xref: /freebsd/sys/kern/kern_umtx.c (revision 56070dd2e4df70e4de2a10c7ace684ceb8b0f0bd)
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, wmesg,
812 		    sbt, 0, flags);
813 		if (error == EINTR || error == ERESTART)
814 			break;
815 		if (error == EWOULDBLOCK && (flags & C_ABSOLUTE) != 0) {
816 			error = ETIMEDOUT;
817 			break;
818 		}
819 	}
820 
821 	curthread->td_rtcgen = 0;
822 	return (error);
823 }
824 
825 /*
826  * Convert userspace address into unique logical address.
827  */
828 int
829 umtx_key_get(const void *addr, int type, int share, struct umtx_key *key)
830 {
831 	struct thread *td = curthread;
832 	vm_map_t map;
833 	vm_map_entry_t entry;
834 	vm_pindex_t pindex;
835 	vm_prot_t prot;
836 	boolean_t wired;
837 
838 	key->type = type;
839 	if (share == THREAD_SHARE) {
840 		key->shared = 0;
841 		key->info.private.vs = td->td_proc->p_vmspace;
842 		key->info.private.addr = (uintptr_t)addr;
843 	} else {
844 		MPASS(share == PROCESS_SHARE || share == AUTO_SHARE);
845 		map = &td->td_proc->p_vmspace->vm_map;
846 		if (vm_map_lookup(&map, (vm_offset_t)addr, VM_PROT_WRITE,
847 		    &entry, &key->info.shared.object, &pindex, &prot,
848 		    &wired) != KERN_SUCCESS) {
849 			return (EFAULT);
850 		}
851 
852 		if ((share == PROCESS_SHARE) ||
853 		    (share == AUTO_SHARE &&
854 		     VM_INHERIT_SHARE == entry->inheritance)) {
855 			key->shared = 1;
856 			key->info.shared.offset = (vm_offset_t)addr -
857 			    entry->start + entry->offset;
858 			vm_object_reference(key->info.shared.object);
859 		} else {
860 			key->shared = 0;
861 			key->info.private.vs = td->td_proc->p_vmspace;
862 			key->info.private.addr = (uintptr_t)addr;
863 		}
864 		vm_map_lookup_done(map, entry);
865 	}
866 
867 	umtxq_hash(key);
868 	return (0);
869 }
870 
871 /*
872  * Release key.
873  */
874 void
875 umtx_key_release(struct umtx_key *key)
876 {
877 	if (key->shared)
878 		vm_object_deallocate(key->info.shared.object);
879 }
880 
881 #ifdef COMPAT_FREEBSD10
882 /*
883  * Lock a umtx object.
884  */
885 static int
886 do_lock_umtx(struct thread *td, struct umtx *umtx, u_long id,
887     const struct timespec *timeout)
888 {
889 	struct umtx_abs_timeout timo;
890 	struct umtx_q *uq;
891 	u_long owner;
892 	u_long old;
893 	int error = 0;
894 
895 	uq = td->td_umtxq;
896 	if (timeout != NULL)
897 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
898 
899 	/*
900 	 * Care must be exercised when dealing with umtx structure. It
901 	 * can fault on any access.
902 	 */
903 	for (;;) {
904 		/*
905 		 * Try the uncontested case.  This should be done in userland.
906 		 */
907 		owner = casuword(&umtx->u_owner, UMTX_UNOWNED, id);
908 
909 		/* The acquire succeeded. */
910 		if (owner == UMTX_UNOWNED)
911 			return (0);
912 
913 		/* The address was invalid. */
914 		if (owner == -1)
915 			return (EFAULT);
916 
917 		/* If no one owns it but it is contested try to acquire it. */
918 		if (owner == UMTX_CONTESTED) {
919 			owner = casuword(&umtx->u_owner,
920 			    UMTX_CONTESTED, id | UMTX_CONTESTED);
921 
922 			if (owner == UMTX_CONTESTED)
923 				return (0);
924 
925 			/* The address was invalid. */
926 			if (owner == -1)
927 				return (EFAULT);
928 
929 			error = thread_check_susp(td, false);
930 			if (error != 0)
931 				break;
932 
933 			/* If this failed the lock has changed, restart. */
934 			continue;
935 		}
936 
937 		/*
938 		 * If we caught a signal, we have retried and now
939 		 * exit immediately.
940 		 */
941 		if (error != 0)
942 			break;
943 
944 		if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK,
945 			AUTO_SHARE, &uq->uq_key)) != 0)
946 			return (error);
947 
948 		umtxq_lock(&uq->uq_key);
949 		umtxq_busy(&uq->uq_key);
950 		umtxq_insert(uq);
951 		umtxq_unbusy(&uq->uq_key);
952 		umtxq_unlock(&uq->uq_key);
953 
954 		/*
955 		 * Set the contested bit so that a release in user space
956 		 * knows to use the system call for unlock.  If this fails
957 		 * either some one else has acquired the lock or it has been
958 		 * released.
959 		 */
960 		old = casuword(&umtx->u_owner, owner, owner | UMTX_CONTESTED);
961 
962 		/* The address was invalid. */
963 		if (old == -1) {
964 			umtxq_lock(&uq->uq_key);
965 			umtxq_remove(uq);
966 			umtxq_unlock(&uq->uq_key);
967 			umtx_key_release(&uq->uq_key);
968 			return (EFAULT);
969 		}
970 
971 		/*
972 		 * We set the contested bit, sleep. Otherwise the lock changed
973 		 * and we need to retry or we lost a race to the thread
974 		 * unlocking the umtx.
975 		 */
976 		umtxq_lock(&uq->uq_key);
977 		if (old == owner)
978 			error = umtxq_sleep(uq, "umtx", timeout == NULL ? NULL :
979 			    &timo);
980 		umtxq_remove(uq);
981 		umtxq_unlock(&uq->uq_key);
982 		umtx_key_release(&uq->uq_key);
983 
984 		if (error == 0)
985 			error = thread_check_susp(td, false);
986 	}
987 
988 	if (timeout == NULL) {
989 		/* Mutex locking is restarted if it is interrupted. */
990 		if (error == EINTR)
991 			error = ERESTART;
992 	} else {
993 		/* Timed-locking is not restarted. */
994 		if (error == ERESTART)
995 			error = EINTR;
996 	}
997 	return (error);
998 }
999 
1000 /*
1001  * Unlock a umtx object.
1002  */
1003 static int
1004 do_unlock_umtx(struct thread *td, struct umtx *umtx, u_long id)
1005 {
1006 	struct umtx_key key;
1007 	u_long owner;
1008 	u_long old;
1009 	int error;
1010 	int count;
1011 
1012 	/*
1013 	 * Make sure we own this mtx.
1014 	 */
1015 	owner = fuword(__DEVOLATILE(u_long *, &umtx->u_owner));
1016 	if (owner == -1)
1017 		return (EFAULT);
1018 
1019 	if ((owner & ~UMTX_CONTESTED) != id)
1020 		return (EPERM);
1021 
1022 	/* This should be done in userland */
1023 	if ((owner & UMTX_CONTESTED) == 0) {
1024 		old = casuword(&umtx->u_owner, owner, UMTX_UNOWNED);
1025 		if (old == -1)
1026 			return (EFAULT);
1027 		if (old == owner)
1028 			return (0);
1029 		owner = old;
1030 	}
1031 
1032 	/* We should only ever be in here for contested locks */
1033 	if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1034 	    &key)) != 0)
1035 		return (error);
1036 
1037 	umtxq_lock(&key);
1038 	umtxq_busy(&key);
1039 	count = umtxq_count(&key);
1040 	umtxq_unlock(&key);
1041 
1042 	/*
1043 	 * When unlocking the umtx, it must be marked as unowned if
1044 	 * there is zero or one thread only waiting for it.
1045 	 * Otherwise, it must be marked as contested.
1046 	 */
1047 	old = casuword(&umtx->u_owner, owner,
1048 	    count <= 1 ? UMTX_UNOWNED : UMTX_CONTESTED);
1049 	umtxq_lock(&key);
1050 	umtxq_signal(&key,1);
1051 	umtxq_unbusy(&key);
1052 	umtxq_unlock(&key);
1053 	umtx_key_release(&key);
1054 	if (old == -1)
1055 		return (EFAULT);
1056 	if (old != owner)
1057 		return (EINVAL);
1058 	return (0);
1059 }
1060 
1061 #ifdef COMPAT_FREEBSD32
1062 
1063 /*
1064  * Lock a umtx object.
1065  */
1066 static int
1067 do_lock_umtx32(struct thread *td, uint32_t *m, uint32_t id,
1068 	const struct timespec *timeout)
1069 {
1070 	struct umtx_abs_timeout timo;
1071 	struct umtx_q *uq;
1072 	uint32_t owner;
1073 	uint32_t old;
1074 	int error = 0;
1075 
1076 	uq = td->td_umtxq;
1077 
1078 	if (timeout != NULL)
1079 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
1080 
1081 	/*
1082 	 * Care must be exercised when dealing with umtx structure. It
1083 	 * can fault on any access.
1084 	 */
1085 	for (;;) {
1086 		/*
1087 		 * Try the uncontested case.  This should be done in userland.
1088 		 */
1089 		owner = casuword32(m, UMUTEX_UNOWNED, id);
1090 
1091 		/* The acquire succeeded. */
1092 		if (owner == UMUTEX_UNOWNED)
1093 			return (0);
1094 
1095 		/* The address was invalid. */
1096 		if (owner == -1)
1097 			return (EFAULT);
1098 
1099 		/* If no one owns it but it is contested try to acquire it. */
1100 		if (owner == UMUTEX_CONTESTED) {
1101 			owner = casuword32(m,
1102 			    UMUTEX_CONTESTED, id | UMUTEX_CONTESTED);
1103 			if (owner == UMUTEX_CONTESTED)
1104 				return (0);
1105 
1106 			/* The address was invalid. */
1107 			if (owner == -1)
1108 				return (EFAULT);
1109 
1110 			error = thread_check_susp(td, false);
1111 			if (error != 0)
1112 				break;
1113 
1114 			/* If this failed the lock has changed, restart. */
1115 			continue;
1116 		}
1117 
1118 		/*
1119 		 * If we caught a signal, we have retried and now
1120 		 * exit immediately.
1121 		 */
1122 		if (error != 0)
1123 			return (error);
1124 
1125 		if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK,
1126 			AUTO_SHARE, &uq->uq_key)) != 0)
1127 			return (error);
1128 
1129 		umtxq_lock(&uq->uq_key);
1130 		umtxq_busy(&uq->uq_key);
1131 		umtxq_insert(uq);
1132 		umtxq_unbusy(&uq->uq_key);
1133 		umtxq_unlock(&uq->uq_key);
1134 
1135 		/*
1136 		 * Set the contested bit so that a release in user space
1137 		 * knows to use the system call for unlock.  If this fails
1138 		 * either some one else has acquired the lock or it has been
1139 		 * released.
1140 		 */
1141 		old = casuword32(m, owner, owner | UMUTEX_CONTESTED);
1142 
1143 		/* The address was invalid. */
1144 		if (old == -1) {
1145 			umtxq_lock(&uq->uq_key);
1146 			umtxq_remove(uq);
1147 			umtxq_unlock(&uq->uq_key);
1148 			umtx_key_release(&uq->uq_key);
1149 			return (EFAULT);
1150 		}
1151 
1152 		/*
1153 		 * We set the contested bit, sleep. Otherwise the lock changed
1154 		 * and we need to retry or we lost a race to the thread
1155 		 * unlocking the umtx.
1156 		 */
1157 		umtxq_lock(&uq->uq_key);
1158 		if (old == owner)
1159 			error = umtxq_sleep(uq, "umtx", timeout == NULL ?
1160 			    NULL : &timo);
1161 		umtxq_remove(uq);
1162 		umtxq_unlock(&uq->uq_key);
1163 		umtx_key_release(&uq->uq_key);
1164 
1165 		if (error == 0)
1166 			error = thread_check_susp(td, false);
1167 	}
1168 
1169 	if (timeout == NULL) {
1170 		/* Mutex locking is restarted if it is interrupted. */
1171 		if (error == EINTR)
1172 			error = ERESTART;
1173 	} else {
1174 		/* Timed-locking is not restarted. */
1175 		if (error == ERESTART)
1176 			error = EINTR;
1177 	}
1178 	return (error);
1179 }
1180 
1181 /*
1182  * Unlock a umtx object.
1183  */
1184 static int
1185 do_unlock_umtx32(struct thread *td, uint32_t *m, uint32_t id)
1186 {
1187 	struct umtx_key key;
1188 	uint32_t owner;
1189 	uint32_t old;
1190 	int error;
1191 	int count;
1192 
1193 	/*
1194 	 * Make sure we own this mtx.
1195 	 */
1196 	owner = fuword32(m);
1197 	if (owner == -1)
1198 		return (EFAULT);
1199 
1200 	if ((owner & ~UMUTEX_CONTESTED) != id)
1201 		return (EPERM);
1202 
1203 	/* This should be done in userland */
1204 	if ((owner & UMUTEX_CONTESTED) == 0) {
1205 		old = casuword32(m, owner, UMUTEX_UNOWNED);
1206 		if (old == -1)
1207 			return (EFAULT);
1208 		if (old == owner)
1209 			return (0);
1210 		owner = old;
1211 	}
1212 
1213 	/* We should only ever be in here for contested locks */
1214 	if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1215 		&key)) != 0)
1216 		return (error);
1217 
1218 	umtxq_lock(&key);
1219 	umtxq_busy(&key);
1220 	count = umtxq_count(&key);
1221 	umtxq_unlock(&key);
1222 
1223 	/*
1224 	 * When unlocking the umtx, it must be marked as unowned if
1225 	 * there is zero or one thread only waiting for it.
1226 	 * Otherwise, it must be marked as contested.
1227 	 */
1228 	old = casuword32(m, owner,
1229 		count <= 1 ? UMUTEX_UNOWNED : UMUTEX_CONTESTED);
1230 	umtxq_lock(&key);
1231 	umtxq_signal(&key,1);
1232 	umtxq_unbusy(&key);
1233 	umtxq_unlock(&key);
1234 	umtx_key_release(&key);
1235 	if (old == -1)
1236 		return (EFAULT);
1237 	if (old != owner)
1238 		return (EINVAL);
1239 	return (0);
1240 }
1241 #endif	/* COMPAT_FREEBSD32 */
1242 #endif	/* COMPAT_FREEBSD10 */
1243 
1244 /*
1245  * Fetch and compare value, sleep on the address if value is not changed.
1246  */
1247 static int
1248 do_wait(struct thread *td, void *addr, u_long id,
1249     struct _umtx_time *timeout, int compat32, int is_private)
1250 {
1251 	struct umtx_abs_timeout timo;
1252 	struct umtx_q *uq;
1253 	u_long tmp;
1254 	uint32_t tmp32;
1255 	int error = 0;
1256 
1257 	uq = td->td_umtxq;
1258 	if ((error = umtx_key_get(addr, TYPE_SIMPLE_WAIT,
1259 		is_private ? THREAD_SHARE : AUTO_SHARE, &uq->uq_key)) != 0)
1260 		return (error);
1261 
1262 	if (timeout != NULL)
1263 		umtx_abs_timeout_init2(&timo, timeout);
1264 
1265 	umtxq_lock(&uq->uq_key);
1266 	umtxq_insert(uq);
1267 	umtxq_unlock(&uq->uq_key);
1268 	if (compat32 == 0) {
1269 		error = fueword(addr, &tmp);
1270 		if (error != 0)
1271 			error = EFAULT;
1272 	} else {
1273 		error = fueword32(addr, &tmp32);
1274 		if (error == 0)
1275 			tmp = tmp32;
1276 		else
1277 			error = EFAULT;
1278 	}
1279 	umtxq_lock(&uq->uq_key);
1280 	if (error == 0) {
1281 		if (tmp == id)
1282 			error = umtxq_sleep(uq, "uwait", timeout == NULL ?
1283 			    NULL : &timo);
1284 		if ((uq->uq_flags & UQF_UMTXQ) == 0)
1285 			error = 0;
1286 		else
1287 			umtxq_remove(uq);
1288 	} else if ((uq->uq_flags & UQF_UMTXQ) != 0) {
1289 		umtxq_remove(uq);
1290 	}
1291 	umtxq_unlock(&uq->uq_key);
1292 	umtx_key_release(&uq->uq_key);
1293 	if (error == ERESTART)
1294 		error = EINTR;
1295 	return (error);
1296 }
1297 
1298 /*
1299  * Wake up threads sleeping on the specified address.
1300  */
1301 int
1302 kern_umtx_wake(struct thread *td, void *uaddr, int n_wake, int is_private)
1303 {
1304 	struct umtx_key key;
1305 	int ret;
1306 
1307 	if ((ret = umtx_key_get(uaddr, TYPE_SIMPLE_WAIT,
1308 	    is_private ? THREAD_SHARE : AUTO_SHARE, &key)) != 0)
1309 		return (ret);
1310 	umtxq_lock(&key);
1311 	umtxq_signal(&key, n_wake);
1312 	umtxq_unlock(&key);
1313 	umtx_key_release(&key);
1314 	return (0);
1315 }
1316 
1317 /*
1318  * Lock PTHREAD_PRIO_NONE protocol POSIX mutex.
1319  */
1320 static int
1321 do_lock_normal(struct thread *td, struct umutex *m, uint32_t flags,
1322     struct _umtx_time *timeout, int mode)
1323 {
1324 	struct umtx_abs_timeout timo;
1325 	struct umtx_q *uq;
1326 	uint32_t owner, old, id;
1327 	int error, rv;
1328 
1329 	id = td->td_tid;
1330 	uq = td->td_umtxq;
1331 	error = 0;
1332 	if (timeout != NULL)
1333 		umtx_abs_timeout_init2(&timo, timeout);
1334 
1335 	/*
1336 	 * Care must be exercised when dealing with umtx structure. It
1337 	 * can fault on any access.
1338 	 */
1339 	for (;;) {
1340 		rv = fueword32(&m->m_owner, &owner);
1341 		if (rv == -1)
1342 			return (EFAULT);
1343 		if (mode == _UMUTEX_WAIT) {
1344 			if (owner == UMUTEX_UNOWNED ||
1345 			    owner == UMUTEX_CONTESTED ||
1346 			    owner == UMUTEX_RB_OWNERDEAD ||
1347 			    owner == UMUTEX_RB_NOTRECOV)
1348 				return (0);
1349 		} else {
1350 			/*
1351 			 * Robust mutex terminated.  Kernel duty is to
1352 			 * return EOWNERDEAD to the userspace.  The
1353 			 * umutex.m_flags UMUTEX_NONCONSISTENT is set
1354 			 * by the common userspace code.
1355 			 */
1356 			if (owner == UMUTEX_RB_OWNERDEAD) {
1357 				rv = casueword32(&m->m_owner,
1358 				    UMUTEX_RB_OWNERDEAD, &owner,
1359 				    id | UMUTEX_CONTESTED);
1360 				if (rv == -1)
1361 					return (EFAULT);
1362 				if (rv == 0) {
1363 					MPASS(owner == UMUTEX_RB_OWNERDEAD);
1364 					return (EOWNERDEAD); /* success */
1365 				}
1366 				MPASS(rv == 1);
1367 				rv = thread_check_susp(td, false);
1368 				if (rv != 0)
1369 					return (rv);
1370 				continue;
1371 			}
1372 			if (owner == UMUTEX_RB_NOTRECOV)
1373 				return (ENOTRECOVERABLE);
1374 
1375 			/*
1376 			 * Try the uncontested case.  This should be
1377 			 * done in userland.
1378 			 */
1379 			rv = casueword32(&m->m_owner, UMUTEX_UNOWNED,
1380 			    &owner, id);
1381 			/* The address was invalid. */
1382 			if (rv == -1)
1383 				return (EFAULT);
1384 
1385 			/* The acquire succeeded. */
1386 			if (rv == 0) {
1387 				MPASS(owner == UMUTEX_UNOWNED);
1388 				return (0);
1389 			}
1390 
1391 			/*
1392 			 * If no one owns it but it is contested try
1393 			 * to acquire it.
1394 			 */
1395 			MPASS(rv == 1);
1396 			if (owner == UMUTEX_CONTESTED) {
1397 				rv = casueword32(&m->m_owner,
1398 				    UMUTEX_CONTESTED, &owner,
1399 				    id | UMUTEX_CONTESTED);
1400 				/* The address was invalid. */
1401 				if (rv == -1)
1402 					return (EFAULT);
1403 				if (rv == 0) {
1404 					MPASS(owner == UMUTEX_CONTESTED);
1405 					return (0);
1406 				}
1407 				if (rv == 1) {
1408 					rv = thread_check_susp(td, false);
1409 					if (rv != 0)
1410 						return (rv);
1411 				}
1412 
1413 				/*
1414 				 * If this failed the lock has
1415 				 * changed, restart.
1416 				 */
1417 				continue;
1418 			}
1419 
1420 			/* rv == 1 but not contested, likely store failure */
1421 			rv = thread_check_susp(td, false);
1422 			if (rv != 0)
1423 				return (rv);
1424 		}
1425 
1426 		if (mode == _UMUTEX_TRY)
1427 			return (EBUSY);
1428 
1429 		/*
1430 		 * If we caught a signal, we have retried and now
1431 		 * exit immediately.
1432 		 */
1433 		if (error != 0)
1434 			return (error);
1435 
1436 		if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX,
1437 		    GET_SHARE(flags), &uq->uq_key)) != 0)
1438 			return (error);
1439 
1440 		umtxq_lock(&uq->uq_key);
1441 		umtxq_busy(&uq->uq_key);
1442 		umtxq_insert(uq);
1443 		umtxq_unlock(&uq->uq_key);
1444 
1445 		/*
1446 		 * Set the contested bit so that a release in user space
1447 		 * knows to use the system call for unlock.  If this fails
1448 		 * either some one else has acquired the lock or it has been
1449 		 * released.
1450 		 */
1451 		rv = casueword32(&m->m_owner, owner, &old,
1452 		    owner | UMUTEX_CONTESTED);
1453 
1454 		/* The address was invalid or casueword failed to store. */
1455 		if (rv == -1 || rv == 1) {
1456 			umtxq_lock(&uq->uq_key);
1457 			umtxq_remove(uq);
1458 			umtxq_unbusy(&uq->uq_key);
1459 			umtxq_unlock(&uq->uq_key);
1460 			umtx_key_release(&uq->uq_key);
1461 			if (rv == -1)
1462 				return (EFAULT);
1463 			if (rv == 1) {
1464 				rv = thread_check_susp(td, false);
1465 				if (rv != 0)
1466 					return (rv);
1467 			}
1468 			continue;
1469 		}
1470 
1471 		/*
1472 		 * We set the contested bit, sleep. Otherwise the lock changed
1473 		 * and we need to retry or we lost a race to the thread
1474 		 * unlocking the umtx.
1475 		 */
1476 		umtxq_lock(&uq->uq_key);
1477 		umtxq_unbusy(&uq->uq_key);
1478 		MPASS(old == owner);
1479 		error = umtxq_sleep(uq, "umtxn", timeout == NULL ?
1480 		    NULL : &timo);
1481 		umtxq_remove(uq);
1482 		umtxq_unlock(&uq->uq_key);
1483 		umtx_key_release(&uq->uq_key);
1484 
1485 		if (error == 0)
1486 			error = thread_check_susp(td, false);
1487 	}
1488 
1489 	return (0);
1490 }
1491 
1492 /*
1493  * Unlock PTHREAD_PRIO_NONE protocol POSIX mutex.
1494  */
1495 static int
1496 do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1497 {
1498 	struct umtx_key key;
1499 	uint32_t owner, old, id, newlock;
1500 	int error, count;
1501 
1502 	id = td->td_tid;
1503 
1504 again:
1505 	/*
1506 	 * Make sure we own this mtx.
1507 	 */
1508 	error = fueword32(&m->m_owner, &owner);
1509 	if (error == -1)
1510 		return (EFAULT);
1511 
1512 	if ((owner & ~UMUTEX_CONTESTED) != id)
1513 		return (EPERM);
1514 
1515 	newlock = umtx_unlock_val(flags, rb);
1516 	if ((owner & UMUTEX_CONTESTED) == 0) {
1517 		error = casueword32(&m->m_owner, owner, &old, newlock);
1518 		if (error == -1)
1519 			return (EFAULT);
1520 		if (error == 1) {
1521 			error = thread_check_susp(td, false);
1522 			if (error != 0)
1523 				return (error);
1524 			goto again;
1525 		}
1526 		MPASS(old == owner);
1527 		return (0);
1528 	}
1529 
1530 	/* We should only ever be in here for contested locks */
1531 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1532 	    &key)) != 0)
1533 		return (error);
1534 
1535 	umtxq_lock(&key);
1536 	umtxq_busy(&key);
1537 	count = umtxq_count(&key);
1538 	umtxq_unlock(&key);
1539 
1540 	/*
1541 	 * When unlocking the umtx, it must be marked as unowned if
1542 	 * there is zero or one thread only waiting for it.
1543 	 * Otherwise, it must be marked as contested.
1544 	 */
1545 	if (count > 1)
1546 		newlock |= UMUTEX_CONTESTED;
1547 	error = casueword32(&m->m_owner, owner, &old, newlock);
1548 	umtxq_lock(&key);
1549 	umtxq_signal(&key, 1);
1550 	umtxq_unbusy(&key);
1551 	umtxq_unlock(&key);
1552 	umtx_key_release(&key);
1553 	if (error == -1)
1554 		return (EFAULT);
1555 	if (error == 1) {
1556 		if (old != owner)
1557 			return (EINVAL);
1558 		error = thread_check_susp(td, false);
1559 		if (error != 0)
1560 			return (error);
1561 		goto again;
1562 	}
1563 	return (0);
1564 }
1565 
1566 /*
1567  * Check if the mutex is available and wake up a waiter,
1568  * only for simple mutex.
1569  */
1570 static int
1571 do_wake_umutex(struct thread *td, struct umutex *m)
1572 {
1573 	struct umtx_key key;
1574 	uint32_t owner;
1575 	uint32_t flags;
1576 	int error;
1577 	int count;
1578 
1579 again:
1580 	error = fueword32(&m->m_owner, &owner);
1581 	if (error == -1)
1582 		return (EFAULT);
1583 
1584 	if ((owner & ~UMUTEX_CONTESTED) != 0 && owner != UMUTEX_RB_OWNERDEAD &&
1585 	    owner != UMUTEX_RB_NOTRECOV)
1586 		return (0);
1587 
1588 	error = fueword32(&m->m_flags, &flags);
1589 	if (error == -1)
1590 		return (EFAULT);
1591 
1592 	/* We should only ever be in here for contested locks */
1593 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1594 	    &key)) != 0)
1595 		return (error);
1596 
1597 	umtxq_lock(&key);
1598 	umtxq_busy(&key);
1599 	count = umtxq_count(&key);
1600 	umtxq_unlock(&key);
1601 
1602 	if (count <= 1 && owner != UMUTEX_RB_OWNERDEAD &&
1603 	    owner != UMUTEX_RB_NOTRECOV) {
1604 		error = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
1605 		    UMUTEX_UNOWNED);
1606 		if (error == -1) {
1607 			error = EFAULT;
1608 		} else if (error == 1) {
1609 			umtxq_lock(&key);
1610 			umtxq_unbusy(&key);
1611 			umtxq_unlock(&key);
1612 			umtx_key_release(&key);
1613 			error = thread_check_susp(td, false);
1614 			if (error != 0)
1615 				return (error);
1616 			goto again;
1617 		}
1618 	}
1619 
1620 	umtxq_lock(&key);
1621 	if (error == 0 && count != 0) {
1622 		MPASS((owner & ~UMUTEX_CONTESTED) == 0 ||
1623 		    owner == UMUTEX_RB_OWNERDEAD ||
1624 		    owner == UMUTEX_RB_NOTRECOV);
1625 		umtxq_signal(&key, 1);
1626 	}
1627 	umtxq_unbusy(&key);
1628 	umtxq_unlock(&key);
1629 	umtx_key_release(&key);
1630 	return (error);
1631 }
1632 
1633 /*
1634  * Check if the mutex has waiters and tries to fix contention bit.
1635  */
1636 static int
1637 do_wake2_umutex(struct thread *td, struct umutex *m, uint32_t flags)
1638 {
1639 	struct umtx_key key;
1640 	uint32_t owner, old;
1641 	int type;
1642 	int error;
1643 	int count;
1644 
1645 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT |
1646 	    UMUTEX_ROBUST)) {
1647 	case 0:
1648 	case UMUTEX_ROBUST:
1649 		type = TYPE_NORMAL_UMUTEX;
1650 		break;
1651 	case UMUTEX_PRIO_INHERIT:
1652 		type = TYPE_PI_UMUTEX;
1653 		break;
1654 	case (UMUTEX_PRIO_INHERIT | UMUTEX_ROBUST):
1655 		type = TYPE_PI_ROBUST_UMUTEX;
1656 		break;
1657 	case UMUTEX_PRIO_PROTECT:
1658 		type = TYPE_PP_UMUTEX;
1659 		break;
1660 	case (UMUTEX_PRIO_PROTECT | UMUTEX_ROBUST):
1661 		type = TYPE_PP_ROBUST_UMUTEX;
1662 		break;
1663 	default:
1664 		return (EINVAL);
1665 	}
1666 	if ((error = umtx_key_get(m, type, GET_SHARE(flags), &key)) != 0)
1667 		return (error);
1668 
1669 	owner = 0;
1670 	umtxq_lock(&key);
1671 	umtxq_busy(&key);
1672 	count = umtxq_count(&key);
1673 	umtxq_unlock(&key);
1674 
1675 	error = fueword32(&m->m_owner, &owner);
1676 	if (error == -1)
1677 		error = EFAULT;
1678 
1679 	/*
1680 	 * Only repair contention bit if there is a waiter, this means
1681 	 * the mutex is still being referenced by userland code,
1682 	 * otherwise don't update any memory.
1683 	 */
1684 	while (error == 0 && (owner & UMUTEX_CONTESTED) == 0 &&
1685 	    (count > 1 || (count == 1 && (owner & ~UMUTEX_CONTESTED) != 0))) {
1686 		error = casueword32(&m->m_owner, owner, &old,
1687 		    owner | UMUTEX_CONTESTED);
1688 		if (error == -1) {
1689 			error = EFAULT;
1690 			break;
1691 		}
1692 		if (error == 0) {
1693 			MPASS(old == owner);
1694 			break;
1695 		}
1696 		owner = old;
1697 		error = thread_check_susp(td, false);
1698 	}
1699 
1700 	umtxq_lock(&key);
1701 	if (error == EFAULT) {
1702 		umtxq_signal(&key, INT_MAX);
1703 	} else if (count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1704 	    owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1705 		umtxq_signal(&key, 1);
1706 	umtxq_unbusy(&key);
1707 	umtxq_unlock(&key);
1708 	umtx_key_release(&key);
1709 	return (error);
1710 }
1711 
1712 struct umtx_pi *
1713 umtx_pi_alloc(int flags)
1714 {
1715 	struct umtx_pi *pi;
1716 
1717 	pi = uma_zalloc(umtx_pi_zone, M_ZERO | flags);
1718 	TAILQ_INIT(&pi->pi_blocked);
1719 	atomic_add_int(&umtx_pi_allocated, 1);
1720 	return (pi);
1721 }
1722 
1723 void
1724 umtx_pi_free(struct umtx_pi *pi)
1725 {
1726 	uma_zfree(umtx_pi_zone, pi);
1727 	atomic_add_int(&umtx_pi_allocated, -1);
1728 }
1729 
1730 /*
1731  * Adjust the thread's position on a pi_state after its priority has been
1732  * changed.
1733  */
1734 static int
1735 umtx_pi_adjust_thread(struct umtx_pi *pi, struct thread *td)
1736 {
1737 	struct umtx_q *uq, *uq1, *uq2;
1738 	struct thread *td1;
1739 
1740 	mtx_assert(&umtx_lock, MA_OWNED);
1741 	if (pi == NULL)
1742 		return (0);
1743 
1744 	uq = td->td_umtxq;
1745 
1746 	/*
1747 	 * Check if the thread needs to be moved on the blocked chain.
1748 	 * It needs to be moved if either its priority is lower than
1749 	 * the previous thread or higher than the next thread.
1750 	 */
1751 	uq1 = TAILQ_PREV(uq, umtxq_head, uq_lockq);
1752 	uq2 = TAILQ_NEXT(uq, uq_lockq);
1753 	if ((uq1 != NULL && UPRI(td) < UPRI(uq1->uq_thread)) ||
1754 	    (uq2 != NULL && UPRI(td) > UPRI(uq2->uq_thread))) {
1755 		/*
1756 		 * Remove thread from blocked chain and determine where
1757 		 * it should be moved to.
1758 		 */
1759 		TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1760 		TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1761 			td1 = uq1->uq_thread;
1762 			MPASS(td1->td_proc->p_magic == P_MAGIC);
1763 			if (UPRI(td1) > UPRI(td))
1764 				break;
1765 		}
1766 
1767 		if (uq1 == NULL)
1768 			TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1769 		else
1770 			TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1771 	}
1772 	return (1);
1773 }
1774 
1775 static struct umtx_pi *
1776 umtx_pi_next(struct umtx_pi *pi)
1777 {
1778 	struct umtx_q *uq_owner;
1779 
1780 	if (pi->pi_owner == NULL)
1781 		return (NULL);
1782 	uq_owner = pi->pi_owner->td_umtxq;
1783 	if (uq_owner == NULL)
1784 		return (NULL);
1785 	return (uq_owner->uq_pi_blocked);
1786 }
1787 
1788 /*
1789  * Floyd's Cycle-Finding Algorithm.
1790  */
1791 static bool
1792 umtx_pi_check_loop(struct umtx_pi *pi)
1793 {
1794 	struct umtx_pi *pi1;	/* fast iterator */
1795 
1796 	mtx_assert(&umtx_lock, MA_OWNED);
1797 	if (pi == NULL)
1798 		return (false);
1799 	pi1 = pi;
1800 	for (;;) {
1801 		pi = umtx_pi_next(pi);
1802 		if (pi == NULL)
1803 			break;
1804 		pi1 = umtx_pi_next(pi1);
1805 		if (pi1 == NULL)
1806 			break;
1807 		pi1 = umtx_pi_next(pi1);
1808 		if (pi1 == NULL)
1809 			break;
1810 		if (pi == pi1)
1811 			return (true);
1812 	}
1813 	return (false);
1814 }
1815 
1816 /*
1817  * Propagate priority when a thread is blocked on POSIX
1818  * PI mutex.
1819  */
1820 static void
1821 umtx_propagate_priority(struct thread *td)
1822 {
1823 	struct umtx_q *uq;
1824 	struct umtx_pi *pi;
1825 	int pri;
1826 
1827 	mtx_assert(&umtx_lock, MA_OWNED);
1828 	pri = UPRI(td);
1829 	uq = td->td_umtxq;
1830 	pi = uq->uq_pi_blocked;
1831 	if (pi == NULL)
1832 		return;
1833 	if (umtx_pi_check_loop(pi))
1834 		return;
1835 
1836 	for (;;) {
1837 		td = pi->pi_owner;
1838 		if (td == NULL || td == curthread)
1839 			return;
1840 
1841 		MPASS(td->td_proc != NULL);
1842 		MPASS(td->td_proc->p_magic == P_MAGIC);
1843 
1844 		thread_lock(td);
1845 		if (td->td_lend_user_pri > pri)
1846 			sched_lend_user_prio(td, pri);
1847 		else {
1848 			thread_unlock(td);
1849 			break;
1850 		}
1851 		thread_unlock(td);
1852 
1853 		/*
1854 		 * Pick up the lock that td is blocked on.
1855 		 */
1856 		uq = td->td_umtxq;
1857 		pi = uq->uq_pi_blocked;
1858 		if (pi == NULL)
1859 			break;
1860 		/* Resort td on the list if needed. */
1861 		umtx_pi_adjust_thread(pi, td);
1862 	}
1863 }
1864 
1865 /*
1866  * Unpropagate priority for a PI mutex when a thread blocked on
1867  * it is interrupted by signal or resumed by others.
1868  */
1869 static void
1870 umtx_repropagate_priority(struct umtx_pi *pi)
1871 {
1872 	struct umtx_q *uq, *uq_owner;
1873 	struct umtx_pi *pi2;
1874 	int pri;
1875 
1876 	mtx_assert(&umtx_lock, MA_OWNED);
1877 
1878 	if (umtx_pi_check_loop(pi))
1879 		return;
1880 	while (pi != NULL && pi->pi_owner != NULL) {
1881 		pri = PRI_MAX;
1882 		uq_owner = pi->pi_owner->td_umtxq;
1883 
1884 		TAILQ_FOREACH(pi2, &uq_owner->uq_pi_contested, pi_link) {
1885 			uq = TAILQ_FIRST(&pi2->pi_blocked);
1886 			if (uq != NULL) {
1887 				if (pri > UPRI(uq->uq_thread))
1888 					pri = UPRI(uq->uq_thread);
1889 			}
1890 		}
1891 
1892 		if (pri > uq_owner->uq_inherited_pri)
1893 			pri = uq_owner->uq_inherited_pri;
1894 		thread_lock(pi->pi_owner);
1895 		sched_lend_user_prio(pi->pi_owner, pri);
1896 		thread_unlock(pi->pi_owner);
1897 		if ((pi = uq_owner->uq_pi_blocked) != NULL)
1898 			umtx_pi_adjust_thread(pi, uq_owner->uq_thread);
1899 	}
1900 }
1901 
1902 /*
1903  * Insert a PI mutex into owned list.
1904  */
1905 static void
1906 umtx_pi_setowner(struct umtx_pi *pi, struct thread *owner)
1907 {
1908 	struct umtx_q *uq_owner;
1909 
1910 	uq_owner = owner->td_umtxq;
1911 	mtx_assert(&umtx_lock, MA_OWNED);
1912 	MPASS(pi->pi_owner == NULL);
1913 	pi->pi_owner = owner;
1914 	TAILQ_INSERT_TAIL(&uq_owner->uq_pi_contested, pi, pi_link);
1915 }
1916 
1917 /*
1918  * Disown a PI mutex, and remove it from the owned list.
1919  */
1920 static void
1921 umtx_pi_disown(struct umtx_pi *pi)
1922 {
1923 
1924 	mtx_assert(&umtx_lock, MA_OWNED);
1925 	TAILQ_REMOVE(&pi->pi_owner->td_umtxq->uq_pi_contested, pi, pi_link);
1926 	pi->pi_owner = NULL;
1927 }
1928 
1929 /*
1930  * Claim ownership of a PI mutex.
1931  */
1932 int
1933 umtx_pi_claim(struct umtx_pi *pi, struct thread *owner)
1934 {
1935 	struct umtx_q *uq;
1936 	int pri;
1937 
1938 	mtx_lock(&umtx_lock);
1939 	if (pi->pi_owner == owner) {
1940 		mtx_unlock(&umtx_lock);
1941 		return (0);
1942 	}
1943 
1944 	if (pi->pi_owner != NULL) {
1945 		/*
1946 		 * userland may have already messed the mutex, sigh.
1947 		 */
1948 		mtx_unlock(&umtx_lock);
1949 		return (EPERM);
1950 	}
1951 	umtx_pi_setowner(pi, owner);
1952 	uq = TAILQ_FIRST(&pi->pi_blocked);
1953 	if (uq != NULL) {
1954 		pri = UPRI(uq->uq_thread);
1955 		thread_lock(owner);
1956 		if (pri < UPRI(owner))
1957 			sched_lend_user_prio(owner, pri);
1958 		thread_unlock(owner);
1959 	}
1960 	mtx_unlock(&umtx_lock);
1961 	return (0);
1962 }
1963 
1964 /*
1965  * Adjust a thread's order position in its blocked PI mutex,
1966  * this may result new priority propagating process.
1967  */
1968 void
1969 umtx_pi_adjust(struct thread *td, u_char oldpri)
1970 {
1971 	struct umtx_q *uq;
1972 	struct umtx_pi *pi;
1973 
1974 	uq = td->td_umtxq;
1975 	mtx_lock(&umtx_lock);
1976 	/*
1977 	 * Pick up the lock that td is blocked on.
1978 	 */
1979 	pi = uq->uq_pi_blocked;
1980 	if (pi != NULL) {
1981 		umtx_pi_adjust_thread(pi, td);
1982 		umtx_repropagate_priority(pi);
1983 	}
1984 	mtx_unlock(&umtx_lock);
1985 }
1986 
1987 /*
1988  * Sleep on a PI mutex.
1989  */
1990 int
1991 umtxq_sleep_pi(struct umtx_q *uq, struct umtx_pi *pi, uint32_t owner,
1992     const char *wmesg, struct umtx_abs_timeout *timo, bool shared)
1993 {
1994 	struct thread *td, *td1;
1995 	struct umtx_q *uq1;
1996 	int error, pri;
1997 #ifdef INVARIANTS
1998 	struct umtxq_chain *uc;
1999 
2000 	uc = umtxq_getchain(&pi->pi_key);
2001 #endif
2002 	error = 0;
2003 	td = uq->uq_thread;
2004 	KASSERT(td == curthread, ("inconsistent uq_thread"));
2005 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
2006 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));
2007 	umtxq_insert(uq);
2008 	mtx_lock(&umtx_lock);
2009 	if (pi->pi_owner == NULL) {
2010 		mtx_unlock(&umtx_lock);
2011 		td1 = tdfind(owner, shared ? -1 : td->td_proc->p_pid);
2012 		mtx_lock(&umtx_lock);
2013 		if (td1 != NULL) {
2014 			if (pi->pi_owner == NULL)
2015 				umtx_pi_setowner(pi, td1);
2016 			PROC_UNLOCK(td1->td_proc);
2017 		}
2018 	}
2019 
2020 	TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
2021 		pri = UPRI(uq1->uq_thread);
2022 		if (pri > UPRI(td))
2023 			break;
2024 	}
2025 
2026 	if (uq1 != NULL)
2027 		TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
2028 	else
2029 		TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
2030 
2031 	uq->uq_pi_blocked = pi;
2032 	thread_lock(td);
2033 	td->td_flags |= TDF_UPIBLOCKED;
2034 	thread_unlock(td);
2035 	umtx_propagate_priority(td);
2036 	mtx_unlock(&umtx_lock);
2037 	umtxq_unbusy(&uq->uq_key);
2038 
2039 	error = umtxq_sleep(uq, wmesg, timo);
2040 	umtxq_remove(uq);
2041 
2042 	mtx_lock(&umtx_lock);
2043 	uq->uq_pi_blocked = NULL;
2044 	thread_lock(td);
2045 	td->td_flags &= ~TDF_UPIBLOCKED;
2046 	thread_unlock(td);
2047 	TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
2048 	umtx_repropagate_priority(pi);
2049 	mtx_unlock(&umtx_lock);
2050 	umtxq_unlock(&uq->uq_key);
2051 
2052 	return (error);
2053 }
2054 
2055 /*
2056  * Add reference count for a PI mutex.
2057  */
2058 void
2059 umtx_pi_ref(struct umtx_pi *pi)
2060 {
2061 
2062 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&pi->pi_key));
2063 	pi->pi_refcount++;
2064 }
2065 
2066 /*
2067  * Decrease reference count for a PI mutex, if the counter
2068  * is decreased to zero, its memory space is freed.
2069  */
2070 void
2071 umtx_pi_unref(struct umtx_pi *pi)
2072 {
2073 	struct umtxq_chain *uc;
2074 
2075 	uc = umtxq_getchain(&pi->pi_key);
2076 	UMTXQ_LOCKED_ASSERT(uc);
2077 	KASSERT(pi->pi_refcount > 0, ("invalid reference count"));
2078 	if (--pi->pi_refcount == 0) {
2079 		mtx_lock(&umtx_lock);
2080 		if (pi->pi_owner != NULL)
2081 			umtx_pi_disown(pi);
2082 		KASSERT(TAILQ_EMPTY(&pi->pi_blocked),
2083 			("blocked queue not empty"));
2084 		mtx_unlock(&umtx_lock);
2085 		TAILQ_REMOVE(&uc->uc_pi_list, pi, pi_hashlink);
2086 		umtx_pi_free(pi);
2087 	}
2088 }
2089 
2090 /*
2091  * Find a PI mutex in hash table.
2092  */
2093 struct umtx_pi *
2094 umtx_pi_lookup(struct umtx_key *key)
2095 {
2096 	struct umtxq_chain *uc;
2097 	struct umtx_pi *pi;
2098 
2099 	uc = umtxq_getchain(key);
2100 	UMTXQ_LOCKED_ASSERT(uc);
2101 
2102 	TAILQ_FOREACH(pi, &uc->uc_pi_list, pi_hashlink) {
2103 		if (umtx_key_match(&pi->pi_key, key)) {
2104 			return (pi);
2105 		}
2106 	}
2107 	return (NULL);
2108 }
2109 
2110 /*
2111  * Insert a PI mutex into hash table.
2112  */
2113 void
2114 umtx_pi_insert(struct umtx_pi *pi)
2115 {
2116 	struct umtxq_chain *uc;
2117 
2118 	uc = umtxq_getchain(&pi->pi_key);
2119 	UMTXQ_LOCKED_ASSERT(uc);
2120 	TAILQ_INSERT_TAIL(&uc->uc_pi_list, pi, pi_hashlink);
2121 }
2122 
2123 /*
2124  * Drop a PI mutex and wakeup a top waiter.
2125  */
2126 int
2127 umtx_pi_drop(struct thread *td, struct umtx_key *key, bool rb, int *count)
2128 {
2129 	struct umtx_q *uq_first, *uq_first2, *uq_me;
2130 	struct umtx_pi *pi, *pi2;
2131 	int pri;
2132 
2133 	UMTXQ_ASSERT_LOCKED_BUSY(key);
2134 	*count = umtxq_count_pi(key, &uq_first);
2135 	if (uq_first != NULL) {
2136 		mtx_lock(&umtx_lock);
2137 		pi = uq_first->uq_pi_blocked;
2138 		KASSERT(pi != NULL, ("pi == NULL?"));
2139 		if (pi->pi_owner != td && !(rb && pi->pi_owner == NULL)) {
2140 			mtx_unlock(&umtx_lock);
2141 			/* userland messed the mutex */
2142 			return (EPERM);
2143 		}
2144 		uq_me = td->td_umtxq;
2145 		if (pi->pi_owner == td)
2146 			umtx_pi_disown(pi);
2147 		/* get highest priority thread which is still sleeping. */
2148 		uq_first = TAILQ_FIRST(&pi->pi_blocked);
2149 		while (uq_first != NULL &&
2150 		    (uq_first->uq_flags & UQF_UMTXQ) == 0) {
2151 			uq_first = TAILQ_NEXT(uq_first, uq_lockq);
2152 		}
2153 		pri = PRI_MAX;
2154 		TAILQ_FOREACH(pi2, &uq_me->uq_pi_contested, pi_link) {
2155 			uq_first2 = TAILQ_FIRST(&pi2->pi_blocked);
2156 			if (uq_first2 != NULL) {
2157 				if (pri > UPRI(uq_first2->uq_thread))
2158 					pri = UPRI(uq_first2->uq_thread);
2159 			}
2160 		}
2161 		thread_lock(td);
2162 		sched_lend_user_prio(td, pri);
2163 		thread_unlock(td);
2164 		mtx_unlock(&umtx_lock);
2165 		if (uq_first)
2166 			umtxq_signal_thread(uq_first);
2167 	} else {
2168 		pi = umtx_pi_lookup(key);
2169 		/*
2170 		 * A umtx_pi can exist if a signal or timeout removed the
2171 		 * last waiter from the umtxq, but there is still
2172 		 * a thread in do_lock_pi() holding the umtx_pi.
2173 		 */
2174 		if (pi != NULL) {
2175 			/*
2176 			 * The umtx_pi can be unowned, such as when a thread
2177 			 * has just entered do_lock_pi(), allocated the
2178 			 * umtx_pi, and unlocked the umtxq.
2179 			 * If the current thread owns it, it must disown it.
2180 			 */
2181 			mtx_lock(&umtx_lock);
2182 			if (pi->pi_owner == td)
2183 				umtx_pi_disown(pi);
2184 			mtx_unlock(&umtx_lock);
2185 		}
2186 	}
2187 	return (0);
2188 }
2189 
2190 /*
2191  * Lock a PI mutex.
2192  */
2193 static int
2194 do_lock_pi(struct thread *td, struct umutex *m, uint32_t flags,
2195     struct _umtx_time *timeout, int try)
2196 {
2197 	struct umtx_abs_timeout timo;
2198 	struct umtx_q *uq;
2199 	struct umtx_pi *pi, *new_pi;
2200 	uint32_t id, old_owner, owner, old;
2201 	int error, rv;
2202 
2203 	id = td->td_tid;
2204 	uq = td->td_umtxq;
2205 
2206 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2207 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2208 	    &uq->uq_key)) != 0)
2209 		return (error);
2210 
2211 	if (timeout != NULL)
2212 		umtx_abs_timeout_init2(&timo, timeout);
2213 
2214 	umtxq_lock(&uq->uq_key);
2215 	pi = umtx_pi_lookup(&uq->uq_key);
2216 	if (pi == NULL) {
2217 		new_pi = umtx_pi_alloc(M_NOWAIT);
2218 		if (new_pi == NULL) {
2219 			umtxq_unlock(&uq->uq_key);
2220 			new_pi = umtx_pi_alloc(M_WAITOK);
2221 			umtxq_lock(&uq->uq_key);
2222 			pi = umtx_pi_lookup(&uq->uq_key);
2223 			if (pi != NULL) {
2224 				umtx_pi_free(new_pi);
2225 				new_pi = NULL;
2226 			}
2227 		}
2228 		if (new_pi != NULL) {
2229 			new_pi->pi_key = uq->uq_key;
2230 			umtx_pi_insert(new_pi);
2231 			pi = new_pi;
2232 		}
2233 	}
2234 	umtx_pi_ref(pi);
2235 	umtxq_unlock(&uq->uq_key);
2236 
2237 	/*
2238 	 * Care must be exercised when dealing with umtx structure.  It
2239 	 * can fault on any access.
2240 	 */
2241 	for (;;) {
2242 		/*
2243 		 * Try the uncontested case.  This should be done in userland.
2244 		 */
2245 		rv = casueword32(&m->m_owner, UMUTEX_UNOWNED, &owner, id);
2246 		/* The address was invalid. */
2247 		if (rv == -1) {
2248 			error = EFAULT;
2249 			break;
2250 		}
2251 		/* The acquire succeeded. */
2252 		if (rv == 0) {
2253 			MPASS(owner == UMUTEX_UNOWNED);
2254 			error = 0;
2255 			break;
2256 		}
2257 
2258 		if (owner == UMUTEX_RB_NOTRECOV) {
2259 			error = ENOTRECOVERABLE;
2260 			break;
2261 		}
2262 
2263 		/*
2264 		 * Avoid overwriting a possible error from sleep due
2265 		 * to the pending signal with suspension check result.
2266 		 */
2267 		if (error == 0) {
2268 			error = thread_check_susp(td, true);
2269 			if (error != 0)
2270 				break;
2271 		}
2272 
2273 		/* If no one owns it but it is contested try to acquire it. */
2274 		if (owner == UMUTEX_CONTESTED || owner == UMUTEX_RB_OWNERDEAD) {
2275 			old_owner = owner;
2276 			rv = casueword32(&m->m_owner, owner, &owner,
2277 			    id | UMUTEX_CONTESTED);
2278 			/* The address was invalid. */
2279 			if (rv == -1) {
2280 				error = EFAULT;
2281 				break;
2282 			}
2283 			if (rv == 1) {
2284 				if (error == 0) {
2285 					error = thread_check_susp(td, true);
2286 					if (error != 0)
2287 						break;
2288 				}
2289 
2290 				/*
2291 				 * If this failed the lock could
2292 				 * changed, restart.
2293 				 */
2294 				continue;
2295 			}
2296 
2297 			MPASS(rv == 0);
2298 			MPASS(owner == old_owner);
2299 			umtxq_lock(&uq->uq_key);
2300 			umtxq_busy(&uq->uq_key);
2301 			error = umtx_pi_claim(pi, td);
2302 			umtxq_unbusy(&uq->uq_key);
2303 			umtxq_unlock(&uq->uq_key);
2304 			if (error != 0) {
2305 				/*
2306 				 * Since we're going to return an
2307 				 * error, restore the m_owner to its
2308 				 * previous, unowned state to avoid
2309 				 * compounding the problem.
2310 				 */
2311 				(void)casuword32(&m->m_owner,
2312 				    id | UMUTEX_CONTESTED, old_owner);
2313 			}
2314 			if (error == 0 && old_owner == UMUTEX_RB_OWNERDEAD)
2315 				error = EOWNERDEAD;
2316 			break;
2317 		}
2318 
2319 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2320 			error = EDEADLK;
2321 			break;
2322 		}
2323 
2324 		if (try != 0) {
2325 			error = EBUSY;
2326 			break;
2327 		}
2328 
2329 		/*
2330 		 * If we caught a signal, we have retried and now
2331 		 * exit immediately.
2332 		 */
2333 		if (error != 0)
2334 			break;
2335 
2336 		umtxq_lock(&uq->uq_key);
2337 		umtxq_busy(&uq->uq_key);
2338 		umtxq_unlock(&uq->uq_key);
2339 
2340 		/*
2341 		 * Set the contested bit so that a release in user space
2342 		 * knows to use the system call for unlock.  If this fails
2343 		 * either some one else has acquired the lock or it has been
2344 		 * released.
2345 		 */
2346 		rv = casueword32(&m->m_owner, owner, &old, owner |
2347 		    UMUTEX_CONTESTED);
2348 
2349 		/* The address was invalid. */
2350 		if (rv == -1) {
2351 			umtxq_unbusy_unlocked(&uq->uq_key);
2352 			error = EFAULT;
2353 			break;
2354 		}
2355 		if (rv == 1) {
2356 			umtxq_unbusy_unlocked(&uq->uq_key);
2357 			error = thread_check_susp(td, true);
2358 			if (error != 0)
2359 				break;
2360 
2361 			/*
2362 			 * The lock changed and we need to retry or we
2363 			 * lost a race to the thread unlocking the
2364 			 * umtx.  Note that the UMUTEX_RB_OWNERDEAD
2365 			 * value for owner is impossible there.
2366 			 */
2367 			continue;
2368 		}
2369 
2370 		umtxq_lock(&uq->uq_key);
2371 
2372 		/* We set the contested bit, sleep. */
2373 		MPASS(old == owner);
2374 		error = umtxq_sleep_pi(uq, pi, owner & ~UMUTEX_CONTESTED,
2375 		    "umtxpi", timeout == NULL ? NULL : &timo,
2376 		    (flags & USYNC_PROCESS_SHARED) != 0);
2377 		if (error != 0)
2378 			continue;
2379 
2380 		error = thread_check_susp(td, false);
2381 		if (error != 0)
2382 			break;
2383 	}
2384 
2385 	umtxq_lock(&uq->uq_key);
2386 	umtx_pi_unref(pi);
2387 	umtxq_unlock(&uq->uq_key);
2388 
2389 	umtx_key_release(&uq->uq_key);
2390 	return (error);
2391 }
2392 
2393 /*
2394  * Unlock a PI mutex.
2395  */
2396 static int
2397 do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2398 {
2399 	struct umtx_key key;
2400 	uint32_t id, new_owner, old, owner;
2401 	int count, error;
2402 
2403 	id = td->td_tid;
2404 
2405 usrloop:
2406 	/*
2407 	 * Make sure we own this mtx.
2408 	 */
2409 	error = fueword32(&m->m_owner, &owner);
2410 	if (error == -1)
2411 		return (EFAULT);
2412 
2413 	if ((owner & ~UMUTEX_CONTESTED) != id)
2414 		return (EPERM);
2415 
2416 	new_owner = umtx_unlock_val(flags, rb);
2417 
2418 	/* This should be done in userland */
2419 	if ((owner & UMUTEX_CONTESTED) == 0) {
2420 		error = casueword32(&m->m_owner, owner, &old, new_owner);
2421 		if (error == -1)
2422 			return (EFAULT);
2423 		if (error == 1) {
2424 			error = thread_check_susp(td, true);
2425 			if (error != 0)
2426 				return (error);
2427 			goto usrloop;
2428 		}
2429 		if (old == owner)
2430 			return (0);
2431 		owner = old;
2432 	}
2433 
2434 	/* We should only ever be in here for contested locks */
2435 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2436 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2437 	    &key)) != 0)
2438 		return (error);
2439 
2440 	umtxq_lock(&key);
2441 	umtxq_busy(&key);
2442 	error = umtx_pi_drop(td, &key, rb, &count);
2443 	if (error != 0) {
2444 		umtxq_unbusy(&key);
2445 		umtxq_unlock(&key);
2446 		umtx_key_release(&key);
2447 		/* userland messed the mutex */
2448 		return (error);
2449 	}
2450 	umtxq_unlock(&key);
2451 
2452 	/*
2453 	 * When unlocking the umtx, it must be marked as unowned if
2454 	 * there is zero or one thread only waiting for it.
2455 	 * Otherwise, it must be marked as contested.
2456 	 */
2457 
2458 	if (count > 1)
2459 		new_owner |= UMUTEX_CONTESTED;
2460 again:
2461 	error = casueword32(&m->m_owner, owner, &old, new_owner);
2462 	if (error == 1) {
2463 		error = thread_check_susp(td, false);
2464 		if (error == 0)
2465 			goto again;
2466 	}
2467 	umtxq_unbusy_unlocked(&key);
2468 	umtx_key_release(&key);
2469 	if (error == -1)
2470 		return (EFAULT);
2471 	if (error == 0 && old != owner)
2472 		return (EINVAL);
2473 	return (error);
2474 }
2475 
2476 /*
2477  * Lock a PP mutex.
2478  */
2479 static int
2480 do_lock_pp(struct thread *td, struct umutex *m, uint32_t flags,
2481     struct _umtx_time *timeout, int try)
2482 {
2483 	struct umtx_abs_timeout timo;
2484 	struct umtx_q *uq, *uq2;
2485 	struct umtx_pi *pi;
2486 	uint32_t ceiling;
2487 	uint32_t owner, id;
2488 	int error, pri, old_inherited_pri, su, rv;
2489 
2490 	id = td->td_tid;
2491 	uq = td->td_umtxq;
2492 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2493 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2494 	    &uq->uq_key)) != 0)
2495 		return (error);
2496 
2497 	if (timeout != NULL)
2498 		umtx_abs_timeout_init2(&timo, timeout);
2499 
2500 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2501 	for (;;) {
2502 		old_inherited_pri = uq->uq_inherited_pri;
2503 		umtxq_lock(&uq->uq_key);
2504 		umtxq_busy(&uq->uq_key);
2505 		umtxq_unlock(&uq->uq_key);
2506 
2507 		rv = fueword32(&m->m_ceilings[0], &ceiling);
2508 		if (rv == -1) {
2509 			error = EFAULT;
2510 			goto out;
2511 		}
2512 		ceiling = RTP_PRIO_MAX - ceiling;
2513 		if (ceiling > RTP_PRIO_MAX) {
2514 			error = EINVAL;
2515 			goto out;
2516 		}
2517 
2518 		mtx_lock(&umtx_lock);
2519 		if (UPRI(td) < PRI_MIN_REALTIME + ceiling) {
2520 			mtx_unlock(&umtx_lock);
2521 			error = EINVAL;
2522 			goto out;
2523 		}
2524 		if (su && PRI_MIN_REALTIME + ceiling < uq->uq_inherited_pri) {
2525 			uq->uq_inherited_pri = PRI_MIN_REALTIME + ceiling;
2526 			thread_lock(td);
2527 			if (uq->uq_inherited_pri < UPRI(td))
2528 				sched_lend_user_prio(td, uq->uq_inherited_pri);
2529 			thread_unlock(td);
2530 		}
2531 		mtx_unlock(&umtx_lock);
2532 
2533 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2534 		    id | UMUTEX_CONTESTED);
2535 		/* The address was invalid. */
2536 		if (rv == -1) {
2537 			error = EFAULT;
2538 			break;
2539 		}
2540 		if (rv == 0) {
2541 			MPASS(owner == UMUTEX_CONTESTED);
2542 			error = 0;
2543 			break;
2544 		}
2545 		/* rv == 1 */
2546 		if (owner == UMUTEX_RB_OWNERDEAD) {
2547 			rv = casueword32(&m->m_owner, UMUTEX_RB_OWNERDEAD,
2548 			    &owner, id | UMUTEX_CONTESTED);
2549 			if (rv == -1) {
2550 				error = EFAULT;
2551 				break;
2552 			}
2553 			if (rv == 0) {
2554 				MPASS(owner == UMUTEX_RB_OWNERDEAD);
2555 				error = EOWNERDEAD; /* success */
2556 				break;
2557 			}
2558 
2559 			/*
2560 			 *  rv == 1, only check for suspension if we
2561 			 *  did not already catched a signal.  If we
2562 			 *  get an error from the check, the same
2563 			 *  condition is checked by the umtxq_sleep()
2564 			 *  call below, so we should obliterate the
2565 			 *  error to not skip the last loop iteration.
2566 			 */
2567 			if (error == 0) {
2568 				error = thread_check_susp(td, false);
2569 				if (error == 0) {
2570 					if (try != 0)
2571 						error = EBUSY;
2572 					else
2573 						continue;
2574 				}
2575 				error = 0;
2576 			}
2577 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2578 			error = ENOTRECOVERABLE;
2579 		}
2580 
2581 		if (try != 0)
2582 			error = EBUSY;
2583 
2584 		/*
2585 		 * If we caught a signal, we have retried and now
2586 		 * exit immediately.
2587 		 */
2588 		if (error != 0)
2589 			break;
2590 
2591 		umtxq_lock(&uq->uq_key);
2592 		umtxq_insert(uq);
2593 		umtxq_unbusy(&uq->uq_key);
2594 		error = umtxq_sleep(uq, "umtxpp", timeout == NULL ?
2595 		    NULL : &timo);
2596 		umtxq_remove(uq);
2597 		umtxq_unlock(&uq->uq_key);
2598 
2599 		mtx_lock(&umtx_lock);
2600 		uq->uq_inherited_pri = old_inherited_pri;
2601 		pri = PRI_MAX;
2602 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2603 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2604 			if (uq2 != NULL) {
2605 				if (pri > UPRI(uq2->uq_thread))
2606 					pri = UPRI(uq2->uq_thread);
2607 			}
2608 		}
2609 		if (pri > uq->uq_inherited_pri)
2610 			pri = uq->uq_inherited_pri;
2611 		thread_lock(td);
2612 		sched_lend_user_prio(td, pri);
2613 		thread_unlock(td);
2614 		mtx_unlock(&umtx_lock);
2615 	}
2616 
2617 	if (error != 0 && error != EOWNERDEAD) {
2618 		mtx_lock(&umtx_lock);
2619 		uq->uq_inherited_pri = old_inherited_pri;
2620 		pri = PRI_MAX;
2621 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2622 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2623 			if (uq2 != NULL) {
2624 				if (pri > UPRI(uq2->uq_thread))
2625 					pri = UPRI(uq2->uq_thread);
2626 			}
2627 		}
2628 		if (pri > uq->uq_inherited_pri)
2629 			pri = uq->uq_inherited_pri;
2630 		thread_lock(td);
2631 		sched_lend_user_prio(td, pri);
2632 		thread_unlock(td);
2633 		mtx_unlock(&umtx_lock);
2634 	}
2635 
2636 out:
2637 	umtxq_unbusy_unlocked(&uq->uq_key);
2638 	umtx_key_release(&uq->uq_key);
2639 	return (error);
2640 }
2641 
2642 /*
2643  * Unlock a PP mutex.
2644  */
2645 static int
2646 do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2647 {
2648 	struct umtx_key key;
2649 	struct umtx_q *uq, *uq2;
2650 	struct umtx_pi *pi;
2651 	uint32_t id, owner, rceiling;
2652 	int error, pri, new_inherited_pri, su;
2653 
2654 	id = td->td_tid;
2655 	uq = td->td_umtxq;
2656 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2657 
2658 	/*
2659 	 * Make sure we own this mtx.
2660 	 */
2661 	error = fueword32(&m->m_owner, &owner);
2662 	if (error == -1)
2663 		return (EFAULT);
2664 
2665 	if ((owner & ~UMUTEX_CONTESTED) != id)
2666 		return (EPERM);
2667 
2668 	error = copyin(&m->m_ceilings[1], &rceiling, sizeof(uint32_t));
2669 	if (error != 0)
2670 		return (error);
2671 
2672 	if (rceiling == -1)
2673 		new_inherited_pri = PRI_MAX;
2674 	else {
2675 		rceiling = RTP_PRIO_MAX - rceiling;
2676 		if (rceiling > RTP_PRIO_MAX)
2677 			return (EINVAL);
2678 		new_inherited_pri = PRI_MIN_REALTIME + rceiling;
2679 	}
2680 
2681 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2682 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2683 	    &key)) != 0)
2684 		return (error);
2685 	umtxq_lock(&key);
2686 	umtxq_busy(&key);
2687 	umtxq_unlock(&key);
2688 	/*
2689 	 * For priority protected mutex, always set unlocked state
2690 	 * to UMUTEX_CONTESTED, so that userland always enters kernel
2691 	 * to lock the mutex, it is necessary because thread priority
2692 	 * has to be adjusted for such mutex.
2693 	 */
2694 	error = suword32(&m->m_owner, umtx_unlock_val(flags, rb) |
2695 	    UMUTEX_CONTESTED);
2696 
2697 	umtxq_lock(&key);
2698 	if (error == 0)
2699 		umtxq_signal(&key, 1);
2700 	umtxq_unbusy(&key);
2701 	umtxq_unlock(&key);
2702 
2703 	if (error == -1)
2704 		error = EFAULT;
2705 	else {
2706 		mtx_lock(&umtx_lock);
2707 		if (su != 0)
2708 			uq->uq_inherited_pri = new_inherited_pri;
2709 		pri = PRI_MAX;
2710 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2711 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2712 			if (uq2 != NULL) {
2713 				if (pri > UPRI(uq2->uq_thread))
2714 					pri = UPRI(uq2->uq_thread);
2715 			}
2716 		}
2717 		if (pri > uq->uq_inherited_pri)
2718 			pri = uq->uq_inherited_pri;
2719 		thread_lock(td);
2720 		sched_lend_user_prio(td, pri);
2721 		thread_unlock(td);
2722 		mtx_unlock(&umtx_lock);
2723 	}
2724 	umtx_key_release(&key);
2725 	return (error);
2726 }
2727 
2728 static int
2729 do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
2730     uint32_t *old_ceiling)
2731 {
2732 	struct umtx_q *uq;
2733 	uint32_t flags, id, owner, save_ceiling;
2734 	int error, rv, rv1;
2735 
2736 	error = fueword32(&m->m_flags, &flags);
2737 	if (error == -1)
2738 		return (EFAULT);
2739 	if ((flags & UMUTEX_PRIO_PROTECT) == 0)
2740 		return (EINVAL);
2741 	if (ceiling > RTP_PRIO_MAX)
2742 		return (EINVAL);
2743 	id = td->td_tid;
2744 	uq = td->td_umtxq;
2745 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2746 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2747 	    &uq->uq_key)) != 0)
2748 		return (error);
2749 	for (;;) {
2750 		umtxq_lock(&uq->uq_key);
2751 		umtxq_busy(&uq->uq_key);
2752 		umtxq_unlock(&uq->uq_key);
2753 
2754 		rv = fueword32(&m->m_ceilings[0], &save_ceiling);
2755 		if (rv == -1) {
2756 			error = EFAULT;
2757 			break;
2758 		}
2759 
2760 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2761 		    id | UMUTEX_CONTESTED);
2762 		if (rv == -1) {
2763 			error = EFAULT;
2764 			break;
2765 		}
2766 
2767 		if (rv == 0) {
2768 			MPASS(owner == UMUTEX_CONTESTED);
2769 			rv = suword32(&m->m_ceilings[0], ceiling);
2770 			rv1 = suword32(&m->m_owner, UMUTEX_CONTESTED);
2771 			error = (rv == 0 && rv1 == 0) ? 0: EFAULT;
2772 			break;
2773 		}
2774 
2775 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2776 			rv = suword32(&m->m_ceilings[0], ceiling);
2777 			error = rv == 0 ? 0 : EFAULT;
2778 			break;
2779 		}
2780 
2781 		if (owner == UMUTEX_RB_OWNERDEAD) {
2782 			error = EOWNERDEAD;
2783 			break;
2784 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2785 			error = ENOTRECOVERABLE;
2786 			break;
2787 		}
2788 
2789 		/*
2790 		 * If we caught a signal, we have retried and now
2791 		 * exit immediately.
2792 		 */
2793 		if (error != 0)
2794 			break;
2795 
2796 		/*
2797 		 * We set the contested bit, sleep. Otherwise the lock changed
2798 		 * and we need to retry or we lost a race to the thread
2799 		 * unlocking the umtx.
2800 		 */
2801 		umtxq_lock(&uq->uq_key);
2802 		umtxq_insert(uq);
2803 		umtxq_unbusy(&uq->uq_key);
2804 		error = umtxq_sleep(uq, "umtxpp", NULL);
2805 		umtxq_remove(uq);
2806 		umtxq_unlock(&uq->uq_key);
2807 	}
2808 	umtxq_lock(&uq->uq_key);
2809 	if (error == 0)
2810 		umtxq_signal(&uq->uq_key, INT_MAX);
2811 	umtxq_unbusy(&uq->uq_key);
2812 	umtxq_unlock(&uq->uq_key);
2813 	umtx_key_release(&uq->uq_key);
2814 	if (error == 0 && old_ceiling != NULL) {
2815 		rv = suword32(old_ceiling, save_ceiling);
2816 		error = rv == 0 ? 0 : EFAULT;
2817 	}
2818 	return (error);
2819 }
2820 
2821 /*
2822  * Lock a userland POSIX mutex.
2823  */
2824 static int
2825 do_lock_umutex(struct thread *td, struct umutex *m,
2826     struct _umtx_time *timeout, int mode)
2827 {
2828 	uint32_t flags;
2829 	int error;
2830 
2831 	error = fueword32(&m->m_flags, &flags);
2832 	if (error == -1)
2833 		return (EFAULT);
2834 
2835 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2836 	case 0:
2837 		error = do_lock_normal(td, m, flags, timeout, mode);
2838 		break;
2839 	case UMUTEX_PRIO_INHERIT:
2840 		error = do_lock_pi(td, m, flags, timeout, mode);
2841 		break;
2842 	case UMUTEX_PRIO_PROTECT:
2843 		error = do_lock_pp(td, m, flags, timeout, mode);
2844 		break;
2845 	default:
2846 		return (EINVAL);
2847 	}
2848 	if (timeout == NULL) {
2849 		if (error == EINTR && mode != _UMUTEX_WAIT)
2850 			error = ERESTART;
2851 	} else {
2852 		/* Timed-locking is not restarted. */
2853 		if (error == ERESTART)
2854 			error = EINTR;
2855 	}
2856 	return (error);
2857 }
2858 
2859 /*
2860  * Unlock a userland POSIX mutex.
2861  */
2862 static int
2863 do_unlock_umutex(struct thread *td, struct umutex *m, bool rb)
2864 {
2865 	uint32_t flags;
2866 	int error;
2867 
2868 	error = fueword32(&m->m_flags, &flags);
2869 	if (error == -1)
2870 		return (EFAULT);
2871 
2872 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2873 	case 0:
2874 		return (do_unlock_normal(td, m, flags, rb));
2875 	case UMUTEX_PRIO_INHERIT:
2876 		return (do_unlock_pi(td, m, flags, rb));
2877 	case UMUTEX_PRIO_PROTECT:
2878 		return (do_unlock_pp(td, m, flags, rb));
2879 	}
2880 
2881 	return (EINVAL);
2882 }
2883 
2884 static int
2885 do_cv_wait(struct thread *td, struct ucond *cv, struct umutex *m,
2886     struct timespec *timeout, u_long wflags)
2887 {
2888 	struct umtx_abs_timeout timo;
2889 	struct umtx_q *uq;
2890 	uint32_t flags, clockid, hasw;
2891 	int error;
2892 
2893 	uq = td->td_umtxq;
2894 	error = fueword32(&cv->c_flags, &flags);
2895 	if (error == -1)
2896 		return (EFAULT);
2897 	error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &uq->uq_key);
2898 	if (error != 0)
2899 		return (error);
2900 
2901 	if ((wflags & CVWAIT_CLOCKID) != 0) {
2902 		error = fueword32(&cv->c_clockid, &clockid);
2903 		if (error == -1) {
2904 			umtx_key_release(&uq->uq_key);
2905 			return (EFAULT);
2906 		}
2907 		if (clockid < CLOCK_REALTIME ||
2908 		    clockid >= CLOCK_THREAD_CPUTIME_ID) {
2909 			/* hmm, only HW clock id will work. */
2910 			umtx_key_release(&uq->uq_key);
2911 			return (EINVAL);
2912 		}
2913 	} else {
2914 		clockid = CLOCK_REALTIME;
2915 	}
2916 
2917 	umtxq_lock(&uq->uq_key);
2918 	umtxq_busy(&uq->uq_key);
2919 	umtxq_insert(uq);
2920 	umtxq_unlock(&uq->uq_key);
2921 
2922 	/*
2923 	 * Set c_has_waiters to 1 before releasing user mutex, also
2924 	 * don't modify cache line when unnecessary.
2925 	 */
2926 	error = fueword32(&cv->c_has_waiters, &hasw);
2927 	if (error == 0 && hasw == 0)
2928 		suword32(&cv->c_has_waiters, 1);
2929 
2930 	umtxq_unbusy_unlocked(&uq->uq_key);
2931 
2932 	error = do_unlock_umutex(td, m, false);
2933 
2934 	if (timeout != NULL)
2935 		umtx_abs_timeout_init(&timo, clockid,
2936 		    (wflags & CVWAIT_ABSTIME) != 0, timeout);
2937 
2938 	umtxq_lock(&uq->uq_key);
2939 	if (error == 0) {
2940 		error = umtxq_sleep(uq, "ucond", timeout == NULL ?
2941 		    NULL : &timo);
2942 	}
2943 
2944 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
2945 		error = 0;
2946 	else {
2947 		/*
2948 		 * This must be timeout,interrupted by signal or
2949 		 * surprious wakeup, clear c_has_waiter flag when
2950 		 * necessary.
2951 		 */
2952 		umtxq_busy(&uq->uq_key);
2953 		if ((uq->uq_flags & UQF_UMTXQ) != 0) {
2954 			int oldlen = uq->uq_cur_queue->length;
2955 			umtxq_remove(uq);
2956 			if (oldlen == 1) {
2957 				umtxq_unlock(&uq->uq_key);
2958 				suword32(&cv->c_has_waiters, 0);
2959 				umtxq_lock(&uq->uq_key);
2960 			}
2961 		}
2962 		umtxq_unbusy(&uq->uq_key);
2963 		if (error == ERESTART)
2964 			error = EINTR;
2965 	}
2966 
2967 	umtxq_unlock(&uq->uq_key);
2968 	umtx_key_release(&uq->uq_key);
2969 	return (error);
2970 }
2971 
2972 /*
2973  * Signal a userland condition variable.
2974  */
2975 static int
2976 do_cv_signal(struct thread *td, struct ucond *cv)
2977 {
2978 	struct umtx_key key;
2979 	int error, cnt, nwake;
2980 	uint32_t flags;
2981 
2982 	error = fueword32(&cv->c_flags, &flags);
2983 	if (error == -1)
2984 		return (EFAULT);
2985 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
2986 		return (error);
2987 	umtxq_lock(&key);
2988 	umtxq_busy(&key);
2989 	cnt = umtxq_count(&key);
2990 	nwake = umtxq_signal(&key, 1);
2991 	if (cnt <= nwake) {
2992 		umtxq_unlock(&key);
2993 		error = suword32(&cv->c_has_waiters, 0);
2994 		if (error == -1)
2995 			error = EFAULT;
2996 		umtxq_lock(&key);
2997 	}
2998 	umtxq_unbusy(&key);
2999 	umtxq_unlock(&key);
3000 	umtx_key_release(&key);
3001 	return (error);
3002 }
3003 
3004 static int
3005 do_cv_broadcast(struct thread *td, struct ucond *cv)
3006 {
3007 	struct umtx_key key;
3008 	int error;
3009 	uint32_t flags;
3010 
3011 	error = fueword32(&cv->c_flags, &flags);
3012 	if (error == -1)
3013 		return (EFAULT);
3014 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
3015 		return (error);
3016 
3017 	umtxq_lock(&key);
3018 	umtxq_busy(&key);
3019 	umtxq_signal(&key, INT_MAX);
3020 	umtxq_unlock(&key);
3021 
3022 	error = suword32(&cv->c_has_waiters, 0);
3023 	if (error == -1)
3024 		error = EFAULT;
3025 
3026 	umtxq_unbusy_unlocked(&key);
3027 
3028 	umtx_key_release(&key);
3029 	return (error);
3030 }
3031 
3032 static int
3033 do_rw_rdlock(struct thread *td, struct urwlock *rwlock, long fflag,
3034     struct _umtx_time *timeout)
3035 {
3036 	struct umtx_abs_timeout timo;
3037 	struct umtx_q *uq;
3038 	uint32_t flags, wrflags;
3039 	int32_t state, oldstate;
3040 	int32_t blocked_readers;
3041 	int error, error1, rv;
3042 
3043 	uq = td->td_umtxq;
3044 	error = fueword32(&rwlock->rw_flags, &flags);
3045 	if (error == -1)
3046 		return (EFAULT);
3047 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3048 	if (error != 0)
3049 		return (error);
3050 
3051 	if (timeout != NULL)
3052 		umtx_abs_timeout_init2(&timo, timeout);
3053 
3054 	wrflags = URWLOCK_WRITE_OWNER;
3055 	if (!(fflag & URWLOCK_PREFER_READER) && !(flags & URWLOCK_PREFER_READER))
3056 		wrflags |= URWLOCK_WRITE_WAITERS;
3057 
3058 	for (;;) {
3059 		rv = fueword32(&rwlock->rw_state, &state);
3060 		if (rv == -1) {
3061 			umtx_key_release(&uq->uq_key);
3062 			return (EFAULT);
3063 		}
3064 
3065 		/* try to lock it */
3066 		while (!(state & wrflags)) {
3067 			if (__predict_false(URWLOCK_READER_COUNT(state) ==
3068 			    URWLOCK_MAX_READERS)) {
3069 				umtx_key_release(&uq->uq_key);
3070 				return (EAGAIN);
3071 			}
3072 			rv = casueword32(&rwlock->rw_state, state,
3073 			    &oldstate, state + 1);
3074 			if (rv == -1) {
3075 				umtx_key_release(&uq->uq_key);
3076 				return (EFAULT);
3077 			}
3078 			if (rv == 0) {
3079 				MPASS(oldstate == state);
3080 				umtx_key_release(&uq->uq_key);
3081 				return (0);
3082 			}
3083 			error = thread_check_susp(td, true);
3084 			if (error != 0)
3085 				break;
3086 			state = oldstate;
3087 		}
3088 
3089 		if (error)
3090 			break;
3091 
3092 		/* grab monitor lock */
3093 		umtxq_lock(&uq->uq_key);
3094 		umtxq_busy(&uq->uq_key);
3095 		umtxq_unlock(&uq->uq_key);
3096 
3097 		/*
3098 		 * re-read the state, in case it changed between the try-lock above
3099 		 * and the check below
3100 		 */
3101 		rv = fueword32(&rwlock->rw_state, &state);
3102 		if (rv == -1)
3103 			error = EFAULT;
3104 
3105 		/* set read contention bit */
3106 		while (error == 0 && (state & wrflags) &&
3107 		    !(state & URWLOCK_READ_WAITERS)) {
3108 			rv = casueword32(&rwlock->rw_state, state,
3109 			    &oldstate, state | URWLOCK_READ_WAITERS);
3110 			if (rv == -1) {
3111 				error = EFAULT;
3112 				break;
3113 			}
3114 			if (rv == 0) {
3115 				MPASS(oldstate == state);
3116 				goto sleep;
3117 			}
3118 			state = oldstate;
3119 			error = thread_check_susp(td, false);
3120 			if (error != 0)
3121 				break;
3122 		}
3123 		if (error != 0) {
3124 			umtxq_unbusy_unlocked(&uq->uq_key);
3125 			break;
3126 		}
3127 
3128 		/* state is changed while setting flags, restart */
3129 		if (!(state & wrflags)) {
3130 			umtxq_unbusy_unlocked(&uq->uq_key);
3131 			error = thread_check_susp(td, true);
3132 			if (error != 0)
3133 				break;
3134 			continue;
3135 		}
3136 
3137 sleep:
3138 		/*
3139 		 * Contention bit is set, before sleeping, increase
3140 		 * read waiter count.
3141 		 */
3142 		rv = fueword32(&rwlock->rw_blocked_readers,
3143 		    &blocked_readers);
3144 		if (rv == -1) {
3145 			umtxq_unbusy_unlocked(&uq->uq_key);
3146 			error = EFAULT;
3147 			break;
3148 		}
3149 		suword32(&rwlock->rw_blocked_readers, blocked_readers+1);
3150 
3151 		while (state & wrflags) {
3152 			umtxq_lock(&uq->uq_key);
3153 			umtxq_insert(uq);
3154 			umtxq_unbusy(&uq->uq_key);
3155 
3156 			error = umtxq_sleep(uq, "urdlck", timeout == NULL ?
3157 			    NULL : &timo);
3158 
3159 			umtxq_busy(&uq->uq_key);
3160 			umtxq_remove(uq);
3161 			umtxq_unlock(&uq->uq_key);
3162 			if (error)
3163 				break;
3164 			rv = fueword32(&rwlock->rw_state, &state);
3165 			if (rv == -1) {
3166 				error = EFAULT;
3167 				break;
3168 			}
3169 		}
3170 
3171 		/* decrease read waiter count, and may clear read contention bit */
3172 		rv = fueword32(&rwlock->rw_blocked_readers,
3173 		    &blocked_readers);
3174 		if (rv == -1) {
3175 			umtxq_unbusy_unlocked(&uq->uq_key);
3176 			error = EFAULT;
3177 			break;
3178 		}
3179 		suword32(&rwlock->rw_blocked_readers, blocked_readers-1);
3180 		if (blocked_readers == 1) {
3181 			rv = fueword32(&rwlock->rw_state, &state);
3182 			if (rv == -1) {
3183 				umtxq_unbusy_unlocked(&uq->uq_key);
3184 				error = EFAULT;
3185 				break;
3186 			}
3187 			for (;;) {
3188 				rv = casueword32(&rwlock->rw_state, state,
3189 				    &oldstate, state & ~URWLOCK_READ_WAITERS);
3190 				if (rv == -1) {
3191 					error = EFAULT;
3192 					break;
3193 				}
3194 				if (rv == 0) {
3195 					MPASS(oldstate == state);
3196 					break;
3197 				}
3198 				state = oldstate;
3199 				error1 = thread_check_susp(td, false);
3200 				if (error1 != 0) {
3201 					if (error == 0)
3202 						error = error1;
3203 					break;
3204 				}
3205 			}
3206 		}
3207 
3208 		umtxq_unbusy_unlocked(&uq->uq_key);
3209 		if (error != 0)
3210 			break;
3211 	}
3212 	umtx_key_release(&uq->uq_key);
3213 	if (error == ERESTART)
3214 		error = EINTR;
3215 	return (error);
3216 }
3217 
3218 static int
3219 do_rw_wrlock(struct thread *td, struct urwlock *rwlock, struct _umtx_time *timeout)
3220 {
3221 	struct umtx_abs_timeout timo;
3222 	struct umtx_q *uq;
3223 	uint32_t flags;
3224 	int32_t state, oldstate;
3225 	int32_t blocked_writers;
3226 	int32_t blocked_readers;
3227 	int error, error1, rv;
3228 
3229 	uq = td->td_umtxq;
3230 	error = fueword32(&rwlock->rw_flags, &flags);
3231 	if (error == -1)
3232 		return (EFAULT);
3233 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3234 	if (error != 0)
3235 		return (error);
3236 
3237 	if (timeout != NULL)
3238 		umtx_abs_timeout_init2(&timo, timeout);
3239 
3240 	blocked_readers = 0;
3241 	for (;;) {
3242 		rv = fueword32(&rwlock->rw_state, &state);
3243 		if (rv == -1) {
3244 			umtx_key_release(&uq->uq_key);
3245 			return (EFAULT);
3246 		}
3247 		while ((state & URWLOCK_WRITE_OWNER) == 0 &&
3248 		    URWLOCK_READER_COUNT(state) == 0) {
3249 			rv = casueword32(&rwlock->rw_state, state,
3250 			    &oldstate, state | URWLOCK_WRITE_OWNER);
3251 			if (rv == -1) {
3252 				umtx_key_release(&uq->uq_key);
3253 				return (EFAULT);
3254 			}
3255 			if (rv == 0) {
3256 				MPASS(oldstate == state);
3257 				umtx_key_release(&uq->uq_key);
3258 				return (0);
3259 			}
3260 			state = oldstate;
3261 			error = thread_check_susp(td, true);
3262 			if (error != 0)
3263 				break;
3264 		}
3265 
3266 		if (error) {
3267 			if ((state & (URWLOCK_WRITE_OWNER |
3268 			    URWLOCK_WRITE_WAITERS)) == 0 &&
3269 			    blocked_readers != 0) {
3270 				umtxq_lock(&uq->uq_key);
3271 				umtxq_busy(&uq->uq_key);
3272 				umtxq_signal_queue(&uq->uq_key, INT_MAX,
3273 				    UMTX_SHARED_QUEUE);
3274 				umtxq_unbusy(&uq->uq_key);
3275 				umtxq_unlock(&uq->uq_key);
3276 			}
3277 
3278 			break;
3279 		}
3280 
3281 		/* grab monitor lock */
3282 		umtxq_lock(&uq->uq_key);
3283 		umtxq_busy(&uq->uq_key);
3284 		umtxq_unlock(&uq->uq_key);
3285 
3286 		/*
3287 		 * Re-read the state, in case it changed between the
3288 		 * try-lock above and the check below.
3289 		 */
3290 		rv = fueword32(&rwlock->rw_state, &state);
3291 		if (rv == -1)
3292 			error = EFAULT;
3293 
3294 		while (error == 0 && ((state & URWLOCK_WRITE_OWNER) ||
3295 		    URWLOCK_READER_COUNT(state) != 0) &&
3296 		    (state & URWLOCK_WRITE_WAITERS) == 0) {
3297 			rv = casueword32(&rwlock->rw_state, state,
3298 			    &oldstate, state | URWLOCK_WRITE_WAITERS);
3299 			if (rv == -1) {
3300 				error = EFAULT;
3301 				break;
3302 			}
3303 			if (rv == 0) {
3304 				MPASS(oldstate == state);
3305 				goto sleep;
3306 			}
3307 			state = oldstate;
3308 			error = thread_check_susp(td, false);
3309 			if (error != 0)
3310 				break;
3311 		}
3312 		if (error != 0) {
3313 			umtxq_unbusy_unlocked(&uq->uq_key);
3314 			break;
3315 		}
3316 
3317 		if ((state & URWLOCK_WRITE_OWNER) == 0 &&
3318 		    URWLOCK_READER_COUNT(state) == 0) {
3319 			umtxq_unbusy_unlocked(&uq->uq_key);
3320 			error = thread_check_susp(td, false);
3321 			if (error != 0)
3322 				break;
3323 			continue;
3324 		}
3325 sleep:
3326 		rv = fueword32(&rwlock->rw_blocked_writers,
3327 		    &blocked_writers);
3328 		if (rv == -1) {
3329 			umtxq_unbusy_unlocked(&uq->uq_key);
3330 			error = EFAULT;
3331 			break;
3332 		}
3333 		suword32(&rwlock->rw_blocked_writers, blocked_writers + 1);
3334 
3335 		while ((state & URWLOCK_WRITE_OWNER) ||
3336 		    URWLOCK_READER_COUNT(state) != 0) {
3337 			umtxq_lock(&uq->uq_key);
3338 			umtxq_insert_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3339 			umtxq_unbusy(&uq->uq_key);
3340 
3341 			error = umtxq_sleep(uq, "uwrlck", timeout == NULL ?
3342 			    NULL : &timo);
3343 
3344 			umtxq_busy(&uq->uq_key);
3345 			umtxq_remove_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3346 			umtxq_unlock(&uq->uq_key);
3347 			if (error)
3348 				break;
3349 			rv = fueword32(&rwlock->rw_state, &state);
3350 			if (rv == -1) {
3351 				error = EFAULT;
3352 				break;
3353 			}
3354 		}
3355 
3356 		rv = fueword32(&rwlock->rw_blocked_writers,
3357 		    &blocked_writers);
3358 		if (rv == -1) {
3359 			umtxq_unbusy_unlocked(&uq->uq_key);
3360 			error = EFAULT;
3361 			break;
3362 		}
3363 		suword32(&rwlock->rw_blocked_writers, blocked_writers-1);
3364 		if (blocked_writers == 1) {
3365 			rv = fueword32(&rwlock->rw_state, &state);
3366 			if (rv == -1) {
3367 				umtxq_unbusy_unlocked(&uq->uq_key);
3368 				error = EFAULT;
3369 				break;
3370 			}
3371 			for (;;) {
3372 				rv = casueword32(&rwlock->rw_state, state,
3373 				    &oldstate, state & ~URWLOCK_WRITE_WAITERS);
3374 				if (rv == -1) {
3375 					error = EFAULT;
3376 					break;
3377 				}
3378 				if (rv == 0) {
3379 					MPASS(oldstate == state);
3380 					break;
3381 				}
3382 				state = oldstate;
3383 				error1 = thread_check_susp(td, false);
3384 				/*
3385 				 * We are leaving the URWLOCK_WRITE_WAITERS
3386 				 * behind, but this should not harm the
3387 				 * correctness.
3388 				 */
3389 				if (error1 != 0) {
3390 					if (error == 0)
3391 						error = error1;
3392 					break;
3393 				}
3394 			}
3395 			rv = fueword32(&rwlock->rw_blocked_readers,
3396 			    &blocked_readers);
3397 			if (rv == -1) {
3398 				umtxq_unbusy_unlocked(&uq->uq_key);
3399 				error = EFAULT;
3400 				break;
3401 			}
3402 		} else
3403 			blocked_readers = 0;
3404 
3405 		umtxq_unbusy_unlocked(&uq->uq_key);
3406 	}
3407 
3408 	umtx_key_release(&uq->uq_key);
3409 	if (error == ERESTART)
3410 		error = EINTR;
3411 	return (error);
3412 }
3413 
3414 static int
3415 do_rw_unlock(struct thread *td, struct urwlock *rwlock)
3416 {
3417 	struct umtx_q *uq;
3418 	uint32_t flags;
3419 	int32_t state, oldstate;
3420 	int error, rv, q, count;
3421 
3422 	uq = td->td_umtxq;
3423 	error = fueword32(&rwlock->rw_flags, &flags);
3424 	if (error == -1)
3425 		return (EFAULT);
3426 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3427 	if (error != 0)
3428 		return (error);
3429 
3430 	error = fueword32(&rwlock->rw_state, &state);
3431 	if (error == -1) {
3432 		error = EFAULT;
3433 		goto out;
3434 	}
3435 	if (state & URWLOCK_WRITE_OWNER) {
3436 		for (;;) {
3437 			rv = casueword32(&rwlock->rw_state, state,
3438 			    &oldstate, state & ~URWLOCK_WRITE_OWNER);
3439 			if (rv == -1) {
3440 				error = EFAULT;
3441 				goto out;
3442 			}
3443 			if (rv == 1) {
3444 				state = oldstate;
3445 				if (!(oldstate & URWLOCK_WRITE_OWNER)) {
3446 					error = EPERM;
3447 					goto out;
3448 				}
3449 				error = thread_check_susp(td, true);
3450 				if (error != 0)
3451 					goto out;
3452 			} else
3453 				break;
3454 		}
3455 	} else if (URWLOCK_READER_COUNT(state) != 0) {
3456 		for (;;) {
3457 			rv = casueword32(&rwlock->rw_state, state,
3458 			    &oldstate, state - 1);
3459 			if (rv == -1) {
3460 				error = EFAULT;
3461 				goto out;
3462 			}
3463 			if (rv == 1) {
3464 				state = oldstate;
3465 				if (URWLOCK_READER_COUNT(oldstate) == 0) {
3466 					error = EPERM;
3467 					goto out;
3468 				}
3469 				error = thread_check_susp(td, true);
3470 				if (error != 0)
3471 					goto out;
3472 			} else
3473 				break;
3474 		}
3475 	} else {
3476 		error = EPERM;
3477 		goto out;
3478 	}
3479 
3480 	count = 0;
3481 
3482 	if (!(flags & URWLOCK_PREFER_READER)) {
3483 		if (state & URWLOCK_WRITE_WAITERS) {
3484 			count = 1;
3485 			q = UMTX_EXCLUSIVE_QUEUE;
3486 		} else if (state & URWLOCK_READ_WAITERS) {
3487 			count = INT_MAX;
3488 			q = UMTX_SHARED_QUEUE;
3489 		}
3490 	} else {
3491 		if (state & URWLOCK_READ_WAITERS) {
3492 			count = INT_MAX;
3493 			q = UMTX_SHARED_QUEUE;
3494 		} else if (state & URWLOCK_WRITE_WAITERS) {
3495 			count = 1;
3496 			q = UMTX_EXCLUSIVE_QUEUE;
3497 		}
3498 	}
3499 
3500 	if (count) {
3501 		umtxq_lock(&uq->uq_key);
3502 		umtxq_busy(&uq->uq_key);
3503 		umtxq_signal_queue(&uq->uq_key, count, q);
3504 		umtxq_unbusy(&uq->uq_key);
3505 		umtxq_unlock(&uq->uq_key);
3506 	}
3507 out:
3508 	umtx_key_release(&uq->uq_key);
3509 	return (error);
3510 }
3511 
3512 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3513 static int
3514 do_sem_wait(struct thread *td, struct _usem *sem, struct _umtx_time *timeout)
3515 {
3516 	struct umtx_abs_timeout timo;
3517 	struct umtx_q *uq;
3518 	uint32_t flags, count, count1;
3519 	int error, rv, rv1;
3520 
3521 	uq = td->td_umtxq;
3522 	error = fueword32(&sem->_flags, &flags);
3523 	if (error == -1)
3524 		return (EFAULT);
3525 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3526 	if (error != 0)
3527 		return (error);
3528 
3529 	if (timeout != NULL)
3530 		umtx_abs_timeout_init2(&timo, timeout);
3531 
3532 again:
3533 	umtxq_lock(&uq->uq_key);
3534 	umtxq_busy(&uq->uq_key);
3535 	umtxq_insert(uq);
3536 	umtxq_unlock(&uq->uq_key);
3537 	rv = casueword32(&sem->_has_waiters, 0, &count1, 1);
3538 	if (rv == 0)
3539 		rv1 = fueword32(&sem->_count, &count);
3540 	if (rv == -1 || (rv == 0 && (rv1 == -1 || count != 0)) ||
3541 	    (rv == 1 && count1 == 0)) {
3542 		umtxq_lock(&uq->uq_key);
3543 		umtxq_unbusy(&uq->uq_key);
3544 		umtxq_remove(uq);
3545 		umtxq_unlock(&uq->uq_key);
3546 		if (rv == 1) {
3547 			rv = thread_check_susp(td, true);
3548 			if (rv == 0)
3549 				goto again;
3550 			error = rv;
3551 			goto out;
3552 		}
3553 		if (rv == 0)
3554 			rv = rv1;
3555 		error = rv == -1 ? EFAULT : 0;
3556 		goto out;
3557 	}
3558 	umtxq_lock(&uq->uq_key);
3559 	umtxq_unbusy(&uq->uq_key);
3560 
3561 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3562 
3563 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3564 		error = 0;
3565 	else {
3566 		umtxq_remove(uq);
3567 		/* A relative timeout cannot be restarted. */
3568 		if (error == ERESTART && timeout != NULL &&
3569 		    (timeout->_flags & UMTX_ABSTIME) == 0)
3570 			error = EINTR;
3571 	}
3572 	umtxq_unlock(&uq->uq_key);
3573 out:
3574 	umtx_key_release(&uq->uq_key);
3575 	return (error);
3576 }
3577 
3578 /*
3579  * Signal a userland semaphore.
3580  */
3581 static int
3582 do_sem_wake(struct thread *td, struct _usem *sem)
3583 {
3584 	struct umtx_key key;
3585 	int error, cnt;
3586 	uint32_t flags;
3587 
3588 	error = fueword32(&sem->_flags, &flags);
3589 	if (error == -1)
3590 		return (EFAULT);
3591 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3592 		return (error);
3593 	umtxq_lock(&key);
3594 	umtxq_busy(&key);
3595 	cnt = umtxq_count(&key);
3596 	if (cnt > 0) {
3597 		/*
3598 		 * Check if count is greater than 0, this means the memory is
3599 		 * still being referenced by user code, so we can safely
3600 		 * update _has_waiters flag.
3601 		 */
3602 		if (cnt == 1) {
3603 			umtxq_unlock(&key);
3604 			error = suword32(&sem->_has_waiters, 0);
3605 			umtxq_lock(&key);
3606 			if (error == -1)
3607 				error = EFAULT;
3608 		}
3609 		umtxq_signal(&key, 1);
3610 	}
3611 	umtxq_unbusy(&key);
3612 	umtxq_unlock(&key);
3613 	umtx_key_release(&key);
3614 	return (error);
3615 }
3616 #endif
3617 
3618 static int
3619 do_sem2_wait(struct thread *td, struct _usem2 *sem, struct _umtx_time *timeout)
3620 {
3621 	struct umtx_abs_timeout timo;
3622 	struct umtx_q *uq;
3623 	uint32_t count, flags;
3624 	int error, rv;
3625 
3626 	uq = td->td_umtxq;
3627 	flags = fuword32(&sem->_flags);
3628 	if (timeout != NULL)
3629 		umtx_abs_timeout_init2(&timo, timeout);
3630 
3631 again:
3632 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3633 	if (error != 0)
3634 		return (error);
3635 	umtxq_lock(&uq->uq_key);
3636 	umtxq_busy(&uq->uq_key);
3637 	umtxq_insert(uq);
3638 	umtxq_unlock(&uq->uq_key);
3639 	rv = fueword32(&sem->_count, &count);
3640 	if (rv == -1) {
3641 		umtxq_lock(&uq->uq_key);
3642 		umtxq_unbusy(&uq->uq_key);
3643 		umtxq_remove(uq);
3644 		umtxq_unlock(&uq->uq_key);
3645 		umtx_key_release(&uq->uq_key);
3646 		return (EFAULT);
3647 	}
3648 	for (;;) {
3649 		if (USEM_COUNT(count) != 0) {
3650 			umtxq_lock(&uq->uq_key);
3651 			umtxq_unbusy(&uq->uq_key);
3652 			umtxq_remove(uq);
3653 			umtxq_unlock(&uq->uq_key);
3654 			umtx_key_release(&uq->uq_key);
3655 			return (0);
3656 		}
3657 		if (count == USEM_HAS_WAITERS)
3658 			break;
3659 		rv = casueword32(&sem->_count, 0, &count, USEM_HAS_WAITERS);
3660 		if (rv == 0)
3661 			break;
3662 		umtxq_lock(&uq->uq_key);
3663 		umtxq_unbusy(&uq->uq_key);
3664 		umtxq_remove(uq);
3665 		umtxq_unlock(&uq->uq_key);
3666 		umtx_key_release(&uq->uq_key);
3667 		if (rv == -1)
3668 			return (EFAULT);
3669 		rv = thread_check_susp(td, true);
3670 		if (rv != 0)
3671 			return (rv);
3672 		goto again;
3673 	}
3674 	umtxq_lock(&uq->uq_key);
3675 	umtxq_unbusy(&uq->uq_key);
3676 
3677 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3678 
3679 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3680 		error = 0;
3681 	else {
3682 		umtxq_remove(uq);
3683 		if (timeout != NULL && (timeout->_flags & UMTX_ABSTIME) == 0) {
3684 			/* A relative timeout cannot be restarted. */
3685 			if (error == ERESTART)
3686 				error = EINTR;
3687 			if (error == EINTR) {
3688 				kern_clock_gettime(curthread, timo.clockid,
3689 				    &timo.cur);
3690 				timespecsub(&timo.end, &timo.cur,
3691 				    &timeout->_timeout);
3692 			}
3693 		}
3694 	}
3695 	umtxq_unlock(&uq->uq_key);
3696 	umtx_key_release(&uq->uq_key);
3697 	return (error);
3698 }
3699 
3700 /*
3701  * Signal a userland semaphore.
3702  */
3703 static int
3704 do_sem2_wake(struct thread *td, struct _usem2 *sem)
3705 {
3706 	struct umtx_key key;
3707 	int error, cnt, rv;
3708 	uint32_t count, flags;
3709 
3710 	rv = fueword32(&sem->_flags, &flags);
3711 	if (rv == -1)
3712 		return (EFAULT);
3713 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3714 		return (error);
3715 	umtxq_lock(&key);
3716 	umtxq_busy(&key);
3717 	cnt = umtxq_count(&key);
3718 	if (cnt > 0) {
3719 		/*
3720 		 * If this was the last sleeping thread, clear the waiters
3721 		 * flag in _count.
3722 		 */
3723 		if (cnt == 1) {
3724 			umtxq_unlock(&key);
3725 			rv = fueword32(&sem->_count, &count);
3726 			while (rv != -1 && count & USEM_HAS_WAITERS) {
3727 				rv = casueword32(&sem->_count, count, &count,
3728 				    count & ~USEM_HAS_WAITERS);
3729 				if (rv == 1) {
3730 					rv = thread_check_susp(td, true);
3731 					if (rv != 0)
3732 						break;
3733 				}
3734 			}
3735 			if (rv == -1)
3736 				error = EFAULT;
3737 			else if (rv > 0) {
3738 				error = rv;
3739 			}
3740 			umtxq_lock(&key);
3741 		}
3742 
3743 		umtxq_signal(&key, 1);
3744 	}
3745 	umtxq_unbusy(&key);
3746 	umtxq_unlock(&key);
3747 	umtx_key_release(&key);
3748 	return (error);
3749 }
3750 
3751 #ifdef COMPAT_FREEBSD10
3752 int
3753 freebsd10__umtx_lock(struct thread *td, struct freebsd10__umtx_lock_args *uap)
3754 {
3755 	return (do_lock_umtx(td, uap->umtx, td->td_tid, 0));
3756 }
3757 
3758 int
3759 freebsd10__umtx_unlock(struct thread *td,
3760     struct freebsd10__umtx_unlock_args *uap)
3761 {
3762 	return (do_unlock_umtx(td, uap->umtx, td->td_tid));
3763 }
3764 #endif
3765 
3766 inline int
3767 umtx_copyin_timeout(const void *uaddr, struct timespec *tsp)
3768 {
3769 	int error;
3770 
3771 	error = copyin(uaddr, tsp, sizeof(*tsp));
3772 	if (error == 0) {
3773 		if (tsp->tv_sec < 0 ||
3774 		    tsp->tv_nsec >= 1000000000 ||
3775 		    tsp->tv_nsec < 0)
3776 			error = EINVAL;
3777 	}
3778 	return (error);
3779 }
3780 
3781 static inline int
3782 umtx_copyin_umtx_time(const void *uaddr, size_t size, struct _umtx_time *tp)
3783 {
3784 	int error;
3785 
3786 	if (size <= sizeof(tp->_timeout)) {
3787 		tp->_clockid = CLOCK_REALTIME;
3788 		tp->_flags = 0;
3789 		error = copyin(uaddr, &tp->_timeout, sizeof(tp->_timeout));
3790 	} else
3791 		error = copyin(uaddr, tp, sizeof(*tp));
3792 	if (error != 0)
3793 		return (error);
3794 	if (tp->_timeout.tv_sec < 0 ||
3795 	    tp->_timeout.tv_nsec >= 1000000000 || tp->_timeout.tv_nsec < 0)
3796 		return (EINVAL);
3797 	return (0);
3798 }
3799 
3800 static int
3801 umtx_copyin_robust_lists(const void *uaddr, size_t size,
3802     struct umtx_robust_lists_params *rb)
3803 {
3804 
3805 	if (size > sizeof(*rb))
3806 		return (EINVAL);
3807 	return (copyin(uaddr, rb, size));
3808 }
3809 
3810 static int
3811 umtx_copyout_timeout(void *uaddr, size_t sz, struct timespec *tsp)
3812 {
3813 
3814 	/*
3815 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
3816 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
3817 	 * copyops.
3818 	 */
3819 	KASSERT(sz >= sizeof(*tsp),
3820 	    ("umtx_copyops specifies incorrect sizes"));
3821 
3822 	return (copyout(tsp, uaddr, sizeof(*tsp)));
3823 }
3824 
3825 #ifdef COMPAT_FREEBSD10
3826 static int
3827 __umtx_op_lock_umtx(struct thread *td, struct _umtx_op_args *uap,
3828     const struct umtx_copyops *ops)
3829 {
3830 	struct timespec *ts, timeout;
3831 	int error;
3832 
3833 	/* Allow a null timespec (wait forever). */
3834 	if (uap->uaddr2 == NULL)
3835 		ts = NULL;
3836 	else {
3837 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
3838 		if (error != 0)
3839 			return (error);
3840 		ts = &timeout;
3841 	}
3842 #ifdef COMPAT_FREEBSD32
3843 	if (ops->compat32)
3844 		return (do_lock_umtx32(td, uap->obj, uap->val, ts));
3845 #endif
3846 	return (do_lock_umtx(td, uap->obj, uap->val, ts));
3847 }
3848 
3849 static int
3850 __umtx_op_unlock_umtx(struct thread *td, struct _umtx_op_args *uap,
3851     const struct umtx_copyops *ops)
3852 {
3853 #ifdef COMPAT_FREEBSD32
3854 	if (ops->compat32)
3855 		return (do_unlock_umtx32(td, uap->obj, uap->val));
3856 #endif
3857 	return (do_unlock_umtx(td, uap->obj, uap->val));
3858 }
3859 #endif	/* COMPAT_FREEBSD10 */
3860 
3861 #if !defined(COMPAT_FREEBSD10)
3862 static int
3863 __umtx_op_unimpl(struct thread *td __unused, struct _umtx_op_args *uap __unused,
3864     const struct umtx_copyops *ops __unused)
3865 {
3866 	return (EOPNOTSUPP);
3867 }
3868 #endif	/* COMPAT_FREEBSD10 */
3869 
3870 static int
3871 __umtx_op_wait(struct thread *td, struct _umtx_op_args *uap,
3872     const struct umtx_copyops *ops)
3873 {
3874 	struct _umtx_time timeout, *tm_p;
3875 	int error;
3876 
3877 	if (uap->uaddr2 == NULL)
3878 		tm_p = NULL;
3879 	else {
3880 		error = ops->copyin_umtx_time(
3881 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3882 		if (error != 0)
3883 			return (error);
3884 		tm_p = &timeout;
3885 	}
3886 	return (do_wait(td, uap->obj, uap->val, tm_p, ops->compat32, 0));
3887 }
3888 
3889 static int
3890 __umtx_op_wait_uint(struct thread *td, struct _umtx_op_args *uap,
3891     const struct umtx_copyops *ops)
3892 {
3893 	struct _umtx_time timeout, *tm_p;
3894 	int error;
3895 
3896 	if (uap->uaddr2 == NULL)
3897 		tm_p = NULL;
3898 	else {
3899 		error = ops->copyin_umtx_time(
3900 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3901 		if (error != 0)
3902 			return (error);
3903 		tm_p = &timeout;
3904 	}
3905 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
3906 }
3907 
3908 static int
3909 __umtx_op_wait_uint_private(struct thread *td, struct _umtx_op_args *uap,
3910     const struct umtx_copyops *ops)
3911 {
3912 	struct _umtx_time *tm_p, timeout;
3913 	int error;
3914 
3915 	if (uap->uaddr2 == NULL)
3916 		tm_p = NULL;
3917 	else {
3918 		error = ops->copyin_umtx_time(
3919 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3920 		if (error != 0)
3921 			return (error);
3922 		tm_p = &timeout;
3923 	}
3924 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
3925 }
3926 
3927 static int
3928 __umtx_op_wake(struct thread *td, struct _umtx_op_args *uap,
3929     const struct umtx_copyops *ops __unused)
3930 {
3931 
3932 	return (kern_umtx_wake(td, uap->obj, uap->val, 0));
3933 }
3934 
3935 #define BATCH_SIZE	128
3936 static int
3937 __umtx_op_nwake_private_native(struct thread *td, struct _umtx_op_args *uap)
3938 {
3939 	char *uaddrs[BATCH_SIZE], **upp;
3940 	int count, error, i, pos, tocopy;
3941 
3942 	upp = (char **)uap->obj;
3943 	error = 0;
3944 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3945 	    pos += tocopy) {
3946 		tocopy = MIN(count, BATCH_SIZE);
3947 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(char *));
3948 		if (error != 0)
3949 			break;
3950 		for (i = 0; i < tocopy; ++i) {
3951 			kern_umtx_wake(td, uaddrs[i], INT_MAX, 1);
3952 		}
3953 		maybe_yield();
3954 	}
3955 	return (error);
3956 }
3957 
3958 static int
3959 __umtx_op_nwake_private_compat32(struct thread *td, struct _umtx_op_args *uap)
3960 {
3961 	uint32_t uaddrs[BATCH_SIZE], *upp;
3962 	int count, error, i, pos, tocopy;
3963 
3964 	upp = (uint32_t *)uap->obj;
3965 	error = 0;
3966 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3967 	    pos += tocopy) {
3968 		tocopy = MIN(count, BATCH_SIZE);
3969 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(uint32_t));
3970 		if (error != 0)
3971 			break;
3972 		for (i = 0; i < tocopy; ++i) {
3973 			kern_umtx_wake(td, (void *)(uintptr_t)uaddrs[i],
3974 			    INT_MAX, 1);
3975 		}
3976 		maybe_yield();
3977 	}
3978 	return (error);
3979 }
3980 
3981 static int
3982 __umtx_op_nwake_private(struct thread *td, struct _umtx_op_args *uap,
3983     const struct umtx_copyops *ops)
3984 {
3985 
3986 	if (ops->compat32)
3987 		return (__umtx_op_nwake_private_compat32(td, uap));
3988 	return (__umtx_op_nwake_private_native(td, uap));
3989 }
3990 
3991 static int
3992 __umtx_op_wake_private(struct thread *td, struct _umtx_op_args *uap,
3993     const struct umtx_copyops *ops __unused)
3994 {
3995 
3996 	return (kern_umtx_wake(td, uap->obj, uap->val, 1));
3997 }
3998 
3999 static int
4000 __umtx_op_lock_umutex(struct thread *td, struct _umtx_op_args *uap,
4001    const struct umtx_copyops *ops)
4002 {
4003 	struct _umtx_time *tm_p, timeout;
4004 	int error;
4005 
4006 	/* Allow a null timespec (wait forever). */
4007 	if (uap->uaddr2 == NULL)
4008 		tm_p = NULL;
4009 	else {
4010 		error = ops->copyin_umtx_time(
4011 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4012 		if (error != 0)
4013 			return (error);
4014 		tm_p = &timeout;
4015 	}
4016 	return (do_lock_umutex(td, uap->obj, tm_p, 0));
4017 }
4018 
4019 static int
4020 __umtx_op_trylock_umutex(struct thread *td, struct _umtx_op_args *uap,
4021     const struct umtx_copyops *ops __unused)
4022 {
4023 
4024 	return (do_lock_umutex(td, uap->obj, NULL, _UMUTEX_TRY));
4025 }
4026 
4027 static int
4028 __umtx_op_wait_umutex(struct thread *td, struct _umtx_op_args *uap,
4029     const struct umtx_copyops *ops)
4030 {
4031 	struct _umtx_time *tm_p, timeout;
4032 	int error;
4033 
4034 	/* Allow a null timespec (wait forever). */
4035 	if (uap->uaddr2 == NULL)
4036 		tm_p = NULL;
4037 	else {
4038 		error = ops->copyin_umtx_time(
4039 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4040 		if (error != 0)
4041 			return (error);
4042 		tm_p = &timeout;
4043 	}
4044 	return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
4045 }
4046 
4047 static int
4048 __umtx_op_wake_umutex(struct thread *td, struct _umtx_op_args *uap,
4049     const struct umtx_copyops *ops __unused)
4050 {
4051 
4052 	return (do_wake_umutex(td, uap->obj));
4053 }
4054 
4055 static int
4056 __umtx_op_unlock_umutex(struct thread *td, struct _umtx_op_args *uap,
4057     const struct umtx_copyops *ops __unused)
4058 {
4059 
4060 	return (do_unlock_umutex(td, uap->obj, false));
4061 }
4062 
4063 static int
4064 __umtx_op_set_ceiling(struct thread *td, struct _umtx_op_args *uap,
4065     const struct umtx_copyops *ops __unused)
4066 {
4067 
4068 	return (do_set_ceiling(td, uap->obj, uap->val, uap->uaddr1));
4069 }
4070 
4071 static int
4072 __umtx_op_cv_wait(struct thread *td, struct _umtx_op_args *uap,
4073     const struct umtx_copyops *ops)
4074 {
4075 	struct timespec *ts, timeout;
4076 	int error;
4077 
4078 	/* Allow a null timespec (wait forever). */
4079 	if (uap->uaddr2 == NULL)
4080 		ts = NULL;
4081 	else {
4082 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
4083 		if (error != 0)
4084 			return (error);
4085 		ts = &timeout;
4086 	}
4087 	return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
4088 }
4089 
4090 static int
4091 __umtx_op_cv_signal(struct thread *td, struct _umtx_op_args *uap,
4092     const struct umtx_copyops *ops __unused)
4093 {
4094 
4095 	return (do_cv_signal(td, uap->obj));
4096 }
4097 
4098 static int
4099 __umtx_op_cv_broadcast(struct thread *td, struct _umtx_op_args *uap,
4100     const struct umtx_copyops *ops __unused)
4101 {
4102 
4103 	return (do_cv_broadcast(td, uap->obj));
4104 }
4105 
4106 static int
4107 __umtx_op_rw_rdlock(struct thread *td, struct _umtx_op_args *uap,
4108     const struct umtx_copyops *ops)
4109 {
4110 	struct _umtx_time timeout;
4111 	int error;
4112 
4113 	/* Allow a null timespec (wait forever). */
4114 	if (uap->uaddr2 == NULL) {
4115 		error = do_rw_rdlock(td, uap->obj, uap->val, 0);
4116 	} else {
4117 		error = ops->copyin_umtx_time(uap->uaddr2,
4118 		   (size_t)uap->uaddr1, &timeout);
4119 		if (error != 0)
4120 			return (error);
4121 		error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
4122 	}
4123 	return (error);
4124 }
4125 
4126 static int
4127 __umtx_op_rw_wrlock(struct thread *td, struct _umtx_op_args *uap,
4128     const struct umtx_copyops *ops)
4129 {
4130 	struct _umtx_time timeout;
4131 	int error;
4132 
4133 	/* Allow a null timespec (wait forever). */
4134 	if (uap->uaddr2 == NULL) {
4135 		error = do_rw_wrlock(td, uap->obj, 0);
4136 	} else {
4137 		error = ops->copyin_umtx_time(uap->uaddr2,
4138 		   (size_t)uap->uaddr1, &timeout);
4139 		if (error != 0)
4140 			return (error);
4141 
4142 		error = do_rw_wrlock(td, uap->obj, &timeout);
4143 	}
4144 	return (error);
4145 }
4146 
4147 static int
4148 __umtx_op_rw_unlock(struct thread *td, struct _umtx_op_args *uap,
4149     const struct umtx_copyops *ops __unused)
4150 {
4151 
4152 	return (do_rw_unlock(td, uap->obj));
4153 }
4154 
4155 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4156 static int
4157 __umtx_op_sem_wait(struct thread *td, struct _umtx_op_args *uap,
4158     const struct umtx_copyops *ops)
4159 {
4160 	struct _umtx_time *tm_p, timeout;
4161 	int error;
4162 
4163 	/* Allow a null timespec (wait forever). */
4164 	if (uap->uaddr2 == NULL)
4165 		tm_p = NULL;
4166 	else {
4167 		error = ops->copyin_umtx_time(
4168 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4169 		if (error != 0)
4170 			return (error);
4171 		tm_p = &timeout;
4172 	}
4173 	return (do_sem_wait(td, uap->obj, tm_p));
4174 }
4175 
4176 static int
4177 __umtx_op_sem_wake(struct thread *td, struct _umtx_op_args *uap,
4178     const struct umtx_copyops *ops __unused)
4179 {
4180 
4181 	return (do_sem_wake(td, uap->obj));
4182 }
4183 #endif
4184 
4185 static int
4186 __umtx_op_wake2_umutex(struct thread *td, struct _umtx_op_args *uap,
4187     const struct umtx_copyops *ops __unused)
4188 {
4189 
4190 	return (do_wake2_umutex(td, uap->obj, uap->val));
4191 }
4192 
4193 static int
4194 __umtx_op_sem2_wait(struct thread *td, struct _umtx_op_args *uap,
4195     const struct umtx_copyops *ops)
4196 {
4197 	struct _umtx_time *tm_p, timeout;
4198 	size_t uasize;
4199 	int error;
4200 
4201 	/* Allow a null timespec (wait forever). */
4202 	if (uap->uaddr2 == NULL) {
4203 		uasize = 0;
4204 		tm_p = NULL;
4205 	} else {
4206 		uasize = (size_t)uap->uaddr1;
4207 		error = ops->copyin_umtx_time(uap->uaddr2, uasize, &timeout);
4208 		if (error != 0)
4209 			return (error);
4210 		tm_p = &timeout;
4211 	}
4212 	error = do_sem2_wait(td, uap->obj, tm_p);
4213 	if (error == EINTR && uap->uaddr2 != NULL &&
4214 	    (timeout._flags & UMTX_ABSTIME) == 0 &&
4215 	    uasize >= ops->umtx_time_sz + ops->timespec_sz) {
4216 		error = ops->copyout_timeout(
4217 		    (void *)((uintptr_t)uap->uaddr2 + ops->umtx_time_sz),
4218 		    uasize - ops->umtx_time_sz, &timeout._timeout);
4219 		if (error == 0) {
4220 			error = EINTR;
4221 		}
4222 	}
4223 
4224 	return (error);
4225 }
4226 
4227 static int
4228 __umtx_op_sem2_wake(struct thread *td, struct _umtx_op_args *uap,
4229     const struct umtx_copyops *ops __unused)
4230 {
4231 
4232 	return (do_sem2_wake(td, uap->obj));
4233 }
4234 
4235 #define	USHM_OBJ_UMTX(o)						\
4236     ((struct umtx_shm_obj_list *)(&(o)->umtx_data))
4237 
4238 #define	USHMF_REG_LINKED	0x0001
4239 #define	USHMF_OBJ_LINKED	0x0002
4240 struct umtx_shm_reg {
4241 	TAILQ_ENTRY(umtx_shm_reg) ushm_reg_link;
4242 	LIST_ENTRY(umtx_shm_reg) ushm_obj_link;
4243 	struct umtx_key		ushm_key;
4244 	struct ucred		*ushm_cred;
4245 	struct shmfd		*ushm_obj;
4246 	u_int			ushm_refcnt;
4247 	u_int			ushm_flags;
4248 };
4249 
4250 LIST_HEAD(umtx_shm_obj_list, umtx_shm_reg);
4251 TAILQ_HEAD(umtx_shm_reg_head, umtx_shm_reg);
4252 
4253 static uma_zone_t umtx_shm_reg_zone;
4254 static struct umtx_shm_reg_head umtx_shm_registry[UMTX_CHAINS];
4255 static struct mtx umtx_shm_lock;
4256 static struct umtx_shm_reg_head umtx_shm_reg_delfree =
4257     TAILQ_HEAD_INITIALIZER(umtx_shm_reg_delfree);
4258 
4259 static void umtx_shm_free_reg(struct umtx_shm_reg *reg);
4260 
4261 static void
4262 umtx_shm_reg_delfree_tq(void *context __unused, int pending __unused)
4263 {
4264 	struct umtx_shm_reg_head d;
4265 	struct umtx_shm_reg *reg, *reg1;
4266 
4267 	TAILQ_INIT(&d);
4268 	mtx_lock(&umtx_shm_lock);
4269 	TAILQ_CONCAT(&d, &umtx_shm_reg_delfree, ushm_reg_link);
4270 	mtx_unlock(&umtx_shm_lock);
4271 	TAILQ_FOREACH_SAFE(reg, &d, ushm_reg_link, reg1) {
4272 		TAILQ_REMOVE(&d, reg, ushm_reg_link);
4273 		umtx_shm_free_reg(reg);
4274 	}
4275 }
4276 
4277 static struct task umtx_shm_reg_delfree_task =
4278     TASK_INITIALIZER(0, umtx_shm_reg_delfree_tq, NULL);
4279 
4280 static struct umtx_shm_reg *
4281 umtx_shm_find_reg_locked(const struct umtx_key *key)
4282 {
4283 	struct umtx_shm_reg *reg;
4284 	struct umtx_shm_reg_head *reg_head;
4285 
4286 	KASSERT(key->shared, ("umtx_p_find_rg: private key"));
4287 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4288 	reg_head = &umtx_shm_registry[key->hash];
4289 	TAILQ_FOREACH(reg, reg_head, ushm_reg_link) {
4290 		KASSERT(reg->ushm_key.shared,
4291 		    ("non-shared key on reg %p %d", reg, reg->ushm_key.shared));
4292 		if (reg->ushm_key.info.shared.object ==
4293 		    key->info.shared.object &&
4294 		    reg->ushm_key.info.shared.offset ==
4295 		    key->info.shared.offset) {
4296 			KASSERT(reg->ushm_key.type == TYPE_SHM, ("TYPE_USHM"));
4297 			KASSERT(reg->ushm_refcnt > 0,
4298 			    ("reg %p refcnt 0 onlist", reg));
4299 			KASSERT((reg->ushm_flags & USHMF_REG_LINKED) != 0,
4300 			    ("reg %p not linked", reg));
4301 			reg->ushm_refcnt++;
4302 			return (reg);
4303 		}
4304 	}
4305 	return (NULL);
4306 }
4307 
4308 static struct umtx_shm_reg *
4309 umtx_shm_find_reg(const struct umtx_key *key)
4310 {
4311 	struct umtx_shm_reg *reg;
4312 
4313 	mtx_lock(&umtx_shm_lock);
4314 	reg = umtx_shm_find_reg_locked(key);
4315 	mtx_unlock(&umtx_shm_lock);
4316 	return (reg);
4317 }
4318 
4319 static void
4320 umtx_shm_free_reg(struct umtx_shm_reg *reg)
4321 {
4322 
4323 	chgumtxcnt(reg->ushm_cred->cr_ruidinfo, -1, 0);
4324 	crfree(reg->ushm_cred);
4325 	shm_drop(reg->ushm_obj);
4326 	uma_zfree(umtx_shm_reg_zone, reg);
4327 }
4328 
4329 static bool
4330 umtx_shm_unref_reg_locked(struct umtx_shm_reg *reg, bool force)
4331 {
4332 	bool res;
4333 
4334 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4335 	KASSERT(reg->ushm_refcnt > 0, ("ushm_reg %p refcnt 0", reg));
4336 	reg->ushm_refcnt--;
4337 	res = reg->ushm_refcnt == 0;
4338 	if (res || force) {
4339 		if ((reg->ushm_flags & USHMF_REG_LINKED) != 0) {
4340 			TAILQ_REMOVE(&umtx_shm_registry[reg->ushm_key.hash],
4341 			    reg, ushm_reg_link);
4342 			reg->ushm_flags &= ~USHMF_REG_LINKED;
4343 		}
4344 		if ((reg->ushm_flags & USHMF_OBJ_LINKED) != 0) {
4345 			LIST_REMOVE(reg, ushm_obj_link);
4346 			reg->ushm_flags &= ~USHMF_OBJ_LINKED;
4347 		}
4348 	}
4349 	return (res);
4350 }
4351 
4352 static void
4353 umtx_shm_unref_reg(struct umtx_shm_reg *reg, bool force)
4354 {
4355 	vm_object_t object;
4356 	bool dofree;
4357 
4358 	if (force) {
4359 		object = reg->ushm_obj->shm_object;
4360 		VM_OBJECT_WLOCK(object);
4361 		object->flags |= OBJ_UMTXDEAD;
4362 		VM_OBJECT_WUNLOCK(object);
4363 	}
4364 	mtx_lock(&umtx_shm_lock);
4365 	dofree = umtx_shm_unref_reg_locked(reg, force);
4366 	mtx_unlock(&umtx_shm_lock);
4367 	if (dofree)
4368 		umtx_shm_free_reg(reg);
4369 }
4370 
4371 void
4372 umtx_shm_object_init(vm_object_t object)
4373 {
4374 
4375 	LIST_INIT(USHM_OBJ_UMTX(object));
4376 }
4377 
4378 void
4379 umtx_shm_object_terminated(vm_object_t object)
4380 {
4381 	struct umtx_shm_reg *reg, *reg1;
4382 	bool dofree;
4383 
4384 	if (LIST_EMPTY(USHM_OBJ_UMTX(object)))
4385 		return;
4386 
4387 	dofree = false;
4388 	mtx_lock(&umtx_shm_lock);
4389 	LIST_FOREACH_SAFE(reg, USHM_OBJ_UMTX(object), ushm_obj_link, reg1) {
4390 		if (umtx_shm_unref_reg_locked(reg, true)) {
4391 			TAILQ_INSERT_TAIL(&umtx_shm_reg_delfree, reg,
4392 			    ushm_reg_link);
4393 			dofree = true;
4394 		}
4395 	}
4396 	mtx_unlock(&umtx_shm_lock);
4397 	if (dofree)
4398 		taskqueue_enqueue(taskqueue_thread, &umtx_shm_reg_delfree_task);
4399 }
4400 
4401 static int
4402 umtx_shm_create_reg(struct thread *td, const struct umtx_key *key,
4403     struct umtx_shm_reg **res)
4404 {
4405 	struct umtx_shm_reg *reg, *reg1;
4406 	struct ucred *cred;
4407 	int error;
4408 
4409 	reg = umtx_shm_find_reg(key);
4410 	if (reg != NULL) {
4411 		*res = reg;
4412 		return (0);
4413 	}
4414 	cred = td->td_ucred;
4415 	if (!chgumtxcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_UMTXP)))
4416 		return (ENOMEM);
4417 	reg = uma_zalloc(umtx_shm_reg_zone, M_WAITOK | M_ZERO);
4418 	reg->ushm_refcnt = 1;
4419 	bcopy(key, &reg->ushm_key, sizeof(*key));
4420 	reg->ushm_obj = shm_alloc(td->td_ucred, O_RDWR, false);
4421 	reg->ushm_cred = crhold(cred);
4422 	error = shm_dotruncate(reg->ushm_obj, PAGE_SIZE);
4423 	if (error != 0) {
4424 		umtx_shm_free_reg(reg);
4425 		return (error);
4426 	}
4427 	mtx_lock(&umtx_shm_lock);
4428 	reg1 = umtx_shm_find_reg_locked(key);
4429 	if (reg1 != NULL) {
4430 		mtx_unlock(&umtx_shm_lock);
4431 		umtx_shm_free_reg(reg);
4432 		*res = reg1;
4433 		return (0);
4434 	}
4435 	reg->ushm_refcnt++;
4436 	TAILQ_INSERT_TAIL(&umtx_shm_registry[key->hash], reg, ushm_reg_link);
4437 	LIST_INSERT_HEAD(USHM_OBJ_UMTX(key->info.shared.object), reg,
4438 	    ushm_obj_link);
4439 	reg->ushm_flags = USHMF_REG_LINKED | USHMF_OBJ_LINKED;
4440 	mtx_unlock(&umtx_shm_lock);
4441 	*res = reg;
4442 	return (0);
4443 }
4444 
4445 static int
4446 umtx_shm_alive(struct thread *td, void *addr)
4447 {
4448 	vm_map_t map;
4449 	vm_map_entry_t entry;
4450 	vm_object_t object;
4451 	vm_pindex_t pindex;
4452 	vm_prot_t prot;
4453 	int res, ret;
4454 	boolean_t wired;
4455 
4456 	map = &td->td_proc->p_vmspace->vm_map;
4457 	res = vm_map_lookup(&map, (uintptr_t)addr, VM_PROT_READ, &entry,
4458 	    &object, &pindex, &prot, &wired);
4459 	if (res != KERN_SUCCESS)
4460 		return (EFAULT);
4461 	if (object == NULL)
4462 		ret = EINVAL;
4463 	else
4464 		ret = (object->flags & OBJ_UMTXDEAD) != 0 ? ENOTTY : 0;
4465 	vm_map_lookup_done(map, entry);
4466 	return (ret);
4467 }
4468 
4469 static void
4470 umtx_shm_init(void)
4471 {
4472 	int i;
4473 
4474 	umtx_shm_reg_zone = uma_zcreate("umtx_shm", sizeof(struct umtx_shm_reg),
4475 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
4476 	mtx_init(&umtx_shm_lock, "umtxshm", NULL, MTX_DEF);
4477 	for (i = 0; i < nitems(umtx_shm_registry); i++)
4478 		TAILQ_INIT(&umtx_shm_registry[i]);
4479 }
4480 
4481 static int
4482 umtx_shm(struct thread *td, void *addr, u_int flags)
4483 {
4484 	struct umtx_key key;
4485 	struct umtx_shm_reg *reg;
4486 	struct file *fp;
4487 	int error, fd;
4488 
4489 	if (__bitcount(flags & (UMTX_SHM_CREAT | UMTX_SHM_LOOKUP |
4490 	    UMTX_SHM_DESTROY| UMTX_SHM_ALIVE)) != 1)
4491 		return (EINVAL);
4492 	if ((flags & UMTX_SHM_ALIVE) != 0)
4493 		return (umtx_shm_alive(td, addr));
4494 	error = umtx_key_get(addr, TYPE_SHM, PROCESS_SHARE, &key);
4495 	if (error != 0)
4496 		return (error);
4497 	KASSERT(key.shared == 1, ("non-shared key"));
4498 	if ((flags & UMTX_SHM_CREAT) != 0) {
4499 		error = umtx_shm_create_reg(td, &key, &reg);
4500 	} else {
4501 		reg = umtx_shm_find_reg(&key);
4502 		if (reg == NULL)
4503 			error = ESRCH;
4504 	}
4505 	umtx_key_release(&key);
4506 	if (error != 0)
4507 		return (error);
4508 	KASSERT(reg != NULL, ("no reg"));
4509 	if ((flags & UMTX_SHM_DESTROY) != 0) {
4510 		umtx_shm_unref_reg(reg, true);
4511 	} else {
4512 #if 0
4513 #ifdef MAC
4514 		error = mac_posixshm_check_open(td->td_ucred,
4515 		    reg->ushm_obj, FFLAGS(O_RDWR));
4516 		if (error == 0)
4517 #endif
4518 			error = shm_access(reg->ushm_obj, td->td_ucred,
4519 			    FFLAGS(O_RDWR));
4520 		if (error == 0)
4521 #endif
4522 			error = falloc_caps(td, &fp, &fd, O_CLOEXEC, NULL);
4523 		if (error == 0) {
4524 			shm_hold(reg->ushm_obj);
4525 			finit(fp, FFLAGS(O_RDWR), DTYPE_SHM, reg->ushm_obj,
4526 			    &shm_ops);
4527 			td->td_retval[0] = fd;
4528 			fdrop(fp, td);
4529 		}
4530 	}
4531 	umtx_shm_unref_reg(reg, false);
4532 	return (error);
4533 }
4534 
4535 static int
4536 __umtx_op_shm(struct thread *td, struct _umtx_op_args *uap,
4537     const struct umtx_copyops *ops __unused)
4538 {
4539 
4540 	return (umtx_shm(td, uap->uaddr1, uap->val));
4541 }
4542 
4543 static int
4544 __umtx_op_robust_lists(struct thread *td, struct _umtx_op_args *uap,
4545     const struct umtx_copyops *ops)
4546 {
4547 	struct umtx_robust_lists_params rb;
4548 	int error;
4549 
4550 	if (ops->compat32) {
4551 		if ((td->td_pflags2 & TDP2_COMPAT32RB) == 0 &&
4552 		    (td->td_rb_list != 0 || td->td_rbp_list != 0 ||
4553 		    td->td_rb_inact != 0))
4554 			return (EBUSY);
4555 	} else if ((td->td_pflags2 & TDP2_COMPAT32RB) != 0) {
4556 		return (EBUSY);
4557 	}
4558 
4559 	bzero(&rb, sizeof(rb));
4560 	error = ops->copyin_robust_lists(uap->uaddr1, uap->val, &rb);
4561 	if (error != 0)
4562 		return (error);
4563 
4564 	if (ops->compat32)
4565 		td->td_pflags2 |= TDP2_COMPAT32RB;
4566 
4567 	td->td_rb_list = rb.robust_list_offset;
4568 	td->td_rbp_list = rb.robust_priv_list_offset;
4569 	td->td_rb_inact = rb.robust_inact_offset;
4570 	return (0);
4571 }
4572 
4573 #if defined(__i386__) || defined(__amd64__)
4574 /*
4575  * Provide the standard 32-bit definitions for x86, since native/compat32 use a
4576  * 32-bit time_t there.  Other architectures just need the i386 definitions
4577  * along with their standard compat32.
4578  */
4579 struct timespecx32 {
4580 	int64_t			tv_sec;
4581 	int32_t			tv_nsec;
4582 };
4583 
4584 struct umtx_timex32 {
4585 	struct	timespecx32	_timeout;
4586 	uint32_t		_flags;
4587 	uint32_t		_clockid;
4588 };
4589 
4590 #ifndef __i386__
4591 #define	timespeci386	timespec32
4592 #define	umtx_timei386	umtx_time32
4593 #endif
4594 #else /* !__i386__ && !__amd64__ */
4595 /* 32-bit architectures can emulate i386, so define these almost everywhere. */
4596 struct timespeci386 {
4597 	int32_t			tv_sec;
4598 	int32_t			tv_nsec;
4599 };
4600 
4601 struct umtx_timei386 {
4602 	struct	timespeci386	_timeout;
4603 	uint32_t		_flags;
4604 	uint32_t		_clockid;
4605 };
4606 
4607 #if defined(__LP64__)
4608 #define	timespecx32	timespec32
4609 #define	umtx_timex32	umtx_time32
4610 #endif
4611 #endif
4612 
4613 static int
4614 umtx_copyin_robust_lists32(const void *uaddr, size_t size,
4615     struct umtx_robust_lists_params *rbp)
4616 {
4617 	struct umtx_robust_lists_params_compat32 rb32;
4618 	int error;
4619 
4620 	if (size > sizeof(rb32))
4621 		return (EINVAL);
4622 	bzero(&rb32, sizeof(rb32));
4623 	error = copyin(uaddr, &rb32, size);
4624 	if (error != 0)
4625 		return (error);
4626 	CP(rb32, *rbp, robust_list_offset);
4627 	CP(rb32, *rbp, robust_priv_list_offset);
4628 	CP(rb32, *rbp, robust_inact_offset);
4629 	return (0);
4630 }
4631 
4632 #ifndef __i386__
4633 static inline int
4634 umtx_copyin_timeouti386(const void *uaddr, struct timespec *tsp)
4635 {
4636 	struct timespeci386 ts32;
4637 	int error;
4638 
4639 	error = copyin(uaddr, &ts32, sizeof(ts32));
4640 	if (error == 0) {
4641 		if (ts32.tv_sec < 0 ||
4642 		    ts32.tv_nsec >= 1000000000 ||
4643 		    ts32.tv_nsec < 0)
4644 			error = EINVAL;
4645 		else {
4646 			CP(ts32, *tsp, tv_sec);
4647 			CP(ts32, *tsp, tv_nsec);
4648 		}
4649 	}
4650 	return (error);
4651 }
4652 
4653 static inline int
4654 umtx_copyin_umtx_timei386(const void *uaddr, size_t size, struct _umtx_time *tp)
4655 {
4656 	struct umtx_timei386 t32;
4657 	int error;
4658 
4659 	t32._clockid = CLOCK_REALTIME;
4660 	t32._flags   = 0;
4661 	if (size <= sizeof(t32._timeout))
4662 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4663 	else
4664 		error = copyin(uaddr, &t32, sizeof(t32));
4665 	if (error != 0)
4666 		return (error);
4667 	if (t32._timeout.tv_sec < 0 ||
4668 	    t32._timeout.tv_nsec >= 1000000000 || t32._timeout.tv_nsec < 0)
4669 		return (EINVAL);
4670 	TS_CP(t32, *tp, _timeout);
4671 	CP(t32, *tp, _flags);
4672 	CP(t32, *tp, _clockid);
4673 	return (0);
4674 }
4675 
4676 static int
4677 umtx_copyout_timeouti386(void *uaddr, size_t sz, struct timespec *tsp)
4678 {
4679 	struct timespeci386 remain32 = {
4680 		.tv_sec = tsp->tv_sec,
4681 		.tv_nsec = tsp->tv_nsec,
4682 	};
4683 
4684 	/*
4685 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4686 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4687 	 * copyops.
4688 	 */
4689 	KASSERT(sz >= sizeof(remain32),
4690 	    ("umtx_copyops specifies incorrect sizes"));
4691 
4692 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4693 }
4694 #endif /* !__i386__ */
4695 
4696 #if defined(__i386__) || defined(__LP64__)
4697 static inline int
4698 umtx_copyin_timeoutx32(const void *uaddr, struct timespec *tsp)
4699 {
4700 	struct timespecx32 ts32;
4701 	int error;
4702 
4703 	error = copyin(uaddr, &ts32, sizeof(ts32));
4704 	if (error == 0) {
4705 		if (ts32.tv_sec < 0 ||
4706 		    ts32.tv_nsec >= 1000000000 ||
4707 		    ts32.tv_nsec < 0)
4708 			error = EINVAL;
4709 		else {
4710 			CP(ts32, *tsp, tv_sec);
4711 			CP(ts32, *tsp, tv_nsec);
4712 		}
4713 	}
4714 	return (error);
4715 }
4716 
4717 static inline int
4718 umtx_copyin_umtx_timex32(const void *uaddr, size_t size, struct _umtx_time *tp)
4719 {
4720 	struct umtx_timex32 t32;
4721 	int error;
4722 
4723 	t32._clockid = CLOCK_REALTIME;
4724 	t32._flags   = 0;
4725 	if (size <= sizeof(t32._timeout))
4726 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4727 	else
4728 		error = copyin(uaddr, &t32, sizeof(t32));
4729 	if (error != 0)
4730 		return (error);
4731 	if (t32._timeout.tv_sec < 0 ||
4732 	    t32._timeout.tv_nsec >= 1000000000 || t32._timeout.tv_nsec < 0)
4733 		return (EINVAL);
4734 	TS_CP(t32, *tp, _timeout);
4735 	CP(t32, *tp, _flags);
4736 	CP(t32, *tp, _clockid);
4737 	return (0);
4738 }
4739 
4740 static int
4741 umtx_copyout_timeoutx32(void *uaddr, size_t sz, struct timespec *tsp)
4742 {
4743 	struct timespecx32 remain32 = {
4744 		.tv_sec = tsp->tv_sec,
4745 		.tv_nsec = tsp->tv_nsec,
4746 	};
4747 
4748 	/*
4749 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4750 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4751 	 * copyops.
4752 	 */
4753 	KASSERT(sz >= sizeof(remain32),
4754 	    ("umtx_copyops specifies incorrect sizes"));
4755 
4756 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4757 }
4758 #endif /* __i386__ || __LP64__ */
4759 
4760 typedef int (*_umtx_op_func)(struct thread *td, struct _umtx_op_args *uap,
4761     const struct umtx_copyops *umtx_ops);
4762 
4763 static const _umtx_op_func op_table[] = {
4764 #ifdef COMPAT_FREEBSD10
4765 	[UMTX_OP_LOCK]		= __umtx_op_lock_umtx,
4766 	[UMTX_OP_UNLOCK]	= __umtx_op_unlock_umtx,
4767 #else
4768 	[UMTX_OP_LOCK]		= __umtx_op_unimpl,
4769 	[UMTX_OP_UNLOCK]	= __umtx_op_unimpl,
4770 #endif
4771 	[UMTX_OP_WAIT]		= __umtx_op_wait,
4772 	[UMTX_OP_WAKE]		= __umtx_op_wake,
4773 	[UMTX_OP_MUTEX_TRYLOCK]	= __umtx_op_trylock_umutex,
4774 	[UMTX_OP_MUTEX_LOCK]	= __umtx_op_lock_umutex,
4775 	[UMTX_OP_MUTEX_UNLOCK]	= __umtx_op_unlock_umutex,
4776 	[UMTX_OP_SET_CEILING]	= __umtx_op_set_ceiling,
4777 	[UMTX_OP_CV_WAIT]	= __umtx_op_cv_wait,
4778 	[UMTX_OP_CV_SIGNAL]	= __umtx_op_cv_signal,
4779 	[UMTX_OP_CV_BROADCAST]	= __umtx_op_cv_broadcast,
4780 	[UMTX_OP_WAIT_UINT]	= __umtx_op_wait_uint,
4781 	[UMTX_OP_RW_RDLOCK]	= __umtx_op_rw_rdlock,
4782 	[UMTX_OP_RW_WRLOCK]	= __umtx_op_rw_wrlock,
4783 	[UMTX_OP_RW_UNLOCK]	= __umtx_op_rw_unlock,
4784 	[UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private,
4785 	[UMTX_OP_WAKE_PRIVATE]	= __umtx_op_wake_private,
4786 	[UMTX_OP_MUTEX_WAIT]	= __umtx_op_wait_umutex,
4787 	[UMTX_OP_MUTEX_WAKE]	= __umtx_op_wake_umutex,
4788 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4789 	[UMTX_OP_SEM_WAIT]	= __umtx_op_sem_wait,
4790 	[UMTX_OP_SEM_WAKE]	= __umtx_op_sem_wake,
4791 #else
4792 	[UMTX_OP_SEM_WAIT]	= __umtx_op_unimpl,
4793 	[UMTX_OP_SEM_WAKE]	= __umtx_op_unimpl,
4794 #endif
4795 	[UMTX_OP_NWAKE_PRIVATE]	= __umtx_op_nwake_private,
4796 	[UMTX_OP_MUTEX_WAKE2]	= __umtx_op_wake2_umutex,
4797 	[UMTX_OP_SEM2_WAIT]	= __umtx_op_sem2_wait,
4798 	[UMTX_OP_SEM2_WAKE]	= __umtx_op_sem2_wake,
4799 	[UMTX_OP_SHM]		= __umtx_op_shm,
4800 	[UMTX_OP_ROBUST_LISTS]	= __umtx_op_robust_lists,
4801 };
4802 
4803 static const struct umtx_copyops umtx_native_ops = {
4804 	.copyin_timeout = umtx_copyin_timeout,
4805 	.copyin_umtx_time = umtx_copyin_umtx_time,
4806 	.copyin_robust_lists = umtx_copyin_robust_lists,
4807 	.copyout_timeout = umtx_copyout_timeout,
4808 	.timespec_sz = sizeof(struct timespec),
4809 	.umtx_time_sz = sizeof(struct _umtx_time),
4810 };
4811 
4812 #ifndef __i386__
4813 static const struct umtx_copyops umtx_native_opsi386 = {
4814 	.copyin_timeout = umtx_copyin_timeouti386,
4815 	.copyin_umtx_time = umtx_copyin_umtx_timei386,
4816 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4817 	.copyout_timeout = umtx_copyout_timeouti386,
4818 	.timespec_sz = sizeof(struct timespeci386),
4819 	.umtx_time_sz = sizeof(struct umtx_timei386),
4820 	.compat32 = true,
4821 };
4822 #endif
4823 
4824 #if defined(__i386__) || defined(__LP64__)
4825 /* i386 can emulate other 32-bit archs, too! */
4826 static const struct umtx_copyops umtx_native_opsx32 = {
4827 	.copyin_timeout = umtx_copyin_timeoutx32,
4828 	.copyin_umtx_time = umtx_copyin_umtx_timex32,
4829 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4830 	.copyout_timeout = umtx_copyout_timeoutx32,
4831 	.timespec_sz = sizeof(struct timespecx32),
4832 	.umtx_time_sz = sizeof(struct umtx_timex32),
4833 	.compat32 = true,
4834 };
4835 
4836 #ifdef COMPAT_FREEBSD32
4837 #ifdef __amd64__
4838 #define	umtx_native_ops32	umtx_native_opsi386
4839 #else
4840 #define	umtx_native_ops32	umtx_native_opsx32
4841 #endif
4842 #endif /* COMPAT_FREEBSD32 */
4843 #endif /* __i386__ || __LP64__ */
4844 
4845 #define	UMTX_OP__FLAGS	(UMTX_OP__32BIT | UMTX_OP__I386)
4846 
4847 static int
4848 kern__umtx_op(struct thread *td, void *obj, int op, unsigned long val,
4849     void *uaddr1, void *uaddr2, const struct umtx_copyops *ops)
4850 {
4851 	struct _umtx_op_args uap = {
4852 		.obj = obj,
4853 		.op = op & ~UMTX_OP__FLAGS,
4854 		.val = val,
4855 		.uaddr1 = uaddr1,
4856 		.uaddr2 = uaddr2
4857 	};
4858 
4859 	if ((uap.op >= nitems(op_table)))
4860 		return (EINVAL);
4861 	return ((*op_table[uap.op])(td, &uap, ops));
4862 }
4863 
4864 int
4865 sys__umtx_op(struct thread *td, struct _umtx_op_args *uap)
4866 {
4867 	static const struct umtx_copyops *umtx_ops;
4868 
4869 	umtx_ops = &umtx_native_ops;
4870 #ifdef __LP64__
4871 	if ((uap->op & (UMTX_OP__32BIT | UMTX_OP__I386)) != 0) {
4872 		if ((uap->op & UMTX_OP__I386) != 0)
4873 			umtx_ops = &umtx_native_opsi386;
4874 		else
4875 			umtx_ops = &umtx_native_opsx32;
4876 	}
4877 #elif !defined(__i386__)
4878 	/* We consider UMTX_OP__32BIT a nop on !i386 ILP32. */
4879 	if ((uap->op & UMTX_OP__I386) != 0)
4880 		umtx_ops = &umtx_native_opsi386;
4881 #else
4882 	/* Likewise, UMTX_OP__I386 is a nop on i386. */
4883 	if ((uap->op & UMTX_OP__32BIT) != 0)
4884 		umtx_ops = &umtx_native_opsx32;
4885 #endif
4886 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4887 	    uap->uaddr2, umtx_ops));
4888 }
4889 
4890 #ifdef COMPAT_FREEBSD32
4891 #ifdef COMPAT_FREEBSD10
4892 int
4893 freebsd10_freebsd32__umtx_lock(struct thread *td,
4894     struct freebsd10_freebsd32__umtx_lock_args *uap)
4895 {
4896 	return (do_lock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid, NULL));
4897 }
4898 
4899 int
4900 freebsd10_freebsd32__umtx_unlock(struct thread *td,
4901     struct freebsd10_freebsd32__umtx_unlock_args *uap)
4902 {
4903 	return (do_unlock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid));
4904 }
4905 #endif /* COMPAT_FREEBSD10 */
4906 
4907 int
4908 freebsd32__umtx_op(struct thread *td, struct freebsd32__umtx_op_args *uap)
4909 {
4910 
4911 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4912 	    uap->uaddr2, &umtx_native_ops32));
4913 }
4914 #endif /* COMPAT_FREEBSD32 */
4915 
4916 void
4917 umtx_thread_init(struct thread *td)
4918 {
4919 
4920 	td->td_umtxq = umtxq_alloc();
4921 	td->td_umtxq->uq_thread = td;
4922 }
4923 
4924 void
4925 umtx_thread_fini(struct thread *td)
4926 {
4927 
4928 	umtxq_free(td->td_umtxq);
4929 }
4930 
4931 /*
4932  * It will be called when new thread is created, e.g fork().
4933  */
4934 void
4935 umtx_thread_alloc(struct thread *td)
4936 {
4937 	struct umtx_q *uq;
4938 
4939 	uq = td->td_umtxq;
4940 	uq->uq_inherited_pri = PRI_MAX;
4941 
4942 	KASSERT(uq->uq_flags == 0, ("uq_flags != 0"));
4943 	KASSERT(uq->uq_thread == td, ("uq_thread != td"));
4944 	KASSERT(uq->uq_pi_blocked == NULL, ("uq_pi_blocked != NULL"));
4945 	KASSERT(TAILQ_EMPTY(&uq->uq_pi_contested), ("uq_pi_contested is not empty"));
4946 }
4947 
4948 /*
4949  * exec() hook.
4950  *
4951  * Clear robust lists for all process' threads, not delaying the
4952  * cleanup to thread exit, since the relevant address space is
4953  * destroyed right now.
4954  */
4955 void
4956 umtx_exec(struct proc *p)
4957 {
4958 	struct thread *td;
4959 
4960 	KASSERT(p == curproc, ("need curproc"));
4961 	KASSERT((p->p_flag & P_HADTHREADS) == 0 ||
4962 	    (p->p_flag & P_STOPPED_SINGLE) != 0,
4963 	    ("curproc must be single-threaded"));
4964 	/*
4965 	 * There is no need to lock the list as only this thread can be
4966 	 * running.
4967 	 */
4968 	FOREACH_THREAD_IN_PROC(p, td) {
4969 		KASSERT(td == curthread ||
4970 		    ((td->td_flags & TDF_BOUNDARY) != 0 && TD_IS_SUSPENDED(td)),
4971 		    ("running thread %p %p", p, td));
4972 		umtx_thread_cleanup(td);
4973 		td->td_rb_list = td->td_rbp_list = td->td_rb_inact = 0;
4974 	}
4975 }
4976 
4977 /*
4978  * thread exit hook.
4979  */
4980 void
4981 umtx_thread_exit(struct thread *td)
4982 {
4983 
4984 	umtx_thread_cleanup(td);
4985 }
4986 
4987 static int
4988 umtx_read_uptr(struct thread *td, uintptr_t ptr, uintptr_t *res, bool compat32)
4989 {
4990 	u_long res1;
4991 	uint32_t res32;
4992 	int error;
4993 
4994 	if (compat32) {
4995 		error = fueword32((void *)ptr, &res32);
4996 		if (error == 0)
4997 			res1 = res32;
4998 	} else {
4999 		error = fueword((void *)ptr, &res1);
5000 	}
5001 	if (error == 0)
5002 		*res = res1;
5003 	else
5004 		error = EFAULT;
5005 	return (error);
5006 }
5007 
5008 static void
5009 umtx_read_rb_list(struct thread *td, struct umutex *m, uintptr_t *rb_list,
5010     bool compat32)
5011 {
5012 	struct umutex32 m32;
5013 
5014 	if (compat32) {
5015 		memcpy(&m32, m, sizeof(m32));
5016 		*rb_list = m32.m_rb_lnk;
5017 	} else {
5018 		*rb_list = m->m_rb_lnk;
5019 	}
5020 }
5021 
5022 static int
5023 umtx_handle_rb(struct thread *td, uintptr_t rbp, uintptr_t *rb_list, bool inact,
5024     bool compat32)
5025 {
5026 	struct umutex m;
5027 	int error;
5028 
5029 	KASSERT(td->td_proc == curproc, ("need current vmspace"));
5030 	error = copyin((void *)rbp, &m, sizeof(m));
5031 	if (error != 0)
5032 		return (error);
5033 	if (rb_list != NULL)
5034 		umtx_read_rb_list(td, &m, rb_list, compat32);
5035 	if ((m.m_flags & UMUTEX_ROBUST) == 0)
5036 		return (EINVAL);
5037 	if ((m.m_owner & ~UMUTEX_CONTESTED) != td->td_tid)
5038 		/* inact is cleared after unlock, allow the inconsistency */
5039 		return (inact ? 0 : EINVAL);
5040 	return (do_unlock_umutex(td, (struct umutex *)rbp, true));
5041 }
5042 
5043 static void
5044 umtx_cleanup_rb_list(struct thread *td, uintptr_t rb_list, uintptr_t *rb_inact,
5045     const char *name, bool compat32)
5046 {
5047 	int error, i;
5048 	uintptr_t rbp;
5049 	bool inact;
5050 
5051 	if (rb_list == 0)
5052 		return;
5053 	error = umtx_read_uptr(td, rb_list, &rbp, compat32);
5054 	for (i = 0; error == 0 && rbp != 0 && i < umtx_max_rb; i++) {
5055 		if (rbp == *rb_inact) {
5056 			inact = true;
5057 			*rb_inact = 0;
5058 		} else
5059 			inact = false;
5060 		error = umtx_handle_rb(td, rbp, &rbp, inact, compat32);
5061 	}
5062 	if (i == umtx_max_rb && umtx_verbose_rb) {
5063 		uprintf("comm %s pid %d: reached umtx %smax rb %d\n",
5064 		    td->td_proc->p_comm, td->td_proc->p_pid, name, umtx_max_rb);
5065 	}
5066 	if (error != 0 && umtx_verbose_rb) {
5067 		uprintf("comm %s pid %d: handling %srb error %d\n",
5068 		    td->td_proc->p_comm, td->td_proc->p_pid, name, error);
5069 	}
5070 }
5071 
5072 /*
5073  * Clean up umtx data.
5074  */
5075 static void
5076 umtx_thread_cleanup(struct thread *td)
5077 {
5078 	struct umtx_q *uq;
5079 	struct umtx_pi *pi;
5080 	uintptr_t rb_inact;
5081 	bool compat32;
5082 
5083 	/*
5084 	 * Disown pi mutexes.
5085 	 */
5086 	uq = td->td_umtxq;
5087 	if (uq != NULL) {
5088 		if (uq->uq_inherited_pri != PRI_MAX ||
5089 		    !TAILQ_EMPTY(&uq->uq_pi_contested)) {
5090 			mtx_lock(&umtx_lock);
5091 			uq->uq_inherited_pri = PRI_MAX;
5092 			while ((pi = TAILQ_FIRST(&uq->uq_pi_contested)) != NULL) {
5093 				pi->pi_owner = NULL;
5094 				TAILQ_REMOVE(&uq->uq_pi_contested, pi, pi_link);
5095 			}
5096 			mtx_unlock(&umtx_lock);
5097 		}
5098 		sched_lend_user_prio_cond(td, PRI_MAX);
5099 	}
5100 
5101 	compat32 = (td->td_pflags2 & TDP2_COMPAT32RB) != 0;
5102 	td->td_pflags2 &= ~TDP2_COMPAT32RB;
5103 
5104 	if (td->td_rb_inact == 0 && td->td_rb_list == 0 && td->td_rbp_list == 0)
5105 		return;
5106 
5107 	/*
5108 	 * Handle terminated robust mutexes.  Must be done after
5109 	 * robust pi disown, otherwise unlock could see unowned
5110 	 * entries.
5111 	 */
5112 	rb_inact = td->td_rb_inact;
5113 	if (rb_inact != 0)
5114 		(void)umtx_read_uptr(td, rb_inact, &rb_inact, compat32);
5115 	umtx_cleanup_rb_list(td, td->td_rb_list, &rb_inact, "", compat32);
5116 	umtx_cleanup_rb_list(td, td->td_rbp_list, &rb_inact, "priv ", compat32);
5117 	if (rb_inact != 0)
5118 		(void)umtx_handle_rb(td, rb_inact, NULL, true, compat32);
5119 }
5120