xref: /freebsd/sys/kern/kern_umtx.c (revision 6ba2210ee039f2f12878c217bcf058e9c8b26b29)
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 static int umtx_abs_timeout_gethz(struct umtx_abs_timeout *timo);
179 static inline void umtx_abs_timeout_update(struct umtx_abs_timeout *timo);
180 
181 static void umtx_shm_init(void);
182 static void umtxq_sysinit(void *);
183 static void umtxq_hash(struct umtx_key *key);
184 static int do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags,
185     bool rb);
186 static void umtx_thread_cleanup(struct thread *td);
187 SYSINIT(umtx, SI_SUB_EVENTHANDLER+1, SI_ORDER_MIDDLE, umtxq_sysinit, NULL);
188 
189 #define umtxq_signal(key, nwake)	umtxq_signal_queue((key), (nwake), UMTX_SHARED_QUEUE)
190 
191 static struct mtx umtx_lock;
192 
193 #ifdef UMTX_PROFILING
194 static void
195 umtx_init_profiling(void)
196 {
197 	struct sysctl_oid *chain_oid;
198 	char chain_name[10];
199 	int i;
200 
201 	for (i = 0; i < UMTX_CHAINS; ++i) {
202 		snprintf(chain_name, sizeof(chain_name), "%d", i);
203 		chain_oid = SYSCTL_ADD_NODE(NULL,
204 		    SYSCTL_STATIC_CHILDREN(_debug_umtx_chains), OID_AUTO,
205 		    chain_name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
206 		    "umtx hash stats");
207 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
208 		    "max_length0", CTLFLAG_RD, &umtxq_chains[0][i].max_length, 0, NULL);
209 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
210 		    "max_length1", CTLFLAG_RD, &umtxq_chains[1][i].max_length, 0, NULL);
211 	}
212 }
213 
214 static int
215 sysctl_debug_umtx_chains_peaks(SYSCTL_HANDLER_ARGS)
216 {
217 	char buf[512];
218 	struct sbuf sb;
219 	struct umtxq_chain *uc;
220 	u_int fract, i, j, tot, whole;
221 	u_int sf0, sf1, sf2, sf3, sf4;
222 	u_int si0, si1, si2, si3, si4;
223 	u_int sw0, sw1, sw2, sw3, sw4;
224 
225 	sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
226 	for (i = 0; i < 2; i++) {
227 		tot = 0;
228 		for (j = 0; j < UMTX_CHAINS; ++j) {
229 			uc = &umtxq_chains[i][j];
230 			mtx_lock(&uc->uc_lock);
231 			tot += uc->max_length;
232 			mtx_unlock(&uc->uc_lock);
233 		}
234 		if (tot == 0)
235 			sbuf_printf(&sb, "%u) Empty ", i);
236 		else {
237 			sf0 = sf1 = sf2 = sf3 = sf4 = 0;
238 			si0 = si1 = si2 = si3 = si4 = 0;
239 			sw0 = sw1 = sw2 = sw3 = sw4 = 0;
240 			for (j = 0; j < UMTX_CHAINS; j++) {
241 				uc = &umtxq_chains[i][j];
242 				mtx_lock(&uc->uc_lock);
243 				whole = uc->max_length * 100;
244 				mtx_unlock(&uc->uc_lock);
245 				fract = (whole % tot) * 100;
246 				if (UPROF_PERC_BIGGER(whole, fract, sw0, sf0)) {
247 					sf0 = fract;
248 					si0 = j;
249 					sw0 = whole;
250 				} else if (UPROF_PERC_BIGGER(whole, fract, sw1,
251 				    sf1)) {
252 					sf1 = fract;
253 					si1 = j;
254 					sw1 = whole;
255 				} else if (UPROF_PERC_BIGGER(whole, fract, sw2,
256 				    sf2)) {
257 					sf2 = fract;
258 					si2 = j;
259 					sw2 = whole;
260 				} else if (UPROF_PERC_BIGGER(whole, fract, sw3,
261 				    sf3)) {
262 					sf3 = fract;
263 					si3 = j;
264 					sw3 = whole;
265 				} else if (UPROF_PERC_BIGGER(whole, fract, sw4,
266 				    sf4)) {
267 					sf4 = fract;
268 					si4 = j;
269 					sw4 = whole;
270 				}
271 			}
272 			sbuf_printf(&sb, "queue %u:\n", i);
273 			sbuf_printf(&sb, "1st: %u.%u%% idx: %u\n", sw0 / tot,
274 			    sf0 / tot, si0);
275 			sbuf_printf(&sb, "2nd: %u.%u%% idx: %u\n", sw1 / tot,
276 			    sf1 / tot, si1);
277 			sbuf_printf(&sb, "3rd: %u.%u%% idx: %u\n", sw2 / tot,
278 			    sf2 / tot, si2);
279 			sbuf_printf(&sb, "4th: %u.%u%% idx: %u\n", sw3 / tot,
280 			    sf3 / tot, si3);
281 			sbuf_printf(&sb, "5th: %u.%u%% idx: %u\n", sw4 / tot,
282 			    sf4 / tot, si4);
283 		}
284 	}
285 	sbuf_trim(&sb);
286 	sbuf_finish(&sb);
287 	sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
288 	sbuf_delete(&sb);
289 	return (0);
290 }
291 
292 static int
293 sysctl_debug_umtx_chains_clear(SYSCTL_HANDLER_ARGS)
294 {
295 	struct umtxq_chain *uc;
296 	u_int i, j;
297 	int clear, error;
298 
299 	clear = 0;
300 	error = sysctl_handle_int(oidp, &clear, 0, req);
301 	if (error != 0 || req->newptr == NULL)
302 		return (error);
303 
304 	if (clear != 0) {
305 		for (i = 0; i < 2; ++i) {
306 			for (j = 0; j < UMTX_CHAINS; ++j) {
307 				uc = &umtxq_chains[i][j];
308 				mtx_lock(&uc->uc_lock);
309 				uc->length = 0;
310 				uc->max_length = 0;
311 				mtx_unlock(&uc->uc_lock);
312 			}
313 		}
314 	}
315 	return (0);
316 }
317 
318 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, clear,
319     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
320     sysctl_debug_umtx_chains_clear, "I",
321     "Clear umtx chains statistics");
322 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, peaks,
323     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
324     sysctl_debug_umtx_chains_peaks, "A",
325     "Highest peaks in chains max length");
326 #endif
327 
328 static void
329 umtxq_sysinit(void *arg __unused)
330 {
331 	int i, j;
332 
333 	umtx_pi_zone = uma_zcreate("umtx pi", sizeof(struct umtx_pi),
334 		NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
335 	for (i = 0; i < 2; ++i) {
336 		for (j = 0; j < UMTX_CHAINS; ++j) {
337 			mtx_init(&umtxq_chains[i][j].uc_lock, "umtxql", NULL,
338 				 MTX_DEF | MTX_DUPOK);
339 			LIST_INIT(&umtxq_chains[i][j].uc_queue[0]);
340 			LIST_INIT(&umtxq_chains[i][j].uc_queue[1]);
341 			LIST_INIT(&umtxq_chains[i][j].uc_spare_queue);
342 			TAILQ_INIT(&umtxq_chains[i][j].uc_pi_list);
343 			umtxq_chains[i][j].uc_busy = 0;
344 			umtxq_chains[i][j].uc_waiters = 0;
345 #ifdef UMTX_PROFILING
346 			umtxq_chains[i][j].length = 0;
347 			umtxq_chains[i][j].max_length = 0;
348 #endif
349 		}
350 	}
351 #ifdef UMTX_PROFILING
352 	umtx_init_profiling();
353 #endif
354 	mtx_init(&umtx_lock, "umtx lock", NULL, MTX_DEF);
355 	umtx_shm_init();
356 }
357 
358 struct umtx_q *
359 umtxq_alloc(void)
360 {
361 	struct umtx_q *uq;
362 
363 	uq = malloc(sizeof(struct umtx_q), M_UMTX, M_WAITOK | M_ZERO);
364 	uq->uq_spare_queue = malloc(sizeof(struct umtxq_queue), M_UMTX,
365 	    M_WAITOK | M_ZERO);
366 	TAILQ_INIT(&uq->uq_spare_queue->head);
367 	TAILQ_INIT(&uq->uq_pi_contested);
368 	uq->uq_inherited_pri = PRI_MAX;
369 	return (uq);
370 }
371 
372 void
373 umtxq_free(struct umtx_q *uq)
374 {
375 
376 	MPASS(uq->uq_spare_queue != NULL);
377 	free(uq->uq_spare_queue, M_UMTX);
378 	free(uq, M_UMTX);
379 }
380 
381 static inline void
382 umtxq_hash(struct umtx_key *key)
383 {
384 	unsigned n;
385 
386 	n = (uintptr_t)key->info.both.a + key->info.both.b;
387 	key->hash = ((n * GOLDEN_RATIO_PRIME) >> UMTX_SHIFTS) % UMTX_CHAINS;
388 }
389 
390 struct umtxq_chain *
391 umtxq_getchain(struct umtx_key *key)
392 {
393 
394 	if (key->type <= TYPE_SEM)
395 		return (&umtxq_chains[1][key->hash]);
396 	return (&umtxq_chains[0][key->hash]);
397 }
398 
399 /*
400  * Set chain to busy state when following operation
401  * may be blocked (kernel mutex can not be used).
402  */
403 void
404 umtxq_busy(struct umtx_key *key)
405 {
406 	struct umtxq_chain *uc;
407 
408 	uc = umtxq_getchain(key);
409 	mtx_assert(&uc->uc_lock, MA_OWNED);
410 	if (uc->uc_busy) {
411 #ifdef SMP
412 		if (smp_cpus > 1) {
413 			int count = BUSY_SPINS;
414 			if (count > 0) {
415 				umtxq_unlock(key);
416 				while (uc->uc_busy && --count > 0)
417 					cpu_spinwait();
418 				umtxq_lock(key);
419 			}
420 		}
421 #endif
422 		while (uc->uc_busy) {
423 			uc->uc_waiters++;
424 			msleep(uc, &uc->uc_lock, 0, "umtxqb", 0);
425 			uc->uc_waiters--;
426 		}
427 	}
428 	uc->uc_busy = 1;
429 }
430 
431 /*
432  * Unbusy a chain.
433  */
434 void
435 umtxq_unbusy(struct umtx_key *key)
436 {
437 	struct umtxq_chain *uc;
438 
439 	uc = umtxq_getchain(key);
440 	mtx_assert(&uc->uc_lock, MA_OWNED);
441 	KASSERT(uc->uc_busy != 0, ("not busy"));
442 	uc->uc_busy = 0;
443 	if (uc->uc_waiters)
444 		wakeup_one(uc);
445 }
446 
447 void
448 umtxq_unbusy_unlocked(struct umtx_key *key)
449 {
450 
451 	umtxq_lock(key);
452 	umtxq_unbusy(key);
453 	umtxq_unlock(key);
454 }
455 
456 static struct umtxq_queue *
457 umtxq_queue_lookup(struct umtx_key *key, int q)
458 {
459 	struct umtxq_queue *uh;
460 	struct umtxq_chain *uc;
461 
462 	uc = umtxq_getchain(key);
463 	UMTXQ_LOCKED_ASSERT(uc);
464 	LIST_FOREACH(uh, &uc->uc_queue[q], link) {
465 		if (umtx_key_match(&uh->key, key))
466 			return (uh);
467 	}
468 
469 	return (NULL);
470 }
471 
472 void
473 umtxq_insert_queue(struct umtx_q *uq, int q)
474 {
475 	struct umtxq_queue *uh;
476 	struct umtxq_chain *uc;
477 
478 	uc = umtxq_getchain(&uq->uq_key);
479 	UMTXQ_LOCKED_ASSERT(uc);
480 	KASSERT((uq->uq_flags & UQF_UMTXQ) == 0, ("umtx_q is already on queue"));
481 	uh = umtxq_queue_lookup(&uq->uq_key, q);
482 	if (uh != NULL) {
483 		LIST_INSERT_HEAD(&uc->uc_spare_queue, uq->uq_spare_queue, link);
484 	} else {
485 		uh = uq->uq_spare_queue;
486 		uh->key = uq->uq_key;
487 		LIST_INSERT_HEAD(&uc->uc_queue[q], uh, link);
488 #ifdef UMTX_PROFILING
489 		uc->length++;
490 		if (uc->length > uc->max_length) {
491 			uc->max_length = uc->length;
492 			if (uc->max_length > max_length)
493 				max_length = uc->max_length;
494 		}
495 #endif
496 	}
497 	uq->uq_spare_queue = NULL;
498 
499 	TAILQ_INSERT_TAIL(&uh->head, uq, uq_link);
500 	uh->length++;
501 	uq->uq_flags |= UQF_UMTXQ;
502 	uq->uq_cur_queue = uh;
503 	return;
504 }
505 
506 void
507 umtxq_remove_queue(struct umtx_q *uq, int q)
508 {
509 	struct umtxq_chain *uc;
510 	struct umtxq_queue *uh;
511 
512 	uc = umtxq_getchain(&uq->uq_key);
513 	UMTXQ_LOCKED_ASSERT(uc);
514 	if (uq->uq_flags & UQF_UMTXQ) {
515 		uh = uq->uq_cur_queue;
516 		TAILQ_REMOVE(&uh->head, uq, uq_link);
517 		uh->length--;
518 		uq->uq_flags &= ~UQF_UMTXQ;
519 		if (TAILQ_EMPTY(&uh->head)) {
520 			KASSERT(uh->length == 0,
521 			    ("inconsistent umtxq_queue length"));
522 #ifdef UMTX_PROFILING
523 			uc->length--;
524 #endif
525 			LIST_REMOVE(uh, link);
526 		} else {
527 			uh = LIST_FIRST(&uc->uc_spare_queue);
528 			KASSERT(uh != NULL, ("uc_spare_queue is empty"));
529 			LIST_REMOVE(uh, link);
530 		}
531 		uq->uq_spare_queue = uh;
532 		uq->uq_cur_queue = NULL;
533 	}
534 }
535 
536 /*
537  * Check if there are multiple waiters
538  */
539 int
540 umtxq_count(struct umtx_key *key)
541 {
542 	struct umtxq_queue *uh;
543 
544 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
545 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
546 	if (uh != NULL)
547 		return (uh->length);
548 	return (0);
549 }
550 
551 /*
552  * Check if there are multiple PI waiters and returns first
553  * waiter.
554  */
555 static int
556 umtxq_count_pi(struct umtx_key *key, struct umtx_q **first)
557 {
558 	struct umtxq_queue *uh;
559 
560 	*first = NULL;
561 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
562 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
563 	if (uh != NULL) {
564 		*first = TAILQ_FIRST(&uh->head);
565 		return (uh->length);
566 	}
567 	return (0);
568 }
569 
570 /*
571  * Wake up threads waiting on an userland object by a bit mask.
572  */
573 int
574 umtxq_signal_mask(struct umtx_key *key, int n_wake, u_int bitset)
575 {
576 	struct umtxq_queue *uh;
577 	struct umtx_q *uq, *uq_temp;
578 	int ret;
579 
580 	ret = 0;
581 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
582 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
583 	if (uh == NULL)
584 		return (0);
585 	TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
586 		if ((uq->uq_bitset & bitset) == 0)
587 			continue;
588 		umtxq_remove_queue(uq, UMTX_SHARED_QUEUE);
589 		wakeup_one(uq);
590 		if (++ret >= n_wake)
591 			break;
592 	}
593 	return (ret);
594 }
595 
596 /*
597  * Wake up threads waiting on an userland object.
598  */
599 
600 static int
601 umtxq_signal_queue(struct umtx_key *key, int n_wake, int q)
602 {
603 	struct umtxq_queue *uh;
604 	struct umtx_q *uq;
605 	int ret;
606 
607 	ret = 0;
608 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
609 	uh = umtxq_queue_lookup(key, q);
610 	if (uh != NULL) {
611 		while ((uq = TAILQ_FIRST(&uh->head)) != NULL) {
612 			umtxq_remove_queue(uq, q);
613 			wakeup(uq);
614 			if (++ret >= n_wake)
615 				return (ret);
616 		}
617 	}
618 	return (ret);
619 }
620 
621 /*
622  * Wake up specified thread.
623  */
624 static inline void
625 umtxq_signal_thread(struct umtx_q *uq)
626 {
627 
628 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
629 	umtxq_remove(uq);
630 	wakeup(uq);
631 }
632 
633 /*
634  * Wake up a maximum of n_wake threads that are waiting on an userland
635  * object identified by key. The remaining threads are removed from queue
636  * identified by key and added to the queue identified by key2 (requeued).
637  * The n_requeue specifies an upper limit on the number of threads that
638  * are requeued to the second queue.
639  */
640 int
641 umtxq_requeue(struct umtx_key *key, int n_wake, struct umtx_key *key2,
642     int n_requeue)
643 {
644 	struct umtxq_queue *uh, *uh2;
645 	struct umtx_q *uq, *uq_temp;
646 	int ret;
647 
648 	ret = 0;
649 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
650 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key2));
651 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
652 	uh2 = umtxq_queue_lookup(key2, UMTX_SHARED_QUEUE);
653 	if (uh == NULL)
654 		return (0);
655 	TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
656 		if (++ret <= n_wake) {
657 			umtxq_remove(uq);
658 			wakeup_one(uq);
659 		} else {
660 			umtxq_remove(uq);
661 			uq->uq_key = *key2;
662 			umtxq_insert(uq);
663 			if (ret - n_wake == n_requeue)
664 				break;
665 		}
666 	}
667 	return (ret);
668 }
669 
670 static inline int
671 tstohz(const struct timespec *tsp)
672 {
673 	struct timeval tv;
674 
675 	TIMESPEC_TO_TIMEVAL(&tv, tsp);
676 	return tvtohz(&tv);
677 }
678 
679 void
680 umtx_abs_timeout_init(struct umtx_abs_timeout *timo, int clockid,
681     int absolute, const struct timespec *timeout)
682 {
683 
684 	timo->clockid = clockid;
685 	if (!absolute) {
686 		timo->is_abs_real = false;
687 		umtx_abs_timeout_update(timo);
688 		timespecadd(&timo->cur, timeout, &timo->end);
689 	} else {
690 		timo->end = *timeout;
691 		timo->is_abs_real = clockid == CLOCK_REALTIME ||
692 		    clockid == CLOCK_REALTIME_FAST ||
693 		    clockid == CLOCK_REALTIME_PRECISE;
694 		/*
695 		 * If is_abs_real, umtxq_sleep will read the clock
696 		 * after setting td_rtcgen; otherwise, read it here.
697 		 */
698 		if (!timo->is_abs_real) {
699 			umtx_abs_timeout_update(timo);
700 		}
701 	}
702 }
703 
704 static void
705 umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
706     const struct _umtx_time *umtxtime)
707 {
708 
709 	umtx_abs_timeout_init(timo, umtxtime->_clockid,
710 	    (umtxtime->_flags & UMTX_ABSTIME) != 0, &umtxtime->_timeout);
711 }
712 
713 static void
714 umtx_abs_timeout_update(struct umtx_abs_timeout *timo)
715 {
716 
717 	kern_clock_gettime(curthread, timo->clockid, &timo->cur);
718 }
719 
720 static int
721 umtx_abs_timeout_gethz(struct umtx_abs_timeout *timo)
722 {
723 	struct timespec tts;
724 
725 	if (timespeccmp(&timo->end, &timo->cur, <=))
726 		return (-1);
727 	timespecsub(&timo->end, &timo->cur, &tts);
728 	return (tstohz(&tts));
729 }
730 
731 static uint32_t
732 umtx_unlock_val(uint32_t flags, bool rb)
733 {
734 
735 	if (rb)
736 		return (UMUTEX_RB_OWNERDEAD);
737 	else if ((flags & UMUTEX_NONCONSISTENT) != 0)
738 		return (UMUTEX_RB_NOTRECOV);
739 	else
740 		return (UMUTEX_UNOWNED);
741 
742 }
743 
744 /*
745  * Put thread into sleep state, before sleeping, check if
746  * thread was removed from umtx queue.
747  */
748 int
749 umtxq_sleep(struct umtx_q *uq, const char *wmesg,
750     struct umtx_abs_timeout *abstime)
751 {
752 	struct umtxq_chain *uc;
753 	int error, timo;
754 
755 	if (abstime != NULL && abstime->is_abs_real) {
756 		curthread->td_rtcgen = atomic_load_acq_int(&rtc_generation);
757 		umtx_abs_timeout_update(abstime);
758 	}
759 
760 	uc = umtxq_getchain(&uq->uq_key);
761 	UMTXQ_LOCKED_ASSERT(uc);
762 	for (;;) {
763 		if (!(uq->uq_flags & UQF_UMTXQ)) {
764 			error = 0;
765 			break;
766 		}
767 		if (abstime != NULL) {
768 			timo = umtx_abs_timeout_gethz(abstime);
769 			if (timo < 0) {
770 				error = ETIMEDOUT;
771 				break;
772 			}
773 		} else
774 			timo = 0;
775 		error = msleep(uq, &uc->uc_lock, PCATCH | PDROP, wmesg, timo);
776 		if (error == EINTR || error == ERESTART) {
777 			umtxq_lock(&uq->uq_key);
778 			break;
779 		}
780 		if (abstime != NULL) {
781 			if (abstime->is_abs_real)
782 				curthread->td_rtcgen =
783 				    atomic_load_acq_int(&rtc_generation);
784 			umtx_abs_timeout_update(abstime);
785 		}
786 		umtxq_lock(&uq->uq_key);
787 	}
788 
789 	curthread->td_rtcgen = 0;
790 	return (error);
791 }
792 
793 /*
794  * Convert userspace address into unique logical address.
795  */
796 int
797 umtx_key_get(const void *addr, int type, int share, struct umtx_key *key)
798 {
799 	struct thread *td = curthread;
800 	vm_map_t map;
801 	vm_map_entry_t entry;
802 	vm_pindex_t pindex;
803 	vm_prot_t prot;
804 	boolean_t wired;
805 
806 	key->type = type;
807 	if (share == THREAD_SHARE) {
808 		key->shared = 0;
809 		key->info.private.vs = td->td_proc->p_vmspace;
810 		key->info.private.addr = (uintptr_t)addr;
811 	} else {
812 		MPASS(share == PROCESS_SHARE || share == AUTO_SHARE);
813 		map = &td->td_proc->p_vmspace->vm_map;
814 		if (vm_map_lookup(&map, (vm_offset_t)addr, VM_PROT_WRITE,
815 		    &entry, &key->info.shared.object, &pindex, &prot,
816 		    &wired) != KERN_SUCCESS) {
817 			return (EFAULT);
818 		}
819 
820 		if ((share == PROCESS_SHARE) ||
821 		    (share == AUTO_SHARE &&
822 		     VM_INHERIT_SHARE == entry->inheritance)) {
823 			key->shared = 1;
824 			key->info.shared.offset = (vm_offset_t)addr -
825 			    entry->start + entry->offset;
826 			vm_object_reference(key->info.shared.object);
827 		} else {
828 			key->shared = 0;
829 			key->info.private.vs = td->td_proc->p_vmspace;
830 			key->info.private.addr = (uintptr_t)addr;
831 		}
832 		vm_map_lookup_done(map, entry);
833 	}
834 
835 	umtxq_hash(key);
836 	return (0);
837 }
838 
839 /*
840  * Release key.
841  */
842 void
843 umtx_key_release(struct umtx_key *key)
844 {
845 	if (key->shared)
846 		vm_object_deallocate(key->info.shared.object);
847 }
848 
849 #ifdef COMPAT_FREEBSD10
850 /*
851  * Lock a umtx object.
852  */
853 static int
854 do_lock_umtx(struct thread *td, struct umtx *umtx, u_long id,
855     const struct timespec *timeout)
856 {
857 	struct umtx_abs_timeout timo;
858 	struct umtx_q *uq;
859 	u_long owner;
860 	u_long old;
861 	int error = 0;
862 
863 	uq = td->td_umtxq;
864 	if (timeout != NULL)
865 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
866 
867 	/*
868 	 * Care must be exercised when dealing with umtx structure. It
869 	 * can fault on any access.
870 	 */
871 	for (;;) {
872 		/*
873 		 * Try the uncontested case.  This should be done in userland.
874 		 */
875 		owner = casuword(&umtx->u_owner, UMTX_UNOWNED, id);
876 
877 		/* The acquire succeeded. */
878 		if (owner == UMTX_UNOWNED)
879 			return (0);
880 
881 		/* The address was invalid. */
882 		if (owner == -1)
883 			return (EFAULT);
884 
885 		/* If no one owns it but it is contested try to acquire it. */
886 		if (owner == UMTX_CONTESTED) {
887 			owner = casuword(&umtx->u_owner,
888 			    UMTX_CONTESTED, id | UMTX_CONTESTED);
889 
890 			if (owner == UMTX_CONTESTED)
891 				return (0);
892 
893 			/* The address was invalid. */
894 			if (owner == -1)
895 				return (EFAULT);
896 
897 			error = thread_check_susp(td, false);
898 			if (error != 0)
899 				break;
900 
901 			/* If this failed the lock has changed, restart. */
902 			continue;
903 		}
904 
905 		/*
906 		 * If we caught a signal, we have retried and now
907 		 * exit immediately.
908 		 */
909 		if (error != 0)
910 			break;
911 
912 		if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK,
913 			AUTO_SHARE, &uq->uq_key)) != 0)
914 			return (error);
915 
916 		umtxq_lock(&uq->uq_key);
917 		umtxq_busy(&uq->uq_key);
918 		umtxq_insert(uq);
919 		umtxq_unbusy(&uq->uq_key);
920 		umtxq_unlock(&uq->uq_key);
921 
922 		/*
923 		 * Set the contested bit so that a release in user space
924 		 * knows to use the system call for unlock.  If this fails
925 		 * either some one else has acquired the lock or it has been
926 		 * released.
927 		 */
928 		old = casuword(&umtx->u_owner, owner, owner | UMTX_CONTESTED);
929 
930 		/* The address was invalid. */
931 		if (old == -1) {
932 			umtxq_lock(&uq->uq_key);
933 			umtxq_remove(uq);
934 			umtxq_unlock(&uq->uq_key);
935 			umtx_key_release(&uq->uq_key);
936 			return (EFAULT);
937 		}
938 
939 		/*
940 		 * We set the contested bit, sleep. Otherwise the lock changed
941 		 * and we need to retry or we lost a race to the thread
942 		 * unlocking the umtx.
943 		 */
944 		umtxq_lock(&uq->uq_key);
945 		if (old == owner)
946 			error = umtxq_sleep(uq, "umtx", timeout == NULL ? NULL :
947 			    &timo);
948 		umtxq_remove(uq);
949 		umtxq_unlock(&uq->uq_key);
950 		umtx_key_release(&uq->uq_key);
951 
952 		if (error == 0)
953 			error = thread_check_susp(td, false);
954 	}
955 
956 	if (timeout == NULL) {
957 		/* Mutex locking is restarted if it is interrupted. */
958 		if (error == EINTR)
959 			error = ERESTART;
960 	} else {
961 		/* Timed-locking is not restarted. */
962 		if (error == ERESTART)
963 			error = EINTR;
964 	}
965 	return (error);
966 }
967 
968 /*
969  * Unlock a umtx object.
970  */
971 static int
972 do_unlock_umtx(struct thread *td, struct umtx *umtx, u_long id)
973 {
974 	struct umtx_key key;
975 	u_long owner;
976 	u_long old;
977 	int error;
978 	int count;
979 
980 	/*
981 	 * Make sure we own this mtx.
982 	 */
983 	owner = fuword(__DEVOLATILE(u_long *, &umtx->u_owner));
984 	if (owner == -1)
985 		return (EFAULT);
986 
987 	if ((owner & ~UMTX_CONTESTED) != id)
988 		return (EPERM);
989 
990 	/* This should be done in userland */
991 	if ((owner & UMTX_CONTESTED) == 0) {
992 		old = casuword(&umtx->u_owner, owner, UMTX_UNOWNED);
993 		if (old == -1)
994 			return (EFAULT);
995 		if (old == owner)
996 			return (0);
997 		owner = old;
998 	}
999 
1000 	/* We should only ever be in here for contested locks */
1001 	if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1002 	    &key)) != 0)
1003 		return (error);
1004 
1005 	umtxq_lock(&key);
1006 	umtxq_busy(&key);
1007 	count = umtxq_count(&key);
1008 	umtxq_unlock(&key);
1009 
1010 	/*
1011 	 * When unlocking the umtx, it must be marked as unowned if
1012 	 * there is zero or one thread only waiting for it.
1013 	 * Otherwise, it must be marked as contested.
1014 	 */
1015 	old = casuword(&umtx->u_owner, owner,
1016 	    count <= 1 ? UMTX_UNOWNED : UMTX_CONTESTED);
1017 	umtxq_lock(&key);
1018 	umtxq_signal(&key,1);
1019 	umtxq_unbusy(&key);
1020 	umtxq_unlock(&key);
1021 	umtx_key_release(&key);
1022 	if (old == -1)
1023 		return (EFAULT);
1024 	if (old != owner)
1025 		return (EINVAL);
1026 	return (0);
1027 }
1028 
1029 #ifdef COMPAT_FREEBSD32
1030 
1031 /*
1032  * Lock a umtx object.
1033  */
1034 static int
1035 do_lock_umtx32(struct thread *td, uint32_t *m, uint32_t id,
1036 	const struct timespec *timeout)
1037 {
1038 	struct umtx_abs_timeout timo;
1039 	struct umtx_q *uq;
1040 	uint32_t owner;
1041 	uint32_t old;
1042 	int error = 0;
1043 
1044 	uq = td->td_umtxq;
1045 
1046 	if (timeout != NULL)
1047 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
1048 
1049 	/*
1050 	 * Care must be exercised when dealing with umtx structure. It
1051 	 * can fault on any access.
1052 	 */
1053 	for (;;) {
1054 		/*
1055 		 * Try the uncontested case.  This should be done in userland.
1056 		 */
1057 		owner = casuword32(m, UMUTEX_UNOWNED, id);
1058 
1059 		/* The acquire succeeded. */
1060 		if (owner == UMUTEX_UNOWNED)
1061 			return (0);
1062 
1063 		/* The address was invalid. */
1064 		if (owner == -1)
1065 			return (EFAULT);
1066 
1067 		/* If no one owns it but it is contested try to acquire it. */
1068 		if (owner == UMUTEX_CONTESTED) {
1069 			owner = casuword32(m,
1070 			    UMUTEX_CONTESTED, id | UMUTEX_CONTESTED);
1071 			if (owner == UMUTEX_CONTESTED)
1072 				return (0);
1073 
1074 			/* The address was invalid. */
1075 			if (owner == -1)
1076 				return (EFAULT);
1077 
1078 			error = thread_check_susp(td, false);
1079 			if (error != 0)
1080 				break;
1081 
1082 			/* If this failed the lock has changed, restart. */
1083 			continue;
1084 		}
1085 
1086 		/*
1087 		 * If we caught a signal, we have retried and now
1088 		 * exit immediately.
1089 		 */
1090 		if (error != 0)
1091 			return (error);
1092 
1093 		if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK,
1094 			AUTO_SHARE, &uq->uq_key)) != 0)
1095 			return (error);
1096 
1097 		umtxq_lock(&uq->uq_key);
1098 		umtxq_busy(&uq->uq_key);
1099 		umtxq_insert(uq);
1100 		umtxq_unbusy(&uq->uq_key);
1101 		umtxq_unlock(&uq->uq_key);
1102 
1103 		/*
1104 		 * Set the contested bit so that a release in user space
1105 		 * knows to use the system call for unlock.  If this fails
1106 		 * either some one else has acquired the lock or it has been
1107 		 * released.
1108 		 */
1109 		old = casuword32(m, owner, owner | UMUTEX_CONTESTED);
1110 
1111 		/* The address was invalid. */
1112 		if (old == -1) {
1113 			umtxq_lock(&uq->uq_key);
1114 			umtxq_remove(uq);
1115 			umtxq_unlock(&uq->uq_key);
1116 			umtx_key_release(&uq->uq_key);
1117 			return (EFAULT);
1118 		}
1119 
1120 		/*
1121 		 * We set the contested bit, sleep. Otherwise the lock changed
1122 		 * and we need to retry or we lost a race to the thread
1123 		 * unlocking the umtx.
1124 		 */
1125 		umtxq_lock(&uq->uq_key);
1126 		if (old == owner)
1127 			error = umtxq_sleep(uq, "umtx", timeout == NULL ?
1128 			    NULL : &timo);
1129 		umtxq_remove(uq);
1130 		umtxq_unlock(&uq->uq_key);
1131 		umtx_key_release(&uq->uq_key);
1132 
1133 		if (error == 0)
1134 			error = thread_check_susp(td, false);
1135 	}
1136 
1137 	if (timeout == NULL) {
1138 		/* Mutex locking is restarted if it is interrupted. */
1139 		if (error == EINTR)
1140 			error = ERESTART;
1141 	} else {
1142 		/* Timed-locking is not restarted. */
1143 		if (error == ERESTART)
1144 			error = EINTR;
1145 	}
1146 	return (error);
1147 }
1148 
1149 /*
1150  * Unlock a umtx object.
1151  */
1152 static int
1153 do_unlock_umtx32(struct thread *td, uint32_t *m, uint32_t id)
1154 {
1155 	struct umtx_key key;
1156 	uint32_t owner;
1157 	uint32_t old;
1158 	int error;
1159 	int count;
1160 
1161 	/*
1162 	 * Make sure we own this mtx.
1163 	 */
1164 	owner = fuword32(m);
1165 	if (owner == -1)
1166 		return (EFAULT);
1167 
1168 	if ((owner & ~UMUTEX_CONTESTED) != id)
1169 		return (EPERM);
1170 
1171 	/* This should be done in userland */
1172 	if ((owner & UMUTEX_CONTESTED) == 0) {
1173 		old = casuword32(m, owner, UMUTEX_UNOWNED);
1174 		if (old == -1)
1175 			return (EFAULT);
1176 		if (old == owner)
1177 			return (0);
1178 		owner = old;
1179 	}
1180 
1181 	/* We should only ever be in here for contested locks */
1182 	if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1183 		&key)) != 0)
1184 		return (error);
1185 
1186 	umtxq_lock(&key);
1187 	umtxq_busy(&key);
1188 	count = umtxq_count(&key);
1189 	umtxq_unlock(&key);
1190 
1191 	/*
1192 	 * When unlocking the umtx, it must be marked as unowned if
1193 	 * there is zero or one thread only waiting for it.
1194 	 * Otherwise, it must be marked as contested.
1195 	 */
1196 	old = casuword32(m, owner,
1197 		count <= 1 ? UMUTEX_UNOWNED : UMUTEX_CONTESTED);
1198 	umtxq_lock(&key);
1199 	umtxq_signal(&key,1);
1200 	umtxq_unbusy(&key);
1201 	umtxq_unlock(&key);
1202 	umtx_key_release(&key);
1203 	if (old == -1)
1204 		return (EFAULT);
1205 	if (old != owner)
1206 		return (EINVAL);
1207 	return (0);
1208 }
1209 #endif	/* COMPAT_FREEBSD32 */
1210 #endif	/* COMPAT_FREEBSD10 */
1211 
1212 /*
1213  * Fetch and compare value, sleep on the address if value is not changed.
1214  */
1215 static int
1216 do_wait(struct thread *td, void *addr, u_long id,
1217     struct _umtx_time *timeout, int compat32, int is_private)
1218 {
1219 	struct umtx_abs_timeout timo;
1220 	struct umtx_q *uq;
1221 	u_long tmp;
1222 	uint32_t tmp32;
1223 	int error = 0;
1224 
1225 	uq = td->td_umtxq;
1226 	if ((error = umtx_key_get(addr, TYPE_SIMPLE_WAIT,
1227 		is_private ? THREAD_SHARE : AUTO_SHARE, &uq->uq_key)) != 0)
1228 		return (error);
1229 
1230 	if (timeout != NULL)
1231 		umtx_abs_timeout_init2(&timo, timeout);
1232 
1233 	umtxq_lock(&uq->uq_key);
1234 	umtxq_insert(uq);
1235 	umtxq_unlock(&uq->uq_key);
1236 	if (compat32 == 0) {
1237 		error = fueword(addr, &tmp);
1238 		if (error != 0)
1239 			error = EFAULT;
1240 	} else {
1241 		error = fueword32(addr, &tmp32);
1242 		if (error == 0)
1243 			tmp = tmp32;
1244 		else
1245 			error = EFAULT;
1246 	}
1247 	umtxq_lock(&uq->uq_key);
1248 	if (error == 0) {
1249 		if (tmp == id)
1250 			error = umtxq_sleep(uq, "uwait", timeout == NULL ?
1251 			    NULL : &timo);
1252 		if ((uq->uq_flags & UQF_UMTXQ) == 0)
1253 			error = 0;
1254 		else
1255 			umtxq_remove(uq);
1256 	} else if ((uq->uq_flags & UQF_UMTXQ) != 0) {
1257 		umtxq_remove(uq);
1258 	}
1259 	umtxq_unlock(&uq->uq_key);
1260 	umtx_key_release(&uq->uq_key);
1261 	if (error == ERESTART)
1262 		error = EINTR;
1263 	return (error);
1264 }
1265 
1266 /*
1267  * Wake up threads sleeping on the specified address.
1268  */
1269 int
1270 kern_umtx_wake(struct thread *td, void *uaddr, int n_wake, int is_private)
1271 {
1272 	struct umtx_key key;
1273 	int ret;
1274 
1275 	if ((ret = umtx_key_get(uaddr, TYPE_SIMPLE_WAIT,
1276 	    is_private ? THREAD_SHARE : AUTO_SHARE, &key)) != 0)
1277 		return (ret);
1278 	umtxq_lock(&key);
1279 	umtxq_signal(&key, n_wake);
1280 	umtxq_unlock(&key);
1281 	umtx_key_release(&key);
1282 	return (0);
1283 }
1284 
1285 /*
1286  * Lock PTHREAD_PRIO_NONE protocol POSIX mutex.
1287  */
1288 static int
1289 do_lock_normal(struct thread *td, struct umutex *m, uint32_t flags,
1290     struct _umtx_time *timeout, int mode)
1291 {
1292 	struct umtx_abs_timeout timo;
1293 	struct umtx_q *uq;
1294 	uint32_t owner, old, id;
1295 	int error, rv;
1296 
1297 	id = td->td_tid;
1298 	uq = td->td_umtxq;
1299 	error = 0;
1300 	if (timeout != NULL)
1301 		umtx_abs_timeout_init2(&timo, timeout);
1302 
1303 	/*
1304 	 * Care must be exercised when dealing with umtx structure. It
1305 	 * can fault on any access.
1306 	 */
1307 	for (;;) {
1308 		rv = fueword32(&m->m_owner, &owner);
1309 		if (rv == -1)
1310 			return (EFAULT);
1311 		if (mode == _UMUTEX_WAIT) {
1312 			if (owner == UMUTEX_UNOWNED ||
1313 			    owner == UMUTEX_CONTESTED ||
1314 			    owner == UMUTEX_RB_OWNERDEAD ||
1315 			    owner == UMUTEX_RB_NOTRECOV)
1316 				return (0);
1317 		} else {
1318 			/*
1319 			 * Robust mutex terminated.  Kernel duty is to
1320 			 * return EOWNERDEAD to the userspace.  The
1321 			 * umutex.m_flags UMUTEX_NONCONSISTENT is set
1322 			 * by the common userspace code.
1323 			 */
1324 			if (owner == UMUTEX_RB_OWNERDEAD) {
1325 				rv = casueword32(&m->m_owner,
1326 				    UMUTEX_RB_OWNERDEAD, &owner,
1327 				    id | UMUTEX_CONTESTED);
1328 				if (rv == -1)
1329 					return (EFAULT);
1330 				if (rv == 0) {
1331 					MPASS(owner == UMUTEX_RB_OWNERDEAD);
1332 					return (EOWNERDEAD); /* success */
1333 				}
1334 				MPASS(rv == 1);
1335 				rv = thread_check_susp(td, false);
1336 				if (rv != 0)
1337 					return (rv);
1338 				continue;
1339 			}
1340 			if (owner == UMUTEX_RB_NOTRECOV)
1341 				return (ENOTRECOVERABLE);
1342 
1343 			/*
1344 			 * Try the uncontested case.  This should be
1345 			 * done in userland.
1346 			 */
1347 			rv = casueword32(&m->m_owner, UMUTEX_UNOWNED,
1348 			    &owner, id);
1349 			/* The address was invalid. */
1350 			if (rv == -1)
1351 				return (EFAULT);
1352 
1353 			/* The acquire succeeded. */
1354 			if (rv == 0) {
1355 				MPASS(owner == UMUTEX_UNOWNED);
1356 				return (0);
1357 			}
1358 
1359 			/*
1360 			 * If no one owns it but it is contested try
1361 			 * to acquire it.
1362 			 */
1363 			MPASS(rv == 1);
1364 			if (owner == UMUTEX_CONTESTED) {
1365 				rv = casueword32(&m->m_owner,
1366 				    UMUTEX_CONTESTED, &owner,
1367 				    id | UMUTEX_CONTESTED);
1368 				/* The address was invalid. */
1369 				if (rv == -1)
1370 					return (EFAULT);
1371 				if (rv == 0) {
1372 					MPASS(owner == UMUTEX_CONTESTED);
1373 					return (0);
1374 				}
1375 				if (rv == 1) {
1376 					rv = thread_check_susp(td, false);
1377 					if (rv != 0)
1378 						return (rv);
1379 				}
1380 
1381 				/*
1382 				 * If this failed the lock has
1383 				 * changed, restart.
1384 				 */
1385 				continue;
1386 			}
1387 
1388 			/* rv == 1 but not contested, likely store failure */
1389 			rv = thread_check_susp(td, false);
1390 			if (rv != 0)
1391 				return (rv);
1392 		}
1393 
1394 		if (mode == _UMUTEX_TRY)
1395 			return (EBUSY);
1396 
1397 		/*
1398 		 * If we caught a signal, we have retried and now
1399 		 * exit immediately.
1400 		 */
1401 		if (error != 0)
1402 			return (error);
1403 
1404 		if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX,
1405 		    GET_SHARE(flags), &uq->uq_key)) != 0)
1406 			return (error);
1407 
1408 		umtxq_lock(&uq->uq_key);
1409 		umtxq_busy(&uq->uq_key);
1410 		umtxq_insert(uq);
1411 		umtxq_unlock(&uq->uq_key);
1412 
1413 		/*
1414 		 * Set the contested bit so that a release in user space
1415 		 * knows to use the system call for unlock.  If this fails
1416 		 * either some one else has acquired the lock or it has been
1417 		 * released.
1418 		 */
1419 		rv = casueword32(&m->m_owner, owner, &old,
1420 		    owner | UMUTEX_CONTESTED);
1421 
1422 		/* The address was invalid or casueword failed to store. */
1423 		if (rv == -1 || rv == 1) {
1424 			umtxq_lock(&uq->uq_key);
1425 			umtxq_remove(uq);
1426 			umtxq_unbusy(&uq->uq_key);
1427 			umtxq_unlock(&uq->uq_key);
1428 			umtx_key_release(&uq->uq_key);
1429 			if (rv == -1)
1430 				return (EFAULT);
1431 			if (rv == 1) {
1432 				rv = thread_check_susp(td, false);
1433 				if (rv != 0)
1434 					return (rv);
1435 			}
1436 			continue;
1437 		}
1438 
1439 		/*
1440 		 * We set the contested bit, sleep. Otherwise the lock changed
1441 		 * and we need to retry or we lost a race to the thread
1442 		 * unlocking the umtx.
1443 		 */
1444 		umtxq_lock(&uq->uq_key);
1445 		umtxq_unbusy(&uq->uq_key);
1446 		MPASS(old == owner);
1447 		error = umtxq_sleep(uq, "umtxn", timeout == NULL ?
1448 		    NULL : &timo);
1449 		umtxq_remove(uq);
1450 		umtxq_unlock(&uq->uq_key);
1451 		umtx_key_release(&uq->uq_key);
1452 
1453 		if (error == 0)
1454 			error = thread_check_susp(td, false);
1455 	}
1456 
1457 	return (0);
1458 }
1459 
1460 /*
1461  * Unlock PTHREAD_PRIO_NONE protocol POSIX mutex.
1462  */
1463 static int
1464 do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1465 {
1466 	struct umtx_key key;
1467 	uint32_t owner, old, id, newlock;
1468 	int error, count;
1469 
1470 	id = td->td_tid;
1471 
1472 again:
1473 	/*
1474 	 * Make sure we own this mtx.
1475 	 */
1476 	error = fueword32(&m->m_owner, &owner);
1477 	if (error == -1)
1478 		return (EFAULT);
1479 
1480 	if ((owner & ~UMUTEX_CONTESTED) != id)
1481 		return (EPERM);
1482 
1483 	newlock = umtx_unlock_val(flags, rb);
1484 	if ((owner & UMUTEX_CONTESTED) == 0) {
1485 		error = casueword32(&m->m_owner, owner, &old, newlock);
1486 		if (error == -1)
1487 			return (EFAULT);
1488 		if (error == 1) {
1489 			error = thread_check_susp(td, false);
1490 			if (error != 0)
1491 				return (error);
1492 			goto again;
1493 		}
1494 		MPASS(old == owner);
1495 		return (0);
1496 	}
1497 
1498 	/* We should only ever be in here for contested locks */
1499 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1500 	    &key)) != 0)
1501 		return (error);
1502 
1503 	umtxq_lock(&key);
1504 	umtxq_busy(&key);
1505 	count = umtxq_count(&key);
1506 	umtxq_unlock(&key);
1507 
1508 	/*
1509 	 * When unlocking the umtx, it must be marked as unowned if
1510 	 * there is zero or one thread only waiting for it.
1511 	 * Otherwise, it must be marked as contested.
1512 	 */
1513 	if (count > 1)
1514 		newlock |= UMUTEX_CONTESTED;
1515 	error = casueword32(&m->m_owner, owner, &old, newlock);
1516 	umtxq_lock(&key);
1517 	umtxq_signal(&key, 1);
1518 	umtxq_unbusy(&key);
1519 	umtxq_unlock(&key);
1520 	umtx_key_release(&key);
1521 	if (error == -1)
1522 		return (EFAULT);
1523 	if (error == 1) {
1524 		if (old != owner)
1525 			return (EINVAL);
1526 		error = thread_check_susp(td, false);
1527 		if (error != 0)
1528 			return (error);
1529 		goto again;
1530 	}
1531 	return (0);
1532 }
1533 
1534 /*
1535  * Check if the mutex is available and wake up a waiter,
1536  * only for simple mutex.
1537  */
1538 static int
1539 do_wake_umutex(struct thread *td, struct umutex *m)
1540 {
1541 	struct umtx_key key;
1542 	uint32_t owner;
1543 	uint32_t flags;
1544 	int error;
1545 	int count;
1546 
1547 again:
1548 	error = fueword32(&m->m_owner, &owner);
1549 	if (error == -1)
1550 		return (EFAULT);
1551 
1552 	if ((owner & ~UMUTEX_CONTESTED) != 0 && owner != UMUTEX_RB_OWNERDEAD &&
1553 	    owner != UMUTEX_RB_NOTRECOV)
1554 		return (0);
1555 
1556 	error = fueword32(&m->m_flags, &flags);
1557 	if (error == -1)
1558 		return (EFAULT);
1559 
1560 	/* We should only ever be in here for contested locks */
1561 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1562 	    &key)) != 0)
1563 		return (error);
1564 
1565 	umtxq_lock(&key);
1566 	umtxq_busy(&key);
1567 	count = umtxq_count(&key);
1568 	umtxq_unlock(&key);
1569 
1570 	if (count <= 1 && owner != UMUTEX_RB_OWNERDEAD &&
1571 	    owner != UMUTEX_RB_NOTRECOV) {
1572 		error = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
1573 		    UMUTEX_UNOWNED);
1574 		if (error == -1) {
1575 			error = EFAULT;
1576 		} else if (error == 1) {
1577 			umtxq_lock(&key);
1578 			umtxq_unbusy(&key);
1579 			umtxq_unlock(&key);
1580 			umtx_key_release(&key);
1581 			error = thread_check_susp(td, false);
1582 			if (error != 0)
1583 				return (error);
1584 			goto again;
1585 		}
1586 	}
1587 
1588 	umtxq_lock(&key);
1589 	if (error == 0 && count != 0) {
1590 		MPASS((owner & ~UMUTEX_CONTESTED) == 0 ||
1591 		    owner == UMUTEX_RB_OWNERDEAD ||
1592 		    owner == UMUTEX_RB_NOTRECOV);
1593 		umtxq_signal(&key, 1);
1594 	}
1595 	umtxq_unbusy(&key);
1596 	umtxq_unlock(&key);
1597 	umtx_key_release(&key);
1598 	return (error);
1599 }
1600 
1601 /*
1602  * Check if the mutex has waiters and tries to fix contention bit.
1603  */
1604 static int
1605 do_wake2_umutex(struct thread *td, struct umutex *m, uint32_t flags)
1606 {
1607 	struct umtx_key key;
1608 	uint32_t owner, old;
1609 	int type;
1610 	int error;
1611 	int count;
1612 
1613 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT |
1614 	    UMUTEX_ROBUST)) {
1615 	case 0:
1616 	case UMUTEX_ROBUST:
1617 		type = TYPE_NORMAL_UMUTEX;
1618 		break;
1619 	case UMUTEX_PRIO_INHERIT:
1620 		type = TYPE_PI_UMUTEX;
1621 		break;
1622 	case (UMUTEX_PRIO_INHERIT | UMUTEX_ROBUST):
1623 		type = TYPE_PI_ROBUST_UMUTEX;
1624 		break;
1625 	case UMUTEX_PRIO_PROTECT:
1626 		type = TYPE_PP_UMUTEX;
1627 		break;
1628 	case (UMUTEX_PRIO_PROTECT | UMUTEX_ROBUST):
1629 		type = TYPE_PP_ROBUST_UMUTEX;
1630 		break;
1631 	default:
1632 		return (EINVAL);
1633 	}
1634 	if ((error = umtx_key_get(m, type, GET_SHARE(flags), &key)) != 0)
1635 		return (error);
1636 
1637 	owner = 0;
1638 	umtxq_lock(&key);
1639 	umtxq_busy(&key);
1640 	count = umtxq_count(&key);
1641 	umtxq_unlock(&key);
1642 
1643 	error = fueword32(&m->m_owner, &owner);
1644 	if (error == -1)
1645 		error = EFAULT;
1646 
1647 	/*
1648 	 * Only repair contention bit if there is a waiter, this means
1649 	 * the mutex is still being referenced by userland code,
1650 	 * otherwise don't update any memory.
1651 	 */
1652 	while (error == 0 && (owner & UMUTEX_CONTESTED) == 0 &&
1653 	    (count > 1 || (count == 1 && (owner & ~UMUTEX_CONTESTED) != 0))) {
1654 		error = casueword32(&m->m_owner, owner, &old,
1655 		    owner | UMUTEX_CONTESTED);
1656 		if (error == -1) {
1657 			error = EFAULT;
1658 			break;
1659 		}
1660 		if (error == 0) {
1661 			MPASS(old == owner);
1662 			break;
1663 		}
1664 		owner = old;
1665 		error = thread_check_susp(td, false);
1666 	}
1667 
1668 	umtxq_lock(&key);
1669 	if (error == EFAULT) {
1670 		umtxq_signal(&key, INT_MAX);
1671 	} else if (count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1672 	    owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1673 		umtxq_signal(&key, 1);
1674 	umtxq_unbusy(&key);
1675 	umtxq_unlock(&key);
1676 	umtx_key_release(&key);
1677 	return (error);
1678 }
1679 
1680 struct umtx_pi *
1681 umtx_pi_alloc(int flags)
1682 {
1683 	struct umtx_pi *pi;
1684 
1685 	pi = uma_zalloc(umtx_pi_zone, M_ZERO | flags);
1686 	TAILQ_INIT(&pi->pi_blocked);
1687 	atomic_add_int(&umtx_pi_allocated, 1);
1688 	return (pi);
1689 }
1690 
1691 void
1692 umtx_pi_free(struct umtx_pi *pi)
1693 {
1694 	uma_zfree(umtx_pi_zone, pi);
1695 	atomic_add_int(&umtx_pi_allocated, -1);
1696 }
1697 
1698 /*
1699  * Adjust the thread's position on a pi_state after its priority has been
1700  * changed.
1701  */
1702 static int
1703 umtx_pi_adjust_thread(struct umtx_pi *pi, struct thread *td)
1704 {
1705 	struct umtx_q *uq, *uq1, *uq2;
1706 	struct thread *td1;
1707 
1708 	mtx_assert(&umtx_lock, MA_OWNED);
1709 	if (pi == NULL)
1710 		return (0);
1711 
1712 	uq = td->td_umtxq;
1713 
1714 	/*
1715 	 * Check if the thread needs to be moved on the blocked chain.
1716 	 * It needs to be moved if either its priority is lower than
1717 	 * the previous thread or higher than the next thread.
1718 	 */
1719 	uq1 = TAILQ_PREV(uq, umtxq_head, uq_lockq);
1720 	uq2 = TAILQ_NEXT(uq, uq_lockq);
1721 	if ((uq1 != NULL && UPRI(td) < UPRI(uq1->uq_thread)) ||
1722 	    (uq2 != NULL && UPRI(td) > UPRI(uq2->uq_thread))) {
1723 		/*
1724 		 * Remove thread from blocked chain and determine where
1725 		 * it should be moved to.
1726 		 */
1727 		TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1728 		TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1729 			td1 = uq1->uq_thread;
1730 			MPASS(td1->td_proc->p_magic == P_MAGIC);
1731 			if (UPRI(td1) > UPRI(td))
1732 				break;
1733 		}
1734 
1735 		if (uq1 == NULL)
1736 			TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1737 		else
1738 			TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1739 	}
1740 	return (1);
1741 }
1742 
1743 static struct umtx_pi *
1744 umtx_pi_next(struct umtx_pi *pi)
1745 {
1746 	struct umtx_q *uq_owner;
1747 
1748 	if (pi->pi_owner == NULL)
1749 		return (NULL);
1750 	uq_owner = pi->pi_owner->td_umtxq;
1751 	if (uq_owner == NULL)
1752 		return (NULL);
1753 	return (uq_owner->uq_pi_blocked);
1754 }
1755 
1756 /*
1757  * Floyd's Cycle-Finding Algorithm.
1758  */
1759 static bool
1760 umtx_pi_check_loop(struct umtx_pi *pi)
1761 {
1762 	struct umtx_pi *pi1;	/* fast iterator */
1763 
1764 	mtx_assert(&umtx_lock, MA_OWNED);
1765 	if (pi == NULL)
1766 		return (false);
1767 	pi1 = pi;
1768 	for (;;) {
1769 		pi = umtx_pi_next(pi);
1770 		if (pi == NULL)
1771 			break;
1772 		pi1 = umtx_pi_next(pi1);
1773 		if (pi1 == NULL)
1774 			break;
1775 		pi1 = umtx_pi_next(pi1);
1776 		if (pi1 == NULL)
1777 			break;
1778 		if (pi == pi1)
1779 			return (true);
1780 	}
1781 	return (false);
1782 }
1783 
1784 /*
1785  * Propagate priority when a thread is blocked on POSIX
1786  * PI mutex.
1787  */
1788 static void
1789 umtx_propagate_priority(struct thread *td)
1790 {
1791 	struct umtx_q *uq;
1792 	struct umtx_pi *pi;
1793 	int pri;
1794 
1795 	mtx_assert(&umtx_lock, MA_OWNED);
1796 	pri = UPRI(td);
1797 	uq = td->td_umtxq;
1798 	pi = uq->uq_pi_blocked;
1799 	if (pi == NULL)
1800 		return;
1801 	if (umtx_pi_check_loop(pi))
1802 		return;
1803 
1804 	for (;;) {
1805 		td = pi->pi_owner;
1806 		if (td == NULL || td == curthread)
1807 			return;
1808 
1809 		MPASS(td->td_proc != NULL);
1810 		MPASS(td->td_proc->p_magic == P_MAGIC);
1811 
1812 		thread_lock(td);
1813 		if (td->td_lend_user_pri > pri)
1814 			sched_lend_user_prio(td, pri);
1815 		else {
1816 			thread_unlock(td);
1817 			break;
1818 		}
1819 		thread_unlock(td);
1820 
1821 		/*
1822 		 * Pick up the lock that td is blocked on.
1823 		 */
1824 		uq = td->td_umtxq;
1825 		pi = uq->uq_pi_blocked;
1826 		if (pi == NULL)
1827 			break;
1828 		/* Resort td on the list if needed. */
1829 		umtx_pi_adjust_thread(pi, td);
1830 	}
1831 }
1832 
1833 /*
1834  * Unpropagate priority for a PI mutex when a thread blocked on
1835  * it is interrupted by signal or resumed by others.
1836  */
1837 static void
1838 umtx_repropagate_priority(struct umtx_pi *pi)
1839 {
1840 	struct umtx_q *uq, *uq_owner;
1841 	struct umtx_pi *pi2;
1842 	int pri;
1843 
1844 	mtx_assert(&umtx_lock, MA_OWNED);
1845 
1846 	if (umtx_pi_check_loop(pi))
1847 		return;
1848 	while (pi != NULL && pi->pi_owner != NULL) {
1849 		pri = PRI_MAX;
1850 		uq_owner = pi->pi_owner->td_umtxq;
1851 
1852 		TAILQ_FOREACH(pi2, &uq_owner->uq_pi_contested, pi_link) {
1853 			uq = TAILQ_FIRST(&pi2->pi_blocked);
1854 			if (uq != NULL) {
1855 				if (pri > UPRI(uq->uq_thread))
1856 					pri = UPRI(uq->uq_thread);
1857 			}
1858 		}
1859 
1860 		if (pri > uq_owner->uq_inherited_pri)
1861 			pri = uq_owner->uq_inherited_pri;
1862 		thread_lock(pi->pi_owner);
1863 		sched_lend_user_prio(pi->pi_owner, pri);
1864 		thread_unlock(pi->pi_owner);
1865 		if ((pi = uq_owner->uq_pi_blocked) != NULL)
1866 			umtx_pi_adjust_thread(pi, uq_owner->uq_thread);
1867 	}
1868 }
1869 
1870 /*
1871  * Insert a PI mutex into owned list.
1872  */
1873 static void
1874 umtx_pi_setowner(struct umtx_pi *pi, struct thread *owner)
1875 {
1876 	struct umtx_q *uq_owner;
1877 
1878 	uq_owner = owner->td_umtxq;
1879 	mtx_assert(&umtx_lock, MA_OWNED);
1880 	MPASS(pi->pi_owner == NULL);
1881 	pi->pi_owner = owner;
1882 	TAILQ_INSERT_TAIL(&uq_owner->uq_pi_contested, pi, pi_link);
1883 }
1884 
1885 /*
1886  * Disown a PI mutex, and remove it from the owned list.
1887  */
1888 static void
1889 umtx_pi_disown(struct umtx_pi *pi)
1890 {
1891 
1892 	mtx_assert(&umtx_lock, MA_OWNED);
1893 	TAILQ_REMOVE(&pi->pi_owner->td_umtxq->uq_pi_contested, pi, pi_link);
1894 	pi->pi_owner = NULL;
1895 }
1896 
1897 /*
1898  * Claim ownership of a PI mutex.
1899  */
1900 int
1901 umtx_pi_claim(struct umtx_pi *pi, struct thread *owner)
1902 {
1903 	struct umtx_q *uq;
1904 	int pri;
1905 
1906 	mtx_lock(&umtx_lock);
1907 	if (pi->pi_owner == owner) {
1908 		mtx_unlock(&umtx_lock);
1909 		return (0);
1910 	}
1911 
1912 	if (pi->pi_owner != NULL) {
1913 		/*
1914 		 * userland may have already messed the mutex, sigh.
1915 		 */
1916 		mtx_unlock(&umtx_lock);
1917 		return (EPERM);
1918 	}
1919 	umtx_pi_setowner(pi, owner);
1920 	uq = TAILQ_FIRST(&pi->pi_blocked);
1921 	if (uq != NULL) {
1922 		pri = UPRI(uq->uq_thread);
1923 		thread_lock(owner);
1924 		if (pri < UPRI(owner))
1925 			sched_lend_user_prio(owner, pri);
1926 		thread_unlock(owner);
1927 	}
1928 	mtx_unlock(&umtx_lock);
1929 	return (0);
1930 }
1931 
1932 /*
1933  * Adjust a thread's order position in its blocked PI mutex,
1934  * this may result new priority propagating process.
1935  */
1936 void
1937 umtx_pi_adjust(struct thread *td, u_char oldpri)
1938 {
1939 	struct umtx_q *uq;
1940 	struct umtx_pi *pi;
1941 
1942 	uq = td->td_umtxq;
1943 	mtx_lock(&umtx_lock);
1944 	/*
1945 	 * Pick up the lock that td is blocked on.
1946 	 */
1947 	pi = uq->uq_pi_blocked;
1948 	if (pi != NULL) {
1949 		umtx_pi_adjust_thread(pi, td);
1950 		umtx_repropagate_priority(pi);
1951 	}
1952 	mtx_unlock(&umtx_lock);
1953 }
1954 
1955 /*
1956  * Sleep on a PI mutex.
1957  */
1958 int
1959 umtxq_sleep_pi(struct umtx_q *uq, struct umtx_pi *pi, uint32_t owner,
1960     const char *wmesg, struct umtx_abs_timeout *timo, bool shared)
1961 {
1962 	struct thread *td, *td1;
1963 	struct umtx_q *uq1;
1964 	int error, pri;
1965 #ifdef INVARIANTS
1966 	struct umtxq_chain *uc;
1967 
1968 	uc = umtxq_getchain(&pi->pi_key);
1969 #endif
1970 	error = 0;
1971 	td = uq->uq_thread;
1972 	KASSERT(td == curthread, ("inconsistent uq_thread"));
1973 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
1974 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));
1975 	umtxq_insert(uq);
1976 	mtx_lock(&umtx_lock);
1977 	if (pi->pi_owner == NULL) {
1978 		mtx_unlock(&umtx_lock);
1979 		td1 = tdfind(owner, shared ? -1 : td->td_proc->p_pid);
1980 		mtx_lock(&umtx_lock);
1981 		if (td1 != NULL) {
1982 			if (pi->pi_owner == NULL)
1983 				umtx_pi_setowner(pi, td1);
1984 			PROC_UNLOCK(td1->td_proc);
1985 		}
1986 	}
1987 
1988 	TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1989 		pri = UPRI(uq1->uq_thread);
1990 		if (pri > UPRI(td))
1991 			break;
1992 	}
1993 
1994 	if (uq1 != NULL)
1995 		TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1996 	else
1997 		TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1998 
1999 	uq->uq_pi_blocked = pi;
2000 	thread_lock(td);
2001 	td->td_flags |= TDF_UPIBLOCKED;
2002 	thread_unlock(td);
2003 	umtx_propagate_priority(td);
2004 	mtx_unlock(&umtx_lock);
2005 	umtxq_unbusy(&uq->uq_key);
2006 
2007 	error = umtxq_sleep(uq, wmesg, timo);
2008 	umtxq_remove(uq);
2009 
2010 	mtx_lock(&umtx_lock);
2011 	uq->uq_pi_blocked = NULL;
2012 	thread_lock(td);
2013 	td->td_flags &= ~TDF_UPIBLOCKED;
2014 	thread_unlock(td);
2015 	TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
2016 	umtx_repropagate_priority(pi);
2017 	mtx_unlock(&umtx_lock);
2018 	umtxq_unlock(&uq->uq_key);
2019 
2020 	return (error);
2021 }
2022 
2023 /*
2024  * Add reference count for a PI mutex.
2025  */
2026 void
2027 umtx_pi_ref(struct umtx_pi *pi)
2028 {
2029 
2030 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&pi->pi_key));
2031 	pi->pi_refcount++;
2032 }
2033 
2034 /*
2035  * Decrease reference count for a PI mutex, if the counter
2036  * is decreased to zero, its memory space is freed.
2037  */
2038 void
2039 umtx_pi_unref(struct umtx_pi *pi)
2040 {
2041 	struct umtxq_chain *uc;
2042 
2043 	uc = umtxq_getchain(&pi->pi_key);
2044 	UMTXQ_LOCKED_ASSERT(uc);
2045 	KASSERT(pi->pi_refcount > 0, ("invalid reference count"));
2046 	if (--pi->pi_refcount == 0) {
2047 		mtx_lock(&umtx_lock);
2048 		if (pi->pi_owner != NULL)
2049 			umtx_pi_disown(pi);
2050 		KASSERT(TAILQ_EMPTY(&pi->pi_blocked),
2051 			("blocked queue not empty"));
2052 		mtx_unlock(&umtx_lock);
2053 		TAILQ_REMOVE(&uc->uc_pi_list, pi, pi_hashlink);
2054 		umtx_pi_free(pi);
2055 	}
2056 }
2057 
2058 /*
2059  * Find a PI mutex in hash table.
2060  */
2061 struct umtx_pi *
2062 umtx_pi_lookup(struct umtx_key *key)
2063 {
2064 	struct umtxq_chain *uc;
2065 	struct umtx_pi *pi;
2066 
2067 	uc = umtxq_getchain(key);
2068 	UMTXQ_LOCKED_ASSERT(uc);
2069 
2070 	TAILQ_FOREACH(pi, &uc->uc_pi_list, pi_hashlink) {
2071 		if (umtx_key_match(&pi->pi_key, key)) {
2072 			return (pi);
2073 		}
2074 	}
2075 	return (NULL);
2076 }
2077 
2078 /*
2079  * Insert a PI mutex into hash table.
2080  */
2081 void
2082 umtx_pi_insert(struct umtx_pi *pi)
2083 {
2084 	struct umtxq_chain *uc;
2085 
2086 	uc = umtxq_getchain(&pi->pi_key);
2087 	UMTXQ_LOCKED_ASSERT(uc);
2088 	TAILQ_INSERT_TAIL(&uc->uc_pi_list, pi, pi_hashlink);
2089 }
2090 
2091 /*
2092  * Drop a PI mutex and wakeup a top waiter.
2093  */
2094 int
2095 umtx_pi_drop(struct thread *td, struct umtx_key *key, bool rb, int *count)
2096 {
2097 	struct umtx_q *uq_first, *uq_first2, *uq_me;
2098 	struct umtx_pi *pi, *pi2;
2099 	int pri;
2100 
2101 	UMTXQ_ASSERT_LOCKED_BUSY(key);
2102 	*count = umtxq_count_pi(key, &uq_first);
2103 	if (uq_first != NULL) {
2104 		mtx_lock(&umtx_lock);
2105 		pi = uq_first->uq_pi_blocked;
2106 		KASSERT(pi != NULL, ("pi == NULL?"));
2107 		if (pi->pi_owner != td && !(rb && pi->pi_owner == NULL)) {
2108 			mtx_unlock(&umtx_lock);
2109 			/* userland messed the mutex */
2110 			return (EPERM);
2111 		}
2112 		uq_me = td->td_umtxq;
2113 		if (pi->pi_owner == td)
2114 			umtx_pi_disown(pi);
2115 		/* get highest priority thread which is still sleeping. */
2116 		uq_first = TAILQ_FIRST(&pi->pi_blocked);
2117 		while (uq_first != NULL &&
2118 		    (uq_first->uq_flags & UQF_UMTXQ) == 0) {
2119 			uq_first = TAILQ_NEXT(uq_first, uq_lockq);
2120 		}
2121 		pri = PRI_MAX;
2122 		TAILQ_FOREACH(pi2, &uq_me->uq_pi_contested, pi_link) {
2123 			uq_first2 = TAILQ_FIRST(&pi2->pi_blocked);
2124 			if (uq_first2 != NULL) {
2125 				if (pri > UPRI(uq_first2->uq_thread))
2126 					pri = UPRI(uq_first2->uq_thread);
2127 			}
2128 		}
2129 		thread_lock(td);
2130 		sched_lend_user_prio(td, pri);
2131 		thread_unlock(td);
2132 		mtx_unlock(&umtx_lock);
2133 		if (uq_first)
2134 			umtxq_signal_thread(uq_first);
2135 	} else {
2136 		pi = umtx_pi_lookup(key);
2137 		/*
2138 		 * A umtx_pi can exist if a signal or timeout removed the
2139 		 * last waiter from the umtxq, but there is still
2140 		 * a thread in do_lock_pi() holding the umtx_pi.
2141 		 */
2142 		if (pi != NULL) {
2143 			/*
2144 			 * The umtx_pi can be unowned, such as when a thread
2145 			 * has just entered do_lock_pi(), allocated the
2146 			 * umtx_pi, and unlocked the umtxq.
2147 			 * If the current thread owns it, it must disown it.
2148 			 */
2149 			mtx_lock(&umtx_lock);
2150 			if (pi->pi_owner == td)
2151 				umtx_pi_disown(pi);
2152 			mtx_unlock(&umtx_lock);
2153 		}
2154 	}
2155 	return (0);
2156 }
2157 
2158 /*
2159  * Lock a PI mutex.
2160  */
2161 static int
2162 do_lock_pi(struct thread *td, struct umutex *m, uint32_t flags,
2163     struct _umtx_time *timeout, int try)
2164 {
2165 	struct umtx_abs_timeout timo;
2166 	struct umtx_q *uq;
2167 	struct umtx_pi *pi, *new_pi;
2168 	uint32_t id, old_owner, owner, old;
2169 	int error, rv;
2170 
2171 	id = td->td_tid;
2172 	uq = td->td_umtxq;
2173 
2174 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2175 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2176 	    &uq->uq_key)) != 0)
2177 		return (error);
2178 
2179 	if (timeout != NULL)
2180 		umtx_abs_timeout_init2(&timo, timeout);
2181 
2182 	umtxq_lock(&uq->uq_key);
2183 	pi = umtx_pi_lookup(&uq->uq_key);
2184 	if (pi == NULL) {
2185 		new_pi = umtx_pi_alloc(M_NOWAIT);
2186 		if (new_pi == NULL) {
2187 			umtxq_unlock(&uq->uq_key);
2188 			new_pi = umtx_pi_alloc(M_WAITOK);
2189 			umtxq_lock(&uq->uq_key);
2190 			pi = umtx_pi_lookup(&uq->uq_key);
2191 			if (pi != NULL) {
2192 				umtx_pi_free(new_pi);
2193 				new_pi = NULL;
2194 			}
2195 		}
2196 		if (new_pi != NULL) {
2197 			new_pi->pi_key = uq->uq_key;
2198 			umtx_pi_insert(new_pi);
2199 			pi = new_pi;
2200 		}
2201 	}
2202 	umtx_pi_ref(pi);
2203 	umtxq_unlock(&uq->uq_key);
2204 
2205 	/*
2206 	 * Care must be exercised when dealing with umtx structure.  It
2207 	 * can fault on any access.
2208 	 */
2209 	for (;;) {
2210 		/*
2211 		 * Try the uncontested case.  This should be done in userland.
2212 		 */
2213 		rv = casueword32(&m->m_owner, UMUTEX_UNOWNED, &owner, id);
2214 		/* The address was invalid. */
2215 		if (rv == -1) {
2216 			error = EFAULT;
2217 			break;
2218 		}
2219 		/* The acquire succeeded. */
2220 		if (rv == 0) {
2221 			MPASS(owner == UMUTEX_UNOWNED);
2222 			error = 0;
2223 			break;
2224 		}
2225 
2226 		if (owner == UMUTEX_RB_NOTRECOV) {
2227 			error = ENOTRECOVERABLE;
2228 			break;
2229 		}
2230 
2231 		/*
2232 		 * Avoid overwriting a possible error from sleep due
2233 		 * to the pending signal with suspension check result.
2234 		 */
2235 		if (error == 0) {
2236 			error = thread_check_susp(td, true);
2237 			if (error != 0)
2238 				break;
2239 		}
2240 
2241 		/* If no one owns it but it is contested try to acquire it. */
2242 		if (owner == UMUTEX_CONTESTED || owner == UMUTEX_RB_OWNERDEAD) {
2243 			old_owner = owner;
2244 			rv = casueword32(&m->m_owner, owner, &owner,
2245 			    id | UMUTEX_CONTESTED);
2246 			/* The address was invalid. */
2247 			if (rv == -1) {
2248 				error = EFAULT;
2249 				break;
2250 			}
2251 			if (rv == 1) {
2252 				if (error == 0) {
2253 					error = thread_check_susp(td, true);
2254 					if (error != 0)
2255 						break;
2256 				}
2257 
2258 				/*
2259 				 * If this failed the lock could
2260 				 * changed, restart.
2261 				 */
2262 				continue;
2263 			}
2264 
2265 			MPASS(rv == 0);
2266 			MPASS(owner == old_owner);
2267 			umtxq_lock(&uq->uq_key);
2268 			umtxq_busy(&uq->uq_key);
2269 			error = umtx_pi_claim(pi, td);
2270 			umtxq_unbusy(&uq->uq_key);
2271 			umtxq_unlock(&uq->uq_key);
2272 			if (error != 0) {
2273 				/*
2274 				 * Since we're going to return an
2275 				 * error, restore the m_owner to its
2276 				 * previous, unowned state to avoid
2277 				 * compounding the problem.
2278 				 */
2279 				(void)casuword32(&m->m_owner,
2280 				    id | UMUTEX_CONTESTED, old_owner);
2281 			}
2282 			if (error == 0 && old_owner == UMUTEX_RB_OWNERDEAD)
2283 				error = EOWNERDEAD;
2284 			break;
2285 		}
2286 
2287 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2288 			error = EDEADLK;
2289 			break;
2290 		}
2291 
2292 		if (try != 0) {
2293 			error = EBUSY;
2294 			break;
2295 		}
2296 
2297 		/*
2298 		 * If we caught a signal, we have retried and now
2299 		 * exit immediately.
2300 		 */
2301 		if (error != 0)
2302 			break;
2303 
2304 		umtxq_lock(&uq->uq_key);
2305 		umtxq_busy(&uq->uq_key);
2306 		umtxq_unlock(&uq->uq_key);
2307 
2308 		/*
2309 		 * Set the contested bit so that a release in user space
2310 		 * knows to use the system call for unlock.  If this fails
2311 		 * either some one else has acquired the lock or it has been
2312 		 * released.
2313 		 */
2314 		rv = casueword32(&m->m_owner, owner, &old, owner |
2315 		    UMUTEX_CONTESTED);
2316 
2317 		/* The address was invalid. */
2318 		if (rv == -1) {
2319 			umtxq_unbusy_unlocked(&uq->uq_key);
2320 			error = EFAULT;
2321 			break;
2322 		}
2323 		if (rv == 1) {
2324 			umtxq_unbusy_unlocked(&uq->uq_key);
2325 			error = thread_check_susp(td, true);
2326 			if (error != 0)
2327 				break;
2328 
2329 			/*
2330 			 * The lock changed and we need to retry or we
2331 			 * lost a race to the thread unlocking the
2332 			 * umtx.  Note that the UMUTEX_RB_OWNERDEAD
2333 			 * value for owner is impossible there.
2334 			 */
2335 			continue;
2336 		}
2337 
2338 		umtxq_lock(&uq->uq_key);
2339 
2340 		/* We set the contested bit, sleep. */
2341 		MPASS(old == owner);
2342 		error = umtxq_sleep_pi(uq, pi, owner & ~UMUTEX_CONTESTED,
2343 		    "umtxpi", timeout == NULL ? NULL : &timo,
2344 		    (flags & USYNC_PROCESS_SHARED) != 0);
2345 		if (error != 0)
2346 			continue;
2347 
2348 		error = thread_check_susp(td, false);
2349 		if (error != 0)
2350 			break;
2351 	}
2352 
2353 	umtxq_lock(&uq->uq_key);
2354 	umtx_pi_unref(pi);
2355 	umtxq_unlock(&uq->uq_key);
2356 
2357 	umtx_key_release(&uq->uq_key);
2358 	return (error);
2359 }
2360 
2361 /*
2362  * Unlock a PI mutex.
2363  */
2364 static int
2365 do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2366 {
2367 	struct umtx_key key;
2368 	uint32_t id, new_owner, old, owner;
2369 	int count, error;
2370 
2371 	id = td->td_tid;
2372 
2373 usrloop:
2374 	/*
2375 	 * Make sure we own this mtx.
2376 	 */
2377 	error = fueword32(&m->m_owner, &owner);
2378 	if (error == -1)
2379 		return (EFAULT);
2380 
2381 	if ((owner & ~UMUTEX_CONTESTED) != id)
2382 		return (EPERM);
2383 
2384 	new_owner = umtx_unlock_val(flags, rb);
2385 
2386 	/* This should be done in userland */
2387 	if ((owner & UMUTEX_CONTESTED) == 0) {
2388 		error = casueword32(&m->m_owner, owner, &old, new_owner);
2389 		if (error == -1)
2390 			return (EFAULT);
2391 		if (error == 1) {
2392 			error = thread_check_susp(td, true);
2393 			if (error != 0)
2394 				return (error);
2395 			goto usrloop;
2396 		}
2397 		if (old == owner)
2398 			return (0);
2399 		owner = old;
2400 	}
2401 
2402 	/* We should only ever be in here for contested locks */
2403 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2404 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2405 	    &key)) != 0)
2406 		return (error);
2407 
2408 	umtxq_lock(&key);
2409 	umtxq_busy(&key);
2410 	error = umtx_pi_drop(td, &key, rb, &count);
2411 	if (error != 0) {
2412 		umtxq_unbusy(&key);
2413 		umtxq_unlock(&key);
2414 		umtx_key_release(&key);
2415 		/* userland messed the mutex */
2416 		return (error);
2417 	}
2418 	umtxq_unlock(&key);
2419 
2420 	/*
2421 	 * When unlocking the umtx, it must be marked as unowned if
2422 	 * there is zero or one thread only waiting for it.
2423 	 * Otherwise, it must be marked as contested.
2424 	 */
2425 
2426 	if (count > 1)
2427 		new_owner |= UMUTEX_CONTESTED;
2428 again:
2429 	error = casueword32(&m->m_owner, owner, &old, new_owner);
2430 	if (error == 1) {
2431 		error = thread_check_susp(td, false);
2432 		if (error == 0)
2433 			goto again;
2434 	}
2435 	umtxq_unbusy_unlocked(&key);
2436 	umtx_key_release(&key);
2437 	if (error == -1)
2438 		return (EFAULT);
2439 	if (error == 0 && old != owner)
2440 		return (EINVAL);
2441 	return (error);
2442 }
2443 
2444 /*
2445  * Lock a PP mutex.
2446  */
2447 static int
2448 do_lock_pp(struct thread *td, struct umutex *m, uint32_t flags,
2449     struct _umtx_time *timeout, int try)
2450 {
2451 	struct umtx_abs_timeout timo;
2452 	struct umtx_q *uq, *uq2;
2453 	struct umtx_pi *pi;
2454 	uint32_t ceiling;
2455 	uint32_t owner, id;
2456 	int error, pri, old_inherited_pri, su, rv;
2457 
2458 	id = td->td_tid;
2459 	uq = td->td_umtxq;
2460 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2461 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2462 	    &uq->uq_key)) != 0)
2463 		return (error);
2464 
2465 	if (timeout != NULL)
2466 		umtx_abs_timeout_init2(&timo, timeout);
2467 
2468 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2469 	for (;;) {
2470 		old_inherited_pri = uq->uq_inherited_pri;
2471 		umtxq_lock(&uq->uq_key);
2472 		umtxq_busy(&uq->uq_key);
2473 		umtxq_unlock(&uq->uq_key);
2474 
2475 		rv = fueword32(&m->m_ceilings[0], &ceiling);
2476 		if (rv == -1) {
2477 			error = EFAULT;
2478 			goto out;
2479 		}
2480 		ceiling = RTP_PRIO_MAX - ceiling;
2481 		if (ceiling > RTP_PRIO_MAX) {
2482 			error = EINVAL;
2483 			goto out;
2484 		}
2485 
2486 		mtx_lock(&umtx_lock);
2487 		if (UPRI(td) < PRI_MIN_REALTIME + ceiling) {
2488 			mtx_unlock(&umtx_lock);
2489 			error = EINVAL;
2490 			goto out;
2491 		}
2492 		if (su && PRI_MIN_REALTIME + ceiling < uq->uq_inherited_pri) {
2493 			uq->uq_inherited_pri = PRI_MIN_REALTIME + ceiling;
2494 			thread_lock(td);
2495 			if (uq->uq_inherited_pri < UPRI(td))
2496 				sched_lend_user_prio(td, uq->uq_inherited_pri);
2497 			thread_unlock(td);
2498 		}
2499 		mtx_unlock(&umtx_lock);
2500 
2501 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2502 		    id | UMUTEX_CONTESTED);
2503 		/* The address was invalid. */
2504 		if (rv == -1) {
2505 			error = EFAULT;
2506 			break;
2507 		}
2508 		if (rv == 0) {
2509 			MPASS(owner == UMUTEX_CONTESTED);
2510 			error = 0;
2511 			break;
2512 		}
2513 		/* rv == 1 */
2514 		if (owner == UMUTEX_RB_OWNERDEAD) {
2515 			rv = casueword32(&m->m_owner, UMUTEX_RB_OWNERDEAD,
2516 			    &owner, id | UMUTEX_CONTESTED);
2517 			if (rv == -1) {
2518 				error = EFAULT;
2519 				break;
2520 			}
2521 			if (rv == 0) {
2522 				MPASS(owner == UMUTEX_RB_OWNERDEAD);
2523 				error = EOWNERDEAD; /* success */
2524 				break;
2525 			}
2526 
2527 			/*
2528 			 *  rv == 1, only check for suspension if we
2529 			 *  did not already catched a signal.  If we
2530 			 *  get an error from the check, the same
2531 			 *  condition is checked by the umtxq_sleep()
2532 			 *  call below, so we should obliterate the
2533 			 *  error to not skip the last loop iteration.
2534 			 */
2535 			if (error == 0) {
2536 				error = thread_check_susp(td, false);
2537 				if (error == 0) {
2538 					if (try != 0)
2539 						error = EBUSY;
2540 					else
2541 						continue;
2542 				}
2543 				error = 0;
2544 			}
2545 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2546 			error = ENOTRECOVERABLE;
2547 		}
2548 
2549 		if (try != 0)
2550 			error = EBUSY;
2551 
2552 		/*
2553 		 * If we caught a signal, we have retried and now
2554 		 * exit immediately.
2555 		 */
2556 		if (error != 0)
2557 			break;
2558 
2559 		umtxq_lock(&uq->uq_key);
2560 		umtxq_insert(uq);
2561 		umtxq_unbusy(&uq->uq_key);
2562 		error = umtxq_sleep(uq, "umtxpp", timeout == NULL ?
2563 		    NULL : &timo);
2564 		umtxq_remove(uq);
2565 		umtxq_unlock(&uq->uq_key);
2566 
2567 		mtx_lock(&umtx_lock);
2568 		uq->uq_inherited_pri = old_inherited_pri;
2569 		pri = PRI_MAX;
2570 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2571 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2572 			if (uq2 != NULL) {
2573 				if (pri > UPRI(uq2->uq_thread))
2574 					pri = UPRI(uq2->uq_thread);
2575 			}
2576 		}
2577 		if (pri > uq->uq_inherited_pri)
2578 			pri = uq->uq_inherited_pri;
2579 		thread_lock(td);
2580 		sched_lend_user_prio(td, pri);
2581 		thread_unlock(td);
2582 		mtx_unlock(&umtx_lock);
2583 	}
2584 
2585 	if (error != 0 && error != EOWNERDEAD) {
2586 		mtx_lock(&umtx_lock);
2587 		uq->uq_inherited_pri = old_inherited_pri;
2588 		pri = PRI_MAX;
2589 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2590 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2591 			if (uq2 != NULL) {
2592 				if (pri > UPRI(uq2->uq_thread))
2593 					pri = UPRI(uq2->uq_thread);
2594 			}
2595 		}
2596 		if (pri > uq->uq_inherited_pri)
2597 			pri = uq->uq_inherited_pri;
2598 		thread_lock(td);
2599 		sched_lend_user_prio(td, pri);
2600 		thread_unlock(td);
2601 		mtx_unlock(&umtx_lock);
2602 	}
2603 
2604 out:
2605 	umtxq_unbusy_unlocked(&uq->uq_key);
2606 	umtx_key_release(&uq->uq_key);
2607 	return (error);
2608 }
2609 
2610 /*
2611  * Unlock a PP mutex.
2612  */
2613 static int
2614 do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2615 {
2616 	struct umtx_key key;
2617 	struct umtx_q *uq, *uq2;
2618 	struct umtx_pi *pi;
2619 	uint32_t id, owner, rceiling;
2620 	int error, pri, new_inherited_pri, su;
2621 
2622 	id = td->td_tid;
2623 	uq = td->td_umtxq;
2624 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2625 
2626 	/*
2627 	 * Make sure we own this mtx.
2628 	 */
2629 	error = fueword32(&m->m_owner, &owner);
2630 	if (error == -1)
2631 		return (EFAULT);
2632 
2633 	if ((owner & ~UMUTEX_CONTESTED) != id)
2634 		return (EPERM);
2635 
2636 	error = copyin(&m->m_ceilings[1], &rceiling, sizeof(uint32_t));
2637 	if (error != 0)
2638 		return (error);
2639 
2640 	if (rceiling == -1)
2641 		new_inherited_pri = PRI_MAX;
2642 	else {
2643 		rceiling = RTP_PRIO_MAX - rceiling;
2644 		if (rceiling > RTP_PRIO_MAX)
2645 			return (EINVAL);
2646 		new_inherited_pri = PRI_MIN_REALTIME + rceiling;
2647 	}
2648 
2649 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2650 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2651 	    &key)) != 0)
2652 		return (error);
2653 	umtxq_lock(&key);
2654 	umtxq_busy(&key);
2655 	umtxq_unlock(&key);
2656 	/*
2657 	 * For priority protected mutex, always set unlocked state
2658 	 * to UMUTEX_CONTESTED, so that userland always enters kernel
2659 	 * to lock the mutex, it is necessary because thread priority
2660 	 * has to be adjusted for such mutex.
2661 	 */
2662 	error = suword32(&m->m_owner, umtx_unlock_val(flags, rb) |
2663 	    UMUTEX_CONTESTED);
2664 
2665 	umtxq_lock(&key);
2666 	if (error == 0)
2667 		umtxq_signal(&key, 1);
2668 	umtxq_unbusy(&key);
2669 	umtxq_unlock(&key);
2670 
2671 	if (error == -1)
2672 		error = EFAULT;
2673 	else {
2674 		mtx_lock(&umtx_lock);
2675 		if (su != 0)
2676 			uq->uq_inherited_pri = new_inherited_pri;
2677 		pri = PRI_MAX;
2678 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2679 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2680 			if (uq2 != NULL) {
2681 				if (pri > UPRI(uq2->uq_thread))
2682 					pri = UPRI(uq2->uq_thread);
2683 			}
2684 		}
2685 		if (pri > uq->uq_inherited_pri)
2686 			pri = uq->uq_inherited_pri;
2687 		thread_lock(td);
2688 		sched_lend_user_prio(td, pri);
2689 		thread_unlock(td);
2690 		mtx_unlock(&umtx_lock);
2691 	}
2692 	umtx_key_release(&key);
2693 	return (error);
2694 }
2695 
2696 static int
2697 do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
2698     uint32_t *old_ceiling)
2699 {
2700 	struct umtx_q *uq;
2701 	uint32_t flags, id, owner, save_ceiling;
2702 	int error, rv, rv1;
2703 
2704 	error = fueword32(&m->m_flags, &flags);
2705 	if (error == -1)
2706 		return (EFAULT);
2707 	if ((flags & UMUTEX_PRIO_PROTECT) == 0)
2708 		return (EINVAL);
2709 	if (ceiling > RTP_PRIO_MAX)
2710 		return (EINVAL);
2711 	id = td->td_tid;
2712 	uq = td->td_umtxq;
2713 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2714 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2715 	    &uq->uq_key)) != 0)
2716 		return (error);
2717 	for (;;) {
2718 		umtxq_lock(&uq->uq_key);
2719 		umtxq_busy(&uq->uq_key);
2720 		umtxq_unlock(&uq->uq_key);
2721 
2722 		rv = fueword32(&m->m_ceilings[0], &save_ceiling);
2723 		if (rv == -1) {
2724 			error = EFAULT;
2725 			break;
2726 		}
2727 
2728 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2729 		    id | UMUTEX_CONTESTED);
2730 		if (rv == -1) {
2731 			error = EFAULT;
2732 			break;
2733 		}
2734 
2735 		if (rv == 0) {
2736 			MPASS(owner == UMUTEX_CONTESTED);
2737 			rv = suword32(&m->m_ceilings[0], ceiling);
2738 			rv1 = suword32(&m->m_owner, UMUTEX_CONTESTED);
2739 			error = (rv == 0 && rv1 == 0) ? 0: EFAULT;
2740 			break;
2741 		}
2742 
2743 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2744 			rv = suword32(&m->m_ceilings[0], ceiling);
2745 			error = rv == 0 ? 0 : EFAULT;
2746 			break;
2747 		}
2748 
2749 		if (owner == UMUTEX_RB_OWNERDEAD) {
2750 			error = EOWNERDEAD;
2751 			break;
2752 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2753 			error = ENOTRECOVERABLE;
2754 			break;
2755 		}
2756 
2757 		/*
2758 		 * If we caught a signal, we have retried and now
2759 		 * exit immediately.
2760 		 */
2761 		if (error != 0)
2762 			break;
2763 
2764 		/*
2765 		 * We set the contested bit, sleep. Otherwise the lock changed
2766 		 * and we need to retry or we lost a race to the thread
2767 		 * unlocking the umtx.
2768 		 */
2769 		umtxq_lock(&uq->uq_key);
2770 		umtxq_insert(uq);
2771 		umtxq_unbusy(&uq->uq_key);
2772 		error = umtxq_sleep(uq, "umtxpp", NULL);
2773 		umtxq_remove(uq);
2774 		umtxq_unlock(&uq->uq_key);
2775 	}
2776 	umtxq_lock(&uq->uq_key);
2777 	if (error == 0)
2778 		umtxq_signal(&uq->uq_key, INT_MAX);
2779 	umtxq_unbusy(&uq->uq_key);
2780 	umtxq_unlock(&uq->uq_key);
2781 	umtx_key_release(&uq->uq_key);
2782 	if (error == 0 && old_ceiling != NULL) {
2783 		rv = suword32(old_ceiling, save_ceiling);
2784 		error = rv == 0 ? 0 : EFAULT;
2785 	}
2786 	return (error);
2787 }
2788 
2789 /*
2790  * Lock a userland POSIX mutex.
2791  */
2792 static int
2793 do_lock_umutex(struct thread *td, struct umutex *m,
2794     struct _umtx_time *timeout, int mode)
2795 {
2796 	uint32_t flags;
2797 	int error;
2798 
2799 	error = fueword32(&m->m_flags, &flags);
2800 	if (error == -1)
2801 		return (EFAULT);
2802 
2803 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2804 	case 0:
2805 		error = do_lock_normal(td, m, flags, timeout, mode);
2806 		break;
2807 	case UMUTEX_PRIO_INHERIT:
2808 		error = do_lock_pi(td, m, flags, timeout, mode);
2809 		break;
2810 	case UMUTEX_PRIO_PROTECT:
2811 		error = do_lock_pp(td, m, flags, timeout, mode);
2812 		break;
2813 	default:
2814 		return (EINVAL);
2815 	}
2816 	if (timeout == NULL) {
2817 		if (error == EINTR && mode != _UMUTEX_WAIT)
2818 			error = ERESTART;
2819 	} else {
2820 		/* Timed-locking is not restarted. */
2821 		if (error == ERESTART)
2822 			error = EINTR;
2823 	}
2824 	return (error);
2825 }
2826 
2827 /*
2828  * Unlock a userland POSIX mutex.
2829  */
2830 static int
2831 do_unlock_umutex(struct thread *td, struct umutex *m, bool rb)
2832 {
2833 	uint32_t flags;
2834 	int error;
2835 
2836 	error = fueword32(&m->m_flags, &flags);
2837 	if (error == -1)
2838 		return (EFAULT);
2839 
2840 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2841 	case 0:
2842 		return (do_unlock_normal(td, m, flags, rb));
2843 	case UMUTEX_PRIO_INHERIT:
2844 		return (do_unlock_pi(td, m, flags, rb));
2845 	case UMUTEX_PRIO_PROTECT:
2846 		return (do_unlock_pp(td, m, flags, rb));
2847 	}
2848 
2849 	return (EINVAL);
2850 }
2851 
2852 static int
2853 do_cv_wait(struct thread *td, struct ucond *cv, struct umutex *m,
2854     struct timespec *timeout, u_long wflags)
2855 {
2856 	struct umtx_abs_timeout timo;
2857 	struct umtx_q *uq;
2858 	uint32_t flags, clockid, hasw;
2859 	int error;
2860 
2861 	uq = td->td_umtxq;
2862 	error = fueword32(&cv->c_flags, &flags);
2863 	if (error == -1)
2864 		return (EFAULT);
2865 	error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &uq->uq_key);
2866 	if (error != 0)
2867 		return (error);
2868 
2869 	if ((wflags & CVWAIT_CLOCKID) != 0) {
2870 		error = fueword32(&cv->c_clockid, &clockid);
2871 		if (error == -1) {
2872 			umtx_key_release(&uq->uq_key);
2873 			return (EFAULT);
2874 		}
2875 		if (clockid < CLOCK_REALTIME ||
2876 		    clockid >= CLOCK_THREAD_CPUTIME_ID) {
2877 			/* hmm, only HW clock id will work. */
2878 			umtx_key_release(&uq->uq_key);
2879 			return (EINVAL);
2880 		}
2881 	} else {
2882 		clockid = CLOCK_REALTIME;
2883 	}
2884 
2885 	umtxq_lock(&uq->uq_key);
2886 	umtxq_busy(&uq->uq_key);
2887 	umtxq_insert(uq);
2888 	umtxq_unlock(&uq->uq_key);
2889 
2890 	/*
2891 	 * Set c_has_waiters to 1 before releasing user mutex, also
2892 	 * don't modify cache line when unnecessary.
2893 	 */
2894 	error = fueword32(&cv->c_has_waiters, &hasw);
2895 	if (error == 0 && hasw == 0)
2896 		suword32(&cv->c_has_waiters, 1);
2897 
2898 	umtxq_unbusy_unlocked(&uq->uq_key);
2899 
2900 	error = do_unlock_umutex(td, m, false);
2901 
2902 	if (timeout != NULL)
2903 		umtx_abs_timeout_init(&timo, clockid,
2904 		    (wflags & CVWAIT_ABSTIME) != 0, timeout);
2905 
2906 	umtxq_lock(&uq->uq_key);
2907 	if (error == 0) {
2908 		error = umtxq_sleep(uq, "ucond", timeout == NULL ?
2909 		    NULL : &timo);
2910 	}
2911 
2912 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
2913 		error = 0;
2914 	else {
2915 		/*
2916 		 * This must be timeout,interrupted by signal or
2917 		 * surprious wakeup, clear c_has_waiter flag when
2918 		 * necessary.
2919 		 */
2920 		umtxq_busy(&uq->uq_key);
2921 		if ((uq->uq_flags & UQF_UMTXQ) != 0) {
2922 			int oldlen = uq->uq_cur_queue->length;
2923 			umtxq_remove(uq);
2924 			if (oldlen == 1) {
2925 				umtxq_unlock(&uq->uq_key);
2926 				suword32(&cv->c_has_waiters, 0);
2927 				umtxq_lock(&uq->uq_key);
2928 			}
2929 		}
2930 		umtxq_unbusy(&uq->uq_key);
2931 		if (error == ERESTART)
2932 			error = EINTR;
2933 	}
2934 
2935 	umtxq_unlock(&uq->uq_key);
2936 	umtx_key_release(&uq->uq_key);
2937 	return (error);
2938 }
2939 
2940 /*
2941  * Signal a userland condition variable.
2942  */
2943 static int
2944 do_cv_signal(struct thread *td, struct ucond *cv)
2945 {
2946 	struct umtx_key key;
2947 	int error, cnt, nwake;
2948 	uint32_t flags;
2949 
2950 	error = fueword32(&cv->c_flags, &flags);
2951 	if (error == -1)
2952 		return (EFAULT);
2953 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
2954 		return (error);
2955 	umtxq_lock(&key);
2956 	umtxq_busy(&key);
2957 	cnt = umtxq_count(&key);
2958 	nwake = umtxq_signal(&key, 1);
2959 	if (cnt <= nwake) {
2960 		umtxq_unlock(&key);
2961 		error = suword32(&cv->c_has_waiters, 0);
2962 		if (error == -1)
2963 			error = EFAULT;
2964 		umtxq_lock(&key);
2965 	}
2966 	umtxq_unbusy(&key);
2967 	umtxq_unlock(&key);
2968 	umtx_key_release(&key);
2969 	return (error);
2970 }
2971 
2972 static int
2973 do_cv_broadcast(struct thread *td, struct ucond *cv)
2974 {
2975 	struct umtx_key key;
2976 	int error;
2977 	uint32_t flags;
2978 
2979 	error = fueword32(&cv->c_flags, &flags);
2980 	if (error == -1)
2981 		return (EFAULT);
2982 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
2983 		return (error);
2984 
2985 	umtxq_lock(&key);
2986 	umtxq_busy(&key);
2987 	umtxq_signal(&key, INT_MAX);
2988 	umtxq_unlock(&key);
2989 
2990 	error = suword32(&cv->c_has_waiters, 0);
2991 	if (error == -1)
2992 		error = EFAULT;
2993 
2994 	umtxq_unbusy_unlocked(&key);
2995 
2996 	umtx_key_release(&key);
2997 	return (error);
2998 }
2999 
3000 static int
3001 do_rw_rdlock(struct thread *td, struct urwlock *rwlock, long fflag,
3002     struct _umtx_time *timeout)
3003 {
3004 	struct umtx_abs_timeout timo;
3005 	struct umtx_q *uq;
3006 	uint32_t flags, wrflags;
3007 	int32_t state, oldstate;
3008 	int32_t blocked_readers;
3009 	int error, error1, rv;
3010 
3011 	uq = td->td_umtxq;
3012 	error = fueword32(&rwlock->rw_flags, &flags);
3013 	if (error == -1)
3014 		return (EFAULT);
3015 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3016 	if (error != 0)
3017 		return (error);
3018 
3019 	if (timeout != NULL)
3020 		umtx_abs_timeout_init2(&timo, timeout);
3021 
3022 	wrflags = URWLOCK_WRITE_OWNER;
3023 	if (!(fflag & URWLOCK_PREFER_READER) && !(flags & URWLOCK_PREFER_READER))
3024 		wrflags |= URWLOCK_WRITE_WAITERS;
3025 
3026 	for (;;) {
3027 		rv = fueword32(&rwlock->rw_state, &state);
3028 		if (rv == -1) {
3029 			umtx_key_release(&uq->uq_key);
3030 			return (EFAULT);
3031 		}
3032 
3033 		/* try to lock it */
3034 		while (!(state & wrflags)) {
3035 			if (__predict_false(URWLOCK_READER_COUNT(state) ==
3036 			    URWLOCK_MAX_READERS)) {
3037 				umtx_key_release(&uq->uq_key);
3038 				return (EAGAIN);
3039 			}
3040 			rv = casueword32(&rwlock->rw_state, state,
3041 			    &oldstate, state + 1);
3042 			if (rv == -1) {
3043 				umtx_key_release(&uq->uq_key);
3044 				return (EFAULT);
3045 			}
3046 			if (rv == 0) {
3047 				MPASS(oldstate == state);
3048 				umtx_key_release(&uq->uq_key);
3049 				return (0);
3050 			}
3051 			error = thread_check_susp(td, true);
3052 			if (error != 0)
3053 				break;
3054 			state = oldstate;
3055 		}
3056 
3057 		if (error)
3058 			break;
3059 
3060 		/* grab monitor lock */
3061 		umtxq_lock(&uq->uq_key);
3062 		umtxq_busy(&uq->uq_key);
3063 		umtxq_unlock(&uq->uq_key);
3064 
3065 		/*
3066 		 * re-read the state, in case it changed between the try-lock above
3067 		 * and the check below
3068 		 */
3069 		rv = fueword32(&rwlock->rw_state, &state);
3070 		if (rv == -1)
3071 			error = EFAULT;
3072 
3073 		/* set read contention bit */
3074 		while (error == 0 && (state & wrflags) &&
3075 		    !(state & URWLOCK_READ_WAITERS)) {
3076 			rv = casueword32(&rwlock->rw_state, state,
3077 			    &oldstate, state | URWLOCK_READ_WAITERS);
3078 			if (rv == -1) {
3079 				error = EFAULT;
3080 				break;
3081 			}
3082 			if (rv == 0) {
3083 				MPASS(oldstate == state);
3084 				goto sleep;
3085 			}
3086 			state = oldstate;
3087 			error = thread_check_susp(td, false);
3088 			if (error != 0)
3089 				break;
3090 		}
3091 		if (error != 0) {
3092 			umtxq_unbusy_unlocked(&uq->uq_key);
3093 			break;
3094 		}
3095 
3096 		/* state is changed while setting flags, restart */
3097 		if (!(state & wrflags)) {
3098 			umtxq_unbusy_unlocked(&uq->uq_key);
3099 			error = thread_check_susp(td, true);
3100 			if (error != 0)
3101 				break;
3102 			continue;
3103 		}
3104 
3105 sleep:
3106 		/*
3107 		 * Contention bit is set, before sleeping, increase
3108 		 * read waiter count.
3109 		 */
3110 		rv = fueword32(&rwlock->rw_blocked_readers,
3111 		    &blocked_readers);
3112 		if (rv == -1) {
3113 			umtxq_unbusy_unlocked(&uq->uq_key);
3114 			error = EFAULT;
3115 			break;
3116 		}
3117 		suword32(&rwlock->rw_blocked_readers, blocked_readers+1);
3118 
3119 		while (state & wrflags) {
3120 			umtxq_lock(&uq->uq_key);
3121 			umtxq_insert(uq);
3122 			umtxq_unbusy(&uq->uq_key);
3123 
3124 			error = umtxq_sleep(uq, "urdlck", timeout == NULL ?
3125 			    NULL : &timo);
3126 
3127 			umtxq_busy(&uq->uq_key);
3128 			umtxq_remove(uq);
3129 			umtxq_unlock(&uq->uq_key);
3130 			if (error)
3131 				break;
3132 			rv = fueword32(&rwlock->rw_state, &state);
3133 			if (rv == -1) {
3134 				error = EFAULT;
3135 				break;
3136 			}
3137 		}
3138 
3139 		/* decrease read waiter count, and may clear read contention bit */
3140 		rv = fueword32(&rwlock->rw_blocked_readers,
3141 		    &blocked_readers);
3142 		if (rv == -1) {
3143 			umtxq_unbusy_unlocked(&uq->uq_key);
3144 			error = EFAULT;
3145 			break;
3146 		}
3147 		suword32(&rwlock->rw_blocked_readers, blocked_readers-1);
3148 		if (blocked_readers == 1) {
3149 			rv = fueword32(&rwlock->rw_state, &state);
3150 			if (rv == -1) {
3151 				umtxq_unbusy_unlocked(&uq->uq_key);
3152 				error = EFAULT;
3153 				break;
3154 			}
3155 			for (;;) {
3156 				rv = casueword32(&rwlock->rw_state, state,
3157 				    &oldstate, state & ~URWLOCK_READ_WAITERS);
3158 				if (rv == -1) {
3159 					error = EFAULT;
3160 					break;
3161 				}
3162 				if (rv == 0) {
3163 					MPASS(oldstate == state);
3164 					break;
3165 				}
3166 				state = oldstate;
3167 				error1 = thread_check_susp(td, false);
3168 				if (error1 != 0) {
3169 					if (error == 0)
3170 						error = error1;
3171 					break;
3172 				}
3173 			}
3174 		}
3175 
3176 		umtxq_unbusy_unlocked(&uq->uq_key);
3177 		if (error != 0)
3178 			break;
3179 	}
3180 	umtx_key_release(&uq->uq_key);
3181 	if (error == ERESTART)
3182 		error = EINTR;
3183 	return (error);
3184 }
3185 
3186 static int
3187 do_rw_wrlock(struct thread *td, struct urwlock *rwlock, struct _umtx_time *timeout)
3188 {
3189 	struct umtx_abs_timeout timo;
3190 	struct umtx_q *uq;
3191 	uint32_t flags;
3192 	int32_t state, oldstate;
3193 	int32_t blocked_writers;
3194 	int32_t blocked_readers;
3195 	int error, error1, rv;
3196 
3197 	uq = td->td_umtxq;
3198 	error = fueword32(&rwlock->rw_flags, &flags);
3199 	if (error == -1)
3200 		return (EFAULT);
3201 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3202 	if (error != 0)
3203 		return (error);
3204 
3205 	if (timeout != NULL)
3206 		umtx_abs_timeout_init2(&timo, timeout);
3207 
3208 	blocked_readers = 0;
3209 	for (;;) {
3210 		rv = fueword32(&rwlock->rw_state, &state);
3211 		if (rv == -1) {
3212 			umtx_key_release(&uq->uq_key);
3213 			return (EFAULT);
3214 		}
3215 		while ((state & URWLOCK_WRITE_OWNER) == 0 &&
3216 		    URWLOCK_READER_COUNT(state) == 0) {
3217 			rv = casueword32(&rwlock->rw_state, state,
3218 			    &oldstate, state | URWLOCK_WRITE_OWNER);
3219 			if (rv == -1) {
3220 				umtx_key_release(&uq->uq_key);
3221 				return (EFAULT);
3222 			}
3223 			if (rv == 0) {
3224 				MPASS(oldstate == state);
3225 				umtx_key_release(&uq->uq_key);
3226 				return (0);
3227 			}
3228 			state = oldstate;
3229 			error = thread_check_susp(td, true);
3230 			if (error != 0)
3231 				break;
3232 		}
3233 
3234 		if (error) {
3235 			if ((state & (URWLOCK_WRITE_OWNER |
3236 			    URWLOCK_WRITE_WAITERS)) == 0 &&
3237 			    blocked_readers != 0) {
3238 				umtxq_lock(&uq->uq_key);
3239 				umtxq_busy(&uq->uq_key);
3240 				umtxq_signal_queue(&uq->uq_key, INT_MAX,
3241 				    UMTX_SHARED_QUEUE);
3242 				umtxq_unbusy(&uq->uq_key);
3243 				umtxq_unlock(&uq->uq_key);
3244 			}
3245 
3246 			break;
3247 		}
3248 
3249 		/* grab monitor lock */
3250 		umtxq_lock(&uq->uq_key);
3251 		umtxq_busy(&uq->uq_key);
3252 		umtxq_unlock(&uq->uq_key);
3253 
3254 		/*
3255 		 * Re-read the state, in case it changed between the
3256 		 * try-lock above and the check below.
3257 		 */
3258 		rv = fueword32(&rwlock->rw_state, &state);
3259 		if (rv == -1)
3260 			error = EFAULT;
3261 
3262 		while (error == 0 && ((state & URWLOCK_WRITE_OWNER) ||
3263 		    URWLOCK_READER_COUNT(state) != 0) &&
3264 		    (state & URWLOCK_WRITE_WAITERS) == 0) {
3265 			rv = casueword32(&rwlock->rw_state, state,
3266 			    &oldstate, state | URWLOCK_WRITE_WAITERS);
3267 			if (rv == -1) {
3268 				error = EFAULT;
3269 				break;
3270 			}
3271 			if (rv == 0) {
3272 				MPASS(oldstate == state);
3273 				goto sleep;
3274 			}
3275 			state = oldstate;
3276 			error = thread_check_susp(td, false);
3277 			if (error != 0)
3278 				break;
3279 		}
3280 		if (error != 0) {
3281 			umtxq_unbusy_unlocked(&uq->uq_key);
3282 			break;
3283 		}
3284 
3285 		if ((state & URWLOCK_WRITE_OWNER) == 0 &&
3286 		    URWLOCK_READER_COUNT(state) == 0) {
3287 			umtxq_unbusy_unlocked(&uq->uq_key);
3288 			error = thread_check_susp(td, false);
3289 			if (error != 0)
3290 				break;
3291 			continue;
3292 		}
3293 sleep:
3294 		rv = fueword32(&rwlock->rw_blocked_writers,
3295 		    &blocked_writers);
3296 		if (rv == -1) {
3297 			umtxq_unbusy_unlocked(&uq->uq_key);
3298 			error = EFAULT;
3299 			break;
3300 		}
3301 		suword32(&rwlock->rw_blocked_writers, blocked_writers + 1);
3302 
3303 		while ((state & URWLOCK_WRITE_OWNER) ||
3304 		    URWLOCK_READER_COUNT(state) != 0) {
3305 			umtxq_lock(&uq->uq_key);
3306 			umtxq_insert_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3307 			umtxq_unbusy(&uq->uq_key);
3308 
3309 			error = umtxq_sleep(uq, "uwrlck", timeout == NULL ?
3310 			    NULL : &timo);
3311 
3312 			umtxq_busy(&uq->uq_key);
3313 			umtxq_remove_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3314 			umtxq_unlock(&uq->uq_key);
3315 			if (error)
3316 				break;
3317 			rv = fueword32(&rwlock->rw_state, &state);
3318 			if (rv == -1) {
3319 				error = EFAULT;
3320 				break;
3321 			}
3322 		}
3323 
3324 		rv = fueword32(&rwlock->rw_blocked_writers,
3325 		    &blocked_writers);
3326 		if (rv == -1) {
3327 			umtxq_unbusy_unlocked(&uq->uq_key);
3328 			error = EFAULT;
3329 			break;
3330 		}
3331 		suword32(&rwlock->rw_blocked_writers, blocked_writers-1);
3332 		if (blocked_writers == 1) {
3333 			rv = fueword32(&rwlock->rw_state, &state);
3334 			if (rv == -1) {
3335 				umtxq_unbusy_unlocked(&uq->uq_key);
3336 				error = EFAULT;
3337 				break;
3338 			}
3339 			for (;;) {
3340 				rv = casueword32(&rwlock->rw_state, state,
3341 				    &oldstate, state & ~URWLOCK_WRITE_WAITERS);
3342 				if (rv == -1) {
3343 					error = EFAULT;
3344 					break;
3345 				}
3346 				if (rv == 0) {
3347 					MPASS(oldstate == state);
3348 					break;
3349 				}
3350 				state = oldstate;
3351 				error1 = thread_check_susp(td, false);
3352 				/*
3353 				 * We are leaving the URWLOCK_WRITE_WAITERS
3354 				 * behind, but this should not harm the
3355 				 * correctness.
3356 				 */
3357 				if (error1 != 0) {
3358 					if (error == 0)
3359 						error = error1;
3360 					break;
3361 				}
3362 			}
3363 			rv = fueword32(&rwlock->rw_blocked_readers,
3364 			    &blocked_readers);
3365 			if (rv == -1) {
3366 				umtxq_unbusy_unlocked(&uq->uq_key);
3367 				error = EFAULT;
3368 				break;
3369 			}
3370 		} else
3371 			blocked_readers = 0;
3372 
3373 		umtxq_unbusy_unlocked(&uq->uq_key);
3374 	}
3375 
3376 	umtx_key_release(&uq->uq_key);
3377 	if (error == ERESTART)
3378 		error = EINTR;
3379 	return (error);
3380 }
3381 
3382 static int
3383 do_rw_unlock(struct thread *td, struct urwlock *rwlock)
3384 {
3385 	struct umtx_q *uq;
3386 	uint32_t flags;
3387 	int32_t state, oldstate;
3388 	int error, rv, q, count;
3389 
3390 	uq = td->td_umtxq;
3391 	error = fueword32(&rwlock->rw_flags, &flags);
3392 	if (error == -1)
3393 		return (EFAULT);
3394 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3395 	if (error != 0)
3396 		return (error);
3397 
3398 	error = fueword32(&rwlock->rw_state, &state);
3399 	if (error == -1) {
3400 		error = EFAULT;
3401 		goto out;
3402 	}
3403 	if (state & URWLOCK_WRITE_OWNER) {
3404 		for (;;) {
3405 			rv = casueword32(&rwlock->rw_state, state,
3406 			    &oldstate, state & ~URWLOCK_WRITE_OWNER);
3407 			if (rv == -1) {
3408 				error = EFAULT;
3409 				goto out;
3410 			}
3411 			if (rv == 1) {
3412 				state = oldstate;
3413 				if (!(oldstate & URWLOCK_WRITE_OWNER)) {
3414 					error = EPERM;
3415 					goto out;
3416 				}
3417 				error = thread_check_susp(td, true);
3418 				if (error != 0)
3419 					goto out;
3420 			} else
3421 				break;
3422 		}
3423 	} else if (URWLOCK_READER_COUNT(state) != 0) {
3424 		for (;;) {
3425 			rv = casueword32(&rwlock->rw_state, state,
3426 			    &oldstate, state - 1);
3427 			if (rv == -1) {
3428 				error = EFAULT;
3429 				goto out;
3430 			}
3431 			if (rv == 1) {
3432 				state = oldstate;
3433 				if (URWLOCK_READER_COUNT(oldstate) == 0) {
3434 					error = EPERM;
3435 					goto out;
3436 				}
3437 				error = thread_check_susp(td, true);
3438 				if (error != 0)
3439 					goto out;
3440 			} else
3441 				break;
3442 		}
3443 	} else {
3444 		error = EPERM;
3445 		goto out;
3446 	}
3447 
3448 	count = 0;
3449 
3450 	if (!(flags & URWLOCK_PREFER_READER)) {
3451 		if (state & URWLOCK_WRITE_WAITERS) {
3452 			count = 1;
3453 			q = UMTX_EXCLUSIVE_QUEUE;
3454 		} else if (state & URWLOCK_READ_WAITERS) {
3455 			count = INT_MAX;
3456 			q = UMTX_SHARED_QUEUE;
3457 		}
3458 	} else {
3459 		if (state & URWLOCK_READ_WAITERS) {
3460 			count = INT_MAX;
3461 			q = UMTX_SHARED_QUEUE;
3462 		} else if (state & URWLOCK_WRITE_WAITERS) {
3463 			count = 1;
3464 			q = UMTX_EXCLUSIVE_QUEUE;
3465 		}
3466 	}
3467 
3468 	if (count) {
3469 		umtxq_lock(&uq->uq_key);
3470 		umtxq_busy(&uq->uq_key);
3471 		umtxq_signal_queue(&uq->uq_key, count, q);
3472 		umtxq_unbusy(&uq->uq_key);
3473 		umtxq_unlock(&uq->uq_key);
3474 	}
3475 out:
3476 	umtx_key_release(&uq->uq_key);
3477 	return (error);
3478 }
3479 
3480 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3481 static int
3482 do_sem_wait(struct thread *td, struct _usem *sem, struct _umtx_time *timeout)
3483 {
3484 	struct umtx_abs_timeout timo;
3485 	struct umtx_q *uq;
3486 	uint32_t flags, count, count1;
3487 	int error, rv, rv1;
3488 
3489 	uq = td->td_umtxq;
3490 	error = fueword32(&sem->_flags, &flags);
3491 	if (error == -1)
3492 		return (EFAULT);
3493 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3494 	if (error != 0)
3495 		return (error);
3496 
3497 	if (timeout != NULL)
3498 		umtx_abs_timeout_init2(&timo, timeout);
3499 
3500 again:
3501 	umtxq_lock(&uq->uq_key);
3502 	umtxq_busy(&uq->uq_key);
3503 	umtxq_insert(uq);
3504 	umtxq_unlock(&uq->uq_key);
3505 	rv = casueword32(&sem->_has_waiters, 0, &count1, 1);
3506 	if (rv == 0)
3507 		rv1 = fueword32(&sem->_count, &count);
3508 	if (rv == -1 || (rv == 0 && (rv1 == -1 || count != 0)) ||
3509 	    (rv == 1 && count1 == 0)) {
3510 		umtxq_lock(&uq->uq_key);
3511 		umtxq_unbusy(&uq->uq_key);
3512 		umtxq_remove(uq);
3513 		umtxq_unlock(&uq->uq_key);
3514 		if (rv == 1) {
3515 			rv = thread_check_susp(td, true);
3516 			if (rv == 0)
3517 				goto again;
3518 			error = rv;
3519 			goto out;
3520 		}
3521 		if (rv == 0)
3522 			rv = rv1;
3523 		error = rv == -1 ? EFAULT : 0;
3524 		goto out;
3525 	}
3526 	umtxq_lock(&uq->uq_key);
3527 	umtxq_unbusy(&uq->uq_key);
3528 
3529 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3530 
3531 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3532 		error = 0;
3533 	else {
3534 		umtxq_remove(uq);
3535 		/* A relative timeout cannot be restarted. */
3536 		if (error == ERESTART && timeout != NULL &&
3537 		    (timeout->_flags & UMTX_ABSTIME) == 0)
3538 			error = EINTR;
3539 	}
3540 	umtxq_unlock(&uq->uq_key);
3541 out:
3542 	umtx_key_release(&uq->uq_key);
3543 	return (error);
3544 }
3545 
3546 /*
3547  * Signal a userland semaphore.
3548  */
3549 static int
3550 do_sem_wake(struct thread *td, struct _usem *sem)
3551 {
3552 	struct umtx_key key;
3553 	int error, cnt;
3554 	uint32_t flags;
3555 
3556 	error = fueword32(&sem->_flags, &flags);
3557 	if (error == -1)
3558 		return (EFAULT);
3559 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3560 		return (error);
3561 	umtxq_lock(&key);
3562 	umtxq_busy(&key);
3563 	cnt = umtxq_count(&key);
3564 	if (cnt > 0) {
3565 		/*
3566 		 * Check if count is greater than 0, this means the memory is
3567 		 * still being referenced by user code, so we can safely
3568 		 * update _has_waiters flag.
3569 		 */
3570 		if (cnt == 1) {
3571 			umtxq_unlock(&key);
3572 			error = suword32(&sem->_has_waiters, 0);
3573 			umtxq_lock(&key);
3574 			if (error == -1)
3575 				error = EFAULT;
3576 		}
3577 		umtxq_signal(&key, 1);
3578 	}
3579 	umtxq_unbusy(&key);
3580 	umtxq_unlock(&key);
3581 	umtx_key_release(&key);
3582 	return (error);
3583 }
3584 #endif
3585 
3586 static int
3587 do_sem2_wait(struct thread *td, struct _usem2 *sem, struct _umtx_time *timeout)
3588 {
3589 	struct umtx_abs_timeout timo;
3590 	struct umtx_q *uq;
3591 	uint32_t count, flags;
3592 	int error, rv;
3593 
3594 	uq = td->td_umtxq;
3595 	flags = fuword32(&sem->_flags);
3596 	if (timeout != NULL)
3597 		umtx_abs_timeout_init2(&timo, timeout);
3598 
3599 again:
3600 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3601 	if (error != 0)
3602 		return (error);
3603 	umtxq_lock(&uq->uq_key);
3604 	umtxq_busy(&uq->uq_key);
3605 	umtxq_insert(uq);
3606 	umtxq_unlock(&uq->uq_key);
3607 	rv = fueword32(&sem->_count, &count);
3608 	if (rv == -1) {
3609 		umtxq_lock(&uq->uq_key);
3610 		umtxq_unbusy(&uq->uq_key);
3611 		umtxq_remove(uq);
3612 		umtxq_unlock(&uq->uq_key);
3613 		umtx_key_release(&uq->uq_key);
3614 		return (EFAULT);
3615 	}
3616 	for (;;) {
3617 		if (USEM_COUNT(count) != 0) {
3618 			umtxq_lock(&uq->uq_key);
3619 			umtxq_unbusy(&uq->uq_key);
3620 			umtxq_remove(uq);
3621 			umtxq_unlock(&uq->uq_key);
3622 			umtx_key_release(&uq->uq_key);
3623 			return (0);
3624 		}
3625 		if (count == USEM_HAS_WAITERS)
3626 			break;
3627 		rv = casueword32(&sem->_count, 0, &count, USEM_HAS_WAITERS);
3628 		if (rv == 0)
3629 			break;
3630 		umtxq_lock(&uq->uq_key);
3631 		umtxq_unbusy(&uq->uq_key);
3632 		umtxq_remove(uq);
3633 		umtxq_unlock(&uq->uq_key);
3634 		umtx_key_release(&uq->uq_key);
3635 		if (rv == -1)
3636 			return (EFAULT);
3637 		rv = thread_check_susp(td, true);
3638 		if (rv != 0)
3639 			return (rv);
3640 		goto again;
3641 	}
3642 	umtxq_lock(&uq->uq_key);
3643 	umtxq_unbusy(&uq->uq_key);
3644 
3645 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3646 
3647 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3648 		error = 0;
3649 	else {
3650 		umtxq_remove(uq);
3651 		if (timeout != NULL && (timeout->_flags & UMTX_ABSTIME) == 0) {
3652 			/* A relative timeout cannot be restarted. */
3653 			if (error == ERESTART)
3654 				error = EINTR;
3655 			if (error == EINTR) {
3656 				umtx_abs_timeout_update(&timo);
3657 				timespecsub(&timo.end, &timo.cur,
3658 				    &timeout->_timeout);
3659 			}
3660 		}
3661 	}
3662 	umtxq_unlock(&uq->uq_key);
3663 	umtx_key_release(&uq->uq_key);
3664 	return (error);
3665 }
3666 
3667 /*
3668  * Signal a userland semaphore.
3669  */
3670 static int
3671 do_sem2_wake(struct thread *td, struct _usem2 *sem)
3672 {
3673 	struct umtx_key key;
3674 	int error, cnt, rv;
3675 	uint32_t count, flags;
3676 
3677 	rv = fueword32(&sem->_flags, &flags);
3678 	if (rv == -1)
3679 		return (EFAULT);
3680 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3681 		return (error);
3682 	umtxq_lock(&key);
3683 	umtxq_busy(&key);
3684 	cnt = umtxq_count(&key);
3685 	if (cnt > 0) {
3686 		/*
3687 		 * If this was the last sleeping thread, clear the waiters
3688 		 * flag in _count.
3689 		 */
3690 		if (cnt == 1) {
3691 			umtxq_unlock(&key);
3692 			rv = fueword32(&sem->_count, &count);
3693 			while (rv != -1 && count & USEM_HAS_WAITERS) {
3694 				rv = casueword32(&sem->_count, count, &count,
3695 				    count & ~USEM_HAS_WAITERS);
3696 				if (rv == 1) {
3697 					rv = thread_check_susp(td, true);
3698 					if (rv != 0)
3699 						break;
3700 				}
3701 			}
3702 			if (rv == -1)
3703 				error = EFAULT;
3704 			else if (rv > 0) {
3705 				error = rv;
3706 			}
3707 			umtxq_lock(&key);
3708 		}
3709 
3710 		umtxq_signal(&key, 1);
3711 	}
3712 	umtxq_unbusy(&key);
3713 	umtxq_unlock(&key);
3714 	umtx_key_release(&key);
3715 	return (error);
3716 }
3717 
3718 #ifdef COMPAT_FREEBSD10
3719 int
3720 freebsd10__umtx_lock(struct thread *td, struct freebsd10__umtx_lock_args *uap)
3721 {
3722 	return (do_lock_umtx(td, uap->umtx, td->td_tid, 0));
3723 }
3724 
3725 int
3726 freebsd10__umtx_unlock(struct thread *td,
3727     struct freebsd10__umtx_unlock_args *uap)
3728 {
3729 	return (do_unlock_umtx(td, uap->umtx, td->td_tid));
3730 }
3731 #endif
3732 
3733 inline int
3734 umtx_copyin_timeout(const void *uaddr, struct timespec *tsp)
3735 {
3736 	int error;
3737 
3738 	error = copyin(uaddr, tsp, sizeof(*tsp));
3739 	if (error == 0) {
3740 		if (tsp->tv_sec < 0 ||
3741 		    tsp->tv_nsec >= 1000000000 ||
3742 		    tsp->tv_nsec < 0)
3743 			error = EINVAL;
3744 	}
3745 	return (error);
3746 }
3747 
3748 static inline int
3749 umtx_copyin_umtx_time(const void *uaddr, size_t size, struct _umtx_time *tp)
3750 {
3751 	int error;
3752 
3753 	if (size <= sizeof(tp->_timeout)) {
3754 		tp->_clockid = CLOCK_REALTIME;
3755 		tp->_flags = 0;
3756 		error = copyin(uaddr, &tp->_timeout, sizeof(tp->_timeout));
3757 	} else
3758 		error = copyin(uaddr, tp, sizeof(*tp));
3759 	if (error != 0)
3760 		return (error);
3761 	if (tp->_timeout.tv_sec < 0 ||
3762 	    tp->_timeout.tv_nsec >= 1000000000 || tp->_timeout.tv_nsec < 0)
3763 		return (EINVAL);
3764 	return (0);
3765 }
3766 
3767 static int
3768 umtx_copyin_robust_lists(const void *uaddr, size_t size,
3769     struct umtx_robust_lists_params *rb)
3770 {
3771 
3772 	if (size > sizeof(*rb))
3773 		return (EINVAL);
3774 	return (copyin(uaddr, rb, size));
3775 }
3776 
3777 static int
3778 umtx_copyout_timeout(void *uaddr, size_t sz, struct timespec *tsp)
3779 {
3780 
3781 	/*
3782 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
3783 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
3784 	 * copyops.
3785 	 */
3786 	KASSERT(sz >= sizeof(*tsp),
3787 	    ("umtx_copyops specifies incorrect sizes"));
3788 
3789 	return (copyout(tsp, uaddr, sizeof(*tsp)));
3790 }
3791 
3792 #ifdef COMPAT_FREEBSD10
3793 static int
3794 __umtx_op_lock_umtx(struct thread *td, struct _umtx_op_args *uap,
3795     const struct umtx_copyops *ops)
3796 {
3797 	struct timespec *ts, timeout;
3798 	int error;
3799 
3800 	/* Allow a null timespec (wait forever). */
3801 	if (uap->uaddr2 == NULL)
3802 		ts = NULL;
3803 	else {
3804 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
3805 		if (error != 0)
3806 			return (error);
3807 		ts = &timeout;
3808 	}
3809 #ifdef COMPAT_FREEBSD32
3810 	if (ops->compat32)
3811 		return (do_lock_umtx32(td, uap->obj, uap->val, ts));
3812 #endif
3813 	return (do_lock_umtx(td, uap->obj, uap->val, ts));
3814 }
3815 
3816 static int
3817 __umtx_op_unlock_umtx(struct thread *td, struct _umtx_op_args *uap,
3818     const struct umtx_copyops *ops)
3819 {
3820 #ifdef COMPAT_FREEBSD32
3821 	if (ops->compat32)
3822 		return (do_unlock_umtx32(td, uap->obj, uap->val));
3823 #endif
3824 	return (do_unlock_umtx(td, uap->obj, uap->val));
3825 }
3826 #endif	/* COMPAT_FREEBSD10 */
3827 
3828 #if !defined(COMPAT_FREEBSD10)
3829 static int
3830 __umtx_op_unimpl(struct thread *td __unused, struct _umtx_op_args *uap __unused,
3831     const struct umtx_copyops *ops __unused)
3832 {
3833 	return (EOPNOTSUPP);
3834 }
3835 #endif	/* COMPAT_FREEBSD10 */
3836 
3837 static int
3838 __umtx_op_wait(struct thread *td, struct _umtx_op_args *uap,
3839     const struct umtx_copyops *ops)
3840 {
3841 	struct _umtx_time timeout, *tm_p;
3842 	int error;
3843 
3844 	if (uap->uaddr2 == NULL)
3845 		tm_p = NULL;
3846 	else {
3847 		error = ops->copyin_umtx_time(
3848 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3849 		if (error != 0)
3850 			return (error);
3851 		tm_p = &timeout;
3852 	}
3853 	return (do_wait(td, uap->obj, uap->val, tm_p, ops->compat32, 0));
3854 }
3855 
3856 static int
3857 __umtx_op_wait_uint(struct thread *td, struct _umtx_op_args *uap,
3858     const struct umtx_copyops *ops)
3859 {
3860 	struct _umtx_time timeout, *tm_p;
3861 	int error;
3862 
3863 	if (uap->uaddr2 == NULL)
3864 		tm_p = NULL;
3865 	else {
3866 		error = ops->copyin_umtx_time(
3867 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3868 		if (error != 0)
3869 			return (error);
3870 		tm_p = &timeout;
3871 	}
3872 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
3873 }
3874 
3875 static int
3876 __umtx_op_wait_uint_private(struct thread *td, struct _umtx_op_args *uap,
3877     const struct umtx_copyops *ops)
3878 {
3879 	struct _umtx_time *tm_p, timeout;
3880 	int error;
3881 
3882 	if (uap->uaddr2 == NULL)
3883 		tm_p = NULL;
3884 	else {
3885 		error = ops->copyin_umtx_time(
3886 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3887 		if (error != 0)
3888 			return (error);
3889 		tm_p = &timeout;
3890 	}
3891 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
3892 }
3893 
3894 static int
3895 __umtx_op_wake(struct thread *td, struct _umtx_op_args *uap,
3896     const struct umtx_copyops *ops __unused)
3897 {
3898 
3899 	return (kern_umtx_wake(td, uap->obj, uap->val, 0));
3900 }
3901 
3902 #define BATCH_SIZE	128
3903 static int
3904 __umtx_op_nwake_private_native(struct thread *td, struct _umtx_op_args *uap)
3905 {
3906 	char *uaddrs[BATCH_SIZE], **upp;
3907 	int count, error, i, pos, tocopy;
3908 
3909 	upp = (char **)uap->obj;
3910 	error = 0;
3911 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3912 	    pos += tocopy) {
3913 		tocopy = MIN(count, BATCH_SIZE);
3914 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(char *));
3915 		if (error != 0)
3916 			break;
3917 		for (i = 0; i < tocopy; ++i) {
3918 			kern_umtx_wake(td, uaddrs[i], INT_MAX, 1);
3919 		}
3920 		maybe_yield();
3921 	}
3922 	return (error);
3923 }
3924 
3925 static int
3926 __umtx_op_nwake_private_compat32(struct thread *td, struct _umtx_op_args *uap)
3927 {
3928 	uint32_t uaddrs[BATCH_SIZE], *upp;
3929 	int count, error, i, pos, tocopy;
3930 
3931 	upp = (uint32_t *)uap->obj;
3932 	error = 0;
3933 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3934 	    pos += tocopy) {
3935 		tocopy = MIN(count, BATCH_SIZE);
3936 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(uint32_t));
3937 		if (error != 0)
3938 			break;
3939 		for (i = 0; i < tocopy; ++i) {
3940 			kern_umtx_wake(td, (void *)(uintptr_t)uaddrs[i],
3941 			    INT_MAX, 1);
3942 		}
3943 		maybe_yield();
3944 	}
3945 	return (error);
3946 }
3947 
3948 static int
3949 __umtx_op_nwake_private(struct thread *td, struct _umtx_op_args *uap,
3950     const struct umtx_copyops *ops)
3951 {
3952 
3953 	if (ops->compat32)
3954 		return (__umtx_op_nwake_private_compat32(td, uap));
3955 	return (__umtx_op_nwake_private_native(td, uap));
3956 }
3957 
3958 static int
3959 __umtx_op_wake_private(struct thread *td, struct _umtx_op_args *uap,
3960     const struct umtx_copyops *ops __unused)
3961 {
3962 
3963 	return (kern_umtx_wake(td, uap->obj, uap->val, 1));
3964 }
3965 
3966 static int
3967 __umtx_op_lock_umutex(struct thread *td, struct _umtx_op_args *uap,
3968    const struct umtx_copyops *ops)
3969 {
3970 	struct _umtx_time *tm_p, timeout;
3971 	int error;
3972 
3973 	/* Allow a null timespec (wait forever). */
3974 	if (uap->uaddr2 == NULL)
3975 		tm_p = NULL;
3976 	else {
3977 		error = ops->copyin_umtx_time(
3978 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3979 		if (error != 0)
3980 			return (error);
3981 		tm_p = &timeout;
3982 	}
3983 	return (do_lock_umutex(td, uap->obj, tm_p, 0));
3984 }
3985 
3986 static int
3987 __umtx_op_trylock_umutex(struct thread *td, struct _umtx_op_args *uap,
3988     const struct umtx_copyops *ops __unused)
3989 {
3990 
3991 	return (do_lock_umutex(td, uap->obj, NULL, _UMUTEX_TRY));
3992 }
3993 
3994 static int
3995 __umtx_op_wait_umutex(struct thread *td, struct _umtx_op_args *uap,
3996     const struct umtx_copyops *ops)
3997 {
3998 	struct _umtx_time *tm_p, timeout;
3999 	int error;
4000 
4001 	/* Allow a null timespec (wait forever). */
4002 	if (uap->uaddr2 == NULL)
4003 		tm_p = NULL;
4004 	else {
4005 		error = ops->copyin_umtx_time(
4006 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4007 		if (error != 0)
4008 			return (error);
4009 		tm_p = &timeout;
4010 	}
4011 	return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
4012 }
4013 
4014 static int
4015 __umtx_op_wake_umutex(struct thread *td, struct _umtx_op_args *uap,
4016     const struct umtx_copyops *ops __unused)
4017 {
4018 
4019 	return (do_wake_umutex(td, uap->obj));
4020 }
4021 
4022 static int
4023 __umtx_op_unlock_umutex(struct thread *td, struct _umtx_op_args *uap,
4024     const struct umtx_copyops *ops __unused)
4025 {
4026 
4027 	return (do_unlock_umutex(td, uap->obj, false));
4028 }
4029 
4030 static int
4031 __umtx_op_set_ceiling(struct thread *td, struct _umtx_op_args *uap,
4032     const struct umtx_copyops *ops __unused)
4033 {
4034 
4035 	return (do_set_ceiling(td, uap->obj, uap->val, uap->uaddr1));
4036 }
4037 
4038 static int
4039 __umtx_op_cv_wait(struct thread *td, struct _umtx_op_args *uap,
4040     const struct umtx_copyops *ops)
4041 {
4042 	struct timespec *ts, timeout;
4043 	int error;
4044 
4045 	/* Allow a null timespec (wait forever). */
4046 	if (uap->uaddr2 == NULL)
4047 		ts = NULL;
4048 	else {
4049 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
4050 		if (error != 0)
4051 			return (error);
4052 		ts = &timeout;
4053 	}
4054 	return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
4055 }
4056 
4057 static int
4058 __umtx_op_cv_signal(struct thread *td, struct _umtx_op_args *uap,
4059     const struct umtx_copyops *ops __unused)
4060 {
4061 
4062 	return (do_cv_signal(td, uap->obj));
4063 }
4064 
4065 static int
4066 __umtx_op_cv_broadcast(struct thread *td, struct _umtx_op_args *uap,
4067     const struct umtx_copyops *ops __unused)
4068 {
4069 
4070 	return (do_cv_broadcast(td, uap->obj));
4071 }
4072 
4073 static int
4074 __umtx_op_rw_rdlock(struct thread *td, struct _umtx_op_args *uap,
4075     const struct umtx_copyops *ops)
4076 {
4077 	struct _umtx_time timeout;
4078 	int error;
4079 
4080 	/* Allow a null timespec (wait forever). */
4081 	if (uap->uaddr2 == NULL) {
4082 		error = do_rw_rdlock(td, uap->obj, uap->val, 0);
4083 	} else {
4084 		error = ops->copyin_umtx_time(uap->uaddr2,
4085 		   (size_t)uap->uaddr1, &timeout);
4086 		if (error != 0)
4087 			return (error);
4088 		error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
4089 	}
4090 	return (error);
4091 }
4092 
4093 static int
4094 __umtx_op_rw_wrlock(struct thread *td, struct _umtx_op_args *uap,
4095     const struct umtx_copyops *ops)
4096 {
4097 	struct _umtx_time timeout;
4098 	int error;
4099 
4100 	/* Allow a null timespec (wait forever). */
4101 	if (uap->uaddr2 == NULL) {
4102 		error = do_rw_wrlock(td, uap->obj, 0);
4103 	} else {
4104 		error = ops->copyin_umtx_time(uap->uaddr2,
4105 		   (size_t)uap->uaddr1, &timeout);
4106 		if (error != 0)
4107 			return (error);
4108 
4109 		error = do_rw_wrlock(td, uap->obj, &timeout);
4110 	}
4111 	return (error);
4112 }
4113 
4114 static int
4115 __umtx_op_rw_unlock(struct thread *td, struct _umtx_op_args *uap,
4116     const struct umtx_copyops *ops __unused)
4117 {
4118 
4119 	return (do_rw_unlock(td, uap->obj));
4120 }
4121 
4122 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4123 static int
4124 __umtx_op_sem_wait(struct thread *td, struct _umtx_op_args *uap,
4125     const struct umtx_copyops *ops)
4126 {
4127 	struct _umtx_time *tm_p, timeout;
4128 	int error;
4129 
4130 	/* Allow a null timespec (wait forever). */
4131 	if (uap->uaddr2 == NULL)
4132 		tm_p = NULL;
4133 	else {
4134 		error = ops->copyin_umtx_time(
4135 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4136 		if (error != 0)
4137 			return (error);
4138 		tm_p = &timeout;
4139 	}
4140 	return (do_sem_wait(td, uap->obj, tm_p));
4141 }
4142 
4143 static int
4144 __umtx_op_sem_wake(struct thread *td, struct _umtx_op_args *uap,
4145     const struct umtx_copyops *ops __unused)
4146 {
4147 
4148 	return (do_sem_wake(td, uap->obj));
4149 }
4150 #endif
4151 
4152 static int
4153 __umtx_op_wake2_umutex(struct thread *td, struct _umtx_op_args *uap,
4154     const struct umtx_copyops *ops __unused)
4155 {
4156 
4157 	return (do_wake2_umutex(td, uap->obj, uap->val));
4158 }
4159 
4160 static int
4161 __umtx_op_sem2_wait(struct thread *td, struct _umtx_op_args *uap,
4162     const struct umtx_copyops *ops)
4163 {
4164 	struct _umtx_time *tm_p, timeout;
4165 	size_t uasize;
4166 	int error;
4167 
4168 	/* Allow a null timespec (wait forever). */
4169 	if (uap->uaddr2 == NULL) {
4170 		uasize = 0;
4171 		tm_p = NULL;
4172 	} else {
4173 		uasize = (size_t)uap->uaddr1;
4174 		error = ops->copyin_umtx_time(uap->uaddr2, uasize, &timeout);
4175 		if (error != 0)
4176 			return (error);
4177 		tm_p = &timeout;
4178 	}
4179 	error = do_sem2_wait(td, uap->obj, tm_p);
4180 	if (error == EINTR && uap->uaddr2 != NULL &&
4181 	    (timeout._flags & UMTX_ABSTIME) == 0 &&
4182 	    uasize >= ops->umtx_time_sz + ops->timespec_sz) {
4183 		error = ops->copyout_timeout(
4184 		    (void *)((uintptr_t)uap->uaddr2 + ops->umtx_time_sz),
4185 		    uasize - ops->umtx_time_sz, &timeout._timeout);
4186 		if (error == 0) {
4187 			error = EINTR;
4188 		}
4189 	}
4190 
4191 	return (error);
4192 }
4193 
4194 static int
4195 __umtx_op_sem2_wake(struct thread *td, struct _umtx_op_args *uap,
4196     const struct umtx_copyops *ops __unused)
4197 {
4198 
4199 	return (do_sem2_wake(td, uap->obj));
4200 }
4201 
4202 #define	USHM_OBJ_UMTX(o)						\
4203     ((struct umtx_shm_obj_list *)(&(o)->umtx_data))
4204 
4205 #define	USHMF_REG_LINKED	0x0001
4206 #define	USHMF_OBJ_LINKED	0x0002
4207 struct umtx_shm_reg {
4208 	TAILQ_ENTRY(umtx_shm_reg) ushm_reg_link;
4209 	LIST_ENTRY(umtx_shm_reg) ushm_obj_link;
4210 	struct umtx_key		ushm_key;
4211 	struct ucred		*ushm_cred;
4212 	struct shmfd		*ushm_obj;
4213 	u_int			ushm_refcnt;
4214 	u_int			ushm_flags;
4215 };
4216 
4217 LIST_HEAD(umtx_shm_obj_list, umtx_shm_reg);
4218 TAILQ_HEAD(umtx_shm_reg_head, umtx_shm_reg);
4219 
4220 static uma_zone_t umtx_shm_reg_zone;
4221 static struct umtx_shm_reg_head umtx_shm_registry[UMTX_CHAINS];
4222 static struct mtx umtx_shm_lock;
4223 static struct umtx_shm_reg_head umtx_shm_reg_delfree =
4224     TAILQ_HEAD_INITIALIZER(umtx_shm_reg_delfree);
4225 
4226 static void umtx_shm_free_reg(struct umtx_shm_reg *reg);
4227 
4228 static void
4229 umtx_shm_reg_delfree_tq(void *context __unused, int pending __unused)
4230 {
4231 	struct umtx_shm_reg_head d;
4232 	struct umtx_shm_reg *reg, *reg1;
4233 
4234 	TAILQ_INIT(&d);
4235 	mtx_lock(&umtx_shm_lock);
4236 	TAILQ_CONCAT(&d, &umtx_shm_reg_delfree, ushm_reg_link);
4237 	mtx_unlock(&umtx_shm_lock);
4238 	TAILQ_FOREACH_SAFE(reg, &d, ushm_reg_link, reg1) {
4239 		TAILQ_REMOVE(&d, reg, ushm_reg_link);
4240 		umtx_shm_free_reg(reg);
4241 	}
4242 }
4243 
4244 static struct task umtx_shm_reg_delfree_task =
4245     TASK_INITIALIZER(0, umtx_shm_reg_delfree_tq, NULL);
4246 
4247 static struct umtx_shm_reg *
4248 umtx_shm_find_reg_locked(const struct umtx_key *key)
4249 {
4250 	struct umtx_shm_reg *reg;
4251 	struct umtx_shm_reg_head *reg_head;
4252 
4253 	KASSERT(key->shared, ("umtx_p_find_rg: private key"));
4254 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4255 	reg_head = &umtx_shm_registry[key->hash];
4256 	TAILQ_FOREACH(reg, reg_head, ushm_reg_link) {
4257 		KASSERT(reg->ushm_key.shared,
4258 		    ("non-shared key on reg %p %d", reg, reg->ushm_key.shared));
4259 		if (reg->ushm_key.info.shared.object ==
4260 		    key->info.shared.object &&
4261 		    reg->ushm_key.info.shared.offset ==
4262 		    key->info.shared.offset) {
4263 			KASSERT(reg->ushm_key.type == TYPE_SHM, ("TYPE_USHM"));
4264 			KASSERT(reg->ushm_refcnt > 0,
4265 			    ("reg %p refcnt 0 onlist", reg));
4266 			KASSERT((reg->ushm_flags & USHMF_REG_LINKED) != 0,
4267 			    ("reg %p not linked", reg));
4268 			reg->ushm_refcnt++;
4269 			return (reg);
4270 		}
4271 	}
4272 	return (NULL);
4273 }
4274 
4275 static struct umtx_shm_reg *
4276 umtx_shm_find_reg(const struct umtx_key *key)
4277 {
4278 	struct umtx_shm_reg *reg;
4279 
4280 	mtx_lock(&umtx_shm_lock);
4281 	reg = umtx_shm_find_reg_locked(key);
4282 	mtx_unlock(&umtx_shm_lock);
4283 	return (reg);
4284 }
4285 
4286 static void
4287 umtx_shm_free_reg(struct umtx_shm_reg *reg)
4288 {
4289 
4290 	chgumtxcnt(reg->ushm_cred->cr_ruidinfo, -1, 0);
4291 	crfree(reg->ushm_cred);
4292 	shm_drop(reg->ushm_obj);
4293 	uma_zfree(umtx_shm_reg_zone, reg);
4294 }
4295 
4296 static bool
4297 umtx_shm_unref_reg_locked(struct umtx_shm_reg *reg, bool force)
4298 {
4299 	bool res;
4300 
4301 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4302 	KASSERT(reg->ushm_refcnt > 0, ("ushm_reg %p refcnt 0", reg));
4303 	reg->ushm_refcnt--;
4304 	res = reg->ushm_refcnt == 0;
4305 	if (res || force) {
4306 		if ((reg->ushm_flags & USHMF_REG_LINKED) != 0) {
4307 			TAILQ_REMOVE(&umtx_shm_registry[reg->ushm_key.hash],
4308 			    reg, ushm_reg_link);
4309 			reg->ushm_flags &= ~USHMF_REG_LINKED;
4310 		}
4311 		if ((reg->ushm_flags & USHMF_OBJ_LINKED) != 0) {
4312 			LIST_REMOVE(reg, ushm_obj_link);
4313 			reg->ushm_flags &= ~USHMF_OBJ_LINKED;
4314 		}
4315 	}
4316 	return (res);
4317 }
4318 
4319 static void
4320 umtx_shm_unref_reg(struct umtx_shm_reg *reg, bool force)
4321 {
4322 	vm_object_t object;
4323 	bool dofree;
4324 
4325 	if (force) {
4326 		object = reg->ushm_obj->shm_object;
4327 		VM_OBJECT_WLOCK(object);
4328 		object->flags |= OBJ_UMTXDEAD;
4329 		VM_OBJECT_WUNLOCK(object);
4330 	}
4331 	mtx_lock(&umtx_shm_lock);
4332 	dofree = umtx_shm_unref_reg_locked(reg, force);
4333 	mtx_unlock(&umtx_shm_lock);
4334 	if (dofree)
4335 		umtx_shm_free_reg(reg);
4336 }
4337 
4338 void
4339 umtx_shm_object_init(vm_object_t object)
4340 {
4341 
4342 	LIST_INIT(USHM_OBJ_UMTX(object));
4343 }
4344 
4345 void
4346 umtx_shm_object_terminated(vm_object_t object)
4347 {
4348 	struct umtx_shm_reg *reg, *reg1;
4349 	bool dofree;
4350 
4351 	if (LIST_EMPTY(USHM_OBJ_UMTX(object)))
4352 		return;
4353 
4354 	dofree = false;
4355 	mtx_lock(&umtx_shm_lock);
4356 	LIST_FOREACH_SAFE(reg, USHM_OBJ_UMTX(object), ushm_obj_link, reg1) {
4357 		if (umtx_shm_unref_reg_locked(reg, true)) {
4358 			TAILQ_INSERT_TAIL(&umtx_shm_reg_delfree, reg,
4359 			    ushm_reg_link);
4360 			dofree = true;
4361 		}
4362 	}
4363 	mtx_unlock(&umtx_shm_lock);
4364 	if (dofree)
4365 		taskqueue_enqueue(taskqueue_thread, &umtx_shm_reg_delfree_task);
4366 }
4367 
4368 static int
4369 umtx_shm_create_reg(struct thread *td, const struct umtx_key *key,
4370     struct umtx_shm_reg **res)
4371 {
4372 	struct umtx_shm_reg *reg, *reg1;
4373 	struct ucred *cred;
4374 	int error;
4375 
4376 	reg = umtx_shm_find_reg(key);
4377 	if (reg != NULL) {
4378 		*res = reg;
4379 		return (0);
4380 	}
4381 	cred = td->td_ucred;
4382 	if (!chgumtxcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_UMTXP)))
4383 		return (ENOMEM);
4384 	reg = uma_zalloc(umtx_shm_reg_zone, M_WAITOK | M_ZERO);
4385 	reg->ushm_refcnt = 1;
4386 	bcopy(key, &reg->ushm_key, sizeof(*key));
4387 	reg->ushm_obj = shm_alloc(td->td_ucred, O_RDWR, false);
4388 	reg->ushm_cred = crhold(cred);
4389 	error = shm_dotruncate(reg->ushm_obj, PAGE_SIZE);
4390 	if (error != 0) {
4391 		umtx_shm_free_reg(reg);
4392 		return (error);
4393 	}
4394 	mtx_lock(&umtx_shm_lock);
4395 	reg1 = umtx_shm_find_reg_locked(key);
4396 	if (reg1 != NULL) {
4397 		mtx_unlock(&umtx_shm_lock);
4398 		umtx_shm_free_reg(reg);
4399 		*res = reg1;
4400 		return (0);
4401 	}
4402 	reg->ushm_refcnt++;
4403 	TAILQ_INSERT_TAIL(&umtx_shm_registry[key->hash], reg, ushm_reg_link);
4404 	LIST_INSERT_HEAD(USHM_OBJ_UMTX(key->info.shared.object), reg,
4405 	    ushm_obj_link);
4406 	reg->ushm_flags = USHMF_REG_LINKED | USHMF_OBJ_LINKED;
4407 	mtx_unlock(&umtx_shm_lock);
4408 	*res = reg;
4409 	return (0);
4410 }
4411 
4412 static int
4413 umtx_shm_alive(struct thread *td, void *addr)
4414 {
4415 	vm_map_t map;
4416 	vm_map_entry_t entry;
4417 	vm_object_t object;
4418 	vm_pindex_t pindex;
4419 	vm_prot_t prot;
4420 	int res, ret;
4421 	boolean_t wired;
4422 
4423 	map = &td->td_proc->p_vmspace->vm_map;
4424 	res = vm_map_lookup(&map, (uintptr_t)addr, VM_PROT_READ, &entry,
4425 	    &object, &pindex, &prot, &wired);
4426 	if (res != KERN_SUCCESS)
4427 		return (EFAULT);
4428 	if (object == NULL)
4429 		ret = EINVAL;
4430 	else
4431 		ret = (object->flags & OBJ_UMTXDEAD) != 0 ? ENOTTY : 0;
4432 	vm_map_lookup_done(map, entry);
4433 	return (ret);
4434 }
4435 
4436 static void
4437 umtx_shm_init(void)
4438 {
4439 	int i;
4440 
4441 	umtx_shm_reg_zone = uma_zcreate("umtx_shm", sizeof(struct umtx_shm_reg),
4442 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
4443 	mtx_init(&umtx_shm_lock, "umtxshm", NULL, MTX_DEF);
4444 	for (i = 0; i < nitems(umtx_shm_registry); i++)
4445 		TAILQ_INIT(&umtx_shm_registry[i]);
4446 }
4447 
4448 static int
4449 umtx_shm(struct thread *td, void *addr, u_int flags)
4450 {
4451 	struct umtx_key key;
4452 	struct umtx_shm_reg *reg;
4453 	struct file *fp;
4454 	int error, fd;
4455 
4456 	if (__bitcount(flags & (UMTX_SHM_CREAT | UMTX_SHM_LOOKUP |
4457 	    UMTX_SHM_DESTROY| UMTX_SHM_ALIVE)) != 1)
4458 		return (EINVAL);
4459 	if ((flags & UMTX_SHM_ALIVE) != 0)
4460 		return (umtx_shm_alive(td, addr));
4461 	error = umtx_key_get(addr, TYPE_SHM, PROCESS_SHARE, &key);
4462 	if (error != 0)
4463 		return (error);
4464 	KASSERT(key.shared == 1, ("non-shared key"));
4465 	if ((flags & UMTX_SHM_CREAT) != 0) {
4466 		error = umtx_shm_create_reg(td, &key, &reg);
4467 	} else {
4468 		reg = umtx_shm_find_reg(&key);
4469 		if (reg == NULL)
4470 			error = ESRCH;
4471 	}
4472 	umtx_key_release(&key);
4473 	if (error != 0)
4474 		return (error);
4475 	KASSERT(reg != NULL, ("no reg"));
4476 	if ((flags & UMTX_SHM_DESTROY) != 0) {
4477 		umtx_shm_unref_reg(reg, true);
4478 	} else {
4479 #if 0
4480 #ifdef MAC
4481 		error = mac_posixshm_check_open(td->td_ucred,
4482 		    reg->ushm_obj, FFLAGS(O_RDWR));
4483 		if (error == 0)
4484 #endif
4485 			error = shm_access(reg->ushm_obj, td->td_ucred,
4486 			    FFLAGS(O_RDWR));
4487 		if (error == 0)
4488 #endif
4489 			error = falloc_caps(td, &fp, &fd, O_CLOEXEC, NULL);
4490 		if (error == 0) {
4491 			shm_hold(reg->ushm_obj);
4492 			finit(fp, FFLAGS(O_RDWR), DTYPE_SHM, reg->ushm_obj,
4493 			    &shm_ops);
4494 			td->td_retval[0] = fd;
4495 			fdrop(fp, td);
4496 		}
4497 	}
4498 	umtx_shm_unref_reg(reg, false);
4499 	return (error);
4500 }
4501 
4502 static int
4503 __umtx_op_shm(struct thread *td, struct _umtx_op_args *uap,
4504     const struct umtx_copyops *ops __unused)
4505 {
4506 
4507 	return (umtx_shm(td, uap->uaddr1, uap->val));
4508 }
4509 
4510 static int
4511 __umtx_op_robust_lists(struct thread *td, struct _umtx_op_args *uap,
4512     const struct umtx_copyops *ops)
4513 {
4514 	struct umtx_robust_lists_params rb;
4515 	int error;
4516 
4517 	if (ops->compat32) {
4518 		if ((td->td_pflags2 & TDP2_COMPAT32RB) == 0 &&
4519 		    (td->td_rb_list != 0 || td->td_rbp_list != 0 ||
4520 		    td->td_rb_inact != 0))
4521 			return (EBUSY);
4522 	} else if ((td->td_pflags2 & TDP2_COMPAT32RB) != 0) {
4523 		return (EBUSY);
4524 	}
4525 
4526 	bzero(&rb, sizeof(rb));
4527 	error = ops->copyin_robust_lists(uap->uaddr1, uap->val, &rb);
4528 	if (error != 0)
4529 		return (error);
4530 
4531 	if (ops->compat32)
4532 		td->td_pflags2 |= TDP2_COMPAT32RB;
4533 
4534 	td->td_rb_list = rb.robust_list_offset;
4535 	td->td_rbp_list = rb.robust_priv_list_offset;
4536 	td->td_rb_inact = rb.robust_inact_offset;
4537 	return (0);
4538 }
4539 
4540 #if defined(__i386__) || defined(__amd64__)
4541 /*
4542  * Provide the standard 32-bit definitions for x86, since native/compat32 use a
4543  * 32-bit time_t there.  Other architectures just need the i386 definitions
4544  * along with their standard compat32.
4545  */
4546 struct timespecx32 {
4547 	int64_t			tv_sec;
4548 	int32_t			tv_nsec;
4549 };
4550 
4551 struct umtx_timex32 {
4552 	struct	timespecx32	_timeout;
4553 	uint32_t		_flags;
4554 	uint32_t		_clockid;
4555 };
4556 
4557 #ifndef __i386__
4558 #define	timespeci386	timespec32
4559 #define	umtx_timei386	umtx_time32
4560 #endif
4561 #else /* !__i386__ && !__amd64__ */
4562 /* 32-bit architectures can emulate i386, so define these almost everywhere. */
4563 struct timespeci386 {
4564 	int32_t			tv_sec;
4565 	int32_t			tv_nsec;
4566 };
4567 
4568 struct umtx_timei386 {
4569 	struct	timespeci386	_timeout;
4570 	uint32_t		_flags;
4571 	uint32_t		_clockid;
4572 };
4573 
4574 #if defined(__LP64__)
4575 #define	timespecx32	timespec32
4576 #define	umtx_timex32	umtx_time32
4577 #endif
4578 #endif
4579 
4580 static int
4581 umtx_copyin_robust_lists32(const void *uaddr, size_t size,
4582     struct umtx_robust_lists_params *rbp)
4583 {
4584 	struct umtx_robust_lists_params_compat32 rb32;
4585 	int error;
4586 
4587 	if (size > sizeof(rb32))
4588 		return (EINVAL);
4589 	bzero(&rb32, sizeof(rb32));
4590 	error = copyin(uaddr, &rb32, size);
4591 	if (error != 0)
4592 		return (error);
4593 	CP(rb32, *rbp, robust_list_offset);
4594 	CP(rb32, *rbp, robust_priv_list_offset);
4595 	CP(rb32, *rbp, robust_inact_offset);
4596 	return (0);
4597 }
4598 
4599 #ifndef __i386__
4600 static inline int
4601 umtx_copyin_timeouti386(const void *uaddr, struct timespec *tsp)
4602 {
4603 	struct timespeci386 ts32;
4604 	int error;
4605 
4606 	error = copyin(uaddr, &ts32, sizeof(ts32));
4607 	if (error == 0) {
4608 		if (ts32.tv_sec < 0 ||
4609 		    ts32.tv_nsec >= 1000000000 ||
4610 		    ts32.tv_nsec < 0)
4611 			error = EINVAL;
4612 		else {
4613 			CP(ts32, *tsp, tv_sec);
4614 			CP(ts32, *tsp, tv_nsec);
4615 		}
4616 	}
4617 	return (error);
4618 }
4619 
4620 static inline int
4621 umtx_copyin_umtx_timei386(const void *uaddr, size_t size, struct _umtx_time *tp)
4622 {
4623 	struct umtx_timei386 t32;
4624 	int error;
4625 
4626 	t32._clockid = CLOCK_REALTIME;
4627 	t32._flags   = 0;
4628 	if (size <= sizeof(t32._timeout))
4629 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4630 	else
4631 		error = copyin(uaddr, &t32, sizeof(t32));
4632 	if (error != 0)
4633 		return (error);
4634 	if (t32._timeout.tv_sec < 0 ||
4635 	    t32._timeout.tv_nsec >= 1000000000 || t32._timeout.tv_nsec < 0)
4636 		return (EINVAL);
4637 	TS_CP(t32, *tp, _timeout);
4638 	CP(t32, *tp, _flags);
4639 	CP(t32, *tp, _clockid);
4640 	return (0);
4641 }
4642 
4643 static int
4644 umtx_copyout_timeouti386(void *uaddr, size_t sz, struct timespec *tsp)
4645 {
4646 	struct timespeci386 remain32 = {
4647 		.tv_sec = tsp->tv_sec,
4648 		.tv_nsec = tsp->tv_nsec,
4649 	};
4650 
4651 	/*
4652 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4653 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4654 	 * copyops.
4655 	 */
4656 	KASSERT(sz >= sizeof(remain32),
4657 	    ("umtx_copyops specifies incorrect sizes"));
4658 
4659 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4660 }
4661 #endif /* !__i386__ */
4662 
4663 #if defined(__i386__) || defined(__LP64__)
4664 static inline int
4665 umtx_copyin_timeoutx32(const void *uaddr, struct timespec *tsp)
4666 {
4667 	struct timespecx32 ts32;
4668 	int error;
4669 
4670 	error = copyin(uaddr, &ts32, sizeof(ts32));
4671 	if (error == 0) {
4672 		if (ts32.tv_sec < 0 ||
4673 		    ts32.tv_nsec >= 1000000000 ||
4674 		    ts32.tv_nsec < 0)
4675 			error = EINVAL;
4676 		else {
4677 			CP(ts32, *tsp, tv_sec);
4678 			CP(ts32, *tsp, tv_nsec);
4679 		}
4680 	}
4681 	return (error);
4682 }
4683 
4684 static inline int
4685 umtx_copyin_umtx_timex32(const void *uaddr, size_t size, struct _umtx_time *tp)
4686 {
4687 	struct umtx_timex32 t32;
4688 	int error;
4689 
4690 	t32._clockid = CLOCK_REALTIME;
4691 	t32._flags   = 0;
4692 	if (size <= sizeof(t32._timeout))
4693 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4694 	else
4695 		error = copyin(uaddr, &t32, sizeof(t32));
4696 	if (error != 0)
4697 		return (error);
4698 	if (t32._timeout.tv_sec < 0 ||
4699 	    t32._timeout.tv_nsec >= 1000000000 || t32._timeout.tv_nsec < 0)
4700 		return (EINVAL);
4701 	TS_CP(t32, *tp, _timeout);
4702 	CP(t32, *tp, _flags);
4703 	CP(t32, *tp, _clockid);
4704 	return (0);
4705 }
4706 
4707 static int
4708 umtx_copyout_timeoutx32(void *uaddr, size_t sz, struct timespec *tsp)
4709 {
4710 	struct timespecx32 remain32 = {
4711 		.tv_sec = tsp->tv_sec,
4712 		.tv_nsec = tsp->tv_nsec,
4713 	};
4714 
4715 	/*
4716 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4717 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4718 	 * copyops.
4719 	 */
4720 	KASSERT(sz >= sizeof(remain32),
4721 	    ("umtx_copyops specifies incorrect sizes"));
4722 
4723 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4724 }
4725 #endif /* __i386__ || __LP64__ */
4726 
4727 typedef int (*_umtx_op_func)(struct thread *td, struct _umtx_op_args *uap,
4728     const struct umtx_copyops *umtx_ops);
4729 
4730 static const _umtx_op_func op_table[] = {
4731 #ifdef COMPAT_FREEBSD10
4732 	[UMTX_OP_LOCK]		= __umtx_op_lock_umtx,
4733 	[UMTX_OP_UNLOCK]	= __umtx_op_unlock_umtx,
4734 #else
4735 	[UMTX_OP_LOCK]		= __umtx_op_unimpl,
4736 	[UMTX_OP_UNLOCK]	= __umtx_op_unimpl,
4737 #endif
4738 	[UMTX_OP_WAIT]		= __umtx_op_wait,
4739 	[UMTX_OP_WAKE]		= __umtx_op_wake,
4740 	[UMTX_OP_MUTEX_TRYLOCK]	= __umtx_op_trylock_umutex,
4741 	[UMTX_OP_MUTEX_LOCK]	= __umtx_op_lock_umutex,
4742 	[UMTX_OP_MUTEX_UNLOCK]	= __umtx_op_unlock_umutex,
4743 	[UMTX_OP_SET_CEILING]	= __umtx_op_set_ceiling,
4744 	[UMTX_OP_CV_WAIT]	= __umtx_op_cv_wait,
4745 	[UMTX_OP_CV_SIGNAL]	= __umtx_op_cv_signal,
4746 	[UMTX_OP_CV_BROADCAST]	= __umtx_op_cv_broadcast,
4747 	[UMTX_OP_WAIT_UINT]	= __umtx_op_wait_uint,
4748 	[UMTX_OP_RW_RDLOCK]	= __umtx_op_rw_rdlock,
4749 	[UMTX_OP_RW_WRLOCK]	= __umtx_op_rw_wrlock,
4750 	[UMTX_OP_RW_UNLOCK]	= __umtx_op_rw_unlock,
4751 	[UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private,
4752 	[UMTX_OP_WAKE_PRIVATE]	= __umtx_op_wake_private,
4753 	[UMTX_OP_MUTEX_WAIT]	= __umtx_op_wait_umutex,
4754 	[UMTX_OP_MUTEX_WAKE]	= __umtx_op_wake_umutex,
4755 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4756 	[UMTX_OP_SEM_WAIT]	= __umtx_op_sem_wait,
4757 	[UMTX_OP_SEM_WAKE]	= __umtx_op_sem_wake,
4758 #else
4759 	[UMTX_OP_SEM_WAIT]	= __umtx_op_unimpl,
4760 	[UMTX_OP_SEM_WAKE]	= __umtx_op_unimpl,
4761 #endif
4762 	[UMTX_OP_NWAKE_PRIVATE]	= __umtx_op_nwake_private,
4763 	[UMTX_OP_MUTEX_WAKE2]	= __umtx_op_wake2_umutex,
4764 	[UMTX_OP_SEM2_WAIT]	= __umtx_op_sem2_wait,
4765 	[UMTX_OP_SEM2_WAKE]	= __umtx_op_sem2_wake,
4766 	[UMTX_OP_SHM]		= __umtx_op_shm,
4767 	[UMTX_OP_ROBUST_LISTS]	= __umtx_op_robust_lists,
4768 };
4769 
4770 static const struct umtx_copyops umtx_native_ops = {
4771 	.copyin_timeout = umtx_copyin_timeout,
4772 	.copyin_umtx_time = umtx_copyin_umtx_time,
4773 	.copyin_robust_lists = umtx_copyin_robust_lists,
4774 	.copyout_timeout = umtx_copyout_timeout,
4775 	.timespec_sz = sizeof(struct timespec),
4776 	.umtx_time_sz = sizeof(struct _umtx_time),
4777 };
4778 
4779 #ifndef __i386__
4780 static const struct umtx_copyops umtx_native_opsi386 = {
4781 	.copyin_timeout = umtx_copyin_timeouti386,
4782 	.copyin_umtx_time = umtx_copyin_umtx_timei386,
4783 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4784 	.copyout_timeout = umtx_copyout_timeouti386,
4785 	.timespec_sz = sizeof(struct timespeci386),
4786 	.umtx_time_sz = sizeof(struct umtx_timei386),
4787 	.compat32 = true,
4788 };
4789 #endif
4790 
4791 #if defined(__i386__) || defined(__LP64__)
4792 /* i386 can emulate other 32-bit archs, too! */
4793 static const struct umtx_copyops umtx_native_opsx32 = {
4794 	.copyin_timeout = umtx_copyin_timeoutx32,
4795 	.copyin_umtx_time = umtx_copyin_umtx_timex32,
4796 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4797 	.copyout_timeout = umtx_copyout_timeoutx32,
4798 	.timespec_sz = sizeof(struct timespecx32),
4799 	.umtx_time_sz = sizeof(struct umtx_timex32),
4800 	.compat32 = true,
4801 };
4802 
4803 #ifdef COMPAT_FREEBSD32
4804 #ifdef __amd64__
4805 #define	umtx_native_ops32	umtx_native_opsi386
4806 #else
4807 #define	umtx_native_ops32	umtx_native_opsx32
4808 #endif
4809 #endif /* COMPAT_FREEBSD32 */
4810 #endif /* __i386__ || __LP64__ */
4811 
4812 #define	UMTX_OP__FLAGS	(UMTX_OP__32BIT | UMTX_OP__I386)
4813 
4814 static int
4815 kern__umtx_op(struct thread *td, void *obj, int op, unsigned long val,
4816     void *uaddr1, void *uaddr2, const struct umtx_copyops *ops)
4817 {
4818 	struct _umtx_op_args uap = {
4819 		.obj = obj,
4820 		.op = op & ~UMTX_OP__FLAGS,
4821 		.val = val,
4822 		.uaddr1 = uaddr1,
4823 		.uaddr2 = uaddr2
4824 	};
4825 
4826 	if ((uap.op >= nitems(op_table)))
4827 		return (EINVAL);
4828 	return ((*op_table[uap.op])(td, &uap, ops));
4829 }
4830 
4831 int
4832 sys__umtx_op(struct thread *td, struct _umtx_op_args *uap)
4833 {
4834 	static const struct umtx_copyops *umtx_ops;
4835 
4836 	umtx_ops = &umtx_native_ops;
4837 #ifdef __LP64__
4838 	if ((uap->op & (UMTX_OP__32BIT | UMTX_OP__I386)) != 0) {
4839 		if ((uap->op & UMTX_OP__I386) != 0)
4840 			umtx_ops = &umtx_native_opsi386;
4841 		else
4842 			umtx_ops = &umtx_native_opsx32;
4843 	}
4844 #elif !defined(__i386__)
4845 	/* We consider UMTX_OP__32BIT a nop on !i386 ILP32. */
4846 	if ((uap->op & UMTX_OP__I386) != 0)
4847 		umtx_ops = &umtx_native_opsi386;
4848 #else
4849 	/* Likewise, UMTX_OP__I386 is a nop on i386. */
4850 	if ((uap->op & UMTX_OP__32BIT) != 0)
4851 		umtx_ops = &umtx_native_opsx32;
4852 #endif
4853 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4854 	    uap->uaddr2, umtx_ops));
4855 }
4856 
4857 #ifdef COMPAT_FREEBSD32
4858 #ifdef COMPAT_FREEBSD10
4859 int
4860 freebsd10_freebsd32_umtx_lock(struct thread *td,
4861     struct freebsd10_freebsd32_umtx_lock_args *uap)
4862 {
4863 	return (do_lock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid, NULL));
4864 }
4865 
4866 int
4867 freebsd10_freebsd32_umtx_unlock(struct thread *td,
4868     struct freebsd10_freebsd32_umtx_unlock_args *uap)
4869 {
4870 	return (do_unlock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid));
4871 }
4872 #endif /* COMPAT_FREEBSD10 */
4873 
4874 int
4875 freebsd32__umtx_op(struct thread *td, struct freebsd32__umtx_op_args *uap)
4876 {
4877 
4878 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr,
4879 	    uap->uaddr2, &umtx_native_ops32));
4880 }
4881 #endif /* COMPAT_FREEBSD32 */
4882 
4883 void
4884 umtx_thread_init(struct thread *td)
4885 {
4886 
4887 	td->td_umtxq = umtxq_alloc();
4888 	td->td_umtxq->uq_thread = td;
4889 }
4890 
4891 void
4892 umtx_thread_fini(struct thread *td)
4893 {
4894 
4895 	umtxq_free(td->td_umtxq);
4896 }
4897 
4898 /*
4899  * It will be called when new thread is created, e.g fork().
4900  */
4901 void
4902 umtx_thread_alloc(struct thread *td)
4903 {
4904 	struct umtx_q *uq;
4905 
4906 	uq = td->td_umtxq;
4907 	uq->uq_inherited_pri = PRI_MAX;
4908 
4909 	KASSERT(uq->uq_flags == 0, ("uq_flags != 0"));
4910 	KASSERT(uq->uq_thread == td, ("uq_thread != td"));
4911 	KASSERT(uq->uq_pi_blocked == NULL, ("uq_pi_blocked != NULL"));
4912 	KASSERT(TAILQ_EMPTY(&uq->uq_pi_contested), ("uq_pi_contested is not empty"));
4913 }
4914 
4915 /*
4916  * exec() hook.
4917  *
4918  * Clear robust lists for all process' threads, not delaying the
4919  * cleanup to thread exit, since the relevant address space is
4920  * destroyed right now.
4921  */
4922 void
4923 umtx_exec(struct proc *p)
4924 {
4925 	struct thread *td;
4926 
4927 	KASSERT(p == curproc, ("need curproc"));
4928 	KASSERT((p->p_flag & P_HADTHREADS) == 0 ||
4929 	    (p->p_flag & P_STOPPED_SINGLE) != 0,
4930 	    ("curproc must be single-threaded"));
4931 	/*
4932 	 * There is no need to lock the list as only this thread can be
4933 	 * running.
4934 	 */
4935 	FOREACH_THREAD_IN_PROC(p, td) {
4936 		KASSERT(td == curthread ||
4937 		    ((td->td_flags & TDF_BOUNDARY) != 0 && TD_IS_SUSPENDED(td)),
4938 		    ("running thread %p %p", p, td));
4939 		umtx_thread_cleanup(td);
4940 		td->td_rb_list = td->td_rbp_list = td->td_rb_inact = 0;
4941 	}
4942 }
4943 
4944 /*
4945  * thread exit hook.
4946  */
4947 void
4948 umtx_thread_exit(struct thread *td)
4949 {
4950 
4951 	umtx_thread_cleanup(td);
4952 }
4953 
4954 static int
4955 umtx_read_uptr(struct thread *td, uintptr_t ptr, uintptr_t *res, bool compat32)
4956 {
4957 	u_long res1;
4958 	uint32_t res32;
4959 	int error;
4960 
4961 	if (compat32) {
4962 		error = fueword32((void *)ptr, &res32);
4963 		if (error == 0)
4964 			res1 = res32;
4965 	} else {
4966 		error = fueword((void *)ptr, &res1);
4967 	}
4968 	if (error == 0)
4969 		*res = res1;
4970 	else
4971 		error = EFAULT;
4972 	return (error);
4973 }
4974 
4975 static void
4976 umtx_read_rb_list(struct thread *td, struct umutex *m, uintptr_t *rb_list,
4977     bool compat32)
4978 {
4979 	struct umutex32 m32;
4980 
4981 	if (compat32) {
4982 		memcpy(&m32, m, sizeof(m32));
4983 		*rb_list = m32.m_rb_lnk;
4984 	} else {
4985 		*rb_list = m->m_rb_lnk;
4986 	}
4987 }
4988 
4989 static int
4990 umtx_handle_rb(struct thread *td, uintptr_t rbp, uintptr_t *rb_list, bool inact,
4991     bool compat32)
4992 {
4993 	struct umutex m;
4994 	int error;
4995 
4996 	KASSERT(td->td_proc == curproc, ("need current vmspace"));
4997 	error = copyin((void *)rbp, &m, sizeof(m));
4998 	if (error != 0)
4999 		return (error);
5000 	if (rb_list != NULL)
5001 		umtx_read_rb_list(td, &m, rb_list, compat32);
5002 	if ((m.m_flags & UMUTEX_ROBUST) == 0)
5003 		return (EINVAL);
5004 	if ((m.m_owner & ~UMUTEX_CONTESTED) != td->td_tid)
5005 		/* inact is cleared after unlock, allow the inconsistency */
5006 		return (inact ? 0 : EINVAL);
5007 	return (do_unlock_umutex(td, (struct umutex *)rbp, true));
5008 }
5009 
5010 static void
5011 umtx_cleanup_rb_list(struct thread *td, uintptr_t rb_list, uintptr_t *rb_inact,
5012     const char *name, bool compat32)
5013 {
5014 	int error, i;
5015 	uintptr_t rbp;
5016 	bool inact;
5017 
5018 	if (rb_list == 0)
5019 		return;
5020 	error = umtx_read_uptr(td, rb_list, &rbp, compat32);
5021 	for (i = 0; error == 0 && rbp != 0 && i < umtx_max_rb; i++) {
5022 		if (rbp == *rb_inact) {
5023 			inact = true;
5024 			*rb_inact = 0;
5025 		} else
5026 			inact = false;
5027 		error = umtx_handle_rb(td, rbp, &rbp, inact, compat32);
5028 	}
5029 	if (i == umtx_max_rb && umtx_verbose_rb) {
5030 		uprintf("comm %s pid %d: reached umtx %smax rb %d\n",
5031 		    td->td_proc->p_comm, td->td_proc->p_pid, name, umtx_max_rb);
5032 	}
5033 	if (error != 0 && umtx_verbose_rb) {
5034 		uprintf("comm %s pid %d: handling %srb error %d\n",
5035 		    td->td_proc->p_comm, td->td_proc->p_pid, name, error);
5036 	}
5037 }
5038 
5039 /*
5040  * Clean up umtx data.
5041  */
5042 static void
5043 umtx_thread_cleanup(struct thread *td)
5044 {
5045 	struct umtx_q *uq;
5046 	struct umtx_pi *pi;
5047 	uintptr_t rb_inact;
5048 	bool compat32;
5049 
5050 	/*
5051 	 * Disown pi mutexes.
5052 	 */
5053 	uq = td->td_umtxq;
5054 	if (uq != NULL) {
5055 		if (uq->uq_inherited_pri != PRI_MAX ||
5056 		    !TAILQ_EMPTY(&uq->uq_pi_contested)) {
5057 			mtx_lock(&umtx_lock);
5058 			uq->uq_inherited_pri = PRI_MAX;
5059 			while ((pi = TAILQ_FIRST(&uq->uq_pi_contested)) != NULL) {
5060 				pi->pi_owner = NULL;
5061 				TAILQ_REMOVE(&uq->uq_pi_contested, pi, pi_link);
5062 			}
5063 			mtx_unlock(&umtx_lock);
5064 		}
5065 		sched_lend_user_prio_cond(td, PRI_MAX);
5066 	}
5067 
5068 	compat32 = (td->td_pflags2 & TDP2_COMPAT32RB) != 0;
5069 	td->td_pflags2 &= ~TDP2_COMPAT32RB;
5070 
5071 	if (td->td_rb_inact == 0 && td->td_rb_list == 0 && td->td_rbp_list == 0)
5072 		return;
5073 
5074 	/*
5075 	 * Handle terminated robust mutexes.  Must be done after
5076 	 * robust pi disown, otherwise unlock could see unowned
5077 	 * entries.
5078 	 */
5079 	rb_inact = td->td_rb_inact;
5080 	if (rb_inact != 0)
5081 		(void)umtx_read_uptr(td, rb_inact, &rb_inact, compat32);
5082 	umtx_cleanup_rb_list(td, td->td_rb_list, &rb_inact, "", compat32);
5083 	umtx_cleanup_rb_list(td, td->td_rbp_list, &rb_inact, "priv ", compat32);
5084 	if (rb_inact != 0)
5085 		(void)umtx_handle_rb(td, rb_inact, NULL, true, compat32);
5086 }
5087