xref: /freebsd/sys/kern/kern_umtx.c (revision 7a8d62cd046942ab34fcf282ecbbb0747a08a4ee)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
36 
37 #include "opt_umtx_profiling.h"
38 
39 #include <sys/param.h>
40 #include <sys/kernel.h>
41 #include <sys/fcntl.h>
42 #include <sys/file.h>
43 #include <sys/filedesc.h>
44 #include <sys/limits.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/mman.h>
48 #include <sys/mutex.h>
49 #include <sys/priv.h>
50 #include <sys/proc.h>
51 #include <sys/resource.h>
52 #include <sys/resourcevar.h>
53 #include <sys/rwlock.h>
54 #include <sys/sbuf.h>
55 #include <sys/sched.h>
56 #include <sys/smp.h>
57 #include <sys/sysctl.h>
58 #include <sys/systm.h>
59 #include <sys/sysproto.h>
60 #include <sys/syscallsubr.h>
61 #include <sys/taskqueue.h>
62 #include <sys/time.h>
63 #include <sys/eventhandler.h>
64 #include <sys/umtx.h>
65 #include <sys/umtxvar.h>
66 
67 #include <security/mac/mac_framework.h>
68 
69 #include <vm/vm.h>
70 #include <vm/vm_param.h>
71 #include <vm/pmap.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_object.h>
74 
75 #include <machine/atomic.h>
76 #include <machine/cpu.h>
77 
78 #include <compat/freebsd32/freebsd32.h>
79 #ifdef COMPAT_FREEBSD32
80 #include <compat/freebsd32/freebsd32_proto.h>
81 #endif
82 
83 #define _UMUTEX_TRY		1
84 #define _UMUTEX_WAIT		2
85 
86 #ifdef UMTX_PROFILING
87 #define	UPROF_PERC_BIGGER(w, f, sw, sf)					\
88 	(((w) > (sw)) || ((w) == (sw) && (f) > (sf)))
89 #endif
90 
91 #define	UMTXQ_LOCKED_ASSERT(uc)		mtx_assert(&(uc)->uc_lock, MA_OWNED)
92 #ifdef INVARIANTS
93 #define	UMTXQ_ASSERT_LOCKED_BUSY(key) do {				\
94 	struct umtxq_chain *uc;						\
95 									\
96 	uc = umtxq_getchain(key);					\
97 	mtx_assert(&uc->uc_lock, MA_OWNED);				\
98 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));		\
99 } while (0)
100 #else
101 #define	UMTXQ_ASSERT_LOCKED_BUSY(key) do {} while (0)
102 #endif
103 
104 /*
105  * Don't propagate time-sharing priority, there is a security reason,
106  * a user can simply introduce PI-mutex, let thread A lock the mutex,
107  * and let another thread B block on the mutex, because B is
108  * sleeping, its priority will be boosted, this causes A's priority to
109  * be boosted via priority propagating too and will never be lowered even
110  * if it is using 100%CPU, this is unfair to other processes.
111  */
112 
113 #define UPRI(td)	(((td)->td_user_pri >= PRI_MIN_TIMESHARE &&\
114 			  (td)->td_user_pri <= PRI_MAX_TIMESHARE) ?\
115 			 PRI_MAX_TIMESHARE : (td)->td_user_pri)
116 
117 #define	GOLDEN_RATIO_PRIME	2654404609U
118 #ifndef	UMTX_CHAINS
119 #define	UMTX_CHAINS		512
120 #endif
121 #define	UMTX_SHIFTS		(__WORD_BIT - 9)
122 
123 #define	GET_SHARE(flags)	\
124     (((flags) & USYNC_PROCESS_SHARED) == 0 ? THREAD_SHARE : PROCESS_SHARE)
125 
126 #define BUSY_SPINS		200
127 
128 struct umtx_copyops {
129 	int	(*copyin_timeout)(const void *uaddr, struct timespec *tsp);
130 	int	(*copyin_umtx_time)(const void *uaddr, size_t size,
131 	    struct _umtx_time *tp);
132 	int	(*copyin_robust_lists)(const void *uaddr, size_t size,
133 	    struct umtx_robust_lists_params *rbp);
134 	int	(*copyout_timeout)(void *uaddr, size_t size,
135 	    struct timespec *tsp);
136 	const size_t	timespec_sz;
137 	const size_t	umtx_time_sz;
138 	const bool	compat32;
139 };
140 
141 _Static_assert(sizeof(struct umutex) == sizeof(struct umutex32), "umutex32");
142 _Static_assert(__offsetof(struct umutex, m_spare[0]) ==
143     __offsetof(struct umutex32, m_spare[0]), "m_spare32");
144 
145 int umtx_shm_vnobj_persistent = 0;
146 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_vnode_persistent, CTLFLAG_RWTUN,
147     &umtx_shm_vnobj_persistent, 0,
148     "False forces destruction of umtx attached to file, on last close");
149 static int umtx_max_rb = 1000;
150 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_max_robust, CTLFLAG_RWTUN,
151     &umtx_max_rb, 0,
152     "Maximum number of robust mutexes allowed for each thread");
153 
154 static uma_zone_t		umtx_pi_zone;
155 static struct umtxq_chain	umtxq_chains[2][UMTX_CHAINS];
156 static MALLOC_DEFINE(M_UMTX, "umtx", "UMTX queue memory");
157 static int			umtx_pi_allocated;
158 
159 static SYSCTL_NODE(_debug, OID_AUTO, umtx, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
160     "umtx debug");
161 SYSCTL_INT(_debug_umtx, OID_AUTO, umtx_pi_allocated, CTLFLAG_RD,
162     &umtx_pi_allocated, 0, "Allocated umtx_pi");
163 static int umtx_verbose_rb = 1;
164 SYSCTL_INT(_debug_umtx, OID_AUTO, robust_faults_verbose, CTLFLAG_RWTUN,
165     &umtx_verbose_rb, 0,
166     "");
167 
168 #ifdef UMTX_PROFILING
169 static long max_length;
170 SYSCTL_LONG(_debug_umtx, OID_AUTO, max_length, CTLFLAG_RD, &max_length, 0, "max_length");
171 static SYSCTL_NODE(_debug_umtx, OID_AUTO, chains, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
172     "umtx chain stats");
173 #endif
174 
175 static inline void umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
176     const struct _umtx_time *umtxtime);
177 
178 static void umtx_shm_init(void);
179 static void umtxq_sysinit(void *);
180 static void umtxq_hash(struct umtx_key *key);
181 static int do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags,
182     bool rb);
183 static void umtx_thread_cleanup(struct thread *td);
184 SYSINIT(umtx, SI_SUB_EVENTHANDLER+1, SI_ORDER_MIDDLE, umtxq_sysinit, NULL);
185 
186 #define umtxq_signal(key, nwake)	umtxq_signal_queue((key), (nwake), UMTX_SHARED_QUEUE)
187 
188 static struct mtx umtx_lock;
189 
190 #ifdef UMTX_PROFILING
191 static void
192 umtx_init_profiling(void)
193 {
194 	struct sysctl_oid *chain_oid;
195 	char chain_name[10];
196 	int i;
197 
198 	for (i = 0; i < UMTX_CHAINS; ++i) {
199 		snprintf(chain_name, sizeof(chain_name), "%d", i);
200 		chain_oid = SYSCTL_ADD_NODE(NULL,
201 		    SYSCTL_STATIC_CHILDREN(_debug_umtx_chains), OID_AUTO,
202 		    chain_name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
203 		    "umtx hash stats");
204 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
205 		    "max_length0", CTLFLAG_RD, &umtxq_chains[0][i].max_length, 0, NULL);
206 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
207 		    "max_length1", CTLFLAG_RD, &umtxq_chains[1][i].max_length, 0, NULL);
208 	}
209 }
210 
211 static int
212 sysctl_debug_umtx_chains_peaks(SYSCTL_HANDLER_ARGS)
213 {
214 	char buf[512];
215 	struct sbuf sb;
216 	struct umtxq_chain *uc;
217 	u_int fract, i, j, tot, whole;
218 	u_int sf0, sf1, sf2, sf3, sf4;
219 	u_int si0, si1, si2, si3, si4;
220 	u_int sw0, sw1, sw2, sw3, sw4;
221 
222 	sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
223 	for (i = 0; i < 2; i++) {
224 		tot = 0;
225 		for (j = 0; j < UMTX_CHAINS; ++j) {
226 			uc = &umtxq_chains[i][j];
227 			mtx_lock(&uc->uc_lock);
228 			tot += uc->max_length;
229 			mtx_unlock(&uc->uc_lock);
230 		}
231 		if (tot == 0)
232 			sbuf_printf(&sb, "%u) Empty ", i);
233 		else {
234 			sf0 = sf1 = sf2 = sf3 = sf4 = 0;
235 			si0 = si1 = si2 = si3 = si4 = 0;
236 			sw0 = sw1 = sw2 = sw3 = sw4 = 0;
237 			for (j = 0; j < UMTX_CHAINS; j++) {
238 				uc = &umtxq_chains[i][j];
239 				mtx_lock(&uc->uc_lock);
240 				whole = uc->max_length * 100;
241 				mtx_unlock(&uc->uc_lock);
242 				fract = (whole % tot) * 100;
243 				if (UPROF_PERC_BIGGER(whole, fract, sw0, sf0)) {
244 					sf0 = fract;
245 					si0 = j;
246 					sw0 = whole;
247 				} else if (UPROF_PERC_BIGGER(whole, fract, sw1,
248 				    sf1)) {
249 					sf1 = fract;
250 					si1 = j;
251 					sw1 = whole;
252 				} else if (UPROF_PERC_BIGGER(whole, fract, sw2,
253 				    sf2)) {
254 					sf2 = fract;
255 					si2 = j;
256 					sw2 = whole;
257 				} else if (UPROF_PERC_BIGGER(whole, fract, sw3,
258 				    sf3)) {
259 					sf3 = fract;
260 					si3 = j;
261 					sw3 = whole;
262 				} else if (UPROF_PERC_BIGGER(whole, fract, sw4,
263 				    sf4)) {
264 					sf4 = fract;
265 					si4 = j;
266 					sw4 = whole;
267 				}
268 			}
269 			sbuf_printf(&sb, "queue %u:\n", i);
270 			sbuf_printf(&sb, "1st: %u.%u%% idx: %u\n", sw0 / tot,
271 			    sf0 / tot, si0);
272 			sbuf_printf(&sb, "2nd: %u.%u%% idx: %u\n", sw1 / tot,
273 			    sf1 / tot, si1);
274 			sbuf_printf(&sb, "3rd: %u.%u%% idx: %u\n", sw2 / tot,
275 			    sf2 / tot, si2);
276 			sbuf_printf(&sb, "4th: %u.%u%% idx: %u\n", sw3 / tot,
277 			    sf3 / tot, si3);
278 			sbuf_printf(&sb, "5th: %u.%u%% idx: %u\n", sw4 / tot,
279 			    sf4 / tot, si4);
280 		}
281 	}
282 	sbuf_trim(&sb);
283 	sbuf_finish(&sb);
284 	sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
285 	sbuf_delete(&sb);
286 	return (0);
287 }
288 
289 static int
290 sysctl_debug_umtx_chains_clear(SYSCTL_HANDLER_ARGS)
291 {
292 	struct umtxq_chain *uc;
293 	u_int i, j;
294 	int clear, error;
295 
296 	clear = 0;
297 	error = sysctl_handle_int(oidp, &clear, 0, req);
298 	if (error != 0 || req->newptr == NULL)
299 		return (error);
300 
301 	if (clear != 0) {
302 		for (i = 0; i < 2; ++i) {
303 			for (j = 0; j < UMTX_CHAINS; ++j) {
304 				uc = &umtxq_chains[i][j];
305 				mtx_lock(&uc->uc_lock);
306 				uc->length = 0;
307 				uc->max_length = 0;
308 				mtx_unlock(&uc->uc_lock);
309 			}
310 		}
311 	}
312 	return (0);
313 }
314 
315 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, clear,
316     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
317     sysctl_debug_umtx_chains_clear, "I",
318     "Clear umtx chains statistics");
319 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, peaks,
320     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
321     sysctl_debug_umtx_chains_peaks, "A",
322     "Highest peaks in chains max length");
323 #endif
324 
325 static void
326 umtxq_sysinit(void *arg __unused)
327 {
328 	int i, j;
329 
330 	umtx_pi_zone = uma_zcreate("umtx pi", sizeof(struct umtx_pi),
331 		NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
332 	for (i = 0; i < 2; ++i) {
333 		for (j = 0; j < UMTX_CHAINS; ++j) {
334 			mtx_init(&umtxq_chains[i][j].uc_lock, "umtxql", NULL,
335 				 MTX_DEF | MTX_DUPOK);
336 			LIST_INIT(&umtxq_chains[i][j].uc_queue[0]);
337 			LIST_INIT(&umtxq_chains[i][j].uc_queue[1]);
338 			LIST_INIT(&umtxq_chains[i][j].uc_spare_queue);
339 			TAILQ_INIT(&umtxq_chains[i][j].uc_pi_list);
340 			umtxq_chains[i][j].uc_busy = 0;
341 			umtxq_chains[i][j].uc_waiters = 0;
342 #ifdef UMTX_PROFILING
343 			umtxq_chains[i][j].length = 0;
344 			umtxq_chains[i][j].max_length = 0;
345 #endif
346 		}
347 	}
348 #ifdef UMTX_PROFILING
349 	umtx_init_profiling();
350 #endif
351 	mtx_init(&umtx_lock, "umtx lock", NULL, MTX_DEF);
352 	umtx_shm_init();
353 }
354 
355 struct umtx_q *
356 umtxq_alloc(void)
357 {
358 	struct umtx_q *uq;
359 
360 	uq = malloc(sizeof(struct umtx_q), M_UMTX, M_WAITOK | M_ZERO);
361 	uq->uq_spare_queue = malloc(sizeof(struct umtxq_queue), M_UMTX,
362 	    M_WAITOK | M_ZERO);
363 	TAILQ_INIT(&uq->uq_spare_queue->head);
364 	TAILQ_INIT(&uq->uq_pi_contested);
365 	uq->uq_inherited_pri = PRI_MAX;
366 	return (uq);
367 }
368 
369 void
370 umtxq_free(struct umtx_q *uq)
371 {
372 
373 	MPASS(uq->uq_spare_queue != NULL);
374 	free(uq->uq_spare_queue, M_UMTX);
375 	free(uq, M_UMTX);
376 }
377 
378 static inline void
379 umtxq_hash(struct umtx_key *key)
380 {
381 	unsigned n;
382 
383 	n = (uintptr_t)key->info.both.a + key->info.both.b;
384 	key->hash = ((n * GOLDEN_RATIO_PRIME) >> UMTX_SHIFTS) % UMTX_CHAINS;
385 }
386 
387 struct umtxq_chain *
388 umtxq_getchain(struct umtx_key *key)
389 {
390 
391 	if (key->type <= TYPE_SEM)
392 		return (&umtxq_chains[1][key->hash]);
393 	return (&umtxq_chains[0][key->hash]);
394 }
395 
396 /*
397  * Set chain to busy state when following operation
398  * may be blocked (kernel mutex can not be used).
399  */
400 void
401 umtxq_busy(struct umtx_key *key)
402 {
403 	struct umtxq_chain *uc;
404 
405 	uc = umtxq_getchain(key);
406 	mtx_assert(&uc->uc_lock, MA_OWNED);
407 	if (uc->uc_busy) {
408 #ifdef SMP
409 		if (smp_cpus > 1) {
410 			int count = BUSY_SPINS;
411 			if (count > 0) {
412 				umtxq_unlock(key);
413 				while (uc->uc_busy && --count > 0)
414 					cpu_spinwait();
415 				umtxq_lock(key);
416 			}
417 		}
418 #endif
419 		while (uc->uc_busy) {
420 			uc->uc_waiters++;
421 			msleep(uc, &uc->uc_lock, 0, "umtxqb", 0);
422 			uc->uc_waiters--;
423 		}
424 	}
425 	uc->uc_busy = 1;
426 }
427 
428 /*
429  * Unbusy a chain.
430  */
431 void
432 umtxq_unbusy(struct umtx_key *key)
433 {
434 	struct umtxq_chain *uc;
435 
436 	uc = umtxq_getchain(key);
437 	mtx_assert(&uc->uc_lock, MA_OWNED);
438 	KASSERT(uc->uc_busy != 0, ("not busy"));
439 	uc->uc_busy = 0;
440 	if (uc->uc_waiters)
441 		wakeup_one(uc);
442 }
443 
444 void
445 umtxq_unbusy_unlocked(struct umtx_key *key)
446 {
447 
448 	umtxq_lock(key);
449 	umtxq_unbusy(key);
450 	umtxq_unlock(key);
451 }
452 
453 static struct umtxq_queue *
454 umtxq_queue_lookup(struct umtx_key *key, int q)
455 {
456 	struct umtxq_queue *uh;
457 	struct umtxq_chain *uc;
458 
459 	uc = umtxq_getchain(key);
460 	UMTXQ_LOCKED_ASSERT(uc);
461 	LIST_FOREACH(uh, &uc->uc_queue[q], link) {
462 		if (umtx_key_match(&uh->key, key))
463 			return (uh);
464 	}
465 
466 	return (NULL);
467 }
468 
469 void
470 umtxq_insert_queue(struct umtx_q *uq, int q)
471 {
472 	struct umtxq_queue *uh;
473 	struct umtxq_chain *uc;
474 
475 	uc = umtxq_getchain(&uq->uq_key);
476 	UMTXQ_LOCKED_ASSERT(uc);
477 	KASSERT((uq->uq_flags & UQF_UMTXQ) == 0, ("umtx_q is already on queue"));
478 	uh = umtxq_queue_lookup(&uq->uq_key, q);
479 	if (uh != NULL) {
480 		LIST_INSERT_HEAD(&uc->uc_spare_queue, uq->uq_spare_queue, link);
481 	} else {
482 		uh = uq->uq_spare_queue;
483 		uh->key = uq->uq_key;
484 		LIST_INSERT_HEAD(&uc->uc_queue[q], uh, link);
485 #ifdef UMTX_PROFILING
486 		uc->length++;
487 		if (uc->length > uc->max_length) {
488 			uc->max_length = uc->length;
489 			if (uc->max_length > max_length)
490 				max_length = uc->max_length;
491 		}
492 #endif
493 	}
494 	uq->uq_spare_queue = NULL;
495 
496 	TAILQ_INSERT_TAIL(&uh->head, uq, uq_link);
497 	uh->length++;
498 	uq->uq_flags |= UQF_UMTXQ;
499 	uq->uq_cur_queue = uh;
500 	return;
501 }
502 
503 void
504 umtxq_remove_queue(struct umtx_q *uq, int q)
505 {
506 	struct umtxq_chain *uc;
507 	struct umtxq_queue *uh;
508 
509 	uc = umtxq_getchain(&uq->uq_key);
510 	UMTXQ_LOCKED_ASSERT(uc);
511 	if (uq->uq_flags & UQF_UMTXQ) {
512 		uh = uq->uq_cur_queue;
513 		TAILQ_REMOVE(&uh->head, uq, uq_link);
514 		uh->length--;
515 		uq->uq_flags &= ~UQF_UMTXQ;
516 		if (TAILQ_EMPTY(&uh->head)) {
517 			KASSERT(uh->length == 0,
518 			    ("inconsistent umtxq_queue length"));
519 #ifdef UMTX_PROFILING
520 			uc->length--;
521 #endif
522 			LIST_REMOVE(uh, link);
523 		} else {
524 			uh = LIST_FIRST(&uc->uc_spare_queue);
525 			KASSERT(uh != NULL, ("uc_spare_queue is empty"));
526 			LIST_REMOVE(uh, link);
527 		}
528 		uq->uq_spare_queue = uh;
529 		uq->uq_cur_queue = NULL;
530 	}
531 }
532 
533 /*
534  * Check if there are multiple waiters
535  */
536 int
537 umtxq_count(struct umtx_key *key)
538 {
539 	struct umtxq_queue *uh;
540 
541 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
542 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
543 	if (uh != NULL)
544 		return (uh->length);
545 	return (0);
546 }
547 
548 /*
549  * Check if there are multiple PI waiters and returns first
550  * waiter.
551  */
552 static int
553 umtxq_count_pi(struct umtx_key *key, struct umtx_q **first)
554 {
555 	struct umtxq_queue *uh;
556 
557 	*first = NULL;
558 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
559 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
560 	if (uh != NULL) {
561 		*first = TAILQ_FIRST(&uh->head);
562 		return (uh->length);
563 	}
564 	return (0);
565 }
566 
567 /*
568  * Wake up threads waiting on an userland object by a bit mask.
569  */
570 int
571 umtxq_signal_mask(struct umtx_key *key, int n_wake, u_int bitset)
572 {
573 	struct umtxq_queue *uh;
574 	struct umtx_q *uq, *uq_temp;
575 	int ret;
576 
577 	ret = 0;
578 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
579 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
580 	if (uh == NULL)
581 		return (0);
582 	TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
583 		if ((uq->uq_bitset & bitset) == 0)
584 			continue;
585 		umtxq_remove_queue(uq, UMTX_SHARED_QUEUE);
586 		wakeup_one(uq);
587 		if (++ret >= n_wake)
588 			break;
589 	}
590 	return (ret);
591 }
592 
593 /*
594  * Wake up threads waiting on an userland object.
595  */
596 
597 static int
598 umtxq_signal_queue(struct umtx_key *key, int n_wake, int q)
599 {
600 	struct umtxq_queue *uh;
601 	struct umtx_q *uq;
602 	int ret;
603 
604 	ret = 0;
605 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
606 	uh = umtxq_queue_lookup(key, q);
607 	if (uh != NULL) {
608 		while ((uq = TAILQ_FIRST(&uh->head)) != NULL) {
609 			umtxq_remove_queue(uq, q);
610 			wakeup(uq);
611 			if (++ret >= n_wake)
612 				return (ret);
613 		}
614 	}
615 	return (ret);
616 }
617 
618 /*
619  * Wake up specified thread.
620  */
621 static inline void
622 umtxq_signal_thread(struct umtx_q *uq)
623 {
624 
625 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
626 	umtxq_remove(uq);
627 	wakeup(uq);
628 }
629 
630 /*
631  * Wake up a maximum of n_wake threads that are waiting on an userland
632  * object identified by key. The remaining threads are removed from queue
633  * identified by key and added to the queue identified by key2 (requeued).
634  * The n_requeue specifies an upper limit on the number of threads that
635  * are requeued to the second queue.
636  */
637 int
638 umtxq_requeue(struct umtx_key *key, int n_wake, struct umtx_key *key2,
639     int n_requeue)
640 {
641 	struct umtxq_queue *uh;
642 	struct umtx_q *uq, *uq_temp;
643 	int ret;
644 
645 	ret = 0;
646 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
647 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key2));
648 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
649 	if (uh == NULL)
650 		return (0);
651 	TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
652 		if (++ret <= n_wake) {
653 			umtxq_remove(uq);
654 			wakeup_one(uq);
655 		} else {
656 			umtxq_remove(uq);
657 			uq->uq_key = *key2;
658 			umtxq_insert(uq);
659 			if (ret - n_wake == n_requeue)
660 				break;
661 		}
662 	}
663 	return (ret);
664 }
665 
666 static inline int
667 tstohz(const struct timespec *tsp)
668 {
669 	struct timeval tv;
670 
671 	TIMESPEC_TO_TIMEVAL(&tv, tsp);
672 	return tvtohz(&tv);
673 }
674 
675 void
676 umtx_abs_timeout_init(struct umtx_abs_timeout *timo, int clockid,
677     int absolute, const struct timespec *timeout)
678 {
679 
680 	timo->clockid = clockid;
681 	if (!absolute) {
682 		timo->is_abs_real = false;
683 		kern_clock_gettime(curthread, timo->clockid, &timo->cur);
684 		timespecadd(&timo->cur, timeout, &timo->end);
685 	} else {
686 		timo->end = *timeout;
687 		timo->is_abs_real = clockid == CLOCK_REALTIME ||
688 		    clockid == CLOCK_REALTIME_FAST ||
689 		    clockid == CLOCK_REALTIME_PRECISE ||
690 		    clockid == CLOCK_SECOND;
691 	}
692 }
693 
694 static void
695 umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
696     const struct _umtx_time *umtxtime)
697 {
698 
699 	umtx_abs_timeout_init(timo, umtxtime->_clockid,
700 	    (umtxtime->_flags & UMTX_ABSTIME) != 0, &umtxtime->_timeout);
701 }
702 
703 static int
704 umtx_abs_timeout_getsbt(struct umtx_abs_timeout *timo, sbintime_t *sbt,
705     int *flags)
706 {
707 	struct bintime bt, bbt;
708 	struct timespec tts;
709 	sbintime_t rem;
710 
711 	switch (timo->clockid) {
712 
713 	/* Clocks that can be converted into absolute time. */
714 	case CLOCK_REALTIME:
715 	case CLOCK_REALTIME_PRECISE:
716 	case CLOCK_REALTIME_FAST:
717 	case CLOCK_MONOTONIC:
718 	case CLOCK_MONOTONIC_PRECISE:
719 	case CLOCK_MONOTONIC_FAST:
720 	case CLOCK_UPTIME:
721 	case CLOCK_UPTIME_PRECISE:
722 	case CLOCK_UPTIME_FAST:
723 	case CLOCK_SECOND:
724 		timespec2bintime(&timo->end, &bt);
725 		switch (timo->clockid) {
726 		case CLOCK_REALTIME:
727 		case CLOCK_REALTIME_PRECISE:
728 		case CLOCK_REALTIME_FAST:
729 		case CLOCK_SECOND:
730 			getboottimebin(&bbt);
731 			bintime_sub(&bt, &bbt);
732 			break;
733 		}
734 		if (bt.sec < 0)
735 			return (ETIMEDOUT);
736 		if (bt.sec >= (SBT_MAX >> 32)) {
737 			*sbt = 0;
738 			*flags = 0;
739 			return (0);
740 		}
741 		*sbt = bttosbt(bt);
742 
743 		/*
744 		 * Check if the absolute time should be aligned to
745 		 * avoid firing multiple timer events in non-periodic
746 		 * timer mode.
747 		 */
748 		switch (timo->clockid) {
749 		case CLOCK_REALTIME_FAST:
750 		case CLOCK_MONOTONIC_FAST:
751 		case CLOCK_UPTIME_FAST:
752 			rem = *sbt % tc_tick_sbt;
753 			if (__predict_true(rem != 0))
754 				*sbt += tc_tick_sbt - rem;
755 			break;
756 		case CLOCK_SECOND:
757 			rem = *sbt % SBT_1S;
758 			if (__predict_true(rem != 0))
759 				*sbt += SBT_1S - rem;
760 			break;
761 		}
762 		*flags = C_ABSOLUTE;
763 		return (0);
764 
765 	/* Clocks that has to be periodically polled. */
766 	case CLOCK_VIRTUAL:
767 	case CLOCK_PROF:
768 	case CLOCK_THREAD_CPUTIME_ID:
769 	case CLOCK_PROCESS_CPUTIME_ID:
770 	default:
771 		kern_clock_gettime(curthread, timo->clockid, &timo->cur);
772 		if (timespeccmp(&timo->end, &timo->cur, <=))
773 			return (ETIMEDOUT);
774 		timespecsub(&timo->end, &timo->cur, &tts);
775 		*sbt = tick_sbt * tstohz(&tts);
776 		*flags = C_HARDCLOCK;
777 		return (0);
778 	}
779 }
780 
781 static uint32_t
782 umtx_unlock_val(uint32_t flags, bool rb)
783 {
784 
785 	if (rb)
786 		return (UMUTEX_RB_OWNERDEAD);
787 	else if ((flags & UMUTEX_NONCONSISTENT) != 0)
788 		return (UMUTEX_RB_NOTRECOV);
789 	else
790 		return (UMUTEX_UNOWNED);
791 
792 }
793 
794 /*
795  * Put thread into sleep state, before sleeping, check if
796  * thread was removed from umtx queue.
797  */
798 int
799 umtxq_sleep(struct umtx_q *uq, const char *wmesg,
800     struct umtx_abs_timeout *timo)
801 {
802 	struct umtxq_chain *uc;
803 	sbintime_t sbt = 0;
804 	int error, flags = 0;
805 
806 	uc = umtxq_getchain(&uq->uq_key);
807 	UMTXQ_LOCKED_ASSERT(uc);
808 	for (;;) {
809 		if (!(uq->uq_flags & UQF_UMTXQ)) {
810 			error = 0;
811 			break;
812 		}
813 		if (timo != NULL) {
814 			if (timo->is_abs_real)
815 				curthread->td_rtcgen =
816 				    atomic_load_acq_int(&rtc_generation);
817 			error = umtx_abs_timeout_getsbt(timo, &sbt, &flags);
818 			if (error != 0)
819 				break;
820 		}
821 		error = msleep_sbt(uq, &uc->uc_lock, PCATCH | PDROP, wmesg,
822 		    sbt, 0, flags);
823 		uc = umtxq_getchain(&uq->uq_key);
824 		mtx_lock(&uc->uc_lock);
825 		if (error == EINTR || error == ERESTART)
826 			break;
827 		if (error == EWOULDBLOCK && (flags & C_ABSOLUTE) != 0) {
828 			error = ETIMEDOUT;
829 			break;
830 		}
831 	}
832 
833 	curthread->td_rtcgen = 0;
834 	return (error);
835 }
836 
837 /*
838  * Convert userspace address into unique logical address.
839  */
840 int
841 umtx_key_get(const void *addr, int type, int share, struct umtx_key *key)
842 {
843 	struct thread *td = curthread;
844 	vm_map_t map;
845 	vm_map_entry_t entry;
846 	vm_pindex_t pindex;
847 	vm_prot_t prot;
848 	boolean_t wired;
849 
850 	key->type = type;
851 	if (share == THREAD_SHARE) {
852 		key->shared = 0;
853 		key->info.private.vs = td->td_proc->p_vmspace;
854 		key->info.private.addr = (uintptr_t)addr;
855 	} else {
856 		MPASS(share == PROCESS_SHARE || share == AUTO_SHARE);
857 		map = &td->td_proc->p_vmspace->vm_map;
858 		if (vm_map_lookup(&map, (vm_offset_t)addr, VM_PROT_WRITE,
859 		    &entry, &key->info.shared.object, &pindex, &prot,
860 		    &wired) != KERN_SUCCESS) {
861 			return (EFAULT);
862 		}
863 
864 		if ((share == PROCESS_SHARE) ||
865 		    (share == AUTO_SHARE &&
866 		     VM_INHERIT_SHARE == entry->inheritance)) {
867 			key->shared = 1;
868 			key->info.shared.offset = (vm_offset_t)addr -
869 			    entry->start + entry->offset;
870 			vm_object_reference(key->info.shared.object);
871 		} else {
872 			key->shared = 0;
873 			key->info.private.vs = td->td_proc->p_vmspace;
874 			key->info.private.addr = (uintptr_t)addr;
875 		}
876 		vm_map_lookup_done(map, entry);
877 	}
878 
879 	umtxq_hash(key);
880 	return (0);
881 }
882 
883 /*
884  * Release key.
885  */
886 void
887 umtx_key_release(struct umtx_key *key)
888 {
889 	if (key->shared)
890 		vm_object_deallocate(key->info.shared.object);
891 }
892 
893 #ifdef COMPAT_FREEBSD10
894 /*
895  * Lock a umtx object.
896  */
897 static int
898 do_lock_umtx(struct thread *td, struct umtx *umtx, u_long id,
899     const struct timespec *timeout)
900 {
901 	struct umtx_abs_timeout timo;
902 	struct umtx_q *uq;
903 	u_long owner;
904 	u_long old;
905 	int error = 0;
906 
907 	uq = td->td_umtxq;
908 	if (timeout != NULL)
909 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
910 
911 	/*
912 	 * Care must be exercised when dealing with umtx structure. It
913 	 * can fault on any access.
914 	 */
915 	for (;;) {
916 		/*
917 		 * Try the uncontested case.  This should be done in userland.
918 		 */
919 		owner = casuword(&umtx->u_owner, UMTX_UNOWNED, id);
920 
921 		/* The acquire succeeded. */
922 		if (owner == UMTX_UNOWNED)
923 			return (0);
924 
925 		/* The address was invalid. */
926 		if (owner == -1)
927 			return (EFAULT);
928 
929 		/* If no one owns it but it is contested try to acquire it. */
930 		if (owner == UMTX_CONTESTED) {
931 			owner = casuword(&umtx->u_owner,
932 			    UMTX_CONTESTED, id | UMTX_CONTESTED);
933 
934 			if (owner == UMTX_CONTESTED)
935 				return (0);
936 
937 			/* The address was invalid. */
938 			if (owner == -1)
939 				return (EFAULT);
940 
941 			error = thread_check_susp(td, false);
942 			if (error != 0)
943 				break;
944 
945 			/* If this failed the lock has changed, restart. */
946 			continue;
947 		}
948 
949 		/*
950 		 * If we caught a signal, we have retried and now
951 		 * exit immediately.
952 		 */
953 		if (error != 0)
954 			break;
955 
956 		if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK,
957 			AUTO_SHARE, &uq->uq_key)) != 0)
958 			return (error);
959 
960 		umtxq_lock(&uq->uq_key);
961 		umtxq_busy(&uq->uq_key);
962 		umtxq_insert(uq);
963 		umtxq_unbusy(&uq->uq_key);
964 		umtxq_unlock(&uq->uq_key);
965 
966 		/*
967 		 * Set the contested bit so that a release in user space
968 		 * knows to use the system call for unlock.  If this fails
969 		 * either some one else has acquired the lock or it has been
970 		 * released.
971 		 */
972 		old = casuword(&umtx->u_owner, owner, owner | UMTX_CONTESTED);
973 
974 		/* The address was invalid. */
975 		if (old == -1) {
976 			umtxq_lock(&uq->uq_key);
977 			umtxq_remove(uq);
978 			umtxq_unlock(&uq->uq_key);
979 			umtx_key_release(&uq->uq_key);
980 			return (EFAULT);
981 		}
982 
983 		/*
984 		 * We set the contested bit, sleep. Otherwise the lock changed
985 		 * and we need to retry or we lost a race to the thread
986 		 * unlocking the umtx.
987 		 */
988 		umtxq_lock(&uq->uq_key);
989 		if (old == owner)
990 			error = umtxq_sleep(uq, "umtx", timeout == NULL ? NULL :
991 			    &timo);
992 		umtxq_remove(uq);
993 		umtxq_unlock(&uq->uq_key);
994 		umtx_key_release(&uq->uq_key);
995 
996 		if (error == 0)
997 			error = thread_check_susp(td, false);
998 	}
999 
1000 	if (timeout == NULL) {
1001 		/* Mutex locking is restarted if it is interrupted. */
1002 		if (error == EINTR)
1003 			error = ERESTART;
1004 	} else {
1005 		/* Timed-locking is not restarted. */
1006 		if (error == ERESTART)
1007 			error = EINTR;
1008 	}
1009 	return (error);
1010 }
1011 
1012 /*
1013  * Unlock a umtx object.
1014  */
1015 static int
1016 do_unlock_umtx(struct thread *td, struct umtx *umtx, u_long id)
1017 {
1018 	struct umtx_key key;
1019 	u_long owner;
1020 	u_long old;
1021 	int error;
1022 	int count;
1023 
1024 	/*
1025 	 * Make sure we own this mtx.
1026 	 */
1027 	owner = fuword(__DEVOLATILE(u_long *, &umtx->u_owner));
1028 	if (owner == -1)
1029 		return (EFAULT);
1030 
1031 	if ((owner & ~UMTX_CONTESTED) != id)
1032 		return (EPERM);
1033 
1034 	/* This should be done in userland */
1035 	if ((owner & UMTX_CONTESTED) == 0) {
1036 		old = casuword(&umtx->u_owner, owner, UMTX_UNOWNED);
1037 		if (old == -1)
1038 			return (EFAULT);
1039 		if (old == owner)
1040 			return (0);
1041 		owner = old;
1042 	}
1043 
1044 	/* We should only ever be in here for contested locks */
1045 	if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1046 	    &key)) != 0)
1047 		return (error);
1048 
1049 	umtxq_lock(&key);
1050 	umtxq_busy(&key);
1051 	count = umtxq_count(&key);
1052 	umtxq_unlock(&key);
1053 
1054 	/*
1055 	 * When unlocking the umtx, it must be marked as unowned if
1056 	 * there is zero or one thread only waiting for it.
1057 	 * Otherwise, it must be marked as contested.
1058 	 */
1059 	old = casuword(&umtx->u_owner, owner,
1060 	    count <= 1 ? UMTX_UNOWNED : UMTX_CONTESTED);
1061 	umtxq_lock(&key);
1062 	umtxq_signal(&key,1);
1063 	umtxq_unbusy(&key);
1064 	umtxq_unlock(&key);
1065 	umtx_key_release(&key);
1066 	if (old == -1)
1067 		return (EFAULT);
1068 	if (old != owner)
1069 		return (EINVAL);
1070 	return (0);
1071 }
1072 
1073 #ifdef COMPAT_FREEBSD32
1074 
1075 /*
1076  * Lock a umtx object.
1077  */
1078 static int
1079 do_lock_umtx32(struct thread *td, uint32_t *m, uint32_t id,
1080 	const struct timespec *timeout)
1081 {
1082 	struct umtx_abs_timeout timo;
1083 	struct umtx_q *uq;
1084 	uint32_t owner;
1085 	uint32_t old;
1086 	int error = 0;
1087 
1088 	uq = td->td_umtxq;
1089 
1090 	if (timeout != NULL)
1091 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
1092 
1093 	/*
1094 	 * Care must be exercised when dealing with umtx structure. It
1095 	 * can fault on any access.
1096 	 */
1097 	for (;;) {
1098 		/*
1099 		 * Try the uncontested case.  This should be done in userland.
1100 		 */
1101 		owner = casuword32(m, UMUTEX_UNOWNED, id);
1102 
1103 		/* The acquire succeeded. */
1104 		if (owner == UMUTEX_UNOWNED)
1105 			return (0);
1106 
1107 		/* The address was invalid. */
1108 		if (owner == -1)
1109 			return (EFAULT);
1110 
1111 		/* If no one owns it but it is contested try to acquire it. */
1112 		if (owner == UMUTEX_CONTESTED) {
1113 			owner = casuword32(m,
1114 			    UMUTEX_CONTESTED, id | UMUTEX_CONTESTED);
1115 			if (owner == UMUTEX_CONTESTED)
1116 				return (0);
1117 
1118 			/* The address was invalid. */
1119 			if (owner == -1)
1120 				return (EFAULT);
1121 
1122 			error = thread_check_susp(td, false);
1123 			if (error != 0)
1124 				break;
1125 
1126 			/* If this failed the lock has changed, restart. */
1127 			continue;
1128 		}
1129 
1130 		/*
1131 		 * If we caught a signal, we have retried and now
1132 		 * exit immediately.
1133 		 */
1134 		if (error != 0)
1135 			return (error);
1136 
1137 		if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK,
1138 			AUTO_SHARE, &uq->uq_key)) != 0)
1139 			return (error);
1140 
1141 		umtxq_lock(&uq->uq_key);
1142 		umtxq_busy(&uq->uq_key);
1143 		umtxq_insert(uq);
1144 		umtxq_unbusy(&uq->uq_key);
1145 		umtxq_unlock(&uq->uq_key);
1146 
1147 		/*
1148 		 * Set the contested bit so that a release in user space
1149 		 * knows to use the system call for unlock.  If this fails
1150 		 * either some one else has acquired the lock or it has been
1151 		 * released.
1152 		 */
1153 		old = casuword32(m, owner, owner | UMUTEX_CONTESTED);
1154 
1155 		/* The address was invalid. */
1156 		if (old == -1) {
1157 			umtxq_lock(&uq->uq_key);
1158 			umtxq_remove(uq);
1159 			umtxq_unlock(&uq->uq_key);
1160 			umtx_key_release(&uq->uq_key);
1161 			return (EFAULT);
1162 		}
1163 
1164 		/*
1165 		 * We set the contested bit, sleep. Otherwise the lock changed
1166 		 * and we need to retry or we lost a race to the thread
1167 		 * unlocking the umtx.
1168 		 */
1169 		umtxq_lock(&uq->uq_key);
1170 		if (old == owner)
1171 			error = umtxq_sleep(uq, "umtx", timeout == NULL ?
1172 			    NULL : &timo);
1173 		umtxq_remove(uq);
1174 		umtxq_unlock(&uq->uq_key);
1175 		umtx_key_release(&uq->uq_key);
1176 
1177 		if (error == 0)
1178 			error = thread_check_susp(td, false);
1179 	}
1180 
1181 	if (timeout == NULL) {
1182 		/* Mutex locking is restarted if it is interrupted. */
1183 		if (error == EINTR)
1184 			error = ERESTART;
1185 	} else {
1186 		/* Timed-locking is not restarted. */
1187 		if (error == ERESTART)
1188 			error = EINTR;
1189 	}
1190 	return (error);
1191 }
1192 
1193 /*
1194  * Unlock a umtx object.
1195  */
1196 static int
1197 do_unlock_umtx32(struct thread *td, uint32_t *m, uint32_t id)
1198 {
1199 	struct umtx_key key;
1200 	uint32_t owner;
1201 	uint32_t old;
1202 	int error;
1203 	int count;
1204 
1205 	/*
1206 	 * Make sure we own this mtx.
1207 	 */
1208 	owner = fuword32(m);
1209 	if (owner == -1)
1210 		return (EFAULT);
1211 
1212 	if ((owner & ~UMUTEX_CONTESTED) != id)
1213 		return (EPERM);
1214 
1215 	/* This should be done in userland */
1216 	if ((owner & UMUTEX_CONTESTED) == 0) {
1217 		old = casuword32(m, owner, UMUTEX_UNOWNED);
1218 		if (old == -1)
1219 			return (EFAULT);
1220 		if (old == owner)
1221 			return (0);
1222 		owner = old;
1223 	}
1224 
1225 	/* We should only ever be in here for contested locks */
1226 	if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1227 		&key)) != 0)
1228 		return (error);
1229 
1230 	umtxq_lock(&key);
1231 	umtxq_busy(&key);
1232 	count = umtxq_count(&key);
1233 	umtxq_unlock(&key);
1234 
1235 	/*
1236 	 * When unlocking the umtx, it must be marked as unowned if
1237 	 * there is zero or one thread only waiting for it.
1238 	 * Otherwise, it must be marked as contested.
1239 	 */
1240 	old = casuword32(m, owner,
1241 		count <= 1 ? UMUTEX_UNOWNED : UMUTEX_CONTESTED);
1242 	umtxq_lock(&key);
1243 	umtxq_signal(&key,1);
1244 	umtxq_unbusy(&key);
1245 	umtxq_unlock(&key);
1246 	umtx_key_release(&key);
1247 	if (old == -1)
1248 		return (EFAULT);
1249 	if (old != owner)
1250 		return (EINVAL);
1251 	return (0);
1252 }
1253 #endif	/* COMPAT_FREEBSD32 */
1254 #endif	/* COMPAT_FREEBSD10 */
1255 
1256 /*
1257  * Fetch and compare value, sleep on the address if value is not changed.
1258  */
1259 static int
1260 do_wait(struct thread *td, void *addr, u_long id,
1261     struct _umtx_time *timeout, int compat32, int is_private)
1262 {
1263 	struct umtx_abs_timeout timo;
1264 	struct umtx_q *uq;
1265 	u_long tmp;
1266 	uint32_t tmp32;
1267 	int error = 0;
1268 
1269 	uq = td->td_umtxq;
1270 	if ((error = umtx_key_get(addr, TYPE_SIMPLE_WAIT,
1271 		is_private ? THREAD_SHARE : AUTO_SHARE, &uq->uq_key)) != 0)
1272 		return (error);
1273 
1274 	if (timeout != NULL)
1275 		umtx_abs_timeout_init2(&timo, timeout);
1276 
1277 	umtxq_lock(&uq->uq_key);
1278 	umtxq_insert(uq);
1279 	umtxq_unlock(&uq->uq_key);
1280 	if (compat32 == 0) {
1281 		error = fueword(addr, &tmp);
1282 		if (error != 0)
1283 			error = EFAULT;
1284 	} else {
1285 		error = fueword32(addr, &tmp32);
1286 		if (error == 0)
1287 			tmp = tmp32;
1288 		else
1289 			error = EFAULT;
1290 	}
1291 	umtxq_lock(&uq->uq_key);
1292 	if (error == 0) {
1293 		if (tmp == id)
1294 			error = umtxq_sleep(uq, "uwait", timeout == NULL ?
1295 			    NULL : &timo);
1296 		if ((uq->uq_flags & UQF_UMTXQ) == 0)
1297 			error = 0;
1298 		else
1299 			umtxq_remove(uq);
1300 	} else if ((uq->uq_flags & UQF_UMTXQ) != 0) {
1301 		umtxq_remove(uq);
1302 	}
1303 	umtxq_unlock(&uq->uq_key);
1304 	umtx_key_release(&uq->uq_key);
1305 	if (error == ERESTART)
1306 		error = EINTR;
1307 	return (error);
1308 }
1309 
1310 /*
1311  * Wake up threads sleeping on the specified address.
1312  */
1313 int
1314 kern_umtx_wake(struct thread *td, void *uaddr, int n_wake, int is_private)
1315 {
1316 	struct umtx_key key;
1317 	int ret;
1318 
1319 	if ((ret = umtx_key_get(uaddr, TYPE_SIMPLE_WAIT,
1320 	    is_private ? THREAD_SHARE : AUTO_SHARE, &key)) != 0)
1321 		return (ret);
1322 	umtxq_lock(&key);
1323 	umtxq_signal(&key, n_wake);
1324 	umtxq_unlock(&key);
1325 	umtx_key_release(&key);
1326 	return (0);
1327 }
1328 
1329 /*
1330  * Lock PTHREAD_PRIO_NONE protocol POSIX mutex.
1331  */
1332 static int
1333 do_lock_normal(struct thread *td, struct umutex *m, uint32_t flags,
1334     struct _umtx_time *timeout, int mode)
1335 {
1336 	struct umtx_abs_timeout timo;
1337 	struct umtx_q *uq;
1338 	uint32_t owner, old, id;
1339 	int error, rv;
1340 
1341 	id = td->td_tid;
1342 	uq = td->td_umtxq;
1343 	error = 0;
1344 	if (timeout != NULL)
1345 		umtx_abs_timeout_init2(&timo, timeout);
1346 
1347 	/*
1348 	 * Care must be exercised when dealing with umtx structure. It
1349 	 * can fault on any access.
1350 	 */
1351 	for (;;) {
1352 		rv = fueword32(&m->m_owner, &owner);
1353 		if (rv == -1)
1354 			return (EFAULT);
1355 		if (mode == _UMUTEX_WAIT) {
1356 			if (owner == UMUTEX_UNOWNED ||
1357 			    owner == UMUTEX_CONTESTED ||
1358 			    owner == UMUTEX_RB_OWNERDEAD ||
1359 			    owner == UMUTEX_RB_NOTRECOV)
1360 				return (0);
1361 		} else {
1362 			/*
1363 			 * Robust mutex terminated.  Kernel duty is to
1364 			 * return EOWNERDEAD to the userspace.  The
1365 			 * umutex.m_flags UMUTEX_NONCONSISTENT is set
1366 			 * by the common userspace code.
1367 			 */
1368 			if (owner == UMUTEX_RB_OWNERDEAD) {
1369 				rv = casueword32(&m->m_owner,
1370 				    UMUTEX_RB_OWNERDEAD, &owner,
1371 				    id | UMUTEX_CONTESTED);
1372 				if (rv == -1)
1373 					return (EFAULT);
1374 				if (rv == 0) {
1375 					MPASS(owner == UMUTEX_RB_OWNERDEAD);
1376 					return (EOWNERDEAD); /* success */
1377 				}
1378 				MPASS(rv == 1);
1379 				rv = thread_check_susp(td, false);
1380 				if (rv != 0)
1381 					return (rv);
1382 				continue;
1383 			}
1384 			if (owner == UMUTEX_RB_NOTRECOV)
1385 				return (ENOTRECOVERABLE);
1386 
1387 			/*
1388 			 * Try the uncontested case.  This should be
1389 			 * done in userland.
1390 			 */
1391 			rv = casueword32(&m->m_owner, UMUTEX_UNOWNED,
1392 			    &owner, id);
1393 			/* The address was invalid. */
1394 			if (rv == -1)
1395 				return (EFAULT);
1396 
1397 			/* The acquire succeeded. */
1398 			if (rv == 0) {
1399 				MPASS(owner == UMUTEX_UNOWNED);
1400 				return (0);
1401 			}
1402 
1403 			/*
1404 			 * If no one owns it but it is contested try
1405 			 * to acquire it.
1406 			 */
1407 			MPASS(rv == 1);
1408 			if (owner == UMUTEX_CONTESTED) {
1409 				rv = casueword32(&m->m_owner,
1410 				    UMUTEX_CONTESTED, &owner,
1411 				    id | UMUTEX_CONTESTED);
1412 				/* The address was invalid. */
1413 				if (rv == -1)
1414 					return (EFAULT);
1415 				if (rv == 0) {
1416 					MPASS(owner == UMUTEX_CONTESTED);
1417 					return (0);
1418 				}
1419 				if (rv == 1) {
1420 					rv = thread_check_susp(td, false);
1421 					if (rv != 0)
1422 						return (rv);
1423 				}
1424 
1425 				/*
1426 				 * If this failed the lock has
1427 				 * changed, restart.
1428 				 */
1429 				continue;
1430 			}
1431 
1432 			/* rv == 1 but not contested, likely store failure */
1433 			rv = thread_check_susp(td, false);
1434 			if (rv != 0)
1435 				return (rv);
1436 		}
1437 
1438 		if (mode == _UMUTEX_TRY)
1439 			return (EBUSY);
1440 
1441 		/*
1442 		 * If we caught a signal, we have retried and now
1443 		 * exit immediately.
1444 		 */
1445 		if (error != 0)
1446 			return (error);
1447 
1448 		if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX,
1449 		    GET_SHARE(flags), &uq->uq_key)) != 0)
1450 			return (error);
1451 
1452 		umtxq_lock(&uq->uq_key);
1453 		umtxq_busy(&uq->uq_key);
1454 		umtxq_insert(uq);
1455 		umtxq_unlock(&uq->uq_key);
1456 
1457 		/*
1458 		 * Set the contested bit so that a release in user space
1459 		 * knows to use the system call for unlock.  If this fails
1460 		 * either some one else has acquired the lock or it has been
1461 		 * released.
1462 		 */
1463 		rv = casueword32(&m->m_owner, owner, &old,
1464 		    owner | UMUTEX_CONTESTED);
1465 
1466 		/* The address was invalid or casueword failed to store. */
1467 		if (rv == -1 || rv == 1) {
1468 			umtxq_lock(&uq->uq_key);
1469 			umtxq_remove(uq);
1470 			umtxq_unbusy(&uq->uq_key);
1471 			umtxq_unlock(&uq->uq_key);
1472 			umtx_key_release(&uq->uq_key);
1473 			if (rv == -1)
1474 				return (EFAULT);
1475 			if (rv == 1) {
1476 				rv = thread_check_susp(td, false);
1477 				if (rv != 0)
1478 					return (rv);
1479 			}
1480 			continue;
1481 		}
1482 
1483 		/*
1484 		 * We set the contested bit, sleep. Otherwise the lock changed
1485 		 * and we need to retry or we lost a race to the thread
1486 		 * unlocking the umtx.
1487 		 */
1488 		umtxq_lock(&uq->uq_key);
1489 		umtxq_unbusy(&uq->uq_key);
1490 		MPASS(old == owner);
1491 		error = umtxq_sleep(uq, "umtxn", timeout == NULL ?
1492 		    NULL : &timo);
1493 		umtxq_remove(uq);
1494 		umtxq_unlock(&uq->uq_key);
1495 		umtx_key_release(&uq->uq_key);
1496 
1497 		if (error == 0)
1498 			error = thread_check_susp(td, false);
1499 	}
1500 
1501 	return (0);
1502 }
1503 
1504 /*
1505  * Unlock PTHREAD_PRIO_NONE protocol POSIX mutex.
1506  */
1507 static int
1508 do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1509 {
1510 	struct umtx_key key;
1511 	uint32_t owner, old, id, newlock;
1512 	int error, count;
1513 
1514 	id = td->td_tid;
1515 
1516 again:
1517 	/*
1518 	 * Make sure we own this mtx.
1519 	 */
1520 	error = fueword32(&m->m_owner, &owner);
1521 	if (error == -1)
1522 		return (EFAULT);
1523 
1524 	if ((owner & ~UMUTEX_CONTESTED) != id)
1525 		return (EPERM);
1526 
1527 	newlock = umtx_unlock_val(flags, rb);
1528 	if ((owner & UMUTEX_CONTESTED) == 0) {
1529 		error = casueword32(&m->m_owner, owner, &old, newlock);
1530 		if (error == -1)
1531 			return (EFAULT);
1532 		if (error == 1) {
1533 			error = thread_check_susp(td, false);
1534 			if (error != 0)
1535 				return (error);
1536 			goto again;
1537 		}
1538 		MPASS(old == owner);
1539 		return (0);
1540 	}
1541 
1542 	/* We should only ever be in here for contested locks */
1543 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1544 	    &key)) != 0)
1545 		return (error);
1546 
1547 	umtxq_lock(&key);
1548 	umtxq_busy(&key);
1549 	count = umtxq_count(&key);
1550 	umtxq_unlock(&key);
1551 
1552 	/*
1553 	 * When unlocking the umtx, it must be marked as unowned if
1554 	 * there is zero or one thread only waiting for it.
1555 	 * Otherwise, it must be marked as contested.
1556 	 */
1557 	if (count > 1)
1558 		newlock |= UMUTEX_CONTESTED;
1559 	error = casueword32(&m->m_owner, owner, &old, newlock);
1560 	umtxq_lock(&key);
1561 	umtxq_signal(&key, 1);
1562 	umtxq_unbusy(&key);
1563 	umtxq_unlock(&key);
1564 	umtx_key_release(&key);
1565 	if (error == -1)
1566 		return (EFAULT);
1567 	if (error == 1) {
1568 		if (old != owner)
1569 			return (EINVAL);
1570 		error = thread_check_susp(td, false);
1571 		if (error != 0)
1572 			return (error);
1573 		goto again;
1574 	}
1575 	return (0);
1576 }
1577 
1578 /*
1579  * Check if the mutex is available and wake up a waiter,
1580  * only for simple mutex.
1581  */
1582 static int
1583 do_wake_umutex(struct thread *td, struct umutex *m)
1584 {
1585 	struct umtx_key key;
1586 	uint32_t owner;
1587 	uint32_t flags;
1588 	int error;
1589 	int count;
1590 
1591 again:
1592 	error = fueword32(&m->m_owner, &owner);
1593 	if (error == -1)
1594 		return (EFAULT);
1595 
1596 	if ((owner & ~UMUTEX_CONTESTED) != 0 && owner != UMUTEX_RB_OWNERDEAD &&
1597 	    owner != UMUTEX_RB_NOTRECOV)
1598 		return (0);
1599 
1600 	error = fueword32(&m->m_flags, &flags);
1601 	if (error == -1)
1602 		return (EFAULT);
1603 
1604 	/* We should only ever be in here for contested locks */
1605 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1606 	    &key)) != 0)
1607 		return (error);
1608 
1609 	umtxq_lock(&key);
1610 	umtxq_busy(&key);
1611 	count = umtxq_count(&key);
1612 	umtxq_unlock(&key);
1613 
1614 	if (count <= 1 && owner != UMUTEX_RB_OWNERDEAD &&
1615 	    owner != UMUTEX_RB_NOTRECOV) {
1616 		error = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
1617 		    UMUTEX_UNOWNED);
1618 		if (error == -1) {
1619 			error = EFAULT;
1620 		} else if (error == 1) {
1621 			umtxq_lock(&key);
1622 			umtxq_unbusy(&key);
1623 			umtxq_unlock(&key);
1624 			umtx_key_release(&key);
1625 			error = thread_check_susp(td, false);
1626 			if (error != 0)
1627 				return (error);
1628 			goto again;
1629 		}
1630 	}
1631 
1632 	umtxq_lock(&key);
1633 	if (error == 0 && count != 0) {
1634 		MPASS((owner & ~UMUTEX_CONTESTED) == 0 ||
1635 		    owner == UMUTEX_RB_OWNERDEAD ||
1636 		    owner == UMUTEX_RB_NOTRECOV);
1637 		umtxq_signal(&key, 1);
1638 	}
1639 	umtxq_unbusy(&key);
1640 	umtxq_unlock(&key);
1641 	umtx_key_release(&key);
1642 	return (error);
1643 }
1644 
1645 /*
1646  * Check if the mutex has waiters and tries to fix contention bit.
1647  */
1648 static int
1649 do_wake2_umutex(struct thread *td, struct umutex *m, uint32_t flags)
1650 {
1651 	struct umtx_key key;
1652 	uint32_t owner, old;
1653 	int type;
1654 	int error;
1655 	int count;
1656 
1657 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT |
1658 	    UMUTEX_ROBUST)) {
1659 	case 0:
1660 	case UMUTEX_ROBUST:
1661 		type = TYPE_NORMAL_UMUTEX;
1662 		break;
1663 	case UMUTEX_PRIO_INHERIT:
1664 		type = TYPE_PI_UMUTEX;
1665 		break;
1666 	case (UMUTEX_PRIO_INHERIT | UMUTEX_ROBUST):
1667 		type = TYPE_PI_ROBUST_UMUTEX;
1668 		break;
1669 	case UMUTEX_PRIO_PROTECT:
1670 		type = TYPE_PP_UMUTEX;
1671 		break;
1672 	case (UMUTEX_PRIO_PROTECT | UMUTEX_ROBUST):
1673 		type = TYPE_PP_ROBUST_UMUTEX;
1674 		break;
1675 	default:
1676 		return (EINVAL);
1677 	}
1678 	if ((error = umtx_key_get(m, type, GET_SHARE(flags), &key)) != 0)
1679 		return (error);
1680 
1681 	owner = 0;
1682 	umtxq_lock(&key);
1683 	umtxq_busy(&key);
1684 	count = umtxq_count(&key);
1685 	umtxq_unlock(&key);
1686 
1687 	error = fueword32(&m->m_owner, &owner);
1688 	if (error == -1)
1689 		error = EFAULT;
1690 
1691 	/*
1692 	 * Only repair contention bit if there is a waiter, this means
1693 	 * the mutex is still being referenced by userland code,
1694 	 * otherwise don't update any memory.
1695 	 */
1696 	while (error == 0 && (owner & UMUTEX_CONTESTED) == 0 &&
1697 	    (count > 1 || (count == 1 && (owner & ~UMUTEX_CONTESTED) != 0))) {
1698 		error = casueword32(&m->m_owner, owner, &old,
1699 		    owner | UMUTEX_CONTESTED);
1700 		if (error == -1) {
1701 			error = EFAULT;
1702 			break;
1703 		}
1704 		if (error == 0) {
1705 			MPASS(old == owner);
1706 			break;
1707 		}
1708 		owner = old;
1709 		error = thread_check_susp(td, false);
1710 	}
1711 
1712 	umtxq_lock(&key);
1713 	if (error == EFAULT) {
1714 		umtxq_signal(&key, INT_MAX);
1715 	} else if (count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1716 	    owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1717 		umtxq_signal(&key, 1);
1718 	umtxq_unbusy(&key);
1719 	umtxq_unlock(&key);
1720 	umtx_key_release(&key);
1721 	return (error);
1722 }
1723 
1724 struct umtx_pi *
1725 umtx_pi_alloc(int flags)
1726 {
1727 	struct umtx_pi *pi;
1728 
1729 	pi = uma_zalloc(umtx_pi_zone, M_ZERO | flags);
1730 	TAILQ_INIT(&pi->pi_blocked);
1731 	atomic_add_int(&umtx_pi_allocated, 1);
1732 	return (pi);
1733 }
1734 
1735 void
1736 umtx_pi_free(struct umtx_pi *pi)
1737 {
1738 	uma_zfree(umtx_pi_zone, pi);
1739 	atomic_add_int(&umtx_pi_allocated, -1);
1740 }
1741 
1742 /*
1743  * Adjust the thread's position on a pi_state after its priority has been
1744  * changed.
1745  */
1746 static int
1747 umtx_pi_adjust_thread(struct umtx_pi *pi, struct thread *td)
1748 {
1749 	struct umtx_q *uq, *uq1, *uq2;
1750 	struct thread *td1;
1751 
1752 	mtx_assert(&umtx_lock, MA_OWNED);
1753 	if (pi == NULL)
1754 		return (0);
1755 
1756 	uq = td->td_umtxq;
1757 
1758 	/*
1759 	 * Check if the thread needs to be moved on the blocked chain.
1760 	 * It needs to be moved if either its priority is lower than
1761 	 * the previous thread or higher than the next thread.
1762 	 */
1763 	uq1 = TAILQ_PREV(uq, umtxq_head, uq_lockq);
1764 	uq2 = TAILQ_NEXT(uq, uq_lockq);
1765 	if ((uq1 != NULL && UPRI(td) < UPRI(uq1->uq_thread)) ||
1766 	    (uq2 != NULL && UPRI(td) > UPRI(uq2->uq_thread))) {
1767 		/*
1768 		 * Remove thread from blocked chain and determine where
1769 		 * it should be moved to.
1770 		 */
1771 		TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1772 		TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1773 			td1 = uq1->uq_thread;
1774 			MPASS(td1->td_proc->p_magic == P_MAGIC);
1775 			if (UPRI(td1) > UPRI(td))
1776 				break;
1777 		}
1778 
1779 		if (uq1 == NULL)
1780 			TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1781 		else
1782 			TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1783 	}
1784 	return (1);
1785 }
1786 
1787 static struct umtx_pi *
1788 umtx_pi_next(struct umtx_pi *pi)
1789 {
1790 	struct umtx_q *uq_owner;
1791 
1792 	if (pi->pi_owner == NULL)
1793 		return (NULL);
1794 	uq_owner = pi->pi_owner->td_umtxq;
1795 	if (uq_owner == NULL)
1796 		return (NULL);
1797 	return (uq_owner->uq_pi_blocked);
1798 }
1799 
1800 /*
1801  * Floyd's Cycle-Finding Algorithm.
1802  */
1803 static bool
1804 umtx_pi_check_loop(struct umtx_pi *pi)
1805 {
1806 	struct umtx_pi *pi1;	/* fast iterator */
1807 
1808 	mtx_assert(&umtx_lock, MA_OWNED);
1809 	if (pi == NULL)
1810 		return (false);
1811 	pi1 = pi;
1812 	for (;;) {
1813 		pi = umtx_pi_next(pi);
1814 		if (pi == NULL)
1815 			break;
1816 		pi1 = umtx_pi_next(pi1);
1817 		if (pi1 == NULL)
1818 			break;
1819 		pi1 = umtx_pi_next(pi1);
1820 		if (pi1 == NULL)
1821 			break;
1822 		if (pi == pi1)
1823 			return (true);
1824 	}
1825 	return (false);
1826 }
1827 
1828 /*
1829  * Propagate priority when a thread is blocked on POSIX
1830  * PI mutex.
1831  */
1832 static void
1833 umtx_propagate_priority(struct thread *td)
1834 {
1835 	struct umtx_q *uq;
1836 	struct umtx_pi *pi;
1837 	int pri;
1838 
1839 	mtx_assert(&umtx_lock, MA_OWNED);
1840 	pri = UPRI(td);
1841 	uq = td->td_umtxq;
1842 	pi = uq->uq_pi_blocked;
1843 	if (pi == NULL)
1844 		return;
1845 	if (umtx_pi_check_loop(pi))
1846 		return;
1847 
1848 	for (;;) {
1849 		td = pi->pi_owner;
1850 		if (td == NULL || td == curthread)
1851 			return;
1852 
1853 		MPASS(td->td_proc != NULL);
1854 		MPASS(td->td_proc->p_magic == P_MAGIC);
1855 
1856 		thread_lock(td);
1857 		if (td->td_lend_user_pri > pri)
1858 			sched_lend_user_prio(td, pri);
1859 		else {
1860 			thread_unlock(td);
1861 			break;
1862 		}
1863 		thread_unlock(td);
1864 
1865 		/*
1866 		 * Pick up the lock that td is blocked on.
1867 		 */
1868 		uq = td->td_umtxq;
1869 		pi = uq->uq_pi_blocked;
1870 		if (pi == NULL)
1871 			break;
1872 		/* Resort td on the list if needed. */
1873 		umtx_pi_adjust_thread(pi, td);
1874 	}
1875 }
1876 
1877 /*
1878  * Unpropagate priority for a PI mutex when a thread blocked on
1879  * it is interrupted by signal or resumed by others.
1880  */
1881 static void
1882 umtx_repropagate_priority(struct umtx_pi *pi)
1883 {
1884 	struct umtx_q *uq, *uq_owner;
1885 	struct umtx_pi *pi2;
1886 	int pri;
1887 
1888 	mtx_assert(&umtx_lock, MA_OWNED);
1889 
1890 	if (umtx_pi_check_loop(pi))
1891 		return;
1892 	while (pi != NULL && pi->pi_owner != NULL) {
1893 		pri = PRI_MAX;
1894 		uq_owner = pi->pi_owner->td_umtxq;
1895 
1896 		TAILQ_FOREACH(pi2, &uq_owner->uq_pi_contested, pi_link) {
1897 			uq = TAILQ_FIRST(&pi2->pi_blocked);
1898 			if (uq != NULL) {
1899 				if (pri > UPRI(uq->uq_thread))
1900 					pri = UPRI(uq->uq_thread);
1901 			}
1902 		}
1903 
1904 		if (pri > uq_owner->uq_inherited_pri)
1905 			pri = uq_owner->uq_inherited_pri;
1906 		thread_lock(pi->pi_owner);
1907 		sched_lend_user_prio(pi->pi_owner, pri);
1908 		thread_unlock(pi->pi_owner);
1909 		if ((pi = uq_owner->uq_pi_blocked) != NULL)
1910 			umtx_pi_adjust_thread(pi, uq_owner->uq_thread);
1911 	}
1912 }
1913 
1914 /*
1915  * Insert a PI mutex into owned list.
1916  */
1917 static void
1918 umtx_pi_setowner(struct umtx_pi *pi, struct thread *owner)
1919 {
1920 	struct umtx_q *uq_owner;
1921 
1922 	uq_owner = owner->td_umtxq;
1923 	mtx_assert(&umtx_lock, MA_OWNED);
1924 	MPASS(pi->pi_owner == NULL);
1925 	pi->pi_owner = owner;
1926 	TAILQ_INSERT_TAIL(&uq_owner->uq_pi_contested, pi, pi_link);
1927 }
1928 
1929 /*
1930  * Disown a PI mutex, and remove it from the owned list.
1931  */
1932 static void
1933 umtx_pi_disown(struct umtx_pi *pi)
1934 {
1935 
1936 	mtx_assert(&umtx_lock, MA_OWNED);
1937 	TAILQ_REMOVE(&pi->pi_owner->td_umtxq->uq_pi_contested, pi, pi_link);
1938 	pi->pi_owner = NULL;
1939 }
1940 
1941 /*
1942  * Claim ownership of a PI mutex.
1943  */
1944 int
1945 umtx_pi_claim(struct umtx_pi *pi, struct thread *owner)
1946 {
1947 	struct umtx_q *uq;
1948 	int pri;
1949 
1950 	mtx_lock(&umtx_lock);
1951 	if (pi->pi_owner == owner) {
1952 		mtx_unlock(&umtx_lock);
1953 		return (0);
1954 	}
1955 
1956 	if (pi->pi_owner != NULL) {
1957 		/*
1958 		 * userland may have already messed the mutex, sigh.
1959 		 */
1960 		mtx_unlock(&umtx_lock);
1961 		return (EPERM);
1962 	}
1963 	umtx_pi_setowner(pi, owner);
1964 	uq = TAILQ_FIRST(&pi->pi_blocked);
1965 	if (uq != NULL) {
1966 		pri = UPRI(uq->uq_thread);
1967 		thread_lock(owner);
1968 		if (pri < UPRI(owner))
1969 			sched_lend_user_prio(owner, pri);
1970 		thread_unlock(owner);
1971 	}
1972 	mtx_unlock(&umtx_lock);
1973 	return (0);
1974 }
1975 
1976 /*
1977  * Adjust a thread's order position in its blocked PI mutex,
1978  * this may result new priority propagating process.
1979  */
1980 void
1981 umtx_pi_adjust(struct thread *td, u_char oldpri)
1982 {
1983 	struct umtx_q *uq;
1984 	struct umtx_pi *pi;
1985 
1986 	uq = td->td_umtxq;
1987 	mtx_lock(&umtx_lock);
1988 	/*
1989 	 * Pick up the lock that td is blocked on.
1990 	 */
1991 	pi = uq->uq_pi_blocked;
1992 	if (pi != NULL) {
1993 		umtx_pi_adjust_thread(pi, td);
1994 		umtx_repropagate_priority(pi);
1995 	}
1996 	mtx_unlock(&umtx_lock);
1997 }
1998 
1999 /*
2000  * Sleep on a PI mutex.
2001  */
2002 int
2003 umtxq_sleep_pi(struct umtx_q *uq, struct umtx_pi *pi, uint32_t owner,
2004     const char *wmesg, struct umtx_abs_timeout *timo, bool shared)
2005 {
2006 	struct thread *td, *td1;
2007 	struct umtx_q *uq1;
2008 	int error, pri;
2009 #ifdef INVARIANTS
2010 	struct umtxq_chain *uc;
2011 
2012 	uc = umtxq_getchain(&pi->pi_key);
2013 #endif
2014 	error = 0;
2015 	td = uq->uq_thread;
2016 	KASSERT(td == curthread, ("inconsistent uq_thread"));
2017 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
2018 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));
2019 	umtxq_insert(uq);
2020 	mtx_lock(&umtx_lock);
2021 	if (pi->pi_owner == NULL) {
2022 		mtx_unlock(&umtx_lock);
2023 		td1 = tdfind(owner, shared ? -1 : td->td_proc->p_pid);
2024 		mtx_lock(&umtx_lock);
2025 		if (td1 != NULL) {
2026 			if (pi->pi_owner == NULL)
2027 				umtx_pi_setowner(pi, td1);
2028 			PROC_UNLOCK(td1->td_proc);
2029 		}
2030 	}
2031 
2032 	TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
2033 		pri = UPRI(uq1->uq_thread);
2034 		if (pri > UPRI(td))
2035 			break;
2036 	}
2037 
2038 	if (uq1 != NULL)
2039 		TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
2040 	else
2041 		TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
2042 
2043 	uq->uq_pi_blocked = pi;
2044 	thread_lock(td);
2045 	td->td_flags |= TDF_UPIBLOCKED;
2046 	thread_unlock(td);
2047 	umtx_propagate_priority(td);
2048 	mtx_unlock(&umtx_lock);
2049 	umtxq_unbusy(&uq->uq_key);
2050 
2051 	error = umtxq_sleep(uq, wmesg, timo);
2052 	umtxq_remove(uq);
2053 
2054 	mtx_lock(&umtx_lock);
2055 	uq->uq_pi_blocked = NULL;
2056 	thread_lock(td);
2057 	td->td_flags &= ~TDF_UPIBLOCKED;
2058 	thread_unlock(td);
2059 	TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
2060 	umtx_repropagate_priority(pi);
2061 	mtx_unlock(&umtx_lock);
2062 	umtxq_unlock(&uq->uq_key);
2063 
2064 	return (error);
2065 }
2066 
2067 /*
2068  * Add reference count for a PI mutex.
2069  */
2070 void
2071 umtx_pi_ref(struct umtx_pi *pi)
2072 {
2073 
2074 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&pi->pi_key));
2075 	pi->pi_refcount++;
2076 }
2077 
2078 /*
2079  * Decrease reference count for a PI mutex, if the counter
2080  * is decreased to zero, its memory space is freed.
2081  */
2082 void
2083 umtx_pi_unref(struct umtx_pi *pi)
2084 {
2085 	struct umtxq_chain *uc;
2086 
2087 	uc = umtxq_getchain(&pi->pi_key);
2088 	UMTXQ_LOCKED_ASSERT(uc);
2089 	KASSERT(pi->pi_refcount > 0, ("invalid reference count"));
2090 	if (--pi->pi_refcount == 0) {
2091 		mtx_lock(&umtx_lock);
2092 		if (pi->pi_owner != NULL)
2093 			umtx_pi_disown(pi);
2094 		KASSERT(TAILQ_EMPTY(&pi->pi_blocked),
2095 			("blocked queue not empty"));
2096 		mtx_unlock(&umtx_lock);
2097 		TAILQ_REMOVE(&uc->uc_pi_list, pi, pi_hashlink);
2098 		umtx_pi_free(pi);
2099 	}
2100 }
2101 
2102 /*
2103  * Find a PI mutex in hash table.
2104  */
2105 struct umtx_pi *
2106 umtx_pi_lookup(struct umtx_key *key)
2107 {
2108 	struct umtxq_chain *uc;
2109 	struct umtx_pi *pi;
2110 
2111 	uc = umtxq_getchain(key);
2112 	UMTXQ_LOCKED_ASSERT(uc);
2113 
2114 	TAILQ_FOREACH(pi, &uc->uc_pi_list, pi_hashlink) {
2115 		if (umtx_key_match(&pi->pi_key, key)) {
2116 			return (pi);
2117 		}
2118 	}
2119 	return (NULL);
2120 }
2121 
2122 /*
2123  * Insert a PI mutex into hash table.
2124  */
2125 void
2126 umtx_pi_insert(struct umtx_pi *pi)
2127 {
2128 	struct umtxq_chain *uc;
2129 
2130 	uc = umtxq_getchain(&pi->pi_key);
2131 	UMTXQ_LOCKED_ASSERT(uc);
2132 	TAILQ_INSERT_TAIL(&uc->uc_pi_list, pi, pi_hashlink);
2133 }
2134 
2135 /*
2136  * Drop a PI mutex and wakeup a top waiter.
2137  */
2138 int
2139 umtx_pi_drop(struct thread *td, struct umtx_key *key, bool rb, int *count)
2140 {
2141 	struct umtx_q *uq_first, *uq_first2, *uq_me;
2142 	struct umtx_pi *pi, *pi2;
2143 	int pri;
2144 
2145 	UMTXQ_ASSERT_LOCKED_BUSY(key);
2146 	*count = umtxq_count_pi(key, &uq_first);
2147 	if (uq_first != NULL) {
2148 		mtx_lock(&umtx_lock);
2149 		pi = uq_first->uq_pi_blocked;
2150 		KASSERT(pi != NULL, ("pi == NULL?"));
2151 		if (pi->pi_owner != td && !(rb && pi->pi_owner == NULL)) {
2152 			mtx_unlock(&umtx_lock);
2153 			/* userland messed the mutex */
2154 			return (EPERM);
2155 		}
2156 		uq_me = td->td_umtxq;
2157 		if (pi->pi_owner == td)
2158 			umtx_pi_disown(pi);
2159 		/* get highest priority thread which is still sleeping. */
2160 		uq_first = TAILQ_FIRST(&pi->pi_blocked);
2161 		while (uq_first != NULL &&
2162 		    (uq_first->uq_flags & UQF_UMTXQ) == 0) {
2163 			uq_first = TAILQ_NEXT(uq_first, uq_lockq);
2164 		}
2165 		pri = PRI_MAX;
2166 		TAILQ_FOREACH(pi2, &uq_me->uq_pi_contested, pi_link) {
2167 			uq_first2 = TAILQ_FIRST(&pi2->pi_blocked);
2168 			if (uq_first2 != NULL) {
2169 				if (pri > UPRI(uq_first2->uq_thread))
2170 					pri = UPRI(uq_first2->uq_thread);
2171 			}
2172 		}
2173 		thread_lock(td);
2174 		sched_lend_user_prio(td, pri);
2175 		thread_unlock(td);
2176 		mtx_unlock(&umtx_lock);
2177 		if (uq_first)
2178 			umtxq_signal_thread(uq_first);
2179 	} else {
2180 		pi = umtx_pi_lookup(key);
2181 		/*
2182 		 * A umtx_pi can exist if a signal or timeout removed the
2183 		 * last waiter from the umtxq, but there is still
2184 		 * a thread in do_lock_pi() holding the umtx_pi.
2185 		 */
2186 		if (pi != NULL) {
2187 			/*
2188 			 * The umtx_pi can be unowned, such as when a thread
2189 			 * has just entered do_lock_pi(), allocated the
2190 			 * umtx_pi, and unlocked the umtxq.
2191 			 * If the current thread owns it, it must disown it.
2192 			 */
2193 			mtx_lock(&umtx_lock);
2194 			if (pi->pi_owner == td)
2195 				umtx_pi_disown(pi);
2196 			mtx_unlock(&umtx_lock);
2197 		}
2198 	}
2199 	return (0);
2200 }
2201 
2202 /*
2203  * Lock a PI mutex.
2204  */
2205 static int
2206 do_lock_pi(struct thread *td, struct umutex *m, uint32_t flags,
2207     struct _umtx_time *timeout, int try)
2208 {
2209 	struct umtx_abs_timeout timo;
2210 	struct umtx_q *uq;
2211 	struct umtx_pi *pi, *new_pi;
2212 	uint32_t id, old_owner, owner, old;
2213 	int error, rv;
2214 
2215 	id = td->td_tid;
2216 	uq = td->td_umtxq;
2217 
2218 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2219 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2220 	    &uq->uq_key)) != 0)
2221 		return (error);
2222 
2223 	if (timeout != NULL)
2224 		umtx_abs_timeout_init2(&timo, timeout);
2225 
2226 	umtxq_lock(&uq->uq_key);
2227 	pi = umtx_pi_lookup(&uq->uq_key);
2228 	if (pi == NULL) {
2229 		new_pi = umtx_pi_alloc(M_NOWAIT);
2230 		if (new_pi == NULL) {
2231 			umtxq_unlock(&uq->uq_key);
2232 			new_pi = umtx_pi_alloc(M_WAITOK);
2233 			umtxq_lock(&uq->uq_key);
2234 			pi = umtx_pi_lookup(&uq->uq_key);
2235 			if (pi != NULL) {
2236 				umtx_pi_free(new_pi);
2237 				new_pi = NULL;
2238 			}
2239 		}
2240 		if (new_pi != NULL) {
2241 			new_pi->pi_key = uq->uq_key;
2242 			umtx_pi_insert(new_pi);
2243 			pi = new_pi;
2244 		}
2245 	}
2246 	umtx_pi_ref(pi);
2247 	umtxq_unlock(&uq->uq_key);
2248 
2249 	/*
2250 	 * Care must be exercised when dealing with umtx structure.  It
2251 	 * can fault on any access.
2252 	 */
2253 	for (;;) {
2254 		/*
2255 		 * Try the uncontested case.  This should be done in userland.
2256 		 */
2257 		rv = casueword32(&m->m_owner, UMUTEX_UNOWNED, &owner, id);
2258 		/* The address was invalid. */
2259 		if (rv == -1) {
2260 			error = EFAULT;
2261 			break;
2262 		}
2263 		/* The acquire succeeded. */
2264 		if (rv == 0) {
2265 			MPASS(owner == UMUTEX_UNOWNED);
2266 			error = 0;
2267 			break;
2268 		}
2269 
2270 		if (owner == UMUTEX_RB_NOTRECOV) {
2271 			error = ENOTRECOVERABLE;
2272 			break;
2273 		}
2274 
2275 		/*
2276 		 * Nobody owns it, but the acquire failed. This can happen
2277 		 * with ll/sc atomics.
2278 		 */
2279 		if (owner == UMUTEX_UNOWNED) {
2280 			error = thread_check_susp(td, true);
2281 			if (error != 0)
2282 				break;
2283 			continue;
2284 		}
2285 
2286 		/*
2287 		 * Avoid overwriting a possible error from sleep due
2288 		 * to the pending signal with suspension check result.
2289 		 */
2290 		if (error == 0) {
2291 			error = thread_check_susp(td, true);
2292 			if (error != 0)
2293 				break;
2294 		}
2295 
2296 		/* If no one owns it but it is contested try to acquire it. */
2297 		if (owner == UMUTEX_CONTESTED || owner == UMUTEX_RB_OWNERDEAD) {
2298 			old_owner = owner;
2299 			rv = casueword32(&m->m_owner, owner, &owner,
2300 			    id | UMUTEX_CONTESTED);
2301 			/* The address was invalid. */
2302 			if (rv == -1) {
2303 				error = EFAULT;
2304 				break;
2305 			}
2306 			if (rv == 1) {
2307 				if (error == 0) {
2308 					error = thread_check_susp(td, true);
2309 					if (error != 0)
2310 						break;
2311 				}
2312 
2313 				/*
2314 				 * If this failed the lock could
2315 				 * changed, restart.
2316 				 */
2317 				continue;
2318 			}
2319 
2320 			MPASS(rv == 0);
2321 			MPASS(owner == old_owner);
2322 			umtxq_lock(&uq->uq_key);
2323 			umtxq_busy(&uq->uq_key);
2324 			error = umtx_pi_claim(pi, td);
2325 			umtxq_unbusy(&uq->uq_key);
2326 			umtxq_unlock(&uq->uq_key);
2327 			if (error != 0) {
2328 				/*
2329 				 * Since we're going to return an
2330 				 * error, restore the m_owner to its
2331 				 * previous, unowned state to avoid
2332 				 * compounding the problem.
2333 				 */
2334 				(void)casuword32(&m->m_owner,
2335 				    id | UMUTEX_CONTESTED, old_owner);
2336 			}
2337 			if (error == 0 && old_owner == UMUTEX_RB_OWNERDEAD)
2338 				error = EOWNERDEAD;
2339 			break;
2340 		}
2341 
2342 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2343 			error = EDEADLK;
2344 			break;
2345 		}
2346 
2347 		if (try != 0) {
2348 			error = EBUSY;
2349 			break;
2350 		}
2351 
2352 		/*
2353 		 * If we caught a signal, we have retried and now
2354 		 * exit immediately.
2355 		 */
2356 		if (error != 0)
2357 			break;
2358 
2359 		umtxq_lock(&uq->uq_key);
2360 		umtxq_busy(&uq->uq_key);
2361 		umtxq_unlock(&uq->uq_key);
2362 
2363 		/*
2364 		 * Set the contested bit so that a release in user space
2365 		 * knows to use the system call for unlock.  If this fails
2366 		 * either some one else has acquired the lock or it has been
2367 		 * released.
2368 		 */
2369 		rv = casueword32(&m->m_owner, owner, &old, owner |
2370 		    UMUTEX_CONTESTED);
2371 
2372 		/* The address was invalid. */
2373 		if (rv == -1) {
2374 			umtxq_unbusy_unlocked(&uq->uq_key);
2375 			error = EFAULT;
2376 			break;
2377 		}
2378 		if (rv == 1) {
2379 			umtxq_unbusy_unlocked(&uq->uq_key);
2380 			error = thread_check_susp(td, true);
2381 			if (error != 0)
2382 				break;
2383 
2384 			/*
2385 			 * The lock changed and we need to retry or we
2386 			 * lost a race to the thread unlocking the
2387 			 * umtx.  Note that the UMUTEX_RB_OWNERDEAD
2388 			 * value for owner is impossible there.
2389 			 */
2390 			continue;
2391 		}
2392 
2393 		umtxq_lock(&uq->uq_key);
2394 
2395 		/* We set the contested bit, sleep. */
2396 		MPASS(old == owner);
2397 		error = umtxq_sleep_pi(uq, pi, owner & ~UMUTEX_CONTESTED,
2398 		    "umtxpi", timeout == NULL ? NULL : &timo,
2399 		    (flags & USYNC_PROCESS_SHARED) != 0);
2400 		if (error != 0)
2401 			continue;
2402 
2403 		error = thread_check_susp(td, false);
2404 		if (error != 0)
2405 			break;
2406 	}
2407 
2408 	umtxq_lock(&uq->uq_key);
2409 	umtx_pi_unref(pi);
2410 	umtxq_unlock(&uq->uq_key);
2411 
2412 	umtx_key_release(&uq->uq_key);
2413 	return (error);
2414 }
2415 
2416 /*
2417  * Unlock a PI mutex.
2418  */
2419 static int
2420 do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2421 {
2422 	struct umtx_key key;
2423 	uint32_t id, new_owner, old, owner;
2424 	int count, error;
2425 
2426 	id = td->td_tid;
2427 
2428 usrloop:
2429 	/*
2430 	 * Make sure we own this mtx.
2431 	 */
2432 	error = fueword32(&m->m_owner, &owner);
2433 	if (error == -1)
2434 		return (EFAULT);
2435 
2436 	if ((owner & ~UMUTEX_CONTESTED) != id)
2437 		return (EPERM);
2438 
2439 	new_owner = umtx_unlock_val(flags, rb);
2440 
2441 	/* This should be done in userland */
2442 	if ((owner & UMUTEX_CONTESTED) == 0) {
2443 		error = casueword32(&m->m_owner, owner, &old, new_owner);
2444 		if (error == -1)
2445 			return (EFAULT);
2446 		if (error == 1) {
2447 			error = thread_check_susp(td, true);
2448 			if (error != 0)
2449 				return (error);
2450 			goto usrloop;
2451 		}
2452 		if (old == owner)
2453 			return (0);
2454 		owner = old;
2455 	}
2456 
2457 	/* We should only ever be in here for contested locks */
2458 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2459 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2460 	    &key)) != 0)
2461 		return (error);
2462 
2463 	umtxq_lock(&key);
2464 	umtxq_busy(&key);
2465 	error = umtx_pi_drop(td, &key, rb, &count);
2466 	if (error != 0) {
2467 		umtxq_unbusy(&key);
2468 		umtxq_unlock(&key);
2469 		umtx_key_release(&key);
2470 		/* userland messed the mutex */
2471 		return (error);
2472 	}
2473 	umtxq_unlock(&key);
2474 
2475 	/*
2476 	 * When unlocking the umtx, it must be marked as unowned if
2477 	 * there is zero or one thread only waiting for it.
2478 	 * Otherwise, it must be marked as contested.
2479 	 */
2480 
2481 	if (count > 1)
2482 		new_owner |= UMUTEX_CONTESTED;
2483 again:
2484 	error = casueword32(&m->m_owner, owner, &old, new_owner);
2485 	if (error == 1) {
2486 		error = thread_check_susp(td, false);
2487 		if (error == 0)
2488 			goto again;
2489 	}
2490 	umtxq_unbusy_unlocked(&key);
2491 	umtx_key_release(&key);
2492 	if (error == -1)
2493 		return (EFAULT);
2494 	if (error == 0 && old != owner)
2495 		return (EINVAL);
2496 	return (error);
2497 }
2498 
2499 /*
2500  * Lock a PP mutex.
2501  */
2502 static int
2503 do_lock_pp(struct thread *td, struct umutex *m, uint32_t flags,
2504     struct _umtx_time *timeout, int try)
2505 {
2506 	struct umtx_abs_timeout timo;
2507 	struct umtx_q *uq, *uq2;
2508 	struct umtx_pi *pi;
2509 	uint32_t ceiling;
2510 	uint32_t owner, id;
2511 	int error, pri, old_inherited_pri, su, rv;
2512 
2513 	id = td->td_tid;
2514 	uq = td->td_umtxq;
2515 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2516 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2517 	    &uq->uq_key)) != 0)
2518 		return (error);
2519 
2520 	if (timeout != NULL)
2521 		umtx_abs_timeout_init2(&timo, timeout);
2522 
2523 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2524 	for (;;) {
2525 		old_inherited_pri = uq->uq_inherited_pri;
2526 		umtxq_lock(&uq->uq_key);
2527 		umtxq_busy(&uq->uq_key);
2528 		umtxq_unlock(&uq->uq_key);
2529 
2530 		rv = fueword32(&m->m_ceilings[0], &ceiling);
2531 		if (rv == -1) {
2532 			error = EFAULT;
2533 			goto out;
2534 		}
2535 		ceiling = RTP_PRIO_MAX - ceiling;
2536 		if (ceiling > RTP_PRIO_MAX) {
2537 			error = EINVAL;
2538 			goto out;
2539 		}
2540 
2541 		mtx_lock(&umtx_lock);
2542 		if (UPRI(td) < PRI_MIN_REALTIME + ceiling) {
2543 			mtx_unlock(&umtx_lock);
2544 			error = EINVAL;
2545 			goto out;
2546 		}
2547 		if (su && PRI_MIN_REALTIME + ceiling < uq->uq_inherited_pri) {
2548 			uq->uq_inherited_pri = PRI_MIN_REALTIME + ceiling;
2549 			thread_lock(td);
2550 			if (uq->uq_inherited_pri < UPRI(td))
2551 				sched_lend_user_prio(td, uq->uq_inherited_pri);
2552 			thread_unlock(td);
2553 		}
2554 		mtx_unlock(&umtx_lock);
2555 
2556 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2557 		    id | UMUTEX_CONTESTED);
2558 		/* The address was invalid. */
2559 		if (rv == -1) {
2560 			error = EFAULT;
2561 			break;
2562 		}
2563 		if (rv == 0) {
2564 			MPASS(owner == UMUTEX_CONTESTED);
2565 			error = 0;
2566 			break;
2567 		}
2568 		/* rv == 1 */
2569 		if (owner == UMUTEX_RB_OWNERDEAD) {
2570 			rv = casueword32(&m->m_owner, UMUTEX_RB_OWNERDEAD,
2571 			    &owner, id | UMUTEX_CONTESTED);
2572 			if (rv == -1) {
2573 				error = EFAULT;
2574 				break;
2575 			}
2576 			if (rv == 0) {
2577 				MPASS(owner == UMUTEX_RB_OWNERDEAD);
2578 				error = EOWNERDEAD; /* success */
2579 				break;
2580 			}
2581 
2582 			/*
2583 			 *  rv == 1, only check for suspension if we
2584 			 *  did not already catched a signal.  If we
2585 			 *  get an error from the check, the same
2586 			 *  condition is checked by the umtxq_sleep()
2587 			 *  call below, so we should obliterate the
2588 			 *  error to not skip the last loop iteration.
2589 			 */
2590 			if (error == 0) {
2591 				error = thread_check_susp(td, false);
2592 				if (error == 0) {
2593 					if (try != 0)
2594 						error = EBUSY;
2595 					else
2596 						continue;
2597 				}
2598 				error = 0;
2599 			}
2600 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2601 			error = ENOTRECOVERABLE;
2602 		}
2603 
2604 		if (try != 0)
2605 			error = EBUSY;
2606 
2607 		/*
2608 		 * If we caught a signal, we have retried and now
2609 		 * exit immediately.
2610 		 */
2611 		if (error != 0)
2612 			break;
2613 
2614 		umtxq_lock(&uq->uq_key);
2615 		umtxq_insert(uq);
2616 		umtxq_unbusy(&uq->uq_key);
2617 		error = umtxq_sleep(uq, "umtxpp", timeout == NULL ?
2618 		    NULL : &timo);
2619 		umtxq_remove(uq);
2620 		umtxq_unlock(&uq->uq_key);
2621 
2622 		mtx_lock(&umtx_lock);
2623 		uq->uq_inherited_pri = old_inherited_pri;
2624 		pri = PRI_MAX;
2625 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2626 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2627 			if (uq2 != NULL) {
2628 				if (pri > UPRI(uq2->uq_thread))
2629 					pri = UPRI(uq2->uq_thread);
2630 			}
2631 		}
2632 		if (pri > uq->uq_inherited_pri)
2633 			pri = uq->uq_inherited_pri;
2634 		thread_lock(td);
2635 		sched_lend_user_prio(td, pri);
2636 		thread_unlock(td);
2637 		mtx_unlock(&umtx_lock);
2638 	}
2639 
2640 	if (error != 0 && error != EOWNERDEAD) {
2641 		mtx_lock(&umtx_lock);
2642 		uq->uq_inherited_pri = old_inherited_pri;
2643 		pri = PRI_MAX;
2644 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2645 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2646 			if (uq2 != NULL) {
2647 				if (pri > UPRI(uq2->uq_thread))
2648 					pri = UPRI(uq2->uq_thread);
2649 			}
2650 		}
2651 		if (pri > uq->uq_inherited_pri)
2652 			pri = uq->uq_inherited_pri;
2653 		thread_lock(td);
2654 		sched_lend_user_prio(td, pri);
2655 		thread_unlock(td);
2656 		mtx_unlock(&umtx_lock);
2657 	}
2658 
2659 out:
2660 	umtxq_unbusy_unlocked(&uq->uq_key);
2661 	umtx_key_release(&uq->uq_key);
2662 	return (error);
2663 }
2664 
2665 /*
2666  * Unlock a PP mutex.
2667  */
2668 static int
2669 do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2670 {
2671 	struct umtx_key key;
2672 	struct umtx_q *uq, *uq2;
2673 	struct umtx_pi *pi;
2674 	uint32_t id, owner, rceiling;
2675 	int error, pri, new_inherited_pri, su;
2676 
2677 	id = td->td_tid;
2678 	uq = td->td_umtxq;
2679 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2680 
2681 	/*
2682 	 * Make sure we own this mtx.
2683 	 */
2684 	error = fueword32(&m->m_owner, &owner);
2685 	if (error == -1)
2686 		return (EFAULT);
2687 
2688 	if ((owner & ~UMUTEX_CONTESTED) != id)
2689 		return (EPERM);
2690 
2691 	error = copyin(&m->m_ceilings[1], &rceiling, sizeof(uint32_t));
2692 	if (error != 0)
2693 		return (error);
2694 
2695 	if (rceiling == -1)
2696 		new_inherited_pri = PRI_MAX;
2697 	else {
2698 		rceiling = RTP_PRIO_MAX - rceiling;
2699 		if (rceiling > RTP_PRIO_MAX)
2700 			return (EINVAL);
2701 		new_inherited_pri = PRI_MIN_REALTIME + rceiling;
2702 	}
2703 
2704 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2705 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2706 	    &key)) != 0)
2707 		return (error);
2708 	umtxq_lock(&key);
2709 	umtxq_busy(&key);
2710 	umtxq_unlock(&key);
2711 	/*
2712 	 * For priority protected mutex, always set unlocked state
2713 	 * to UMUTEX_CONTESTED, so that userland always enters kernel
2714 	 * to lock the mutex, it is necessary because thread priority
2715 	 * has to be adjusted for such mutex.
2716 	 */
2717 	error = suword32(&m->m_owner, umtx_unlock_val(flags, rb) |
2718 	    UMUTEX_CONTESTED);
2719 
2720 	umtxq_lock(&key);
2721 	if (error == 0)
2722 		umtxq_signal(&key, 1);
2723 	umtxq_unbusy(&key);
2724 	umtxq_unlock(&key);
2725 
2726 	if (error == -1)
2727 		error = EFAULT;
2728 	else {
2729 		mtx_lock(&umtx_lock);
2730 		if (su != 0)
2731 			uq->uq_inherited_pri = new_inherited_pri;
2732 		pri = PRI_MAX;
2733 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2734 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2735 			if (uq2 != NULL) {
2736 				if (pri > UPRI(uq2->uq_thread))
2737 					pri = UPRI(uq2->uq_thread);
2738 			}
2739 		}
2740 		if (pri > uq->uq_inherited_pri)
2741 			pri = uq->uq_inherited_pri;
2742 		thread_lock(td);
2743 		sched_lend_user_prio(td, pri);
2744 		thread_unlock(td);
2745 		mtx_unlock(&umtx_lock);
2746 	}
2747 	umtx_key_release(&key);
2748 	return (error);
2749 }
2750 
2751 static int
2752 do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
2753     uint32_t *old_ceiling)
2754 {
2755 	struct umtx_q *uq;
2756 	uint32_t flags, id, owner, save_ceiling;
2757 	int error, rv, rv1;
2758 
2759 	error = fueword32(&m->m_flags, &flags);
2760 	if (error == -1)
2761 		return (EFAULT);
2762 	if ((flags & UMUTEX_PRIO_PROTECT) == 0)
2763 		return (EINVAL);
2764 	if (ceiling > RTP_PRIO_MAX)
2765 		return (EINVAL);
2766 	id = td->td_tid;
2767 	uq = td->td_umtxq;
2768 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2769 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2770 	    &uq->uq_key)) != 0)
2771 		return (error);
2772 	for (;;) {
2773 		umtxq_lock(&uq->uq_key);
2774 		umtxq_busy(&uq->uq_key);
2775 		umtxq_unlock(&uq->uq_key);
2776 
2777 		rv = fueword32(&m->m_ceilings[0], &save_ceiling);
2778 		if (rv == -1) {
2779 			error = EFAULT;
2780 			break;
2781 		}
2782 
2783 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2784 		    id | UMUTEX_CONTESTED);
2785 		if (rv == -1) {
2786 			error = EFAULT;
2787 			break;
2788 		}
2789 
2790 		if (rv == 0) {
2791 			MPASS(owner == UMUTEX_CONTESTED);
2792 			rv = suword32(&m->m_ceilings[0], ceiling);
2793 			rv1 = suword32(&m->m_owner, UMUTEX_CONTESTED);
2794 			error = (rv == 0 && rv1 == 0) ? 0: EFAULT;
2795 			break;
2796 		}
2797 
2798 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2799 			rv = suword32(&m->m_ceilings[0], ceiling);
2800 			error = rv == 0 ? 0 : EFAULT;
2801 			break;
2802 		}
2803 
2804 		if (owner == UMUTEX_RB_OWNERDEAD) {
2805 			error = EOWNERDEAD;
2806 			break;
2807 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2808 			error = ENOTRECOVERABLE;
2809 			break;
2810 		}
2811 
2812 		/*
2813 		 * If we caught a signal, we have retried and now
2814 		 * exit immediately.
2815 		 */
2816 		if (error != 0)
2817 			break;
2818 
2819 		/*
2820 		 * We set the contested bit, sleep. Otherwise the lock changed
2821 		 * and we need to retry or we lost a race to the thread
2822 		 * unlocking the umtx.
2823 		 */
2824 		umtxq_lock(&uq->uq_key);
2825 		umtxq_insert(uq);
2826 		umtxq_unbusy(&uq->uq_key);
2827 		error = umtxq_sleep(uq, "umtxpp", NULL);
2828 		umtxq_remove(uq);
2829 		umtxq_unlock(&uq->uq_key);
2830 	}
2831 	umtxq_lock(&uq->uq_key);
2832 	if (error == 0)
2833 		umtxq_signal(&uq->uq_key, INT_MAX);
2834 	umtxq_unbusy(&uq->uq_key);
2835 	umtxq_unlock(&uq->uq_key);
2836 	umtx_key_release(&uq->uq_key);
2837 	if (error == 0 && old_ceiling != NULL) {
2838 		rv = suword32(old_ceiling, save_ceiling);
2839 		error = rv == 0 ? 0 : EFAULT;
2840 	}
2841 	return (error);
2842 }
2843 
2844 /*
2845  * Lock a userland POSIX mutex.
2846  */
2847 static int
2848 do_lock_umutex(struct thread *td, struct umutex *m,
2849     struct _umtx_time *timeout, int mode)
2850 {
2851 	uint32_t flags;
2852 	int error;
2853 
2854 	error = fueword32(&m->m_flags, &flags);
2855 	if (error == -1)
2856 		return (EFAULT);
2857 
2858 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2859 	case 0:
2860 		error = do_lock_normal(td, m, flags, timeout, mode);
2861 		break;
2862 	case UMUTEX_PRIO_INHERIT:
2863 		error = do_lock_pi(td, m, flags, timeout, mode);
2864 		break;
2865 	case UMUTEX_PRIO_PROTECT:
2866 		error = do_lock_pp(td, m, flags, timeout, mode);
2867 		break;
2868 	default:
2869 		return (EINVAL);
2870 	}
2871 	if (timeout == NULL) {
2872 		if (error == EINTR && mode != _UMUTEX_WAIT)
2873 			error = ERESTART;
2874 	} else {
2875 		/* Timed-locking is not restarted. */
2876 		if (error == ERESTART)
2877 			error = EINTR;
2878 	}
2879 	return (error);
2880 }
2881 
2882 /*
2883  * Unlock a userland POSIX mutex.
2884  */
2885 static int
2886 do_unlock_umutex(struct thread *td, struct umutex *m, bool rb)
2887 {
2888 	uint32_t flags;
2889 	int error;
2890 
2891 	error = fueword32(&m->m_flags, &flags);
2892 	if (error == -1)
2893 		return (EFAULT);
2894 
2895 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2896 	case 0:
2897 		return (do_unlock_normal(td, m, flags, rb));
2898 	case UMUTEX_PRIO_INHERIT:
2899 		return (do_unlock_pi(td, m, flags, rb));
2900 	case UMUTEX_PRIO_PROTECT:
2901 		return (do_unlock_pp(td, m, flags, rb));
2902 	}
2903 
2904 	return (EINVAL);
2905 }
2906 
2907 static int
2908 do_cv_wait(struct thread *td, struct ucond *cv, struct umutex *m,
2909     struct timespec *timeout, u_long wflags)
2910 {
2911 	struct umtx_abs_timeout timo;
2912 	struct umtx_q *uq;
2913 	uint32_t flags, clockid, hasw;
2914 	int error;
2915 
2916 	uq = td->td_umtxq;
2917 	error = fueword32(&cv->c_flags, &flags);
2918 	if (error == -1)
2919 		return (EFAULT);
2920 	error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &uq->uq_key);
2921 	if (error != 0)
2922 		return (error);
2923 
2924 	if ((wflags & CVWAIT_CLOCKID) != 0) {
2925 		error = fueword32(&cv->c_clockid, &clockid);
2926 		if (error == -1) {
2927 			umtx_key_release(&uq->uq_key);
2928 			return (EFAULT);
2929 		}
2930 		if (clockid < CLOCK_REALTIME ||
2931 		    clockid >= CLOCK_THREAD_CPUTIME_ID) {
2932 			/* hmm, only HW clock id will work. */
2933 			umtx_key_release(&uq->uq_key);
2934 			return (EINVAL);
2935 		}
2936 	} else {
2937 		clockid = CLOCK_REALTIME;
2938 	}
2939 
2940 	umtxq_lock(&uq->uq_key);
2941 	umtxq_busy(&uq->uq_key);
2942 	umtxq_insert(uq);
2943 	umtxq_unlock(&uq->uq_key);
2944 
2945 	/*
2946 	 * Set c_has_waiters to 1 before releasing user mutex, also
2947 	 * don't modify cache line when unnecessary.
2948 	 */
2949 	error = fueword32(&cv->c_has_waiters, &hasw);
2950 	if (error == 0 && hasw == 0)
2951 		suword32(&cv->c_has_waiters, 1);
2952 
2953 	umtxq_unbusy_unlocked(&uq->uq_key);
2954 
2955 	error = do_unlock_umutex(td, m, false);
2956 
2957 	if (timeout != NULL)
2958 		umtx_abs_timeout_init(&timo, clockid,
2959 		    (wflags & CVWAIT_ABSTIME) != 0, timeout);
2960 
2961 	umtxq_lock(&uq->uq_key);
2962 	if (error == 0) {
2963 		error = umtxq_sleep(uq, "ucond", timeout == NULL ?
2964 		    NULL : &timo);
2965 	}
2966 
2967 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
2968 		error = 0;
2969 	else {
2970 		/*
2971 		 * This must be timeout,interrupted by signal or
2972 		 * surprious wakeup, clear c_has_waiter flag when
2973 		 * necessary.
2974 		 */
2975 		umtxq_busy(&uq->uq_key);
2976 		if ((uq->uq_flags & UQF_UMTXQ) != 0) {
2977 			int oldlen = uq->uq_cur_queue->length;
2978 			umtxq_remove(uq);
2979 			if (oldlen == 1) {
2980 				umtxq_unlock(&uq->uq_key);
2981 				suword32(&cv->c_has_waiters, 0);
2982 				umtxq_lock(&uq->uq_key);
2983 			}
2984 		}
2985 		umtxq_unbusy(&uq->uq_key);
2986 		if (error == ERESTART)
2987 			error = EINTR;
2988 	}
2989 
2990 	umtxq_unlock(&uq->uq_key);
2991 	umtx_key_release(&uq->uq_key);
2992 	return (error);
2993 }
2994 
2995 /*
2996  * Signal a userland condition variable.
2997  */
2998 static int
2999 do_cv_signal(struct thread *td, struct ucond *cv)
3000 {
3001 	struct umtx_key key;
3002 	int error, cnt, nwake;
3003 	uint32_t flags;
3004 
3005 	error = fueword32(&cv->c_flags, &flags);
3006 	if (error == -1)
3007 		return (EFAULT);
3008 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
3009 		return (error);
3010 	umtxq_lock(&key);
3011 	umtxq_busy(&key);
3012 	cnt = umtxq_count(&key);
3013 	nwake = umtxq_signal(&key, 1);
3014 	if (cnt <= nwake) {
3015 		umtxq_unlock(&key);
3016 		error = suword32(&cv->c_has_waiters, 0);
3017 		if (error == -1)
3018 			error = EFAULT;
3019 		umtxq_lock(&key);
3020 	}
3021 	umtxq_unbusy(&key);
3022 	umtxq_unlock(&key);
3023 	umtx_key_release(&key);
3024 	return (error);
3025 }
3026 
3027 static int
3028 do_cv_broadcast(struct thread *td, struct ucond *cv)
3029 {
3030 	struct umtx_key key;
3031 	int error;
3032 	uint32_t flags;
3033 
3034 	error = fueword32(&cv->c_flags, &flags);
3035 	if (error == -1)
3036 		return (EFAULT);
3037 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
3038 		return (error);
3039 
3040 	umtxq_lock(&key);
3041 	umtxq_busy(&key);
3042 	umtxq_signal(&key, INT_MAX);
3043 	umtxq_unlock(&key);
3044 
3045 	error = suword32(&cv->c_has_waiters, 0);
3046 	if (error == -1)
3047 		error = EFAULT;
3048 
3049 	umtxq_unbusy_unlocked(&key);
3050 
3051 	umtx_key_release(&key);
3052 	return (error);
3053 }
3054 
3055 static int
3056 do_rw_rdlock(struct thread *td, struct urwlock *rwlock, long fflag,
3057     struct _umtx_time *timeout)
3058 {
3059 	struct umtx_abs_timeout timo;
3060 	struct umtx_q *uq;
3061 	uint32_t flags, wrflags;
3062 	int32_t state, oldstate;
3063 	int32_t blocked_readers;
3064 	int error, error1, rv;
3065 
3066 	uq = td->td_umtxq;
3067 	error = fueword32(&rwlock->rw_flags, &flags);
3068 	if (error == -1)
3069 		return (EFAULT);
3070 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3071 	if (error != 0)
3072 		return (error);
3073 
3074 	if (timeout != NULL)
3075 		umtx_abs_timeout_init2(&timo, timeout);
3076 
3077 	wrflags = URWLOCK_WRITE_OWNER;
3078 	if (!(fflag & URWLOCK_PREFER_READER) && !(flags & URWLOCK_PREFER_READER))
3079 		wrflags |= URWLOCK_WRITE_WAITERS;
3080 
3081 	for (;;) {
3082 		rv = fueword32(&rwlock->rw_state, &state);
3083 		if (rv == -1) {
3084 			umtx_key_release(&uq->uq_key);
3085 			return (EFAULT);
3086 		}
3087 
3088 		/* try to lock it */
3089 		while (!(state & wrflags)) {
3090 			if (__predict_false(URWLOCK_READER_COUNT(state) ==
3091 			    URWLOCK_MAX_READERS)) {
3092 				umtx_key_release(&uq->uq_key);
3093 				return (EAGAIN);
3094 			}
3095 			rv = casueword32(&rwlock->rw_state, state,
3096 			    &oldstate, state + 1);
3097 			if (rv == -1) {
3098 				umtx_key_release(&uq->uq_key);
3099 				return (EFAULT);
3100 			}
3101 			if (rv == 0) {
3102 				MPASS(oldstate == state);
3103 				umtx_key_release(&uq->uq_key);
3104 				return (0);
3105 			}
3106 			error = thread_check_susp(td, true);
3107 			if (error != 0)
3108 				break;
3109 			state = oldstate;
3110 		}
3111 
3112 		if (error)
3113 			break;
3114 
3115 		/* grab monitor lock */
3116 		umtxq_lock(&uq->uq_key);
3117 		umtxq_busy(&uq->uq_key);
3118 		umtxq_unlock(&uq->uq_key);
3119 
3120 		/*
3121 		 * re-read the state, in case it changed between the try-lock above
3122 		 * and the check below
3123 		 */
3124 		rv = fueword32(&rwlock->rw_state, &state);
3125 		if (rv == -1)
3126 			error = EFAULT;
3127 
3128 		/* set read contention bit */
3129 		while (error == 0 && (state & wrflags) &&
3130 		    !(state & URWLOCK_READ_WAITERS)) {
3131 			rv = casueword32(&rwlock->rw_state, state,
3132 			    &oldstate, state | URWLOCK_READ_WAITERS);
3133 			if (rv == -1) {
3134 				error = EFAULT;
3135 				break;
3136 			}
3137 			if (rv == 0) {
3138 				MPASS(oldstate == state);
3139 				goto sleep;
3140 			}
3141 			state = oldstate;
3142 			error = thread_check_susp(td, false);
3143 			if (error != 0)
3144 				break;
3145 		}
3146 		if (error != 0) {
3147 			umtxq_unbusy_unlocked(&uq->uq_key);
3148 			break;
3149 		}
3150 
3151 		/* state is changed while setting flags, restart */
3152 		if (!(state & wrflags)) {
3153 			umtxq_unbusy_unlocked(&uq->uq_key);
3154 			error = thread_check_susp(td, true);
3155 			if (error != 0)
3156 				break;
3157 			continue;
3158 		}
3159 
3160 sleep:
3161 		/*
3162 		 * Contention bit is set, before sleeping, increase
3163 		 * read waiter count.
3164 		 */
3165 		rv = fueword32(&rwlock->rw_blocked_readers,
3166 		    &blocked_readers);
3167 		if (rv == -1) {
3168 			umtxq_unbusy_unlocked(&uq->uq_key);
3169 			error = EFAULT;
3170 			break;
3171 		}
3172 		suword32(&rwlock->rw_blocked_readers, blocked_readers+1);
3173 
3174 		while (state & wrflags) {
3175 			umtxq_lock(&uq->uq_key);
3176 			umtxq_insert(uq);
3177 			umtxq_unbusy(&uq->uq_key);
3178 
3179 			error = umtxq_sleep(uq, "urdlck", timeout == NULL ?
3180 			    NULL : &timo);
3181 
3182 			umtxq_busy(&uq->uq_key);
3183 			umtxq_remove(uq);
3184 			umtxq_unlock(&uq->uq_key);
3185 			if (error)
3186 				break;
3187 			rv = fueword32(&rwlock->rw_state, &state);
3188 			if (rv == -1) {
3189 				error = EFAULT;
3190 				break;
3191 			}
3192 		}
3193 
3194 		/* decrease read waiter count, and may clear read contention bit */
3195 		rv = fueword32(&rwlock->rw_blocked_readers,
3196 		    &blocked_readers);
3197 		if (rv == -1) {
3198 			umtxq_unbusy_unlocked(&uq->uq_key);
3199 			error = EFAULT;
3200 			break;
3201 		}
3202 		suword32(&rwlock->rw_blocked_readers, blocked_readers-1);
3203 		if (blocked_readers == 1) {
3204 			rv = fueword32(&rwlock->rw_state, &state);
3205 			if (rv == -1) {
3206 				umtxq_unbusy_unlocked(&uq->uq_key);
3207 				error = EFAULT;
3208 				break;
3209 			}
3210 			for (;;) {
3211 				rv = casueword32(&rwlock->rw_state, state,
3212 				    &oldstate, state & ~URWLOCK_READ_WAITERS);
3213 				if (rv == -1) {
3214 					error = EFAULT;
3215 					break;
3216 				}
3217 				if (rv == 0) {
3218 					MPASS(oldstate == state);
3219 					break;
3220 				}
3221 				state = oldstate;
3222 				error1 = thread_check_susp(td, false);
3223 				if (error1 != 0) {
3224 					if (error == 0)
3225 						error = error1;
3226 					break;
3227 				}
3228 			}
3229 		}
3230 
3231 		umtxq_unbusy_unlocked(&uq->uq_key);
3232 		if (error != 0)
3233 			break;
3234 	}
3235 	umtx_key_release(&uq->uq_key);
3236 	if (error == ERESTART)
3237 		error = EINTR;
3238 	return (error);
3239 }
3240 
3241 static int
3242 do_rw_wrlock(struct thread *td, struct urwlock *rwlock, struct _umtx_time *timeout)
3243 {
3244 	struct umtx_abs_timeout timo;
3245 	struct umtx_q *uq;
3246 	uint32_t flags;
3247 	int32_t state, oldstate;
3248 	int32_t blocked_writers;
3249 	int32_t blocked_readers;
3250 	int error, error1, rv;
3251 
3252 	uq = td->td_umtxq;
3253 	error = fueword32(&rwlock->rw_flags, &flags);
3254 	if (error == -1)
3255 		return (EFAULT);
3256 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3257 	if (error != 0)
3258 		return (error);
3259 
3260 	if (timeout != NULL)
3261 		umtx_abs_timeout_init2(&timo, timeout);
3262 
3263 	blocked_readers = 0;
3264 	for (;;) {
3265 		rv = fueword32(&rwlock->rw_state, &state);
3266 		if (rv == -1) {
3267 			umtx_key_release(&uq->uq_key);
3268 			return (EFAULT);
3269 		}
3270 		while ((state & URWLOCK_WRITE_OWNER) == 0 &&
3271 		    URWLOCK_READER_COUNT(state) == 0) {
3272 			rv = casueword32(&rwlock->rw_state, state,
3273 			    &oldstate, state | URWLOCK_WRITE_OWNER);
3274 			if (rv == -1) {
3275 				umtx_key_release(&uq->uq_key);
3276 				return (EFAULT);
3277 			}
3278 			if (rv == 0) {
3279 				MPASS(oldstate == state);
3280 				umtx_key_release(&uq->uq_key);
3281 				return (0);
3282 			}
3283 			state = oldstate;
3284 			error = thread_check_susp(td, true);
3285 			if (error != 0)
3286 				break;
3287 		}
3288 
3289 		if (error) {
3290 			if ((state & (URWLOCK_WRITE_OWNER |
3291 			    URWLOCK_WRITE_WAITERS)) == 0 &&
3292 			    blocked_readers != 0) {
3293 				umtxq_lock(&uq->uq_key);
3294 				umtxq_busy(&uq->uq_key);
3295 				umtxq_signal_queue(&uq->uq_key, INT_MAX,
3296 				    UMTX_SHARED_QUEUE);
3297 				umtxq_unbusy(&uq->uq_key);
3298 				umtxq_unlock(&uq->uq_key);
3299 			}
3300 
3301 			break;
3302 		}
3303 
3304 		/* grab monitor lock */
3305 		umtxq_lock(&uq->uq_key);
3306 		umtxq_busy(&uq->uq_key);
3307 		umtxq_unlock(&uq->uq_key);
3308 
3309 		/*
3310 		 * Re-read the state, in case it changed between the
3311 		 * try-lock above and the check below.
3312 		 */
3313 		rv = fueword32(&rwlock->rw_state, &state);
3314 		if (rv == -1)
3315 			error = EFAULT;
3316 
3317 		while (error == 0 && ((state & URWLOCK_WRITE_OWNER) ||
3318 		    URWLOCK_READER_COUNT(state) != 0) &&
3319 		    (state & URWLOCK_WRITE_WAITERS) == 0) {
3320 			rv = casueword32(&rwlock->rw_state, state,
3321 			    &oldstate, state | URWLOCK_WRITE_WAITERS);
3322 			if (rv == -1) {
3323 				error = EFAULT;
3324 				break;
3325 			}
3326 			if (rv == 0) {
3327 				MPASS(oldstate == state);
3328 				goto sleep;
3329 			}
3330 			state = oldstate;
3331 			error = thread_check_susp(td, false);
3332 			if (error != 0)
3333 				break;
3334 		}
3335 		if (error != 0) {
3336 			umtxq_unbusy_unlocked(&uq->uq_key);
3337 			break;
3338 		}
3339 
3340 		if ((state & URWLOCK_WRITE_OWNER) == 0 &&
3341 		    URWLOCK_READER_COUNT(state) == 0) {
3342 			umtxq_unbusy_unlocked(&uq->uq_key);
3343 			error = thread_check_susp(td, false);
3344 			if (error != 0)
3345 				break;
3346 			continue;
3347 		}
3348 sleep:
3349 		rv = fueword32(&rwlock->rw_blocked_writers,
3350 		    &blocked_writers);
3351 		if (rv == -1) {
3352 			umtxq_unbusy_unlocked(&uq->uq_key);
3353 			error = EFAULT;
3354 			break;
3355 		}
3356 		suword32(&rwlock->rw_blocked_writers, blocked_writers + 1);
3357 
3358 		while ((state & URWLOCK_WRITE_OWNER) ||
3359 		    URWLOCK_READER_COUNT(state) != 0) {
3360 			umtxq_lock(&uq->uq_key);
3361 			umtxq_insert_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3362 			umtxq_unbusy(&uq->uq_key);
3363 
3364 			error = umtxq_sleep(uq, "uwrlck", timeout == NULL ?
3365 			    NULL : &timo);
3366 
3367 			umtxq_busy(&uq->uq_key);
3368 			umtxq_remove_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3369 			umtxq_unlock(&uq->uq_key);
3370 			if (error)
3371 				break;
3372 			rv = fueword32(&rwlock->rw_state, &state);
3373 			if (rv == -1) {
3374 				error = EFAULT;
3375 				break;
3376 			}
3377 		}
3378 
3379 		rv = fueword32(&rwlock->rw_blocked_writers,
3380 		    &blocked_writers);
3381 		if (rv == -1) {
3382 			umtxq_unbusy_unlocked(&uq->uq_key);
3383 			error = EFAULT;
3384 			break;
3385 		}
3386 		suword32(&rwlock->rw_blocked_writers, blocked_writers-1);
3387 		if (blocked_writers == 1) {
3388 			rv = fueword32(&rwlock->rw_state, &state);
3389 			if (rv == -1) {
3390 				umtxq_unbusy_unlocked(&uq->uq_key);
3391 				error = EFAULT;
3392 				break;
3393 			}
3394 			for (;;) {
3395 				rv = casueword32(&rwlock->rw_state, state,
3396 				    &oldstate, state & ~URWLOCK_WRITE_WAITERS);
3397 				if (rv == -1) {
3398 					error = EFAULT;
3399 					break;
3400 				}
3401 				if (rv == 0) {
3402 					MPASS(oldstate == state);
3403 					break;
3404 				}
3405 				state = oldstate;
3406 				error1 = thread_check_susp(td, false);
3407 				/*
3408 				 * We are leaving the URWLOCK_WRITE_WAITERS
3409 				 * behind, but this should not harm the
3410 				 * correctness.
3411 				 */
3412 				if (error1 != 0) {
3413 					if (error == 0)
3414 						error = error1;
3415 					break;
3416 				}
3417 			}
3418 			rv = fueword32(&rwlock->rw_blocked_readers,
3419 			    &blocked_readers);
3420 			if (rv == -1) {
3421 				umtxq_unbusy_unlocked(&uq->uq_key);
3422 				error = EFAULT;
3423 				break;
3424 			}
3425 		} else
3426 			blocked_readers = 0;
3427 
3428 		umtxq_unbusy_unlocked(&uq->uq_key);
3429 	}
3430 
3431 	umtx_key_release(&uq->uq_key);
3432 	if (error == ERESTART)
3433 		error = EINTR;
3434 	return (error);
3435 }
3436 
3437 static int
3438 do_rw_unlock(struct thread *td, struct urwlock *rwlock)
3439 {
3440 	struct umtx_q *uq;
3441 	uint32_t flags;
3442 	int32_t state, oldstate;
3443 	int error, rv, q, count;
3444 
3445 	uq = td->td_umtxq;
3446 	error = fueword32(&rwlock->rw_flags, &flags);
3447 	if (error == -1)
3448 		return (EFAULT);
3449 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3450 	if (error != 0)
3451 		return (error);
3452 
3453 	error = fueword32(&rwlock->rw_state, &state);
3454 	if (error == -1) {
3455 		error = EFAULT;
3456 		goto out;
3457 	}
3458 	if (state & URWLOCK_WRITE_OWNER) {
3459 		for (;;) {
3460 			rv = casueword32(&rwlock->rw_state, state,
3461 			    &oldstate, state & ~URWLOCK_WRITE_OWNER);
3462 			if (rv == -1) {
3463 				error = EFAULT;
3464 				goto out;
3465 			}
3466 			if (rv == 1) {
3467 				state = oldstate;
3468 				if (!(oldstate & URWLOCK_WRITE_OWNER)) {
3469 					error = EPERM;
3470 					goto out;
3471 				}
3472 				error = thread_check_susp(td, true);
3473 				if (error != 0)
3474 					goto out;
3475 			} else
3476 				break;
3477 		}
3478 	} else if (URWLOCK_READER_COUNT(state) != 0) {
3479 		for (;;) {
3480 			rv = casueword32(&rwlock->rw_state, state,
3481 			    &oldstate, state - 1);
3482 			if (rv == -1) {
3483 				error = EFAULT;
3484 				goto out;
3485 			}
3486 			if (rv == 1) {
3487 				state = oldstate;
3488 				if (URWLOCK_READER_COUNT(oldstate) == 0) {
3489 					error = EPERM;
3490 					goto out;
3491 				}
3492 				error = thread_check_susp(td, true);
3493 				if (error != 0)
3494 					goto out;
3495 			} else
3496 				break;
3497 		}
3498 	} else {
3499 		error = EPERM;
3500 		goto out;
3501 	}
3502 
3503 	count = 0;
3504 
3505 	if (!(flags & URWLOCK_PREFER_READER)) {
3506 		if (state & URWLOCK_WRITE_WAITERS) {
3507 			count = 1;
3508 			q = UMTX_EXCLUSIVE_QUEUE;
3509 		} else if (state & URWLOCK_READ_WAITERS) {
3510 			count = INT_MAX;
3511 			q = UMTX_SHARED_QUEUE;
3512 		}
3513 	} else {
3514 		if (state & URWLOCK_READ_WAITERS) {
3515 			count = INT_MAX;
3516 			q = UMTX_SHARED_QUEUE;
3517 		} else if (state & URWLOCK_WRITE_WAITERS) {
3518 			count = 1;
3519 			q = UMTX_EXCLUSIVE_QUEUE;
3520 		}
3521 	}
3522 
3523 	if (count) {
3524 		umtxq_lock(&uq->uq_key);
3525 		umtxq_busy(&uq->uq_key);
3526 		umtxq_signal_queue(&uq->uq_key, count, q);
3527 		umtxq_unbusy(&uq->uq_key);
3528 		umtxq_unlock(&uq->uq_key);
3529 	}
3530 out:
3531 	umtx_key_release(&uq->uq_key);
3532 	return (error);
3533 }
3534 
3535 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3536 static int
3537 do_sem_wait(struct thread *td, struct _usem *sem, struct _umtx_time *timeout)
3538 {
3539 	struct umtx_abs_timeout timo;
3540 	struct umtx_q *uq;
3541 	uint32_t flags, count, count1;
3542 	int error, rv, rv1;
3543 
3544 	uq = td->td_umtxq;
3545 	error = fueword32(&sem->_flags, &flags);
3546 	if (error == -1)
3547 		return (EFAULT);
3548 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3549 	if (error != 0)
3550 		return (error);
3551 
3552 	if (timeout != NULL)
3553 		umtx_abs_timeout_init2(&timo, timeout);
3554 
3555 again:
3556 	umtxq_lock(&uq->uq_key);
3557 	umtxq_busy(&uq->uq_key);
3558 	umtxq_insert(uq);
3559 	umtxq_unlock(&uq->uq_key);
3560 	rv = casueword32(&sem->_has_waiters, 0, &count1, 1);
3561 	if (rv != -1)
3562 		rv1 = fueword32(&sem->_count, &count);
3563 	if (rv == -1 || rv1 == -1 || count != 0 || (rv == 1 && count1 == 0)) {
3564 		if (rv == 0)
3565 			suword32(&sem->_has_waiters, 0);
3566 		umtxq_lock(&uq->uq_key);
3567 		umtxq_unbusy(&uq->uq_key);
3568 		umtxq_remove(uq);
3569 		umtxq_unlock(&uq->uq_key);
3570 		if (rv == -1 || rv1 == -1) {
3571 			error = EFAULT;
3572 			goto out;
3573 		}
3574 		if (count != 0) {
3575 			error = 0;
3576 			goto out;
3577 		}
3578 		MPASS(rv == 1 && count1 == 0);
3579 		rv = thread_check_susp(td, true);
3580 		if (rv == 0)
3581 			goto again;
3582 		error = rv;
3583 		goto out;
3584 	}
3585 	umtxq_lock(&uq->uq_key);
3586 	umtxq_unbusy(&uq->uq_key);
3587 
3588 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3589 
3590 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3591 		error = 0;
3592 	else {
3593 		umtxq_remove(uq);
3594 		/* A relative timeout cannot be restarted. */
3595 		if (error == ERESTART && timeout != NULL &&
3596 		    (timeout->_flags & UMTX_ABSTIME) == 0)
3597 			error = EINTR;
3598 	}
3599 	umtxq_unlock(&uq->uq_key);
3600 out:
3601 	umtx_key_release(&uq->uq_key);
3602 	return (error);
3603 }
3604 
3605 /*
3606  * Signal a userland semaphore.
3607  */
3608 static int
3609 do_sem_wake(struct thread *td, struct _usem *sem)
3610 {
3611 	struct umtx_key key;
3612 	int error, cnt;
3613 	uint32_t flags;
3614 
3615 	error = fueword32(&sem->_flags, &flags);
3616 	if (error == -1)
3617 		return (EFAULT);
3618 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3619 		return (error);
3620 	umtxq_lock(&key);
3621 	umtxq_busy(&key);
3622 	cnt = umtxq_count(&key);
3623 	if (cnt > 0) {
3624 		/*
3625 		 * Check if count is greater than 0, this means the memory is
3626 		 * still being referenced by user code, so we can safely
3627 		 * update _has_waiters flag.
3628 		 */
3629 		if (cnt == 1) {
3630 			umtxq_unlock(&key);
3631 			error = suword32(&sem->_has_waiters, 0);
3632 			umtxq_lock(&key);
3633 			if (error == -1)
3634 				error = EFAULT;
3635 		}
3636 		umtxq_signal(&key, 1);
3637 	}
3638 	umtxq_unbusy(&key);
3639 	umtxq_unlock(&key);
3640 	umtx_key_release(&key);
3641 	return (error);
3642 }
3643 #endif
3644 
3645 static int
3646 do_sem2_wait(struct thread *td, struct _usem2 *sem, struct _umtx_time *timeout)
3647 {
3648 	struct umtx_abs_timeout timo;
3649 	struct umtx_q *uq;
3650 	uint32_t count, flags;
3651 	int error, rv;
3652 
3653 	uq = td->td_umtxq;
3654 	flags = fuword32(&sem->_flags);
3655 	if (timeout != NULL)
3656 		umtx_abs_timeout_init2(&timo, timeout);
3657 
3658 again:
3659 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3660 	if (error != 0)
3661 		return (error);
3662 	umtxq_lock(&uq->uq_key);
3663 	umtxq_busy(&uq->uq_key);
3664 	umtxq_insert(uq);
3665 	umtxq_unlock(&uq->uq_key);
3666 	rv = fueword32(&sem->_count, &count);
3667 	if (rv == -1) {
3668 		umtxq_lock(&uq->uq_key);
3669 		umtxq_unbusy(&uq->uq_key);
3670 		umtxq_remove(uq);
3671 		umtxq_unlock(&uq->uq_key);
3672 		umtx_key_release(&uq->uq_key);
3673 		return (EFAULT);
3674 	}
3675 	for (;;) {
3676 		if (USEM_COUNT(count) != 0) {
3677 			umtxq_lock(&uq->uq_key);
3678 			umtxq_unbusy(&uq->uq_key);
3679 			umtxq_remove(uq);
3680 			umtxq_unlock(&uq->uq_key);
3681 			umtx_key_release(&uq->uq_key);
3682 			return (0);
3683 		}
3684 		if (count == USEM_HAS_WAITERS)
3685 			break;
3686 		rv = casueword32(&sem->_count, 0, &count, USEM_HAS_WAITERS);
3687 		if (rv == 0)
3688 			break;
3689 		umtxq_lock(&uq->uq_key);
3690 		umtxq_unbusy(&uq->uq_key);
3691 		umtxq_remove(uq);
3692 		umtxq_unlock(&uq->uq_key);
3693 		umtx_key_release(&uq->uq_key);
3694 		if (rv == -1)
3695 			return (EFAULT);
3696 		rv = thread_check_susp(td, true);
3697 		if (rv != 0)
3698 			return (rv);
3699 		goto again;
3700 	}
3701 	umtxq_lock(&uq->uq_key);
3702 	umtxq_unbusy(&uq->uq_key);
3703 
3704 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3705 
3706 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3707 		error = 0;
3708 	else {
3709 		umtxq_remove(uq);
3710 		if (timeout != NULL && (timeout->_flags & UMTX_ABSTIME) == 0) {
3711 			/* A relative timeout cannot be restarted. */
3712 			if (error == ERESTART)
3713 				error = EINTR;
3714 			if (error == EINTR) {
3715 				kern_clock_gettime(curthread, timo.clockid,
3716 				    &timo.cur);
3717 				timespecsub(&timo.end, &timo.cur,
3718 				    &timeout->_timeout);
3719 			}
3720 		}
3721 	}
3722 	umtxq_unlock(&uq->uq_key);
3723 	umtx_key_release(&uq->uq_key);
3724 	return (error);
3725 }
3726 
3727 /*
3728  * Signal a userland semaphore.
3729  */
3730 static int
3731 do_sem2_wake(struct thread *td, struct _usem2 *sem)
3732 {
3733 	struct umtx_key key;
3734 	int error, cnt, rv;
3735 	uint32_t count, flags;
3736 
3737 	rv = fueword32(&sem->_flags, &flags);
3738 	if (rv == -1)
3739 		return (EFAULT);
3740 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3741 		return (error);
3742 	umtxq_lock(&key);
3743 	umtxq_busy(&key);
3744 	cnt = umtxq_count(&key);
3745 	if (cnt > 0) {
3746 		/*
3747 		 * If this was the last sleeping thread, clear the waiters
3748 		 * flag in _count.
3749 		 */
3750 		if (cnt == 1) {
3751 			umtxq_unlock(&key);
3752 			rv = fueword32(&sem->_count, &count);
3753 			while (rv != -1 && count & USEM_HAS_WAITERS) {
3754 				rv = casueword32(&sem->_count, count, &count,
3755 				    count & ~USEM_HAS_WAITERS);
3756 				if (rv == 1) {
3757 					rv = thread_check_susp(td, true);
3758 					if (rv != 0)
3759 						break;
3760 				}
3761 			}
3762 			if (rv == -1)
3763 				error = EFAULT;
3764 			else if (rv > 0) {
3765 				error = rv;
3766 			}
3767 			umtxq_lock(&key);
3768 		}
3769 
3770 		umtxq_signal(&key, 1);
3771 	}
3772 	umtxq_unbusy(&key);
3773 	umtxq_unlock(&key);
3774 	umtx_key_release(&key);
3775 	return (error);
3776 }
3777 
3778 #ifdef COMPAT_FREEBSD10
3779 int
3780 freebsd10__umtx_lock(struct thread *td, struct freebsd10__umtx_lock_args *uap)
3781 {
3782 	return (do_lock_umtx(td, uap->umtx, td->td_tid, 0));
3783 }
3784 
3785 int
3786 freebsd10__umtx_unlock(struct thread *td,
3787     struct freebsd10__umtx_unlock_args *uap)
3788 {
3789 	return (do_unlock_umtx(td, uap->umtx, td->td_tid));
3790 }
3791 #endif
3792 
3793 inline int
3794 umtx_copyin_timeout(const void *uaddr, struct timespec *tsp)
3795 {
3796 	int error;
3797 
3798 	error = copyin(uaddr, tsp, sizeof(*tsp));
3799 	if (error == 0) {
3800 		if (!timespecvalid_interval(tsp))
3801 			error = EINVAL;
3802 	}
3803 	return (error);
3804 }
3805 
3806 static inline int
3807 umtx_copyin_umtx_time(const void *uaddr, size_t size, struct _umtx_time *tp)
3808 {
3809 	int error;
3810 
3811 	if (size <= sizeof(tp->_timeout)) {
3812 		tp->_clockid = CLOCK_REALTIME;
3813 		tp->_flags = 0;
3814 		error = copyin(uaddr, &tp->_timeout, sizeof(tp->_timeout));
3815 	} else
3816 		error = copyin(uaddr, tp, sizeof(*tp));
3817 	if (error != 0)
3818 		return (error);
3819 	if (!timespecvalid_interval(&tp->_timeout))
3820 		return (EINVAL);
3821 	return (0);
3822 }
3823 
3824 static int
3825 umtx_copyin_robust_lists(const void *uaddr, size_t size,
3826     struct umtx_robust_lists_params *rb)
3827 {
3828 
3829 	if (size > sizeof(*rb))
3830 		return (EINVAL);
3831 	return (copyin(uaddr, rb, size));
3832 }
3833 
3834 static int
3835 umtx_copyout_timeout(void *uaddr, size_t sz, struct timespec *tsp)
3836 {
3837 
3838 	/*
3839 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
3840 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
3841 	 * copyops.
3842 	 */
3843 	KASSERT(sz >= sizeof(*tsp),
3844 	    ("umtx_copyops specifies incorrect sizes"));
3845 
3846 	return (copyout(tsp, uaddr, sizeof(*tsp)));
3847 }
3848 
3849 #ifdef COMPAT_FREEBSD10
3850 static int
3851 __umtx_op_lock_umtx(struct thread *td, struct _umtx_op_args *uap,
3852     const struct umtx_copyops *ops)
3853 {
3854 	struct timespec *ts, timeout;
3855 	int error;
3856 
3857 	/* Allow a null timespec (wait forever). */
3858 	if (uap->uaddr2 == NULL)
3859 		ts = NULL;
3860 	else {
3861 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
3862 		if (error != 0)
3863 			return (error);
3864 		ts = &timeout;
3865 	}
3866 #ifdef COMPAT_FREEBSD32
3867 	if (ops->compat32)
3868 		return (do_lock_umtx32(td, uap->obj, uap->val, ts));
3869 #endif
3870 	return (do_lock_umtx(td, uap->obj, uap->val, ts));
3871 }
3872 
3873 static int
3874 __umtx_op_unlock_umtx(struct thread *td, struct _umtx_op_args *uap,
3875     const struct umtx_copyops *ops)
3876 {
3877 #ifdef COMPAT_FREEBSD32
3878 	if (ops->compat32)
3879 		return (do_unlock_umtx32(td, uap->obj, uap->val));
3880 #endif
3881 	return (do_unlock_umtx(td, uap->obj, uap->val));
3882 }
3883 #endif	/* COMPAT_FREEBSD10 */
3884 
3885 #if !defined(COMPAT_FREEBSD10)
3886 static int
3887 __umtx_op_unimpl(struct thread *td __unused, struct _umtx_op_args *uap __unused,
3888     const struct umtx_copyops *ops __unused)
3889 {
3890 	return (EOPNOTSUPP);
3891 }
3892 #endif	/* COMPAT_FREEBSD10 */
3893 
3894 static int
3895 __umtx_op_wait(struct thread *td, struct _umtx_op_args *uap,
3896     const struct umtx_copyops *ops)
3897 {
3898 	struct _umtx_time timeout, *tm_p;
3899 	int error;
3900 
3901 	if (uap->uaddr2 == NULL)
3902 		tm_p = NULL;
3903 	else {
3904 		error = ops->copyin_umtx_time(
3905 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3906 		if (error != 0)
3907 			return (error);
3908 		tm_p = &timeout;
3909 	}
3910 	return (do_wait(td, uap->obj, uap->val, tm_p, ops->compat32, 0));
3911 }
3912 
3913 static int
3914 __umtx_op_wait_uint(struct thread *td, struct _umtx_op_args *uap,
3915     const struct umtx_copyops *ops)
3916 {
3917 	struct _umtx_time timeout, *tm_p;
3918 	int error;
3919 
3920 	if (uap->uaddr2 == NULL)
3921 		tm_p = NULL;
3922 	else {
3923 		error = ops->copyin_umtx_time(
3924 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3925 		if (error != 0)
3926 			return (error);
3927 		tm_p = &timeout;
3928 	}
3929 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
3930 }
3931 
3932 static int
3933 __umtx_op_wait_uint_private(struct thread *td, struct _umtx_op_args *uap,
3934     const struct umtx_copyops *ops)
3935 {
3936 	struct _umtx_time *tm_p, timeout;
3937 	int error;
3938 
3939 	if (uap->uaddr2 == NULL)
3940 		tm_p = NULL;
3941 	else {
3942 		error = ops->copyin_umtx_time(
3943 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3944 		if (error != 0)
3945 			return (error);
3946 		tm_p = &timeout;
3947 	}
3948 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
3949 }
3950 
3951 static int
3952 __umtx_op_wake(struct thread *td, struct _umtx_op_args *uap,
3953     const struct umtx_copyops *ops __unused)
3954 {
3955 
3956 	return (kern_umtx_wake(td, uap->obj, uap->val, 0));
3957 }
3958 
3959 #define BATCH_SIZE	128
3960 static int
3961 __umtx_op_nwake_private_native(struct thread *td, struct _umtx_op_args *uap)
3962 {
3963 	char *uaddrs[BATCH_SIZE], **upp;
3964 	int count, error, i, pos, tocopy;
3965 
3966 	upp = (char **)uap->obj;
3967 	error = 0;
3968 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3969 	    pos += tocopy) {
3970 		tocopy = MIN(count, BATCH_SIZE);
3971 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(char *));
3972 		if (error != 0)
3973 			break;
3974 		for (i = 0; i < tocopy; ++i) {
3975 			kern_umtx_wake(td, uaddrs[i], INT_MAX, 1);
3976 		}
3977 		maybe_yield();
3978 	}
3979 	return (error);
3980 }
3981 
3982 static int
3983 __umtx_op_nwake_private_compat32(struct thread *td, struct _umtx_op_args *uap)
3984 {
3985 	uint32_t uaddrs[BATCH_SIZE], *upp;
3986 	int count, error, i, pos, tocopy;
3987 
3988 	upp = (uint32_t *)uap->obj;
3989 	error = 0;
3990 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3991 	    pos += tocopy) {
3992 		tocopy = MIN(count, BATCH_SIZE);
3993 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(uint32_t));
3994 		if (error != 0)
3995 			break;
3996 		for (i = 0; i < tocopy; ++i) {
3997 			kern_umtx_wake(td, (void *)(uintptr_t)uaddrs[i],
3998 			    INT_MAX, 1);
3999 		}
4000 		maybe_yield();
4001 	}
4002 	return (error);
4003 }
4004 
4005 static int
4006 __umtx_op_nwake_private(struct thread *td, struct _umtx_op_args *uap,
4007     const struct umtx_copyops *ops)
4008 {
4009 
4010 	if (ops->compat32)
4011 		return (__umtx_op_nwake_private_compat32(td, uap));
4012 	return (__umtx_op_nwake_private_native(td, uap));
4013 }
4014 
4015 static int
4016 __umtx_op_wake_private(struct thread *td, struct _umtx_op_args *uap,
4017     const struct umtx_copyops *ops __unused)
4018 {
4019 
4020 	return (kern_umtx_wake(td, uap->obj, uap->val, 1));
4021 }
4022 
4023 static int
4024 __umtx_op_lock_umutex(struct thread *td, struct _umtx_op_args *uap,
4025    const struct umtx_copyops *ops)
4026 {
4027 	struct _umtx_time *tm_p, timeout;
4028 	int error;
4029 
4030 	/* Allow a null timespec (wait forever). */
4031 	if (uap->uaddr2 == NULL)
4032 		tm_p = NULL;
4033 	else {
4034 		error = ops->copyin_umtx_time(
4035 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4036 		if (error != 0)
4037 			return (error);
4038 		tm_p = &timeout;
4039 	}
4040 	return (do_lock_umutex(td, uap->obj, tm_p, 0));
4041 }
4042 
4043 static int
4044 __umtx_op_trylock_umutex(struct thread *td, struct _umtx_op_args *uap,
4045     const struct umtx_copyops *ops __unused)
4046 {
4047 
4048 	return (do_lock_umutex(td, uap->obj, NULL, _UMUTEX_TRY));
4049 }
4050 
4051 static int
4052 __umtx_op_wait_umutex(struct thread *td, struct _umtx_op_args *uap,
4053     const struct umtx_copyops *ops)
4054 {
4055 	struct _umtx_time *tm_p, timeout;
4056 	int error;
4057 
4058 	/* Allow a null timespec (wait forever). */
4059 	if (uap->uaddr2 == NULL)
4060 		tm_p = NULL;
4061 	else {
4062 		error = ops->copyin_umtx_time(
4063 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4064 		if (error != 0)
4065 			return (error);
4066 		tm_p = &timeout;
4067 	}
4068 	return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
4069 }
4070 
4071 static int
4072 __umtx_op_wake_umutex(struct thread *td, struct _umtx_op_args *uap,
4073     const struct umtx_copyops *ops __unused)
4074 {
4075 
4076 	return (do_wake_umutex(td, uap->obj));
4077 }
4078 
4079 static int
4080 __umtx_op_unlock_umutex(struct thread *td, struct _umtx_op_args *uap,
4081     const struct umtx_copyops *ops __unused)
4082 {
4083 
4084 	return (do_unlock_umutex(td, uap->obj, false));
4085 }
4086 
4087 static int
4088 __umtx_op_set_ceiling(struct thread *td, struct _umtx_op_args *uap,
4089     const struct umtx_copyops *ops __unused)
4090 {
4091 
4092 	return (do_set_ceiling(td, uap->obj, uap->val, uap->uaddr1));
4093 }
4094 
4095 static int
4096 __umtx_op_cv_wait(struct thread *td, struct _umtx_op_args *uap,
4097     const struct umtx_copyops *ops)
4098 {
4099 	struct timespec *ts, timeout;
4100 	int error;
4101 
4102 	/* Allow a null timespec (wait forever). */
4103 	if (uap->uaddr2 == NULL)
4104 		ts = NULL;
4105 	else {
4106 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
4107 		if (error != 0)
4108 			return (error);
4109 		ts = &timeout;
4110 	}
4111 	return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
4112 }
4113 
4114 static int
4115 __umtx_op_cv_signal(struct thread *td, struct _umtx_op_args *uap,
4116     const struct umtx_copyops *ops __unused)
4117 {
4118 
4119 	return (do_cv_signal(td, uap->obj));
4120 }
4121 
4122 static int
4123 __umtx_op_cv_broadcast(struct thread *td, struct _umtx_op_args *uap,
4124     const struct umtx_copyops *ops __unused)
4125 {
4126 
4127 	return (do_cv_broadcast(td, uap->obj));
4128 }
4129 
4130 static int
4131 __umtx_op_rw_rdlock(struct thread *td, struct _umtx_op_args *uap,
4132     const struct umtx_copyops *ops)
4133 {
4134 	struct _umtx_time timeout;
4135 	int error;
4136 
4137 	/* Allow a null timespec (wait forever). */
4138 	if (uap->uaddr2 == NULL) {
4139 		error = do_rw_rdlock(td, uap->obj, uap->val, 0);
4140 	} else {
4141 		error = ops->copyin_umtx_time(uap->uaddr2,
4142 		   (size_t)uap->uaddr1, &timeout);
4143 		if (error != 0)
4144 			return (error);
4145 		error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
4146 	}
4147 	return (error);
4148 }
4149 
4150 static int
4151 __umtx_op_rw_wrlock(struct thread *td, struct _umtx_op_args *uap,
4152     const struct umtx_copyops *ops)
4153 {
4154 	struct _umtx_time timeout;
4155 	int error;
4156 
4157 	/* Allow a null timespec (wait forever). */
4158 	if (uap->uaddr2 == NULL) {
4159 		error = do_rw_wrlock(td, uap->obj, 0);
4160 	} else {
4161 		error = ops->copyin_umtx_time(uap->uaddr2,
4162 		   (size_t)uap->uaddr1, &timeout);
4163 		if (error != 0)
4164 			return (error);
4165 
4166 		error = do_rw_wrlock(td, uap->obj, &timeout);
4167 	}
4168 	return (error);
4169 }
4170 
4171 static int
4172 __umtx_op_rw_unlock(struct thread *td, struct _umtx_op_args *uap,
4173     const struct umtx_copyops *ops __unused)
4174 {
4175 
4176 	return (do_rw_unlock(td, uap->obj));
4177 }
4178 
4179 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4180 static int
4181 __umtx_op_sem_wait(struct thread *td, struct _umtx_op_args *uap,
4182     const struct umtx_copyops *ops)
4183 {
4184 	struct _umtx_time *tm_p, timeout;
4185 	int error;
4186 
4187 	/* Allow a null timespec (wait forever). */
4188 	if (uap->uaddr2 == NULL)
4189 		tm_p = NULL;
4190 	else {
4191 		error = ops->copyin_umtx_time(
4192 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4193 		if (error != 0)
4194 			return (error);
4195 		tm_p = &timeout;
4196 	}
4197 	return (do_sem_wait(td, uap->obj, tm_p));
4198 }
4199 
4200 static int
4201 __umtx_op_sem_wake(struct thread *td, struct _umtx_op_args *uap,
4202     const struct umtx_copyops *ops __unused)
4203 {
4204 
4205 	return (do_sem_wake(td, uap->obj));
4206 }
4207 #endif
4208 
4209 static int
4210 __umtx_op_wake2_umutex(struct thread *td, struct _umtx_op_args *uap,
4211     const struct umtx_copyops *ops __unused)
4212 {
4213 
4214 	return (do_wake2_umutex(td, uap->obj, uap->val));
4215 }
4216 
4217 static int
4218 __umtx_op_sem2_wait(struct thread *td, struct _umtx_op_args *uap,
4219     const struct umtx_copyops *ops)
4220 {
4221 	struct _umtx_time *tm_p, timeout;
4222 	size_t uasize;
4223 	int error;
4224 
4225 	/* Allow a null timespec (wait forever). */
4226 	if (uap->uaddr2 == NULL) {
4227 		uasize = 0;
4228 		tm_p = NULL;
4229 	} else {
4230 		uasize = (size_t)uap->uaddr1;
4231 		error = ops->copyin_umtx_time(uap->uaddr2, uasize, &timeout);
4232 		if (error != 0)
4233 			return (error);
4234 		tm_p = &timeout;
4235 	}
4236 	error = do_sem2_wait(td, uap->obj, tm_p);
4237 	if (error == EINTR && uap->uaddr2 != NULL &&
4238 	    (timeout._flags & UMTX_ABSTIME) == 0 &&
4239 	    uasize >= ops->umtx_time_sz + ops->timespec_sz) {
4240 		error = ops->copyout_timeout(
4241 		    (void *)((uintptr_t)uap->uaddr2 + ops->umtx_time_sz),
4242 		    uasize - ops->umtx_time_sz, &timeout._timeout);
4243 		if (error == 0) {
4244 			error = EINTR;
4245 		}
4246 	}
4247 
4248 	return (error);
4249 }
4250 
4251 static int
4252 __umtx_op_sem2_wake(struct thread *td, struct _umtx_op_args *uap,
4253     const struct umtx_copyops *ops __unused)
4254 {
4255 
4256 	return (do_sem2_wake(td, uap->obj));
4257 }
4258 
4259 #define	USHM_OBJ_UMTX(o)						\
4260     ((struct umtx_shm_obj_list *)(&(o)->umtx_data))
4261 
4262 #define	USHMF_REG_LINKED	0x0001
4263 #define	USHMF_OBJ_LINKED	0x0002
4264 struct umtx_shm_reg {
4265 	TAILQ_ENTRY(umtx_shm_reg) ushm_reg_link;
4266 	LIST_ENTRY(umtx_shm_reg) ushm_obj_link;
4267 	struct umtx_key		ushm_key;
4268 	struct ucred		*ushm_cred;
4269 	struct shmfd		*ushm_obj;
4270 	u_int			ushm_refcnt;
4271 	u_int			ushm_flags;
4272 };
4273 
4274 LIST_HEAD(umtx_shm_obj_list, umtx_shm_reg);
4275 TAILQ_HEAD(umtx_shm_reg_head, umtx_shm_reg);
4276 
4277 static uma_zone_t umtx_shm_reg_zone;
4278 static struct umtx_shm_reg_head umtx_shm_registry[UMTX_CHAINS];
4279 static struct mtx umtx_shm_lock;
4280 static struct umtx_shm_reg_head umtx_shm_reg_delfree =
4281     TAILQ_HEAD_INITIALIZER(umtx_shm_reg_delfree);
4282 
4283 static void umtx_shm_free_reg(struct umtx_shm_reg *reg);
4284 
4285 static void
4286 umtx_shm_reg_delfree_tq(void *context __unused, int pending __unused)
4287 {
4288 	struct umtx_shm_reg_head d;
4289 	struct umtx_shm_reg *reg, *reg1;
4290 
4291 	TAILQ_INIT(&d);
4292 	mtx_lock(&umtx_shm_lock);
4293 	TAILQ_CONCAT(&d, &umtx_shm_reg_delfree, ushm_reg_link);
4294 	mtx_unlock(&umtx_shm_lock);
4295 	TAILQ_FOREACH_SAFE(reg, &d, ushm_reg_link, reg1) {
4296 		TAILQ_REMOVE(&d, reg, ushm_reg_link);
4297 		umtx_shm_free_reg(reg);
4298 	}
4299 }
4300 
4301 static struct task umtx_shm_reg_delfree_task =
4302     TASK_INITIALIZER(0, umtx_shm_reg_delfree_tq, NULL);
4303 
4304 static struct umtx_shm_reg *
4305 umtx_shm_find_reg_locked(const struct umtx_key *key)
4306 {
4307 	struct umtx_shm_reg *reg;
4308 	struct umtx_shm_reg_head *reg_head;
4309 
4310 	KASSERT(key->shared, ("umtx_p_find_rg: private key"));
4311 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4312 	reg_head = &umtx_shm_registry[key->hash];
4313 	TAILQ_FOREACH(reg, reg_head, ushm_reg_link) {
4314 		KASSERT(reg->ushm_key.shared,
4315 		    ("non-shared key on reg %p %d", reg, reg->ushm_key.shared));
4316 		if (reg->ushm_key.info.shared.object ==
4317 		    key->info.shared.object &&
4318 		    reg->ushm_key.info.shared.offset ==
4319 		    key->info.shared.offset) {
4320 			KASSERT(reg->ushm_key.type == TYPE_SHM, ("TYPE_USHM"));
4321 			KASSERT(reg->ushm_refcnt > 0,
4322 			    ("reg %p refcnt 0 onlist", reg));
4323 			KASSERT((reg->ushm_flags & USHMF_REG_LINKED) != 0,
4324 			    ("reg %p not linked", reg));
4325 			reg->ushm_refcnt++;
4326 			return (reg);
4327 		}
4328 	}
4329 	return (NULL);
4330 }
4331 
4332 static struct umtx_shm_reg *
4333 umtx_shm_find_reg(const struct umtx_key *key)
4334 {
4335 	struct umtx_shm_reg *reg;
4336 
4337 	mtx_lock(&umtx_shm_lock);
4338 	reg = umtx_shm_find_reg_locked(key);
4339 	mtx_unlock(&umtx_shm_lock);
4340 	return (reg);
4341 }
4342 
4343 static void
4344 umtx_shm_free_reg(struct umtx_shm_reg *reg)
4345 {
4346 
4347 	chgumtxcnt(reg->ushm_cred->cr_ruidinfo, -1, 0);
4348 	crfree(reg->ushm_cred);
4349 	shm_drop(reg->ushm_obj);
4350 	uma_zfree(umtx_shm_reg_zone, reg);
4351 }
4352 
4353 static bool
4354 umtx_shm_unref_reg_locked(struct umtx_shm_reg *reg, bool force)
4355 {
4356 	bool res;
4357 
4358 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4359 	KASSERT(reg->ushm_refcnt > 0, ("ushm_reg %p refcnt 0", reg));
4360 	reg->ushm_refcnt--;
4361 	res = reg->ushm_refcnt == 0;
4362 	if (res || force) {
4363 		if ((reg->ushm_flags & USHMF_REG_LINKED) != 0) {
4364 			TAILQ_REMOVE(&umtx_shm_registry[reg->ushm_key.hash],
4365 			    reg, ushm_reg_link);
4366 			reg->ushm_flags &= ~USHMF_REG_LINKED;
4367 		}
4368 		if ((reg->ushm_flags & USHMF_OBJ_LINKED) != 0) {
4369 			LIST_REMOVE(reg, ushm_obj_link);
4370 			reg->ushm_flags &= ~USHMF_OBJ_LINKED;
4371 		}
4372 	}
4373 	return (res);
4374 }
4375 
4376 static void
4377 umtx_shm_unref_reg(struct umtx_shm_reg *reg, bool force)
4378 {
4379 	vm_object_t object;
4380 	bool dofree;
4381 
4382 	if (force) {
4383 		object = reg->ushm_obj->shm_object;
4384 		VM_OBJECT_WLOCK(object);
4385 		vm_object_set_flag(object, OBJ_UMTXDEAD);
4386 		VM_OBJECT_WUNLOCK(object);
4387 	}
4388 	mtx_lock(&umtx_shm_lock);
4389 	dofree = umtx_shm_unref_reg_locked(reg, force);
4390 	mtx_unlock(&umtx_shm_lock);
4391 	if (dofree)
4392 		umtx_shm_free_reg(reg);
4393 }
4394 
4395 void
4396 umtx_shm_object_init(vm_object_t object)
4397 {
4398 
4399 	LIST_INIT(USHM_OBJ_UMTX(object));
4400 }
4401 
4402 void
4403 umtx_shm_object_terminated(vm_object_t object)
4404 {
4405 	struct umtx_shm_reg *reg, *reg1;
4406 	bool dofree;
4407 
4408 	if (LIST_EMPTY(USHM_OBJ_UMTX(object)))
4409 		return;
4410 
4411 	dofree = false;
4412 	mtx_lock(&umtx_shm_lock);
4413 	LIST_FOREACH_SAFE(reg, USHM_OBJ_UMTX(object), ushm_obj_link, reg1) {
4414 		if (umtx_shm_unref_reg_locked(reg, true)) {
4415 			TAILQ_INSERT_TAIL(&umtx_shm_reg_delfree, reg,
4416 			    ushm_reg_link);
4417 			dofree = true;
4418 		}
4419 	}
4420 	mtx_unlock(&umtx_shm_lock);
4421 	if (dofree)
4422 		taskqueue_enqueue(taskqueue_thread, &umtx_shm_reg_delfree_task);
4423 }
4424 
4425 static int
4426 umtx_shm_create_reg(struct thread *td, const struct umtx_key *key,
4427     struct umtx_shm_reg **res)
4428 {
4429 	struct umtx_shm_reg *reg, *reg1;
4430 	struct ucred *cred;
4431 	int error;
4432 
4433 	reg = umtx_shm_find_reg(key);
4434 	if (reg != NULL) {
4435 		*res = reg;
4436 		return (0);
4437 	}
4438 	cred = td->td_ucred;
4439 	if (!chgumtxcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_UMTXP)))
4440 		return (ENOMEM);
4441 	reg = uma_zalloc(umtx_shm_reg_zone, M_WAITOK | M_ZERO);
4442 	reg->ushm_refcnt = 1;
4443 	bcopy(key, &reg->ushm_key, sizeof(*key));
4444 	reg->ushm_obj = shm_alloc(td->td_ucred, O_RDWR, false);
4445 	reg->ushm_cred = crhold(cred);
4446 	error = shm_dotruncate(reg->ushm_obj, PAGE_SIZE);
4447 	if (error != 0) {
4448 		umtx_shm_free_reg(reg);
4449 		return (error);
4450 	}
4451 	mtx_lock(&umtx_shm_lock);
4452 	reg1 = umtx_shm_find_reg_locked(key);
4453 	if (reg1 != NULL) {
4454 		mtx_unlock(&umtx_shm_lock);
4455 		umtx_shm_free_reg(reg);
4456 		*res = reg1;
4457 		return (0);
4458 	}
4459 	reg->ushm_refcnt++;
4460 	TAILQ_INSERT_TAIL(&umtx_shm_registry[key->hash], reg, ushm_reg_link);
4461 	LIST_INSERT_HEAD(USHM_OBJ_UMTX(key->info.shared.object), reg,
4462 	    ushm_obj_link);
4463 	reg->ushm_flags = USHMF_REG_LINKED | USHMF_OBJ_LINKED;
4464 	mtx_unlock(&umtx_shm_lock);
4465 	*res = reg;
4466 	return (0);
4467 }
4468 
4469 static int
4470 umtx_shm_alive(struct thread *td, void *addr)
4471 {
4472 	vm_map_t map;
4473 	vm_map_entry_t entry;
4474 	vm_object_t object;
4475 	vm_pindex_t pindex;
4476 	vm_prot_t prot;
4477 	int res, ret;
4478 	boolean_t wired;
4479 
4480 	map = &td->td_proc->p_vmspace->vm_map;
4481 	res = vm_map_lookup(&map, (uintptr_t)addr, VM_PROT_READ, &entry,
4482 	    &object, &pindex, &prot, &wired);
4483 	if (res != KERN_SUCCESS)
4484 		return (EFAULT);
4485 	if (object == NULL)
4486 		ret = EINVAL;
4487 	else
4488 		ret = (object->flags & OBJ_UMTXDEAD) != 0 ? ENOTTY : 0;
4489 	vm_map_lookup_done(map, entry);
4490 	return (ret);
4491 }
4492 
4493 static void
4494 umtx_shm_init(void)
4495 {
4496 	int i;
4497 
4498 	umtx_shm_reg_zone = uma_zcreate("umtx_shm", sizeof(struct umtx_shm_reg),
4499 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
4500 	mtx_init(&umtx_shm_lock, "umtxshm", NULL, MTX_DEF);
4501 	for (i = 0; i < nitems(umtx_shm_registry); i++)
4502 		TAILQ_INIT(&umtx_shm_registry[i]);
4503 }
4504 
4505 static int
4506 umtx_shm(struct thread *td, void *addr, u_int flags)
4507 {
4508 	struct umtx_key key;
4509 	struct umtx_shm_reg *reg;
4510 	struct file *fp;
4511 	int error, fd;
4512 
4513 	if (__bitcount(flags & (UMTX_SHM_CREAT | UMTX_SHM_LOOKUP |
4514 	    UMTX_SHM_DESTROY| UMTX_SHM_ALIVE)) != 1)
4515 		return (EINVAL);
4516 	if ((flags & UMTX_SHM_ALIVE) != 0)
4517 		return (umtx_shm_alive(td, addr));
4518 	error = umtx_key_get(addr, TYPE_SHM, PROCESS_SHARE, &key);
4519 	if (error != 0)
4520 		return (error);
4521 	KASSERT(key.shared == 1, ("non-shared key"));
4522 	if ((flags & UMTX_SHM_CREAT) != 0) {
4523 		error = umtx_shm_create_reg(td, &key, &reg);
4524 	} else {
4525 		reg = umtx_shm_find_reg(&key);
4526 		if (reg == NULL)
4527 			error = ESRCH;
4528 	}
4529 	umtx_key_release(&key);
4530 	if (error != 0)
4531 		return (error);
4532 	KASSERT(reg != NULL, ("no reg"));
4533 	if ((flags & UMTX_SHM_DESTROY) != 0) {
4534 		umtx_shm_unref_reg(reg, true);
4535 	} else {
4536 #if 0
4537 #ifdef MAC
4538 		error = mac_posixshm_check_open(td->td_ucred,
4539 		    reg->ushm_obj, FFLAGS(O_RDWR));
4540 		if (error == 0)
4541 #endif
4542 			error = shm_access(reg->ushm_obj, td->td_ucred,
4543 			    FFLAGS(O_RDWR));
4544 		if (error == 0)
4545 #endif
4546 			error = falloc_caps(td, &fp, &fd, O_CLOEXEC, NULL);
4547 		if (error == 0) {
4548 			shm_hold(reg->ushm_obj);
4549 			finit(fp, FFLAGS(O_RDWR), DTYPE_SHM, reg->ushm_obj,
4550 			    &shm_ops);
4551 			td->td_retval[0] = fd;
4552 			fdrop(fp, td);
4553 		}
4554 	}
4555 	umtx_shm_unref_reg(reg, false);
4556 	return (error);
4557 }
4558 
4559 static int
4560 __umtx_op_shm(struct thread *td, struct _umtx_op_args *uap,
4561     const struct umtx_copyops *ops __unused)
4562 {
4563 
4564 	return (umtx_shm(td, uap->uaddr1, uap->val));
4565 }
4566 
4567 static int
4568 __umtx_op_robust_lists(struct thread *td, struct _umtx_op_args *uap,
4569     const struct umtx_copyops *ops)
4570 {
4571 	struct umtx_robust_lists_params rb;
4572 	int error;
4573 
4574 	if (ops->compat32) {
4575 		if ((td->td_pflags2 & TDP2_COMPAT32RB) == 0 &&
4576 		    (td->td_rb_list != 0 || td->td_rbp_list != 0 ||
4577 		    td->td_rb_inact != 0))
4578 			return (EBUSY);
4579 	} else if ((td->td_pflags2 & TDP2_COMPAT32RB) != 0) {
4580 		return (EBUSY);
4581 	}
4582 
4583 	bzero(&rb, sizeof(rb));
4584 	error = ops->copyin_robust_lists(uap->uaddr1, uap->val, &rb);
4585 	if (error != 0)
4586 		return (error);
4587 
4588 	if (ops->compat32)
4589 		td->td_pflags2 |= TDP2_COMPAT32RB;
4590 
4591 	td->td_rb_list = rb.robust_list_offset;
4592 	td->td_rbp_list = rb.robust_priv_list_offset;
4593 	td->td_rb_inact = rb.robust_inact_offset;
4594 	return (0);
4595 }
4596 
4597 #if defined(__i386__) || defined(__amd64__)
4598 /*
4599  * Provide the standard 32-bit definitions for x86, since native/compat32 use a
4600  * 32-bit time_t there.  Other architectures just need the i386 definitions
4601  * along with their standard compat32.
4602  */
4603 struct timespecx32 {
4604 	int64_t			tv_sec;
4605 	int32_t			tv_nsec;
4606 };
4607 
4608 struct umtx_timex32 {
4609 	struct	timespecx32	_timeout;
4610 	uint32_t		_flags;
4611 	uint32_t		_clockid;
4612 };
4613 
4614 #ifndef __i386__
4615 #define	timespeci386	timespec32
4616 #define	umtx_timei386	umtx_time32
4617 #endif
4618 #else /* !__i386__ && !__amd64__ */
4619 /* 32-bit architectures can emulate i386, so define these almost everywhere. */
4620 struct timespeci386 {
4621 	int32_t			tv_sec;
4622 	int32_t			tv_nsec;
4623 };
4624 
4625 struct umtx_timei386 {
4626 	struct	timespeci386	_timeout;
4627 	uint32_t		_flags;
4628 	uint32_t		_clockid;
4629 };
4630 
4631 #if defined(__LP64__)
4632 #define	timespecx32	timespec32
4633 #define	umtx_timex32	umtx_time32
4634 #endif
4635 #endif
4636 
4637 static int
4638 umtx_copyin_robust_lists32(const void *uaddr, size_t size,
4639     struct umtx_robust_lists_params *rbp)
4640 {
4641 	struct umtx_robust_lists_params_compat32 rb32;
4642 	int error;
4643 
4644 	if (size > sizeof(rb32))
4645 		return (EINVAL);
4646 	bzero(&rb32, sizeof(rb32));
4647 	error = copyin(uaddr, &rb32, size);
4648 	if (error != 0)
4649 		return (error);
4650 	CP(rb32, *rbp, robust_list_offset);
4651 	CP(rb32, *rbp, robust_priv_list_offset);
4652 	CP(rb32, *rbp, robust_inact_offset);
4653 	return (0);
4654 }
4655 
4656 #ifndef __i386__
4657 static inline int
4658 umtx_copyin_timeouti386(const void *uaddr, struct timespec *tsp)
4659 {
4660 	struct timespeci386 ts32;
4661 	int error;
4662 
4663 	error = copyin(uaddr, &ts32, sizeof(ts32));
4664 	if (error == 0) {
4665 		if (!timespecvalid_interval(&ts32))
4666 			error = EINVAL;
4667 		else {
4668 			CP(ts32, *tsp, tv_sec);
4669 			CP(ts32, *tsp, tv_nsec);
4670 		}
4671 	}
4672 	return (error);
4673 }
4674 
4675 static inline int
4676 umtx_copyin_umtx_timei386(const void *uaddr, size_t size, struct _umtx_time *tp)
4677 {
4678 	struct umtx_timei386 t32;
4679 	int error;
4680 
4681 	t32._clockid = CLOCK_REALTIME;
4682 	t32._flags   = 0;
4683 	if (size <= sizeof(t32._timeout))
4684 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4685 	else
4686 		error = copyin(uaddr, &t32, sizeof(t32));
4687 	if (error != 0)
4688 		return (error);
4689 	if (!timespecvalid_interval(&t32._timeout))
4690 		return (EINVAL);
4691 	TS_CP(t32, *tp, _timeout);
4692 	CP(t32, *tp, _flags);
4693 	CP(t32, *tp, _clockid);
4694 	return (0);
4695 }
4696 
4697 static int
4698 umtx_copyout_timeouti386(void *uaddr, size_t sz, struct timespec *tsp)
4699 {
4700 	struct timespeci386 remain32 = {
4701 		.tv_sec = tsp->tv_sec,
4702 		.tv_nsec = tsp->tv_nsec,
4703 	};
4704 
4705 	/*
4706 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4707 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4708 	 * copyops.
4709 	 */
4710 	KASSERT(sz >= sizeof(remain32),
4711 	    ("umtx_copyops specifies incorrect sizes"));
4712 
4713 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4714 }
4715 #endif /* !__i386__ */
4716 
4717 #if defined(__i386__) || defined(__LP64__)
4718 static inline int
4719 umtx_copyin_timeoutx32(const void *uaddr, struct timespec *tsp)
4720 {
4721 	struct timespecx32 ts32;
4722 	int error;
4723 
4724 	error = copyin(uaddr, &ts32, sizeof(ts32));
4725 	if (error == 0) {
4726 		if (!timespecvalid_interval(&ts32))
4727 			error = EINVAL;
4728 		else {
4729 			CP(ts32, *tsp, tv_sec);
4730 			CP(ts32, *tsp, tv_nsec);
4731 		}
4732 	}
4733 	return (error);
4734 }
4735 
4736 static inline int
4737 umtx_copyin_umtx_timex32(const void *uaddr, size_t size, struct _umtx_time *tp)
4738 {
4739 	struct umtx_timex32 t32;
4740 	int error;
4741 
4742 	t32._clockid = CLOCK_REALTIME;
4743 	t32._flags   = 0;
4744 	if (size <= sizeof(t32._timeout))
4745 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4746 	else
4747 		error = copyin(uaddr, &t32, sizeof(t32));
4748 	if (error != 0)
4749 		return (error);
4750 	if (!timespecvalid_interval(&t32._timeout))
4751 		return (EINVAL);
4752 	TS_CP(t32, *tp, _timeout);
4753 	CP(t32, *tp, _flags);
4754 	CP(t32, *tp, _clockid);
4755 	return (0);
4756 }
4757 
4758 static int
4759 umtx_copyout_timeoutx32(void *uaddr, size_t sz, struct timespec *tsp)
4760 {
4761 	struct timespecx32 remain32 = {
4762 		.tv_sec = tsp->tv_sec,
4763 		.tv_nsec = tsp->tv_nsec,
4764 	};
4765 
4766 	/*
4767 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4768 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4769 	 * copyops.
4770 	 */
4771 	KASSERT(sz >= sizeof(remain32),
4772 	    ("umtx_copyops specifies incorrect sizes"));
4773 
4774 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4775 }
4776 #endif /* __i386__ || __LP64__ */
4777 
4778 typedef int (*_umtx_op_func)(struct thread *td, struct _umtx_op_args *uap,
4779     const struct umtx_copyops *umtx_ops);
4780 
4781 static const _umtx_op_func op_table[] = {
4782 #ifdef COMPAT_FREEBSD10
4783 	[UMTX_OP_LOCK]		= __umtx_op_lock_umtx,
4784 	[UMTX_OP_UNLOCK]	= __umtx_op_unlock_umtx,
4785 #else
4786 	[UMTX_OP_LOCK]		= __umtx_op_unimpl,
4787 	[UMTX_OP_UNLOCK]	= __umtx_op_unimpl,
4788 #endif
4789 	[UMTX_OP_WAIT]		= __umtx_op_wait,
4790 	[UMTX_OP_WAKE]		= __umtx_op_wake,
4791 	[UMTX_OP_MUTEX_TRYLOCK]	= __umtx_op_trylock_umutex,
4792 	[UMTX_OP_MUTEX_LOCK]	= __umtx_op_lock_umutex,
4793 	[UMTX_OP_MUTEX_UNLOCK]	= __umtx_op_unlock_umutex,
4794 	[UMTX_OP_SET_CEILING]	= __umtx_op_set_ceiling,
4795 	[UMTX_OP_CV_WAIT]	= __umtx_op_cv_wait,
4796 	[UMTX_OP_CV_SIGNAL]	= __umtx_op_cv_signal,
4797 	[UMTX_OP_CV_BROADCAST]	= __umtx_op_cv_broadcast,
4798 	[UMTX_OP_WAIT_UINT]	= __umtx_op_wait_uint,
4799 	[UMTX_OP_RW_RDLOCK]	= __umtx_op_rw_rdlock,
4800 	[UMTX_OP_RW_WRLOCK]	= __umtx_op_rw_wrlock,
4801 	[UMTX_OP_RW_UNLOCK]	= __umtx_op_rw_unlock,
4802 	[UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private,
4803 	[UMTX_OP_WAKE_PRIVATE]	= __umtx_op_wake_private,
4804 	[UMTX_OP_MUTEX_WAIT]	= __umtx_op_wait_umutex,
4805 	[UMTX_OP_MUTEX_WAKE]	= __umtx_op_wake_umutex,
4806 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4807 	[UMTX_OP_SEM_WAIT]	= __umtx_op_sem_wait,
4808 	[UMTX_OP_SEM_WAKE]	= __umtx_op_sem_wake,
4809 #else
4810 	[UMTX_OP_SEM_WAIT]	= __umtx_op_unimpl,
4811 	[UMTX_OP_SEM_WAKE]	= __umtx_op_unimpl,
4812 #endif
4813 	[UMTX_OP_NWAKE_PRIVATE]	= __umtx_op_nwake_private,
4814 	[UMTX_OP_MUTEX_WAKE2]	= __umtx_op_wake2_umutex,
4815 	[UMTX_OP_SEM2_WAIT]	= __umtx_op_sem2_wait,
4816 	[UMTX_OP_SEM2_WAKE]	= __umtx_op_sem2_wake,
4817 	[UMTX_OP_SHM]		= __umtx_op_shm,
4818 	[UMTX_OP_ROBUST_LISTS]	= __umtx_op_robust_lists,
4819 };
4820 
4821 static const struct umtx_copyops umtx_native_ops = {
4822 	.copyin_timeout = umtx_copyin_timeout,
4823 	.copyin_umtx_time = umtx_copyin_umtx_time,
4824 	.copyin_robust_lists = umtx_copyin_robust_lists,
4825 	.copyout_timeout = umtx_copyout_timeout,
4826 	.timespec_sz = sizeof(struct timespec),
4827 	.umtx_time_sz = sizeof(struct _umtx_time),
4828 };
4829 
4830 #ifndef __i386__
4831 static const struct umtx_copyops umtx_native_opsi386 = {
4832 	.copyin_timeout = umtx_copyin_timeouti386,
4833 	.copyin_umtx_time = umtx_copyin_umtx_timei386,
4834 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4835 	.copyout_timeout = umtx_copyout_timeouti386,
4836 	.timespec_sz = sizeof(struct timespeci386),
4837 	.umtx_time_sz = sizeof(struct umtx_timei386),
4838 	.compat32 = true,
4839 };
4840 #endif
4841 
4842 #if defined(__i386__) || defined(__LP64__)
4843 /* i386 can emulate other 32-bit archs, too! */
4844 static const struct umtx_copyops umtx_native_opsx32 = {
4845 	.copyin_timeout = umtx_copyin_timeoutx32,
4846 	.copyin_umtx_time = umtx_copyin_umtx_timex32,
4847 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4848 	.copyout_timeout = umtx_copyout_timeoutx32,
4849 	.timespec_sz = sizeof(struct timespecx32),
4850 	.umtx_time_sz = sizeof(struct umtx_timex32),
4851 	.compat32 = true,
4852 };
4853 
4854 #ifdef COMPAT_FREEBSD32
4855 #ifdef __amd64__
4856 #define	umtx_native_ops32	umtx_native_opsi386
4857 #else
4858 #define	umtx_native_ops32	umtx_native_opsx32
4859 #endif
4860 #endif /* COMPAT_FREEBSD32 */
4861 #endif /* __i386__ || __LP64__ */
4862 
4863 #define	UMTX_OP__FLAGS	(UMTX_OP__32BIT | UMTX_OP__I386)
4864 
4865 static int
4866 kern__umtx_op(struct thread *td, void *obj, int op, unsigned long val,
4867     void *uaddr1, void *uaddr2, const struct umtx_copyops *ops)
4868 {
4869 	struct _umtx_op_args uap = {
4870 		.obj = obj,
4871 		.op = op & ~UMTX_OP__FLAGS,
4872 		.val = val,
4873 		.uaddr1 = uaddr1,
4874 		.uaddr2 = uaddr2
4875 	};
4876 
4877 	if ((uap.op >= nitems(op_table)))
4878 		return (EINVAL);
4879 	return ((*op_table[uap.op])(td, &uap, ops));
4880 }
4881 
4882 int
4883 sys__umtx_op(struct thread *td, struct _umtx_op_args *uap)
4884 {
4885 	static const struct umtx_copyops *umtx_ops;
4886 
4887 	umtx_ops = &umtx_native_ops;
4888 #ifdef __LP64__
4889 	if ((uap->op & (UMTX_OP__32BIT | UMTX_OP__I386)) != 0) {
4890 		if ((uap->op & UMTX_OP__I386) != 0)
4891 			umtx_ops = &umtx_native_opsi386;
4892 		else
4893 			umtx_ops = &umtx_native_opsx32;
4894 	}
4895 #elif !defined(__i386__)
4896 	/* We consider UMTX_OP__32BIT a nop on !i386 ILP32. */
4897 	if ((uap->op & UMTX_OP__I386) != 0)
4898 		umtx_ops = &umtx_native_opsi386;
4899 #else
4900 	/* Likewise, UMTX_OP__I386 is a nop on i386. */
4901 	if ((uap->op & UMTX_OP__32BIT) != 0)
4902 		umtx_ops = &umtx_native_opsx32;
4903 #endif
4904 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4905 	    uap->uaddr2, umtx_ops));
4906 }
4907 
4908 #ifdef COMPAT_FREEBSD32
4909 #ifdef COMPAT_FREEBSD10
4910 int
4911 freebsd10_freebsd32__umtx_lock(struct thread *td,
4912     struct freebsd10_freebsd32__umtx_lock_args *uap)
4913 {
4914 	return (do_lock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid, NULL));
4915 }
4916 
4917 int
4918 freebsd10_freebsd32__umtx_unlock(struct thread *td,
4919     struct freebsd10_freebsd32__umtx_unlock_args *uap)
4920 {
4921 	return (do_unlock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid));
4922 }
4923 #endif /* COMPAT_FREEBSD10 */
4924 
4925 int
4926 freebsd32__umtx_op(struct thread *td, struct freebsd32__umtx_op_args *uap)
4927 {
4928 
4929 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4930 	    uap->uaddr2, &umtx_native_ops32));
4931 }
4932 #endif /* COMPAT_FREEBSD32 */
4933 
4934 void
4935 umtx_thread_init(struct thread *td)
4936 {
4937 
4938 	td->td_umtxq = umtxq_alloc();
4939 	td->td_umtxq->uq_thread = td;
4940 }
4941 
4942 void
4943 umtx_thread_fini(struct thread *td)
4944 {
4945 
4946 	umtxq_free(td->td_umtxq);
4947 }
4948 
4949 /*
4950  * It will be called when new thread is created, e.g fork().
4951  */
4952 void
4953 umtx_thread_alloc(struct thread *td)
4954 {
4955 	struct umtx_q *uq;
4956 
4957 	uq = td->td_umtxq;
4958 	uq->uq_inherited_pri = PRI_MAX;
4959 
4960 	KASSERT(uq->uq_flags == 0, ("uq_flags != 0"));
4961 	KASSERT(uq->uq_thread == td, ("uq_thread != td"));
4962 	KASSERT(uq->uq_pi_blocked == NULL, ("uq_pi_blocked != NULL"));
4963 	KASSERT(TAILQ_EMPTY(&uq->uq_pi_contested), ("uq_pi_contested is not empty"));
4964 }
4965 
4966 /*
4967  * exec() hook.
4968  *
4969  * Clear robust lists for all process' threads, not delaying the
4970  * cleanup to thread exit, since the relevant address space is
4971  * destroyed right now.
4972  */
4973 void
4974 umtx_exec(struct proc *p)
4975 {
4976 	struct thread *td;
4977 
4978 	KASSERT(p == curproc, ("need curproc"));
4979 	KASSERT((p->p_flag & P_HADTHREADS) == 0 ||
4980 	    (p->p_flag & P_STOPPED_SINGLE) != 0,
4981 	    ("curproc must be single-threaded"));
4982 	/*
4983 	 * There is no need to lock the list as only this thread can be
4984 	 * running.
4985 	 */
4986 	FOREACH_THREAD_IN_PROC(p, td) {
4987 		KASSERT(td == curthread ||
4988 		    ((td->td_flags & TDF_BOUNDARY) != 0 && TD_IS_SUSPENDED(td)),
4989 		    ("running thread %p %p", p, td));
4990 		umtx_thread_cleanup(td);
4991 		td->td_rb_list = td->td_rbp_list = td->td_rb_inact = 0;
4992 	}
4993 }
4994 
4995 /*
4996  * thread exit hook.
4997  */
4998 void
4999 umtx_thread_exit(struct thread *td)
5000 {
5001 
5002 	umtx_thread_cleanup(td);
5003 }
5004 
5005 static int
5006 umtx_read_uptr(struct thread *td, uintptr_t ptr, uintptr_t *res, bool compat32)
5007 {
5008 	u_long res1;
5009 	uint32_t res32;
5010 	int error;
5011 
5012 	if (compat32) {
5013 		error = fueword32((void *)ptr, &res32);
5014 		if (error == 0)
5015 			res1 = res32;
5016 	} else {
5017 		error = fueword((void *)ptr, &res1);
5018 	}
5019 	if (error == 0)
5020 		*res = res1;
5021 	else
5022 		error = EFAULT;
5023 	return (error);
5024 }
5025 
5026 static void
5027 umtx_read_rb_list(struct thread *td, struct umutex *m, uintptr_t *rb_list,
5028     bool compat32)
5029 {
5030 	struct umutex32 m32;
5031 
5032 	if (compat32) {
5033 		memcpy(&m32, m, sizeof(m32));
5034 		*rb_list = m32.m_rb_lnk;
5035 	} else {
5036 		*rb_list = m->m_rb_lnk;
5037 	}
5038 }
5039 
5040 static int
5041 umtx_handle_rb(struct thread *td, uintptr_t rbp, uintptr_t *rb_list, bool inact,
5042     bool compat32)
5043 {
5044 	struct umutex m;
5045 	int error;
5046 
5047 	KASSERT(td->td_proc == curproc, ("need current vmspace"));
5048 	error = copyin((void *)rbp, &m, sizeof(m));
5049 	if (error != 0)
5050 		return (error);
5051 	if (rb_list != NULL)
5052 		umtx_read_rb_list(td, &m, rb_list, compat32);
5053 	if ((m.m_flags & UMUTEX_ROBUST) == 0)
5054 		return (EINVAL);
5055 	if ((m.m_owner & ~UMUTEX_CONTESTED) != td->td_tid)
5056 		/* inact is cleared after unlock, allow the inconsistency */
5057 		return (inact ? 0 : EINVAL);
5058 	return (do_unlock_umutex(td, (struct umutex *)rbp, true));
5059 }
5060 
5061 static void
5062 umtx_cleanup_rb_list(struct thread *td, uintptr_t rb_list, uintptr_t *rb_inact,
5063     const char *name, bool compat32)
5064 {
5065 	int error, i;
5066 	uintptr_t rbp;
5067 	bool inact;
5068 
5069 	if (rb_list == 0)
5070 		return;
5071 	error = umtx_read_uptr(td, rb_list, &rbp, compat32);
5072 	for (i = 0; error == 0 && rbp != 0 && i < umtx_max_rb; i++) {
5073 		if (rbp == *rb_inact) {
5074 			inact = true;
5075 			*rb_inact = 0;
5076 		} else
5077 			inact = false;
5078 		error = umtx_handle_rb(td, rbp, &rbp, inact, compat32);
5079 	}
5080 	if (i == umtx_max_rb && umtx_verbose_rb) {
5081 		uprintf("comm %s pid %d: reached umtx %smax rb %d\n",
5082 		    td->td_proc->p_comm, td->td_proc->p_pid, name, umtx_max_rb);
5083 	}
5084 	if (error != 0 && umtx_verbose_rb) {
5085 		uprintf("comm %s pid %d: handling %srb error %d\n",
5086 		    td->td_proc->p_comm, td->td_proc->p_pid, name, error);
5087 	}
5088 }
5089 
5090 /*
5091  * Clean up umtx data.
5092  */
5093 static void
5094 umtx_thread_cleanup(struct thread *td)
5095 {
5096 	struct umtx_q *uq;
5097 	struct umtx_pi *pi;
5098 	uintptr_t rb_inact;
5099 	bool compat32;
5100 
5101 	/*
5102 	 * Disown pi mutexes.
5103 	 */
5104 	uq = td->td_umtxq;
5105 	if (uq != NULL) {
5106 		if (uq->uq_inherited_pri != PRI_MAX ||
5107 		    !TAILQ_EMPTY(&uq->uq_pi_contested)) {
5108 			mtx_lock(&umtx_lock);
5109 			uq->uq_inherited_pri = PRI_MAX;
5110 			while ((pi = TAILQ_FIRST(&uq->uq_pi_contested)) != NULL) {
5111 				pi->pi_owner = NULL;
5112 				TAILQ_REMOVE(&uq->uq_pi_contested, pi, pi_link);
5113 			}
5114 			mtx_unlock(&umtx_lock);
5115 		}
5116 		sched_lend_user_prio_cond(td, PRI_MAX);
5117 	}
5118 
5119 	compat32 = (td->td_pflags2 & TDP2_COMPAT32RB) != 0;
5120 	td->td_pflags2 &= ~TDP2_COMPAT32RB;
5121 
5122 	if (td->td_rb_inact == 0 && td->td_rb_list == 0 && td->td_rbp_list == 0)
5123 		return;
5124 
5125 	/*
5126 	 * Handle terminated robust mutexes.  Must be done after
5127 	 * robust pi disown, otherwise unlock could see unowned
5128 	 * entries.
5129 	 */
5130 	rb_inact = td->td_rb_inact;
5131 	if (rb_inact != 0)
5132 		(void)umtx_read_uptr(td, rb_inact, &rb_inact, compat32);
5133 	umtx_cleanup_rb_list(td, td->td_rb_list, &rb_inact, "", compat32);
5134 	umtx_cleanup_rb_list(td, td->td_rbp_list, &rb_inact, "priv ", compat32);
5135 	if (rb_inact != 0)
5136 		(void)umtx_handle_rb(td, rb_inact, NULL, true, compat32);
5137 }
5138