling);
2812 			error = rv == 0 ? 0 : EFAULT;
2813 			break;
2814 		}
2815 
2816 		if (owner == UMUTEX_RB_OWNERDEAD) {
2817 			error = EOWNERDEAD;
2818 			break;
2819 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2820 			error = ENOTRECOVERABLE;
2821 			break;
2822 		}
2823 
2824 		/*
2825 		 * If we caught a signal, we have retried and now
2826 		 * exit immediately.
2827 		 */
2828 		if (error != 0)
2829 			break;
2830 
2831 		/*
2832 		 * We set the contested bit, sleep. Otherwise the lock changed
2833 		 * and we need to retry or we lost a race to the thread
2834 		 * unlocking the umtx.
2835 		 */
2836 		umtxq_lock(&uq->uq_key);
2837 		umtxq_insert(uq);
2838 		umtxq_unbusy(&uq->uq_key);
2839 		error = umtxq_sleep(uq, "umtxpp", NULL);
2840 		umtxq_remove(uq);
2841 		umtxq_unlock(&uq->uq_key);
2842 	}
2843 	umtxq_lock(&uq->uq_key);
2844 	if (error == 0)
2845 		umtxq_signal(&uq->uq_key, INT_MAX);
2846 	umtxq_unbusy(&uq->uq_key);
2847 	umtxq_unlock(&uq->uq_key);
2848 	umtx_key_release(&uq->uq_key);
2849 	if (error == 0 && old_ceiling != NULL) {
2850 		rv = suword32(old_ceiling, save_ceiling);
2851 		error = rv == 0 ? 0 : EFAULT;
2852 	}
2853 	return (error);
2854 }
2855 
2856 /*
2857  * Lock a userland POSIX mutex.
2858  */
2859 static int
2860 do_lock_umutex(struct thread *td, struct umutex *m,
2861     struct _umtx_time *timeout, int mode)
2862 {
2863 	uint32_t flags;
2864 	int error;
2865 
2866 	error = fueword32(&m->m_flags, &flags);
2867 	if (error == -1)
2868 		return (EFAULT);
2869 
2870 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2871 	case 0:
2872 		error = do_lock_normal(td, m, flags, timeout, mode);
2873 		break;
2874 	case UMUTEX_PRIO_INHERIT:
2875 		error = do_lock_pi(td, m, flags, timeout, mode);
2876 		break;
2877 	case UMUTEX_PRIO_PROTECT:
2878 		error = do_lock_pp(td, m, flags, timeout, mode);
2879 		break;
2880 	default:
2881 		return (EINVAL);
2882 	}
2883 	if (timeout == NULL) {
2884 		if (error == EINTR && mode != _UMUTEX_WAIT)
2885 			error = ERESTART;
2886 	} else {
2887 		/* Timed-locking is not restarted. */
2888 		if (error == ERESTART)
2889 			error = EINTR;
2890 	}
2891 	return (error);
2892 }
2893 
2894 /*
2895  * Unlock a userland POSIX mutex.
2896  */
2897 static int
2898 do_unlock_umutex(struct thread *td, struct umutex *m, bool rb)
2899 {
2900 	uint32_t flags;
2901 	int error;
2902 
2903 	error = fueword32(&m->m_flags, &flags);
2904 	if (error == -1)
2905 		return (EFAULT);
2906 
2907 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2908 	case 0:
2909 		return (do_unlock_normal(td, m, flags, rb));
2910 	case UMUTEX_PRIO_INHERIT:
2911 		return (do_unlock_pi(td, m, flags, rb));
2912 	case UMUTEX_PRIO_PROTECT:
2913 		return (do_unlock_pp(td, m, flags, rb));
2914 	}
2915 
2916 	return (EINVAL);
2917 }
2918 
2919 static int
2920 do_cv_wait(struct thread *td, struct ucond *cv, struct umutex *m,
2921     struct timespec *timeout, u_long wflags)
2922 {
2923 	struct umtx_abs_timeout timo;
2924 	struct umtx_q *uq;
2925 	uint32_t flags, clockid, hasw;
2926 	int error;
2927 
2928 	uq = td->td_umtxq;
2929 	error = fueword32(&cv->c_flags, &flags);
2930 	if (error == -1)
2931 		return (EFAULT);
2932 	error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &uq->uq_key);
2933 	if (error != 0)
2934 		return (error);
2935 
2936 	if ((wflags & CVWAIT_CLOCKID) != 0) {
2937 		error = fueword32(&cv->c_clockid, &clockid);
2938 		if (error == -1) {
2939 			umtx_key_release(&uq->uq_key);
2940 			return (EFAULT);
2941 		}
2942 		if (clockid < CLOCK_REALTIME ||
2943 		    clockid >= CLOCK_THREAD_CPUTIME_ID) {
2944 			/* hmm, only HW clock id will work. */
2945 			umtx_key_release(&uq->uq_key);
2946 			return (EINVAL);
2947 		}
2948 	} else {
2949 		clockid = CLOCK_REALTIME;
2950 	}
2951 
2952 	umtxq_lock(&uq->uq_key);
2953 	umtxq_busy(&uq->uq_key);
2954 	umtxq_insert(uq);
2955 	umtxq_unlock(&uq->uq_key);
2956 
2957 	/*
2958 	 * Set c_has_waiters to 1 before releasing user mutex, also
2959 	 * don't modify cache line when unnecessary.
2960 	 */
2961 	error = fueword32(&cv->c_has_waiters, &hasw);
2962 	if (error == 0 && hasw == 0)
2963 		error = suword32(&cv->c_has_waiters, 1);
2964 	if (error != 0) {
2965 		umtxq_lock(&uq->uq_key);
2966 		umtxq_remove(uq);
2967 		umtxq_unbusy(&uq->uq_key);
2968 		error = EFAULT;
2969 		goto out;
2970 	}
2971 
2972 	umtxq_unbusy_unlocked(&uq->uq_key);
2973 
2974 	error = do_unlock_umutex(td, m, false);
2975 
2976 	if (timeout != NULL)
2977 		umtx_abs_timeout_init(&timo, clockid,
2978 		    (wflags & CVWAIT_ABSTIME) != 0, timeout);
2979 
2980 	umtxq_lock(&uq->uq_key);
2981 	if (error == 0) {
2982 		error = umtxq_sleep(uq, "ucond", timeout == NULL ?
2983 		    NULL : &timo);
2984 	}
2985 
2986 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
2987 		error = 0;
2988 	else {
2989 		/*
2990 		 * This must be timeout,interrupted by signal or
2991 		 * surprious wakeup, clear c_has_waiter flag when
2992 		 * necessary.
2993 		 */
2994 		umtxq_busy(&uq->uq_key);
2995 		if ((uq->uq_flags & UQF_UMTXQ) != 0) {
2996 			int oldlen = uq->uq_cur_queue->length;
2997 			umtxq_remove(uq);
2998 			if (oldlen == 1) {
2999 				umtxq_unlock(&uq->uq_key);
3000 				if (suword32(&cv->c_has_waiters, 0) != 0 &&
3001 				    error == 0)
3002 					error = EFAULT;
3003 				umtxq_lock(&uq->uq_key);
3004 			}
3005 		}
3006 		umtxq_unbusy(&uq->uq_key);
3007 		if (error == ERESTART)
3008 			error = EINTR;
3009 	}
3010 out:
3011 	umtxq_unlock(&uq->uq_key);
3012 	umtx_key_release(&uq->uq_key);
3013 	return (error);
3014 }
3015 
3016 /*
3017  * Signal a userland condition variable.
3018  */
3019 static int
3020 do_cv_signal(struct thread *td, struct ucond *cv)
3021 {
3022 	struct umtx_key key;
3023 	int error, cnt, nwake;
3024 	uint32_t flags;
3025 
3026 	error = fueword32(&cv->c_flags, &flags);
3027 	if (error == -1)
3028 		return (EFAULT);
3029 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
3030 		return (error);
3031 	umtxq_lock(&key);
3032 	umtxq_busy(&key);
3033 	cnt = umtxq_count(&key);
3034 	nwake = umtxq_signal(&key, 1);
3035 	if (cnt <= nwake) {
3036 		umtxq_unlock(&key);
3037 		error = suword32(&cv->c_has_waiters, 0);
3038 		if (error == -1)
3039 			error = EFAULT;
3040 		umtxq_lock(&key);
3041 	}
3042 	umtxq_unbusy(&key);
3043 	umtxq_unlock(&key);
3044 	umtx_key_release(&key);
3045 	return (error);
3046 }
3047 
3048 static int
3049 do_cv_broadcast(struct thread *td, struct ucond *cv)
3050 {
3051 	struct umtx_key key;
3052 	int error;
3053 	uint32_t flags;
3054 
3055 	error = fueword32(&cv->c_flags, &flags);
3056 	if (error == -1)
3057 		return (EFAULT);
3058 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
3059 		return (error);
3060 
3061 	umtxq_lock(&key);
3062 	umtxq_busy(&key);
3063 	umtxq_signal(&key, INT_MAX);
3064 	umtxq_unlock(&key);
3065 
3066 	error = suword32(&cv->c_has_waiters, 0);
3067 	if (error == -1)
3068 		error = EFAULT;
3069 
3070 	umtxq_unbusy_unlocked(&key);
3071 
3072 	umtx_key_release(&key);
3073 	return (error);
3074 }
3075 
3076 static int
3077 do_rw_rdlock(struct thread *td, struct urwlock *rwlock, long fflag,
3078     struct _umtx_time *timeout)
3079 {
3080 	struct umtx_abs_timeout timo;
3081 	struct umtx_q *uq;
3082 	uint32_t flags, wrflags;
3083 	int32_t state, oldstate;
3084 	int32_t blocked_readers;
3085 	int error, error1, rv;
3086 
3087 	uq = td->td_umtxq;
3088 	error = fueword32(&rwlock->rw_flags, &flags);
3089 	if (error == -1)
3090 		return (EFAULT);
3091 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3092 	if (error != 0)
3093 		return (error);
3094 
3095 	if (timeout != NULL)
3096 		umtx_abs_timeout_init2(&timo, timeout);
3097 
3098 	wrflags = URWLOCK_WRITE_OWNER;
3099 	if (!(fflag & URWLOCK_PREFER_READER) && !(flags & URWLOCK_PREFER_READER))
3100 		wrflags |= URWLOCK_WRITE_WAITERS;
3101 
3102 	for (;;) {
3103 		rv = fueword32(&rwlock->rw_state, &state);
3104 		if (rv == -1) {
3105 			umtx_key_release(&uq->uq_key);
3106 			return (EFAULT);
3107 		}
3108 
3109 		/* try to lock it */
3110 		while (!(state & wrflags)) {
3111 			if (__predict_false(URWLOCK_READER_COUNT(state) ==
3112 			    URWLOCK_MAX_READERS)) {
3113 				umtx_key_release(&uq->uq_key);
3114 				return (EAGAIN);
3115 			}
3116 			rv = casueword32(&rwlock->rw_state, state,
3117 			    &oldstate, state + 1);
3118 			if (rv == -1) {
3119 				umtx_key_release(&uq->uq_key);
3120 				return (EFAULT);
3121 			}
3122 			if (rv == 0) {
3123 				MPASS(oldstate == state);
3124 				umtx_key_release(&uq->uq_key);
3125 				return (0);
3126 			}
3127 			error = thread_check_susp(td, true);
3128 			if (error != 0)
3129 				break;
3130 			state = oldstate;
3131 		}
3132 
3133 		if (error)
3134 			break;
3135 
3136 		/* grab monitor lock */
3137 		umtxq_lock(&uq->uq_key);
3138 		umtxq_busy(&uq->uq_key);
3139 		umtxq_unlock(&uq->uq_key);
3140 
3141 		/*
3142 		 * re-read the state, in case it changed between the try-lock above
3143 		 * and the check below
3144 		 */
3145 		rv = fueword32(&rwlock->rw_state, &state);
3146 		if (rv == -1)
3147 			error = EFAULT;
3148 
3149 		/* set read contention bit */
3150 		while (error == 0 && (state & wrflags) &&
3151 		    !(state & URWLOCK_READ_WAITERS)) {
3152 			rv = casueword32(&rwlock->rw_state, state,
3153 			    &oldstate, state | URWLOCK_READ_WAITERS);
3154 			if (rv == -1) {
3155 				error = EFAULT;
3156 				break;
3157 			}
3158 			if (rv == 0) {
3159 				MPASS(oldstate == state);
3160 				goto sleep;
3161 			}
3162 			state = oldstate;
3163 			error = thread_check_susp(td, false);
3164 			if (error != 0)
3165 				break;
3166 		}
3167 		if (error != 0) {
3168 			umtxq_unbusy_unlocked(&uq->uq_key);
3169 			break;
3170 		}
3171 
3172 		/* state is changed while setting flags, restart */
3173 		if (!(state & wrflags)) {
3174 			umtxq_unbusy_unlocked(&uq->uq_key);
3175 			error = thread_check_susp(td, true);
3176 			if (error != 0)
3177 				break;
3178 			continue;
3179 		}
3180 
3181 sleep:
3182 		/*
3183 		 * Contention bit is set, before sleeping, increase
3184 		 * read waiter count.
3185 		 */
3186 		rv = fueword32(&rwlock->rw_blocked_readers,
3187 		    &blocked_readers);
3188 		if (rv == 0)
3189 			rv = suword32(&rwlock->rw_blocked_readers,
3190 			    blocked_readers + 1);
3191 		if (rv == -1) {
3192 			umtxq_unbusy_unlocked(&uq->uq_key);
3193 			error = EFAULT;
3194 			break;
3195 		}
3196 
3197 		while (state & wrflags) {
3198 			umtxq_lock(&uq->uq_key);
3199 			umtxq_insert(uq);
3200 			umtxq_unbusy(&uq->uq_key);
3201 
3202 			error = umtxq_sleep(uq, "urdlck", timeout == NULL ?
3203 			    NULL : &timo);
3204 
3205 			umtxq_busy(&uq->uq_key);
3206 			umtxq_remove(uq);
3207 			umtxq_unlock(&uq->uq_key);
3208 			if (error)
3209 				break;
3210 			rv = fueword32(&rwlock->rw_state, &state);
3211 			if (rv == -1) {
3212 				error = EFAULT;
3213 				break;
3214 			}
3215 		}
3216 
3217 		/* decrease read waiter count, and may clear read contention bit */
3218 		rv = fueword32(&rwlock->rw_blocked_readers,
3219 		    &blocked_readers);
3220 		if (rv == 0)
3221 			rv = suword32(&rwlock->rw_blocked_readers,
3222 			    blocked_readers - 1);
3223 		if (rv == -1) {
3224 			umtxq_unbusy_unlocked(&uq->uq_key);
3225 			error = EFAULT;
3226 			break;
3227 		}
3228 		if (blocked_readers == 1) {
3229 			rv = fueword32(&rwlock->rw_state, &state);
3230 			if (rv == -1) {
3231 				umtxq_unbusy_unlocked(&uq->uq_key);
3232 				error = EFAULT;
3233 				break;
3234 			}
3235 			for (;;) {
3236 				rv = casueword32(&rwlock->rw_state, state,
3237 				    &oldstate, state & ~URWLOCK_READ_WAITERS);
3238 				if (rv == -1) {
3239 					error = EFAULT;
3240 					break;
3241 				}
3242 				if (rv == 0) {
3243 					MPASS(oldstate == state);
3244 					break;
3245 				}
3246 				state = oldstate;
3247 				error1 = thread_check_susp(td, false);
3248 				if (error1 != 0) {
3249 					if (error == 0)
3250 						error = error1;
3251 					break;
3252 				}
3253 			}
3254 		}
3255 
3256 		umtxq_unbusy_unlocked(&uq->uq_key);
3257 		if (error != 0)
3258 			break;
3259 	}
3260 	umtx_key_release(&uq->uq_key);
3261 	if (error == ERESTART)
3262 		error = EINTR;
3263 	return (error);
3264 }
3265 
3266 static int
3267 do_rw_wrlock(struct thread *td, struct urwlock *rwlock, struct _umtx_time *timeout)
3268 {
3269 	struct umtx_abs_timeout timo;
3270 	struct umtx_q *uq;
3271 	uint32_t flags;
3272 	int32_t state, oldstate;
3273 	int32_t blocked_writers;
3274 	int32_t blocked_readers;
3275 	int error, error1, rv;
3276 
3277 	uq = td->td_umtxq;
3278 	error = fueword32(&rwlock->rw_flags, &flags);
3279 	if (error == -1)
3280 		return (EFAULT);
3281 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3282 	if (error != 0)
3283 		return (error);
3284 
3285 	if (timeout != NULL)
3286 		umtx_abs_timeout_init2(&timo, timeout);
3287 
3288 	blocked_readers = 0;
3289 	for (;;) {
3290 		rv = fueword32(&rwlock->rw_state, &state);
3291 		if (rv == -1) {
3292 			umtx_key_release(&uq->uq_key);
3293 			return (EFAULT);
3294 		}
3295 		while ((state & URWLOCK_WRITE_OWNER) == 0 &&
3296 		    URWLOCK_READER_COUNT(state) == 0) {
3297 			rv = casueword32(&rwlock->rw_state, state,
3298 			    &oldstate, state | URWLOCK_WRITE_OWNER);
3299 			if (rv == -1) {
3300 				umtx_key_release(&uq->uq_key);
3301 				return (EFAULT);
3302 			}
3303 			if (rv == 0) {
3304 				MPASS(oldstate == state);
3305 				umtx_key_release(&uq->uq_key);
3306 				return (0);
3307 			}
3308 			state = oldstate;
3309 			error = thread_check_susp(td, true);
3310 			if (error != 0)
3311 				break;
3312 		}
3313 
3314 		if (error) {
3315 			if ((state & (URWLOCK_WRITE_OWNER |
3316 			    URWLOCK_WRITE_WAITERS)) == 0 &&
3317 			    blocked_readers != 0) {
3318 				umtxq_lock(&uq->uq_key);
3319 				umtxq_busy(&uq->uq_key);
3320 				umtxq_signal_queue(&uq->uq_key, INT_MAX,
3321 				    UMTX_SHARED_QUEUE);
3322 				umtxq_unbusy(&uq->uq_key);
3323 				umtxq_unlock(&uq->uq_key);
3324 			}
3325 
3326 			break;
3327 		}
3328 
3329 		/* grab monitor lock */
3330 		umtxq_lock(&uq->uq_key);
3331 		umtxq_busy(&uq->uq_key);
3332 		umtxq_unlock(&uq->uq_key);
3333 
3334 		/*
3335 		 * Re-read the state, in case it changed between the
3336 		 * try-lock above and the check below.
3337 		 */
3338 		rv = fueword32(&rwlock->rw_state, &state);
3339 		if (rv == -1)
3340 			error = EFAULT;
3341 
3342 		while (error == 0 && ((state & URWLOCK_WRITE_OWNER) ||
3343 		    URWLOCK_READER_COUNT(state) != 0) &&
3344 		    (state & URWLOCK_WRITE_WAITERS) == 0) {
3345 			rv = casueword32(&rwlock->rw_state, state,
3346 			    &oldstate, state | URWLOCK_WRITE_WAITERS);
3347 			if (rv == -1) {
3348 				error = EFAULT;
3349 				break;
3350 			}
3351 			if (rv == 0) {
3352 				MPASS(oldstate == state);
3353 				goto sleep;
3354 			}
3355 			state = oldstate;
3356 			error = thread_check_susp(td, false);
3357 			if (error != 0)
3358 				break;
3359 		}
3360 		if (error != 0) {
3361 			umtxq_unbusy_unlocked(&uq->uq_key);
3362 			break;
3363 		}
3364 
3365 		if ((state & URWLOCK_WRITE_OWNER) == 0 &&
3366 		    URWLOCK_READER_COUNT(state) == 0) {
3367 			umtxq_unbusy_unlocked(&uq->uq_key);
3368 			error = thread_check_susp(td, false);
3369 			if (error != 0)
3370 				break;
3371 			continue;
3372 		}
3373 sleep:
3374 		rv = fueword32(&rwlock->rw_blocked_writers,
3375 		    &blocked_writers);
3376 		if (rv == 0)
3377 			rv = suword32(&rwlock->rw_blocked_writers,
3378 			    blocked_writers + 1);
3379 		if (rv == -1) {
3380 			umtxq_unbusy_unlocked(&uq->uq_key);
3381 			error = EFAULT;
3382 			break;
3383 		}
3384 
3385 		while ((state & URWLOCK_WRITE_OWNER) ||
3386 		    URWLOCK_READER_COUNT(state) != 0) {
3387 			umtxq_lock(&uq->uq_key);
3388 			umtxq_insert_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3389 			umtxq_unbusy(&uq->uq_key);
3390 
3391 			error = umtxq_sleep(uq, "uwrlck", timeout == NULL ?
3392 			    NULL : &timo);
3393 
3394 			umtxq_busy(&uq->uq_key);
3395 			umtxq_remove_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3396 			umtxq_unlock(&uq->uq_key);
3397 			if (error)
3398 				break;
3399 			rv = fueword32(&rwlock->rw_state, &state);
3400 			if (rv == -1) {
3401 				error = EFAULT;
3402 				break;
3403 			}
3404 		}
3405 
3406 		rv = fueword32(&rwlock->rw_blocked_writers,
3407 		    &blocked_writers);
3408 		if (rv == 0)
3409 			rv = suword32(&rwlock->rw_blocked_writers,
3410 			    blocked_writers - 1);
3411 		if (rv == -1) {
3412 			umtxq_unbusy_unlocked(&uq->uq_key);
3413 			error = EFAULT;
3414 			break;
3415 		}
3416 		if (blocked_writers == 1) {
3417 			rv = fueword32(&rwlock->rw_state, &state);
3418 			if (rv == -1) {
3419 				umtxq_unbusy_unlocked(&uq->uq_key);
3420 				error = EFAULT;
3421 				break;
3422 			}
3423 			for (;;) {
3424 				rv = casueword32(&rwlock->rw_state, state,
3425 				    &oldstate, state & ~URWLOCK_WRITE_WAITERS);
3426 				if (rv == -1) {
3427 					error = EFAULT;
3428 					break;
3429 				}
3430 				if (rv == 0) {
3431 					MPASS(oldstate == state);
3432 					break;
3433 				}
3434 				state = oldstate;
3435 				error1 = thread_check_susp(td, false);
3436 				/*
3437 				 * We are leaving the URWLOCK_WRITE_WAITERS
3438 				 * behind, but this should not harm the
3439 				 * correctness.
3440 				 */
3441 				if (error1 != 0) {
3442 					if (error == 0)
3443 						error = error1;
3444 					break;
3445 				}
3446 			}
3447 			rv = fueword32(&rwlock->rw_blocked_readers,
3448 			    &blocked_readers);
3449 			if (rv == -1) {
3450 				umtxq_unbusy_unlocked(&uq->uq_key);
3451 				error = EFAULT;
3452 				break;
3453 			}
3454 		} else
3455 			blocked_readers = 0;
3456 
3457 		umtxq_unbusy_unlocked(&uq->uq_key);
3458 	}
3459 
3460 	umtx_key_release(&uq->uq_key);
3461 	if (error == ERESTART)
3462 		error = EINTR;
3463 	return (error);
3464 }
3465 
3466 static int
3467 do_rw_unlock(struct thread *td, struct urwlock *rwlock)
3468 {
3469 	struct umtx_q *uq;
3470 	uint32_t flags;
3471 	int32_t state, oldstate;
3472 	int error, rv, q, count;
3473 
3474 	uq = td->td_umtxq;
3475 	error = fueword32(&rwlock->rw_flags, &flags);
3476 	if (error == -1)
3477 		return (EFAULT);
3478 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3479 	if (error != 0)
3480 		return (error);
3481 
3482 	error = fueword32(&rwlock->rw_state, &state);
3483 	if (error == -1) {
3484 		error = EFAULT;
3485 		goto out;
3486 	}
3487 	if (state & URWLOCK_WRITE_OWNER) {
3488 		for (;;) {
3489 			rv = casueword32(&rwlock->rw_state, state,
3490 			    &oldstate, state & ~URWLOCK_WRITE_OWNER);
3491 			if (rv == -1) {
3492 				error = EFAULT;
3493 				goto out;
3494 			}
3495 			if (rv == 1) {
3496 				state = oldstate;
3497 				if (!(oldstate & URWLOCK_WRITE_OWNER)) {
3498 					error = EPERM;
3499 					goto out;
3500 				}
3501 				error = thread_check_susp(td, true);
3502 				if (error != 0)
3503 					goto out;
3504 			} else
3505 				break;
3506 		}
3507 	} else if (URWLOCK_READER_COUNT(state) != 0) {
3508 		for (;;) {
3509 			rv = casueword32(&rwlock->rw_state, state,
3510 			    &oldstate, state - 1);
3511 			if (rv == -1) {
3512 				error = EFAULT;
3513 				goto out;
3514 			}
3515 			if (rv == 1) {
3516 				state = oldstate;
3517 				if (URWLOCK_READER_COUNT(oldstate) == 0) {
3518 					error = EPERM;
3519 					goto out;
3520 				}
3521 				error = thread_check_susp(td, true);
3522 				if (error != 0)
3523 					goto out;
3524 			} else
3525 				break;
3526 		}
3527 	} else {
3528 		error = EPERM;
3529 		goto out;
3530 	}
3531 
3532 	count = 0;
3533 
3534 	if (!(flags & URWLOCK_PREFER_READER)) {
3535 		if (state & URWLOCK_WRITE_WAITERS) {
3536 			count = 1;
3537 			q = UMTX_EXCLUSIVE_QUEUE;
3538 		} else if (state & URWLOCK_READ_WAITERS) {
3539 			count = INT_MAX;
3540 			q = UMTX_SHARED_QUEUE;
3541 		}
3542 	} else {
3543 		if (state & URWLOCK_READ_WAITERS) {
3544 			count = INT_MAX;
3545 			q = UMTX_SHARED_QUEUE;
3546 		} else if (state & URWLOCK_WRITE_WAITERS) {
3547 			count = 1;
3548 			q = UMTX_EXCLUSIVE_QUEUE;
3549 		}
3550 	}
3551 
3552 	if (count) {
3553 		umtxq_lock(&uq->uq_key);
3554 		umtxq_busy(&uq->uq_key);
3555 		umtxq_signal_queue(&uq->uq_key, count, q);
3556 		umtxq_unbusy(&uq->uq_key);
3557 		umtxq_unlock(&uq->uq_key);
3558 	}
3559 out:
3560 	umtx_key_release(&uq->uq_key);
3561 	return (error);
3562 }
3563 
3564 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3565 static int
3566 do_sem_wait(struct thread *td, struct _usem *sem, struct _umtx_time *timeout)
3567 {
3568 	struct umtx_abs_timeout timo;
3569 	struct umtx_q *uq;
3570 	uint32_t flags, count, count1;
3571 	int error, rv, rv1;
3572 
3573 	uq = td->td_umtxq;
3574 	error = fueword32(&sem->_flags, &flags);
3575 	if (error == -1)
3576 		return (EFAULT);
3577 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3578 	if (error != 0)
3579 		return (error);
3580 
3581 	if (timeout != NULL)
3582 		umtx_abs_timeout_init2(&timo, timeout);
3583 
3584 again:
3585 	umtxq_lock(&uq->uq_key);
3586 	umtxq_busy(&uq->uq_key);
3587 	umtxq_insert(uq);
3588 	umtxq_unlock(&uq->uq_key);
3589 	rv = casueword32(&sem->_has_waiters, 0, &count1, 1);
3590 	if (rv != -1)
3591 		rv1 = fueword32(&sem->_count, &count);
3592 	if (rv == -1 || rv1 == -1 || count != 0 || (rv == 1 && count1 == 0)) {
3593 		if (rv == 0)
3594 			rv = suword32(&sem->_has_waiters, 0);
3595 		umtxq_lock(&uq->uq_key);
3596 		umtxq_unbusy(&uq->uq_key);
3597 		umtxq_remove(uq);
3598 		umtxq_unlock(&uq->uq_key);
3599 		if (rv == -1 || rv1 == -1) {
3600 			error = EFAULT;
3601 			goto out;
3602 		}
3603 		if (count != 0) {
3604 			error = 0;
3605 			goto out;
3606 		}
3607 		MPASS(rv == 1 && count1 == 0);
3608 		rv = thread_check_susp(td, true);
3609 		if (rv == 0)
3610 			goto again;
3611 		error = rv;
3612 		goto out;
3613 	}
3614 	umtxq_lock(&uq->uq_key);
3615 	umtxq_unbusy(&uq->uq_key);
3616 
3617 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3618 
3619 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3620 		error = 0;
3621 	else {
3622 		umtxq_remove(uq);
3623 		/* A relative timeout cannot be restarted. */
3624 		if (error == ERESTART && timeout != NULL &&
3625 		    (timeout->_flags & UMTX_ABSTIME) == 0)
3626 			error = EINTR;
3627 	}
3628 	umtxq_unlock(&uq->uq_key);
3629 out:
3630 	umtx_key_release(&uq->uq_key);
3631 	return (error);
3632 }
3633 
3634 /*
3635  * Signal a userland semaphore.
3636  */
3637 static int
3638 do_sem_wake(struct thread *td, struct _usem *sem)
3639 {
3640 	struct umtx_key key;
3641 	int error, cnt;
3642 	uint32_t flags;
3643 
3644 	error = fueword32(&sem->_flags, &flags);
3645 	if (error == -1)
3646 		return (EFAULT);
3647 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3648 		return (error);
3649 	umtxq_lock(&key);
3650 	umtxq_busy(&key);
3651 	cnt = umtxq_count(&key);
3652 	if (cnt > 0) {
3653 		/*
3654 		 * Check if count is greater than 0, this means the memory is
3655 		 * still being referenced by user code, so we can safely
3656 		 * update _has_waiters flag.
3657 		 */
3658 		if (cnt == 1) {
3659 			umtxq_unlock(&key);
3660 			error = suword32(&sem->_has_waiters, 0);
3661 			umtxq_lock(&key);
3662 			if (error == -1)
3663 				error = EFAULT;
3664 		}
3665 		umtxq_signal(&key, 1);
3666 	}
3667 	umtxq_unbusy(&key);
3668 	umtxq_unlock(&key);
3669 	umtx_key_release(&key);
3670 	return (error);
3671 }
3672 #endif
3673 
3674 static int
3675 do_sem2_wait(struct thread *td, struct _usem2 *sem, struct _umtx_time *timeout)
3676 {
3677 	struct umtx_abs_timeout timo;
3678 	struct umtx_q *uq;
3679 	uint32_t count, flags;
3680 	int error, rv;
3681 
3682 	uq = td->td_umtxq;
3683 	flags = fuword32(&sem->_flags);
3684 	if (timeout != NULL)
3685 		umtx_abs_timeout_init2(&timo, timeout);
3686 
3687 again:
3688 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3689 	if (error != 0)
3690 		return (error);
3691 	umtxq_lock(&uq->uq_key);
3692 	umtxq_busy(&uq->uq_key);
3693 	umtxq_insert(uq);
3694 	umtxq_unlock(&uq->uq_key);
3695 	rv = fueword32(&sem->_count, &count);
3696 	if (rv == -1) {
3697 		umtxq_lock(&uq->uq_key);
3698 		umtxq_unbusy(&uq->uq_key);
3699 		umtxq_remove(uq);
3700 		umtxq_unlock(&uq->uq_key);
3701 		umtx_key_release(&uq->uq_key);
3702 		return (EFAULT);
3703 	}
3704 	for (;;) {
3705 		if (USEM_COUNT(count) != 0) {
3706 			umtxq_lock(&uq->uq_key);
3707 			umtxq_unbusy(&uq->uq_key);
3708 			umtxq_remove(uq);
3709 			umtxq_unlock(&uq->uq_key);
3710 			umtx_key_release(&uq->uq_key);
3711 			return (0);
3712 		}
3713 		if (count == USEM_HAS_WAITERS)
3714 			break;
3715 		rv = casueword32(&sem->_count, 0, &count, USEM_HAS_WAITERS);
3716 		if (rv == 0)
3717 			break;
3718 		umtxq_lock(&uq->uq_key);
3719 		umtxq_unbusy(&uq->uq_key);
3720 		umtxq_remove(uq);
3721 		umtxq_unlock(&uq->uq_key);
3722 		umtx_key_release(&uq->uq_key);
3723 		if (rv == -1)
3724 			return (EFAULT);
3725 		rv = thread_check_susp(td, true);
3726 		if (rv != 0)
3727 			return (rv);
3728 		goto again;
3729 	}
3730 	umtxq_lock(&uq->uq_key);
3731 	umtxq_unbusy(&uq->uq_key);
3732 
3733 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3734 
3735 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3736 		error = 0;
3737 	else {
3738 		umtxq_remove(uq);
3739 		if (timeout != NULL && (timeout->_flags & UMTX_ABSTIME) == 0) {
3740 			/* A relative timeout cannot be restarted. */
3741 			if (error == ERESTART)
3742 				error = EINTR;
3743 			if (error == EINTR) {
3744 				kern_clock_gettime(curthread, timo.clockid,
3745 				    &timo.cur);
3746 				timespecsub(&timo.end, &timo.cur,
3747 				    &timeout->_timeout);
3748 			}
3749 		}
3750 	}
3751 	umtxq_unlock(&uq->uq_key);
3752 	umtx_key_release(&uq->uq_key);
3753 	return (error);
3754 }
3755 
3756 /*
3757  * Signal a userland semaphore.
3758  */
3759 static int
3760 do_sem2_wake(struct thread *td, struct _usem2 *sem)
3761 {
3762 	struct umtx_key key;
3763 	int error, cnt, rv;
3764 	uint32_t count, flags;
3765 
3766 	rv = fueword32(&sem->_flags, &flags);
3767 	if (rv == -1)
3768 		return (EFAULT);
3769 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3770 		return (error);
3771 	umtxq_lock(&key);
3772 	umtxq_busy(&key);
3773 	cnt = umtxq_count(&key);
3774 	if (cnt > 0) {
3775 		/*
3776 		 * If this was the last sleeping thread, clear the waiters
3777 		 * flag in _count.
3778 		 */
3779 		if (cnt == 1) {
3780 			umtxq_unlock(&key);
3781 			rv = fueword32(&sem->_count, &count);
3782 			while (rv != -1 && count & USEM_HAS_WAITERS) {
3783 				rv = casueword32(&sem->_count, count, &count,
3784 				    count & ~USEM_HAS_WAITERS);
3785 				if (rv == 1) {
3786 					rv = thread_check_susp(td, true);
3787 					if (rv != 0)
3788 						break;
3789 				}
3790 			}
3791 			if (rv == -1)
3792 				error = EFAULT;
3793 			else if (rv > 0) {
3794 				error = rv;
3795 			}
3796 			umtxq_lock(&key);
3797 		}
3798 
3799 		umtxq_signal(&key, 1);
3800 	}
3801 	umtxq_unbusy(&key);
3802 	umtxq_unlock(&key);
3803 	umtx_key_release(&key);
3804 	return (error);
3805 }
3806 
3807 #ifdef COMPAT_FREEBSD10
3808 int
3809 freebsd10__umtx_lock(struct thread *td, struct freebsd10__umtx_lock_args *uap)
3810 {
3811 	return (do_lock_umtx(td, uap->umtx, td->td_tid, 0));
3812 }
3813 
3814 int
3815 freebsd10__umtx_unlock(struct thread *td,
3816     struct freebsd10__umtx_unlock_args *uap)
3817 {
3818 	return (do_unlock_umtx(td, uap->umtx, td->td_tid));
3819 }
3820 #endif
3821 
3822 inline int
3823 umtx_copyin_timeout(const void *uaddr, struct timespec *tsp)
3824 {
3825 	int error;
3826 
3827 	error = copyin(uaddr, tsp, sizeof(*tsp));
3828 	if (error == 0) {
3829 		if (!timespecvalid_interval(tsp))
3830 			error = EINVAL;
3831 	}
3832 	return (error);
3833 }
3834 
3835 static inline int
3836 umtx_copyin_umtx_time(const void *uaddr, size_t size, struct _umtx_time *tp)
3837 {
3838 	int error;
3839 
3840 	if (size <= sizeof(tp->_timeout)) {
3841 		tp->_clockid = CLOCK_REALTIME;
3842 		tp->_flags = 0;
3843 		error = copyin(uaddr, &tp->_timeout, sizeof(tp->_timeout));
3844 	} else
3845 		error = copyin(uaddr, tp, sizeof(*tp));
3846 	if (error != 0)
3847 		return (error);
3848 	if (!timespecvalid_interval(&tp->_timeout))
3849 		return (EINVAL);
3850 	return (0);
3851 }
3852 
3853 static int
3854 umtx_copyin_robust_lists(const void *uaddr, size_t size,
3855     struct umtx_robust_lists_params *rb)
3856 {
3857 
3858 	if (size > sizeof(*rb))
3859 		return (EINVAL);
3860 	return (copyin(uaddr, rb, size));
3861 }
3862 
3863 static int
3864 umtx_copyout_timeout(void *uaddr, size_t sz, struct timespec *tsp)
3865 {
3866 
3867 	/*
3868 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
3869 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
3870 	 * copyops.
3871 	 */
3872 	KASSERT(sz >= sizeof(*tsp),
3873 	    ("umtx_copyops specifies incorrect sizes"));
3874 
3875 	return (copyout(tsp, uaddr, sizeof(*tsp)));
3876 }
3877 
3878 #ifdef COMPAT_FREEBSD10
3879 static int
3880 __umtx_op_lock_umtx(struct thread *td, struct _umtx_op_args *uap,
3881     const struct umtx_copyops *ops)
3882 {
3883 	struct timespec *ts, timeout;
3884 	int error;
3885 
3886 	/* Allow a null timespec (wait forever). */
3887 	if (uap->uaddr2 == NULL)
3888 		ts = NULL;
3889 	else {
3890 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
3891 		if (error != 0)
3892 			return (error);
3893 		ts = &timeout;
3894 	}
3895 #ifdef COMPAT_FREEBSD32
3896 	if (ops->compat32)
3897 		return (do_lock_umtx32(td, uap->obj, uap->val, ts));
3898 #endif
3899 	return (do_lock_umtx(td, uap->obj, uap->val, ts));
3900 }
3901 
3902 static int
3903 __umtx_op_unlock_umtx(struct thread *td, struct _umtx_op_args *uap,
3904     const struct umtx_copyops *ops)
3905 {
3906 #ifdef COMPAT_FREEBSD32
3907 	if (ops->compat32)
3908 		return (do_unlock_umtx32(td, uap->obj, uap->val));
3909 #endif
3910 	return (do_unlock_umtx(td, uap->obj, uap->val));
3911 }
3912 #endif	/* COMPAT_FREEBSD10 */
3913 
3914 #if !defined(COMPAT_FREEBSD10)
3915 static int
3916 __umtx_op_unimpl(struct thread *td __unused, struct _umtx_op_args *uap __unused,
3917     const struct umtx_copyops *ops __unused)
3918 {
3919 	return (EOPNOTSUPP);
3920 }
3921 #endif	/* COMPAT_FREEBSD10 */
3922 
3923 static int
3924 __umtx_op_wait(struct thread *td, struct _umtx_op_args *uap,
3925     const struct umtx_copyops *ops)
3926 {
3927 	struct _umtx_time timeout, *tm_p;
3928 	int error;
3929 
3930 	if (uap->uaddr2 == NULL)
3931 		tm_p = NULL;
3932 	else {
3933 		error = ops->copyin_umtx_time(
3934 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3935 		if (error != 0)
3936 			return (error);
3937 		tm_p = &timeout;
3938 	}
3939 	return (do_wait(td, uap->obj, uap->val, tm_p, ops->compat32, 0));
3940 }
3941 
3942 static int
3943 __umtx_op_wait_uint(struct thread *td, struct _umtx_op_args *uap,
3944     const struct umtx_copyops *ops)
3945 {
3946 	struct _umtx_time timeout, *tm_p;
3947 	int error;
3948 
3949 	if (uap->uaddr2 == NULL)
3950 		tm_p = NULL;
3951 	else {
3952 		error = ops->copyin_umtx_time(
3953 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3954 		if (error != 0)
3955 			return (error);
3956 		tm_p = &timeout;
3957 	}
3958 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
3959 }
3960 
3961 static int
3962 __umtx_op_wait_uint_private(struct thread *td, struct _umtx_op_args *uap,
3963     const struct umtx_copyops *ops)
3964 {
3965 	struct _umtx_time *tm_p, timeout;
3966 	int error;
3967 
3968 	if (uap->uaddr2 == NULL)
3969 		tm_p = NULL;
3970 	else {
3971 		error = ops->copyin_umtx_time(
3972 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3973 		if (error != 0)
3974 			return (error);
3975 		tm_p = &timeout;
3976 	}
3977 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
3978 }
3979 
3980 static int
3981 __umtx_op_wake(struct thread *td, struct _umtx_op_args *uap,
3982     const struct umtx_copyops *ops __unused)
3983 {
3984 
3985 	return (kern_umtx_wake(td, uap->obj, uap->val, 0));
3986 }
3987 
3988 #define BATCH_SIZE	128
3989 static int
3990 __umtx_op_nwake_private_native(struct thread *td, struct _umtx_op_args *uap)
3991 {
3992 	char *uaddrs[BATCH_SIZE], **upp;
3993 	int count, error, i, pos, tocopy;
3994 
3995 	upp = (char **)uap->obj;
3996 	error = 0;
3997 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3998 	    pos += tocopy) {
3999 		tocopy = MIN(count, BATCH_SIZE);
4000 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(char *));
4001 		if (error != 0)
4002 			break;
4003 		for (i = 0; i < tocopy; ++i) {
4004 			kern_umtx_wake(td, uaddrs[i], INT_MAX, 1);
4005 		}
4006 		maybe_yield();
4007 	}
4008 	return (error);
4009 }
4010 
4011 static int
4012 __umtx_op_nwake_private_compat32(struct thread *td, struct _umtx_op_args *uap)
4013 {
4014 	uint32_t uaddrs[BATCH_SIZE], *upp;
4015 	int count, error, i, pos, tocopy;
4016 
4017 	upp = (uint32_t *)uap->obj;
4018 	error = 0;
4019 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
4020 	    pos += tocopy) {
4021 		tocopy = MIN(count, BATCH_SIZE);
4022 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(uint32_t));
4023 		if (error != 0)
4024 			break;
4025 		for (i = 0; i < tocopy; ++i) {
4026 			kern_umtx_wake(td, (void *)(uintptr_t)uaddrs[i],
4027 			    INT_MAX, 1);
4028 		}
4029 		maybe_yield();
4030 	}
4031 	return (error);
4032 }
4033 
4034 static int
4035 __umtx_op_nwake_private(struct thread *td, struct _umtx_op_args *uap,
4036     const struct umtx_copyops *ops)
4037 {
4038 
4039 	if (ops->compat32)
4040 		return (__umtx_op_nwake_private_compat32(td, uap));
4041 	return (__umtx_op_nwake_private_native(td, uap));
4042 }
4043 
4044 static int
4045 __umtx_op_wake_private(struct thread *td, struct _umtx_op_args *uap,
4046     const struct umtx_copyops *ops __unused)
4047 {
4048 
4049 	return (kern_umtx_wake(td, uap->obj, uap->val, 1));
4050 }
4051 
4052 static int
4053 __umtx_op_lock_umutex(struct thread *td, struct _umtx_op_args *uap,
4054    const struct umtx_copyops *ops)
4055 {
4056 	struct _umtx_time *tm_p, timeout;
4057 	int error;
4058 
4059 	/* Allow a null timespec (wait forever). */
4060 	if (uap->uaddr2 == NULL)
4061 		tm_p = NULL;
4062 	else {
4063 		error = ops->copyin_umtx_time(
4064 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4065 		if (error != 0)
4066 			return (error);
4067 		tm_p = &timeout;
4068 	}
4069 	return (do_lock_umutex(td, uap->obj, tm_p, 0));
4070 }
4071 
4072 static int
4073 __umtx_op_trylock_umutex(struct thread *td, struct _umtx_op_args *uap,
4074     const struct umtx_copyops *ops __unused)
4075 {
4076 
4077 	return (do_lock_umutex(td, uap->obj, NULL, _UMUTEX_TRY));
4078 }
4079 
4080 static int
4081 __umtx_op_wait_umutex(struct thread *td, struct _umtx_op_args *uap,
4082     const struct umtx_copyops *ops)
4083 {
4084 	struct _umtx_time *tm_p, timeout;
4085 	int error;
4086 
4087 	/* Allow a null timespec (wait forever). */
4088 	if (uap->uaddr2 == NULL)
4089 		tm_p = NULL;
4090 	else {
4091 		error = ops->copyin_umtx_time(
4092 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4093 		if (error != 0)
4094 			return (error);
4095 		tm_p = &timeout;
4096 	}
4097 	return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
4098 }
4099 
4100 static int
4101 __umtx_op_wake_umutex(struct thread *td, struct _umtx_op_args *uap,
4102     const struct umtx_copyops *ops __unused)
4103 {
4104 
4105 	return (do_wake_umutex(td, uap->obj));
4106 }
4107 
4108 static int
4109 __umtx_op_unlock_umutex(struct thread *td, struct _umtx_op_args *uap,
4110     const struct umtx_copyops *ops __unused)
4111 {
4112 
4113 	return (do_unlock_umutex(td, uap->obj, false));
4114 }
4115 
4116 static int
4117 __umtx_op_set_ceiling(struct thread *td, struct _umtx_op_args *uap,
4118     const struct umtx_copyops *ops __unused)
4119 {
4120 
4121 	return (do_set_ceiling(td, uap->obj, uap->val, uap->uaddr1));
4122 }
4123 
4124 static int
4125 __umtx_op_cv_wait(struct thread *td, struct _umtx_op_args *uap,
4126     const struct umtx_copyops *ops)
4127 {
4128 	struct timespec *ts, timeout;
4129 	int error;
4130 
4131 	/* Allow a null timespec (wait forever). */
4132 	if (uap->uaddr2 == NULL)
4133 		ts = NULL;
4134 	else {
4135 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
4136 		if (error != 0)
4137 			return (error);
4138 		ts = &timeout;
4139 	}
4140 	return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
4141 }
4142 
4143 static int
4144 __umtx_op_cv_signal(struct thread *td, struct _umtx_op_args *uap,
4145     const struct umtx_copyops *ops __unused)
4146 {
4147 
4148 	return (do_cv_signal(td, uap->obj));
4149 }
4150 
4151 static int
4152 __umtx_op_cv_broadcast(struct thread *td, struct _umtx_op_args *uap,
4153     const struct umtx_copyops *ops __unused)
4154 {
4155 
4156 	return (do_cv_broadcast(td, uap->obj));
4157 }
4158 
4159 static int
4160 __umtx_op_rw_rdlock(struct thread *td, struct _umtx_op_args *uap,
4161     const struct umtx_copyops *ops)
4162 {
4163 	struct _umtx_time timeout;
4164 	int error;
4165 
4166 	/* Allow a null timespec (wait forever). */
4167 	if (uap->uaddr2 == NULL) {
4168 		error = do_rw_rdlock(td, uap->obj, uap->val, 0);
4169 	} else {
4170 		error = ops->copyin_umtx_time(uap->uaddr2,
4171 		   (size_t)uap->uaddr1, &timeout);
4172 		if (error != 0)
4173 			return (error);
4174 		error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
4175 	}
4176 	return (error);
4177 }
4178 
4179 static int
4180 __umtx_op_rw_wrlock(struct thread *td, struct _umtx_op_args *uap,
4181     const struct umtx_copyops *ops)
4182 {
4183 	struct _umtx_time timeout;
4184 	int error;
4185 
4186 	/* Allow a null timespec (wait forever). */
4187 	if (uap->uaddr2 == NULL) {
4188 		error = do_rw_wrlock(td, uap->obj, 0);
4189 	} else {
4190 		error = ops->copyin_umtx_time(uap->uaddr2,
4191 		   (size_t)uap->uaddr1, &timeout);
4192 		if (error != 0)
4193 			return (error);
4194 
4195 		error = do_rw_wrlock(td, uap->obj, &timeout);
4196 	}
4197 	return (error);
4198 }
4199 
4200 static int
4201 __umtx_op_rw_unlock(struct thread *td, struct _umtx_op_args *uap,
4202     const struct umtx_copyops *ops __unused)
4203 {
4204 
4205 	return (do_rw_unlock(td, uap->obj));
4206 }
4207 
4208 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4209 static int
4210 __umtx_op_sem_wait(struct thread *td, struct _umtx_op_args *uap,
4211     const struct umtx_copyops *ops)
4212 {
4213 	struct _umtx_time *tm_p, timeout;
4214 	int error;
4215 
4216 	/* Allow a null timespec (wait forever). */
4217 	if (uap->uaddr2 == NULL)
4218 		tm_p = NULL;
4219 	else {
4220 		error = ops->copyin_umtx_time(
4221 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4222 		if (error != 0)
4223 			return (error);
4224 		tm_p = &timeout;
4225 	}
4226 	return (do_sem_wait(td, uap->obj, tm_p));
4227 }
4228 
4229 static int
4230 __umtx_op_sem_wake(struct thread *td, struct _umtx_op_args *uap,
4231     const struct umtx_copyops *ops __unused)
4232 {
4233 
4234 	return (do_sem_wake(td, uap->obj));
4235 }
4236 #endif
4237 
4238 static int
4239 __umtx_op_wake2_umutex(struct thread *td, struct _umtx_op_args *uap,
4240     const struct umtx_copyops *ops __unused)
4241 {
4242 
4243 	return (do_wake2_umutex(td, uap->obj, uap->val));
4244 }
4245 
4246 static int
4247 __umtx_op_sem2_wait(struct thread *td, struct _umtx_op_args *uap,
4248     const struct umtx_copyops *ops)
4249 {
4250 	struct _umtx_time *tm_p, timeout;
4251 	size_t uasize;
4252 	int error;
4253 
4254 	/* Allow a null timespec (wait forever). */
4255 	if (uap->uaddr2 == NULL) {
4256 		uasize = 0;
4257 		tm_p = NULL;
4258 	} else {
4259 		uasize = (size_t)uap->uaddr1;
4260 		error = ops->copyin_umtx_time(uap->uaddr2, uasize, &timeout);
4261 		if (error != 0)
4262 			return (error);
4263 		tm_p = &timeout;
4264 	}
4265 	error = do_sem2_wait(td, uap->obj, tm_p);
4266 	if (error == EINTR && uap->uaddr2 != NULL &&
4267 	    (timeout._flags & UMTX_ABSTIME) == 0 &&
4268 	    uasize >= ops->umtx_time_sz + ops->timespec_sz) {
4269 		error = ops->copyout_timeout(
4270 		    (void *)((uintptr_t)uap->uaddr2 + ops->umtx_time_sz),
4271 		    uasize - ops->umtx_time_sz, &timeout._timeout);
4272 		if (error == 0) {
4273 			error = EINTR;
4274 		}
4275 	}
4276 
4277 	return (error);
4278 }
4279 
4280 static int
4281 __umtx_op_sem2_wake(struct thread *td, struct _umtx_op_args *uap,
4282     const struct umtx_copyops *ops __unused)
4283 {
4284 
4285 	return (do_sem2_wake(td, uap->obj));
4286 }
4287 
4288 #define	USHM_OBJ_UMTX(o)						\
4289     ((struct umtx_shm_obj_list *)(&(o)->umtx_data))
4290 
4291 #define	USHMF_REG_LINKED	0x0001
4292 #define	USHMF_OBJ_LINKED	0x0002
4293 struct umtx_shm_reg {
4294 	TAILQ_ENTRY(umtx_shm_reg) ushm_reg_link;
4295 	LIST_ENTRY(umtx_shm_reg) ushm_obj_link;
4296 	struct umtx_key		ushm_key;
4297 	struct ucred		*ushm_cred;
4298 	struct shmfd		*ushm_obj;
4299 	u_int			ushm_refcnt;
4300 	u_int			ushm_flags;
4301 };
4302 
4303 LIST_HEAD(umtx_shm_obj_list, umtx_shm_reg);
4304 TAILQ_HEAD(umtx_shm_reg_head, umtx_shm_reg);
4305 
4306 static uma_zone_t umtx_shm_reg_zone;
4307 static struct umtx_shm_reg_head umtx_shm_registry[UMTX_CHAINS];
4308 static struct mtx umtx_shm_lock;
4309 static struct umtx_shm_reg_head umtx_shm_reg_delfree =
4310     TAILQ_HEAD_INITIALIZER(umtx_shm_reg_delfree);
4311 
4312 static void umtx_shm_free_reg(struct umtx_shm_reg *reg);
4313 
4314 static void
4315 umtx_shm_reg_delfree_tq(void *context __unused, int pending __unused)
4316 {
4317 	struct umtx_shm_reg_head d;
4318 	struct umtx_shm_reg *reg, *reg1;
4319 
4320 	TAILQ_INIT(&d);
4321 	mtx_lock(&umtx_shm_lock);
4322 	TAILQ_CONCAT(&d, &umtx_shm_reg_delfree, ushm_reg_link);
4323 	mtx_unlock(&umtx_shm_lock);
4324 	TAILQ_FOREACH_SAFE(reg, &d, ushm_reg_link, reg1) {
4325 		TAILQ_REMOVE(&d, reg, ushm_reg_link);
4326 		umtx_shm_free_reg(reg);
4327 	}
4328 }
4329 
4330 static struct task umtx_shm_reg_delfree_task =
4331     TASK_INITIALIZER(0, umtx_shm_reg_delfree_tq, NULL);
4332 
4333 static struct umtx_shm_reg *
4334 umtx_shm_find_reg_locked(const struct umtx_key *key)
4335 {
4336 	struct umtx_shm_reg *reg;
4337 	struct umtx_shm_reg_head *reg_head;
4338 
4339 	KASSERT(key->shared, ("umtx_p_find_rg: private key"));
4340 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4341 	reg_head = &umtx_shm_registry[key->hash];
4342 	TAILQ_FOREACH(reg, reg_head, ushm_reg_link) {
4343 		KASSERT(reg->ushm_key.shared,
4344 		    ("non-shared key on reg %p %d", reg, reg->ushm_key.shared));
4345 		if (reg->ushm_key.info.shared.object ==
4346 		    key->info.shared.object &&
4347 		    reg->ushm_key.info.shared.offset ==
4348 		    key->info.shared.offset) {
4349 			KASSERT(reg->ushm_key.type == TYPE_SHM, ("TYPE_USHM"));
4350 			KASSERT(reg->ushm_refcnt > 0,
4351 			    ("reg %p refcnt 0 onlist", reg));
4352 			KASSERT((reg->ushm_flags & USHMF_REG_LINKED) != 0,
4353 			    ("reg %p not linked", reg));
4354 			reg->ushm_refcnt++;
4355 			return (reg);
4356 		}
4357 	}
4358 	return (NULL);
4359 }
4360 
4361 static struct umtx_shm_reg *
4362 umtx_shm_find_reg(const struct umtx_key *key)
4363 {
4364 	struct umtx_shm_reg *reg;
4365 
4366 	mtx_lock(&umtx_shm_lock);
4367 	reg = umtx_shm_find_reg_locked(key);
4368 	mtx_unlock(&umtx_shm_lock);
4369 	return (reg);
4370 }
4371 
4372 static void
4373 umtx_shm_free_reg(struct umtx_shm_reg *reg)
4374 {
4375 
4376 	chgumtxcnt(reg->ushm_cred->cr_ruidinfo, -1, 0);
4377 	crfree(reg->ushm_cred);
4378 	shm_drop(reg->ushm_obj);
4379 	uma_zfree(umtx_shm_reg_zone, reg);
4380 }
4381 
4382 static bool
4383 umtx_shm_unref_reg_locked(struct umtx_shm_reg *reg, bool force)
4384 {
4385 	bool res;
4386 
4387 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4388 	KASSERT(reg->ushm_refcnt > 0, ("ushm_reg %p refcnt 0", reg));
4389 	reg->ushm_refcnt--;
4390 	res = reg->ushm_refcnt == 0;
4391 	if (res || force) {
4392 		if ((reg->ushm_flags & USHMF_REG_LINKED) != 0) {
4393 			TAILQ_REMOVE(&umtx_shm_registry[reg->ushm_key.hash],
4394 			    reg, ushm_reg_link);
4395 			reg->ushm_flags &= ~USHMF_REG_LINKED;
4396 		}
4397 		if ((reg->ushm_flags & USHMF_OBJ_LINKED) != 0) {
4398 			LIST_REMOVE(reg, ushm_obj_link);
4399 			reg->ushm_flags &= ~USHMF_OBJ_LINKED;
4400 		}
4401 	}
4402 	return (res);
4403 }
4404 
4405 static void
4406 umtx_shm_unref_reg(struct umtx_shm_reg *reg, bool force)
4407 {
4408 	vm_object_t object;
4409 	bool dofree;
4410 
4411 	if (force) {
4412 		object = reg->ushm_obj->shm_object;
4413 		VM_OBJECT_WLOCK(object);
4414 		vm_object_set_flag(object, OBJ_UMTXDEAD);
4415 		VM_OBJECT_WUNLOCK(object);
4416 	}
4417 	mtx_lock(&umtx_shm_lock);
4418 	dofree = umtx_shm_unref_reg_locked(reg, force);
4419 	mtx_unlock(&umtx_shm_lock);
4420 	if (dofree)
4421 		umtx_shm_free_reg(reg);
4422 }
4423 
4424 void
4425 umtx_shm_object_init(vm_object_t object)
4426 {
4427 
4428 	LIST_INIT(USHM_OBJ_UMTX(object));
4429 }
4430 
4431 void
4432 umtx_shm_object_terminated(vm_object_t object)
4433 {
4434 	struct umtx_shm_reg *reg, *reg1;
4435 	bool dofree;
4436 
4437 	if (LIST_EMPTY(USHM_OBJ_UMTX(object)))
4438 		return;
4439 
4440 	dofree = false;
4441 	mtx_lock(&umtx_shm_lock);
4442 	LIST_FOREACH_SAFE(reg, USHM_OBJ_UMTX(object), ushm_obj_link, reg1) {
4443 		if (umtx_shm_unref_reg_locked(reg, true)) {
4444 			TAILQ_INSERT_TAIL(&umtx_shm_reg_delfree, reg,
4445 			    ushm_reg_link);
4446 			dofree = true;
4447 		}
4448 	}
4449 	mtx_unlock(&umtx_shm_lock);
4450 	if (dofree)
4451 		taskqueue_enqueue(taskqueue_thread, &umtx_shm_reg_delfree_task);
4452 }
4453 
4454 static int
4455 umtx_shm_create_reg(struct thread *td, const struct umtx_key *key,
4456     struct umtx_shm_reg **res)
4457 {
4458 	struct umtx_shm_reg *reg, *reg1;
4459 	struct ucred *cred;
4460 	int error;
4461 
4462 	reg = umtx_shm_find_reg(key);
4463 	if (reg != NULL) {
4464 		*res = reg;
4465 		return (0);
4466 	}
4467 	cred = td->td_ucred;
4468 	if (!chgumtxcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_UMTXP)))
4469 		return (ENOMEM);
4470 	reg = uma_zalloc(umtx_shm_reg_zone, M_WAITOK | M_ZERO);
4471 	reg->ushm_refcnt = 1;
4472 	bcopy(key, &reg->ushm_key, sizeof(*key));
4473 	reg->ushm_obj = shm_alloc(td->td_ucred, O_RDWR, false);
4474 	reg->ushm_cred = crhold(cred);
4475 	error = shm_dotruncate(reg->ushm_obj, PAGE_SIZE);
4476 	if (error != 0) {
4477 		umtx_shm_free_reg(reg);
4478 		return (error);
4479 	}
4480 	mtx_lock(&umtx_shm_lock);
4481 	reg1 = umtx_shm_find_reg_locked(key);
4482 	if (reg1 != NULL) {
4483 		mtx_unlock(&umtx_shm_lock);
4484 		umtx_shm_free_reg(reg);
4485 		*res = reg1;
4486 		return (0);
4487 	}
4488 	reg->ushm_refcnt++;
4489 	TAILQ_INSERT_TAIL(&umtx_shm_registry[key->hash], reg, ushm_reg_link);
4490 	LIST_INSERT_HEAD(USHM_OBJ_UMTX(key->info.shared.object), reg,
4491 	    ushm_obj_link);
4492 	reg->ushm_flags = USHMF_REG_LINKED | USHMF_OBJ_LINKED;
4493 	mtx_unlock(&umtx_shm_lock);
4494 	*res = reg;
4495 	return (0);
4496 }
4497 
4498 static int
4499 umtx_shm_alive(struct thread *td, void *addr)
4500 {
4501 	vm_map_t map;
4502 	vm_map_entry_t entry;
4503 	vm_object_t object;
4504 	vm_pindex_t pindex;
4505 	vm_prot_t prot;
4506 	int res, ret;
4507 	boolean_t wired;
4508 
4509 	map = &td->td_proc->p_vmspace->vm_map;
4510 	res = vm_map_lookup(&map, (uintptr_t)addr, VM_PROT_READ, &entry,
4511 	    &object, &pindex, &prot, &wired);
4512 	if (res != KERN_SUCCESS)
4513 		return (EFAULT);
4514 	if (object == NULL)
4515 		ret = EINVAL;
4516 	else
4517 		ret = (object->flags & OBJ_UMTXDEAD) != 0 ? ENOTTY : 0;
4518 	vm_map_lookup_done(map, entry);
4519 	return (ret);
4520 }
4521 
4522 static void
4523 umtx_shm_init(void)
4524 {
4525 	int i;
4526 
4527 	umtx_shm_reg_zone = uma_zcreate("umtx_shm", sizeof(struct umtx_shm_reg),
4528 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
4529 	mtx_init(&umtx_shm_lock, "umtxshm", NULL, MTX_DEF);
4530 	for (i = 0; i < nitems(umtx_shm_registry); i++)
4531 		TAILQ_INIT(&umtx_shm_registry[i]);
4532 }
4533 
4534 static int
4535 umtx_shm(struct thread *td, void *addr, u_int flags)
4536 {
4537 	struct umtx_key key;
4538 	struct umtx_shm_reg *reg;
4539 	struct file *fp;
4540 	int error, fd;
4541 
4542 	if (__bitcount(flags & (UMTX_SHM_CREAT | UMTX_SHM_LOOKUP |
4543 	    UMTX_SHM_DESTROY| UMTX_SHM_ALIVE)) != 1)
4544 		return (EINVAL);
4545 	if ((flags & UMTX_SHM_ALIVE) != 0)
4546 		return (umtx_shm_alive(td, addr));
4547 	error = umtx_key_get(addr, TYPE_SHM, PROCESS_SHARE, &key);
4548 	if (error != 0)
4549 		return (error);
4550 	KASSERT(key.shared == 1, ("non-shared key"));
4551 	if ((flags & UMTX_SHM_CREAT) != 0) {
4552 		error = umtx_shm_create_reg(td, &key, &reg);
4553 	} else {
4554 		reg = umtx_shm_find_reg(&key);
4555 		if (reg == NULL)
4556 			error = ESRCH;
4557 	}
4558 	umtx_key_release(&key);
4559 	if (error != 0)
4560 		return (error);
4561 	KASSERT(reg != NULL, ("no reg"));
4562 	if ((flags & UMTX_SHM_DESTROY) != 0) {
4563 		umtx_shm_unref_reg(reg, true);
4564 	} else {
4565 #if 0
4566 #ifdef MAC
4567 		error = mac_posixshm_check_open(td->td_ucred,
4568 		    reg->ushm_obj, FFLAGS(O_RDWR));
4569 		if (error == 0)
4570 #endif
4571 			error = shm_access(reg->ushm_obj, td->td_ucred,
4572 			    FFLAGS(O_RDWR));
4573 		if (error == 0)
4574 #endif
4575 			error = falloc_caps(td, &fp, &fd, O_CLOEXEC, NULL);
4576 		if (error == 0) {
4577 			shm_hold(reg->ushm_obj);
4578 			finit(fp, FFLAGS(O_RDWR), DTYPE_SHM, reg->ushm_obj,
4579 			    &shm_ops);
4580 			td->td_retval[0] = fd;
4581 			fdrop(fp, td);
4582 		}
4583 	}
4584 	umtx_shm_unref_reg(reg, false);
4585 	return (error);
4586 }
4587 
4588 static int
4589 __umtx_op_shm(struct thread *td, struct _umtx_op_args *uap,
4590     const struct umtx_copyops *ops __unused)
4591 {
4592 
4593 	return (umtx_shm(td, uap->uaddr1, uap->val));
4594 }
4595 
4596 static int
4597 __umtx_op_robust_lists(struct thread *td, struct _umtx_op_args *uap,
4598     const struct umtx_copyops *ops)
4599 {
4600 	struct umtx_robust_lists_params rb;
4601 	int error;
4602 
4603 	if (ops->compat32) {
4604 		if ((td->td_pflags2 & TDP2_COMPAT32RB) == 0 &&
4605 		    (td->td_rb_list != 0 || td->td_rbp_list != 0 ||
4606 		    td->td_rb_inact != 0))
4607 			return (EBUSY);
4608 	} else if ((td->td_pflags2 & TDP2_COMPAT32RB) != 0) {
4609 		return (EBUSY);
4610 	}
4611 
4612 	bzero(&rb, sizeof(rb));
4613 	error = ops->copyin_robust_lists(uap->uaddr1, uap->val, &rb);
4614 	if (error != 0)
4615 		return (error);
4616 
4617 	if (ops->compat32)
4618 		td->td_pflags2 |= TDP2_COMPAT32RB;
4619 
4620 	td->td_rb_list = rb.robust_list_offset;
4621 	td->td_rbp_list = rb.robust_priv_list_offset;
4622 	td->td_rb_inact = rb.robust_inact_offset;
4623 	return (0);
4624 }
4625 
4626 static int
4627 __umtx_op_get_min_timeout(struct thread *td, struct _umtx_op_args *uap,
4628     const struct umtx_copyops *ops)
4629 {
4630 	long val;
4631 	int error, val1;
4632 
4633 	val = sbttons(td->td_proc->p_umtx_min_timeout);
4634 	if (ops->compat32) {
4635 		val1 = (int)val;
4636 		error = copyout(&val1, uap->uaddr1, sizeof(val1));
4637 	} else {
4638 		error = copyout(&val, uap->uaddr1, sizeof(val));
4639 	}
4640 	return (error);
4641 }
4642 
4643 static int
4644 __umtx_op_set_min_timeout(struct thread *td, struct _umtx_op_args *uap,
4645     const struct umtx_copyops *ops)
4646 {
4647 	if (uap->val < 0)
4648 		return (EINVAL);
4649 	td->td_proc->p_umtx_min_timeout = nstosbt(uap->val);
4650 	return (0);
4651 }
4652 
4653 #if defined(__i386__) || defined(__amd64__)
4654 /*
4655  * Provide the standard 32-bit definitions for x86, since native/compat32 use a
4656  * 32-bit time_t there.  Other architectures just need the i386 definitions
4657  * along with their standard compat32.
4658  */
4659 struct timespecx32 {
4660 	int64_t			tv_sec;
4661 	int32_t			tv_nsec;
4662 };
4663 
4664 struct umtx_timex32 {
4665 	struct	timespecx32	_timeout;
4666 	uint32_t		_flags;
4667 	uint32_t		_clockid;
4668 };
4669 
4670 #ifndef __i386__
4671 #define	timespeci386	timespec32
4672 #define	umtx_timei386	umtx_time32
4673 #endif
4674 #else /* !__i386__ && !__amd64__ */
4675 /* 32-bit architectures can emulate i386, so define these almost everywhere. */
4676 struct timespeci386 {
4677 	int32_t			tv_sec;
4678 	int32_t			tv_nsec;
4679 };
4680 
4681 struct umtx_timei386 {
4682 	struct	timespeci386	_timeout;
4683 	uint32_t		_flags;
4684 	uint32_t		_clockid;
4685 };
4686 
4687 #if defined(__LP64__)
4688 #define	timespecx32	timespec32
4689 #define	umtx_timex32	umtx_time32
4690 #endif
4691 #endif
4692 
4693 static int
4694 umtx_copyin_robust_lists32(const void *uaddr, size_t size,
4695     struct umtx_robust_lists_params *rbp)
4696 {
4697 	struct umtx_robust_lists_params_compat32 rb32;
4698 	int error;
4699 
4700 	if (size > sizeof(rb32))
4701 		return (EINVAL);
4702 	bzero(&rb32, sizeof(rb32));
4703 	error = copyin(uaddr, &rb32, size);
4704 	if (error != 0)
4705 		return (error);
4706 	CP(rb32, *rbp, robust_list_offset);
4707 	CP(rb32, *rbp, robust_priv_list_offset);
4708 	CP(rb32, *rbp, robust_inact_offset);
4709 	return (0);
4710 }
4711 
4712 #ifndef __i386__
4713 static inline int
4714 umtx_copyin_timeouti386(const void *uaddr, struct timespec *tsp)
4715 {
4716 	struct timespeci386 ts32;
4717 	int error;
4718 
4719 	error = copyin(uaddr, &ts32, sizeof(ts32));
4720 	if (error == 0) {
4721 		if (!timespecvalid_interval(&ts32))
4722 			error = EINVAL;
4723 		else {
4724 			CP(ts32, *tsp, tv_sec);
4725 			CP(ts32, *tsp, tv_nsec);
4726 		}
4727 	}
4728 	return (error);
4729 }
4730 
4731 static inline int
4732 umtx_copyin_umtx_timei386(const void *uaddr, size_t size, struct _umtx_time *tp)
4733 {
4734 	struct umtx_timei386 t32;
4735 	int error;
4736 
4737 	t32._clockid = CLOCK_REALTIME;
4738 	t32._flags   = 0;
4739 	if (size <= sizeof(t32._timeout))
4740 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4741 	else
4742 		error = copyin(uaddr, &t32, sizeof(t32));
4743 	if (error != 0)
4744 		return (error);
4745 	if (!timespecvalid_interval(&t32._timeout))
4746 		return (EINVAL);
4747 	TS_CP(t32, *tp, _timeout);
4748 	CP(t32, *tp, _flags);
4749 	CP(t32, *tp, _clockid);
4750 	return (0);
4751 }
4752 
4753 static int
4754 umtx_copyout_timeouti386(void *uaddr, size_t sz, struct timespec *tsp)
4755 {
4756 	struct timespeci386 remain32 = {
4757 		.tv_sec = tsp->tv_sec,
4758 		.tv_nsec = tsp->tv_nsec,
4759 	};
4760 
4761 	/*
4762 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4763 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4764 	 * copyops.
4765 	 */
4766 	KASSERT(sz >= sizeof(remain32),
4767 	    ("umtx_copyops specifies incorrect sizes"));
4768 
4769 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4770 }
4771 #endif /* !__i386__ */
4772 
4773 #if defined(__i386__) || defined(__LP64__)
4774 static inline int
4775 umtx_copyin_timeoutx32(const void *uaddr, struct timespec *tsp)
4776 {
4777 	struct timespecx32 ts32;
4778 	int error;
4779 
4780 	error = copyin(uaddr, &ts32, sizeof(ts32));
4781 	if (error == 0) {
4782 		if (!timespecvalid_interval(&ts32))
4783 			error = EINVAL;
4784 		else {
4785 			CP(ts32, *tsp, tv_sec);
4786 			CP(ts32, *tsp, tv_nsec);
4787 		}
4788 	}
4789 	return (error);
4790 }
4791 
4792 static inline int
4793 umtx_copyin_umtx_timex32(const void *uaddr, size_t size, struct _umtx_time *tp)
4794 {
4795 	struct umtx_timex32 t32;
4796 	int error;
4797 
4798 	t32._clockid = CLOCK_REALTIME;
4799 	t32._flags   = 0;
4800 	if (size <= sizeof(t32._timeout))
4801 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4802 	else
4803 		error = copyin(uaddr, &t32, sizeof(t32));
4804 	if (error != 0)
4805 		return (error);
4806 	if (!timespecvalid_interval(&t32._timeout))
4807 		return (EINVAL);
4808 	TS_CP(t32, *tp, _timeout);
4809 	CP(t32, *tp, _flags);
4810 	CP(t32, *tp, _clockid);
4811 	return (0);
4812 }
4813 
4814 static int
4815 umtx_copyout_timeoutx32(void *uaddr, size_t sz, struct timespec *tsp)
4816 {
4817 	struct timespecx32 remain32 = {
4818 		.tv_sec = tsp->tv_sec,
4819 		.tv_nsec = tsp->tv_nsec,
4820 	};
4821 
4822 	/*
4823 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4824 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4825 	 * copyops.
4826 	 */
4827 	KASSERT(sz >= sizeof(remain32),
4828 	    ("umtx_copyops specifies incorrect sizes"));
4829 
4830 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4831 }
4832 #endif /* __i386__ || __LP64__ */
4833 
4834 typedef int (*_umtx_op_func)(struct thread *td, struct _umtx_op_args *uap,
4835     const struct umtx_copyops *umtx_ops);
4836 
4837 static const _umtx_op_func op_table[] = {
4838 #ifdef COMPAT_FREEBSD10
4839 	[UMTX_OP_LOCK]		= __umtx_op_lock_umtx,
4840 	[UMTX_OP_UNLOCK]	= __umtx_op_unlock_umtx,
4841 #else
4842 	[UMTX_OP_LOCK]		= __umtx_op_unimpl,
4843 	[UMTX_OP_UNLOCK]	= __umtx_op_unimpl,
4844 #endif
4845 	[UMTX_OP_WAIT]		= __umtx_op_wait,
4846 	[UMTX_OP_WAKE]		= __umtx_op_wake,
4847 	[UMTX_OP_MUTEX_TRYLOCK]	= __umtx_op_trylock_umutex,
4848 	[UMTX_OP_MUTEX_LOCK]	= __umtx_op_lock_umutex,
4849 	[UMTX_OP_MUTEX_UNLOCK]	= __umtx_op_unlock_umutex,
4850 	[UMTX_OP_SET_CEILING]	= __umtx_op_set_ceiling,
4851 	[UMTX_OP_CV_WAIT]	= __umtx_op_cv_wait,
4852 	[UMTX_OP_CV_SIGNAL]	= __umtx_op_cv_signal,
4853 	[UMTX_OP_CV_BROADCAST]	= __umtx_op_cv_broadcast,
4854 	[UMTX_OP_WAIT_UINT]	= __umtx_op_wait_uint,
4855 	[UMTX_OP_RW_RDLOCK]	= __umtx_op_rw_rdlock,
4856 	[UMTX_OP_RW_WRLOCK]	= __umtx_op_rw_wrlock,
4857 	[UMTX_OP_RW_UNLOCK]	= __umtx_op_rw_unlock,
4858 	[UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private,
4859 	[UMTX_OP_WAKE_PRIVATE]	= __umtx_op_wake_private,
4860 	[UMTX_OP_MUTEX_WAIT]	= __umtx_op_wait_umutex,
4861 	[UMTX_OP_MUTEX_WAKE]	= __umtx_op_wake_umutex,
4862 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4863 	[UMTX_OP_SEM_WAIT]	= __umtx_op_sem_wait,
4864 	[UMTX_OP_SEM_WAKE]	= __umtx_op_sem_wake,
4865 #else
4866 	[UMTX_OP_SEM_WAIT]	= __umtx_op_unimpl,
4867 	[UMTX_OP_SEM_WAKE]	= __umtx_op_unimpl,
4868 #endif
4869 	[UMTX_OP_NWAKE_PRIVATE]	= __umtx_op_nwake_private,
4870 	[UMTX_OP_MUTEX_WAKE2]	= __umtx_op_wake2_umutex,
4871 	[UMTX_OP_SEM2_WAIT]	= __umtx_op_sem2_wait,
4872 	[UMTX_OP_SEM2_WAKE]	= __umtx_op_sem2_wake,
4873 	[UMTX_OP_SHM]		= __umtx_op_shm,
4874 	[UMTX_OP_ROBUST_LISTS]	= __umtx_op_robust_lists,
4875 	[UMTX_OP_GET_MIN_TIMEOUT] = __umtx_op_get_min_timeout,
4876 	[UMTX_OP_SET_MIN_TIMEOUT] = __umtx_op_set_min_timeout,
4877 };
4878 
4879 static const struct umtx_copyops umtx_native_ops = {
4880 	.copyin_timeout = umtx_copyin_timeout,
4881 	.copyin_umtx_time = umtx_copyin_umtx_time,
4882 	.copyin_robust_lists = umtx_copyin_robust_lists,
4883 	.copyout_timeout = umtx_copyout_timeout,
4884 	.timespec_sz = sizeof(struct timespec),
4885 	.umtx_time_sz = sizeof(struct _umtx_time),
4886 };
4887 
4888 #ifndef __i386__
4889 static const struct umtx_copyops umtx_native_opsi386 = {
4890 	.copyin_timeout = umtx_copyin_timeouti386,
4891 	.copyin_umtx_time = umtx_copyin_umtx_timei386,
4892 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4893 	.copyout_timeout = umtx_copyout_timeouti386,
4894 	.timespec_sz = sizeof(struct timespeci386),
4895 	.umtx_time_sz = sizeof(struct umtx_timei386),
4896 	.compat32 = true,
4897 };
4898 #endif
4899 
4900 #if defined(__i386__) || defined(__LP64__)
4901 /* i386 can emulate other 32-bit archs, too! */
4902 static const struct umtx_copyops umtx_native_opsx32 = {
4903 	.copyin_timeout = umtx_copyin_timeoutx32,
4904 	.copyin_umtx_time = umtx_copyin_umtx_timex32,
4905 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4906 	.copyout_timeout = umtx_copyout_timeoutx32,
4907 	.timespec_sz = sizeof(struct timespecx32),
4908 	.umtx_time_sz = sizeof(struct umtx_timex32),
4909 	.compat32 = true,
4910 };
4911 
4912 #ifdef COMPAT_FREEBSD32
4913 #ifdef __amd64__
4914 #define	umtx_native_ops32	umtx_native_opsi386
4915 #else
4916 #define	umtx_native_ops32	umtx_native_opsx32
4917 #endif
4918 #endif /* COMPAT_FREEBSD32 */
4919 #endif /* __i386__ || __LP64__ */
4920 
4921 #define	UMTX_OP__FLAGS	(UMTX_OP__32BIT | UMTX_OP__I386)
4922 
4923 static int
4924 kern__umtx_op(struct thread *td, void *obj, int op, unsigned long val,
4925     void *uaddr1, void *uaddr2, const struct umtx_copyops *ops)
4926 {
4927 	struct _umtx_op_args uap = {
4928 		.obj = obj,
4929 		.op = op & ~UMTX_OP__FLAGS,
4930 		.val = val,
4931 		.uaddr1 = uaddr1,
4932 		.uaddr2 = uaddr2
4933 	};
4934 
4935 	if ((uap.op >= nitems(op_table)))
4936 		return (EINVAL);
4937 	return ((*op_table[uap.op])(td, &uap, ops));
4938 }
4939 
4940 int
4941 sys__umtx_op(struct thread *td, struct _umtx_op_args *uap)
4942 {
4943 	static const struct umtx_copyops *umtx_ops;
4944 
4945 	umtx_ops = &umtx_native_ops;
4946 #ifdef __LP64__
4947 	if ((uap->op & (UMTX_OP__32BIT | UMTX_OP__I386)) != 0) {
4948 		if ((uap->op & UMTX_OP__I386) != 0)
4949 			umtx_ops = &umtx_native_opsi386;
4950 		else
4951 			umtx_ops = &umtx_native_opsx32;
4952 	}
4953 #elif !defined(__i386__)
4954 	/* We consider UMTX_OP__32BIT a nop on !i386 ILP32. */
4955 	if ((uap->op & UMTX_OP__I386) != 0)
4956 		umtx_ops = &umtx_native_opsi386;
4957 #else
4958 	/* Likewise, UMTX_OP__I386 is a nop on i386. */
4959 	if ((uap->op & UMTX_OP__32BIT) != 0)
4960 		umtx_ops = &umtx_native_opsx32;
4961 #endif
4962 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4963 	    uap->uaddr2, umtx_ops));
4964 }
4965 
4966 #ifdef COMPAT_FREEBSD32
4967 #ifdef COMPAT_FREEBSD10
4968 int
4969 freebsd10_freebsd32__umtx_lock(struct thread *td,
4970     struct freebsd10_freebsd32__umtx_lock_args *uap)
4971 {
4972 	return (do_lock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid, NULL));
4973 }
4974 
4975 int
4976 freebsd10_freebsd32__umtx_unlock(struct thread *td,
4977     struct freebsd10_freebsd32__umtx_unlock_args *uap)
4978 {
4979 	return (do_unlock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid));
4980 }
4981 #endif /* COMPAT_FREEBSD10 */
4982 
4983 int
4984 freebsd32__umtx_op(struct thread *td, struct freebsd32__umtx_op_args *uap)
4985 {
4986 
4987 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4988 	    uap->uaddr2, &umtx_native_ops32));
4989 }
4990 #endif /* COMPAT_FREEBSD32 */
4991 
4992 void
4993 umtx_thread_init(struct thread *td)
4994 {
4995 
4996 	td->td_umtxq = umtxq_alloc();
4997 	td->td_umtxq->uq_thread = td;
4998 }
4999 
5000 void
5001 umtx_thread_fini(struct thread *td)
5002 {
5003 
5004 	umtxq_free(td->td_umtxq);
5005 }
5006 
5007 /*
5008  * It will be called when new thread is created, e.g fork().
5009  */
5010 void
5011 umtx_thread_alloc(struct thread *td)
5012 {
5013 	struct umtx_q *uq;
5014 
5015 	uq = td->td_umtxq;
5016 	uq->uq_inherited_pri = PRI_MAX;
5017 
5018 	KASSERT(uq->uq_flags == 0, ("uq_flags != 0"));
5019 	KASSERT(uq->uq_thread == td, ("uq_thread != td"));
5020 	KASSERT(uq->uq_pi_blocked == NULL, ("uq_pi_blocked != NULL"));
5021 	KASSERT(TAILQ_EMPTY(&uq->uq_pi_contested), ("uq_pi_contested is not empty"));
5022 }
5023 
5024 /*
5025  * exec() hook.
5026  *
5027  * Clear robust lists for all process' threads, not delaying the
5028  * cleanup to thread exit, since the relevant address space is
5029  * destroyed right now.
5030  */
5031 void
5032 umtx_exec(struct proc *p)
5033 {
5034 	struct thread *td;
5035 
5036 	KASSERT(p == curproc, ("need curproc"));
5037 	KASSERT((p->p_flag & P_HADTHREADS) == 0 ||
5038 	    (p->p_flag & P_STOPPED_SINGLE) != 0,
5039 	    ("curproc must be single-threaded"));
5040 	/*
5041 	 * There is no need to lock the list as only this thread can be
5042 	 * running.
5043 	 */
5044 	FOREACH_THREAD_IN_PROC(p, td) {
5045 		KASSERT(td == curthread ||
5046 		    ((td->td_flags & TDF_BOUNDARY) != 0 && TD_IS_SUSPENDED(td)),
5047 		    ("running thread %p %p", p, td));
5048 		umtx_thread_cleanup(td);
5049 		td->td_rb_list = td->td_rbp_list = td->td_rb_inact = 0;
5050 	}
5051 
5052 	p->p_umtx_min_timeout = 0;
5053 }
5054 
5055 /*
5056  * thread exit hook.
5057  */
5058 void
5059 umtx_thread_exit(struct thread *td)
5060 {
5061 
5062 	umtx_thread_cleanup(td);
5063 }
5064 
5065 static int
5066 umtx_read_uptr(struct thread *td, uintptr_t ptr, uintptr_t *res, bool compat32)
5067 {
5068 	u_long res1;
5069 	uint32_t res32;
5070 	int error;
5071 
5072 	if (compat32) {
5073 		error = fueword32((void *)ptr, &res32);
5074 		if (error == 0)
5075 			res1 = res32;
5076 	} else {
5077 		error = fueword((void *)ptr, &res1);
5078 	}
5079 	if (error == 0)
5080 		*res = res1;
5081 	else
5082 		error = EFAULT;
5083 	return (error);
5084 }
5085 
5086 static void
5087 umtx_read_rb_list(struct thread *td, struct umutex *m, uintptr_t *rb_list,
5088     bool compat32)
5089 {
5090 	struct umutex32 m32;
5091 
5092 	if (compat32) {
5093 		memcpy(&m32, m, sizeof(m32));
5094 		*rb_list = m32.m_rb_lnk;
5095 	} else {
5096 		*rb_list = m->m_rb_lnk;
5097 	}
5098 }
5099 
5100 static int
5101 umtx_handle_rb(struct thread *td, uintptr_t rbp, uintptr_t *rb_list, bool inact,
5102     bool compat32)
5103 {
5104 	struct umutex m;
5105 	int error;
5106 
5107 	KASSERT(td->td_proc == curproc, ("need current vmspace"));
5108 	error = copyin((void *)rbp, &m, sizeof(m));
5109 	if (error != 0)
5110 		return (error);
5111 	if (rb_list != NULL)
5112 		umtx_read_rb_list(td, &m, rb_list, compat32);
5113 	if ((m.m_flags & UMUTEX_ROBUST) == 0)
5114 		return (EINVAL);
5115 	if ((m.m_owner & ~UMUTEX_CONTESTED) != td->td_tid)
5116 		/* inact is cleared after unlock, allow the inconsistency */
5117 		return (inact ? 0 : EINVAL);
5118 	return (do_unlock_umutex(td, (struct umutex *)rbp, true));
5119 }
5120 
5121 static void
5122 umtx_cleanup_rb_list(struct thread *td, uintptr_t rb_list, uintptr_t *rb_inact,
5123     const char *name, bool compat32)
5124 {
5125 	int error, i;
5126 	uintptr_t rbp;
5127 	bool inact;
5128 
5129 	if (rb_list == 0)
5130 		return;
5131 	error = umtx_read_uptr(td, rb_list, &rbp, compat32);
5132 	for (i = 0; error == 0 && rbp != 0 && i < umtx_max_rb; i++) {
5133 		if (rbp == *rb_inact) {
5134 			inact = true;
5135 			*rb_inact = 0;
5136 		} else
5137 			inact = false;
5138 		error = umtx_handle_rb(td, rbp, &rbp, inact, compat32);
5139 	}
5140 	if (i == umtx_max_rb && umtx_verbose_rb) {
5141 		uprintf("comm %s pid %d: reached umtx %smax rb %d\n",
5142 		    td->td_proc->p_comm, td->td_proc->p_pid, name, umtx_max_rb);
5143 	}
5144 	if (error != 0 && umtx_verbose_rb) {
5145 		uprintf("comm %s pid %d: handling %srb error %d\n",
5146 		    td->td_proc->p_comm, td->td_proc->p_pid, name, error);
5147 	}
5148 }
5149 
5150 /*
5151  * Clean up umtx data.
5152  */
5153 static void
5154 umtx_thread_cleanup(struct thread *td)
5155 {
5156 	struct umtx_q *uq;
5157 	struct umtx_pi *pi;
5158 	uintptr_t rb_inact;
5159 	bool compat32;
5160 
5161 	/*
5162 	 * Disown pi mutexes.
5163 	 */
5164 	uq = td->td_umtxq;
5165 	if (uq != NULL) {
5166 		if (uq->uq_inherited_pri != PRI_MAX ||
5167 		    !TAILQ_EMPTY(&uq->uq_pi_contested)) {
5168 			mtx_lock(&umtx_lock);
5169 			uq->uq_inherited_pri = PRI_MAX;
5170 			while ((pi = TAILQ_FIRST(&uq->uq_pi_contested)) != NULL) {
5171 				pi->pi_owner = NULL;
5172 				TAILQ_REMOVE(&uq->uq_pi_contested, pi, pi_link);
5173 			}
5174 			mtx_unlock(&umtx_lock);
5175 		}
5176 		sched_lend_user_prio_cond(td, PRI_MAX);
5177 	}
5178 
5179 	compat32 = (td->td_pflags2 & TDP2_COMPAT32RB) != 0;
5180 	td->td_pflags2 &= ~TDP2_COMPAT32RB;
5181 
5182 	if (td->td_rb_inact == 0 && td->td_rb_list == 0 && td->td_rbp_list == 0)
5183 		return;
5184 
5185 	/*
5186 	 * Handle terminated robust mutexes.  Must be done after
5187 	 * robust pi disown, otherwise unlock could see unowned
5188 	 * entries.
5189 	 */
5190 	rb_inact = td->td_rb_inact;
5191 	if (rb_inact != 0)
5192 		(void)umtx_read_uptr(td, rb_inact, &rb_inact, compat32);
5193 	umtx_cleanup_rb_list(td, td->td_rb_list, &rb_inact, "", compat32);
5194 	umtx_cleanup_rb_list(td, td->td_rbp_list, &rb_inact, "priv ", compat32);
5195 	if (rb_inact != 0)
5196 		(void)umtx_handle_rb(td, rb_inact, NULL, true, compat32);
5197 }
5198