xref: /freebsd/sys/kern/kern_umtx.c (revision 6e563a1b608438504d963c2d7c70e50d2e75af46)
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/sysent.h>
59 #include <sys/systm.h>
60 #include <sys/sysproto.h>
61 #include <sys/syscallsubr.h>
62 #include <sys/taskqueue.h>
63 #include <sys/time.h>
64 #include <sys/eventhandler.h>
65 #include <sys/umtx.h>
66 #include <sys/umtxvar.h>
67 
68 #include <security/mac/mac_framework.h>
69 
70 #include <vm/vm.h>
71 #include <vm/vm_param.h>
72 #include <vm/pmap.h>
73 #include <vm/vm_map.h>
74 #include <vm/vm_object.h>
75 
76 #include <machine/atomic.h>
77 #include <machine/cpu.h>
78 
79 #include <compat/freebsd32/freebsd32.h>
80 #ifdef COMPAT_FREEBSD32
81 #include <compat/freebsd32/freebsd32_proto.h>
82 #endif
83 
84 #define _UMUTEX_TRY		1
85 #define _UMUTEX_WAIT		2
86 
87 #ifdef UMTX_PROFILING
88 #define	UPROF_PERC_BIGGER(w, f, sw, sf)					\
89 	(((w) > (sw)) || ((w) == (sw) && (f) > (sf)))
90 #endif
91 
92 #define	UMTXQ_LOCKED_ASSERT(uc)		mtx_assert(&(uc)->uc_lock, MA_OWNED)
93 #ifdef INVARIANTS
94 #define	UMTXQ_ASSERT_LOCKED_BUSY(key) do {				\
95 	struct umtxq_chain *uc;						\
96 									\
97 	uc = umtxq_getchain(key);					\
98 	mtx_assert(&uc->uc_lock, MA_OWNED);				\
99 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));		\
100 } while (0)
101 #else
102 #define	UMTXQ_ASSERT_LOCKED_BUSY(key) do {} while (0)
103 #endif
104 
105 /*
106  * Don't propagate time-sharing priority, there is a security reason,
107  * a user can simply introduce PI-mutex, let thread A lock the mutex,
108  * and let another thread B block on the mutex, because B is
109  * sleeping, its priority will be boosted, this causes A's priority to
110  * be boosted via priority propagating too and will never be lowered even
111  * if it is using 100%CPU, this is unfair to other processes.
112  */
113 
114 #define UPRI(td)	(((td)->td_user_pri >= PRI_MIN_TIMESHARE &&\
115 			  (td)->td_user_pri <= PRI_MAX_TIMESHARE) ?\
116 			 PRI_MAX_TIMESHARE : (td)->td_user_pri)
117 
118 #define	GOLDEN_RATIO_PRIME	2654404609U
119 #ifndef	UMTX_CHAINS
120 #define	UMTX_CHAINS		512
121 #endif
122 #define	UMTX_SHIFTS		(__WORD_BIT - 9)
123 
124 #define	GET_SHARE(flags)	\
125     (((flags) & USYNC_PROCESS_SHARED) == 0 ? THREAD_SHARE : PROCESS_SHARE)
126 
127 #define BUSY_SPINS		200
128 
129 struct umtx_copyops {
130 	int	(*copyin_timeout)(const void *uaddr, struct timespec *tsp);
131 	int	(*copyin_umtx_time)(const void *uaddr, size_t size,
132 	    struct _umtx_time *tp);
133 	int	(*copyin_robust_lists)(const void *uaddr, size_t size,
134 	    struct umtx_robust_lists_params *rbp);
135 	int	(*copyout_timeout)(void *uaddr, size_t size,
136 	    struct timespec *tsp);
137 	const size_t	timespec_sz;
138 	const size_t	umtx_time_sz;
139 	const bool	compat32;
140 };
141 
142 _Static_assert(sizeof(struct umutex) == sizeof(struct umutex32), "umutex32");
143 _Static_assert(__offsetof(struct umutex, m_spare[0]) ==
144     __offsetof(struct umutex32, m_spare[0]), "m_spare32");
145 
146 int umtx_shm_vnobj_persistent = 0;
147 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_vnode_persistent, CTLFLAG_RWTUN,
148     &umtx_shm_vnobj_persistent, 0,
149     "False forces destruction of umtx attached to file, on last close");
150 static int umtx_max_rb = 1000;
151 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_max_robust, CTLFLAG_RWTUN,
152     &umtx_max_rb, 0,
153     "Maximum number of robust mutexes allowed for each thread");
154 
155 static uma_zone_t		umtx_pi_zone;
156 static struct umtxq_chain	umtxq_chains[2][UMTX_CHAINS];
157 static MALLOC_DEFINE(M_UMTX, "umtx", "UMTX queue memory");
158 static int			umtx_pi_allocated;
159 
160 static SYSCTL_NODE(_debug, OID_AUTO, umtx, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
161     "umtx debug");
162 SYSCTL_INT(_debug_umtx, OID_AUTO, umtx_pi_allocated, CTLFLAG_RD,
163     &umtx_pi_allocated, 0, "Allocated umtx_pi");
164 static int umtx_verbose_rb = 1;
165 SYSCTL_INT(_debug_umtx, OID_AUTO, robust_faults_verbose, CTLFLAG_RWTUN,
166     &umtx_verbose_rb, 0,
167     "");
168 
169 #ifdef UMTX_PROFILING
170 static long max_length;
171 SYSCTL_LONG(_debug_umtx, OID_AUTO, max_length, CTLFLAG_RD, &max_length, 0, "max_length");
172 static SYSCTL_NODE(_debug_umtx, OID_AUTO, chains, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
173     "umtx chain stats");
174 #endif
175 
176 static inline void umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
177     const struct _umtx_time *umtxtime);
178 
179 static void umtx_shm_init(void);
180 static void umtxq_sysinit(void *);
181 static void umtxq_hash(struct umtx_key *key);
182 static int do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags,
183     bool rb);
184 static void umtx_thread_cleanup(struct thread *td);
185 SYSINIT(umtx, SI_SUB_EVENTHANDLER+1, SI_ORDER_MIDDLE, umtxq_sysinit, NULL);
186 
187 #define umtxq_signal(key, nwake)	umtxq_signal_queue((key), (nwake), UMTX_SHARED_QUEUE)
188 
189 static struct mtx umtx_lock;
190 
191 #ifdef UMTX_PROFILING
192 static void
193 umtx_init_profiling(void)
194 {
195 	struct sysctl_oid *chain_oid;
196 	char chain_name[10];
197 	int i;
198 
199 	for (i = 0; i < UMTX_CHAINS; ++i) {
200 		snprintf(chain_name, sizeof(chain_name), "%d", i);
201 		chain_oid = SYSCTL_ADD_NODE(NULL,
202 		    SYSCTL_STATIC_CHILDREN(_debug_umtx_chains), OID_AUTO,
203 		    chain_name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
204 		    "umtx hash stats");
205 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
206 		    "max_length0", CTLFLAG_RD, &umtxq_chains[0][i].max_length, 0, NULL);
207 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
208 		    "max_length1", CTLFLAG_RD, &umtxq_chains[1][i].max_length, 0, NULL);
209 	}
210 }
211 
212 static int
213 sysctl_debug_umtx_chains_peaks(SYSCTL_HANDLER_ARGS)
214 {
215 	char buf[512];
216 	struct sbuf sb;
217 	struct umtxq_chain *uc;
218 	u_int fract, i, j, tot, whole;
219 	u_int sf0, sf1, sf2, sf3, sf4;
220 	u_int si0, si1, si2, si3, si4;
221 	u_int sw0, sw1, sw2, sw3, sw4;
222 
223 	sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
224 	for (i = 0; i < 2; i++) {
225 		tot = 0;
226 		for (j = 0; j < UMTX_CHAINS; ++j) {
227 			uc = &umtxq_chains[i][j];
228 			mtx_lock(&uc->uc_lock);
229 			tot += uc->max_length;
230 			mtx_unlock(&uc->uc_lock);
231 		}
232 		if (tot == 0)
233 			sbuf_printf(&sb, "%u) Empty ", i);
234 		else {
235 			sf0 = sf1 = sf2 = sf3 = sf4 = 0;
236 			si0 = si1 = si2 = si3 = si4 = 0;
237 			sw0 = sw1 = sw2 = sw3 = sw4 = 0;
238 			for (j = 0; j < UMTX_CHAINS; j++) {
239 				uc = &umtxq_chains[i][j];
240 				mtx_lock(&uc->uc_lock);
241 				whole = uc->max_length * 100;
242 				mtx_unlock(&uc->uc_lock);
243 				fract = (whole % tot) * 100;
244 				if (UPROF_PERC_BIGGER(whole, fract, sw0, sf0)) {
245 					sf0 = fract;
246 					si0 = j;
247 					sw0 = whole;
248 				} else if (UPROF_PERC_BIGGER(whole, fract, sw1,
249 				    sf1)) {
250 					sf1 = fract;
251 					si1 = j;
252 					sw1 = whole;
253 				} else if (UPROF_PERC_BIGGER(whole, fract, sw2,
254 				    sf2)) {
255 					sf2 = fract;
256 					si2 = j;
257 					sw2 = whole;
258 				} else if (UPROF_PERC_BIGGER(whole, fract, sw3,
259 				    sf3)) {
260 					sf3 = fract;
261 					si3 = j;
262 					sw3 = whole;
263 				} else if (UPROF_PERC_BIGGER(whole, fract, sw4,
264 				    sf4)) {
265 					sf4 = fract;
266 					si4 = j;
267 					sw4 = whole;
268 				}
269 			}
270 			sbuf_printf(&sb, "queue %u:\n", i);
271 			sbuf_printf(&sb, "1st: %u.%u%% idx: %u\n", sw0 / tot,
272 			    sf0 / tot, si0);
273 			sbuf_printf(&sb, "2nd: %u.%u%% idx: %u\n", sw1 / tot,
274 			    sf1 / tot, si1);
275 			sbuf_printf(&sb, "3rd: %u.%u%% idx: %u\n", sw2 / tot,
276 			    sf2 / tot, si2);
277 			sbuf_printf(&sb, "4th: %u.%u%% idx: %u\n", sw3 / tot,
278 			    sf3 / tot, si3);
279 			sbuf_printf(&sb, "5th: %u.%u%% idx: %u\n", sw4 / tot,
280 			    sf4 / tot, si4);
281 		}
282 	}
283 	sbuf_trim(&sb);
284 	sbuf_finish(&sb);
285 	sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
286 	sbuf_delete(&sb);
287 	return (0);
288 }
289 
290 static int
291 sysctl_debug_umtx_chains_clear(SYSCTL_HANDLER_ARGS)
292 {
293 	struct umtxq_chain *uc;
294 	u_int i, j;
295 	int clear, error;
296 
297 	clear = 0;
298 	error = sysctl_handle_int(oidp, &clear, 0, req);
299 	if (error != 0 || req->newptr == NULL)
300 		return (error);
301 
302 	if (clear != 0) {
303 		for (i = 0; i < 2; ++i) {
304 			for (j = 0; j < UMTX_CHAINS; ++j) {
305 				uc = &umtxq_chains[i][j];
306 				mtx_lock(&uc->uc_lock);
307 				uc->length = 0;
308 				uc->max_length = 0;
309 				mtx_unlock(&uc->uc_lock);
310 			}
311 		}
312 	}
313 	return (0);
314 }
315 
316 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, clear,
317     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
318     sysctl_debug_umtx_chains_clear, "I",
319     "Clear umtx chains statistics");
320 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, peaks,
321     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
322     sysctl_debug_umtx_chains_peaks, "A",
323     "Highest peaks in chains max length");
324 #endif
325 
326 static void
327 umtxq_sysinit(void *arg __unused)
328 {
329 	int i, j;
330 
331 	umtx_pi_zone = uma_zcreate("umtx pi", sizeof(struct umtx_pi),
332 		NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
333 	for (i = 0; i < 2; ++i) {
334 		for (j = 0; j < UMTX_CHAINS; ++j) {
335 			mtx_init(&umtxq_chains[i][j].uc_lock, "umtxql", NULL,
336 				 MTX_DEF | MTX_DUPOK);
337 			LIST_INIT(&umtxq_chains[i][j].uc_queue[0]);
338 			LIST_INIT(&umtxq_chains[i][j].uc_queue[1]);
339 			LIST_INIT(&umtxq_chains[i][j].uc_spare_queue);
340 			TAILQ_INIT(&umtxq_chains[i][j].uc_pi_list);
341 			umtxq_chains[i][j].uc_busy = 0;
342 			umtxq_chains[i][j].uc_waiters = 0;
343 #ifdef UMTX_PROFILING
344 			umtxq_chains[i][j].length = 0;
345 			umtxq_chains[i][j].max_length = 0;
346 #endif
347 		}
348 	}
349 #ifdef UMTX_PROFILING
350 	umtx_init_profiling();
351 #endif
352 	mtx_init(&umtx_lock, "umtx lock", NULL, MTX_DEF);
353 	umtx_shm_init();
354 }
355 
356 struct umtx_q *
357 umtxq_alloc(void)
358 {
359 	struct umtx_q *uq;
360 
361 	uq = malloc(sizeof(struct umtx_q), M_UMTX, M_WAITOK | M_ZERO);
362 	uq->uq_spare_queue = malloc(sizeof(struct umtxq_queue), M_UMTX,
363 	    M_WAITOK | M_ZERO);
364 	TAILQ_INIT(&uq->uq_spare_queue->head);
365 	TAILQ_INIT(&uq->uq_pi_contested);
366 	uq->uq_inherited_pri = PRI_MAX;
367 	return (uq);
368 }
369 
370 void
371 umtxq_free(struct umtx_q *uq)
372 {
373 
374 	MPASS(uq->uq_spare_queue != NULL);
375 	free(uq->uq_spare_queue, M_UMTX);
376 	free(uq, M_UMTX);
377 }
378 
379 static inline void
380 umtxq_hash(struct umtx_key *key)
381 {
382 	unsigned n;
383 
384 	n = (uintptr_t)key->info.both.a + key->info.both.b;
385 	key->hash = ((n * GOLDEN_RATIO_PRIME) >> UMTX_SHIFTS) % UMTX_CHAINS;
386 }
387 
388 struct umtxq_chain *
389 umtxq_getchain(struct umtx_key *key)
390 {
391 
392 	if (key->type <= TYPE_SEM)
393 		return (&umtxq_chains[1][key->hash]);
394 	return (&umtxq_chains[0][key->hash]);
395 }
396 
397 /*
398  * Set chain to busy state when following operation
399  * may be blocked (kernel mutex can not be used).
400  */
401 void
402 umtxq_busy(struct umtx_key *key)
403 {
404 	struct umtxq_chain *uc;
405 
406 	uc = umtxq_getchain(key);
407 	mtx_assert(&uc->uc_lock, MA_OWNED);
408 	if (uc->uc_busy) {
409 #ifdef SMP
410 		if (smp_cpus > 1) {
411 			int count = BUSY_SPINS;
412 			if (count > 0) {
413 				umtxq_unlock(key);
414 				while (uc->uc_busy && --count > 0)
415 					cpu_spinwait();
416 				umtxq_lock(key);
417 			}
418 		}
419 #endif
420 		while (uc->uc_busy) {
421 			uc->uc_waiters++;
422 			msleep(uc, &uc->uc_lock, 0, "umtxqb", 0);
423 			uc->uc_waiters--;
424 		}
425 	}
426 	uc->uc_busy = 1;
427 }
428 
429 /*
430  * Unbusy a chain.
431  */
432 void
433 umtxq_unbusy(struct umtx_key *key)
434 {
435 	struct umtxq_chain *uc;
436 
437 	uc = umtxq_getchain(key);
438 	mtx_assert(&uc->uc_lock, MA_OWNED);
439 	KASSERT(uc->uc_busy != 0, ("not busy"));
440 	uc->uc_busy = 0;
441 	if (uc->uc_waiters)
442 		wakeup_one(uc);
443 }
444 
445 void
446 umtxq_unbusy_unlocked(struct umtx_key *key)
447 {
448 
449 	umtxq_lock(key);
450 	umtxq_unbusy(key);
451 	umtxq_unlock(key);
452 }
453 
454 static struct umtxq_queue *
455 umtxq_queue_lookup(struct umtx_key *key, int q)
456 {
457 	struct umtxq_queue *uh;
458 	struct umtxq_chain *uc;
459 
460 	uc = umtxq_getchain(key);
461 	UMTXQ_LOCKED_ASSERT(uc);
462 	LIST_FOREACH(uh, &uc->uc_queue[q], link) {
463 		if (umtx_key_match(&uh->key, key))
464 			return (uh);
465 	}
466 
467 	return (NULL);
468 }
469 
470 void
471 umtxq_insert_queue(struct umtx_q *uq, int q)
472 {
473 	struct umtxq_queue *uh;
474 	struct umtxq_chain *uc;
475 
476 	uc = umtxq_getchain(&uq->uq_key);
477 	UMTXQ_LOCKED_ASSERT(uc);
478 	KASSERT((uq->uq_flags & UQF_UMTXQ) == 0, ("umtx_q is already on queue"));
479 	uh = umtxq_queue_lookup(&uq->uq_key, q);
480 	if (uh != NULL) {
481 		LIST_INSERT_HEAD(&uc->uc_spare_queue, uq->uq_spare_queue, link);
482 	} else {
483 		uh = uq->uq_spare_queue;
484 		uh->key = uq->uq_key;
485 		LIST_INSERT_HEAD(&uc->uc_queue[q], uh, link);
486 #ifdef UMTX_PROFILING
487 		uc->length++;
488 		if (uc->length > uc->max_length) {
489 			uc->max_length = uc->length;
490 			if (uc->max_length > max_length)
491 				max_length = uc->max_length;
492 		}
493 #endif
494 	}
495 	uq->uq_spare_queue = NULL;
496 
497 	TAILQ_INSERT_TAIL(&uh->head, uq, uq_link);
498 	uh->length++;
499 	uq->uq_flags |= UQF_UMTXQ;
500 	uq->uq_cur_queue = uh;
501 	return;
502 }
503 
504 void
505 umtxq_remove_queue(struct umtx_q *uq, int q)
506 {
507 	struct umtxq_chain *uc;
508 	struct umtxq_queue *uh;
509 
510 	uc = umtxq_getchain(&uq->uq_key);
511 	UMTXQ_LOCKED_ASSERT(uc);
512 	if (uq->uq_flags & UQF_UMTXQ) {
513 		uh = uq->uq_cur_queue;
514 		TAILQ_REMOVE(&uh->head, uq, uq_link);
515 		uh->length--;
516 		uq->uq_flags &= ~UQF_UMTXQ;
517 		if (TAILQ_EMPTY(&uh->head)) {
518 			KASSERT(uh->length == 0,
519 			    ("inconsistent umtxq_queue length"));
520 #ifdef UMTX_PROFILING
521 			uc->length--;
522 #endif
523 			LIST_REMOVE(uh, link);
524 		} else {
525 			uh = LIST_FIRST(&uc->uc_spare_queue);
526 			KASSERT(uh != NULL, ("uc_spare_queue is empty"));
527 			LIST_REMOVE(uh, link);
528 		}
529 		uq->uq_spare_queue = uh;
530 		uq->uq_cur_queue = NULL;
531 	}
532 }
533 
534 /*
535  * Check if there are multiple waiters
536  */
537 int
538 umtxq_count(struct umtx_key *key)
539 {
540 	struct umtxq_queue *uh;
541 
542 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
543 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
544 	if (uh != NULL)
545 		return (uh->length);
546 	return (0);
547 }
548 
549 /*
550  * Check if there are multiple PI waiters and returns first
551  * waiter.
552  */
553 static int
554 umtxq_count_pi(struct umtx_key *key, struct umtx_q **first)
555 {
556 	struct umtxq_queue *uh;
557 
558 	*first = NULL;
559 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
560 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
561 	if (uh != NULL) {
562 		*first = TAILQ_FIRST(&uh->head);
563 		return (uh->length);
564 	}
565 	return (0);
566 }
567 
568 /*
569  * Wake up threads waiting on an userland object by a bit mask.
570  */
571 int
572 umtxq_signal_mask(struct umtx_key *key, int n_wake, u_int bitset)
573 {
574 	struct umtxq_queue *uh;
575 	struct umtx_q *uq, *uq_temp;
576 	int ret;
577 
578 	ret = 0;
579 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
580 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
581 	if (uh == NULL)
582 		return (0);
583 	TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
584 		if ((uq->uq_bitset & bitset) == 0)
585 			continue;
586 		umtxq_remove_queue(uq, UMTX_SHARED_QUEUE);
587 		wakeup_one(uq);
588 		if (++ret >= n_wake)
589 			break;
590 	}
591 	return (ret);
592 }
593 
594 /*
595  * Wake up threads waiting on an userland object.
596  */
597 
598 static int
599 umtxq_signal_queue(struct umtx_key *key, int n_wake, int q)
600 {
601 	struct umtxq_queue *uh;
602 	struct umtx_q *uq;
603 	int ret;
604 
605 	ret = 0;
606 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
607 	uh = umtxq_queue_lookup(key, q);
608 	if (uh != NULL) {
609 		while ((uq = TAILQ_FIRST(&uh->head)) != NULL) {
610 			umtxq_remove_queue(uq, q);
611 			wakeup(uq);
612 			if (++ret >= n_wake)
613 				return (ret);
614 		}
615 	}
616 	return (ret);
617 }
618 
619 /*
620  * Wake up specified thread.
621  */
622 static inline void
623 umtxq_signal_thread(struct umtx_q *uq)
624 {
625 
626 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
627 	umtxq_remove(uq);
628 	wakeup(uq);
629 }
630 
631 /*
632  * Wake up a maximum of n_wake threads that are waiting on an userland
633  * object identified by key. The remaining threads are removed from queue
634  * identified by key and added to the queue identified by key2 (requeued).
635  * The n_requeue specifies an upper limit on the number of threads that
636  * are requeued to the second queue.
637  */
638 int
639 umtxq_requeue(struct umtx_key *key, int n_wake, struct umtx_key *key2,
640     int n_requeue)
641 {
642 	struct umtxq_queue *uh;
643 	struct umtx_q *uq, *uq_temp;
644 	int ret;
645 
646 	ret = 0;
647 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
648 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key2));
649 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
650 	if (uh == NULL)
651 		return (0);
652 	TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
653 		if (++ret <= n_wake) {
654 			umtxq_remove(uq);
655 			wakeup_one(uq);
656 		} else {
657 			umtxq_remove(uq);
658 			uq->uq_key = *key2;
659 			umtxq_insert(uq);
660 			if (ret - n_wake == n_requeue)
661 				break;
662 		}
663 	}
664 	return (ret);
665 }
666 
667 static inline int
668 tstohz(const struct timespec *tsp)
669 {
670 	struct timeval tv;
671 
672 	TIMESPEC_TO_TIMEVAL(&tv, tsp);
673 	return tvtohz(&tv);
674 }
675 
676 void
677 umtx_abs_timeout_init(struct umtx_abs_timeout *timo, int clockid,
678     int absolute, const struct timespec *timeout)
679 {
680 
681 	timo->clockid = clockid;
682 	if (!absolute) {
683 		timo->is_abs_real = false;
684 		kern_clock_gettime(curthread, timo->clockid, &timo->cur);
685 		timespecadd(&timo->cur, timeout, &timo->end);
686 	} else {
687 		timo->end = *timeout;
688 		timo->is_abs_real = clockid == CLOCK_REALTIME ||
689 		    clockid == CLOCK_REALTIME_FAST ||
690 		    clockid == CLOCK_REALTIME_PRECISE ||
691 		    clockid == CLOCK_SECOND;
692 	}
693 }
694 
695 static void
696 umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
697     const struct _umtx_time *umtxtime)
698 {
699 
700 	umtx_abs_timeout_init(timo, umtxtime->_clockid,
701 	    (umtxtime->_flags & UMTX_ABSTIME) != 0, &umtxtime->_timeout);
702 }
703 
704 static int
705 umtx_abs_timeout_getsbt(struct umtx_abs_timeout *timo, sbintime_t *sbt,
706     int *flags)
707 {
708 	struct bintime bt, bbt;
709 	struct timespec tts;
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 		switch (timo->clockid) {
743 		case CLOCK_REALTIME_FAST:
744 		case CLOCK_MONOTONIC_FAST:
745 		case CLOCK_UPTIME_FAST:
746 			*sbt += tc_tick_sbt;
747 			break;
748 		case CLOCK_SECOND:
749 			*sbt += SBT_1S;
750 			break;
751 		}
752 		*flags = C_ABSOLUTE;
753 		return (0);
754 
755 	/* Clocks that has to be periodically polled. */
756 	case CLOCK_VIRTUAL:
757 	case CLOCK_PROF:
758 	case CLOCK_THREAD_CPUTIME_ID:
759 	case CLOCK_PROCESS_CPUTIME_ID:
760 	default:
761 		kern_clock_gettime(curthread, timo->clockid, &timo->cur);
762 		if (timespeccmp(&timo->end, &timo->cur, <=))
763 			return (ETIMEDOUT);
764 		timespecsub(&timo->end, &timo->cur, &tts);
765 		*sbt = tick_sbt * tstohz(&tts);
766 		*flags = C_HARDCLOCK;
767 		return (0);
768 	}
769 }
770 
771 static uint32_t
772 umtx_unlock_val(uint32_t flags, bool rb)
773 {
774 
775 	if (rb)
776 		return (UMUTEX_RB_OWNERDEAD);
777 	else if ((flags & UMUTEX_NONCONSISTENT) != 0)
778 		return (UMUTEX_RB_NOTRECOV);
779 	else
780 		return (UMUTEX_UNOWNED);
781 
782 }
783 
784 /*
785  * Put thread into sleep state, before sleeping, check if
786  * thread was removed from umtx queue.
787  */
788 int
789 umtxq_sleep(struct umtx_q *uq, const char *wmesg,
790     struct umtx_abs_timeout *timo)
791 {
792 	struct umtxq_chain *uc;
793 	sbintime_t sbt = 0;
794 	int error, flags = 0;
795 
796 	uc = umtxq_getchain(&uq->uq_key);
797 	UMTXQ_LOCKED_ASSERT(uc);
798 	for (;;) {
799 		if (!(uq->uq_flags & UQF_UMTXQ)) {
800 			error = 0;
801 			break;
802 		}
803 		if (timo != NULL) {
804 			if (timo->is_abs_real)
805 				curthread->td_rtcgen =
806 				    atomic_load_acq_int(&rtc_generation);
807 			error = umtx_abs_timeout_getsbt(timo, &sbt, &flags);
808 			if (error != 0)
809 				break;
810 		}
811 		error = msleep_sbt(uq, &uc->uc_lock, PCATCH | PDROP, wmesg,
812 		    sbt, 0, flags);
813 		uc = umtxq_getchain(&uq->uq_key);
814 		mtx_lock(&uc->uc_lock);
815 		if (error == EINTR || error == ERESTART)
816 			break;
817 		if (error == EWOULDBLOCK && (flags & C_ABSOLUTE) != 0) {
818 			error = ETIMEDOUT;
819 			break;
820 		}
821 	}
822 
823 	curthread->td_rtcgen = 0;
824 	return (error);
825 }
826 
827 /*
828  * Convert userspace address into unique logical address.
829  */
830 int
831 umtx_key_get(const void *addr, int type, int share, struct umtx_key *key)
832 {
833 	struct thread *td = curthread;
834 	vm_map_t map;
835 	vm_map_entry_t entry;
836 	vm_pindex_t pindex;
837 	vm_prot_t prot;
838 	boolean_t wired;
839 
840 	key->type = type;
841 	if (share == THREAD_SHARE) {
842 		key->shared = 0;
843 		key->info.private.vs = td->td_proc->p_vmspace;
844 		key->info.private.addr = (uintptr_t)addr;
845 	} else {
846 		MPASS(share == PROCESS_SHARE || share == AUTO_SHARE);
847 		map = &td->td_proc->p_vmspace->vm_map;
848 		if (vm_map_lookup(&map, (vm_offset_t)addr, VM_PROT_WRITE,
849 		    &entry, &key->info.shared.object, &pindex, &prot,
850 		    &wired) != KERN_SUCCESS) {
851 			return (EFAULT);
852 		}
853 
854 		if ((share == PROCESS_SHARE) ||
855 		    (share == AUTO_SHARE &&
856 		     VM_INHERIT_SHARE == entry->inheritance)) {
857 			key->shared = 1;
858 			key->info.shared.offset = (vm_offset_t)addr -
859 			    entry->start + entry->offset;
860 			vm_object_reference(key->info.shared.object);
861 		} else {
862 			key->shared = 0;
863 			key->info.private.vs = td->td_proc->p_vmspace;
864 			key->info.private.addr = (uintptr_t)addr;
865 		}
866 		vm_map_lookup_done(map, entry);
867 	}
868 
869 	umtxq_hash(key);
870 	return (0);
871 }
872 
873 /*
874  * Release key.
875  */
876 void
877 umtx_key_release(struct umtx_key *key)
878 {
879 	if (key->shared)
880 		vm_object_deallocate(key->info.shared.object);
881 }
882 
883 #ifdef COMPAT_FREEBSD10
884 /*
885  * Lock a umtx object.
886  */
887 static int
888 do_lock_umtx(struct thread *td, struct umtx *umtx, u_long id,
889     const struct timespec *timeout)
890 {
891 	struct umtx_abs_timeout timo;
892 	struct umtx_q *uq;
893 	u_long owner;
894 	u_long old;
895 	int error = 0;
896 
897 	uq = td->td_umtxq;
898 	if (timeout != NULL)
899 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
900 
901 	/*
902 	 * Care must be exercised when dealing with umtx structure. It
903 	 * can fault on any access.
904 	 */
905 	for (;;) {
906 		/*
907 		 * Try the uncontested case.  This should be done in userland.
908 		 */
909 		owner = casuword(&umtx->u_owner, UMTX_UNOWNED, id);
910 
911 		/* The acquire succeeded. */
912 		if (owner == UMTX_UNOWNED)
913 			return (0);
914 
915 		/* The address was invalid. */
916 		if (owner == -1)
917 			return (EFAULT);
918 
919 		/* If no one owns it but it is contested try to acquire it. */
920 		if (owner == UMTX_CONTESTED) {
921 			owner = casuword(&umtx->u_owner,
922 			    UMTX_CONTESTED, id | UMTX_CONTESTED);
923 
924 			if (owner == UMTX_CONTESTED)
925 				return (0);
926 
927 			/* The address was invalid. */
928 			if (owner == -1)
929 				return (EFAULT);
930 
931 			error = thread_check_susp(td, false);
932 			if (error != 0)
933 				break;
934 
935 			/* If this failed the lock has changed, restart. */
936 			continue;
937 		}
938 
939 		/*
940 		 * If we caught a signal, we have retried and now
941 		 * exit immediately.
942 		 */
943 		if (error != 0)
944 			break;
945 
946 		if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK,
947 			AUTO_SHARE, &uq->uq_key)) != 0)
948 			return (error);
949 
950 		umtxq_lock(&uq->uq_key);
951 		umtxq_busy(&uq->uq_key);
952 		umtxq_insert(uq);
953 		umtxq_unbusy(&uq->uq_key);
954 		umtxq_unlock(&uq->uq_key);
955 
956 		/*
957 		 * Set the contested bit so that a release in user space
958 		 * knows to use the system call for unlock.  If this fails
959 		 * either some one else has acquired the lock or it has been
960 		 * released.
961 		 */
962 		old = casuword(&umtx->u_owner, owner, owner | UMTX_CONTESTED);
963 
964 		/* The address was invalid. */
965 		if (old == -1) {
966 			umtxq_lock(&uq->uq_key);
967 			umtxq_remove(uq);
968 			umtxq_unlock(&uq->uq_key);
969 			umtx_key_release(&uq->uq_key);
970 			return (EFAULT);
971 		}
972 
973 		/*
974 		 * We set the contested bit, sleep. Otherwise the lock changed
975 		 * and we need to retry or we lost a race to the thread
976 		 * unlocking the umtx.
977 		 */
978 		umtxq_lock(&uq->uq_key);
979 		if (old == owner)
980 			error = umtxq_sleep(uq, "umtx", timeout == NULL ? NULL :
981 			    &timo);
982 		umtxq_remove(uq);
983 		umtxq_unlock(&uq->uq_key);
984 		umtx_key_release(&uq->uq_key);
985 
986 		if (error == 0)
987 			error = thread_check_susp(td, false);
988 	}
989 
990 	if (timeout == NULL) {
991 		/* Mutex locking is restarted if it is interrupted. */
992 		if (error == EINTR)
993 			error = ERESTART;
994 	} else {
995 		/* Timed-locking is not restarted. */
996 		if (error == ERESTART)
997 			error = EINTR;
998 	}
999 	return (error);
1000 }
1001 
1002 /*
1003  * Unlock a umtx object.
1004  */
1005 static int
1006 do_unlock_umtx(struct thread *td, struct umtx *umtx, u_long id)
1007 {
1008 	struct umtx_key key;
1009 	u_long owner;
1010 	u_long old;
1011 	int error;
1012 	int count;
1013 
1014 	/*
1015 	 * Make sure we own this mtx.
1016 	 */
1017 	owner = fuword(__DEVOLATILE(u_long *, &umtx->u_owner));
1018 	if (owner == -1)
1019 		return (EFAULT);
1020 
1021 	if ((owner & ~UMTX_CONTESTED) != id)
1022 		return (EPERM);
1023 
1024 	/* This should be done in userland */
1025 	if ((owner & UMTX_CONTESTED) == 0) {
1026 		old = casuword(&umtx->u_owner, owner, UMTX_UNOWNED);
1027 		if (old == -1)
1028 			return (EFAULT);
1029 		if (old == owner)
1030 			return (0);
1031 		owner = old;
1032 	}
1033 
1034 	/* We should only ever be in here for contested locks */
1035 	if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1036 	    &key)) != 0)
1037 		return (error);
1038 
1039 	umtxq_lock(&key);
1040 	umtxq_busy(&key);
1041 	count = umtxq_count(&key);
1042 	umtxq_unlock(&key);
1043 
1044 	/*
1045 	 * When unlocking the umtx, it must be marked as unowned if
1046 	 * there is zero or one thread only waiting for it.
1047 	 * Otherwise, it must be marked as contested.
1048 	 */
1049 	old = casuword(&umtx->u_owner, owner,
1050 	    count <= 1 ? UMTX_UNOWNED : UMTX_CONTESTED);
1051 	umtxq_lock(&key);
1052 	umtxq_signal(&key,1);
1053 	umtxq_unbusy(&key);
1054 	umtxq_unlock(&key);
1055 	umtx_key_release(&key);
1056 	if (old == -1)
1057 		return (EFAULT);
1058 	if (old != owner)
1059 		return (EINVAL);
1060 	return (0);
1061 }
1062 
1063 #ifdef COMPAT_FREEBSD32
1064 
1065 /*
1066  * Lock a umtx object.
1067  */
1068 static int
1069 do_lock_umtx32(struct thread *td, uint32_t *m, uint32_t id,
1070 	const struct timespec *timeout)
1071 {
1072 	struct umtx_abs_timeout timo;
1073 	struct umtx_q *uq;
1074 	uint32_t owner;
1075 	uint32_t old;
1076 	int error = 0;
1077 
1078 	uq = td->td_umtxq;
1079 
1080 	if (timeout != NULL)
1081 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
1082 
1083 	/*
1084 	 * Care must be exercised when dealing with umtx structure. It
1085 	 * can fault on any access.
1086 	 */
1087 	for (;;) {
1088 		/*
1089 		 * Try the uncontested case.  This should be done in userland.
1090 		 */
1091 		owner = casuword32(m, UMUTEX_UNOWNED, id);
1092 
1093 		/* The acquire succeeded. */
1094 		if (owner == UMUTEX_UNOWNED)
1095 			return (0);
1096 
1097 		/* The address was invalid. */
1098 		if (owner == -1)
1099 			return (EFAULT);
1100 
1101 		/* If no one owns it but it is contested try to acquire it. */
1102 		if (owner == UMUTEX_CONTESTED) {
1103 			owner = casuword32(m,
1104 			    UMUTEX_CONTESTED, id | UMUTEX_CONTESTED);
1105 			if (owner == UMUTEX_CONTESTED)
1106 				return (0);
1107 
1108 			/* The address was invalid. */
1109 			if (owner == -1)
1110 				return (EFAULT);
1111 
1112 			error = thread_check_susp(td, false);
1113 			if (error != 0)
1114 				break;
1115 
1116 			/* If this failed the lock has changed, restart. */
1117 			continue;
1118 		}
1119 
1120 		/*
1121 		 * If we caught a signal, we have retried and now
1122 		 * exit immediately.
1123 		 */
1124 		if (error != 0)
1125 			return (error);
1126 
1127 		if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK,
1128 			AUTO_SHARE, &uq->uq_key)) != 0)
1129 			return (error);
1130 
1131 		umtxq_lock(&uq->uq_key);
1132 		umtxq_busy(&uq->uq_key);
1133 		umtxq_insert(uq);
1134 		umtxq_unbusy(&uq->uq_key);
1135 		umtxq_unlock(&uq->uq_key);
1136 
1137 		/*
1138 		 * Set the contested bit so that a release in user space
1139 		 * knows to use the system call for unlock.  If this fails
1140 		 * either some one else has acquired the lock or it has been
1141 		 * released.
1142 		 */
1143 		old = casuword32(m, owner, owner | UMUTEX_CONTESTED);
1144 
1145 		/* The address was invalid. */
1146 		if (old == -1) {
1147 			umtxq_lock(&uq->uq_key);
1148 			umtxq_remove(uq);
1149 			umtxq_unlock(&uq->uq_key);
1150 			umtx_key_release(&uq->uq_key);
1151 			return (EFAULT);
1152 		}
1153 
1154 		/*
1155 		 * We set the contested bit, sleep. Otherwise the lock changed
1156 		 * and we need to retry or we lost a race to the thread
1157 		 * unlocking the umtx.
1158 		 */
1159 		umtxq_lock(&uq->uq_key);
1160 		if (old == owner)
1161 			error = umtxq_sleep(uq, "umtx", timeout == NULL ?
1162 			    NULL : &timo);
1163 		umtxq_remove(uq);
1164 		umtxq_unlock(&uq->uq_key);
1165 		umtx_key_release(&uq->uq_key);
1166 
1167 		if (error == 0)
1168 			error = thread_check_susp(td, false);
1169 	}
1170 
1171 	if (timeout == NULL) {
1172 		/* Mutex locking is restarted if it is interrupted. */
1173 		if (error == EINTR)
1174 			error = ERESTART;
1175 	} else {
1176 		/* Timed-locking is not restarted. */
1177 		if (error == ERESTART)
1178 			error = EINTR;
1179 	}
1180 	return (error);
1181 }
1182 
1183 /*
1184  * Unlock a umtx object.
1185  */
1186 static int
1187 do_unlock_umtx32(struct thread *td, uint32_t *m, uint32_t id)
1188 {
1189 	struct umtx_key key;
1190 	uint32_t owner;
1191 	uint32_t old;
1192 	int error;
1193 	int count;
1194 
1195 	/*
1196 	 * Make sure we own this mtx.
1197 	 */
1198 	owner = fuword32(m);
1199 	if (owner == -1)
1200 		return (EFAULT);
1201 
1202 	if ((owner & ~UMUTEX_CONTESTED) != id)
1203 		return (EPERM);
1204 
1205 	/* This should be done in userland */
1206 	if ((owner & UMUTEX_CONTESTED) == 0) {
1207 		old = casuword32(m, owner, UMUTEX_UNOWNED);
1208 		if (old == -1)
1209 			return (EFAULT);
1210 		if (old == owner)
1211 			return (0);
1212 		owner = old;
1213 	}
1214 
1215 	/* We should only ever be in here for contested locks */
1216 	if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1217 		&key)) != 0)
1218 		return (error);
1219 
1220 	umtxq_lock(&key);
1221 	umtxq_busy(&key);
1222 	count = umtxq_count(&key);
1223 	umtxq_unlock(&key);
1224 
1225 	/*
1226 	 * When unlocking the umtx, it must be marked as unowned if
1227 	 * there is zero or one thread only waiting for it.
1228 	 * Otherwise, it must be marked as contested.
1229 	 */
1230 	old = casuword32(m, owner,
1231 		count <= 1 ? UMUTEX_UNOWNED : UMUTEX_CONTESTED);
1232 	umtxq_lock(&key);
1233 	umtxq_signal(&key,1);
1234 	umtxq_unbusy(&key);
1235 	umtxq_unlock(&key);
1236 	umtx_key_release(&key);
1237 	if (old == -1)
1238 		return (EFAULT);
1239 	if (old != owner)
1240 		return (EINVAL);
1241 	return (0);
1242 }
1243 #endif	/* COMPAT_FREEBSD32 */
1244 #endif	/* COMPAT_FREEBSD10 */
1245 
1246 /*
1247  * Fetch and compare value, sleep on the address if value is not changed.
1248  */
1249 static int
1250 do_wait(struct thread *td, void *addr, u_long id,
1251     struct _umtx_time *timeout, int compat32, int is_private)
1252 {
1253 	struct umtx_abs_timeout timo;
1254 	struct umtx_q *uq;
1255 	u_long tmp;
1256 	uint32_t tmp32;
1257 	int error = 0;
1258 
1259 	uq = td->td_umtxq;
1260 	if ((error = umtx_key_get(addr, TYPE_SIMPLE_WAIT,
1261 		is_private ? THREAD_SHARE : AUTO_SHARE, &uq->uq_key)) != 0)
1262 		return (error);
1263 
1264 	if (timeout != NULL)
1265 		umtx_abs_timeout_init2(&timo, timeout);
1266 
1267 	umtxq_lock(&uq->uq_key);
1268 	umtxq_insert(uq);
1269 	umtxq_unlock(&uq->uq_key);
1270 	if (compat32 == 0) {
1271 		error = fueword(addr, &tmp);
1272 		if (error != 0)
1273 			error = EFAULT;
1274 	} else {
1275 		error = fueword32(addr, &tmp32);
1276 		if (error == 0)
1277 			tmp = tmp32;
1278 		else
1279 			error = EFAULT;
1280 	}
1281 	umtxq_lock(&uq->uq_key);
1282 	if (error == 0) {
1283 		if (tmp == id)
1284 			error = umtxq_sleep(uq, "uwait", timeout == NULL ?
1285 			    NULL : &timo);
1286 		if ((uq->uq_flags & UQF_UMTXQ) == 0)
1287 			error = 0;
1288 		else
1289 			umtxq_remove(uq);
1290 	} else if ((uq->uq_flags & UQF_UMTXQ) != 0) {
1291 		umtxq_remove(uq);
1292 	}
1293 	umtxq_unlock(&uq->uq_key);
1294 	umtx_key_release(&uq->uq_key);
1295 	if (error == ERESTART)
1296 		error = EINTR;
1297 	return (error);
1298 }
1299 
1300 /*
1301  * Wake up threads sleeping on the specified address.
1302  */
1303 int
1304 kern_umtx_wake(struct thread *td, void *uaddr, int n_wake, int is_private)
1305 {
1306 	struct umtx_key key;
1307 	int ret;
1308 
1309 	if ((ret = umtx_key_get(uaddr, TYPE_SIMPLE_WAIT,
1310 	    is_private ? THREAD_SHARE : AUTO_SHARE, &key)) != 0)
1311 		return (ret);
1312 	umtxq_lock(&key);
1313 	umtxq_signal(&key, n_wake);
1314 	umtxq_unlock(&key);
1315 	umtx_key_release(&key);
1316 	return (0);
1317 }
1318 
1319 /*
1320  * Lock PTHREAD_PRIO_NONE protocol POSIX mutex.
1321  */
1322 static int
1323 do_lock_normal(struct thread *td, struct umutex *m, uint32_t flags,
1324     struct _umtx_time *timeout, int mode)
1325 {
1326 	struct umtx_abs_timeout timo;
1327 	struct umtx_q *uq;
1328 	uint32_t owner, old, id;
1329 	int error, rv;
1330 
1331 	id = td->td_tid;
1332 	uq = td->td_umtxq;
1333 	error = 0;
1334 	if (timeout != NULL)
1335 		umtx_abs_timeout_init2(&timo, timeout);
1336 
1337 	/*
1338 	 * Care must be exercised when dealing with umtx structure. It
1339 	 * can fault on any access.
1340 	 */
1341 	for (;;) {
1342 		rv = fueword32(&m->m_owner, &owner);
1343 		if (rv == -1)
1344 			return (EFAULT);
1345 		if (mode == _UMUTEX_WAIT) {
1346 			if (owner == UMUTEX_UNOWNED ||
1347 			    owner == UMUTEX_CONTESTED ||
1348 			    owner == UMUTEX_RB_OWNERDEAD ||
1349 			    owner == UMUTEX_RB_NOTRECOV)
1350 				return (0);
1351 		} else {
1352 			/*
1353 			 * Robust mutex terminated.  Kernel duty is to
1354 			 * return EOWNERDEAD to the userspace.  The
1355 			 * umutex.m_flags UMUTEX_NONCONSISTENT is set
1356 			 * by the common userspace code.
1357 			 */
1358 			if (owner == UMUTEX_RB_OWNERDEAD) {
1359 				rv = casueword32(&m->m_owner,
1360 				    UMUTEX_RB_OWNERDEAD, &owner,
1361 				    id | UMUTEX_CONTESTED);
1362 				if (rv == -1)
1363 					return (EFAULT);
1364 				if (rv == 0) {
1365 					MPASS(owner == UMUTEX_RB_OWNERDEAD);
1366 					return (EOWNERDEAD); /* success */
1367 				}
1368 				MPASS(rv == 1);
1369 				rv = thread_check_susp(td, false);
1370 				if (rv != 0)
1371 					return (rv);
1372 				continue;
1373 			}
1374 			if (owner == UMUTEX_RB_NOTRECOV)
1375 				return (ENOTRECOVERABLE);
1376 
1377 			/*
1378 			 * Try the uncontested case.  This should be
1379 			 * done in userland.
1380 			 */
1381 			rv = casueword32(&m->m_owner, UMUTEX_UNOWNED,
1382 			    &owner, id);
1383 			/* The address was invalid. */
1384 			if (rv == -1)
1385 				return (EFAULT);
1386 
1387 			/* The acquire succeeded. */
1388 			if (rv == 0) {
1389 				MPASS(owner == UMUTEX_UNOWNED);
1390 				return (0);
1391 			}
1392 
1393 			/*
1394 			 * If no one owns it but it is contested try
1395 			 * to acquire it.
1396 			 */
1397 			MPASS(rv == 1);
1398 			if (owner == UMUTEX_CONTESTED) {
1399 				rv = casueword32(&m->m_owner,
1400 				    UMUTEX_CONTESTED, &owner,
1401 				    id | UMUTEX_CONTESTED);
1402 				/* The address was invalid. */
1403 				if (rv == -1)
1404 					return (EFAULT);
1405 				if (rv == 0) {
1406 					MPASS(owner == UMUTEX_CONTESTED);
1407 					return (0);
1408 				}
1409 				if (rv == 1) {
1410 					rv = thread_check_susp(td, false);
1411 					if (rv != 0)
1412 						return (rv);
1413 				}
1414 
1415 				/*
1416 				 * If this failed the lock has
1417 				 * changed, restart.
1418 				 */
1419 				continue;
1420 			}
1421 
1422 			/* rv == 1 but not contested, likely store failure */
1423 			rv = thread_check_susp(td, false);
1424 			if (rv != 0)
1425 				return (rv);
1426 		}
1427 
1428 		if (mode == _UMUTEX_TRY)
1429 			return (EBUSY);
1430 
1431 		/*
1432 		 * If we caught a signal, we have retried and now
1433 		 * exit immediately.
1434 		 */
1435 		if (error != 0)
1436 			return (error);
1437 
1438 		if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX,
1439 		    GET_SHARE(flags), &uq->uq_key)) != 0)
1440 			return (error);
1441 
1442 		umtxq_lock(&uq->uq_key);
1443 		umtxq_busy(&uq->uq_key);
1444 		umtxq_insert(uq);
1445 		umtxq_unlock(&uq->uq_key);
1446 
1447 		/*
1448 		 * Set the contested bit so that a release in user space
1449 		 * knows to use the system call for unlock.  If this fails
1450 		 * either some one else has acquired the lock or it has been
1451 		 * released.
1452 		 */
1453 		rv = casueword32(&m->m_owner, owner, &old,
1454 		    owner | UMUTEX_CONTESTED);
1455 
1456 		/* The address was invalid or casueword failed to store. */
1457 		if (rv == -1 || rv == 1) {
1458 			umtxq_lock(&uq->uq_key);
1459 			umtxq_remove(uq);
1460 			umtxq_unbusy(&uq->uq_key);
1461 			umtxq_unlock(&uq->uq_key);
1462 			umtx_key_release(&uq->uq_key);
1463 			if (rv == -1)
1464 				return (EFAULT);
1465 			if (rv == 1) {
1466 				rv = thread_check_susp(td, false);
1467 				if (rv != 0)
1468 					return (rv);
1469 			}
1470 			continue;
1471 		}
1472 
1473 		/*
1474 		 * We set the contested bit, sleep. Otherwise the lock changed
1475 		 * and we need to retry or we lost a race to the thread
1476 		 * unlocking the umtx.
1477 		 */
1478 		umtxq_lock(&uq->uq_key);
1479 		umtxq_unbusy(&uq->uq_key);
1480 		MPASS(old == owner);
1481 		error = umtxq_sleep(uq, "umtxn", timeout == NULL ?
1482 		    NULL : &timo);
1483 		umtxq_remove(uq);
1484 		umtxq_unlock(&uq->uq_key);
1485 		umtx_key_release(&uq->uq_key);
1486 
1487 		if (error == 0)
1488 			error = thread_check_susp(td, false);
1489 	}
1490 
1491 	return (0);
1492 }
1493 
1494 /*
1495  * Unlock PTHREAD_PRIO_NONE protocol POSIX mutex.
1496  */
1497 static int
1498 do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1499 {
1500 	struct umtx_key key;
1501 	uint32_t owner, old, id, newlock;
1502 	int error, count;
1503 
1504 	id = td->td_tid;
1505 
1506 again:
1507 	/*
1508 	 * Make sure we own this mtx.
1509 	 */
1510 	error = fueword32(&m->m_owner, &owner);
1511 	if (error == -1)
1512 		return (EFAULT);
1513 
1514 	if ((owner & ~UMUTEX_CONTESTED) != id)
1515 		return (EPERM);
1516 
1517 	newlock = umtx_unlock_val(flags, rb);
1518 	if ((owner & UMUTEX_CONTESTED) == 0) {
1519 		error = casueword32(&m->m_owner, owner, &old, newlock);
1520 		if (error == -1)
1521 			return (EFAULT);
1522 		if (error == 1) {
1523 			error = thread_check_susp(td, false);
1524 			if (error != 0)
1525 				return (error);
1526 			goto again;
1527 		}
1528 		MPASS(old == owner);
1529 		return (0);
1530 	}
1531 
1532 	/* We should only ever be in here for contested locks */
1533 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1534 	    &key)) != 0)
1535 		return (error);
1536 
1537 	umtxq_lock(&key);
1538 	umtxq_busy(&key);
1539 	count = umtxq_count(&key);
1540 	umtxq_unlock(&key);
1541 
1542 	/*
1543 	 * When unlocking the umtx, it must be marked as unowned if
1544 	 * there is zero or one thread only waiting for it.
1545 	 * Otherwise, it must be marked as contested.
1546 	 */
1547 	if (count > 1)
1548 		newlock |= UMUTEX_CONTESTED;
1549 	error = casueword32(&m->m_owner, owner, &old, newlock);
1550 	umtxq_lock(&key);
1551 	umtxq_signal(&key, 1);
1552 	umtxq_unbusy(&key);
1553 	umtxq_unlock(&key);
1554 	umtx_key_release(&key);
1555 	if (error == -1)
1556 		return (EFAULT);
1557 	if (error == 1) {
1558 		if (old != owner)
1559 			return (EINVAL);
1560 		error = thread_check_susp(td, false);
1561 		if (error != 0)
1562 			return (error);
1563 		goto again;
1564 	}
1565 	return (0);
1566 }
1567 
1568 /*
1569  * Check if the mutex is available and wake up a waiter,
1570  * only for simple mutex.
1571  */
1572 static int
1573 do_wake_umutex(struct thread *td, struct umutex *m)
1574 {
1575 	struct umtx_key key;
1576 	uint32_t owner;
1577 	uint32_t flags;
1578 	int error;
1579 	int count;
1580 
1581 again:
1582 	error = fueword32(&m->m_owner, &owner);
1583 	if (error == -1)
1584 		return (EFAULT);
1585 
1586 	if ((owner & ~UMUTEX_CONTESTED) != 0 && owner != UMUTEX_RB_OWNERDEAD &&
1587 	    owner != UMUTEX_RB_NOTRECOV)
1588 		return (0);
1589 
1590 	error = fueword32(&m->m_flags, &flags);
1591 	if (error == -1)
1592 		return (EFAULT);
1593 
1594 	/* We should only ever be in here for contested locks */
1595 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1596 	    &key)) != 0)
1597 		return (error);
1598 
1599 	umtxq_lock(&key);
1600 	umtxq_busy(&key);
1601 	count = umtxq_count(&key);
1602 	umtxq_unlock(&key);
1603 
1604 	if (count <= 1 && owner != UMUTEX_RB_OWNERDEAD &&
1605 	    owner != UMUTEX_RB_NOTRECOV) {
1606 		error = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
1607 		    UMUTEX_UNOWNED);
1608 		if (error == -1) {
1609 			error = EFAULT;
1610 		} else if (error == 1) {
1611 			umtxq_lock(&key);
1612 			umtxq_unbusy(&key);
1613 			umtxq_unlock(&key);
1614 			umtx_key_release(&key);
1615 			error = thread_check_susp(td, false);
1616 			if (error != 0)
1617 				return (error);
1618 			goto again;
1619 		}
1620 	}
1621 
1622 	umtxq_lock(&key);
1623 	if (error == 0 && count != 0) {
1624 		MPASS((owner & ~UMUTEX_CONTESTED) == 0 ||
1625 		    owner == UMUTEX_RB_OWNERDEAD ||
1626 		    owner == UMUTEX_RB_NOTRECOV);
1627 		umtxq_signal(&key, 1);
1628 	}
1629 	umtxq_unbusy(&key);
1630 	umtxq_unlock(&key);
1631 	umtx_key_release(&key);
1632 	return (error);
1633 }
1634 
1635 /*
1636  * Check if the mutex has waiters and tries to fix contention bit.
1637  */
1638 static int
1639 do_wake2_umutex(struct thread *td, struct umutex *m, uint32_t flags)
1640 {
1641 	struct umtx_key key;
1642 	uint32_t owner, old;
1643 	int type;
1644 	int error;
1645 	int count;
1646 
1647 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT |
1648 	    UMUTEX_ROBUST)) {
1649 	case 0:
1650 	case UMUTEX_ROBUST:
1651 		type = TYPE_NORMAL_UMUTEX;
1652 		break;
1653 	case UMUTEX_PRIO_INHERIT:
1654 		type = TYPE_PI_UMUTEX;
1655 		break;
1656 	case (UMUTEX_PRIO_INHERIT | UMUTEX_ROBUST):
1657 		type = TYPE_PI_ROBUST_UMUTEX;
1658 		break;
1659 	case UMUTEX_PRIO_PROTECT:
1660 		type = TYPE_PP_UMUTEX;
1661 		break;
1662 	case (UMUTEX_PRIO_PROTECT | UMUTEX_ROBUST):
1663 		type = TYPE_PP_ROBUST_UMUTEX;
1664 		break;
1665 	default:
1666 		return (EINVAL);
1667 	}
1668 	if ((error = umtx_key_get(m, type, GET_SHARE(flags), &key)) != 0)
1669 		return (error);
1670 
1671 	owner = 0;
1672 	umtxq_lock(&key);
1673 	umtxq_busy(&key);
1674 	count = umtxq_count(&key);
1675 	umtxq_unlock(&key);
1676 
1677 	error = fueword32(&m->m_owner, &owner);
1678 	if (error == -1)
1679 		error = EFAULT;
1680 
1681 	/*
1682 	 * Only repair contention bit if there is a waiter, this means
1683 	 * the mutex is still being referenced by userland code,
1684 	 * otherwise don't update any memory.
1685 	 */
1686 	while (error == 0 && (owner & UMUTEX_CONTESTED) == 0 &&
1687 	    (count > 1 || (count == 1 && (owner & ~UMUTEX_CONTESTED) != 0))) {
1688 		error = casueword32(&m->m_owner, owner, &old,
1689 		    owner | UMUTEX_CONTESTED);
1690 		if (error == -1) {
1691 			error = EFAULT;
1692 			break;
1693 		}
1694 		if (error == 0) {
1695 			MPASS(old == owner);
1696 			break;
1697 		}
1698 		owner = old;
1699 		error = thread_check_susp(td, false);
1700 	}
1701 
1702 	umtxq_lock(&key);
1703 	if (error == EFAULT) {
1704 		umtxq_signal(&key, INT_MAX);
1705 	} else if (count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1706 	    owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1707 		umtxq_signal(&key, 1);
1708 	umtxq_unbusy(&key);
1709 	umtxq_unlock(&key);
1710 	umtx_key_release(&key);
1711 	return (error);
1712 }
1713 
1714 struct umtx_pi *
1715 umtx_pi_alloc(int flags)
1716 {
1717 	struct umtx_pi *pi;
1718 
1719 	pi = uma_zalloc(umtx_pi_zone, M_ZERO | flags);
1720 	TAILQ_INIT(&pi->pi_blocked);
1721 	atomic_add_int(&umtx_pi_allocated, 1);
1722 	return (pi);
1723 }
1724 
1725 void
1726 umtx_pi_free(struct umtx_pi *pi)
1727 {
1728 	uma_zfree(umtx_pi_zone, pi);
1729 	atomic_add_int(&umtx_pi_allocated, -1);
1730 }
1731 
1732 /*
1733  * Adjust the thread's position on a pi_state after its priority has been
1734  * changed.
1735  */
1736 static int
1737 umtx_pi_adjust_thread(struct umtx_pi *pi, struct thread *td)
1738 {
1739 	struct umtx_q *uq, *uq1, *uq2;
1740 	struct thread *td1;
1741 
1742 	mtx_assert(&umtx_lock, MA_OWNED);
1743 	if (pi == NULL)
1744 		return (0);
1745 
1746 	uq = td->td_umtxq;
1747 
1748 	/*
1749 	 * Check if the thread needs to be moved on the blocked chain.
1750 	 * It needs to be moved if either its priority is lower than
1751 	 * the previous thread or higher than the next thread.
1752 	 */
1753 	uq1 = TAILQ_PREV(uq, umtxq_head, uq_lockq);
1754 	uq2 = TAILQ_NEXT(uq, uq_lockq);
1755 	if ((uq1 != NULL && UPRI(td) < UPRI(uq1->uq_thread)) ||
1756 	    (uq2 != NULL && UPRI(td) > UPRI(uq2->uq_thread))) {
1757 		/*
1758 		 * Remove thread from blocked chain and determine where
1759 		 * it should be moved to.
1760 		 */
1761 		TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1762 		TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1763 			td1 = uq1->uq_thread;
1764 			MPASS(td1->td_proc->p_magic == P_MAGIC);
1765 			if (UPRI(td1) > UPRI(td))
1766 				break;
1767 		}
1768 
1769 		if (uq1 == NULL)
1770 			TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1771 		else
1772 			TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1773 	}
1774 	return (1);
1775 }
1776 
1777 static struct umtx_pi *
1778 umtx_pi_next(struct umtx_pi *pi)
1779 {
1780 	struct umtx_q *uq_owner;
1781 
1782 	if (pi->pi_owner == NULL)
1783 		return (NULL);
1784 	uq_owner = pi->pi_owner->td_umtxq;
1785 	if (uq_owner == NULL)
1786 		return (NULL);
1787 	return (uq_owner->uq_pi_blocked);
1788 }
1789 
1790 /*
1791  * Floyd's Cycle-Finding Algorithm.
1792  */
1793 static bool
1794 umtx_pi_check_loop(struct umtx_pi *pi)
1795 {
1796 	struct umtx_pi *pi1;	/* fast iterator */
1797 
1798 	mtx_assert(&umtx_lock, MA_OWNED);
1799 	if (pi == NULL)
1800 		return (false);
1801 	pi1 = pi;
1802 	for (;;) {
1803 		pi = umtx_pi_next(pi);
1804 		if (pi == NULL)
1805 			break;
1806 		pi1 = umtx_pi_next(pi1);
1807 		if (pi1 == NULL)
1808 			break;
1809 		pi1 = umtx_pi_next(pi1);
1810 		if (pi1 == NULL)
1811 			break;
1812 		if (pi == pi1)
1813 			return (true);
1814 	}
1815 	return (false);
1816 }
1817 
1818 /*
1819  * Propagate priority when a thread is blocked on POSIX
1820  * PI mutex.
1821  */
1822 static void
1823 umtx_propagate_priority(struct thread *td)
1824 {
1825 	struct umtx_q *uq;
1826 	struct umtx_pi *pi;
1827 	int pri;
1828 
1829 	mtx_assert(&umtx_lock, MA_OWNED);
1830 	pri = UPRI(td);
1831 	uq = td->td_umtxq;
1832 	pi = uq->uq_pi_blocked;
1833 	if (pi == NULL)
1834 		return;
1835 	if (umtx_pi_check_loop(pi))
1836 		return;
1837 
1838 	for (;;) {
1839 		td = pi->pi_owner;
1840 		if (td == NULL || td == curthread)
1841 			return;
1842 
1843 		MPASS(td->td_proc != NULL);
1844 		MPASS(td->td_proc->p_magic == P_MAGIC);
1845 
1846 		thread_lock(td);
1847 		if (td->td_lend_user_pri > pri)
1848 			sched_lend_user_prio(td, pri);
1849 		else {
1850 			thread_unlock(td);
1851 			break;
1852 		}
1853 		thread_unlock(td);
1854 
1855 		/*
1856 		 * Pick up the lock that td is blocked on.
1857 		 */
1858 		uq = td->td_umtxq;
1859 		pi = uq->uq_pi_blocked;
1860 		if (pi == NULL)
1861 			break;
1862 		/* Resort td on the list if needed. */
1863 		umtx_pi_adjust_thread(pi, td);
1864 	}
1865 }
1866 
1867 /*
1868  * Unpropagate priority for a PI mutex when a thread blocked on
1869  * it is interrupted by signal or resumed by others.
1870  */
1871 static void
1872 umtx_repropagate_priority(struct umtx_pi *pi)
1873 {
1874 	struct umtx_q *uq, *uq_owner;
1875 	struct umtx_pi *pi2;
1876 	int pri;
1877 
1878 	mtx_assert(&umtx_lock, MA_OWNED);
1879 
1880 	if (umtx_pi_check_loop(pi))
1881 		return;
1882 	while (pi != NULL && pi->pi_owner != NULL) {
1883 		pri = PRI_MAX;
1884 		uq_owner = pi->pi_owner->td_umtxq;
1885 
1886 		TAILQ_FOREACH(pi2, &uq_owner->uq_pi_contested, pi_link) {
1887 			uq = TAILQ_FIRST(&pi2->pi_blocked);
1888 			if (uq != NULL) {
1889 				if (pri > UPRI(uq->uq_thread))
1890 					pri = UPRI(uq->uq_thread);
1891 			}
1892 		}
1893 
1894 		if (pri > uq_owner->uq_inherited_pri)
1895 			pri = uq_owner->uq_inherited_pri;
1896 		thread_lock(pi->pi_owner);
1897 		sched_lend_user_prio(pi->pi_owner, pri);
1898 		thread_unlock(pi->pi_owner);
1899 		if ((pi = uq_owner->uq_pi_blocked) != NULL)
1900 			umtx_pi_adjust_thread(pi, uq_owner->uq_thread);
1901 	}
1902 }
1903 
1904 /*
1905  * Insert a PI mutex into owned list.
1906  */
1907 static void
1908 umtx_pi_setowner(struct umtx_pi *pi, struct thread *owner)
1909 {
1910 	struct umtx_q *uq_owner;
1911 
1912 	uq_owner = owner->td_umtxq;
1913 	mtx_assert(&umtx_lock, MA_OWNED);
1914 	MPASS(pi->pi_owner == NULL);
1915 	pi->pi_owner = owner;
1916 	TAILQ_INSERT_TAIL(&uq_owner->uq_pi_contested, pi, pi_link);
1917 }
1918 
1919 /*
1920  * Disown a PI mutex, and remove it from the owned list.
1921  */
1922 static void
1923 umtx_pi_disown(struct umtx_pi *pi)
1924 {
1925 
1926 	mtx_assert(&umtx_lock, MA_OWNED);
1927 	TAILQ_REMOVE(&pi->pi_owner->td_umtxq->uq_pi_contested, pi, pi_link);
1928 	pi->pi_owner = NULL;
1929 }
1930 
1931 /*
1932  * Claim ownership of a PI mutex.
1933  */
1934 int
1935 umtx_pi_claim(struct umtx_pi *pi, struct thread *owner)
1936 {
1937 	struct umtx_q *uq;
1938 	int pri;
1939 
1940 	mtx_lock(&umtx_lock);
1941 	if (pi->pi_owner == owner) {
1942 		mtx_unlock(&umtx_lock);
1943 		return (0);
1944 	}
1945 
1946 	if (pi->pi_owner != NULL) {
1947 		/*
1948 		 * userland may have already messed the mutex, sigh.
1949 		 */
1950 		mtx_unlock(&umtx_lock);
1951 		return (EPERM);
1952 	}
1953 	umtx_pi_setowner(pi, owner);
1954 	uq = TAILQ_FIRST(&pi->pi_blocked);
1955 	if (uq != NULL) {
1956 		pri = UPRI(uq->uq_thread);
1957 		thread_lock(owner);
1958 		if (pri < UPRI(owner))
1959 			sched_lend_user_prio(owner, pri);
1960 		thread_unlock(owner);
1961 	}
1962 	mtx_unlock(&umtx_lock);
1963 	return (0);
1964 }
1965 
1966 /*
1967  * Adjust a thread's order position in its blocked PI mutex,
1968  * this may result new priority propagating process.
1969  */
1970 void
1971 umtx_pi_adjust(struct thread *td, u_char oldpri)
1972 {
1973 	struct umtx_q *uq;
1974 	struct umtx_pi *pi;
1975 
1976 	uq = td->td_umtxq;
1977 	mtx_lock(&umtx_lock);
1978 	/*
1979 	 * Pick up the lock that td is blocked on.
1980 	 */
1981 	pi = uq->uq_pi_blocked;
1982 	if (pi != NULL) {
1983 		umtx_pi_adjust_thread(pi, td);
1984 		umtx_repropagate_priority(pi);
1985 	}
1986 	mtx_unlock(&umtx_lock);
1987 }
1988 
1989 /*
1990  * Sleep on a PI mutex.
1991  */
1992 int
1993 umtxq_sleep_pi(struct umtx_q *uq, struct umtx_pi *pi, uint32_t owner,
1994     const char *wmesg, struct umtx_abs_timeout *timo, bool shared)
1995 {
1996 	struct thread *td, *td1;
1997 	struct umtx_q *uq1;
1998 	int error, pri;
1999 #ifdef INVARIANTS
2000 	struct umtxq_chain *uc;
2001 
2002 	uc = umtxq_getchain(&pi->pi_key);
2003 #endif
2004 	error = 0;
2005 	td = uq->uq_thread;
2006 	KASSERT(td == curthread, ("inconsistent uq_thread"));
2007 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
2008 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));
2009 	umtxq_insert(uq);
2010 	mtx_lock(&umtx_lock);
2011 	if (pi->pi_owner == NULL) {
2012 		mtx_unlock(&umtx_lock);
2013 		td1 = tdfind(owner, shared ? -1 : td->td_proc->p_pid);
2014 		mtx_lock(&umtx_lock);
2015 		if (td1 != NULL) {
2016 			if (pi->pi_owner == NULL)
2017 				umtx_pi_setowner(pi, td1);
2018 			PROC_UNLOCK(td1->td_proc);
2019 		}
2020 	}
2021 
2022 	TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
2023 		pri = UPRI(uq1->uq_thread);
2024 		if (pri > UPRI(td))
2025 			break;
2026 	}
2027 
2028 	if (uq1 != NULL)
2029 		TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
2030 	else
2031 		TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
2032 
2033 	uq->uq_pi_blocked = pi;
2034 	thread_lock(td);
2035 	td->td_flags |= TDF_UPIBLOCKED;
2036 	thread_unlock(td);
2037 	umtx_propagate_priority(td);
2038 	mtx_unlock(&umtx_lock);
2039 	umtxq_unbusy(&uq->uq_key);
2040 
2041 	error = umtxq_sleep(uq, wmesg, timo);
2042 	umtxq_remove(uq);
2043 
2044 	mtx_lock(&umtx_lock);
2045 	uq->uq_pi_blocked = NULL;
2046 	thread_lock(td);
2047 	td->td_flags &= ~TDF_UPIBLOCKED;
2048 	thread_unlock(td);
2049 	TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
2050 	umtx_repropagate_priority(pi);
2051 	mtx_unlock(&umtx_lock);
2052 	umtxq_unlock(&uq->uq_key);
2053 
2054 	return (error);
2055 }
2056 
2057 /*
2058  * Add reference count for a PI mutex.
2059  */
2060 void
2061 umtx_pi_ref(struct umtx_pi *pi)
2062 {
2063 
2064 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&pi->pi_key));
2065 	pi->pi_refcount++;
2066 }
2067 
2068 /*
2069  * Decrease reference count for a PI mutex, if the counter
2070  * is decreased to zero, its memory space is freed.
2071  */
2072 void
2073 umtx_pi_unref(struct umtx_pi *pi)
2074 {
2075 	struct umtxq_chain *uc;
2076 
2077 	uc = umtxq_getchain(&pi->pi_key);
2078 	UMTXQ_LOCKED_ASSERT(uc);
2079 	KASSERT(pi->pi_refcount > 0, ("invalid reference count"));
2080 	if (--pi->pi_refcount == 0) {
2081 		mtx_lock(&umtx_lock);
2082 		if (pi->pi_owner != NULL)
2083 			umtx_pi_disown(pi);
2084 		KASSERT(TAILQ_EMPTY(&pi->pi_blocked),
2085 			("blocked queue not empty"));
2086 		mtx_unlock(&umtx_lock);
2087 		TAILQ_REMOVE(&uc->uc_pi_list, pi, pi_hashlink);
2088 		umtx_pi_free(pi);
2089 	}
2090 }
2091 
2092 /*
2093  * Find a PI mutex in hash table.
2094  */
2095 struct umtx_pi *
2096 umtx_pi_lookup(struct umtx_key *key)
2097 {
2098 	struct umtxq_chain *uc;
2099 	struct umtx_pi *pi;
2100 
2101 	uc = umtxq_getchain(key);
2102 	UMTXQ_LOCKED_ASSERT(uc);
2103 
2104 	TAILQ_FOREACH(pi, &uc->uc_pi_list, pi_hashlink) {
2105 		if (umtx_key_match(&pi->pi_key, key)) {
2106 			return (pi);
2107 		}
2108 	}
2109 	return (NULL);
2110 }
2111 
2112 /*
2113  * Insert a PI mutex into hash table.
2114  */
2115 void
2116 umtx_pi_insert(struct umtx_pi *pi)
2117 {
2118 	struct umtxq_chain *uc;
2119 
2120 	uc = umtxq_getchain(&pi->pi_key);
2121 	UMTXQ_LOCKED_ASSERT(uc);
2122 	TAILQ_INSERT_TAIL(&uc->uc_pi_list, pi, pi_hashlink);
2123 }
2124 
2125 /*
2126  * Drop a PI mutex and wakeup a top waiter.
2127  */
2128 int
2129 umtx_pi_drop(struct thread *td, struct umtx_key *key, bool rb, int *count)
2130 {
2131 	struct umtx_q *uq_first, *uq_first2, *uq_me;
2132 	struct umtx_pi *pi, *pi2;
2133 	int pri;
2134 
2135 	UMTXQ_ASSERT_LOCKED_BUSY(key);
2136 	*count = umtxq_count_pi(key, &uq_first);
2137 	if (uq_first != NULL) {
2138 		mtx_lock(&umtx_lock);
2139 		pi = uq_first->uq_pi_blocked;
2140 		KASSERT(pi != NULL, ("pi == NULL?"));
2141 		if (pi->pi_owner != td && !(rb && pi->pi_owner == NULL)) {
2142 			mtx_unlock(&umtx_lock);
2143 			/* userland messed the mutex */
2144 			return (EPERM);
2145 		}
2146 		uq_me = td->td_umtxq;
2147 		if (pi->pi_owner == td)
2148 			umtx_pi_disown(pi);
2149 		/* get highest priority thread which is still sleeping. */
2150 		uq_first = TAILQ_FIRST(&pi->pi_blocked);
2151 		while (uq_first != NULL &&
2152 		    (uq_first->uq_flags & UQF_UMTXQ) == 0) {
2153 			uq_first = TAILQ_NEXT(uq_first, uq_lockq);
2154 		}
2155 		pri = PRI_MAX;
2156 		TAILQ_FOREACH(pi2, &uq_me->uq_pi_contested, pi_link) {
2157 			uq_first2 = TAILQ_FIRST(&pi2->pi_blocked);
2158 			if (uq_first2 != NULL) {
2159 				if (pri > UPRI(uq_first2->uq_thread))
2160 					pri = UPRI(uq_first2->uq_thread);
2161 			}
2162 		}
2163 		thread_lock(td);
2164 		sched_lend_user_prio(td, pri);
2165 		thread_unlock(td);
2166 		mtx_unlock(&umtx_lock);
2167 		if (uq_first)
2168 			umtxq_signal_thread(uq_first);
2169 	} else {
2170 		pi = umtx_pi_lookup(key);
2171 		/*
2172 		 * A umtx_pi can exist if a signal or timeout removed the
2173 		 * last waiter from the umtxq, but there is still
2174 		 * a thread in do_lock_pi() holding the umtx_pi.
2175 		 */
2176 		if (pi != NULL) {
2177 			/*
2178 			 * The umtx_pi can be unowned, such as when a thread
2179 			 * has just entered do_lock_pi(), allocated the
2180 			 * umtx_pi, and unlocked the umtxq.
2181 			 * If the current thread owns it, it must disown it.
2182 			 */
2183 			mtx_lock(&umtx_lock);
2184 			if (pi->pi_owner == td)
2185 				umtx_pi_disown(pi);
2186 			mtx_unlock(&umtx_lock);
2187 		}
2188 	}
2189 	return (0);
2190 }
2191 
2192 /*
2193  * Lock a PI mutex.
2194  */
2195 static int
2196 do_lock_pi(struct thread *td, struct umutex *m, uint32_t flags,
2197     struct _umtx_time *timeout, int try)
2198 {
2199 	struct umtx_abs_timeout timo;
2200 	struct umtx_q *uq;
2201 	struct umtx_pi *pi, *new_pi;
2202 	uint32_t id, old_owner, owner, old;
2203 	int error, rv;
2204 
2205 	id = td->td_tid;
2206 	uq = td->td_umtxq;
2207 
2208 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2209 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2210 	    &uq->uq_key)) != 0)
2211 		return (error);
2212 
2213 	if (timeout != NULL)
2214 		umtx_abs_timeout_init2(&timo, timeout);
2215 
2216 	umtxq_lock(&uq->uq_key);
2217 	pi = umtx_pi_lookup(&uq->uq_key);
2218 	if (pi == NULL) {
2219 		new_pi = umtx_pi_alloc(M_NOWAIT);
2220 		if (new_pi == NULL) {
2221 			umtxq_unlock(&uq->uq_key);
2222 			new_pi = umtx_pi_alloc(M_WAITOK);
2223 			umtxq_lock(&uq->uq_key);
2224 			pi = umtx_pi_lookup(&uq->uq_key);
2225 			if (pi != NULL) {
2226 				umtx_pi_free(new_pi);
2227 				new_pi = NULL;
2228 			}
2229 		}
2230 		if (new_pi != NULL) {
2231 			new_pi->pi_key = uq->uq_key;
2232 			umtx_pi_insert(new_pi);
2233 			pi = new_pi;
2234 		}
2235 	}
2236 	umtx_pi_ref(pi);
2237 	umtxq_unlock(&uq->uq_key);
2238 
2239 	/*
2240 	 * Care must be exercised when dealing with umtx structure.  It
2241 	 * can fault on any access.
2242 	 */
2243 	for (;;) {
2244 		/*
2245 		 * Try the uncontested case.  This should be done in userland.
2246 		 */
2247 		rv = casueword32(&m->m_owner, UMUTEX_UNOWNED, &owner, id);
2248 		/* The address was invalid. */
2249 		if (rv == -1) {
2250 			error = EFAULT;
2251 			break;
2252 		}
2253 		/* The acquire succeeded. */
2254 		if (rv == 0) {
2255 			MPASS(owner == UMUTEX_UNOWNED);
2256 			error = 0;
2257 			break;
2258 		}
2259 
2260 		if (owner == UMUTEX_RB_NOTRECOV) {
2261 			error = ENOTRECOVERABLE;
2262 			break;
2263 		}
2264 
2265 		/*
2266 		 * Avoid overwriting a possible error from sleep due
2267 		 * to the pending signal with suspension check result.
2268 		 */
2269 		if (error == 0) {
2270 			error = thread_check_susp(td, true);
2271 			if (error != 0)
2272 				break;
2273 		}
2274 
2275 		/* If no one owns it but it is contested try to acquire it. */
2276 		if (owner == UMUTEX_CONTESTED || owner == UMUTEX_RB_OWNERDEAD) {
2277 			old_owner = owner;
2278 			rv = casueword32(&m->m_owner, owner, &owner,
2279 			    id | UMUTEX_CONTESTED);
2280 			/* The address was invalid. */
2281 			if (rv == -1) {
2282 				error = EFAULT;
2283 				break;
2284 			}
2285 			if (rv == 1) {
2286 				if (error == 0) {
2287 					error = thread_check_susp(td, true);
2288 					if (error != 0)
2289 						break;
2290 				}
2291 
2292 				/*
2293 				 * If this failed the lock could
2294 				 * changed, restart.
2295 				 */
2296 				continue;
2297 			}
2298 
2299 			MPASS(rv == 0);
2300 			MPASS(owner == old_owner);
2301 			umtxq_lock(&uq->uq_key);
2302 			umtxq_busy(&uq->uq_key);
2303 			error = umtx_pi_claim(pi, td);
2304 			umtxq_unbusy(&uq->uq_key);
2305 			umtxq_unlock(&uq->uq_key);
2306 			if (error != 0) {
2307 				/*
2308 				 * Since we're going to return an
2309 				 * error, restore the m_owner to its
2310 				 * previous, unowned state to avoid
2311 				 * compounding the problem.
2312 				 */
2313 				(void)casuword32(&m->m_owner,
2314 				    id | UMUTEX_CONTESTED, old_owner);
2315 			}
2316 			if (error == 0 && old_owner == UMUTEX_RB_OWNERDEAD)
2317 				error = EOWNERDEAD;
2318 			break;
2319 		}
2320 
2321 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2322 			error = EDEADLK;
2323 			break;
2324 		}
2325 
2326 		if (try != 0) {
2327 			error = EBUSY;
2328 			break;
2329 		}
2330 
2331 		/*
2332 		 * If we caught a signal, we have retried and now
2333 		 * exit immediately.
2334 		 */
2335 		if (error != 0)
2336 			break;
2337 
2338 		umtxq_lock(&uq->uq_key);
2339 		umtxq_busy(&uq->uq_key);
2340 		umtxq_unlock(&uq->uq_key);
2341 
2342 		/*
2343 		 * Set the contested bit so that a release in user space
2344 		 * knows to use the system call for unlock.  If this fails
2345 		 * either some one else has acquired the lock or it has been
2346 		 * released.
2347 		 */
2348 		rv = casueword32(&m->m_owner, owner, &old, owner |
2349 		    UMUTEX_CONTESTED);
2350 
2351 		/* The address was invalid. */
2352 		if (rv == -1) {
2353 			umtxq_unbusy_unlocked(&uq->uq_key);
2354 			error = EFAULT;
2355 			break;
2356 		}
2357 		if (rv == 1) {
2358 			umtxq_unbusy_unlocked(&uq->uq_key);
2359 			error = thread_check_susp(td, true);
2360 			if (error != 0)
2361 				break;
2362 
2363 			/*
2364 			 * The lock changed and we need to retry or we
2365 			 * lost a race to the thread unlocking the
2366 			 * umtx.  Note that the UMUTEX_RB_OWNERDEAD
2367 			 * value for owner is impossible there.
2368 			 */
2369 			continue;
2370 		}
2371 
2372 		umtxq_lock(&uq->uq_key);
2373 
2374 		/* We set the contested bit, sleep. */
2375 		MPASS(old == owner);
2376 		error = umtxq_sleep_pi(uq, pi, owner & ~UMUTEX_CONTESTED,
2377 		    "umtxpi", timeout == NULL ? NULL : &timo,
2378 		    (flags & USYNC_PROCESS_SHARED) != 0);
2379 		if (error != 0)
2380 			continue;
2381 
2382 		error = thread_check_susp(td, false);
2383 		if (error != 0)
2384 			break;
2385 	}
2386 
2387 	umtxq_lock(&uq->uq_key);
2388 	umtx_pi_unref(pi);
2389 	umtxq_unlock(&uq->uq_key);
2390 
2391 	umtx_key_release(&uq->uq_key);
2392 	return (error);
2393 }
2394 
2395 /*
2396  * Unlock a PI mutex.
2397  */
2398 static int
2399 do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2400 {
2401 	struct umtx_key key;
2402 	uint32_t id, new_owner, old, owner;
2403 	int count, error;
2404 
2405 	id = td->td_tid;
2406 
2407 usrloop:
2408 	/*
2409 	 * Make sure we own this mtx.
2410 	 */
2411 	error = fueword32(&m->m_owner, &owner);
2412 	if (error == -1)
2413 		return (EFAULT);
2414 
2415 	if ((owner & ~UMUTEX_CONTESTED) != id)
2416 		return (EPERM);
2417 
2418 	new_owner = umtx_unlock_val(flags, rb);
2419 
2420 	/* This should be done in userland */
2421 	if ((owner & UMUTEX_CONTESTED) == 0) {
2422 		error = casueword32(&m->m_owner, owner, &old, new_owner);
2423 		if (error == -1)
2424 			return (EFAULT);
2425 		if (error == 1) {
2426 			error = thread_check_susp(td, true);
2427 			if (error != 0)
2428 				return (error);
2429 			goto usrloop;
2430 		}
2431 		if (old == owner)
2432 			return (0);
2433 		owner = old;
2434 	}
2435 
2436 	/* We should only ever be in here for contested locks */
2437 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2438 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2439 	    &key)) != 0)
2440 		return (error);
2441 
2442 	umtxq_lock(&key);
2443 	umtxq_busy(&key);
2444 	error = umtx_pi_drop(td, &key, rb, &count);
2445 	if (error != 0) {
2446 		umtxq_unbusy(&key);
2447 		umtxq_unlock(&key);
2448 		umtx_key_release(&key);
2449 		/* userland messed the mutex */
2450 		return (error);
2451 	}
2452 	umtxq_unlock(&key);
2453 
2454 	/*
2455 	 * When unlocking the umtx, it must be marked as unowned if
2456 	 * there is zero or one thread only waiting for it.
2457 	 * Otherwise, it must be marked as contested.
2458 	 */
2459 
2460 	if (count > 1)
2461 		new_owner |= UMUTEX_CONTESTED;
2462 again:
2463 	error = casueword32(&m->m_owner, owner, &old, new_owner);
2464 	if (error == 1) {
2465 		error = thread_check_susp(td, false);
2466 		if (error == 0)
2467 			goto again;
2468 	}
2469 	umtxq_unbusy_unlocked(&key);
2470 	umtx_key_release(&key);
2471 	if (error == -1)
2472 		return (EFAULT);
2473 	if (error == 0 && old != owner)
2474 		return (EINVAL);
2475 	return (error);
2476 }
2477 
2478 /*
2479  * Lock a PP mutex.
2480  */
2481 static int
2482 do_lock_pp(struct thread *td, struct umutex *m, uint32_t flags,
2483     struct _umtx_time *timeout, int try)
2484 {
2485 	struct umtx_abs_timeout timo;
2486 	struct umtx_q *uq, *uq2;
2487 	struct umtx_pi *pi;
2488 	uint32_t ceiling;
2489 	uint32_t owner, id;
2490 	int error, pri, old_inherited_pri, su, rv;
2491 
2492 	id = td->td_tid;
2493 	uq = td->td_umtxq;
2494 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2495 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2496 	    &uq->uq_key)) != 0)
2497 		return (error);
2498 
2499 	if (timeout != NULL)
2500 		umtx_abs_timeout_init2(&timo, timeout);
2501 
2502 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2503 	for (;;) {
2504 		old_inherited_pri = uq->uq_inherited_pri;
2505 		umtxq_lock(&uq->uq_key);
2506 		umtxq_busy(&uq->uq_key);
2507 		umtxq_unlock(&uq->uq_key);
2508 
2509 		rv = fueword32(&m->m_ceilings[0], &ceiling);
2510 		if (rv == -1) {
2511 			error = EFAULT;
2512 			goto out;
2513 		}
2514 		ceiling = RTP_PRIO_MAX - ceiling;
2515 		if (ceiling > RTP_PRIO_MAX) {
2516 			error = EINVAL;
2517 			goto out;
2518 		}
2519 
2520 		mtx_lock(&umtx_lock);
2521 		if (UPRI(td) < PRI_MIN_REALTIME + ceiling) {
2522 			mtx_unlock(&umtx_lock);
2523 			error = EINVAL;
2524 			goto out;
2525 		}
2526 		if (su && PRI_MIN_REALTIME + ceiling < uq->uq_inherited_pri) {
2527 			uq->uq_inherited_pri = PRI_MIN_REALTIME + ceiling;
2528 			thread_lock(td);
2529 			if (uq->uq_inherited_pri < UPRI(td))
2530 				sched_lend_user_prio(td, uq->uq_inherited_pri);
2531 			thread_unlock(td);
2532 		}
2533 		mtx_unlock(&umtx_lock);
2534 
2535 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2536 		    id | UMUTEX_CONTESTED);
2537 		/* The address was invalid. */
2538 		if (rv == -1) {
2539 			error = EFAULT;
2540 			break;
2541 		}
2542 		if (rv == 0) {
2543 			MPASS(owner == UMUTEX_CONTESTED);
2544 			error = 0;
2545 			break;
2546 		}
2547 		/* rv == 1 */
2548 		if (owner == UMUTEX_RB_OWNERDEAD) {
2549 			rv = casueword32(&m->m_owner, UMUTEX_RB_OWNERDEAD,
2550 			    &owner, id | UMUTEX_CONTESTED);
2551 			if (rv == -1) {
2552 				error = EFAULT;
2553 				break;
2554 			}
2555 			if (rv == 0) {
2556 				MPASS(owner == UMUTEX_RB_OWNERDEAD);
2557 				error = EOWNERDEAD; /* success */
2558 				break;
2559 			}
2560 
2561 			/*
2562 			 *  rv == 1, only check for suspension if we
2563 			 *  did not already catched a signal.  If we
2564 			 *  get an error from the check, the same
2565 			 *  condition is checked by the umtxq_sleep()
2566 			 *  call below, so we should obliterate the
2567 			 *  error to not skip the last loop iteration.
2568 			 */
2569 			if (error == 0) {
2570 				error = thread_check_susp(td, false);
2571 				if (error == 0) {
2572 					if (try != 0)
2573 						error = EBUSY;
2574 					else
2575 						continue;
2576 				}
2577 				error = 0;
2578 			}
2579 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2580 			error = ENOTRECOVERABLE;
2581 		}
2582 
2583 		if (try != 0)
2584 			error = EBUSY;
2585 
2586 		/*
2587 		 * If we caught a signal, we have retried and now
2588 		 * exit immediately.
2589 		 */
2590 		if (error != 0)
2591 			break;
2592 
2593 		umtxq_lock(&uq->uq_key);
2594 		umtxq_insert(uq);
2595 		umtxq_unbusy(&uq->uq_key);
2596 		error = umtxq_sleep(uq, "umtxpp", timeout == NULL ?
2597 		    NULL : &timo);
2598 		umtxq_remove(uq);
2599 		umtxq_unlock(&uq->uq_key);
2600 
2601 		mtx_lock(&umtx_lock);
2602 		uq->uq_inherited_pri = old_inherited_pri;
2603 		pri = PRI_MAX;
2604 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2605 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2606 			if (uq2 != NULL) {
2607 				if (pri > UPRI(uq2->uq_thread))
2608 					pri = UPRI(uq2->uq_thread);
2609 			}
2610 		}
2611 		if (pri > uq->uq_inherited_pri)
2612 			pri = uq->uq_inherited_pri;
2613 		thread_lock(td);
2614 		sched_lend_user_prio(td, pri);
2615 		thread_unlock(td);
2616 		mtx_unlock(&umtx_lock);
2617 	}
2618 
2619 	if (error != 0 && error != EOWNERDEAD) {
2620 		mtx_lock(&umtx_lock);
2621 		uq->uq_inherited_pri = old_inherited_pri;
2622 		pri = PRI_MAX;
2623 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2624 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2625 			if (uq2 != NULL) {
2626 				if (pri > UPRI(uq2->uq_thread))
2627 					pri = UPRI(uq2->uq_thread);
2628 			}
2629 		}
2630 		if (pri > uq->uq_inherited_pri)
2631 			pri = uq->uq_inherited_pri;
2632 		thread_lock(td);
2633 		sched_lend_user_prio(td, pri);
2634 		thread_unlock(td);
2635 		mtx_unlock(&umtx_lock);
2636 	}
2637 
2638 out:
2639 	umtxq_unbusy_unlocked(&uq->uq_key);
2640 	umtx_key_release(&uq->uq_key);
2641 	return (error);
2642 }
2643 
2644 /*
2645  * Unlock a PP mutex.
2646  */
2647 static int
2648 do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2649 {
2650 	struct umtx_key key;
2651 	struct umtx_q *uq, *uq2;
2652 	struct umtx_pi *pi;
2653 	uint32_t id, owner, rceiling;
2654 	int error, pri, new_inherited_pri, su;
2655 
2656 	id = td->td_tid;
2657 	uq = td->td_umtxq;
2658 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2659 
2660 	/*
2661 	 * Make sure we own this mtx.
2662 	 */
2663 	error = fueword32(&m->m_owner, &owner);
2664 	if (error == -1)
2665 		return (EFAULT);
2666 
2667 	if ((owner & ~UMUTEX_CONTESTED) != id)
2668 		return (EPERM);
2669 
2670 	error = copyin(&m->m_ceilings[1], &rceiling, sizeof(uint32_t));
2671 	if (error != 0)
2672 		return (error);
2673 
2674 	if (rceiling == -1)
2675 		new_inherited_pri = PRI_MAX;
2676 	else {
2677 		rceiling = RTP_PRIO_MAX - rceiling;
2678 		if (rceiling > RTP_PRIO_MAX)
2679 			return (EINVAL);
2680 		new_inherited_pri = PRI_MIN_REALTIME + rceiling;
2681 	}
2682 
2683 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2684 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2685 	    &key)) != 0)
2686 		return (error);
2687 	umtxq_lock(&key);
2688 	umtxq_busy(&key);
2689 	umtxq_unlock(&key);
2690 	/*
2691 	 * For priority protected mutex, always set unlocked state
2692 	 * to UMUTEX_CONTESTED, so that userland always enters kernel
2693 	 * to lock the mutex, it is necessary because thread priority
2694 	 * has to be adjusted for such mutex.
2695 	 */
2696 	error = suword32(&m->m_owner, umtx_unlock_val(flags, rb) |
2697 	    UMUTEX_CONTESTED);
2698 
2699 	umtxq_lock(&key);
2700 	if (error == 0)
2701 		umtxq_signal(&key, 1);
2702 	umtxq_unbusy(&key);
2703 	umtxq_unlock(&key);
2704 
2705 	if (error == -1)
2706 		error = EFAULT;
2707 	else {
2708 		mtx_lock(&umtx_lock);
2709 		if (su != 0)
2710 			uq->uq_inherited_pri = new_inherited_pri;
2711 		pri = PRI_MAX;
2712 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2713 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2714 			if (uq2 != NULL) {
2715 				if (pri > UPRI(uq2->uq_thread))
2716 					pri = UPRI(uq2->uq_thread);
2717 			}
2718 		}
2719 		if (pri > uq->uq_inherited_pri)
2720 			pri = uq->uq_inherited_pri;
2721 		thread_lock(td);
2722 		sched_lend_user_prio(td, pri);
2723 		thread_unlock(td);
2724 		mtx_unlock(&umtx_lock);
2725 	}
2726 	umtx_key_release(&key);
2727 	return (error);
2728 }
2729 
2730 static int
2731 do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
2732     uint32_t *old_ceiling)
2733 {
2734 	struct umtx_q *uq;
2735 	uint32_t flags, id, owner, save_ceiling;
2736 	int error, rv, rv1;
2737 
2738 	error = fueword32(&m->m_flags, &flags);
2739 	if (error == -1)
2740 		return (EFAULT);
2741 	if ((flags & UMUTEX_PRIO_PROTECT) == 0)
2742 		return (EINVAL);
2743 	if (ceiling > RTP_PRIO_MAX)
2744 		return (EINVAL);
2745 	id = td->td_tid;
2746 	uq = td->td_umtxq;
2747 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2748 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2749 	    &uq->uq_key)) != 0)
2750 		return (error);
2751 	for (;;) {
2752 		umtxq_lock(&uq->uq_key);
2753 		umtxq_busy(&uq->uq_key);
2754 		umtxq_unlock(&uq->uq_key);
2755 
2756 		rv = fueword32(&m->m_ceilings[0], &save_ceiling);
2757 		if (rv == -1) {
2758 			error = EFAULT;
2759 			break;
2760 		}
2761 
2762 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2763 		    id | UMUTEX_CONTESTED);
2764 		if (rv == -1) {
2765 			error = EFAULT;
2766 			break;
2767 		}
2768 
2769 		if (rv == 0) {
2770 			MPASS(owner == UMUTEX_CONTESTED);
2771 			rv = suword32(&m->m_ceilings[0], ceiling);
2772 			rv1 = suword32(&m->m_owner, UMUTEX_CONTESTED);
2773 			error = (rv == 0 && rv1 == 0) ? 0: EFAULT;
2774 			break;
2775 		}
2776 
2777 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2778 			rv = suword32(&m->m_ceilings[0], ceiling);
2779 			error = rv == 0 ? 0 : EFAULT;
2780 			break;
2781 		}
2782 
2783 		if (owner == UMUTEX_RB_OWNERDEAD) {
2784 			error = EOWNERDEAD;
2785 			break;
2786 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2787 			error = ENOTRECOVERABLE;
2788 			break;
2789 		}
2790 
2791 		/*
2792 		 * If we caught a signal, we have retried and now
2793 		 * exit immediately.
2794 		 */
2795 		if (error != 0)
2796 			break;
2797 
2798 		/*
2799 		 * We set the contested bit, sleep. Otherwise the lock changed
2800 		 * and we need to retry or we lost a race to the thread
2801 		 * unlocking the umtx.
2802 		 */
2803 		umtxq_lock(&uq->uq_key);
2804 		umtxq_insert(uq);
2805 		umtxq_unbusy(&uq->uq_key);
2806 		error = umtxq_sleep(uq, "umtxpp", NULL);
2807 		umtxq_remove(uq);
2808 		umtxq_unlock(&uq->uq_key);
2809 	}
2810 	umtxq_lock(&uq->uq_key);
2811 	if (error == 0)
2812 		umtxq_signal(&uq->uq_key, INT_MAX);
2813 	umtxq_unbusy(&uq->uq_key);
2814 	umtxq_unlock(&uq->uq_key);
2815 	umtx_key_release(&uq->uq_key);
2816 	if (error == 0 && old_ceiling != NULL) {
2817 		rv = suword32(old_ceiling, save_ceiling);
2818 		error = rv == 0 ? 0 : EFAULT;
2819 	}
2820 	return (error);
2821 }
2822 
2823 /*
2824  * Lock a userland POSIX mutex.
2825  */
2826 static int
2827 do_lock_umutex(struct thread *td, struct umutex *m,
2828     struct _umtx_time *timeout, int mode)
2829 {
2830 	uint32_t flags;
2831 	int error;
2832 
2833 	error = fueword32(&m->m_flags, &flags);
2834 	if (error == -1)
2835 		return (EFAULT);
2836 
2837 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2838 	case 0:
2839 		error = do_lock_normal(td, m, flags, timeout, mode);
2840 		break;
2841 	case UMUTEX_PRIO_INHERIT:
2842 		error = do_lock_pi(td, m, flags, timeout, mode);
2843 		break;
2844 	case UMUTEX_PRIO_PROTECT:
2845 		error = do_lock_pp(td, m, flags, timeout, mode);
2846 		break;
2847 	default:
2848 		return (EINVAL);
2849 	}
2850 	if (timeout == NULL) {
2851 		if (error == EINTR && mode != _UMUTEX_WAIT)
2852 			error = ERESTART;
2853 	} else {
2854 		/* Timed-locking is not restarted. */
2855 		if (error == ERESTART)
2856 			error = EINTR;
2857 	}
2858 	return (error);
2859 }
2860 
2861 /*
2862  * Unlock a userland POSIX mutex.
2863  */
2864 static int
2865 do_unlock_umutex(struct thread *td, struct umutex *m, bool rb)
2866 {
2867 	uint32_t flags;
2868 	int error;
2869 
2870 	error = fueword32(&m->m_flags, &flags);
2871 	if (error == -1)
2872 		return (EFAULT);
2873 
2874 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2875 	case 0:
2876 		return (do_unlock_normal(td, m, flags, rb));
2877 	case UMUTEX_PRIO_INHERIT:
2878 		return (do_unlock_pi(td, m, flags, rb));
2879 	case UMUTEX_PRIO_PROTECT:
2880 		return (do_unlock_pp(td, m, flags, rb));
2881 	}
2882 
2883 	return (EINVAL);
2884 }
2885 
2886 static int
2887 do_cv_wait(struct thread *td, struct ucond *cv, struct umutex *m,
2888     struct timespec *timeout, u_long wflags)
2889 {
2890 	struct umtx_abs_timeout timo;
2891 	struct umtx_q *uq;
2892 	uint32_t flags, clockid, hasw;
2893 	int error;
2894 
2895 	uq = td->td_umtxq;
2896 	error = fueword32(&cv->c_flags, &flags);
2897 	if (error == -1)
2898 		return (EFAULT);
2899 	error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &uq->uq_key);
2900 	if (error != 0)
2901 		return (error);
2902 
2903 	if ((wflags & CVWAIT_CLOCKID) != 0) {
2904 		error = fueword32(&cv->c_clockid, &clockid);
2905 		if (error == -1) {
2906 			umtx_key_release(&uq->uq_key);
2907 			return (EFAULT);
2908 		}
2909 		if (clockid < CLOCK_REALTIME ||
2910 		    clockid >= CLOCK_THREAD_CPUTIME_ID) {
2911 			/* hmm, only HW clock id will work. */
2912 			umtx_key_release(&uq->uq_key);
2913 			return (EINVAL);
2914 		}
2915 	} else {
2916 		clockid = CLOCK_REALTIME;
2917 	}
2918 
2919 	umtxq_lock(&uq->uq_key);
2920 	umtxq_busy(&uq->uq_key);
2921 	umtxq_insert(uq);
2922 	umtxq_unlock(&uq->uq_key);
2923 
2924 	/*
2925 	 * Set c_has_waiters to 1 before releasing user mutex, also
2926 	 * don't modify cache line when unnecessary.
2927 	 */
2928 	error = fueword32(&cv->c_has_waiters, &hasw);
2929 	if (error == 0 && hasw == 0)
2930 		suword32(&cv->c_has_waiters, 1);
2931 
2932 	umtxq_unbusy_unlocked(&uq->uq_key);
2933 
2934 	error = do_unlock_umutex(td, m, false);
2935 
2936 	if (timeout != NULL)
2937 		umtx_abs_timeout_init(&timo, clockid,
2938 		    (wflags & CVWAIT_ABSTIME) != 0, timeout);
2939 
2940 	umtxq_lock(&uq->uq_key);
2941 	if (error == 0) {
2942 		error = umtxq_sleep(uq, "ucond", timeout == NULL ?
2943 		    NULL : &timo);
2944 	}
2945 
2946 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
2947 		error = 0;
2948 	else {
2949 		/*
2950 		 * This must be timeout,interrupted by signal or
2951 		 * surprious wakeup, clear c_has_waiter flag when
2952 		 * necessary.
2953 		 */
2954 		umtxq_busy(&uq->uq_key);
2955 		if ((uq->uq_flags & UQF_UMTXQ) != 0) {
2956 			int oldlen = uq->uq_cur_queue->length;
2957 			umtxq_remove(uq);
2958 			if (oldlen == 1) {
2959 				umtxq_unlock(&uq->uq_key);
2960 				suword32(&cv->c_has_waiters, 0);
2961 				umtxq_lock(&uq->uq_key);
2962 			}
2963 		}
2964 		umtxq_unbusy(&uq->uq_key);
2965 		if (error == ERESTART)
2966 			error = EINTR;
2967 	}
2968 
2969 	umtxq_unlock(&uq->uq_key);
2970 	umtx_key_release(&uq->uq_key);
2971 	return (error);
2972 }
2973 
2974 /*
2975  * Signal a userland condition variable.
2976  */
2977 static int
2978 do_cv_signal(struct thread *td, struct ucond *cv)
2979 {
2980 	struct umtx_key key;
2981 	int error, cnt, nwake;
2982 	uint32_t flags;
2983 
2984 	error = fueword32(&cv->c_flags, &flags);
2985 	if (error == -1)
2986 		return (EFAULT);
2987 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
2988 		return (error);
2989 	umtxq_lock(&key);
2990 	umtxq_busy(&key);
2991 	cnt = umtxq_count(&key);
2992 	nwake = umtxq_signal(&key, 1);
2993 	if (cnt <= nwake) {
2994 		umtxq_unlock(&key);
2995 		error = suword32(&cv->c_has_waiters, 0);
2996 		if (error == -1)
2997 			error = EFAULT;
2998 		umtxq_lock(&key);
2999 	}
3000 	umtxq_unbusy(&key);
3001 	umtxq_unlock(&key);
3002 	umtx_key_release(&key);
3003 	return (error);
3004 }
3005 
3006 static int
3007 do_cv_broadcast(struct thread *td, struct ucond *cv)
3008 {
3009 	struct umtx_key key;
3010 	int error;
3011 	uint32_t flags;
3012 
3013 	error = fueword32(&cv->c_flags, &flags);
3014 	if (error == -1)
3015 		return (EFAULT);
3016 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
3017 		return (error);
3018 
3019 	umtxq_lock(&key);
3020 	umtxq_busy(&key);
3021 	umtxq_signal(&key, INT_MAX);
3022 	umtxq_unlock(&key);
3023 
3024 	error = suword32(&cv->c_has_waiters, 0);
3025 	if (error == -1)
3026 		error = EFAULT;
3027 
3028 	umtxq_unbusy_unlocked(&key);
3029 
3030 	umtx_key_release(&key);
3031 	return (error);
3032 }
3033 
3034 static int
3035 do_rw_rdlock(struct thread *td, struct urwlock *rwlock, long fflag,
3036     struct _umtx_time *timeout)
3037 {
3038 	struct umtx_abs_timeout timo;
3039 	struct umtx_q *uq;
3040 	uint32_t flags, wrflags;
3041 	int32_t state, oldstate;
3042 	int32_t blocked_readers;
3043 	int error, error1, rv;
3044 
3045 	uq = td->td_umtxq;
3046 	error = fueword32(&rwlock->rw_flags, &flags);
3047 	if (error == -1)
3048 		return (EFAULT);
3049 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3050 	if (error != 0)
3051 		return (error);
3052 
3053 	if (timeout != NULL)
3054 		umtx_abs_timeout_init2(&timo, timeout);
3055 
3056 	wrflags = URWLOCK_WRITE_OWNER;
3057 	if (!(fflag & URWLOCK_PREFER_READER) && !(flags & URWLOCK_PREFER_READER))
3058 		wrflags |= URWLOCK_WRITE_WAITERS;
3059 
3060 	for (;;) {
3061 		rv = fueword32(&rwlock->rw_state, &state);
3062 		if (rv == -1) {
3063 			umtx_key_release(&uq->uq_key);
3064 			return (EFAULT);
3065 		}
3066 
3067 		/* try to lock it */
3068 		while (!(state & wrflags)) {
3069 			if (__predict_false(URWLOCK_READER_COUNT(state) ==
3070 			    URWLOCK_MAX_READERS)) {
3071 				umtx_key_release(&uq->uq_key);
3072 				return (EAGAIN);
3073 			}
3074 			rv = casueword32(&rwlock->rw_state, state,
3075 			    &oldstate, state + 1);
3076 			if (rv == -1) {
3077 				umtx_key_release(&uq->uq_key);
3078 				return (EFAULT);
3079 			}
3080 			if (rv == 0) {
3081 				MPASS(oldstate == state);
3082 				umtx_key_release(&uq->uq_key);
3083 				return (0);
3084 			}
3085 			error = thread_check_susp(td, true);
3086 			if (error != 0)
3087 				break;
3088 			state = oldstate;
3089 		}
3090 
3091 		if (error)
3092 			break;
3093 
3094 		/* grab monitor lock */
3095 		umtxq_lock(&uq->uq_key);
3096 		umtxq_busy(&uq->uq_key);
3097 		umtxq_unlock(&uq->uq_key);
3098 
3099 		/*
3100 		 * re-read the state, in case it changed between the try-lock above
3101 		 * and the check below
3102 		 */
3103 		rv = fueword32(&rwlock->rw_state, &state);
3104 		if (rv == -1)
3105 			error = EFAULT;
3106 
3107 		/* set read contention bit */
3108 		while (error == 0 && (state & wrflags) &&
3109 		    !(state & URWLOCK_READ_WAITERS)) {
3110 			rv = casueword32(&rwlock->rw_state, state,
3111 			    &oldstate, state | URWLOCK_READ_WAITERS);
3112 			if (rv == -1) {
3113 				error = EFAULT;
3114 				break;
3115 			}
3116 			if (rv == 0) {
3117 				MPASS(oldstate == state);
3118 				goto sleep;
3119 			}
3120 			state = oldstate;
3121 			error = thread_check_susp(td, false);
3122 			if (error != 0)
3123 				break;
3124 		}
3125 		if (error != 0) {
3126 			umtxq_unbusy_unlocked(&uq->uq_key);
3127 			break;
3128 		}
3129 
3130 		/* state is changed while setting flags, restart */
3131 		if (!(state & wrflags)) {
3132 			umtxq_unbusy_unlocked(&uq->uq_key);
3133 			error = thread_check_susp(td, true);
3134 			if (error != 0)
3135 				break;
3136 			continue;
3137 		}
3138 
3139 sleep:
3140 		/*
3141 		 * Contention bit is set, before sleeping, increase
3142 		 * read waiter count.
3143 		 */
3144 		rv = fueword32(&rwlock->rw_blocked_readers,
3145 		    &blocked_readers);
3146 		if (rv == -1) {
3147 			umtxq_unbusy_unlocked(&uq->uq_key);
3148 			error = EFAULT;
3149 			break;
3150 		}
3151 		suword32(&rwlock->rw_blocked_readers, blocked_readers+1);
3152 
3153 		while (state & wrflags) {
3154 			umtxq_lock(&uq->uq_key);
3155 			umtxq_insert(uq);
3156 			umtxq_unbusy(&uq->uq_key);
3157 
3158 			error = umtxq_sleep(uq, "urdlck", timeout == NULL ?
3159 			    NULL : &timo);
3160 
3161 			umtxq_busy(&uq->uq_key);
3162 			umtxq_remove(uq);
3163 			umtxq_unlock(&uq->uq_key);
3164 			if (error)
3165 				break;
3166 			rv = fueword32(&rwlock->rw_state, &state);
3167 			if (rv == -1) {
3168 				error = EFAULT;
3169 				break;
3170 			}
3171 		}
3172 
3173 		/* decrease read waiter count, and may clear read contention bit */
3174 		rv = fueword32(&rwlock->rw_blocked_readers,
3175 		    &blocked_readers);
3176 		if (rv == -1) {
3177 			umtxq_unbusy_unlocked(&uq->uq_key);
3178 			error = EFAULT;
3179 			break;
3180 		}
3181 		suword32(&rwlock->rw_blocked_readers, blocked_readers-1);
3182 		if (blocked_readers == 1) {
3183 			rv = fueword32(&rwlock->rw_state, &state);
3184 			if (rv == -1) {
3185 				umtxq_unbusy_unlocked(&uq->uq_key);
3186 				error = EFAULT;
3187 				break;
3188 			}
3189 			for (;;) {
3190 				rv = casueword32(&rwlock->rw_state, state,
3191 				    &oldstate, state & ~URWLOCK_READ_WAITERS);
3192 				if (rv == -1) {
3193 					error = EFAULT;
3194 					break;
3195 				}
3196 				if (rv == 0) {
3197 					MPASS(oldstate == state);
3198 					break;
3199 				}
3200 				state = oldstate;
3201 				error1 = thread_check_susp(td, false);
3202 				if (error1 != 0) {
3203 					if (error == 0)
3204 						error = error1;
3205 					break;
3206 				}
3207 			}
3208 		}
3209 
3210 		umtxq_unbusy_unlocked(&uq->uq_key);
3211 		if (error != 0)
3212 			break;
3213 	}
3214 	umtx_key_release(&uq->uq_key);
3215 	if (error == ERESTART)
3216 		error = EINTR;
3217 	return (error);
3218 }
3219 
3220 static int
3221 do_rw_wrlock(struct thread *td, struct urwlock *rwlock, struct _umtx_time *timeout)
3222 {
3223 	struct umtx_abs_timeout timo;
3224 	struct umtx_q *uq;
3225 	uint32_t flags;
3226 	int32_t state, oldstate;
3227 	int32_t blocked_writers;
3228 	int32_t blocked_readers;
3229 	int error, error1, rv;
3230 
3231 	uq = td->td_umtxq;
3232 	error = fueword32(&rwlock->rw_flags, &flags);
3233 	if (error == -1)
3234 		return (EFAULT);
3235 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3236 	if (error != 0)
3237 		return (error);
3238 
3239 	if (timeout != NULL)
3240 		umtx_abs_timeout_init2(&timo, timeout);
3241 
3242 	blocked_readers = 0;
3243 	for (;;) {
3244 		rv = fueword32(&rwlock->rw_state, &state);
3245 		if (rv == -1) {
3246 			umtx_key_release(&uq->uq_key);
3247 			return (EFAULT);
3248 		}
3249 		while ((state & URWLOCK_WRITE_OWNER) == 0 &&
3250 		    URWLOCK_READER_COUNT(state) == 0) {
3251 			rv = casueword32(&rwlock->rw_state, state,
3252 			    &oldstate, state | URWLOCK_WRITE_OWNER);
3253 			if (rv == -1) {
3254 				umtx_key_release(&uq->uq_key);
3255 				return (EFAULT);
3256 			}
3257 			if (rv == 0) {
3258 				MPASS(oldstate == state);
3259 				umtx_key_release(&uq->uq_key);
3260 				return (0);
3261 			}
3262 			state = oldstate;
3263 			error = thread_check_susp(td, true);
3264 			if (error != 0)
3265 				break;
3266 		}
3267 
3268 		if (error) {
3269 			if ((state & (URWLOCK_WRITE_OWNER |
3270 			    URWLOCK_WRITE_WAITERS)) == 0 &&
3271 			    blocked_readers != 0) {
3272 				umtxq_lock(&uq->uq_key);
3273 				umtxq_busy(&uq->uq_key);
3274 				umtxq_signal_queue(&uq->uq_key, INT_MAX,
3275 				    UMTX_SHARED_QUEUE);
3276 				umtxq_unbusy(&uq->uq_key);
3277 				umtxq_unlock(&uq->uq_key);
3278 			}
3279 
3280 			break;
3281 		}
3282 
3283 		/* grab monitor lock */
3284 		umtxq_lock(&uq->uq_key);
3285 		umtxq_busy(&uq->uq_key);
3286 		umtxq_unlock(&uq->uq_key);
3287 
3288 		/*
3289 		 * Re-read the state, in case it changed between the
3290 		 * try-lock above and the check below.
3291 		 */
3292 		rv = fueword32(&rwlock->rw_state, &state);
3293 		if (rv == -1)
3294 			error = EFAULT;
3295 
3296 		while (error == 0 && ((state & URWLOCK_WRITE_OWNER) ||
3297 		    URWLOCK_READER_COUNT(state) != 0) &&
3298 		    (state & URWLOCK_WRITE_WAITERS) == 0) {
3299 			rv = casueword32(&rwlock->rw_state, state,
3300 			    &oldstate, state | URWLOCK_WRITE_WAITERS);
3301 			if (rv == -1) {
3302 				error = EFAULT;
3303 				break;
3304 			}
3305 			if (rv == 0) {
3306 				MPASS(oldstate == state);
3307 				goto sleep;
3308 			}
3309 			state = oldstate;
3310 			error = thread_check_susp(td, false);
3311 			if (error != 0)
3312 				break;
3313 		}
3314 		if (error != 0) {
3315 			umtxq_unbusy_unlocked(&uq->uq_key);
3316 			break;
3317 		}
3318 
3319 		if ((state & URWLOCK_WRITE_OWNER) == 0 &&
3320 		    URWLOCK_READER_COUNT(state) == 0) {
3321 			umtxq_unbusy_unlocked(&uq->uq_key);
3322 			error = thread_check_susp(td, false);
3323 			if (error != 0)
3324 				break;
3325 			continue;
3326 		}
3327 sleep:
3328 		rv = fueword32(&rwlock->rw_blocked_writers,
3329 		    &blocked_writers);
3330 		if (rv == -1) {
3331 			umtxq_unbusy_unlocked(&uq->uq_key);
3332 			error = EFAULT;
3333 			break;
3334 		}
3335 		suword32(&rwlock->rw_blocked_writers, blocked_writers + 1);
3336 
3337 		while ((state & URWLOCK_WRITE_OWNER) ||
3338 		    URWLOCK_READER_COUNT(state) != 0) {
3339 			umtxq_lock(&uq->uq_key);
3340 			umtxq_insert_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3341 			umtxq_unbusy(&uq->uq_key);
3342 
3343 			error = umtxq_sleep(uq, "uwrlck", timeout == NULL ?
3344 			    NULL : &timo);
3345 
3346 			umtxq_busy(&uq->uq_key);
3347 			umtxq_remove_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3348 			umtxq_unlock(&uq->uq_key);
3349 			if (error)
3350 				break;
3351 			rv = fueword32(&rwlock->rw_state, &state);
3352 			if (rv == -1) {
3353 				error = EFAULT;
3354 				break;
3355 			}
3356 		}
3357 
3358 		rv = fueword32(&rwlock->rw_blocked_writers,
3359 		    &blocked_writers);
3360 		if (rv == -1) {
3361 			umtxq_unbusy_unlocked(&uq->uq_key);
3362 			error = EFAULT;
3363 			break;
3364 		}
3365 		suword32(&rwlock->rw_blocked_writers, blocked_writers-1);
3366 		if (blocked_writers == 1) {
3367 			rv = fueword32(&rwlock->rw_state, &state);
3368 			if (rv == -1) {
3369 				umtxq_unbusy_unlocked(&uq->uq_key);
3370 				error = EFAULT;
3371 				break;
3372 			}
3373 			for (;;) {
3374 				rv = casueword32(&rwlock->rw_state, state,
3375 				    &oldstate, state & ~URWLOCK_WRITE_WAITERS);
3376 				if (rv == -1) {
3377 					error = EFAULT;
3378 					break;
3379 				}
3380 				if (rv == 0) {
3381 					MPASS(oldstate == state);
3382 					break;
3383 				}
3384 				state = oldstate;
3385 				error1 = thread_check_susp(td, false);
3386 				/*
3387 				 * We are leaving the URWLOCK_WRITE_WAITERS
3388 				 * behind, but this should not harm the
3389 				 * correctness.
3390 				 */
3391 				if (error1 != 0) {
3392 					if (error == 0)
3393 						error = error1;
3394 					break;
3395 				}
3396 			}
3397 			rv = fueword32(&rwlock->rw_blocked_readers,
3398 			    &blocked_readers);
3399 			if (rv == -1) {
3400 				umtxq_unbusy_unlocked(&uq->uq_key);
3401 				error = EFAULT;
3402 				break;
3403 			}
3404 		} else
3405 			blocked_readers = 0;
3406 
3407 		umtxq_unbusy_unlocked(&uq->uq_key);
3408 	}
3409 
3410 	umtx_key_release(&uq->uq_key);
3411 	if (error == ERESTART)
3412 		error = EINTR;
3413 	return (error);
3414 }
3415 
3416 static int
3417 do_rw_unlock(struct thread *td, struct urwlock *rwlock)
3418 {
3419 	struct umtx_q *uq;
3420 	uint32_t flags;
3421 	int32_t state, oldstate;
3422 	int error, rv, q, count;
3423 
3424 	uq = td->td_umtxq;
3425 	error = fueword32(&rwlock->rw_flags, &flags);
3426 	if (error == -1)
3427 		return (EFAULT);
3428 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3429 	if (error != 0)
3430 		return (error);
3431 
3432 	error = fueword32(&rwlock->rw_state, &state);
3433 	if (error == -1) {
3434 		error = EFAULT;
3435 		goto out;
3436 	}
3437 	if (state & URWLOCK_WRITE_OWNER) {
3438 		for (;;) {
3439 			rv = casueword32(&rwlock->rw_state, state,
3440 			    &oldstate, state & ~URWLOCK_WRITE_OWNER);
3441 			if (rv == -1) {
3442 				error = EFAULT;
3443 				goto out;
3444 			}
3445 			if (rv == 1) {
3446 				state = oldstate;
3447 				if (!(oldstate & URWLOCK_WRITE_OWNER)) {
3448 					error = EPERM;
3449 					goto out;
3450 				}
3451 				error = thread_check_susp(td, true);
3452 				if (error != 0)
3453 					goto out;
3454 			} else
3455 				break;
3456 		}
3457 	} else if (URWLOCK_READER_COUNT(state) != 0) {
3458 		for (;;) {
3459 			rv = casueword32(&rwlock->rw_state, state,
3460 			    &oldstate, state - 1);
3461 			if (rv == -1) {
3462 				error = EFAULT;
3463 				goto out;
3464 			}
3465 			if (rv == 1) {
3466 				state = oldstate;
3467 				if (URWLOCK_READER_COUNT(oldstate) == 0) {
3468 					error = EPERM;
3469 					goto out;
3470 				}
3471 				error = thread_check_susp(td, true);
3472 				if (error != 0)
3473 					goto out;
3474 			} else
3475 				break;
3476 		}
3477 	} else {
3478 		error = EPERM;
3479 		goto out;
3480 	}
3481 
3482 	count = 0;
3483 
3484 	if (!(flags & URWLOCK_PREFER_READER)) {
3485 		if (state & URWLOCK_WRITE_WAITERS) {
3486 			count = 1;
3487 			q = UMTX_EXCLUSIVE_QUEUE;
3488 		} else if (state & URWLOCK_READ_WAITERS) {
3489 			count = INT_MAX;
3490 			q = UMTX_SHARED_QUEUE;
3491 		}
3492 	} else {
3493 		if (state & URWLOCK_READ_WAITERS) {
3494 			count = INT_MAX;
3495 			q = UMTX_SHARED_QUEUE;
3496 		} else if (state & URWLOCK_WRITE_WAITERS) {
3497 			count = 1;
3498 			q = UMTX_EXCLUSIVE_QUEUE;
3499 		}
3500 	}
3501 
3502 	if (count) {
3503 		umtxq_lock(&uq->uq_key);
3504 		umtxq_busy(&uq->uq_key);
3505 		umtxq_signal_queue(&uq->uq_key, count, q);
3506 		umtxq_unbusy(&uq->uq_key);
3507 		umtxq_unlock(&uq->uq_key);
3508 	}
3509 out:
3510 	umtx_key_release(&uq->uq_key);
3511 	return (error);
3512 }
3513 
3514 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3515 static int
3516 do_sem_wait(struct thread *td, struct _usem *sem, struct _umtx_time *timeout)
3517 {
3518 	struct umtx_abs_timeout timo;
3519 	struct umtx_q *uq;
3520 	uint32_t flags, count, count1;
3521 	int error, rv, rv1;
3522 
3523 	uq = td->td_umtxq;
3524 	error = fueword32(&sem->_flags, &flags);
3525 	if (error == -1)
3526 		return (EFAULT);
3527 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3528 	if (error != 0)
3529 		return (error);
3530 
3531 	if (timeout != NULL)
3532 		umtx_abs_timeout_init2(&timo, timeout);
3533 
3534 again:
3535 	umtxq_lock(&uq->uq_key);
3536 	umtxq_busy(&uq->uq_key);
3537 	umtxq_insert(uq);
3538 	umtxq_unlock(&uq->uq_key);
3539 	rv = casueword32(&sem->_has_waiters, 0, &count1, 1);
3540 	if (rv == 0)
3541 		rv1 = fueword32(&sem->_count, &count);
3542 	if (rv == -1 || (rv == 0 && (rv1 == -1 || count != 0)) ||
3543 	    (rv == 1 && count1 == 0)) {
3544 		umtxq_lock(&uq->uq_key);
3545 		umtxq_unbusy(&uq->uq_key);
3546 		umtxq_remove(uq);
3547 		umtxq_unlock(&uq->uq_key);
3548 		if (rv == 1) {
3549 			rv = thread_check_susp(td, true);
3550 			if (rv == 0)
3551 				goto again;
3552 			error = rv;
3553 			goto out;
3554 		}
3555 		if (rv == 0)
3556 			rv = rv1;
3557 		error = rv == -1 ? EFAULT : 0;
3558 		goto out;
3559 	}
3560 	umtxq_lock(&uq->uq_key);
3561 	umtxq_unbusy(&uq->uq_key);
3562 
3563 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3564 
3565 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3566 		error = 0;
3567 	else {
3568 		umtxq_remove(uq);
3569 		/* A relative timeout cannot be restarted. */
3570 		if (error == ERESTART && timeout != NULL &&
3571 		    (timeout->_flags & UMTX_ABSTIME) == 0)
3572 			error = EINTR;
3573 	}
3574 	umtxq_unlock(&uq->uq_key);
3575 out:
3576 	umtx_key_release(&uq->uq_key);
3577 	return (error);
3578 }
3579 
3580 /*
3581  * Signal a userland semaphore.
3582  */
3583 static int
3584 do_sem_wake(struct thread *td, struct _usem *sem)
3585 {
3586 	struct umtx_key key;
3587 	int error, cnt;
3588 	uint32_t flags;
3589 
3590 	error = fueword32(&sem->_flags, &flags);
3591 	if (error == -1)
3592 		return (EFAULT);
3593 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3594 		return (error);
3595 	umtxq_lock(&key);
3596 	umtxq_busy(&key);
3597 	cnt = umtxq_count(&key);
3598 	if (cnt > 0) {
3599 		/*
3600 		 * Check if count is greater than 0, this means the memory is
3601 		 * still being referenced by user code, so we can safely
3602 		 * update _has_waiters flag.
3603 		 */
3604 		if (cnt == 1) {
3605 			umtxq_unlock(&key);
3606 			error = suword32(&sem->_has_waiters, 0);
3607 			umtxq_lock(&key);
3608 			if (error == -1)
3609 				error = EFAULT;
3610 		}
3611 		umtxq_signal(&key, 1);
3612 	}
3613 	umtxq_unbusy(&key);
3614 	umtxq_unlock(&key);
3615 	umtx_key_release(&key);
3616 	return (error);
3617 }
3618 #endif
3619 
3620 static int
3621 do_sem2_wait(struct thread *td, struct _usem2 *sem, struct _umtx_time *timeout)
3622 {
3623 	struct umtx_abs_timeout timo;
3624 	struct umtx_q *uq;
3625 	uint32_t count, flags;
3626 	int error, rv;
3627 
3628 	uq = td->td_umtxq;
3629 	flags = fuword32(&sem->_flags);
3630 	if (timeout != NULL)
3631 		umtx_abs_timeout_init2(&timo, timeout);
3632 
3633 again:
3634 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3635 	if (error != 0)
3636 		return (error);
3637 	umtxq_lock(&uq->uq_key);
3638 	umtxq_busy(&uq->uq_key);
3639 	umtxq_insert(uq);
3640 	umtxq_unlock(&uq->uq_key);
3641 	rv = fueword32(&sem->_count, &count);
3642 	if (rv == -1) {
3643 		umtxq_lock(&uq->uq_key);
3644 		umtxq_unbusy(&uq->uq_key);
3645 		umtxq_remove(uq);
3646 		umtxq_unlock(&uq->uq_key);
3647 		umtx_key_release(&uq->uq_key);
3648 		return (EFAULT);
3649 	}
3650 	for (;;) {
3651 		if (USEM_COUNT(count) != 0) {
3652 			umtxq_lock(&uq->uq_key);
3653 			umtxq_unbusy(&uq->uq_key);
3654 			umtxq_remove(uq);
3655 			umtxq_unlock(&uq->uq_key);
3656 			umtx_key_release(&uq->uq_key);
3657 			return (0);
3658 		}
3659 		if (count == USEM_HAS_WAITERS)
3660 			break;
3661 		rv = casueword32(&sem->_count, 0, &count, USEM_HAS_WAITERS);
3662 		if (rv == 0)
3663 			break;
3664 		umtxq_lock(&uq->uq_key);
3665 		umtxq_unbusy(&uq->uq_key);
3666 		umtxq_remove(uq);
3667 		umtxq_unlock(&uq->uq_key);
3668 		umtx_key_release(&uq->uq_key);
3669 		if (rv == -1)
3670 			return (EFAULT);
3671 		rv = thread_check_susp(td, true);
3672 		if (rv != 0)
3673 			return (rv);
3674 		goto again;
3675 	}
3676 	umtxq_lock(&uq->uq_key);
3677 	umtxq_unbusy(&uq->uq_key);
3678 
3679 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3680 
3681 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3682 		error = 0;
3683 	else {
3684 		umtxq_remove(uq);
3685 		if (timeout != NULL && (timeout->_flags & UMTX_ABSTIME) == 0) {
3686 			/* A relative timeout cannot be restarted. */
3687 			if (error == ERESTART)
3688 				error = EINTR;
3689 			if (error == EINTR) {
3690 				kern_clock_gettime(curthread, timo.clockid,
3691 				    &timo.cur);
3692 				timespecsub(&timo.end, &timo.cur,
3693 				    &timeout->_timeout);
3694 			}
3695 		}
3696 	}
3697 	umtxq_unlock(&uq->uq_key);
3698 	umtx_key_release(&uq->uq_key);
3699 	return (error);
3700 }
3701 
3702 /*
3703  * Signal a userland semaphore.
3704  */
3705 static int
3706 do_sem2_wake(struct thread *td, struct _usem2 *sem)
3707 {
3708 	struct umtx_key key;
3709 	int error, cnt, rv;
3710 	uint32_t count, flags;
3711 
3712 	rv = fueword32(&sem->_flags, &flags);
3713 	if (rv == -1)
3714 		return (EFAULT);
3715 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3716 		return (error);
3717 	umtxq_lock(&key);
3718 	umtxq_busy(&key);
3719 	cnt = umtxq_count(&key);
3720 	if (cnt > 0) {
3721 		/*
3722 		 * If this was the last sleeping thread, clear the waiters
3723 		 * flag in _count.
3724 		 */
3725 		if (cnt == 1) {
3726 			umtxq_unlock(&key);
3727 			rv = fueword32(&sem->_count, &count);
3728 			while (rv != -1 && count & USEM_HAS_WAITERS) {
3729 				rv = casueword32(&sem->_count, count, &count,
3730 				    count & ~USEM_HAS_WAITERS);
3731 				if (rv == 1) {
3732 					rv = thread_check_susp(td, true);
3733 					if (rv != 0)
3734 						break;
3735 				}
3736 			}
3737 			if (rv == -1)
3738 				error = EFAULT;
3739 			else if (rv > 0) {
3740 				error = rv;
3741 			}
3742 			umtxq_lock(&key);
3743 		}
3744 
3745 		umtxq_signal(&key, 1);
3746 	}
3747 	umtxq_unbusy(&key);
3748 	umtxq_unlock(&key);
3749 	umtx_key_release(&key);
3750 	return (error);
3751 }
3752 
3753 #ifdef COMPAT_FREEBSD10
3754 int
3755 freebsd10__umtx_lock(struct thread *td, struct freebsd10__umtx_lock_args *uap)
3756 {
3757 	return (do_lock_umtx(td, uap->umtx, td->td_tid, 0));
3758 }
3759 
3760 int
3761 freebsd10__umtx_unlock(struct thread *td,
3762     struct freebsd10__umtx_unlock_args *uap)
3763 {
3764 	return (do_unlock_umtx(td, uap->umtx, td->td_tid));
3765 }
3766 #endif
3767 
3768 inline int
3769 umtx_copyin_timeout(const void *uaddr, struct timespec *tsp)
3770 {
3771 	int error;
3772 
3773 	error = copyin(uaddr, tsp, sizeof(*tsp));
3774 	if (error == 0) {
3775 		if (tsp->tv_sec < 0 ||
3776 		    tsp->tv_nsec >= 1000000000 ||
3777 		    tsp->tv_nsec < 0)
3778 			error = EINVAL;
3779 	}
3780 	return (error);
3781 }
3782 
3783 static inline int
3784 umtx_copyin_umtx_time(const void *uaddr, size_t size, struct _umtx_time *tp)
3785 {
3786 	int error;
3787 
3788 	if (size <= sizeof(tp->_timeout)) {
3789 		tp->_clockid = CLOCK_REALTIME;
3790 		tp->_flags = 0;
3791 		error = copyin(uaddr, &tp->_timeout, sizeof(tp->_timeout));
3792 	} else
3793 		error = copyin(uaddr, tp, sizeof(*tp));
3794 	if (error != 0)
3795 		return (error);
3796 	if (tp->_timeout.tv_sec < 0 ||
3797 	    tp->_timeout.tv_nsec >= 1000000000 || tp->_timeout.tv_nsec < 0)
3798 		return (EINVAL);
3799 	return (0);
3800 }
3801 
3802 static int
3803 umtx_copyin_robust_lists(const void *uaddr, size_t size,
3804     struct umtx_robust_lists_params *rb)
3805 {
3806 
3807 	if (size > sizeof(*rb))
3808 		return (EINVAL);
3809 	return (copyin(uaddr, rb, size));
3810 }
3811 
3812 static int
3813 umtx_copyout_timeout(void *uaddr, size_t sz, struct timespec *tsp)
3814 {
3815 
3816 	/*
3817 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
3818 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
3819 	 * copyops.
3820 	 */
3821 	KASSERT(sz >= sizeof(*tsp),
3822 	    ("umtx_copyops specifies incorrect sizes"));
3823 
3824 	return (copyout(tsp, uaddr, sizeof(*tsp)));
3825 }
3826 
3827 #ifdef COMPAT_FREEBSD10
3828 static int
3829 __umtx_op_lock_umtx(struct thread *td, struct _umtx_op_args *uap,
3830     const struct umtx_copyops *ops)
3831 {
3832 	struct timespec *ts, timeout;
3833 	int error;
3834 
3835 	/* Allow a null timespec (wait forever). */
3836 	if (uap->uaddr2 == NULL)
3837 		ts = NULL;
3838 	else {
3839 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
3840 		if (error != 0)
3841 			return (error);
3842 		ts = &timeout;
3843 	}
3844 #ifdef COMPAT_FREEBSD32
3845 	if (ops->compat32)
3846 		return (do_lock_umtx32(td, uap->obj, uap->val, ts));
3847 #endif
3848 	return (do_lock_umtx(td, uap->obj, uap->val, ts));
3849 }
3850 
3851 static int
3852 __umtx_op_unlock_umtx(struct thread *td, struct _umtx_op_args *uap,
3853     const struct umtx_copyops *ops)
3854 {
3855 #ifdef COMPAT_FREEBSD32
3856 	if (ops->compat32)
3857 		return (do_unlock_umtx32(td, uap->obj, uap->val));
3858 #endif
3859 	return (do_unlock_umtx(td, uap->obj, uap->val));
3860 }
3861 #endif	/* COMPAT_FREEBSD10 */
3862 
3863 #if !defined(COMPAT_FREEBSD10)
3864 static int
3865 __umtx_op_unimpl(struct thread *td __unused, struct _umtx_op_args *uap __unused,
3866     const struct umtx_copyops *ops __unused)
3867 {
3868 	return (EOPNOTSUPP);
3869 }
3870 #endif	/* COMPAT_FREEBSD10 */
3871 
3872 static int
3873 __umtx_op_wait(struct thread *td, struct _umtx_op_args *uap,
3874     const struct umtx_copyops *ops)
3875 {
3876 	struct _umtx_time timeout, *tm_p;
3877 	int error;
3878 
3879 	if (uap->uaddr2 == NULL)
3880 		tm_p = NULL;
3881 	else {
3882 		error = ops->copyin_umtx_time(
3883 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3884 		if (error != 0)
3885 			return (error);
3886 		tm_p = &timeout;
3887 	}
3888 	return (do_wait(td, uap->obj, uap->val, tm_p, ops->compat32, 0));
3889 }
3890 
3891 static int
3892 __umtx_op_wait_uint(struct thread *td, struct _umtx_op_args *uap,
3893     const struct umtx_copyops *ops)
3894 {
3895 	struct _umtx_time timeout, *tm_p;
3896 	int error;
3897 
3898 	if (uap->uaddr2 == NULL)
3899 		tm_p = NULL;
3900 	else {
3901 		error = ops->copyin_umtx_time(
3902 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3903 		if (error != 0)
3904 			return (error);
3905 		tm_p = &timeout;
3906 	}
3907 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
3908 }
3909 
3910 static int
3911 __umtx_op_wait_uint_private(struct thread *td, struct _umtx_op_args *uap,
3912     const struct umtx_copyops *ops)
3913 {
3914 	struct _umtx_time *tm_p, timeout;
3915 	int error;
3916 
3917 	if (uap->uaddr2 == NULL)
3918 		tm_p = NULL;
3919 	else {
3920 		error = ops->copyin_umtx_time(
3921 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3922 		if (error != 0)
3923 			return (error);
3924 		tm_p = &timeout;
3925 	}
3926 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
3927 }
3928 
3929 static int
3930 __umtx_op_wake(struct thread *td, struct _umtx_op_args *uap,
3931     const struct umtx_copyops *ops __unused)
3932 {
3933 
3934 	return (kern_umtx_wake(td, uap->obj, uap->val, 0));
3935 }
3936 
3937 #define BATCH_SIZE	128
3938 static int
3939 __umtx_op_nwake_private_native(struct thread *td, struct _umtx_op_args *uap)
3940 {
3941 	char *uaddrs[BATCH_SIZE], **upp;
3942 	int count, error, i, pos, tocopy;
3943 
3944 	upp = (char **)uap->obj;
3945 	error = 0;
3946 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3947 	    pos += tocopy) {
3948 		tocopy = MIN(count, BATCH_SIZE);
3949 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(char *));
3950 		if (error != 0)
3951 			break;
3952 		for (i = 0; i < tocopy; ++i) {
3953 			kern_umtx_wake(td, uaddrs[i], INT_MAX, 1);
3954 		}
3955 		maybe_yield();
3956 	}
3957 	return (error);
3958 }
3959 
3960 static int
3961 __umtx_op_nwake_private_compat32(struct thread *td, struct _umtx_op_args *uap)
3962 {
3963 	uint32_t uaddrs[BATCH_SIZE], *upp;
3964 	int count, error, i, pos, tocopy;
3965 
3966 	upp = (uint32_t *)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(uint32_t));
3972 		if (error != 0)
3973 			break;
3974 		for (i = 0; i < tocopy; ++i) {
3975 			kern_umtx_wake(td, (void *)(uintptr_t)uaddrs[i],
3976 			    INT_MAX, 1);
3977 		}
3978 		maybe_yield();
3979 	}
3980 	return (error);
3981 }
3982 
3983 static int
3984 __umtx_op_nwake_private(struct thread *td, struct _umtx_op_args *uap,
3985     const struct umtx_copyops *ops)
3986 {
3987 
3988 	if (ops->compat32)
3989 		return (__umtx_op_nwake_private_compat32(td, uap));
3990 	return (__umtx_op_nwake_private_native(td, uap));
3991 }
3992 
3993 static int
3994 __umtx_op_wake_private(struct thread *td, struct _umtx_op_args *uap,
3995     const struct umtx_copyops *ops __unused)
3996 {
3997 
3998 	return (kern_umtx_wake(td, uap->obj, uap->val, 1));
3999 }
4000 
4001 static int
4002 __umtx_op_lock_umutex(struct thread *td, struct _umtx_op_args *uap,
4003    const struct umtx_copyops *ops)
4004 {
4005 	struct _umtx_time *tm_p, timeout;
4006 	int error;
4007 
4008 	/* Allow a null timespec (wait forever). */
4009 	if (uap->uaddr2 == NULL)
4010 		tm_p = NULL;
4011 	else {
4012 		error = ops->copyin_umtx_time(
4013 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4014 		if (error != 0)
4015 			return (error);
4016 		tm_p = &timeout;
4017 	}
4018 	return (do_lock_umutex(td, uap->obj, tm_p, 0));
4019 }
4020 
4021 static int
4022 __umtx_op_trylock_umutex(struct thread *td, struct _umtx_op_args *uap,
4023     const struct umtx_copyops *ops __unused)
4024 {
4025 
4026 	return (do_lock_umutex(td, uap->obj, NULL, _UMUTEX_TRY));
4027 }
4028 
4029 static int
4030 __umtx_op_wait_umutex(struct thread *td, struct _umtx_op_args *uap,
4031     const struct umtx_copyops *ops)
4032 {
4033 	struct _umtx_time *tm_p, timeout;
4034 	int error;
4035 
4036 	/* Allow a null timespec (wait forever). */
4037 	if (uap->uaddr2 == NULL)
4038 		tm_p = NULL;
4039 	else {
4040 		error = ops->copyin_umtx_time(
4041 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4042 		if (error != 0)
4043 			return (error);
4044 		tm_p = &timeout;
4045 	}
4046 	return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
4047 }
4048 
4049 static int
4050 __umtx_op_wake_umutex(struct thread *td, struct _umtx_op_args *uap,
4051     const struct umtx_copyops *ops __unused)
4052 {
4053 
4054 	return (do_wake_umutex(td, uap->obj));
4055 }
4056 
4057 static int
4058 __umtx_op_unlock_umutex(struct thread *td, struct _umtx_op_args *uap,
4059     const struct umtx_copyops *ops __unused)
4060 {
4061 
4062 	return (do_unlock_umutex(td, uap->obj, false));
4063 }
4064 
4065 static int
4066 __umtx_op_set_ceiling(struct thread *td, struct _umtx_op_args *uap,
4067     const struct umtx_copyops *ops __unused)
4068 {
4069 
4070 	return (do_set_ceiling(td, uap->obj, uap->val, uap->uaddr1));
4071 }
4072 
4073 static int
4074 __umtx_op_cv_wait(struct thread *td, struct _umtx_op_args *uap,
4075     const struct umtx_copyops *ops)
4076 {
4077 	struct timespec *ts, timeout;
4078 	int error;
4079 
4080 	/* Allow a null timespec (wait forever). */
4081 	if (uap->uaddr2 == NULL)
4082 		ts = NULL;
4083 	else {
4084 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
4085 		if (error != 0)
4086 			return (error);
4087 		ts = &timeout;
4088 	}
4089 	return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
4090 }
4091 
4092 static int
4093 __umtx_op_cv_signal(struct thread *td, struct _umtx_op_args *uap,
4094     const struct umtx_copyops *ops __unused)
4095 {
4096 
4097 	return (do_cv_signal(td, uap->obj));
4098 }
4099 
4100 static int
4101 __umtx_op_cv_broadcast(struct thread *td, struct _umtx_op_args *uap,
4102     const struct umtx_copyops *ops __unused)
4103 {
4104 
4105 	return (do_cv_broadcast(td, uap->obj));
4106 }
4107 
4108 static int
4109 __umtx_op_rw_rdlock(struct thread *td, struct _umtx_op_args *uap,
4110     const struct umtx_copyops *ops)
4111 {
4112 	struct _umtx_time timeout;
4113 	int error;
4114 
4115 	/* Allow a null timespec (wait forever). */
4116 	if (uap->uaddr2 == NULL) {
4117 		error = do_rw_rdlock(td, uap->obj, uap->val, 0);
4118 	} else {
4119 		error = ops->copyin_umtx_time(uap->uaddr2,
4120 		   (size_t)uap->uaddr1, &timeout);
4121 		if (error != 0)
4122 			return (error);
4123 		error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
4124 	}
4125 	return (error);
4126 }
4127 
4128 static int
4129 __umtx_op_rw_wrlock(struct thread *td, struct _umtx_op_args *uap,
4130     const struct umtx_copyops *ops)
4131 {
4132 	struct _umtx_time timeout;
4133 	int error;
4134 
4135 	/* Allow a null timespec (wait forever). */
4136 	if (uap->uaddr2 == NULL) {
4137 		error = do_rw_wrlock(td, uap->obj, 0);
4138 	} else {
4139 		error = ops->copyin_umtx_time(uap->uaddr2,
4140 		   (size_t)uap->uaddr1, &timeout);
4141 		if (error != 0)
4142 			return (error);
4143 
4144 		error = do_rw_wrlock(td, uap->obj, &timeout);
4145 	}
4146 	return (error);
4147 }
4148 
4149 static int
4150 __umtx_op_rw_unlock(struct thread *td, struct _umtx_op_args *uap,
4151     const struct umtx_copyops *ops __unused)
4152 {
4153 
4154 	return (do_rw_unlock(td, uap->obj));
4155 }
4156 
4157 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4158 static int
4159 __umtx_op_sem_wait(struct thread *td, struct _umtx_op_args *uap,
4160     const struct umtx_copyops *ops)
4161 {
4162 	struct _umtx_time *tm_p, timeout;
4163 	int error;
4164 
4165 	/* Allow a null timespec (wait forever). */
4166 	if (uap->uaddr2 == NULL)
4167 		tm_p = NULL;
4168 	else {
4169 		error = ops->copyin_umtx_time(
4170 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4171 		if (error != 0)
4172 			return (error);
4173 		tm_p = &timeout;
4174 	}
4175 	return (do_sem_wait(td, uap->obj, tm_p));
4176 }
4177 
4178 static int
4179 __umtx_op_sem_wake(struct thread *td, struct _umtx_op_args *uap,
4180     const struct umtx_copyops *ops __unused)
4181 {
4182 
4183 	return (do_sem_wake(td, uap->obj));
4184 }
4185 #endif
4186 
4187 static int
4188 __umtx_op_wake2_umutex(struct thread *td, struct _umtx_op_args *uap,
4189     const struct umtx_copyops *ops __unused)
4190 {
4191 
4192 	return (do_wake2_umutex(td, uap->obj, uap->val));
4193 }
4194 
4195 static int
4196 __umtx_op_sem2_wait(struct thread *td, struct _umtx_op_args *uap,
4197     const struct umtx_copyops *ops)
4198 {
4199 	struct _umtx_time *tm_p, timeout;
4200 	size_t uasize;
4201 	int error;
4202 
4203 	/* Allow a null timespec (wait forever). */
4204 	if (uap->uaddr2 == NULL) {
4205 		uasize = 0;
4206 		tm_p = NULL;
4207 	} else {
4208 		uasize = (size_t)uap->uaddr1;
4209 		error = ops->copyin_umtx_time(uap->uaddr2, uasize, &timeout);
4210 		if (error != 0)
4211 			return (error);
4212 		tm_p = &timeout;
4213 	}
4214 	error = do_sem2_wait(td, uap->obj, tm_p);
4215 	if (error == EINTR && uap->uaddr2 != NULL &&
4216 	    (timeout._flags & UMTX_ABSTIME) == 0 &&
4217 	    uasize >= ops->umtx_time_sz + ops->timespec_sz) {
4218 		error = ops->copyout_timeout(
4219 		    (void *)((uintptr_t)uap->uaddr2 + ops->umtx_time_sz),
4220 		    uasize - ops->umtx_time_sz, &timeout._timeout);
4221 		if (error == 0) {
4222 			error = EINTR;
4223 		}
4224 	}
4225 
4226 	return (error);
4227 }
4228 
4229 static int
4230 __umtx_op_sem2_wake(struct thread *td, struct _umtx_op_args *uap,
4231     const struct umtx_copyops *ops __unused)
4232 {
4233 
4234 	return (do_sem2_wake(td, uap->obj));
4235 }
4236 
4237 #define	USHM_OBJ_UMTX(o)						\
4238     ((struct umtx_shm_obj_list *)(&(o)->umtx_data))
4239 
4240 #define	USHMF_REG_LINKED	0x0001
4241 #define	USHMF_OBJ_LINKED	0x0002
4242 struct umtx_shm_reg {
4243 	TAILQ_ENTRY(umtx_shm_reg) ushm_reg_link;
4244 	LIST_ENTRY(umtx_shm_reg) ushm_obj_link;
4245 	struct umtx_key		ushm_key;
4246 	struct ucred		*ushm_cred;
4247 	struct shmfd		*ushm_obj;
4248 	u_int			ushm_refcnt;
4249 	u_int			ushm_flags;
4250 };
4251 
4252 LIST_HEAD(umtx_shm_obj_list, umtx_shm_reg);
4253 TAILQ_HEAD(umtx_shm_reg_head, umtx_shm_reg);
4254 
4255 static uma_zone_t umtx_shm_reg_zone;
4256 static struct umtx_shm_reg_head umtx_shm_registry[UMTX_CHAINS];
4257 static struct mtx umtx_shm_lock;
4258 static struct umtx_shm_reg_head umtx_shm_reg_delfree =
4259     TAILQ_HEAD_INITIALIZER(umtx_shm_reg_delfree);
4260 
4261 static void umtx_shm_free_reg(struct umtx_shm_reg *reg);
4262 
4263 static void
4264 umtx_shm_reg_delfree_tq(void *context __unused, int pending __unused)
4265 {
4266 	struct umtx_shm_reg_head d;
4267 	struct umtx_shm_reg *reg, *reg1;
4268 
4269 	TAILQ_INIT(&d);
4270 	mtx_lock(&umtx_shm_lock);
4271 	TAILQ_CONCAT(&d, &umtx_shm_reg_delfree, ushm_reg_link);
4272 	mtx_unlock(&umtx_shm_lock);
4273 	TAILQ_FOREACH_SAFE(reg, &d, ushm_reg_link, reg1) {
4274 		TAILQ_REMOVE(&d, reg, ushm_reg_link);
4275 		umtx_shm_free_reg(reg);
4276 	}
4277 }
4278 
4279 static struct task umtx_shm_reg_delfree_task =
4280     TASK_INITIALIZER(0, umtx_shm_reg_delfree_tq, NULL);
4281 
4282 static struct umtx_shm_reg *
4283 umtx_shm_find_reg_locked(const struct umtx_key *key)
4284 {
4285 	struct umtx_shm_reg *reg;
4286 	struct umtx_shm_reg_head *reg_head;
4287 
4288 	KASSERT(key->shared, ("umtx_p_find_rg: private key"));
4289 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4290 	reg_head = &umtx_shm_registry[key->hash];
4291 	TAILQ_FOREACH(reg, reg_head, ushm_reg_link) {
4292 		KASSERT(reg->ushm_key.shared,
4293 		    ("non-shared key on reg %p %d", reg, reg->ushm_key.shared));
4294 		if (reg->ushm_key.info.shared.object ==
4295 		    key->info.shared.object &&
4296 		    reg->ushm_key.info.shared.offset ==
4297 		    key->info.shared.offset) {
4298 			KASSERT(reg->ushm_key.type == TYPE_SHM, ("TYPE_USHM"));
4299 			KASSERT(reg->ushm_refcnt > 0,
4300 			    ("reg %p refcnt 0 onlist", reg));
4301 			KASSERT((reg->ushm_flags & USHMF_REG_LINKED) != 0,
4302 			    ("reg %p not linked", reg));
4303 			reg->ushm_refcnt++;
4304 			return (reg);
4305 		}
4306 	}
4307 	return (NULL);
4308 }
4309 
4310 static struct umtx_shm_reg *
4311 umtx_shm_find_reg(const struct umtx_key *key)
4312 {
4313 	struct umtx_shm_reg *reg;
4314 
4315 	mtx_lock(&umtx_shm_lock);
4316 	reg = umtx_shm_find_reg_locked(key);
4317 	mtx_unlock(&umtx_shm_lock);
4318 	return (reg);
4319 }
4320 
4321 static void
4322 umtx_shm_free_reg(struct umtx_shm_reg *reg)
4323 {
4324 
4325 	chgumtxcnt(reg->ushm_cred->cr_ruidinfo, -1, 0);
4326 	crfree(reg->ushm_cred);
4327 	shm_drop(reg->ushm_obj);
4328 	uma_zfree(umtx_shm_reg_zone, reg);
4329 }
4330 
4331 static bool
4332 umtx_shm_unref_reg_locked(struct umtx_shm_reg *reg, bool force)
4333 {
4334 	bool res;
4335 
4336 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4337 	KASSERT(reg->ushm_refcnt > 0, ("ushm_reg %p refcnt 0", reg));
4338 	reg->ushm_refcnt--;
4339 	res = reg->ushm_refcnt == 0;
4340 	if (res || force) {
4341 		if ((reg->ushm_flags & USHMF_REG_LINKED) != 0) {
4342 			TAILQ_REMOVE(&umtx_shm_registry[reg->ushm_key.hash],
4343 			    reg, ushm_reg_link);
4344 			reg->ushm_flags &= ~USHMF_REG_LINKED;
4345 		}
4346 		if ((reg->ushm_flags & USHMF_OBJ_LINKED) != 0) {
4347 			LIST_REMOVE(reg, ushm_obj_link);
4348 			reg->ushm_flags &= ~USHMF_OBJ_LINKED;
4349 		}
4350 	}
4351 	return (res);
4352 }
4353 
4354 static void
4355 umtx_shm_unref_reg(struct umtx_shm_reg *reg, bool force)
4356 {
4357 	vm_object_t object;
4358 	bool dofree;
4359 
4360 	if (force) {
4361 		object = reg->ushm_obj->shm_object;
4362 		VM_OBJECT_WLOCK(object);
4363 		object->flags |= OBJ_UMTXDEAD;
4364 		VM_OBJECT_WUNLOCK(object);
4365 	}
4366 	mtx_lock(&umtx_shm_lock);
4367 	dofree = umtx_shm_unref_reg_locked(reg, force);
4368 	mtx_unlock(&umtx_shm_lock);
4369 	if (dofree)
4370 		umtx_shm_free_reg(reg);
4371 }
4372 
4373 void
4374 umtx_shm_object_init(vm_object_t object)
4375 {
4376 
4377 	LIST_INIT(USHM_OBJ_UMTX(object));
4378 }
4379 
4380 void
4381 umtx_shm_object_terminated(vm_object_t object)
4382 {
4383 	struct umtx_shm_reg *reg, *reg1;
4384 	bool dofree;
4385 
4386 	if (LIST_EMPTY(USHM_OBJ_UMTX(object)))
4387 		return;
4388 
4389 	dofree = false;
4390 	mtx_lock(&umtx_shm_lock);
4391 	LIST_FOREACH_SAFE(reg, USHM_OBJ_UMTX(object), ushm_obj_link, reg1) {
4392 		if (umtx_shm_unref_reg_locked(reg, true)) {
4393 			TAILQ_INSERT_TAIL(&umtx_shm_reg_delfree, reg,
4394 			    ushm_reg_link);
4395 			dofree = true;
4396 		}
4397 	}
4398 	mtx_unlock(&umtx_shm_lock);
4399 	if (dofree)
4400 		taskqueue_enqueue(taskqueue_thread, &umtx_shm_reg_delfree_task);
4401 }
4402 
4403 static int
4404 umtx_shm_create_reg(struct thread *td, const struct umtx_key *key,
4405     struct umtx_shm_reg **res)
4406 {
4407 	struct umtx_shm_reg *reg, *reg1;
4408 	struct ucred *cred;
4409 	int error;
4410 
4411 	reg = umtx_shm_find_reg(key);
4412 	if (reg != NULL) {
4413 		*res = reg;
4414 		return (0);
4415 	}
4416 	cred = td->td_ucred;
4417 	if (!chgumtxcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_UMTXP)))
4418 		return (ENOMEM);
4419 	reg = uma_zalloc(umtx_shm_reg_zone, M_WAITOK | M_ZERO);
4420 	reg->ushm_refcnt = 1;
4421 	bcopy(key, &reg->ushm_key, sizeof(*key));
4422 	reg->ushm_obj = shm_alloc(td->td_ucred, O_RDWR, false);
4423 	reg->ushm_cred = crhold(cred);
4424 	error = shm_dotruncate(reg->ushm_obj, PAGE_SIZE);
4425 	if (error != 0) {
4426 		umtx_shm_free_reg(reg);
4427 		return (error);
4428 	}
4429 	mtx_lock(&umtx_shm_lock);
4430 	reg1 = umtx_shm_find_reg_locked(key);
4431 	if (reg1 != NULL) {
4432 		mtx_unlock(&umtx_shm_lock);
4433 		umtx_shm_free_reg(reg);
4434 		*res = reg1;
4435 		return (0);
4436 	}
4437 	reg->ushm_refcnt++;
4438 	TAILQ_INSERT_TAIL(&umtx_shm_registry[key->hash], reg, ushm_reg_link);
4439 	LIST_INSERT_HEAD(USHM_OBJ_UMTX(key->info.shared.object), reg,
4440 	    ushm_obj_link);
4441 	reg->ushm_flags = USHMF_REG_LINKED | USHMF_OBJ_LINKED;
4442 	mtx_unlock(&umtx_shm_lock);
4443 	*res = reg;
4444 	return (0);
4445 }
4446 
4447 static int
4448 umtx_shm_alive(struct thread *td, void *addr)
4449 {
4450 	vm_map_t map;
4451 	vm_map_entry_t entry;
4452 	vm_object_t object;
4453 	vm_pindex_t pindex;
4454 	vm_prot_t prot;
4455 	int res, ret;
4456 	boolean_t wired;
4457 
4458 	map = &td->td_proc->p_vmspace->vm_map;
4459 	res = vm_map_lookup(&map, (uintptr_t)addr, VM_PROT_READ, &entry,
4460 	    &object, &pindex, &prot, &wired);
4461 	if (res != KERN_SUCCESS)
4462 		return (EFAULT);
4463 	if (object == NULL)
4464 		ret = EINVAL;
4465 	else
4466 		ret = (object->flags & OBJ_UMTXDEAD) != 0 ? ENOTTY : 0;
4467 	vm_map_lookup_done(map, entry);
4468 	return (ret);
4469 }
4470 
4471 static void
4472 umtx_shm_init(void)
4473 {
4474 	int i;
4475 
4476 	umtx_shm_reg_zone = uma_zcreate("umtx_shm", sizeof(struct umtx_shm_reg),
4477 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
4478 	mtx_init(&umtx_shm_lock, "umtxshm", NULL, MTX_DEF);
4479 	for (i = 0; i < nitems(umtx_shm_registry); i++)
4480 		TAILQ_INIT(&umtx_shm_registry[i]);
4481 }
4482 
4483 static int
4484 umtx_shm(struct thread *td, void *addr, u_int flags)
4485 {
4486 	struct umtx_key key;
4487 	struct umtx_shm_reg *reg;
4488 	struct file *fp;
4489 	int error, fd;
4490 
4491 	if (__bitcount(flags & (UMTX_SHM_CREAT | UMTX_SHM_LOOKUP |
4492 	    UMTX_SHM_DESTROY| UMTX_SHM_ALIVE)) != 1)
4493 		return (EINVAL);
4494 	if ((flags & UMTX_SHM_ALIVE) != 0)
4495 		return (umtx_shm_alive(td, addr));
4496 	error = umtx_key_get(addr, TYPE_SHM, PROCESS_SHARE, &key);
4497 	if (error != 0)
4498 		return (error);
4499 	KASSERT(key.shared == 1, ("non-shared key"));
4500 	if ((flags & UMTX_SHM_CREAT) != 0) {
4501 		error = umtx_shm_create_reg(td, &key, &reg);
4502 	} else {
4503 		reg = umtx_shm_find_reg(&key);
4504 		if (reg == NULL)
4505 			error = ESRCH;
4506 	}
4507 	umtx_key_release(&key);
4508 	if (error != 0)
4509 		return (error);
4510 	KASSERT(reg != NULL, ("no reg"));
4511 	if ((flags & UMTX_SHM_DESTROY) != 0) {
4512 		umtx_shm_unref_reg(reg, true);
4513 	} else {
4514 #if 0
4515 #ifdef MAC
4516 		error = mac_posixshm_check_open(td->td_ucred,
4517 		    reg->ushm_obj, FFLAGS(O_RDWR));
4518 		if (error == 0)
4519 #endif
4520 			error = shm_access(reg->ushm_obj, td->td_ucred,
4521 			    FFLAGS(O_RDWR));
4522 		if (error == 0)
4523 #endif
4524 			error = falloc_caps(td, &fp, &fd, O_CLOEXEC, NULL);
4525 		if (error == 0) {
4526 			shm_hold(reg->ushm_obj);
4527 			finit(fp, FFLAGS(O_RDWR), DTYPE_SHM, reg->ushm_obj,
4528 			    &shm_ops);
4529 			td->td_retval[0] = fd;
4530 			fdrop(fp, td);
4531 		}
4532 	}
4533 	umtx_shm_unref_reg(reg, false);
4534 	return (error);
4535 }
4536 
4537 static int
4538 __umtx_op_shm(struct thread *td, struct _umtx_op_args *uap,
4539     const struct umtx_copyops *ops __unused)
4540 {
4541 
4542 	return (umtx_shm(td, uap->uaddr1, uap->val));
4543 }
4544 
4545 static int
4546 __umtx_op_robust_lists(struct thread *td, struct _umtx_op_args *uap,
4547     const struct umtx_copyops *ops)
4548 {
4549 	struct umtx_robust_lists_params rb;
4550 	int error;
4551 
4552 	if (ops->compat32) {
4553 		if ((td->td_pflags2 & TDP2_COMPAT32RB) == 0 &&
4554 		    (td->td_rb_list != 0 || td->td_rbp_list != 0 ||
4555 		    td->td_rb_inact != 0))
4556 			return (EBUSY);
4557 	} else if ((td->td_pflags2 & TDP2_COMPAT32RB) != 0) {
4558 		return (EBUSY);
4559 	}
4560 
4561 	bzero(&rb, sizeof(rb));
4562 	error = ops->copyin_robust_lists(uap->uaddr1, uap->val, &rb);
4563 	if (error != 0)
4564 		return (error);
4565 
4566 	if (ops->compat32)
4567 		td->td_pflags2 |= TDP2_COMPAT32RB;
4568 
4569 	td->td_rb_list = rb.robust_list_offset;
4570 	td->td_rbp_list = rb.robust_priv_list_offset;
4571 	td->td_rb_inact = rb.robust_inact_offset;
4572 	return (0);
4573 }
4574 
4575 #if defined(__i386__) || defined(__amd64__)
4576 /*
4577  * Provide the standard 32-bit definitions for x86, since native/compat32 use a
4578  * 32-bit time_t there.  Other architectures just need the i386 definitions
4579  * along with their standard compat32.
4580  */
4581 struct timespecx32 {
4582 	int64_t			tv_sec;
4583 	int32_t			tv_nsec;
4584 };
4585 
4586 struct umtx_timex32 {
4587 	struct	timespecx32	_timeout;
4588 	uint32_t		_flags;
4589 	uint32_t		_clockid;
4590 };
4591 
4592 #ifndef __i386__
4593 #define	timespeci386	timespec32
4594 #define	umtx_timei386	umtx_time32
4595 #endif
4596 #else /* !__i386__ && !__amd64__ */
4597 /* 32-bit architectures can emulate i386, so define these almost everywhere. */
4598 struct timespeci386 {
4599 	int32_t			tv_sec;
4600 	int32_t			tv_nsec;
4601 };
4602 
4603 struct umtx_timei386 {
4604 	struct	timespeci386	_timeout;
4605 	uint32_t		_flags;
4606 	uint32_t		_clockid;
4607 };
4608 
4609 #if defined(__LP64__)
4610 #define	timespecx32	timespec32
4611 #define	umtx_timex32	umtx_time32
4612 #endif
4613 #endif
4614 
4615 static int
4616 umtx_copyin_robust_lists32(const void *uaddr, size_t size,
4617     struct umtx_robust_lists_params *rbp)
4618 {
4619 	struct umtx_robust_lists_params_compat32 rb32;
4620 	int error;
4621 
4622 	if (size > sizeof(rb32))
4623 		return (EINVAL);
4624 	bzero(&rb32, sizeof(rb32));
4625 	error = copyin(uaddr, &rb32, size);
4626 	if (error != 0)
4627 		return (error);
4628 	CP(rb32, *rbp, robust_list_offset);
4629 	CP(rb32, *rbp, robust_priv_list_offset);
4630 	CP(rb32, *rbp, robust_inact_offset);
4631 	return (0);
4632 }
4633 
4634 #ifndef __i386__
4635 static inline int
4636 umtx_copyin_timeouti386(const void *uaddr, struct timespec *tsp)
4637 {
4638 	struct timespeci386 ts32;
4639 	int error;
4640 
4641 	error = copyin(uaddr, &ts32, sizeof(ts32));
4642 	if (error == 0) {
4643 		if (ts32.tv_sec < 0 ||
4644 		    ts32.tv_nsec >= 1000000000 ||
4645 		    ts32.tv_nsec < 0)
4646 			error = EINVAL;
4647 		else {
4648 			CP(ts32, *tsp, tv_sec);
4649 			CP(ts32, *tsp, tv_nsec);
4650 		}
4651 	}
4652 	return (error);
4653 }
4654 
4655 static inline int
4656 umtx_copyin_umtx_timei386(const void *uaddr, size_t size, struct _umtx_time *tp)
4657 {
4658 	struct umtx_timei386 t32;
4659 	int error;
4660 
4661 	t32._clockid = CLOCK_REALTIME;
4662 	t32._flags   = 0;
4663 	if (size <= sizeof(t32._timeout))
4664 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4665 	else
4666 		error = copyin(uaddr, &t32, sizeof(t32));
4667 	if (error != 0)
4668 		return (error);
4669 	if (t32._timeout.tv_sec < 0 ||
4670 	    t32._timeout.tv_nsec >= 1000000000 || t32._timeout.tv_nsec < 0)
4671 		return (EINVAL);
4672 	TS_CP(t32, *tp, _timeout);
4673 	CP(t32, *tp, _flags);
4674 	CP(t32, *tp, _clockid);
4675 	return (0);
4676 }
4677 
4678 static int
4679 umtx_copyout_timeouti386(void *uaddr, size_t sz, struct timespec *tsp)
4680 {
4681 	struct timespeci386 remain32 = {
4682 		.tv_sec = tsp->tv_sec,
4683 		.tv_nsec = tsp->tv_nsec,
4684 	};
4685 
4686 	/*
4687 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4688 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4689 	 * copyops.
4690 	 */
4691 	KASSERT(sz >= sizeof(remain32),
4692 	    ("umtx_copyops specifies incorrect sizes"));
4693 
4694 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4695 }
4696 #endif /* !__i386__ */
4697 
4698 #if defined(__i386__) || defined(__LP64__)
4699 static inline int
4700 umtx_copyin_timeoutx32(const void *uaddr, struct timespec *tsp)
4701 {
4702 	struct timespecx32 ts32;
4703 	int error;
4704 
4705 	error = copyin(uaddr, &ts32, sizeof(ts32));
4706 	if (error == 0) {
4707 		if (ts32.tv_sec < 0 ||
4708 		    ts32.tv_nsec >= 1000000000 ||
4709 		    ts32.tv_nsec < 0)
4710 			error = EINVAL;
4711 		else {
4712 			CP(ts32, *tsp, tv_sec);
4713 			CP(ts32, *tsp, tv_nsec);
4714 		}
4715 	}
4716 	return (error);
4717 }
4718 
4719 static inline int
4720 umtx_copyin_umtx_timex32(const void *uaddr, size_t size, struct _umtx_time *tp)
4721 {
4722 	struct umtx_timex32 t32;
4723 	int error;
4724 
4725 	t32._clockid = CLOCK_REALTIME;
4726 	t32._flags   = 0;
4727 	if (size <= sizeof(t32._timeout))
4728 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4729 	else
4730 		error = copyin(uaddr, &t32, sizeof(t32));
4731 	if (error != 0)
4732 		return (error);
4733 	if (t32._timeout.tv_sec < 0 ||
4734 	    t32._timeout.tv_nsec >= 1000000000 || t32._timeout.tv_nsec < 0)
4735 		return (EINVAL);
4736 	TS_CP(t32, *tp, _timeout);
4737 	CP(t32, *tp, _flags);
4738 	CP(t32, *tp, _clockid);
4739 	return (0);
4740 }
4741 
4742 static int
4743 umtx_copyout_timeoutx32(void *uaddr, size_t sz, struct timespec *tsp)
4744 {
4745 	struct timespecx32 remain32 = {
4746 		.tv_sec = tsp->tv_sec,
4747 		.tv_nsec = tsp->tv_nsec,
4748 	};
4749 
4750 	/*
4751 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4752 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4753 	 * copyops.
4754 	 */
4755 	KASSERT(sz >= sizeof(remain32),
4756 	    ("umtx_copyops specifies incorrect sizes"));
4757 
4758 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4759 }
4760 #endif /* __i386__ || __LP64__ */
4761 
4762 typedef int (*_umtx_op_func)(struct thread *td, struct _umtx_op_args *uap,
4763     const struct umtx_copyops *umtx_ops);
4764 
4765 static const _umtx_op_func op_table[] = {
4766 #ifdef COMPAT_FREEBSD10
4767 	[UMTX_OP_LOCK]		= __umtx_op_lock_umtx,
4768 	[UMTX_OP_UNLOCK]	= __umtx_op_unlock_umtx,
4769 #else
4770 	[UMTX_OP_LOCK]		= __umtx_op_unimpl,
4771 	[UMTX_OP_UNLOCK]	= __umtx_op_unimpl,
4772 #endif
4773 	[UMTX_OP_WAIT]		= __umtx_op_wait,
4774 	[UMTX_OP_WAKE]		= __umtx_op_wake,
4775 	[UMTX_OP_MUTEX_TRYLOCK]	= __umtx_op_trylock_umutex,
4776 	[UMTX_OP_MUTEX_LOCK]	= __umtx_op_lock_umutex,
4777 	[UMTX_OP_MUTEX_UNLOCK]	= __umtx_op_unlock_umutex,
4778 	[UMTX_OP_SET_CEILING]	= __umtx_op_set_ceiling,
4779 	[UMTX_OP_CV_WAIT]	= __umtx_op_cv_wait,
4780 	[UMTX_OP_CV_SIGNAL]	= __umtx_op_cv_signal,
4781 	[UMTX_OP_CV_BROADCAST]	= __umtx_op_cv_broadcast,
4782 	[UMTX_OP_WAIT_UINT]	= __umtx_op_wait_uint,
4783 	[UMTX_OP_RW_RDLOCK]	= __umtx_op_rw_rdlock,
4784 	[UMTX_OP_RW_WRLOCK]	= __umtx_op_rw_wrlock,
4785 	[UMTX_OP_RW_UNLOCK]	= __umtx_op_rw_unlock,
4786 	[UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private,
4787 	[UMTX_OP_WAKE_PRIVATE]	= __umtx_op_wake_private,
4788 	[UMTX_OP_MUTEX_WAIT]	= __umtx_op_wait_umutex,
4789 	[UMTX_OP_MUTEX_WAKE]	= __umtx_op_wake_umutex,
4790 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4791 	[UMTX_OP_SEM_WAIT]	= __umtx_op_sem_wait,
4792 	[UMTX_OP_SEM_WAKE]	= __umtx_op_sem_wake,
4793 #else
4794 	[UMTX_OP_SEM_WAIT]	= __umtx_op_unimpl,
4795 	[UMTX_OP_SEM_WAKE]	= __umtx_op_unimpl,
4796 #endif
4797 	[UMTX_OP_NWAKE_PRIVATE]	= __umtx_op_nwake_private,
4798 	[UMTX_OP_MUTEX_WAKE2]	= __umtx_op_wake2_umutex,
4799 	[UMTX_OP_SEM2_WAIT]	= __umtx_op_sem2_wait,
4800 	[UMTX_OP_SEM2_WAKE]	= __umtx_op_sem2_wake,
4801 	[UMTX_OP_SHM]		= __umtx_op_shm,
4802 	[UMTX_OP_ROBUST_LISTS]	= __umtx_op_robust_lists,
4803 };
4804 
4805 static const struct umtx_copyops umtx_native_ops = {
4806 	.copyin_timeout = umtx_copyin_timeout,
4807 	.copyin_umtx_time = umtx_copyin_umtx_time,
4808 	.copyin_robust_lists = umtx_copyin_robust_lists,
4809 	.copyout_timeout = umtx_copyout_timeout,
4810 	.timespec_sz = sizeof(struct timespec),
4811 	.umtx_time_sz = sizeof(struct _umtx_time),
4812 };
4813 
4814 #ifndef __i386__
4815 static const struct umtx_copyops umtx_native_opsi386 = {
4816 	.copyin_timeout = umtx_copyin_timeouti386,
4817 	.copyin_umtx_time = umtx_copyin_umtx_timei386,
4818 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4819 	.copyout_timeout = umtx_copyout_timeouti386,
4820 	.timespec_sz = sizeof(struct timespeci386),
4821 	.umtx_time_sz = sizeof(struct umtx_timei386),
4822 	.compat32 = true,
4823 };
4824 #endif
4825 
4826 #if defined(__i386__) || defined(__LP64__)
4827 /* i386 can emulate other 32-bit archs, too! */
4828 static const struct umtx_copyops umtx_native_opsx32 = {
4829 	.copyin_timeout = umtx_copyin_timeoutx32,
4830 	.copyin_umtx_time = umtx_copyin_umtx_timex32,
4831 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4832 	.copyout_timeout = umtx_copyout_timeoutx32,
4833 	.timespec_sz = sizeof(struct timespecx32),
4834 	.umtx_time_sz = sizeof(struct umtx_timex32),
4835 	.compat32 = true,
4836 };
4837 
4838 #ifdef COMPAT_FREEBSD32
4839 #ifdef __amd64__
4840 #define	umtx_native_ops32	umtx_native_opsi386
4841 #else
4842 #define	umtx_native_ops32	umtx_native_opsx32
4843 #endif
4844 #endif /* COMPAT_FREEBSD32 */
4845 #endif /* __i386__ || __LP64__ */
4846 
4847 #define	UMTX_OP__FLAGS	(UMTX_OP__32BIT | UMTX_OP__I386)
4848 
4849 static int
4850 kern__umtx_op(struct thread *td, void *obj, int op, unsigned long val,
4851     void *uaddr1, void *uaddr2, const struct umtx_copyops *ops)
4852 {
4853 	struct _umtx_op_args uap = {
4854 		.obj = obj,
4855 		.op = op & ~UMTX_OP__FLAGS,
4856 		.val = val,
4857 		.uaddr1 = uaddr1,
4858 		.uaddr2 = uaddr2
4859 	};
4860 
4861 	if ((uap.op >= nitems(op_table)))
4862 		return (EINVAL);
4863 	return ((*op_table[uap.op])(td, &uap, ops));
4864 }
4865 
4866 int
4867 sys__umtx_op(struct thread *td, struct _umtx_op_args *uap)
4868 {
4869 	static const struct umtx_copyops *umtx_ops;
4870 
4871 	umtx_ops = &umtx_native_ops;
4872 #ifdef __LP64__
4873 	if ((uap->op & (UMTX_OP__32BIT | UMTX_OP__I386)) != 0) {
4874 		if ((uap->op & UMTX_OP__I386) != 0)
4875 			umtx_ops = &umtx_native_opsi386;
4876 		else
4877 			umtx_ops = &umtx_native_opsx32;
4878 	}
4879 #elif !defined(__i386__)
4880 	/* We consider UMTX_OP__32BIT a nop on !i386 ILP32. */
4881 	if ((uap->op & UMTX_OP__I386) != 0)
4882 		umtx_ops = &umtx_native_opsi386;
4883 #else
4884 	/* Likewise, UMTX_OP__I386 is a nop on i386. */
4885 	if ((uap->op & UMTX_OP__32BIT) != 0)
4886 		umtx_ops = &umtx_native_opsx32;
4887 #endif
4888 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4889 	    uap->uaddr2, umtx_ops));
4890 }
4891 
4892 #ifdef COMPAT_FREEBSD32
4893 #ifdef COMPAT_FREEBSD10
4894 int
4895 freebsd10_freebsd32__umtx_lock(struct thread *td,
4896     struct freebsd10_freebsd32__umtx_lock_args *uap)
4897 {
4898 	return (do_lock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid, NULL));
4899 }
4900 
4901 int
4902 freebsd10_freebsd32__umtx_unlock(struct thread *td,
4903     struct freebsd10_freebsd32__umtx_unlock_args *uap)
4904 {
4905 	return (do_unlock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid));
4906 }
4907 #endif /* COMPAT_FREEBSD10 */
4908 
4909 int
4910 freebsd32__umtx_op(struct thread *td, struct freebsd32__umtx_op_args *uap)
4911 {
4912 
4913 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4914 	    uap->uaddr2, &umtx_native_ops32));
4915 }
4916 #endif /* COMPAT_FREEBSD32 */
4917 
4918 void
4919 umtx_thread_init(struct thread *td)
4920 {
4921 
4922 	td->td_umtxq = umtxq_alloc();
4923 	td->td_umtxq->uq_thread = td;
4924 }
4925 
4926 void
4927 umtx_thread_fini(struct thread *td)
4928 {
4929 
4930 	umtxq_free(td->td_umtxq);
4931 }
4932 
4933 /*
4934  * It will be called when new thread is created, e.g fork().
4935  */
4936 void
4937 umtx_thread_alloc(struct thread *td)
4938 {
4939 	struct umtx_q *uq;
4940 
4941 	uq = td->td_umtxq;
4942 	uq->uq_inherited_pri = PRI_MAX;
4943 
4944 	KASSERT(uq->uq_flags == 0, ("uq_flags != 0"));
4945 	KASSERT(uq->uq_thread == td, ("uq_thread != td"));
4946 	KASSERT(uq->uq_pi_blocked == NULL, ("uq_pi_blocked != NULL"));
4947 	KASSERT(TAILQ_EMPTY(&uq->uq_pi_contested), ("uq_pi_contested is not empty"));
4948 }
4949 
4950 /*
4951  * exec() hook.
4952  *
4953  * Clear robust lists for all process' threads, not delaying the
4954  * cleanup to thread exit, since the relevant address space is
4955  * destroyed right now.
4956  */
4957 void
4958 umtx_exec(struct proc *p)
4959 {
4960 	struct thread *td;
4961 
4962 	KASSERT(p == curproc, ("need curproc"));
4963 	KASSERT((p->p_flag & P_HADTHREADS) == 0 ||
4964 	    (p->p_flag & P_STOPPED_SINGLE) != 0,
4965 	    ("curproc must be single-threaded"));
4966 	/*
4967 	 * There is no need to lock the list as only this thread can be
4968 	 * running.
4969 	 */
4970 	FOREACH_THREAD_IN_PROC(p, td) {
4971 		KASSERT(td == curthread ||
4972 		    ((td->td_flags & TDF_BOUNDARY) != 0 && TD_IS_SUSPENDED(td)),
4973 		    ("running thread %p %p", p, td));
4974 		umtx_thread_cleanup(td);
4975 		td->td_rb_list = td->td_rbp_list = td->td_rb_inact = 0;
4976 	}
4977 }
4978 
4979 /*
4980  * thread exit hook.
4981  */
4982 void
4983 umtx_thread_exit(struct thread *td)
4984 {
4985 
4986 	umtx_thread_cleanup(td);
4987 }
4988 
4989 static int
4990 umtx_read_uptr(struct thread *td, uintptr_t ptr, uintptr_t *res, bool compat32)
4991 {
4992 	u_long res1;
4993 	uint32_t res32;
4994 	int error;
4995 
4996 	if (compat32) {
4997 		error = fueword32((void *)ptr, &res32);
4998 		if (error == 0)
4999 			res1 = res32;
5000 	} else {
5001 		error = fueword((void *)ptr, &res1);
5002 	}
5003 	if (error == 0)
5004 		*res = res1;
5005 	else
5006 		error = EFAULT;
5007 	return (error);
5008 }
5009 
5010 static void
5011 umtx_read_rb_list(struct thread *td, struct umutex *m, uintptr_t *rb_list,
5012     bool compat32)
5013 {
5014 	struct umutex32 m32;
5015 
5016 	if (compat32) {
5017 		memcpy(&m32, m, sizeof(m32));
5018 		*rb_list = m32.m_rb_lnk;
5019 	} else {
5020 		*rb_list = m->m_rb_lnk;
5021 	}
5022 }
5023 
5024 static int
5025 umtx_handle_rb(struct thread *td, uintptr_t rbp, uintptr_t *rb_list, bool inact,
5026     bool compat32)
5027 {
5028 	struct umutex m;
5029 	int error;
5030 
5031 	KASSERT(td->td_proc == curproc, ("need current vmspace"));
5032 	error = copyin((void *)rbp, &m, sizeof(m));
5033 	if (error != 0)
5034 		return (error);
5035 	if (rb_list != NULL)
5036 		umtx_read_rb_list(td, &m, rb_list, compat32);
5037 	if ((m.m_flags & UMUTEX_ROBUST) == 0)
5038 		return (EINVAL);
5039 	if ((m.m_owner & ~UMUTEX_CONTESTED) != td->td_tid)
5040 		/* inact is cleared after unlock, allow the inconsistency */
5041 		return (inact ? 0 : EINVAL);
5042 	return (do_unlock_umutex(td, (struct umutex *)rbp, true));
5043 }
5044 
5045 static void
5046 umtx_cleanup_rb_list(struct thread *td, uintptr_t rb_list, uintptr_t *rb_inact,
5047     const char *name, bool compat32)
5048 {
5049 	int error, i;
5050 	uintptr_t rbp;
5051 	bool inact;
5052 
5053 	if (rb_list == 0)
5054 		return;
5055 	error = umtx_read_uptr(td, rb_list, &rbp, compat32);
5056 	for (i = 0; error == 0 && rbp != 0 && i < umtx_max_rb; i++) {
5057 		if (rbp == *rb_inact) {
5058 			inact = true;
5059 			*rb_inact = 0;
5060 		} else
5061 			inact = false;
5062 		error = umtx_handle_rb(td, rbp, &rbp, inact, compat32);
5063 	}
5064 	if (i == umtx_max_rb && umtx_verbose_rb) {
5065 		uprintf("comm %s pid %d: reached umtx %smax rb %d\n",
5066 		    td->td_proc->p_comm, td->td_proc->p_pid, name, umtx_max_rb);
5067 	}
5068 	if (error != 0 && umtx_verbose_rb) {
5069 		uprintf("comm %s pid %d: handling %srb error %d\n",
5070 		    td->td_proc->p_comm, td->td_proc->p_pid, name, error);
5071 	}
5072 }
5073 
5074 /*
5075  * Clean up umtx data.
5076  */
5077 static void
5078 umtx_thread_cleanup(struct thread *td)
5079 {
5080 	struct umtx_q *uq;
5081 	struct umtx_pi *pi;
5082 	uintptr_t rb_inact;
5083 	bool compat32;
5084 
5085 	/*
5086 	 * Disown pi mutexes.
5087 	 */
5088 	uq = td->td_umtxq;
5089 	if (uq != NULL) {
5090 		if (uq->uq_inherited_pri != PRI_MAX ||
5091 		    !TAILQ_EMPTY(&uq->uq_pi_contested)) {
5092 			mtx_lock(&umtx_lock);
5093 			uq->uq_inherited_pri = PRI_MAX;
5094 			while ((pi = TAILQ_FIRST(&uq->uq_pi_contested)) != NULL) {
5095 				pi->pi_owner = NULL;
5096 				TAILQ_REMOVE(&uq->uq_pi_contested, pi, pi_link);
5097 			}
5098 			mtx_unlock(&umtx_lock);
5099 		}
5100 		sched_lend_user_prio_cond(td, PRI_MAX);
5101 	}
5102 
5103 	compat32 = (td->td_pflags2 & TDP2_COMPAT32RB) != 0;
5104 	td->td_pflags2 &= ~TDP2_COMPAT32RB;
5105 
5106 	if (td->td_rb_inact == 0 && td->td_rb_list == 0 && td->td_rbp_list == 0)
5107 		return;
5108 
5109 	/*
5110 	 * Handle terminated robust mutexes.  Must be done after
5111 	 * robust pi disown, otherwise unlock could see unowned
5112 	 * entries.
5113 	 */
5114 	rb_inact = td->td_rb_inact;
5115 	if (rb_inact != 0)
5116 		(void)umtx_read_uptr(td, rb_inact, &rb_inact, compat32);
5117 	umtx_cleanup_rb_list(td, td->td_rb_list, &rb_inact, "", compat32);
5118 	umtx_cleanup_rb_list(td, td->td_rbp_list, &rb_inact, "priv ", compat32);
5119 	if (rb_inact != 0)
5120 		(void)umtx_handle_rb(td, rb_inact, NULL, true, compat32);
5121 }
5122