xref: /freebsd/sys/kern/kern_umtx.c (revision 3332f1b444d4a73238e9f59cca27bfc95fe936bd)
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;
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 	if (uh == NULL)
653 		return (0);
654 	TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
655 		if (++ret <= n_wake) {
656 			umtxq_remove(uq);
657 			wakeup_one(uq);
658 		} else {
659 			umtxq_remove(uq);
660 			uq->uq_key = *key2;
661 			umtxq_insert(uq);
662 			if (ret - n_wake == n_requeue)
663 				break;
664 		}
665 	}
666 	return (ret);
667 }
668 
669 static inline int
670 tstohz(const struct timespec *tsp)
671 {
672 	struct timeval tv;
673 
674 	TIMESPEC_TO_TIMEVAL(&tv, tsp);
675 	return tvtohz(&tv);
676 }
677 
678 void
679 umtx_abs_timeout_init(struct umtx_abs_timeout *timo, int clockid,
680     int absolute, const struct timespec *timeout)
681 {
682 
683 	timo->clockid = clockid;
684 	if (!absolute) {
685 		timo->is_abs_real = false;
686 		umtx_abs_timeout_update(timo);
687 		timespecadd(&timo->cur, timeout, &timo->end);
688 	} else {
689 		timo->end = *timeout;
690 		timo->is_abs_real = clockid == CLOCK_REALTIME ||
691 		    clockid == CLOCK_REALTIME_FAST ||
692 		    clockid == CLOCK_REALTIME_PRECISE;
693 		/*
694 		 * If is_abs_real, umtxq_sleep will read the clock
695 		 * after setting td_rtcgen; otherwise, read it here.
696 		 */
697 		if (!timo->is_abs_real) {
698 			umtx_abs_timeout_update(timo);
699 		}
700 	}
701 }
702 
703 static void
704 umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
705     const struct _umtx_time *umtxtime)
706 {
707 
708 	umtx_abs_timeout_init(timo, umtxtime->_clockid,
709 	    (umtxtime->_flags & UMTX_ABSTIME) != 0, &umtxtime->_timeout);
710 }
711 
712 static void
713 umtx_abs_timeout_update(struct umtx_abs_timeout *timo)
714 {
715 
716 	kern_clock_gettime(curthread, timo->clockid, &timo->cur);
717 }
718 
719 static int
720 umtx_abs_timeout_gethz(struct umtx_abs_timeout *timo)
721 {
722 	struct timespec tts;
723 
724 	if (timespeccmp(&timo->end, &timo->cur, <=))
725 		return (-1);
726 	timespecsub(&timo->end, &timo->cur, &tts);
727 	return (tstohz(&tts));
728 }
729 
730 static uint32_t
731 umtx_unlock_val(uint32_t flags, bool rb)
732 {
733 
734 	if (rb)
735 		return (UMUTEX_RB_OWNERDEAD);
736 	else if ((flags & UMUTEX_NONCONSISTENT) != 0)
737 		return (UMUTEX_RB_NOTRECOV);
738 	else
739 		return (UMUTEX_UNOWNED);
740 
741 }
742 
743 /*
744  * Put thread into sleep state, before sleeping, check if
745  * thread was removed from umtx queue.
746  */
747 int
748 umtxq_sleep(struct umtx_q *uq, const char *wmesg,
749     struct umtx_abs_timeout *abstime)
750 {
751 	struct umtxq_chain *uc;
752 	int error, timo;
753 
754 	if (abstime != NULL && abstime->is_abs_real) {
755 		curthread->td_rtcgen = atomic_load_acq_int(&rtc_generation);
756 		umtx_abs_timeout_update(abstime);
757 	}
758 
759 	uc = umtxq_getchain(&uq->uq_key);
760 	UMTXQ_LOCKED_ASSERT(uc);
761 	for (;;) {
762 		if (!(uq->uq_flags & UQF_UMTXQ)) {
763 			error = 0;
764 			break;
765 		}
766 		if (abstime != NULL) {
767 			timo = umtx_abs_timeout_gethz(abstime);
768 			if (timo < 0) {
769 				error = ETIMEDOUT;
770 				break;
771 			}
772 		} else
773 			timo = 0;
774 		error = msleep(uq, &uc->uc_lock, PCATCH | PDROP, wmesg, timo);
775 		if (error == EINTR || error == ERESTART) {
776 			umtxq_lock(&uq->uq_key);
777 			break;
778 		}
779 		if (abstime != NULL) {
780 			if (abstime->is_abs_real)
781 				curthread->td_rtcgen =
782 				    atomic_load_acq_int(&rtc_generation);
783 			umtx_abs_timeout_update(abstime);
784 		}
785 		umtxq_lock(&uq->uq_key);
786 	}
787 
788 	curthread->td_rtcgen = 0;
789 	return (error);
790 }
791 
792 /*
793  * Convert userspace address into unique logical address.
794  */
795 int
796 umtx_key_get(const void *addr, int type, int share, struct umtx_key *key)
797 {
798 	struct thread *td = curthread;
799 	vm_map_t map;
800 	vm_map_entry_t entry;
801 	vm_pindex_t pindex;
802 	vm_prot_t prot;
803 	boolean_t wired;
804 
805 	key->type = type;
806 	if (share == THREAD_SHARE) {
807 		key->shared = 0;
808 		key->info.private.vs = td->td_proc->p_vmspace;
809 		key->info.private.addr = (uintptr_t)addr;
810 	} else {
811 		MPASS(share == PROCESS_SHARE || share == AUTO_SHARE);
812 		map = &td->td_proc->p_vmspace->vm_map;
813 		if (vm_map_lookup(&map, (vm_offset_t)addr, VM_PROT_WRITE,
814 		    &entry, &key->info.shared.object, &pindex, &prot,
815 		    &wired) != KERN_SUCCESS) {
816 			return (EFAULT);
817 		}
818 
819 		if ((share == PROCESS_SHARE) ||
820 		    (share == AUTO_SHARE &&
821 		     VM_INHERIT_SHARE == entry->inheritance)) {
822 			key->shared = 1;
823 			key->info.shared.offset = (vm_offset_t)addr -
824 			    entry->start + entry->offset;
825 			vm_object_reference(key->info.shared.object);
826 		} else {
827 			key->shared = 0;
828 			key->info.private.vs = td->td_proc->p_vmspace;
829 			key->info.private.addr = (uintptr_t)addr;
830 		}
831 		vm_map_lookup_done(map, entry);
832 	}
833 
834 	umtxq_hash(key);
835 	return (0);
836 }
837 
838 /*
839  * Release key.
840  */
841 void
842 umtx_key_release(struct umtx_key *key)
843 {
844 	if (key->shared)
845 		vm_object_deallocate(key->info.shared.object);
846 }
847 
848 #ifdef COMPAT_FREEBSD10
849 /*
850  * Lock a umtx object.
851  */
852 static int
853 do_lock_umtx(struct thread *td, struct umtx *umtx, u_long id,
854     const struct timespec *timeout)
855 {
856 	struct umtx_abs_timeout timo;
857 	struct umtx_q *uq;
858 	u_long owner;
859 	u_long old;
860 	int error = 0;
861 
862 	uq = td->td_umtxq;
863 	if (timeout != NULL)
864 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
865 
866 	/*
867 	 * Care must be exercised when dealing with umtx structure. It
868 	 * can fault on any access.
869 	 */
870 	for (;;) {
871 		/*
872 		 * Try the uncontested case.  This should be done in userland.
873 		 */
874 		owner = casuword(&umtx->u_owner, UMTX_UNOWNED, id);
875 
876 		/* The acquire succeeded. */
877 		if (owner == UMTX_UNOWNED)
878 			return (0);
879 
880 		/* The address was invalid. */
881 		if (owner == -1)
882 			return (EFAULT);
883 
884 		/* If no one owns it but it is contested try to acquire it. */
885 		if (owner == UMTX_CONTESTED) {
886 			owner = casuword(&umtx->u_owner,
887 			    UMTX_CONTESTED, id | UMTX_CONTESTED);
888 
889 			if (owner == UMTX_CONTESTED)
890 				return (0);
891 
892 			/* The address was invalid. */
893 			if (owner == -1)
894 				return (EFAULT);
895 
896 			error = thread_check_susp(td, false);
897 			if (error != 0)
898 				break;
899 
900 			/* If this failed the lock has changed, restart. */
901 			continue;
902 		}
903 
904 		/*
905 		 * If we caught a signal, we have retried and now
906 		 * exit immediately.
907 		 */
908 		if (error != 0)
909 			break;
910 
911 		if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK,
912 			AUTO_SHARE, &uq->uq_key)) != 0)
913 			return (error);
914 
915 		umtxq_lock(&uq->uq_key);
916 		umtxq_busy(&uq->uq_key);
917 		umtxq_insert(uq);
918 		umtxq_unbusy(&uq->uq_key);
919 		umtxq_unlock(&uq->uq_key);
920 
921 		/*
922 		 * Set the contested bit so that a release in user space
923 		 * knows to use the system call for unlock.  If this fails
924 		 * either some one else has acquired the lock or it has been
925 		 * released.
926 		 */
927 		old = casuword(&umtx->u_owner, owner, owner | UMTX_CONTESTED);
928 
929 		/* The address was invalid. */
930 		if (old == -1) {
931 			umtxq_lock(&uq->uq_key);
932 			umtxq_remove(uq);
933 			umtxq_unlock(&uq->uq_key);
934 			umtx_key_release(&uq->uq_key);
935 			return (EFAULT);
936 		}
937 
938 		/*
939 		 * We set the contested bit, sleep. Otherwise the lock changed
940 		 * and we need to retry or we lost a race to the thread
941 		 * unlocking the umtx.
942 		 */
943 		umtxq_lock(&uq->uq_key);
944 		if (old == owner)
945 			error = umtxq_sleep(uq, "umtx", timeout == NULL ? NULL :
946 			    &timo);
947 		umtxq_remove(uq);
948 		umtxq_unlock(&uq->uq_key);
949 		umtx_key_release(&uq->uq_key);
950 
951 		if (error == 0)
952 			error = thread_check_susp(td, false);
953 	}
954 
955 	if (timeout == NULL) {
956 		/* Mutex locking is restarted if it is interrupted. */
957 		if (error == EINTR)
958 			error = ERESTART;
959 	} else {
960 		/* Timed-locking is not restarted. */
961 		if (error == ERESTART)
962 			error = EINTR;
963 	}
964 	return (error);
965 }
966 
967 /*
968  * Unlock a umtx object.
969  */
970 static int
971 do_unlock_umtx(struct thread *td, struct umtx *umtx, u_long id)
972 {
973 	struct umtx_key key;
974 	u_long owner;
975 	u_long old;
976 	int error;
977 	int count;
978 
979 	/*
980 	 * Make sure we own this mtx.
981 	 */
982 	owner = fuword(__DEVOLATILE(u_long *, &umtx->u_owner));
983 	if (owner == -1)
984 		return (EFAULT);
985 
986 	if ((owner & ~UMTX_CONTESTED) != id)
987 		return (EPERM);
988 
989 	/* This should be done in userland */
990 	if ((owner & UMTX_CONTESTED) == 0) {
991 		old = casuword(&umtx->u_owner, owner, UMTX_UNOWNED);
992 		if (old == -1)
993 			return (EFAULT);
994 		if (old == owner)
995 			return (0);
996 		owner = old;
997 	}
998 
999 	/* We should only ever be in here for contested locks */
1000 	if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1001 	    &key)) != 0)
1002 		return (error);
1003 
1004 	umtxq_lock(&key);
1005 	umtxq_busy(&key);
1006 	count = umtxq_count(&key);
1007 	umtxq_unlock(&key);
1008 
1009 	/*
1010 	 * When unlocking the umtx, it must be marked as unowned if
1011 	 * there is zero or one thread only waiting for it.
1012 	 * Otherwise, it must be marked as contested.
1013 	 */
1014 	old = casuword(&umtx->u_owner, owner,
1015 	    count <= 1 ? UMTX_UNOWNED : UMTX_CONTESTED);
1016 	umtxq_lock(&key);
1017 	umtxq_signal(&key,1);
1018 	umtxq_unbusy(&key);
1019 	umtxq_unlock(&key);
1020 	umtx_key_release(&key);
1021 	if (old == -1)
1022 		return (EFAULT);
1023 	if (old != owner)
1024 		return (EINVAL);
1025 	return (0);
1026 }
1027 
1028 #ifdef COMPAT_FREEBSD32
1029 
1030 /*
1031  * Lock a umtx object.
1032  */
1033 static int
1034 do_lock_umtx32(struct thread *td, uint32_t *m, uint32_t id,
1035 	const struct timespec *timeout)
1036 {
1037 	struct umtx_abs_timeout timo;
1038 	struct umtx_q *uq;
1039 	uint32_t owner;
1040 	uint32_t old;
1041 	int error = 0;
1042 
1043 	uq = td->td_umtxq;
1044 
1045 	if (timeout != NULL)
1046 		umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
1047 
1048 	/*
1049 	 * Care must be exercised when dealing with umtx structure. It
1050 	 * can fault on any access.
1051 	 */
1052 	for (;;) {
1053 		/*
1054 		 * Try the uncontested case.  This should be done in userland.
1055 		 */
1056 		owner = casuword32(m, UMUTEX_UNOWNED, id);
1057 
1058 		/* The acquire succeeded. */
1059 		if (owner == UMUTEX_UNOWNED)
1060 			return (0);
1061 
1062 		/* The address was invalid. */
1063 		if (owner == -1)
1064 			return (EFAULT);
1065 
1066 		/* If no one owns it but it is contested try to acquire it. */
1067 		if (owner == UMUTEX_CONTESTED) {
1068 			owner = casuword32(m,
1069 			    UMUTEX_CONTESTED, id | UMUTEX_CONTESTED);
1070 			if (owner == UMUTEX_CONTESTED)
1071 				return (0);
1072 
1073 			/* The address was invalid. */
1074 			if (owner == -1)
1075 				return (EFAULT);
1076 
1077 			error = thread_check_susp(td, false);
1078 			if (error != 0)
1079 				break;
1080 
1081 			/* If this failed the lock has changed, restart. */
1082 			continue;
1083 		}
1084 
1085 		/*
1086 		 * If we caught a signal, we have retried and now
1087 		 * exit immediately.
1088 		 */
1089 		if (error != 0)
1090 			return (error);
1091 
1092 		if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK,
1093 			AUTO_SHARE, &uq->uq_key)) != 0)
1094 			return (error);
1095 
1096 		umtxq_lock(&uq->uq_key);
1097 		umtxq_busy(&uq->uq_key);
1098 		umtxq_insert(uq);
1099 		umtxq_unbusy(&uq->uq_key);
1100 		umtxq_unlock(&uq->uq_key);
1101 
1102 		/*
1103 		 * Set the contested bit so that a release in user space
1104 		 * knows to use the system call for unlock.  If this fails
1105 		 * either some one else has acquired the lock or it has been
1106 		 * released.
1107 		 */
1108 		old = casuword32(m, owner, owner | UMUTEX_CONTESTED);
1109 
1110 		/* The address was invalid. */
1111 		if (old == -1) {
1112 			umtxq_lock(&uq->uq_key);
1113 			umtxq_remove(uq);
1114 			umtxq_unlock(&uq->uq_key);
1115 			umtx_key_release(&uq->uq_key);
1116 			return (EFAULT);
1117 		}
1118 
1119 		/*
1120 		 * We set the contested bit, sleep. Otherwise the lock changed
1121 		 * and we need to retry or we lost a race to the thread
1122 		 * unlocking the umtx.
1123 		 */
1124 		umtxq_lock(&uq->uq_key);
1125 		if (old == owner)
1126 			error = umtxq_sleep(uq, "umtx", timeout == NULL ?
1127 			    NULL : &timo);
1128 		umtxq_remove(uq);
1129 		umtxq_unlock(&uq->uq_key);
1130 		umtx_key_release(&uq->uq_key);
1131 
1132 		if (error == 0)
1133 			error = thread_check_susp(td, false);
1134 	}
1135 
1136 	if (timeout == NULL) {
1137 		/* Mutex locking is restarted if it is interrupted. */
1138 		if (error == EINTR)
1139 			error = ERESTART;
1140 	} else {
1141 		/* Timed-locking is not restarted. */
1142 		if (error == ERESTART)
1143 			error = EINTR;
1144 	}
1145 	return (error);
1146 }
1147 
1148 /*
1149  * Unlock a umtx object.
1150  */
1151 static int
1152 do_unlock_umtx32(struct thread *td, uint32_t *m, uint32_t id)
1153 {
1154 	struct umtx_key key;
1155 	uint32_t owner;
1156 	uint32_t old;
1157 	int error;
1158 	int count;
1159 
1160 	/*
1161 	 * Make sure we own this mtx.
1162 	 */
1163 	owner = fuword32(m);
1164 	if (owner == -1)
1165 		return (EFAULT);
1166 
1167 	if ((owner & ~UMUTEX_CONTESTED) != id)
1168 		return (EPERM);
1169 
1170 	/* This should be done in userland */
1171 	if ((owner & UMUTEX_CONTESTED) == 0) {
1172 		old = casuword32(m, owner, UMUTEX_UNOWNED);
1173 		if (old == -1)
1174 			return (EFAULT);
1175 		if (old == owner)
1176 			return (0);
1177 		owner = old;
1178 	}
1179 
1180 	/* We should only ever be in here for contested locks */
1181 	if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1182 		&key)) != 0)
1183 		return (error);
1184 
1185 	umtxq_lock(&key);
1186 	umtxq_busy(&key);
1187 	count = umtxq_count(&key);
1188 	umtxq_unlock(&key);
1189 
1190 	/*
1191 	 * When unlocking the umtx, it must be marked as unowned if
1192 	 * there is zero or one thread only waiting for it.
1193 	 * Otherwise, it must be marked as contested.
1194 	 */
1195 	old = casuword32(m, owner,
1196 		count <= 1 ? UMUTEX_UNOWNED : UMUTEX_CONTESTED);
1197 	umtxq_lock(&key);
1198 	umtxq_signal(&key,1);
1199 	umtxq_unbusy(&key);
1200 	umtxq_unlock(&key);
1201 	umtx_key_release(&key);
1202 	if (old == -1)
1203 		return (EFAULT);
1204 	if (old != owner)
1205 		return (EINVAL);
1206 	return (0);
1207 }
1208 #endif	/* COMPAT_FREEBSD32 */
1209 #endif	/* COMPAT_FREEBSD10 */
1210 
1211 /*
1212  * Fetch and compare value, sleep on the address if value is not changed.
1213  */
1214 static int
1215 do_wait(struct thread *td, void *addr, u_long id,
1216     struct _umtx_time *timeout, int compat32, int is_private)
1217 {
1218 	struct umtx_abs_timeout timo;
1219 	struct umtx_q *uq;
1220 	u_long tmp;
1221 	uint32_t tmp32;
1222 	int error = 0;
1223 
1224 	uq = td->td_umtxq;
1225 	if ((error = umtx_key_get(addr, TYPE_SIMPLE_WAIT,
1226 		is_private ? THREAD_SHARE : AUTO_SHARE, &uq->uq_key)) != 0)
1227 		return (error);
1228 
1229 	if (timeout != NULL)
1230 		umtx_abs_timeout_init2(&timo, timeout);
1231 
1232 	umtxq_lock(&uq->uq_key);
1233 	umtxq_insert(uq);
1234 	umtxq_unlock(&uq->uq_key);
1235 	if (compat32 == 0) {
1236 		error = fueword(addr, &tmp);
1237 		if (error != 0)
1238 			error = EFAULT;
1239 	} else {
1240 		error = fueword32(addr, &tmp32);
1241 		if (error == 0)
1242 			tmp = tmp32;
1243 		else
1244 			error = EFAULT;
1245 	}
1246 	umtxq_lock(&uq->uq_key);
1247 	if (error == 0) {
1248 		if (tmp == id)
1249 			error = umtxq_sleep(uq, "uwait", timeout == NULL ?
1250 			    NULL : &timo);
1251 		if ((uq->uq_flags & UQF_UMTXQ) == 0)
1252 			error = 0;
1253 		else
1254 			umtxq_remove(uq);
1255 	} else if ((uq->uq_flags & UQF_UMTXQ) != 0) {
1256 		umtxq_remove(uq);
1257 	}
1258 	umtxq_unlock(&uq->uq_key);
1259 	umtx_key_release(&uq->uq_key);
1260 	if (error == ERESTART)
1261 		error = EINTR;
1262 	return (error);
1263 }
1264 
1265 /*
1266  * Wake up threads sleeping on the specified address.
1267  */
1268 int
1269 kern_umtx_wake(struct thread *td, void *uaddr, int n_wake, int is_private)
1270 {
1271 	struct umtx_key key;
1272 	int ret;
1273 
1274 	if ((ret = umtx_key_get(uaddr, TYPE_SIMPLE_WAIT,
1275 	    is_private ? THREAD_SHARE : AUTO_SHARE, &key)) != 0)
1276 		return (ret);
1277 	umtxq_lock(&key);
1278 	umtxq_signal(&key, n_wake);
1279 	umtxq_unlock(&key);
1280 	umtx_key_release(&key);
1281 	return (0);
1282 }
1283 
1284 /*
1285  * Lock PTHREAD_PRIO_NONE protocol POSIX mutex.
1286  */
1287 static int
1288 do_lock_normal(struct thread *td, struct umutex *m, uint32_t flags,
1289     struct _umtx_time *timeout, int mode)
1290 {
1291 	struct umtx_abs_timeout timo;
1292 	struct umtx_q *uq;
1293 	uint32_t owner, old, id;
1294 	int error, rv;
1295 
1296 	id = td->td_tid;
1297 	uq = td->td_umtxq;
1298 	error = 0;
1299 	if (timeout != NULL)
1300 		umtx_abs_timeout_init2(&timo, timeout);
1301 
1302 	/*
1303 	 * Care must be exercised when dealing with umtx structure. It
1304 	 * can fault on any access.
1305 	 */
1306 	for (;;) {
1307 		rv = fueword32(&m->m_owner, &owner);
1308 		if (rv == -1)
1309 			return (EFAULT);
1310 		if (mode == _UMUTEX_WAIT) {
1311 			if (owner == UMUTEX_UNOWNED ||
1312 			    owner == UMUTEX_CONTESTED ||
1313 			    owner == UMUTEX_RB_OWNERDEAD ||
1314 			    owner == UMUTEX_RB_NOTRECOV)
1315 				return (0);
1316 		} else {
1317 			/*
1318 			 * Robust mutex terminated.  Kernel duty is to
1319 			 * return EOWNERDEAD to the userspace.  The
1320 			 * umutex.m_flags UMUTEX_NONCONSISTENT is set
1321 			 * by the common userspace code.
1322 			 */
1323 			if (owner == UMUTEX_RB_OWNERDEAD) {
1324 				rv = casueword32(&m->m_owner,
1325 				    UMUTEX_RB_OWNERDEAD, &owner,
1326 				    id | UMUTEX_CONTESTED);
1327 				if (rv == -1)
1328 					return (EFAULT);
1329 				if (rv == 0) {
1330 					MPASS(owner == UMUTEX_RB_OWNERDEAD);
1331 					return (EOWNERDEAD); /* success */
1332 				}
1333 				MPASS(rv == 1);
1334 				rv = thread_check_susp(td, false);
1335 				if (rv != 0)
1336 					return (rv);
1337 				continue;
1338 			}
1339 			if (owner == UMUTEX_RB_NOTRECOV)
1340 				return (ENOTRECOVERABLE);
1341 
1342 			/*
1343 			 * Try the uncontested case.  This should be
1344 			 * done in userland.
1345 			 */
1346 			rv = casueword32(&m->m_owner, UMUTEX_UNOWNED,
1347 			    &owner, id);
1348 			/* The address was invalid. */
1349 			if (rv == -1)
1350 				return (EFAULT);
1351 
1352 			/* The acquire succeeded. */
1353 			if (rv == 0) {
1354 				MPASS(owner == UMUTEX_UNOWNED);
1355 				return (0);
1356 			}
1357 
1358 			/*
1359 			 * If no one owns it but it is contested try
1360 			 * to acquire it.
1361 			 */
1362 			MPASS(rv == 1);
1363 			if (owner == UMUTEX_CONTESTED) {
1364 				rv = casueword32(&m->m_owner,
1365 				    UMUTEX_CONTESTED, &owner,
1366 				    id | UMUTEX_CONTESTED);
1367 				/* The address was invalid. */
1368 				if (rv == -1)
1369 					return (EFAULT);
1370 				if (rv == 0) {
1371 					MPASS(owner == UMUTEX_CONTESTED);
1372 					return (0);
1373 				}
1374 				if (rv == 1) {
1375 					rv = thread_check_susp(td, false);
1376 					if (rv != 0)
1377 						return (rv);
1378 				}
1379 
1380 				/*
1381 				 * If this failed the lock has
1382 				 * changed, restart.
1383 				 */
1384 				continue;
1385 			}
1386 
1387 			/* rv == 1 but not contested, likely store failure */
1388 			rv = thread_check_susp(td, false);
1389 			if (rv != 0)
1390 				return (rv);
1391 		}
1392 
1393 		if (mode == _UMUTEX_TRY)
1394 			return (EBUSY);
1395 
1396 		/*
1397 		 * If we caught a signal, we have retried and now
1398 		 * exit immediately.
1399 		 */
1400 		if (error != 0)
1401 			return (error);
1402 
1403 		if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX,
1404 		    GET_SHARE(flags), &uq->uq_key)) != 0)
1405 			return (error);
1406 
1407 		umtxq_lock(&uq->uq_key);
1408 		umtxq_busy(&uq->uq_key);
1409 		umtxq_insert(uq);
1410 		umtxq_unlock(&uq->uq_key);
1411 
1412 		/*
1413 		 * Set the contested bit so that a release in user space
1414 		 * knows to use the system call for unlock.  If this fails
1415 		 * either some one else has acquired the lock or it has been
1416 		 * released.
1417 		 */
1418 		rv = casueword32(&m->m_owner, owner, &old,
1419 		    owner | UMUTEX_CONTESTED);
1420 
1421 		/* The address was invalid or casueword failed to store. */
1422 		if (rv == -1 || rv == 1) {
1423 			umtxq_lock(&uq->uq_key);
1424 			umtxq_remove(uq);
1425 			umtxq_unbusy(&uq->uq_key);
1426 			umtxq_unlock(&uq->uq_key);
1427 			umtx_key_release(&uq->uq_key);
1428 			if (rv == -1)
1429 				return (EFAULT);
1430 			if (rv == 1) {
1431 				rv = thread_check_susp(td, false);
1432 				if (rv != 0)
1433 					return (rv);
1434 			}
1435 			continue;
1436 		}
1437 
1438 		/*
1439 		 * We set the contested bit, sleep. Otherwise the lock changed
1440 		 * and we need to retry or we lost a race to the thread
1441 		 * unlocking the umtx.
1442 		 */
1443 		umtxq_lock(&uq->uq_key);
1444 		umtxq_unbusy(&uq->uq_key);
1445 		MPASS(old == owner);
1446 		error = umtxq_sleep(uq, "umtxn", timeout == NULL ?
1447 		    NULL : &timo);
1448 		umtxq_remove(uq);
1449 		umtxq_unlock(&uq->uq_key);
1450 		umtx_key_release(&uq->uq_key);
1451 
1452 		if (error == 0)
1453 			error = thread_check_susp(td, false);
1454 	}
1455 
1456 	return (0);
1457 }
1458 
1459 /*
1460  * Unlock PTHREAD_PRIO_NONE protocol POSIX mutex.
1461  */
1462 static int
1463 do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1464 {
1465 	struct umtx_key key;
1466 	uint32_t owner, old, id, newlock;
1467 	int error, count;
1468 
1469 	id = td->td_tid;
1470 
1471 again:
1472 	/*
1473 	 * Make sure we own this mtx.
1474 	 */
1475 	error = fueword32(&m->m_owner, &owner);
1476 	if (error == -1)
1477 		return (EFAULT);
1478 
1479 	if ((owner & ~UMUTEX_CONTESTED) != id)
1480 		return (EPERM);
1481 
1482 	newlock = umtx_unlock_val(flags, rb);
1483 	if ((owner & UMUTEX_CONTESTED) == 0) {
1484 		error = casueword32(&m->m_owner, owner, &old, newlock);
1485 		if (error == -1)
1486 			return (EFAULT);
1487 		if (error == 1) {
1488 			error = thread_check_susp(td, false);
1489 			if (error != 0)
1490 				return (error);
1491 			goto again;
1492 		}
1493 		MPASS(old == owner);
1494 		return (0);
1495 	}
1496 
1497 	/* We should only ever be in here for contested locks */
1498 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1499 	    &key)) != 0)
1500 		return (error);
1501 
1502 	umtxq_lock(&key);
1503 	umtxq_busy(&key);
1504 	count = umtxq_count(&key);
1505 	umtxq_unlock(&key);
1506 
1507 	/*
1508 	 * When unlocking the umtx, it must be marked as unowned if
1509 	 * there is zero or one thread only waiting for it.
1510 	 * Otherwise, it must be marked as contested.
1511 	 */
1512 	if (count > 1)
1513 		newlock |= UMUTEX_CONTESTED;
1514 	error = casueword32(&m->m_owner, owner, &old, newlock);
1515 	umtxq_lock(&key);
1516 	umtxq_signal(&key, 1);
1517 	umtxq_unbusy(&key);
1518 	umtxq_unlock(&key);
1519 	umtx_key_release(&key);
1520 	if (error == -1)
1521 		return (EFAULT);
1522 	if (error == 1) {
1523 		if (old != owner)
1524 			return (EINVAL);
1525 		error = thread_check_susp(td, false);
1526 		if (error != 0)
1527 			return (error);
1528 		goto again;
1529 	}
1530 	return (0);
1531 }
1532 
1533 /*
1534  * Check if the mutex is available and wake up a waiter,
1535  * only for simple mutex.
1536  */
1537 static int
1538 do_wake_umutex(struct thread *td, struct umutex *m)
1539 {
1540 	struct umtx_key key;
1541 	uint32_t owner;
1542 	uint32_t flags;
1543 	int error;
1544 	int count;
1545 
1546 again:
1547 	error = fueword32(&m->m_owner, &owner);
1548 	if (error == -1)
1549 		return (EFAULT);
1550 
1551 	if ((owner & ~UMUTEX_CONTESTED) != 0 && owner != UMUTEX_RB_OWNERDEAD &&
1552 	    owner != UMUTEX_RB_NOTRECOV)
1553 		return (0);
1554 
1555 	error = fueword32(&m->m_flags, &flags);
1556 	if (error == -1)
1557 		return (EFAULT);
1558 
1559 	/* We should only ever be in here for contested locks */
1560 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1561 	    &key)) != 0)
1562 		return (error);
1563 
1564 	umtxq_lock(&key);
1565 	umtxq_busy(&key);
1566 	count = umtxq_count(&key);
1567 	umtxq_unlock(&key);
1568 
1569 	if (count <= 1 && owner != UMUTEX_RB_OWNERDEAD &&
1570 	    owner != UMUTEX_RB_NOTRECOV) {
1571 		error = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
1572 		    UMUTEX_UNOWNED);
1573 		if (error == -1) {
1574 			error = EFAULT;
1575 		} else if (error == 1) {
1576 			umtxq_lock(&key);
1577 			umtxq_unbusy(&key);
1578 			umtxq_unlock(&key);
1579 			umtx_key_release(&key);
1580 			error = thread_check_susp(td, false);
1581 			if (error != 0)
1582 				return (error);
1583 			goto again;
1584 		}
1585 	}
1586 
1587 	umtxq_lock(&key);
1588 	if (error == 0 && count != 0) {
1589 		MPASS((owner & ~UMUTEX_CONTESTED) == 0 ||
1590 		    owner == UMUTEX_RB_OWNERDEAD ||
1591 		    owner == UMUTEX_RB_NOTRECOV);
1592 		umtxq_signal(&key, 1);
1593 	}
1594 	umtxq_unbusy(&key);
1595 	umtxq_unlock(&key);
1596 	umtx_key_release(&key);
1597 	return (error);
1598 }
1599 
1600 /*
1601  * Check if the mutex has waiters and tries to fix contention bit.
1602  */
1603 static int
1604 do_wake2_umutex(struct thread *td, struct umutex *m, uint32_t flags)
1605 {
1606 	struct umtx_key key;
1607 	uint32_t owner, old;
1608 	int type;
1609 	int error;
1610 	int count;
1611 
1612 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT |
1613 	    UMUTEX_ROBUST)) {
1614 	case 0:
1615 	case UMUTEX_ROBUST:
1616 		type = TYPE_NORMAL_UMUTEX;
1617 		break;
1618 	case UMUTEX_PRIO_INHERIT:
1619 		type = TYPE_PI_UMUTEX;
1620 		break;
1621 	case (UMUTEX_PRIO_INHERIT | UMUTEX_ROBUST):
1622 		type = TYPE_PI_ROBUST_UMUTEX;
1623 		break;
1624 	case UMUTEX_PRIO_PROTECT:
1625 		type = TYPE_PP_UMUTEX;
1626 		break;
1627 	case (UMUTEX_PRIO_PROTECT | UMUTEX_ROBUST):
1628 		type = TYPE_PP_ROBUST_UMUTEX;
1629 		break;
1630 	default:
1631 		return (EINVAL);
1632 	}
1633 	if ((error = umtx_key_get(m, type, GET_SHARE(flags), &key)) != 0)
1634 		return (error);
1635 
1636 	owner = 0;
1637 	umtxq_lock(&key);
1638 	umtxq_busy(&key);
1639 	count = umtxq_count(&key);
1640 	umtxq_unlock(&key);
1641 
1642 	error = fueword32(&m->m_owner, &owner);
1643 	if (error == -1)
1644 		error = EFAULT;
1645 
1646 	/*
1647 	 * Only repair contention bit if there is a waiter, this means
1648 	 * the mutex is still being referenced by userland code,
1649 	 * otherwise don't update any memory.
1650 	 */
1651 	while (error == 0 && (owner & UMUTEX_CONTESTED) == 0 &&
1652 	    (count > 1 || (count == 1 && (owner & ~UMUTEX_CONTESTED) != 0))) {
1653 		error = casueword32(&m->m_owner, owner, &old,
1654 		    owner | UMUTEX_CONTESTED);
1655 		if (error == -1) {
1656 			error = EFAULT;
1657 			break;
1658 		}
1659 		if (error == 0) {
1660 			MPASS(old == owner);
1661 			break;
1662 		}
1663 		owner = old;
1664 		error = thread_check_susp(td, false);
1665 	}
1666 
1667 	umtxq_lock(&key);
1668 	if (error == EFAULT) {
1669 		umtxq_signal(&key, INT_MAX);
1670 	} else if (count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1671 	    owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1672 		umtxq_signal(&key, 1);
1673 	umtxq_unbusy(&key);
1674 	umtxq_unlock(&key);
1675 	umtx_key_release(&key);
1676 	return (error);
1677 }
1678 
1679 struct umtx_pi *
1680 umtx_pi_alloc(int flags)
1681 {
1682 	struct umtx_pi *pi;
1683 
1684 	pi = uma_zalloc(umtx_pi_zone, M_ZERO | flags);
1685 	TAILQ_INIT(&pi->pi_blocked);
1686 	atomic_add_int(&umtx_pi_allocated, 1);
1687 	return (pi);
1688 }
1689 
1690 void
1691 umtx_pi_free(struct umtx_pi *pi)
1692 {
1693 	uma_zfree(umtx_pi_zone, pi);
1694 	atomic_add_int(&umtx_pi_allocated, -1);
1695 }
1696 
1697 /*
1698  * Adjust the thread's position on a pi_state after its priority has been
1699  * changed.
1700  */
1701 static int
1702 umtx_pi_adjust_thread(struct umtx_pi *pi, struct thread *td)
1703 {
1704 	struct umtx_q *uq, *uq1, *uq2;
1705 	struct thread *td1;
1706 
1707 	mtx_assert(&umtx_lock, MA_OWNED);
1708 	if (pi == NULL)
1709 		return (0);
1710 
1711 	uq = td->td_umtxq;
1712 
1713 	/*
1714 	 * Check if the thread needs to be moved on the blocked chain.
1715 	 * It needs to be moved if either its priority is lower than
1716 	 * the previous thread or higher than the next thread.
1717 	 */
1718 	uq1 = TAILQ_PREV(uq, umtxq_head, uq_lockq);
1719 	uq2 = TAILQ_NEXT(uq, uq_lockq);
1720 	if ((uq1 != NULL && UPRI(td) < UPRI(uq1->uq_thread)) ||
1721 	    (uq2 != NULL && UPRI(td) > UPRI(uq2->uq_thread))) {
1722 		/*
1723 		 * Remove thread from blocked chain and determine where
1724 		 * it should be moved to.
1725 		 */
1726 		TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1727 		TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1728 			td1 = uq1->uq_thread;
1729 			MPASS(td1->td_proc->p_magic == P_MAGIC);
1730 			if (UPRI(td1) > UPRI(td))
1731 				break;
1732 		}
1733 
1734 		if (uq1 == NULL)
1735 			TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1736 		else
1737 			TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1738 	}
1739 	return (1);
1740 }
1741 
1742 static struct umtx_pi *
1743 umtx_pi_next(struct umtx_pi *pi)
1744 {
1745 	struct umtx_q *uq_owner;
1746 
1747 	if (pi->pi_owner == NULL)
1748 		return (NULL);
1749 	uq_owner = pi->pi_owner->td_umtxq;
1750 	if (uq_owner == NULL)
1751 		return (NULL);
1752 	return (uq_owner->uq_pi_blocked);
1753 }
1754 
1755 /*
1756  * Floyd's Cycle-Finding Algorithm.
1757  */
1758 static bool
1759 umtx_pi_check_loop(struct umtx_pi *pi)
1760 {
1761 	struct umtx_pi *pi1;	/* fast iterator */
1762 
1763 	mtx_assert(&umtx_lock, MA_OWNED);
1764 	if (pi == NULL)
1765 		return (false);
1766 	pi1 = pi;
1767 	for (;;) {
1768 		pi = umtx_pi_next(pi);
1769 		if (pi == NULL)
1770 			break;
1771 		pi1 = umtx_pi_next(pi1);
1772 		if (pi1 == NULL)
1773 			break;
1774 		pi1 = umtx_pi_next(pi1);
1775 		if (pi1 == NULL)
1776 			break;
1777 		if (pi == pi1)
1778 			return (true);
1779 	}
1780 	return (false);
1781 }
1782 
1783 /*
1784  * Propagate priority when a thread is blocked on POSIX
1785  * PI mutex.
1786  */
1787 static void
1788 umtx_propagate_priority(struct thread *td)
1789 {
1790 	struct umtx_q *uq;
1791 	struct umtx_pi *pi;
1792 	int pri;
1793 
1794 	mtx_assert(&umtx_lock, MA_OWNED);
1795 	pri = UPRI(td);
1796 	uq = td->td_umtxq;
1797 	pi = uq->uq_pi_blocked;
1798 	if (pi == NULL)
1799 		return;
1800 	if (umtx_pi_check_loop(pi))
1801 		return;
1802 
1803 	for (;;) {
1804 		td = pi->pi_owner;
1805 		if (td == NULL || td == curthread)
1806 			return;
1807 
1808 		MPASS(td->td_proc != NULL);
1809 		MPASS(td->td_proc->p_magic == P_MAGIC);
1810 
1811 		thread_lock(td);
1812 		if (td->td_lend_user_pri > pri)
1813 			sched_lend_user_prio(td, pri);
1814 		else {
1815 			thread_unlock(td);
1816 			break;
1817 		}
1818 		thread_unlock(td);
1819 
1820 		/*
1821 		 * Pick up the lock that td is blocked on.
1822 		 */
1823 		uq = td->td_umtxq;
1824 		pi = uq->uq_pi_blocked;
1825 		if (pi == NULL)
1826 			break;
1827 		/* Resort td on the list if needed. */
1828 		umtx_pi_adjust_thread(pi, td);
1829 	}
1830 }
1831 
1832 /*
1833  * Unpropagate priority for a PI mutex when a thread blocked on
1834  * it is interrupted by signal or resumed by others.
1835  */
1836 static void
1837 umtx_repropagate_priority(struct umtx_pi *pi)
1838 {
1839 	struct umtx_q *uq, *uq_owner;
1840 	struct umtx_pi *pi2;
1841 	int pri;
1842 
1843 	mtx_assert(&umtx_lock, MA_OWNED);
1844 
1845 	if (umtx_pi_check_loop(pi))
1846 		return;
1847 	while (pi != NULL && pi->pi_owner != NULL) {
1848 		pri = PRI_MAX;
1849 		uq_owner = pi->pi_owner->td_umtxq;
1850 
1851 		TAILQ_FOREACH(pi2, &uq_owner->uq_pi_contested, pi_link) {
1852 			uq = TAILQ_FIRST(&pi2->pi_blocked);
1853 			if (uq != NULL) {
1854 				if (pri > UPRI(uq->uq_thread))
1855 					pri = UPRI(uq->uq_thread);
1856 			}
1857 		}
1858 
1859 		if (pri > uq_owner->uq_inherited_pri)
1860 			pri = uq_owner->uq_inherited_pri;
1861 		thread_lock(pi->pi_owner);
1862 		sched_lend_user_prio(pi->pi_owner, pri);
1863 		thread_unlock(pi->pi_owner);
1864 		if ((pi = uq_owner->uq_pi_blocked) != NULL)
1865 			umtx_pi_adjust_thread(pi, uq_owner->uq_thread);
1866 	}
1867 }
1868 
1869 /*
1870  * Insert a PI mutex into owned list.
1871  */
1872 static void
1873 umtx_pi_setowner(struct umtx_pi *pi, struct thread *owner)
1874 {
1875 	struct umtx_q *uq_owner;
1876 
1877 	uq_owner = owner->td_umtxq;
1878 	mtx_assert(&umtx_lock, MA_OWNED);
1879 	MPASS(pi->pi_owner == NULL);
1880 	pi->pi_owner = owner;
1881 	TAILQ_INSERT_TAIL(&uq_owner->uq_pi_contested, pi, pi_link);
1882 }
1883 
1884 /*
1885  * Disown a PI mutex, and remove it from the owned list.
1886  */
1887 static void
1888 umtx_pi_disown(struct umtx_pi *pi)
1889 {
1890 
1891 	mtx_assert(&umtx_lock, MA_OWNED);
1892 	TAILQ_REMOVE(&pi->pi_owner->td_umtxq->uq_pi_contested, pi, pi_link);
1893 	pi->pi_owner = NULL;
1894 }
1895 
1896 /*
1897  * Claim ownership of a PI mutex.
1898  */
1899 int
1900 umtx_pi_claim(struct umtx_pi *pi, struct thread *owner)
1901 {
1902 	struct umtx_q *uq;
1903 	int pri;
1904 
1905 	mtx_lock(&umtx_lock);
1906 	if (pi->pi_owner == owner) {
1907 		mtx_unlock(&umtx_lock);
1908 		return (0);
1909 	}
1910 
1911 	if (pi->pi_owner != NULL) {
1912 		/*
1913 		 * userland may have already messed the mutex, sigh.
1914 		 */
1915 		mtx_unlock(&umtx_lock);
1916 		return (EPERM);
1917 	}
1918 	umtx_pi_setowner(pi, owner);
1919 	uq = TAILQ_FIRST(&pi->pi_blocked);
1920 	if (uq != NULL) {
1921 		pri = UPRI(uq->uq_thread);
1922 		thread_lock(owner);
1923 		if (pri < UPRI(owner))
1924 			sched_lend_user_prio(owner, pri);
1925 		thread_unlock(owner);
1926 	}
1927 	mtx_unlock(&umtx_lock);
1928 	return (0);
1929 }
1930 
1931 /*
1932  * Adjust a thread's order position in its blocked PI mutex,
1933  * this may result new priority propagating process.
1934  */
1935 void
1936 umtx_pi_adjust(struct thread *td, u_char oldpri)
1937 {
1938 	struct umtx_q *uq;
1939 	struct umtx_pi *pi;
1940 
1941 	uq = td->td_umtxq;
1942 	mtx_lock(&umtx_lock);
1943 	/*
1944 	 * Pick up the lock that td is blocked on.
1945 	 */
1946 	pi = uq->uq_pi_blocked;
1947 	if (pi != NULL) {
1948 		umtx_pi_adjust_thread(pi, td);
1949 		umtx_repropagate_priority(pi);
1950 	}
1951 	mtx_unlock(&umtx_lock);
1952 }
1953 
1954 /*
1955  * Sleep on a PI mutex.
1956  */
1957 int
1958 umtxq_sleep_pi(struct umtx_q *uq, struct umtx_pi *pi, uint32_t owner,
1959     const char *wmesg, struct umtx_abs_timeout *timo, bool shared)
1960 {
1961 	struct thread *td, *td1;
1962 	struct umtx_q *uq1;
1963 	int error, pri;
1964 #ifdef INVARIANTS
1965 	struct umtxq_chain *uc;
1966 
1967 	uc = umtxq_getchain(&pi->pi_key);
1968 #endif
1969 	error = 0;
1970 	td = uq->uq_thread;
1971 	KASSERT(td == curthread, ("inconsistent uq_thread"));
1972 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
1973 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));
1974 	umtxq_insert(uq);
1975 	mtx_lock(&umtx_lock);
1976 	if (pi->pi_owner == NULL) {
1977 		mtx_unlock(&umtx_lock);
1978 		td1 = tdfind(owner, shared ? -1 : td->td_proc->p_pid);
1979 		mtx_lock(&umtx_lock);
1980 		if (td1 != NULL) {
1981 			if (pi->pi_owner == NULL)
1982 				umtx_pi_setowner(pi, td1);
1983 			PROC_UNLOCK(td1->td_proc);
1984 		}
1985 	}
1986 
1987 	TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1988 		pri = UPRI(uq1->uq_thread);
1989 		if (pri > UPRI(td))
1990 			break;
1991 	}
1992 
1993 	if (uq1 != NULL)
1994 		TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1995 	else
1996 		TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1997 
1998 	uq->uq_pi_blocked = pi;
1999 	thread_lock(td);
2000 	td->td_flags |= TDF_UPIBLOCKED;
2001 	thread_unlock(td);
2002 	umtx_propagate_priority(td);
2003 	mtx_unlock(&umtx_lock);
2004 	umtxq_unbusy(&uq->uq_key);
2005 
2006 	error = umtxq_sleep(uq, wmesg, timo);
2007 	umtxq_remove(uq);
2008 
2009 	mtx_lock(&umtx_lock);
2010 	uq->uq_pi_blocked = NULL;
2011 	thread_lock(td);
2012 	td->td_flags &= ~TDF_UPIBLOCKED;
2013 	thread_unlock(td);
2014 	TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
2015 	umtx_repropagate_priority(pi);
2016 	mtx_unlock(&umtx_lock);
2017 	umtxq_unlock(&uq->uq_key);
2018 
2019 	return (error);
2020 }
2021 
2022 /*
2023  * Add reference count for a PI mutex.
2024  */
2025 void
2026 umtx_pi_ref(struct umtx_pi *pi)
2027 {
2028 
2029 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&pi->pi_key));
2030 	pi->pi_refcount++;
2031 }
2032 
2033 /*
2034  * Decrease reference count for a PI mutex, if the counter
2035  * is decreased to zero, its memory space is freed.
2036  */
2037 void
2038 umtx_pi_unref(struct umtx_pi *pi)
2039 {
2040 	struct umtxq_chain *uc;
2041 
2042 	uc = umtxq_getchain(&pi->pi_key);
2043 	UMTXQ_LOCKED_ASSERT(uc);
2044 	KASSERT(pi->pi_refcount > 0, ("invalid reference count"));
2045 	if (--pi->pi_refcount == 0) {
2046 		mtx_lock(&umtx_lock);
2047 		if (pi->pi_owner != NULL)
2048 			umtx_pi_disown(pi);
2049 		KASSERT(TAILQ_EMPTY(&pi->pi_blocked),
2050 			("blocked queue not empty"));
2051 		mtx_unlock(&umtx_lock);
2052 		TAILQ_REMOVE(&uc->uc_pi_list, pi, pi_hashlink);
2053 		umtx_pi_free(pi);
2054 	}
2055 }
2056 
2057 /*
2058  * Find a PI mutex in hash table.
2059  */
2060 struct umtx_pi *
2061 umtx_pi_lookup(struct umtx_key *key)
2062 {
2063 	struct umtxq_chain *uc;
2064 	struct umtx_pi *pi;
2065 
2066 	uc = umtxq_getchain(key);
2067 	UMTXQ_LOCKED_ASSERT(uc);
2068 
2069 	TAILQ_FOREACH(pi, &uc->uc_pi_list, pi_hashlink) {
2070 		if (umtx_key_match(&pi->pi_key, key)) {
2071 			return (pi);
2072 		}
2073 	}
2074 	return (NULL);
2075 }
2076 
2077 /*
2078  * Insert a PI mutex into hash table.
2079  */
2080 void
2081 umtx_pi_insert(struct umtx_pi *pi)
2082 {
2083 	struct umtxq_chain *uc;
2084 
2085 	uc = umtxq_getchain(&pi->pi_key);
2086 	UMTXQ_LOCKED_ASSERT(uc);
2087 	TAILQ_INSERT_TAIL(&uc->uc_pi_list, pi, pi_hashlink);
2088 }
2089 
2090 /*
2091  * Drop a PI mutex and wakeup a top waiter.
2092  */
2093 int
2094 umtx_pi_drop(struct thread *td, struct umtx_key *key, bool rb, int *count)
2095 {
2096 	struct umtx_q *uq_first, *uq_first2, *uq_me;
2097 	struct umtx_pi *pi, *pi2;
2098 	int pri;
2099 
2100 	UMTXQ_ASSERT_LOCKED_BUSY(key);
2101 	*count = umtxq_count_pi(key, &uq_first);
2102 	if (uq_first != NULL) {
2103 		mtx_lock(&umtx_lock);
2104 		pi = uq_first->uq_pi_blocked;
2105 		KASSERT(pi != NULL, ("pi == NULL?"));
2106 		if (pi->pi_owner != td && !(rb && pi->pi_owner == NULL)) {
2107 			mtx_unlock(&umtx_lock);
2108 			/* userland messed the mutex */
2109 			return (EPERM);
2110 		}
2111 		uq_me = td->td_umtxq;
2112 		if (pi->pi_owner == td)
2113 			umtx_pi_disown(pi);
2114 		/* get highest priority thread which is still sleeping. */
2115 		uq_first = TAILQ_FIRST(&pi->pi_blocked);
2116 		while (uq_first != NULL &&
2117 		    (uq_first->uq_flags & UQF_UMTXQ) == 0) {
2118 			uq_first = TAILQ_NEXT(uq_first, uq_lockq);
2119 		}
2120 		pri = PRI_MAX;
2121 		TAILQ_FOREACH(pi2, &uq_me->uq_pi_contested, pi_link) {
2122 			uq_first2 = TAILQ_FIRST(&pi2->pi_blocked);
2123 			if (uq_first2 != NULL) {
2124 				if (pri > UPRI(uq_first2->uq_thread))
2125 					pri = UPRI(uq_first2->uq_thread);
2126 			}
2127 		}
2128 		thread_lock(td);
2129 		sched_lend_user_prio(td, pri);
2130 		thread_unlock(td);
2131 		mtx_unlock(&umtx_lock);
2132 		if (uq_first)
2133 			umtxq_signal_thread(uq_first);
2134 	} else {
2135 		pi = umtx_pi_lookup(key);
2136 		/*
2137 		 * A umtx_pi can exist if a signal or timeout removed the
2138 		 * last waiter from the umtxq, but there is still
2139 		 * a thread in do_lock_pi() holding the umtx_pi.
2140 		 */
2141 		if (pi != NULL) {
2142 			/*
2143 			 * The umtx_pi can be unowned, such as when a thread
2144 			 * has just entered do_lock_pi(), allocated the
2145 			 * umtx_pi, and unlocked the umtxq.
2146 			 * If the current thread owns it, it must disown it.
2147 			 */
2148 			mtx_lock(&umtx_lock);
2149 			if (pi->pi_owner == td)
2150 				umtx_pi_disown(pi);
2151 			mtx_unlock(&umtx_lock);
2152 		}
2153 	}
2154 	return (0);
2155 }
2156 
2157 /*
2158  * Lock a PI mutex.
2159  */
2160 static int
2161 do_lock_pi(struct thread *td, struct umutex *m, uint32_t flags,
2162     struct _umtx_time *timeout, int try)
2163 {
2164 	struct umtx_abs_timeout timo;
2165 	struct umtx_q *uq;
2166 	struct umtx_pi *pi, *new_pi;
2167 	uint32_t id, old_owner, owner, old;
2168 	int error, rv;
2169 
2170 	id = td->td_tid;
2171 	uq = td->td_umtxq;
2172 
2173 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2174 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2175 	    &uq->uq_key)) != 0)
2176 		return (error);
2177 
2178 	if (timeout != NULL)
2179 		umtx_abs_timeout_init2(&timo, timeout);
2180 
2181 	umtxq_lock(&uq->uq_key);
2182 	pi = umtx_pi_lookup(&uq->uq_key);
2183 	if (pi == NULL) {
2184 		new_pi = umtx_pi_alloc(M_NOWAIT);
2185 		if (new_pi == NULL) {
2186 			umtxq_unlock(&uq->uq_key);
2187 			new_pi = umtx_pi_alloc(M_WAITOK);
2188 			umtxq_lock(&uq->uq_key);
2189 			pi = umtx_pi_lookup(&uq->uq_key);
2190 			if (pi != NULL) {
2191 				umtx_pi_free(new_pi);
2192 				new_pi = NULL;
2193 			}
2194 		}
2195 		if (new_pi != NULL) {
2196 			new_pi->pi_key = uq->uq_key;
2197 			umtx_pi_insert(new_pi);
2198 			pi = new_pi;
2199 		}
2200 	}
2201 	umtx_pi_ref(pi);
2202 	umtxq_unlock(&uq->uq_key);
2203 
2204 	/*
2205 	 * Care must be exercised when dealing with umtx structure.  It
2206 	 * can fault on any access.
2207 	 */
2208 	for (;;) {
2209 		/*
2210 		 * Try the uncontested case.  This should be done in userland.
2211 		 */
2212 		rv = casueword32(&m->m_owner, UMUTEX_UNOWNED, &owner, id);
2213 		/* The address was invalid. */
2214 		if (rv == -1) {
2215 			error = EFAULT;
2216 			break;
2217 		}
2218 		/* The acquire succeeded. */
2219 		if (rv == 0) {
2220 			MPASS(owner == UMUTEX_UNOWNED);
2221 			error = 0;
2222 			break;
2223 		}
2224 
2225 		if (owner == UMUTEX_RB_NOTRECOV) {
2226 			error = ENOTRECOVERABLE;
2227 			break;
2228 		}
2229 
2230 		/*
2231 		 * Avoid overwriting a possible error from sleep due
2232 		 * to the pending signal with suspension check result.
2233 		 */
2234 		if (error == 0) {
2235 			error = thread_check_susp(td, true);
2236 			if (error != 0)
2237 				break;
2238 		}
2239 
2240 		/* If no one owns it but it is contested try to acquire it. */
2241 		if (owner == UMUTEX_CONTESTED || owner == UMUTEX_RB_OWNERDEAD) {
2242 			old_owner = owner;
2243 			rv = casueword32(&m->m_owner, owner, &owner,
2244 			    id | UMUTEX_CONTESTED);
2245 			/* The address was invalid. */
2246 			if (rv == -1) {
2247 				error = EFAULT;
2248 				break;
2249 			}
2250 			if (rv == 1) {
2251 				if (error == 0) {
2252 					error = thread_check_susp(td, true);
2253 					if (error != 0)
2254 						break;
2255 				}
2256 
2257 				/*
2258 				 * If this failed the lock could
2259 				 * changed, restart.
2260 				 */
2261 				continue;
2262 			}
2263 
2264 			MPASS(rv == 0);
2265 			MPASS(owner == old_owner);
2266 			umtxq_lock(&uq->uq_key);
2267 			umtxq_busy(&uq->uq_key);
2268 			error = umtx_pi_claim(pi, td);
2269 			umtxq_unbusy(&uq->uq_key);
2270 			umtxq_unlock(&uq->uq_key);
2271 			if (error != 0) {
2272 				/*
2273 				 * Since we're going to return an
2274 				 * error, restore the m_owner to its
2275 				 * previous, unowned state to avoid
2276 				 * compounding the problem.
2277 				 */
2278 				(void)casuword32(&m->m_owner,
2279 				    id | UMUTEX_CONTESTED, old_owner);
2280 			}
2281 			if (error == 0 && old_owner == UMUTEX_RB_OWNERDEAD)
2282 				error = EOWNERDEAD;
2283 			break;
2284 		}
2285 
2286 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2287 			error = EDEADLK;
2288 			break;
2289 		}
2290 
2291 		if (try != 0) {
2292 			error = EBUSY;
2293 			break;
2294 		}
2295 
2296 		/*
2297 		 * If we caught a signal, we have retried and now
2298 		 * exit immediately.
2299 		 */
2300 		if (error != 0)
2301 			break;
2302 
2303 		umtxq_lock(&uq->uq_key);
2304 		umtxq_busy(&uq->uq_key);
2305 		umtxq_unlock(&uq->uq_key);
2306 
2307 		/*
2308 		 * Set the contested bit so that a release in user space
2309 		 * knows to use the system call for unlock.  If this fails
2310 		 * either some one else has acquired the lock or it has been
2311 		 * released.
2312 		 */
2313 		rv = casueword32(&m->m_owner, owner, &old, owner |
2314 		    UMUTEX_CONTESTED);
2315 
2316 		/* The address was invalid. */
2317 		if (rv == -1) {
2318 			umtxq_unbusy_unlocked(&uq->uq_key);
2319 			error = EFAULT;
2320 			break;
2321 		}
2322 		if (rv == 1) {
2323 			umtxq_unbusy_unlocked(&uq->uq_key);
2324 			error = thread_check_susp(td, true);
2325 			if (error != 0)
2326 				break;
2327 
2328 			/*
2329 			 * The lock changed and we need to retry or we
2330 			 * lost a race to the thread unlocking the
2331 			 * umtx.  Note that the UMUTEX_RB_OWNERDEAD
2332 			 * value for owner is impossible there.
2333 			 */
2334 			continue;
2335 		}
2336 
2337 		umtxq_lock(&uq->uq_key);
2338 
2339 		/* We set the contested bit, sleep. */
2340 		MPASS(old == owner);
2341 		error = umtxq_sleep_pi(uq, pi, owner & ~UMUTEX_CONTESTED,
2342 		    "umtxpi", timeout == NULL ? NULL : &timo,
2343 		    (flags & USYNC_PROCESS_SHARED) != 0);
2344 		if (error != 0)
2345 			continue;
2346 
2347 		error = thread_check_susp(td, false);
2348 		if (error != 0)
2349 			break;
2350 	}
2351 
2352 	umtxq_lock(&uq->uq_key);
2353 	umtx_pi_unref(pi);
2354 	umtxq_unlock(&uq->uq_key);
2355 
2356 	umtx_key_release(&uq->uq_key);
2357 	return (error);
2358 }
2359 
2360 /*
2361  * Unlock a PI mutex.
2362  */
2363 static int
2364 do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2365 {
2366 	struct umtx_key key;
2367 	uint32_t id, new_owner, old, owner;
2368 	int count, error;
2369 
2370 	id = td->td_tid;
2371 
2372 usrloop:
2373 	/*
2374 	 * Make sure we own this mtx.
2375 	 */
2376 	error = fueword32(&m->m_owner, &owner);
2377 	if (error == -1)
2378 		return (EFAULT);
2379 
2380 	if ((owner & ~UMUTEX_CONTESTED) != id)
2381 		return (EPERM);
2382 
2383 	new_owner = umtx_unlock_val(flags, rb);
2384 
2385 	/* This should be done in userland */
2386 	if ((owner & UMUTEX_CONTESTED) == 0) {
2387 		error = casueword32(&m->m_owner, owner, &old, new_owner);
2388 		if (error == -1)
2389 			return (EFAULT);
2390 		if (error == 1) {
2391 			error = thread_check_susp(td, true);
2392 			if (error != 0)
2393 				return (error);
2394 			goto usrloop;
2395 		}
2396 		if (old == owner)
2397 			return (0);
2398 		owner = old;
2399 	}
2400 
2401 	/* We should only ever be in here for contested locks */
2402 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2403 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2404 	    &key)) != 0)
2405 		return (error);
2406 
2407 	umtxq_lock(&key);
2408 	umtxq_busy(&key);
2409 	error = umtx_pi_drop(td, &key, rb, &count);
2410 	if (error != 0) {
2411 		umtxq_unbusy(&key);
2412 		umtxq_unlock(&key);
2413 		umtx_key_release(&key);
2414 		/* userland messed the mutex */
2415 		return (error);
2416 	}
2417 	umtxq_unlock(&key);
2418 
2419 	/*
2420 	 * When unlocking the umtx, it must be marked as unowned if
2421 	 * there is zero or one thread only waiting for it.
2422 	 * Otherwise, it must be marked as contested.
2423 	 */
2424 
2425 	if (count > 1)
2426 		new_owner |= UMUTEX_CONTESTED;
2427 again:
2428 	error = casueword32(&m->m_owner, owner, &old, new_owner);
2429 	if (error == 1) {
2430 		error = thread_check_susp(td, false);
2431 		if (error == 0)
2432 			goto again;
2433 	}
2434 	umtxq_unbusy_unlocked(&key);
2435 	umtx_key_release(&key);
2436 	if (error == -1)
2437 		return (EFAULT);
2438 	if (error == 0 && old != owner)
2439 		return (EINVAL);
2440 	return (error);
2441 }
2442 
2443 /*
2444  * Lock a PP mutex.
2445  */
2446 static int
2447 do_lock_pp(struct thread *td, struct umutex *m, uint32_t flags,
2448     struct _umtx_time *timeout, int try)
2449 {
2450 	struct umtx_abs_timeout timo;
2451 	struct umtx_q *uq, *uq2;
2452 	struct umtx_pi *pi;
2453 	uint32_t ceiling;
2454 	uint32_t owner, id;
2455 	int error, pri, old_inherited_pri, su, rv;
2456 
2457 	id = td->td_tid;
2458 	uq = td->td_umtxq;
2459 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2460 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2461 	    &uq->uq_key)) != 0)
2462 		return (error);
2463 
2464 	if (timeout != NULL)
2465 		umtx_abs_timeout_init2(&timo, timeout);
2466 
2467 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2468 	for (;;) {
2469 		old_inherited_pri = uq->uq_inherited_pri;
2470 		umtxq_lock(&uq->uq_key);
2471 		umtxq_busy(&uq->uq_key);
2472 		umtxq_unlock(&uq->uq_key);
2473 
2474 		rv = fueword32(&m->m_ceilings[0], &ceiling);
2475 		if (rv == -1) {
2476 			error = EFAULT;
2477 			goto out;
2478 		}
2479 		ceiling = RTP_PRIO_MAX - ceiling;
2480 		if (ceiling > RTP_PRIO_MAX) {
2481 			error = EINVAL;
2482 			goto out;
2483 		}
2484 
2485 		mtx_lock(&umtx_lock);
2486 		if (UPRI(td) < PRI_MIN_REALTIME + ceiling) {
2487 			mtx_unlock(&umtx_lock);
2488 			error = EINVAL;
2489 			goto out;
2490 		}
2491 		if (su && PRI_MIN_REALTIME + ceiling < uq->uq_inherited_pri) {
2492 			uq->uq_inherited_pri = PRI_MIN_REALTIME + ceiling;
2493 			thread_lock(td);
2494 			if (uq->uq_inherited_pri < UPRI(td))
2495 				sched_lend_user_prio(td, uq->uq_inherited_pri);
2496 			thread_unlock(td);
2497 		}
2498 		mtx_unlock(&umtx_lock);
2499 
2500 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2501 		    id | UMUTEX_CONTESTED);
2502 		/* The address was invalid. */
2503 		if (rv == -1) {
2504 			error = EFAULT;
2505 			break;
2506 		}
2507 		if (rv == 0) {
2508 			MPASS(owner == UMUTEX_CONTESTED);
2509 			error = 0;
2510 			break;
2511 		}
2512 		/* rv == 1 */
2513 		if (owner == UMUTEX_RB_OWNERDEAD) {
2514 			rv = casueword32(&m->m_owner, UMUTEX_RB_OWNERDEAD,
2515 			    &owner, id | UMUTEX_CONTESTED);
2516 			if (rv == -1) {
2517 				error = EFAULT;
2518 				break;
2519 			}
2520 			if (rv == 0) {
2521 				MPASS(owner == UMUTEX_RB_OWNERDEAD);
2522 				error = EOWNERDEAD; /* success */
2523 				break;
2524 			}
2525 
2526 			/*
2527 			 *  rv == 1, only check for suspension if we
2528 			 *  did not already catched a signal.  If we
2529 			 *  get an error from the check, the same
2530 			 *  condition is checked by the umtxq_sleep()
2531 			 *  call below, so we should obliterate the
2532 			 *  error to not skip the last loop iteration.
2533 			 */
2534 			if (error == 0) {
2535 				error = thread_check_susp(td, false);
2536 				if (error == 0) {
2537 					if (try != 0)
2538 						error = EBUSY;
2539 					else
2540 						continue;
2541 				}
2542 				error = 0;
2543 			}
2544 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2545 			error = ENOTRECOVERABLE;
2546 		}
2547 
2548 		if (try != 0)
2549 			error = EBUSY;
2550 
2551 		/*
2552 		 * If we caught a signal, we have retried and now
2553 		 * exit immediately.
2554 		 */
2555 		if (error != 0)
2556 			break;
2557 
2558 		umtxq_lock(&uq->uq_key);
2559 		umtxq_insert(uq);
2560 		umtxq_unbusy(&uq->uq_key);
2561 		error = umtxq_sleep(uq, "umtxpp", timeout == NULL ?
2562 		    NULL : &timo);
2563 		umtxq_remove(uq);
2564 		umtxq_unlock(&uq->uq_key);
2565 
2566 		mtx_lock(&umtx_lock);
2567 		uq->uq_inherited_pri = old_inherited_pri;
2568 		pri = PRI_MAX;
2569 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2570 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2571 			if (uq2 != NULL) {
2572 				if (pri > UPRI(uq2->uq_thread))
2573 					pri = UPRI(uq2->uq_thread);
2574 			}
2575 		}
2576 		if (pri > uq->uq_inherited_pri)
2577 			pri = uq->uq_inherited_pri;
2578 		thread_lock(td);
2579 		sched_lend_user_prio(td, pri);
2580 		thread_unlock(td);
2581 		mtx_unlock(&umtx_lock);
2582 	}
2583 
2584 	if (error != 0 && error != EOWNERDEAD) {
2585 		mtx_lock(&umtx_lock);
2586 		uq->uq_inherited_pri = old_inherited_pri;
2587 		pri = PRI_MAX;
2588 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2589 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2590 			if (uq2 != NULL) {
2591 				if (pri > UPRI(uq2->uq_thread))
2592 					pri = UPRI(uq2->uq_thread);
2593 			}
2594 		}
2595 		if (pri > uq->uq_inherited_pri)
2596 			pri = uq->uq_inherited_pri;
2597 		thread_lock(td);
2598 		sched_lend_user_prio(td, pri);
2599 		thread_unlock(td);
2600 		mtx_unlock(&umtx_lock);
2601 	}
2602 
2603 out:
2604 	umtxq_unbusy_unlocked(&uq->uq_key);
2605 	umtx_key_release(&uq->uq_key);
2606 	return (error);
2607 }
2608 
2609 /*
2610  * Unlock a PP mutex.
2611  */
2612 static int
2613 do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2614 {
2615 	struct umtx_key key;
2616 	struct umtx_q *uq, *uq2;
2617 	struct umtx_pi *pi;
2618 	uint32_t id, owner, rceiling;
2619 	int error, pri, new_inherited_pri, su;
2620 
2621 	id = td->td_tid;
2622 	uq = td->td_umtxq;
2623 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2624 
2625 	/*
2626 	 * Make sure we own this mtx.
2627 	 */
2628 	error = fueword32(&m->m_owner, &owner);
2629 	if (error == -1)
2630 		return (EFAULT);
2631 
2632 	if ((owner & ~UMUTEX_CONTESTED) != id)
2633 		return (EPERM);
2634 
2635 	error = copyin(&m->m_ceilings[1], &rceiling, sizeof(uint32_t));
2636 	if (error != 0)
2637 		return (error);
2638 
2639 	if (rceiling == -1)
2640 		new_inherited_pri = PRI_MAX;
2641 	else {
2642 		rceiling = RTP_PRIO_MAX - rceiling;
2643 		if (rceiling > RTP_PRIO_MAX)
2644 			return (EINVAL);
2645 		new_inherited_pri = PRI_MIN_REALTIME + rceiling;
2646 	}
2647 
2648 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2649 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2650 	    &key)) != 0)
2651 		return (error);
2652 	umtxq_lock(&key);
2653 	umtxq_busy(&key);
2654 	umtxq_unlock(&key);
2655 	/*
2656 	 * For priority protected mutex, always set unlocked state
2657 	 * to UMUTEX_CONTESTED, so that userland always enters kernel
2658 	 * to lock the mutex, it is necessary because thread priority
2659 	 * has to be adjusted for such mutex.
2660 	 */
2661 	error = suword32(&m->m_owner, umtx_unlock_val(flags, rb) |
2662 	    UMUTEX_CONTESTED);
2663 
2664 	umtxq_lock(&key);
2665 	if (error == 0)
2666 		umtxq_signal(&key, 1);
2667 	umtxq_unbusy(&key);
2668 	umtxq_unlock(&key);
2669 
2670 	if (error == -1)
2671 		error = EFAULT;
2672 	else {
2673 		mtx_lock(&umtx_lock);
2674 		if (su != 0)
2675 			uq->uq_inherited_pri = new_inherited_pri;
2676 		pri = PRI_MAX;
2677 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2678 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2679 			if (uq2 != NULL) {
2680 				if (pri > UPRI(uq2->uq_thread))
2681 					pri = UPRI(uq2->uq_thread);
2682 			}
2683 		}
2684 		if (pri > uq->uq_inherited_pri)
2685 			pri = uq->uq_inherited_pri;
2686 		thread_lock(td);
2687 		sched_lend_user_prio(td, pri);
2688 		thread_unlock(td);
2689 		mtx_unlock(&umtx_lock);
2690 	}
2691 	umtx_key_release(&key);
2692 	return (error);
2693 }
2694 
2695 static int
2696 do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
2697     uint32_t *old_ceiling)
2698 {
2699 	struct umtx_q *uq;
2700 	uint32_t flags, id, owner, save_ceiling;
2701 	int error, rv, rv1;
2702 
2703 	error = fueword32(&m->m_flags, &flags);
2704 	if (error == -1)
2705 		return (EFAULT);
2706 	if ((flags & UMUTEX_PRIO_PROTECT) == 0)
2707 		return (EINVAL);
2708 	if (ceiling > RTP_PRIO_MAX)
2709 		return (EINVAL);
2710 	id = td->td_tid;
2711 	uq = td->td_umtxq;
2712 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2713 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2714 	    &uq->uq_key)) != 0)
2715 		return (error);
2716 	for (;;) {
2717 		umtxq_lock(&uq->uq_key);
2718 		umtxq_busy(&uq->uq_key);
2719 		umtxq_unlock(&uq->uq_key);
2720 
2721 		rv = fueword32(&m->m_ceilings[0], &save_ceiling);
2722 		if (rv == -1) {
2723 			error = EFAULT;
2724 			break;
2725 		}
2726 
2727 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2728 		    id | UMUTEX_CONTESTED);
2729 		if (rv == -1) {
2730 			error = EFAULT;
2731 			break;
2732 		}
2733 
2734 		if (rv == 0) {
2735 			MPASS(owner == UMUTEX_CONTESTED);
2736 			rv = suword32(&m->m_ceilings[0], ceiling);
2737 			rv1 = suword32(&m->m_owner, UMUTEX_CONTESTED);
2738 			error = (rv == 0 && rv1 == 0) ? 0: EFAULT;
2739 			break;
2740 		}
2741 
2742 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2743 			rv = suword32(&m->m_ceilings[0], ceiling);
2744 			error = rv == 0 ? 0 : EFAULT;
2745 			break;
2746 		}
2747 
2748 		if (owner == UMUTEX_RB_OWNERDEAD) {
2749 			error = EOWNERDEAD;
2750 			break;
2751 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2752 			error = ENOTRECOVERABLE;
2753 			break;
2754 		}
2755 
2756 		/*
2757 		 * If we caught a signal, we have retried and now
2758 		 * exit immediately.
2759 		 */
2760 		if (error != 0)
2761 			break;
2762 
2763 		/*
2764 		 * We set the contested bit, sleep. Otherwise the lock changed
2765 		 * and we need to retry or we lost a race to the thread
2766 		 * unlocking the umtx.
2767 		 */
2768 		umtxq_lock(&uq->uq_key);
2769 		umtxq_insert(uq);
2770 		umtxq_unbusy(&uq->uq_key);
2771 		error = umtxq_sleep(uq, "umtxpp", NULL);
2772 		umtxq_remove(uq);
2773 		umtxq_unlock(&uq->uq_key);
2774 	}
2775 	umtxq_lock(&uq->uq_key);
2776 	if (error == 0)
2777 		umtxq_signal(&uq->uq_key, INT_MAX);
2778 	umtxq_unbusy(&uq->uq_key);
2779 	umtxq_unlock(&uq->uq_key);
2780 	umtx_key_release(&uq->uq_key);
2781 	if (error == 0 && old_ceiling != NULL) {
2782 		rv = suword32(old_ceiling, save_ceiling);
2783 		error = rv == 0 ? 0 : EFAULT;
2784 	}
2785 	return (error);
2786 }
2787 
2788 /*
2789  * Lock a userland POSIX mutex.
2790  */
2791 static int
2792 do_lock_umutex(struct thread *td, struct umutex *m,
2793     struct _umtx_time *timeout, int mode)
2794 {
2795 	uint32_t flags;
2796 	int error;
2797 
2798 	error = fueword32(&m->m_flags, &flags);
2799 	if (error == -1)
2800 		return (EFAULT);
2801 
2802 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2803 	case 0:
2804 		error = do_lock_normal(td, m, flags, timeout, mode);
2805 		break;
2806 	case UMUTEX_PRIO_INHERIT:
2807 		error = do_lock_pi(td, m, flags, timeout, mode);
2808 		break;
2809 	case UMUTEX_PRIO_PROTECT:
2810 		error = do_lock_pp(td, m, flags, timeout, mode);
2811 		break;
2812 	default:
2813 		return (EINVAL);
2814 	}
2815 	if (timeout == NULL) {
2816 		if (error == EINTR && mode != _UMUTEX_WAIT)
2817 			error = ERESTART;
2818 	} else {
2819 		/* Timed-locking is not restarted. */
2820 		if (error == ERESTART)
2821 			error = EINTR;
2822 	}
2823 	return (error);
2824 }
2825 
2826 /*
2827  * Unlock a userland POSIX mutex.
2828  */
2829 static int
2830 do_unlock_umutex(struct thread *td, struct umutex *m, bool rb)
2831 {
2832 	uint32_t flags;
2833 	int error;
2834 
2835 	error = fueword32(&m->m_flags, &flags);
2836 	if (error == -1)
2837 		return (EFAULT);
2838 
2839 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2840 	case 0:
2841 		return (do_unlock_normal(td, m, flags, rb));
2842 	case UMUTEX_PRIO_INHERIT:
2843 		return (do_unlock_pi(td, m, flags, rb));
2844 	case UMUTEX_PRIO_PROTECT:
2845 		return (do_unlock_pp(td, m, flags, rb));
2846 	}
2847 
2848 	return (EINVAL);
2849 }
2850 
2851 static int
2852 do_cv_wait(struct thread *td, struct ucond *cv, struct umutex *m,
2853     struct timespec *timeout, u_long wflags)
2854 {
2855 	struct umtx_abs_timeout timo;
2856 	struct umtx_q *uq;
2857 	uint32_t flags, clockid, hasw;
2858 	int error;
2859 
2860 	uq = td->td_umtxq;
2861 	error = fueword32(&cv->c_flags, &flags);
2862 	if (error == -1)
2863 		return (EFAULT);
2864 	error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &uq->uq_key);
2865 	if (error != 0)
2866 		return (error);
2867 
2868 	if ((wflags & CVWAIT_CLOCKID) != 0) {
2869 		error = fueword32(&cv->c_clockid, &clockid);
2870 		if (error == -1) {
2871 			umtx_key_release(&uq->uq_key);
2872 			return (EFAULT);
2873 		}
2874 		if (clockid < CLOCK_REALTIME ||
2875 		    clockid >= CLOCK_THREAD_CPUTIME_ID) {
2876 			/* hmm, only HW clock id will work. */
2877 			umtx_key_release(&uq->uq_key);
2878 			return (EINVAL);
2879 		}
2880 	} else {
2881 		clockid = CLOCK_REALTIME;
2882 	}
2883 
2884 	umtxq_lock(&uq->uq_key);
2885 	umtxq_busy(&uq->uq_key);
2886 	umtxq_insert(uq);
2887 	umtxq_unlock(&uq->uq_key);
2888 
2889 	/*
2890 	 * Set c_has_waiters to 1 before releasing user mutex, also
2891 	 * don't modify cache line when unnecessary.
2892 	 */
2893 	error = fueword32(&cv->c_has_waiters, &hasw);
2894 	if (error == 0 && hasw == 0)
2895 		suword32(&cv->c_has_waiters, 1);
2896 
2897 	umtxq_unbusy_unlocked(&uq->uq_key);
2898 
2899 	error = do_unlock_umutex(td, m, false);
2900 
2901 	if (timeout != NULL)
2902 		umtx_abs_timeout_init(&timo, clockid,
2903 		    (wflags & CVWAIT_ABSTIME) != 0, timeout);
2904 
2905 	umtxq_lock(&uq->uq_key);
2906 	if (error == 0) {
2907 		error = umtxq_sleep(uq, "ucond", timeout == NULL ?
2908 		    NULL : &timo);
2909 	}
2910 
2911 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
2912 		error = 0;
2913 	else {
2914 		/*
2915 		 * This must be timeout,interrupted by signal or
2916 		 * surprious wakeup, clear c_has_waiter flag when
2917 		 * necessary.
2918 		 */
2919 		umtxq_busy(&uq->uq_key);
2920 		if ((uq->uq_flags & UQF_UMTXQ) != 0) {
2921 			int oldlen = uq->uq_cur_queue->length;
2922 			umtxq_remove(uq);
2923 			if (oldlen == 1) {
2924 				umtxq_unlock(&uq->uq_key);
2925 				suword32(&cv->c_has_waiters, 0);
2926 				umtxq_lock(&uq->uq_key);
2927 			}
2928 		}
2929 		umtxq_unbusy(&uq->uq_key);
2930 		if (error == ERESTART)
2931 			error = EINTR;
2932 	}
2933 
2934 	umtxq_unlock(&uq->uq_key);
2935 	umtx_key_release(&uq->uq_key);
2936 	return (error);
2937 }
2938 
2939 /*
2940  * Signal a userland condition variable.
2941  */
2942 static int
2943 do_cv_signal(struct thread *td, struct ucond *cv)
2944 {
2945 	struct umtx_key key;
2946 	int error, cnt, nwake;
2947 	uint32_t flags;
2948 
2949 	error = fueword32(&cv->c_flags, &flags);
2950 	if (error == -1)
2951 		return (EFAULT);
2952 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
2953 		return (error);
2954 	umtxq_lock(&key);
2955 	umtxq_busy(&key);
2956 	cnt = umtxq_count(&key);
2957 	nwake = umtxq_signal(&key, 1);
2958 	if (cnt <= nwake) {
2959 		umtxq_unlock(&key);
2960 		error = suword32(&cv->c_has_waiters, 0);
2961 		if (error == -1)
2962 			error = EFAULT;
2963 		umtxq_lock(&key);
2964 	}
2965 	umtxq_unbusy(&key);
2966 	umtxq_unlock(&key);
2967 	umtx_key_release(&key);
2968 	return (error);
2969 }
2970 
2971 static int
2972 do_cv_broadcast(struct thread *td, struct ucond *cv)
2973 {
2974 	struct umtx_key key;
2975 	int error;
2976 	uint32_t flags;
2977 
2978 	error = fueword32(&cv->c_flags, &flags);
2979 	if (error == -1)
2980 		return (EFAULT);
2981 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
2982 		return (error);
2983 
2984 	umtxq_lock(&key);
2985 	umtxq_busy(&key);
2986 	umtxq_signal(&key, INT_MAX);
2987 	umtxq_unlock(&key);
2988 
2989 	error = suword32(&cv->c_has_waiters, 0);
2990 	if (error == -1)
2991 		error = EFAULT;
2992 
2993 	umtxq_unbusy_unlocked(&key);
2994 
2995 	umtx_key_release(&key);
2996 	return (error);
2997 }
2998 
2999 static int
3000 do_rw_rdlock(struct thread *td, struct urwlock *rwlock, long fflag,
3001     struct _umtx_time *timeout)
3002 {
3003 	struct umtx_abs_timeout timo;
3004 	struct umtx_q *uq;
3005 	uint32_t flags, wrflags;
3006 	int32_t state, oldstate;
3007 	int32_t blocked_readers;
3008 	int error, error1, rv;
3009 
3010 	uq = td->td_umtxq;
3011 	error = fueword32(&rwlock->rw_flags, &flags);
3012 	if (error == -1)
3013 		return (EFAULT);
3014 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3015 	if (error != 0)
3016 		return (error);
3017 
3018 	if (timeout != NULL)
3019 		umtx_abs_timeout_init2(&timo, timeout);
3020 
3021 	wrflags = URWLOCK_WRITE_OWNER;
3022 	if (!(fflag & URWLOCK_PREFER_READER) && !(flags & URWLOCK_PREFER_READER))
3023 		wrflags |= URWLOCK_WRITE_WAITERS;
3024 
3025 	for (;;) {
3026 		rv = fueword32(&rwlock->rw_state, &state);
3027 		if (rv == -1) {
3028 			umtx_key_release(&uq->uq_key);
3029 			return (EFAULT);
3030 		}
3031 
3032 		/* try to lock it */
3033 		while (!(state & wrflags)) {
3034 			if (__predict_false(URWLOCK_READER_COUNT(state) ==
3035 			    URWLOCK_MAX_READERS)) {
3036 				umtx_key_release(&uq->uq_key);
3037 				return (EAGAIN);
3038 			}
3039 			rv = casueword32(&rwlock->rw_state, state,
3040 			    &oldstate, state + 1);
3041 			if (rv == -1) {
3042 				umtx_key_release(&uq->uq_key);
3043 				return (EFAULT);
3044 			}
3045 			if (rv == 0) {
3046 				MPASS(oldstate == state);
3047 				umtx_key_release(&uq->uq_key);
3048 				return (0);
3049 			}
3050 			error = thread_check_susp(td, true);
3051 			if (error != 0)
3052 				break;
3053 			state = oldstate;
3054 		}
3055 
3056 		if (error)
3057 			break;
3058 
3059 		/* grab monitor lock */
3060 		umtxq_lock(&uq->uq_key);
3061 		umtxq_busy(&uq->uq_key);
3062 		umtxq_unlock(&uq->uq_key);
3063 
3064 		/*
3065 		 * re-read the state, in case it changed between the try-lock above
3066 		 * and the check below
3067 		 */
3068 		rv = fueword32(&rwlock->rw_state, &state);
3069 		if (rv == -1)
3070 			error = EFAULT;
3071 
3072 		/* set read contention bit */
3073 		while (error == 0 && (state & wrflags) &&
3074 		    !(state & URWLOCK_READ_WAITERS)) {
3075 			rv = casueword32(&rwlock->rw_state, state,
3076 			    &oldstate, state | URWLOCK_READ_WAITERS);
3077 			if (rv == -1) {
3078 				error = EFAULT;
3079 				break;
3080 			}
3081 			if (rv == 0) {
3082 				MPASS(oldstate == state);
3083 				goto sleep;
3084 			}
3085 			state = oldstate;
3086 			error = thread_check_susp(td, false);
3087 			if (error != 0)
3088 				break;
3089 		}
3090 		if (error != 0) {
3091 			umtxq_unbusy_unlocked(&uq->uq_key);
3092 			break;
3093 		}
3094 
3095 		/* state is changed while setting flags, restart */
3096 		if (!(state & wrflags)) {
3097 			umtxq_unbusy_unlocked(&uq->uq_key);
3098 			error = thread_check_susp(td, true);
3099 			if (error != 0)
3100 				break;
3101 			continue;
3102 		}
3103 
3104 sleep:
3105 		/*
3106 		 * Contention bit is set, before sleeping, increase
3107 		 * read waiter count.
3108 		 */
3109 		rv = fueword32(&rwlock->rw_blocked_readers,
3110 		    &blocked_readers);
3111 		if (rv == -1) {
3112 			umtxq_unbusy_unlocked(&uq->uq_key);
3113 			error = EFAULT;
3114 			break;
3115 		}
3116 		suword32(&rwlock->rw_blocked_readers, blocked_readers+1);
3117 
3118 		while (state & wrflags) {
3119 			umtxq_lock(&uq->uq_key);
3120 			umtxq_insert(uq);
3121 			umtxq_unbusy(&uq->uq_key);
3122 
3123 			error = umtxq_sleep(uq, "urdlck", timeout == NULL ?
3124 			    NULL : &timo);
3125 
3126 			umtxq_busy(&uq->uq_key);
3127 			umtxq_remove(uq);
3128 			umtxq_unlock(&uq->uq_key);
3129 			if (error)
3130 				break;
3131 			rv = fueword32(&rwlock->rw_state, &state);
3132 			if (rv == -1) {
3133 				error = EFAULT;
3134 				break;
3135 			}
3136 		}
3137 
3138 		/* decrease read waiter count, and may clear read contention bit */
3139 		rv = fueword32(&rwlock->rw_blocked_readers,
3140 		    &blocked_readers);
3141 		if (rv == -1) {
3142 			umtxq_unbusy_unlocked(&uq->uq_key);
3143 			error = EFAULT;
3144 			break;
3145 		}
3146 		suword32(&rwlock->rw_blocked_readers, blocked_readers-1);
3147 		if (blocked_readers == 1) {
3148 			rv = fueword32(&rwlock->rw_state, &state);
3149 			if (rv == -1) {
3150 				umtxq_unbusy_unlocked(&uq->uq_key);
3151 				error = EFAULT;
3152 				break;
3153 			}
3154 			for (;;) {
3155 				rv = casueword32(&rwlock->rw_state, state,
3156 				    &oldstate, state & ~URWLOCK_READ_WAITERS);
3157 				if (rv == -1) {
3158 					error = EFAULT;
3159 					break;
3160 				}
3161 				if (rv == 0) {
3162 					MPASS(oldstate == state);
3163 					break;
3164 				}
3165 				state = oldstate;
3166 				error1 = thread_check_susp(td, false);
3167 				if (error1 != 0) {
3168 					if (error == 0)
3169 						error = error1;
3170 					break;
3171 				}
3172 			}
3173 		}
3174 
3175 		umtxq_unbusy_unlocked(&uq->uq_key);
3176 		if (error != 0)
3177 			break;
3178 	}
3179 	umtx_key_release(&uq->uq_key);
3180 	if (error == ERESTART)
3181 		error = EINTR;
3182 	return (error);
3183 }
3184 
3185 static int
3186 do_rw_wrlock(struct thread *td, struct urwlock *rwlock, struct _umtx_time *timeout)
3187 {
3188 	struct umtx_abs_timeout timo;
3189 	struct umtx_q *uq;
3190 	uint32_t flags;
3191 	int32_t state, oldstate;
3192 	int32_t blocked_writers;
3193 	int32_t blocked_readers;
3194 	int error, error1, rv;
3195 
3196 	uq = td->td_umtxq;
3197 	error = fueword32(&rwlock->rw_flags, &flags);
3198 	if (error == -1)
3199 		return (EFAULT);
3200 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3201 	if (error != 0)
3202 		return (error);
3203 
3204 	if (timeout != NULL)
3205 		umtx_abs_timeout_init2(&timo, timeout);
3206 
3207 	blocked_readers = 0;
3208 	for (;;) {
3209 		rv = fueword32(&rwlock->rw_state, &state);
3210 		if (rv == -1) {
3211 			umtx_key_release(&uq->uq_key);
3212 			return (EFAULT);
3213 		}
3214 		while ((state & URWLOCK_WRITE_OWNER) == 0 &&
3215 		    URWLOCK_READER_COUNT(state) == 0) {
3216 			rv = casueword32(&rwlock->rw_state, state,
3217 			    &oldstate, state | URWLOCK_WRITE_OWNER);
3218 			if (rv == -1) {
3219 				umtx_key_release(&uq->uq_key);
3220 				return (EFAULT);
3221 			}
3222 			if (rv == 0) {
3223 				MPASS(oldstate == state);
3224 				umtx_key_release(&uq->uq_key);
3225 				return (0);
3226 			}
3227 			state = oldstate;
3228 			error = thread_check_susp(td, true);
3229 			if (error != 0)
3230 				break;
3231 		}
3232 
3233 		if (error) {
3234 			if ((state & (URWLOCK_WRITE_OWNER |
3235 			    URWLOCK_WRITE_WAITERS)) == 0 &&
3236 			    blocked_readers != 0) {
3237 				umtxq_lock(&uq->uq_key);
3238 				umtxq_busy(&uq->uq_key);
3239 				umtxq_signal_queue(&uq->uq_key, INT_MAX,
3240 				    UMTX_SHARED_QUEUE);
3241 				umtxq_unbusy(&uq->uq_key);
3242 				umtxq_unlock(&uq->uq_key);
3243 			}
3244 
3245 			break;
3246 		}
3247 
3248 		/* grab monitor lock */
3249 		umtxq_lock(&uq->uq_key);
3250 		umtxq_busy(&uq->uq_key);
3251 		umtxq_unlock(&uq->uq_key);
3252 
3253 		/*
3254 		 * Re-read the state, in case it changed between the
3255 		 * try-lock above and the check below.
3256 		 */
3257 		rv = fueword32(&rwlock->rw_state, &state);
3258 		if (rv == -1)
3259 			error = EFAULT;
3260 
3261 		while (error == 0 && ((state & URWLOCK_WRITE_OWNER) ||
3262 		    URWLOCK_READER_COUNT(state) != 0) &&
3263 		    (state & URWLOCK_WRITE_WAITERS) == 0) {
3264 			rv = casueword32(&rwlock->rw_state, state,
3265 			    &oldstate, state | URWLOCK_WRITE_WAITERS);
3266 			if (rv == -1) {
3267 				error = EFAULT;
3268 				break;
3269 			}
3270 			if (rv == 0) {
3271 				MPASS(oldstate == state);
3272 				goto sleep;
3273 			}
3274 			state = oldstate;
3275 			error = thread_check_susp(td, false);
3276 			if (error != 0)
3277 				break;
3278 		}
3279 		if (error != 0) {
3280 			umtxq_unbusy_unlocked(&uq->uq_key);
3281 			break;
3282 		}
3283 
3284 		if ((state & URWLOCK_WRITE_OWNER) == 0 &&
3285 		    URWLOCK_READER_COUNT(state) == 0) {
3286 			umtxq_unbusy_unlocked(&uq->uq_key);
3287 			error = thread_check_susp(td, false);
3288 			if (error != 0)
3289 				break;
3290 			continue;
3291 		}
3292 sleep:
3293 		rv = fueword32(&rwlock->rw_blocked_writers,
3294 		    &blocked_writers);
3295 		if (rv == -1) {
3296 			umtxq_unbusy_unlocked(&uq->uq_key);
3297 			error = EFAULT;
3298 			break;
3299 		}
3300 		suword32(&rwlock->rw_blocked_writers, blocked_writers + 1);
3301 
3302 		while ((state & URWLOCK_WRITE_OWNER) ||
3303 		    URWLOCK_READER_COUNT(state) != 0) {
3304 			umtxq_lock(&uq->uq_key);
3305 			umtxq_insert_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3306 			umtxq_unbusy(&uq->uq_key);
3307 
3308 			error = umtxq_sleep(uq, "uwrlck", timeout == NULL ?
3309 			    NULL : &timo);
3310 
3311 			umtxq_busy(&uq->uq_key);
3312 			umtxq_remove_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3313 			umtxq_unlock(&uq->uq_key);
3314 			if (error)
3315 				break;
3316 			rv = fueword32(&rwlock->rw_state, &state);
3317 			if (rv == -1) {
3318 				error = EFAULT;
3319 				break;
3320 			}
3321 		}
3322 
3323 		rv = fueword32(&rwlock->rw_blocked_writers,
3324 		    &blocked_writers);
3325 		if (rv == -1) {
3326 			umtxq_unbusy_unlocked(&uq->uq_key);
3327 			error = EFAULT;
3328 			break;
3329 		}
3330 		suword32(&rwlock->rw_blocked_writers, blocked_writers-1);
3331 		if (blocked_writers == 1) {
3332 			rv = fueword32(&rwlock->rw_state, &state);
3333 			if (rv == -1) {
3334 				umtxq_unbusy_unlocked(&uq->uq_key);
3335 				error = EFAULT;
3336 				break;
3337 			}
3338 			for (;;) {
3339 				rv = casueword32(&rwlock->rw_state, state,
3340 				    &oldstate, state & ~URWLOCK_WRITE_WAITERS);
3341 				if (rv == -1) {
3342 					error = EFAULT;
3343 					break;
3344 				}
3345 				if (rv == 0) {
3346 					MPASS(oldstate == state);
3347 					break;
3348 				}
3349 				state = oldstate;
3350 				error1 = thread_check_susp(td, false);
3351 				/*
3352 				 * We are leaving the URWLOCK_WRITE_WAITERS
3353 				 * behind, but this should not harm the
3354 				 * correctness.
3355 				 */
3356 				if (error1 != 0) {
3357 					if (error == 0)
3358 						error = error1;
3359 					break;
3360 				}
3361 			}
3362 			rv = fueword32(&rwlock->rw_blocked_readers,
3363 			    &blocked_readers);
3364 			if (rv == -1) {
3365 				umtxq_unbusy_unlocked(&uq->uq_key);
3366 				error = EFAULT;
3367 				break;
3368 			}
3369 		} else
3370 			blocked_readers = 0;
3371 
3372 		umtxq_unbusy_unlocked(&uq->uq_key);
3373 	}
3374 
3375 	umtx_key_release(&uq->uq_key);
3376 	if (error == ERESTART)
3377 		error = EINTR;
3378 	return (error);
3379 }
3380 
3381 static int
3382 do_rw_unlock(struct thread *td, struct urwlock *rwlock)
3383 {
3384 	struct umtx_q *uq;
3385 	uint32_t flags;
3386 	int32_t state, oldstate;
3387 	int error, rv, q, count;
3388 
3389 	uq = td->td_umtxq;
3390 	error = fueword32(&rwlock->rw_flags, &flags);
3391 	if (error == -1)
3392 		return (EFAULT);
3393 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3394 	if (error != 0)
3395 		return (error);
3396 
3397 	error = fueword32(&rwlock->rw_state, &state);
3398 	if (error == -1) {
3399 		error = EFAULT;
3400 		goto out;
3401 	}
3402 	if (state & URWLOCK_WRITE_OWNER) {
3403 		for (;;) {
3404 			rv = casueword32(&rwlock->rw_state, state,
3405 			    &oldstate, state & ~URWLOCK_WRITE_OWNER);
3406 			if (rv == -1) {
3407 				error = EFAULT;
3408 				goto out;
3409 			}
3410 			if (rv == 1) {
3411 				state = oldstate;
3412 				if (!(oldstate & URWLOCK_WRITE_OWNER)) {
3413 					error = EPERM;
3414 					goto out;
3415 				}
3416 				error = thread_check_susp(td, true);
3417 				if (error != 0)
3418 					goto out;
3419 			} else
3420 				break;
3421 		}
3422 	} else if (URWLOCK_READER_COUNT(state) != 0) {
3423 		for (;;) {
3424 			rv = casueword32(&rwlock->rw_state, state,
3425 			    &oldstate, state - 1);
3426 			if (rv == -1) {
3427 				error = EFAULT;
3428 				goto out;
3429 			}
3430 			if (rv == 1) {
3431 				state = oldstate;
3432 				if (URWLOCK_READER_COUNT(oldstate) == 0) {
3433 					error = EPERM;
3434 					goto out;
3435 				}
3436 				error = thread_check_susp(td, true);
3437 				if (error != 0)
3438 					goto out;
3439 			} else
3440 				break;
3441 		}
3442 	} else {
3443 		error = EPERM;
3444 		goto out;
3445 	}
3446 
3447 	count = 0;
3448 
3449 	if (!(flags & URWLOCK_PREFER_READER)) {
3450 		if (state & URWLOCK_WRITE_WAITERS) {
3451 			count = 1;
3452 			q = UMTX_EXCLUSIVE_QUEUE;
3453 		} else if (state & URWLOCK_READ_WAITERS) {
3454 			count = INT_MAX;
3455 			q = UMTX_SHARED_QUEUE;
3456 		}
3457 	} else {
3458 		if (state & URWLOCK_READ_WAITERS) {
3459 			count = INT_MAX;
3460 			q = UMTX_SHARED_QUEUE;
3461 		} else if (state & URWLOCK_WRITE_WAITERS) {
3462 			count = 1;
3463 			q = UMTX_EXCLUSIVE_QUEUE;
3464 		}
3465 	}
3466 
3467 	if (count) {
3468 		umtxq_lock(&uq->uq_key);
3469 		umtxq_busy(&uq->uq_key);
3470 		umtxq_signal_queue(&uq->uq_key, count, q);
3471 		umtxq_unbusy(&uq->uq_key);
3472 		umtxq_unlock(&uq->uq_key);
3473 	}
3474 out:
3475 	umtx_key_release(&uq->uq_key);
3476 	return (error);
3477 }
3478 
3479 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3480 static int
3481 do_sem_wait(struct thread *td, struct _usem *sem, struct _umtx_time *timeout)
3482 {
3483 	struct umtx_abs_timeout timo;
3484 	struct umtx_q *uq;
3485 	uint32_t flags, count, count1;
3486 	int error, rv, rv1;
3487 
3488 	uq = td->td_umtxq;
3489 	error = fueword32(&sem->_flags, &flags);
3490 	if (error == -1)
3491 		return (EFAULT);
3492 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3493 	if (error != 0)
3494 		return (error);
3495 
3496 	if (timeout != NULL)
3497 		umtx_abs_timeout_init2(&timo, timeout);
3498 
3499 again:
3500 	umtxq_lock(&uq->uq_key);
3501 	umtxq_busy(&uq->uq_key);
3502 	umtxq_insert(uq);
3503 	umtxq_unlock(&uq->uq_key);
3504 	rv = casueword32(&sem->_has_waiters, 0, &count1, 1);
3505 	if (rv == 0)
3506 		rv1 = fueword32(&sem->_count, &count);
3507 	if (rv == -1 || (rv == 0 && (rv1 == -1 || count != 0)) ||
3508 	    (rv == 1 && count1 == 0)) {
3509 		umtxq_lock(&uq->uq_key);
3510 		umtxq_unbusy(&uq->uq_key);
3511 		umtxq_remove(uq);
3512 		umtxq_unlock(&uq->uq_key);
3513 		if (rv == 1) {
3514 			rv = thread_check_susp(td, true);
3515 			if (rv == 0)
3516 				goto again;
3517 			error = rv;
3518 			goto out;
3519 		}
3520 		if (rv == 0)
3521 			rv = rv1;
3522 		error = rv == -1 ? EFAULT : 0;
3523 		goto out;
3524 	}
3525 	umtxq_lock(&uq->uq_key);
3526 	umtxq_unbusy(&uq->uq_key);
3527 
3528 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3529 
3530 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3531 		error = 0;
3532 	else {
3533 		umtxq_remove(uq);
3534 		/* A relative timeout cannot be restarted. */
3535 		if (error == ERESTART && timeout != NULL &&
3536 		    (timeout->_flags & UMTX_ABSTIME) == 0)
3537 			error = EINTR;
3538 	}
3539 	umtxq_unlock(&uq->uq_key);
3540 out:
3541 	umtx_key_release(&uq->uq_key);
3542 	return (error);
3543 }
3544 
3545 /*
3546  * Signal a userland semaphore.
3547  */
3548 static int
3549 do_sem_wake(struct thread *td, struct _usem *sem)
3550 {
3551 	struct umtx_key key;
3552 	int error, cnt;
3553 	uint32_t flags;
3554 
3555 	error = fueword32(&sem->_flags, &flags);
3556 	if (error == -1)
3557 		return (EFAULT);
3558 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3559 		return (error);
3560 	umtxq_lock(&key);
3561 	umtxq_busy(&key);
3562 	cnt = umtxq_count(&key);
3563 	if (cnt > 0) {
3564 		/*
3565 		 * Check if count is greater than 0, this means the memory is
3566 		 * still being referenced by user code, so we can safely
3567 		 * update _has_waiters flag.
3568 		 */
3569 		if (cnt == 1) {
3570 			umtxq_unlock(&key);
3571 			error = suword32(&sem->_has_waiters, 0);
3572 			umtxq_lock(&key);
3573 			if (error == -1)
3574 				error = EFAULT;
3575 		}
3576 		umtxq_signal(&key, 1);
3577 	}
3578 	umtxq_unbusy(&key);
3579 	umtxq_unlock(&key);
3580 	umtx_key_release(&key);
3581 	return (error);
3582 }
3583 #endif
3584 
3585 static int
3586 do_sem2_wait(struct thread *td, struct _usem2 *sem, struct _umtx_time *timeout)
3587 {
3588 	struct umtx_abs_timeout timo;
3589 	struct umtx_q *uq;
3590 	uint32_t count, flags;
3591 	int error, rv;
3592 
3593 	uq = td->td_umtxq;
3594 	flags = fuword32(&sem->_flags);
3595 	if (timeout != NULL)
3596 		umtx_abs_timeout_init2(&timo, timeout);
3597 
3598 again:
3599 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3600 	if (error != 0)
3601 		return (error);
3602 	umtxq_lock(&uq->uq_key);
3603 	umtxq_busy(&uq->uq_key);
3604 	umtxq_insert(uq);
3605 	umtxq_unlock(&uq->uq_key);
3606 	rv = fueword32(&sem->_count, &count);
3607 	if (rv == -1) {
3608 		umtxq_lock(&uq->uq_key);
3609 		umtxq_unbusy(&uq->uq_key);
3610 		umtxq_remove(uq);
3611 		umtxq_unlock(&uq->uq_key);
3612 		umtx_key_release(&uq->uq_key);
3613 		return (EFAULT);
3614 	}
3615 	for (;;) {
3616 		if (USEM_COUNT(count) != 0) {
3617 			umtxq_lock(&uq->uq_key);
3618 			umtxq_unbusy(&uq->uq_key);
3619 			umtxq_remove(uq);
3620 			umtxq_unlock(&uq->uq_key);
3621 			umtx_key_release(&uq->uq_key);
3622 			return (0);
3623 		}
3624 		if (count == USEM_HAS_WAITERS)
3625 			break;
3626 		rv = casueword32(&sem->_count, 0, &count, USEM_HAS_WAITERS);
3627 		if (rv == 0)
3628 			break;
3629 		umtxq_lock(&uq->uq_key);
3630 		umtxq_unbusy(&uq->uq_key);
3631 		umtxq_remove(uq);
3632 		umtxq_unlock(&uq->uq_key);
3633 		umtx_key_release(&uq->uq_key);
3634 		if (rv == -1)
3635 			return (EFAULT);
3636 		rv = thread_check_susp(td, true);
3637 		if (rv != 0)
3638 			return (rv);
3639 		goto again;
3640 	}
3641 	umtxq_lock(&uq->uq_key);
3642 	umtxq_unbusy(&uq->uq_key);
3643 
3644 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3645 
3646 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3647 		error = 0;
3648 	else {
3649 		umtxq_remove(uq);
3650 		if (timeout != NULL && (timeout->_flags & UMTX_ABSTIME) == 0) {
3651 			/* A relative timeout cannot be restarted. */
3652 			if (error == ERESTART)
3653 				error = EINTR;
3654 			if (error == EINTR) {
3655 				umtx_abs_timeout_update(&timo);
3656 				timespecsub(&timo.end, &timo.cur,
3657 				    &timeout->_timeout);
3658 			}
3659 		}
3660 	}
3661 	umtxq_unlock(&uq->uq_key);
3662 	umtx_key_release(&uq->uq_key);
3663 	return (error);
3664 }
3665 
3666 /*
3667  * Signal a userland semaphore.
3668  */
3669 static int
3670 do_sem2_wake(struct thread *td, struct _usem2 *sem)
3671 {
3672 	struct umtx_key key;
3673 	int error, cnt, rv;
3674 	uint32_t count, flags;
3675 
3676 	rv = fueword32(&sem->_flags, &flags);
3677 	if (rv == -1)
3678 		return (EFAULT);
3679 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3680 		return (error);
3681 	umtxq_lock(&key);
3682 	umtxq_busy(&key);
3683 	cnt = umtxq_count(&key);
3684 	if (cnt > 0) {
3685 		/*
3686 		 * If this was the last sleeping thread, clear the waiters
3687 		 * flag in _count.
3688 		 */
3689 		if (cnt == 1) {
3690 			umtxq_unlock(&key);
3691 			rv = fueword32(&sem->_count, &count);
3692 			while (rv != -1 && count & USEM_HAS_WAITERS) {
3693 				rv = casueword32(&sem->_count, count, &count,
3694 				    count & ~USEM_HAS_WAITERS);
3695 				if (rv == 1) {
3696 					rv = thread_check_susp(td, true);
3697 					if (rv != 0)
3698 						break;
3699 				}
3700 			}
3701 			if (rv == -1)
3702 				error = EFAULT;
3703 			else if (rv > 0) {
3704 				error = rv;
3705 			}
3706 			umtxq_lock(&key);
3707 		}
3708 
3709 		umtxq_signal(&key, 1);
3710 	}
3711 	umtxq_unbusy(&key);
3712 	umtxq_unlock(&key);
3713 	umtx_key_release(&key);
3714 	return (error);
3715 }
3716 
3717 #ifdef COMPAT_FREEBSD10
3718 int
3719 freebsd10__umtx_lock(struct thread *td, struct freebsd10__umtx_lock_args *uap)
3720 {
3721 	return (do_lock_umtx(td, uap->umtx, td->td_tid, 0));
3722 }
3723 
3724 int
3725 freebsd10__umtx_unlock(struct thread *td,
3726     struct freebsd10__umtx_unlock_args *uap)
3727 {
3728 	return (do_unlock_umtx(td, uap->umtx, td->td_tid));
3729 }
3730 #endif
3731 
3732 inline int
3733 umtx_copyin_timeout(const void *uaddr, struct timespec *tsp)
3734 {
3735 	int error;
3736 
3737 	error = copyin(uaddr, tsp, sizeof(*tsp));
3738 	if (error == 0) {
3739 		if (tsp->tv_sec < 0 ||
3740 		    tsp->tv_nsec >= 1000000000 ||
3741 		    tsp->tv_nsec < 0)
3742 			error = EINVAL;
3743 	}
3744 	return (error);
3745 }
3746 
3747 static inline int
3748 umtx_copyin_umtx_time(const void *uaddr, size_t size, struct _umtx_time *tp)
3749 {
3750 	int error;
3751 
3752 	if (size <= sizeof(tp->_timeout)) {
3753 		tp->_clockid = CLOCK_REALTIME;
3754 		tp->_flags = 0;
3755 		error = copyin(uaddr, &tp->_timeout, sizeof(tp->_timeout));
3756 	} else
3757 		error = copyin(uaddr, tp, sizeof(*tp));
3758 	if (error != 0)
3759 		return (error);
3760 	if (tp->_timeout.tv_sec < 0 ||
3761 	    tp->_timeout.tv_nsec >= 1000000000 || tp->_timeout.tv_nsec < 0)
3762 		return (EINVAL);
3763 	return (0);
3764 }
3765 
3766 static int
3767 umtx_copyin_robust_lists(const void *uaddr, size_t size,
3768     struct umtx_robust_lists_params *rb)
3769 {
3770 
3771 	if (size > sizeof(*rb))
3772 		return (EINVAL);
3773 	return (copyin(uaddr, rb, size));
3774 }
3775 
3776 static int
3777 umtx_copyout_timeout(void *uaddr, size_t sz, struct timespec *tsp)
3778 {
3779 
3780 	/*
3781 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
3782 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
3783 	 * copyops.
3784 	 */
3785 	KASSERT(sz >= sizeof(*tsp),
3786 	    ("umtx_copyops specifies incorrect sizes"));
3787 
3788 	return (copyout(tsp, uaddr, sizeof(*tsp)));
3789 }
3790 
3791 #ifdef COMPAT_FREEBSD10
3792 static int
3793 __umtx_op_lock_umtx(struct thread *td, struct _umtx_op_args *uap,
3794     const struct umtx_copyops *ops)
3795 {
3796 	struct timespec *ts, timeout;
3797 	int error;
3798 
3799 	/* Allow a null timespec (wait forever). */
3800 	if (uap->uaddr2 == NULL)
3801 		ts = NULL;
3802 	else {
3803 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
3804 		if (error != 0)
3805 			return (error);
3806 		ts = &timeout;
3807 	}
3808 #ifdef COMPAT_FREEBSD32
3809 	if (ops->compat32)
3810 		return (do_lock_umtx32(td, uap->obj, uap->val, ts));
3811 #endif
3812 	return (do_lock_umtx(td, uap->obj, uap->val, ts));
3813 }
3814 
3815 static int
3816 __umtx_op_unlock_umtx(struct thread *td, struct _umtx_op_args *uap,
3817     const struct umtx_copyops *ops)
3818 {
3819 #ifdef COMPAT_FREEBSD32
3820 	if (ops->compat32)
3821 		return (do_unlock_umtx32(td, uap->obj, uap->val));
3822 #endif
3823 	return (do_unlock_umtx(td, uap->obj, uap->val));
3824 }
3825 #endif	/* COMPAT_FREEBSD10 */
3826 
3827 #if !defined(COMPAT_FREEBSD10)
3828 static int
3829 __umtx_op_unimpl(struct thread *td __unused, struct _umtx_op_args *uap __unused,
3830     const struct umtx_copyops *ops __unused)
3831 {
3832 	return (EOPNOTSUPP);
3833 }
3834 #endif	/* COMPAT_FREEBSD10 */
3835 
3836 static int
3837 __umtx_op_wait(struct thread *td, struct _umtx_op_args *uap,
3838     const struct umtx_copyops *ops)
3839 {
3840 	struct _umtx_time timeout, *tm_p;
3841 	int error;
3842 
3843 	if (uap->uaddr2 == NULL)
3844 		tm_p = NULL;
3845 	else {
3846 		error = ops->copyin_umtx_time(
3847 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3848 		if (error != 0)
3849 			return (error);
3850 		tm_p = &timeout;
3851 	}
3852 	return (do_wait(td, uap->obj, uap->val, tm_p, ops->compat32, 0));
3853 }
3854 
3855 static int
3856 __umtx_op_wait_uint(struct thread *td, struct _umtx_op_args *uap,
3857     const struct umtx_copyops *ops)
3858 {
3859 	struct _umtx_time timeout, *tm_p;
3860 	int error;
3861 
3862 	if (uap->uaddr2 == NULL)
3863 		tm_p = NULL;
3864 	else {
3865 		error = ops->copyin_umtx_time(
3866 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3867 		if (error != 0)
3868 			return (error);
3869 		tm_p = &timeout;
3870 	}
3871 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
3872 }
3873 
3874 static int
3875 __umtx_op_wait_uint_private(struct thread *td, struct _umtx_op_args *uap,
3876     const struct umtx_copyops *ops)
3877 {
3878 	struct _umtx_time *tm_p, timeout;
3879 	int error;
3880 
3881 	if (uap->uaddr2 == NULL)
3882 		tm_p = NULL;
3883 	else {
3884 		error = ops->copyin_umtx_time(
3885 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3886 		if (error != 0)
3887 			return (error);
3888 		tm_p = &timeout;
3889 	}
3890 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
3891 }
3892 
3893 static int
3894 __umtx_op_wake(struct thread *td, struct _umtx_op_args *uap,
3895     const struct umtx_copyops *ops __unused)
3896 {
3897 
3898 	return (kern_umtx_wake(td, uap->obj, uap->val, 0));
3899 }
3900 
3901 #define BATCH_SIZE	128
3902 static int
3903 __umtx_op_nwake_private_native(struct thread *td, struct _umtx_op_args *uap)
3904 {
3905 	char *uaddrs[BATCH_SIZE], **upp;
3906 	int count, error, i, pos, tocopy;
3907 
3908 	upp = (char **)uap->obj;
3909 	error = 0;
3910 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3911 	    pos += tocopy) {
3912 		tocopy = MIN(count, BATCH_SIZE);
3913 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(char *));
3914 		if (error != 0)
3915 			break;
3916 		for (i = 0; i < tocopy; ++i) {
3917 			kern_umtx_wake(td, uaddrs[i], INT_MAX, 1);
3918 		}
3919 		maybe_yield();
3920 	}
3921 	return (error);
3922 }
3923 
3924 static int
3925 __umtx_op_nwake_private_compat32(struct thread *td, struct _umtx_op_args *uap)
3926 {
3927 	uint32_t uaddrs[BATCH_SIZE], *upp;
3928 	int count, error, i, pos, tocopy;
3929 
3930 	upp = (uint32_t *)uap->obj;
3931 	error = 0;
3932 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3933 	    pos += tocopy) {
3934 		tocopy = MIN(count, BATCH_SIZE);
3935 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(uint32_t));
3936 		if (error != 0)
3937 			break;
3938 		for (i = 0; i < tocopy; ++i) {
3939 			kern_umtx_wake(td, (void *)(uintptr_t)uaddrs[i],
3940 			    INT_MAX, 1);
3941 		}
3942 		maybe_yield();
3943 	}
3944 	return (error);
3945 }
3946 
3947 static int
3948 __umtx_op_nwake_private(struct thread *td, struct _umtx_op_args *uap,
3949     const struct umtx_copyops *ops)
3950 {
3951 
3952 	if (ops->compat32)
3953 		return (__umtx_op_nwake_private_compat32(td, uap));
3954 	return (__umtx_op_nwake_private_native(td, uap));
3955 }
3956 
3957 static int
3958 __umtx_op_wake_private(struct thread *td, struct _umtx_op_args *uap,
3959     const struct umtx_copyops *ops __unused)
3960 {
3961 
3962 	return (kern_umtx_wake(td, uap->obj, uap->val, 1));
3963 }
3964 
3965 static int
3966 __umtx_op_lock_umutex(struct thread *td, struct _umtx_op_args *uap,
3967    const struct umtx_copyops *ops)
3968 {
3969 	struct _umtx_time *tm_p, timeout;
3970 	int error;
3971 
3972 	/* Allow a null timespec (wait forever). */
3973 	if (uap->uaddr2 == NULL)
3974 		tm_p = NULL;
3975 	else {
3976 		error = ops->copyin_umtx_time(
3977 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3978 		if (error != 0)
3979 			return (error);
3980 		tm_p = &timeout;
3981 	}
3982 	return (do_lock_umutex(td, uap->obj, tm_p, 0));
3983 }
3984 
3985 static int
3986 __umtx_op_trylock_umutex(struct thread *td, struct _umtx_op_args *uap,
3987     const struct umtx_copyops *ops __unused)
3988 {
3989 
3990 	return (do_lock_umutex(td, uap->obj, NULL, _UMUTEX_TRY));
3991 }
3992 
3993 static int
3994 __umtx_op_wait_umutex(struct thread *td, struct _umtx_op_args *uap,
3995     const struct umtx_copyops *ops)
3996 {
3997 	struct _umtx_time *tm_p, timeout;
3998 	int error;
3999 
4000 	/* Allow a null timespec (wait forever). */
4001 	if (uap->uaddr2 == NULL)
4002 		tm_p = NULL;
4003 	else {
4004 		error = ops->copyin_umtx_time(
4005 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4006 		if (error != 0)
4007 			return (error);
4008 		tm_p = &timeout;
4009 	}
4010 	return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
4011 }
4012 
4013 static int
4014 __umtx_op_wake_umutex(struct thread *td, struct _umtx_op_args *uap,
4015     const struct umtx_copyops *ops __unused)
4016 {
4017 
4018 	return (do_wake_umutex(td, uap->obj));
4019 }
4020 
4021 static int
4022 __umtx_op_unlock_umutex(struct thread *td, struct _umtx_op_args *uap,
4023     const struct umtx_copyops *ops __unused)
4024 {
4025 
4026 	return (do_unlock_umutex(td, uap->obj, false));
4027 }
4028 
4029 static int
4030 __umtx_op_set_ceiling(struct thread *td, struct _umtx_op_args *uap,
4031     const struct umtx_copyops *ops __unused)
4032 {
4033 
4034 	return (do_set_ceiling(td, uap->obj, uap->val, uap->uaddr1));
4035 }
4036 
4037 static int
4038 __umtx_op_cv_wait(struct thread *td, struct _umtx_op_args *uap,
4039     const struct umtx_copyops *ops)
4040 {
4041 	struct timespec *ts, timeout;
4042 	int error;
4043 
4044 	/* Allow a null timespec (wait forever). */
4045 	if (uap->uaddr2 == NULL)
4046 		ts = NULL;
4047 	else {
4048 		error = ops->copyin_timeout(uap->uaddr2, &timeout);
4049 		if (error != 0)
4050 			return (error);
4051 		ts = &timeout;
4052 	}
4053 	return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
4054 }
4055 
4056 static int
4057 __umtx_op_cv_signal(struct thread *td, struct _umtx_op_args *uap,
4058     const struct umtx_copyops *ops __unused)
4059 {
4060 
4061 	return (do_cv_signal(td, uap->obj));
4062 }
4063 
4064 static int
4065 __umtx_op_cv_broadcast(struct thread *td, struct _umtx_op_args *uap,
4066     const struct umtx_copyops *ops __unused)
4067 {
4068 
4069 	return (do_cv_broadcast(td, uap->obj));
4070 }
4071 
4072 static int
4073 __umtx_op_rw_rdlock(struct thread *td, struct _umtx_op_args *uap,
4074     const struct umtx_copyops *ops)
4075 {
4076 	struct _umtx_time timeout;
4077 	int error;
4078 
4079 	/* Allow a null timespec (wait forever). */
4080 	if (uap->uaddr2 == NULL) {
4081 		error = do_rw_rdlock(td, uap->obj, uap->val, 0);
4082 	} else {
4083 		error = ops->copyin_umtx_time(uap->uaddr2,
4084 		   (size_t)uap->uaddr1, &timeout);
4085 		if (error != 0)
4086 			return (error);
4087 		error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
4088 	}
4089 	return (error);
4090 }
4091 
4092 static int
4093 __umtx_op_rw_wrlock(struct thread *td, struct _umtx_op_args *uap,
4094     const struct umtx_copyops *ops)
4095 {
4096 	struct _umtx_time timeout;
4097 	int error;
4098 
4099 	/* Allow a null timespec (wait forever). */
4100 	if (uap->uaddr2 == NULL) {
4101 		error = do_rw_wrlock(td, uap->obj, 0);
4102 	} else {
4103 		error = ops->copyin_umtx_time(uap->uaddr2,
4104 		   (size_t)uap->uaddr1, &timeout);
4105 		if (error != 0)
4106 			return (error);
4107 
4108 		error = do_rw_wrlock(td, uap->obj, &timeout);
4109 	}
4110 	return (error);
4111 }
4112 
4113 static int
4114 __umtx_op_rw_unlock(struct thread *td, struct _umtx_op_args *uap,
4115     const struct umtx_copyops *ops __unused)
4116 {
4117 
4118 	return (do_rw_unlock(td, uap->obj));
4119 }
4120 
4121 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4122 static int
4123 __umtx_op_sem_wait(struct thread *td, struct _umtx_op_args *uap,
4124     const struct umtx_copyops *ops)
4125 {
4126 	struct _umtx_time *tm_p, timeout;
4127 	int error;
4128 
4129 	/* Allow a null timespec (wait forever). */
4130 	if (uap->uaddr2 == NULL)
4131 		tm_p = NULL;
4132 	else {
4133 		error = ops->copyin_umtx_time(
4134 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4135 		if (error != 0)
4136 			return (error);
4137 		tm_p = &timeout;
4138 	}
4139 	return (do_sem_wait(td, uap->obj, tm_p));
4140 }
4141 
4142 static int
4143 __umtx_op_sem_wake(struct thread *td, struct _umtx_op_args *uap,
4144     const struct umtx_copyops *ops __unused)
4145 {
4146 
4147 	return (do_sem_wake(td, uap->obj));
4148 }
4149 #endif
4150 
4151 static int
4152 __umtx_op_wake2_umutex(struct thread *td, struct _umtx_op_args *uap,
4153     const struct umtx_copyops *ops __unused)
4154 {
4155 
4156 	return (do_wake2_umutex(td, uap->obj, uap->val));
4157 }
4158 
4159 static int
4160 __umtx_op_sem2_wait(struct thread *td, struct _umtx_op_args *uap,
4161     const struct umtx_copyops *ops)
4162 {
4163 	struct _umtx_time *tm_p, timeout;
4164 	size_t uasize;
4165 	int error;
4166 
4167 	/* Allow a null timespec (wait forever). */
4168 	if (uap->uaddr2 == NULL) {
4169 		uasize = 0;
4170 		tm_p = NULL;
4171 	} else {
4172 		uasize = (size_t)uap->uaddr1;
4173 		error = ops->copyin_umtx_time(uap->uaddr2, uasize, &timeout);
4174 		if (error != 0)
4175 			return (error);
4176 		tm_p = &timeout;
4177 	}
4178 	error = do_sem2_wait(td, uap->obj, tm_p);
4179 	if (error == EINTR && uap->uaddr2 != NULL &&
4180 	    (timeout._flags & UMTX_ABSTIME) == 0 &&
4181 	    uasize >= ops->umtx_time_sz + ops->timespec_sz) {
4182 		error = ops->copyout_timeout(
4183 		    (void *)((uintptr_t)uap->uaddr2 + ops->umtx_time_sz),
4184 		    uasize - ops->umtx_time_sz, &timeout._timeout);
4185 		if (error == 0) {
4186 			error = EINTR;
4187 		}
4188 	}
4189 
4190 	return (error);
4191 }
4192 
4193 static int
4194 __umtx_op_sem2_wake(struct thread *td, struct _umtx_op_args *uap,
4195     const struct umtx_copyops *ops __unused)
4196 {
4197 
4198 	return (do_sem2_wake(td, uap->obj));
4199 }
4200 
4201 #define	USHM_OBJ_UMTX(o)						\
4202     ((struct umtx_shm_obj_list *)(&(o)->umtx_data))
4203 
4204 #define	USHMF_REG_LINKED	0x0001
4205 #define	USHMF_OBJ_LINKED	0x0002
4206 struct umtx_shm_reg {
4207 	TAILQ_ENTRY(umtx_shm_reg) ushm_reg_link;
4208 	LIST_ENTRY(umtx_shm_reg) ushm_obj_link;
4209 	struct umtx_key		ushm_key;
4210 	struct ucred		*ushm_cred;
4211 	struct shmfd		*ushm_obj;
4212 	u_int			ushm_refcnt;
4213 	u_int			ushm_flags;
4214 };
4215 
4216 LIST_HEAD(umtx_shm_obj_list, umtx_shm_reg);
4217 TAILQ_HEAD(umtx_shm_reg_head, umtx_shm_reg);
4218 
4219 static uma_zone_t umtx_shm_reg_zone;
4220 static struct umtx_shm_reg_head umtx_shm_registry[UMTX_CHAINS];
4221 static struct mtx umtx_shm_lock;
4222 static struct umtx_shm_reg_head umtx_shm_reg_delfree =
4223     TAILQ_HEAD_INITIALIZER(umtx_shm_reg_delfree);
4224 
4225 static void umtx_shm_free_reg(struct umtx_shm_reg *reg);
4226 
4227 static void
4228 umtx_shm_reg_delfree_tq(void *context __unused, int pending __unused)
4229 {
4230 	struct umtx_shm_reg_head d;
4231 	struct umtx_shm_reg *reg, *reg1;
4232 
4233 	TAILQ_INIT(&d);
4234 	mtx_lock(&umtx_shm_lock);
4235 	TAILQ_CONCAT(&d, &umtx_shm_reg_delfree, ushm_reg_link);
4236 	mtx_unlock(&umtx_shm_lock);
4237 	TAILQ_FOREACH_SAFE(reg, &d, ushm_reg_link, reg1) {
4238 		TAILQ_REMOVE(&d, reg, ushm_reg_link);
4239 		umtx_shm_free_reg(reg);
4240 	}
4241 }
4242 
4243 static struct task umtx_shm_reg_delfree_task =
4244     TASK_INITIALIZER(0, umtx_shm_reg_delfree_tq, NULL);
4245 
4246 static struct umtx_shm_reg *
4247 umtx_shm_find_reg_locked(const struct umtx_key *key)
4248 {
4249 	struct umtx_shm_reg *reg;
4250 	struct umtx_shm_reg_head *reg_head;
4251 
4252 	KASSERT(key->shared, ("umtx_p_find_rg: private key"));
4253 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4254 	reg_head = &umtx_shm_registry[key->hash];
4255 	TAILQ_FOREACH(reg, reg_head, ushm_reg_link) {
4256 		KASSERT(reg->ushm_key.shared,
4257 		    ("non-shared key on reg %p %d", reg, reg->ushm_key.shared));
4258 		if (reg->ushm_key.info.shared.object ==
4259 		    key->info.shared.object &&
4260 		    reg->ushm_key.info.shared.offset ==
4261 		    key->info.shared.offset) {
4262 			KASSERT(reg->ushm_key.type == TYPE_SHM, ("TYPE_USHM"));
4263 			KASSERT(reg->ushm_refcnt > 0,
4264 			    ("reg %p refcnt 0 onlist", reg));
4265 			KASSERT((reg->ushm_flags & USHMF_REG_LINKED) != 0,
4266 			    ("reg %p not linked", reg));
4267 			reg->ushm_refcnt++;
4268 			return (reg);
4269 		}
4270 	}
4271 	return (NULL);
4272 }
4273 
4274 static struct umtx_shm_reg *
4275 umtx_shm_find_reg(const struct umtx_key *key)
4276 {
4277 	struct umtx_shm_reg *reg;
4278 
4279 	mtx_lock(&umtx_shm_lock);
4280 	reg = umtx_shm_find_reg_locked(key);
4281 	mtx_unlock(&umtx_shm_lock);
4282 	return (reg);
4283 }
4284 
4285 static void
4286 umtx_shm_free_reg(struct umtx_shm_reg *reg)
4287 {
4288 
4289 	chgumtxcnt(reg->ushm_cred->cr_ruidinfo, -1, 0);
4290 	crfree(reg->ushm_cred);
4291 	shm_drop(reg->ushm_obj);
4292 	uma_zfree(umtx_shm_reg_zone, reg);
4293 }
4294 
4295 static bool
4296 umtx_shm_unref_reg_locked(struct umtx_shm_reg *reg, bool force)
4297 {
4298 	bool res;
4299 
4300 	mtx_assert(&umtx_shm_lock, MA_OWNED);
4301 	KASSERT(reg->ushm_refcnt > 0, ("ushm_reg %p refcnt 0", reg));
4302 	reg->ushm_refcnt--;
4303 	res = reg->ushm_refcnt == 0;
4304 	if (res || force) {
4305 		if ((reg->ushm_flags & USHMF_REG_LINKED) != 0) {
4306 			TAILQ_REMOVE(&umtx_shm_registry[reg->ushm_key.hash],
4307 			    reg, ushm_reg_link);
4308 			reg->ushm_flags &= ~USHMF_REG_LINKED;
4309 		}
4310 		if ((reg->ushm_flags & USHMF_OBJ_LINKED) != 0) {
4311 			LIST_REMOVE(reg, ushm_obj_link);
4312 			reg->ushm_flags &= ~USHMF_OBJ_LINKED;
4313 		}
4314 	}
4315 	return (res);
4316 }
4317 
4318 static void
4319 umtx_shm_unref_reg(struct umtx_shm_reg *reg, bool force)
4320 {
4321 	vm_object_t object;
4322 	bool dofree;
4323 
4324 	if (force) {
4325 		object = reg->ushm_obj->shm_object;
4326 		VM_OBJECT_WLOCK(object);
4327 		object->flags |= OBJ_UMTXDEAD;
4328 		VM_OBJECT_WUNLOCK(object);
4329 	}
4330 	mtx_lock(&umtx_shm_lock);
4331 	dofree = umtx_shm_unref_reg_locked(reg, force);
4332 	mtx_unlock(&umtx_shm_lock);
4333 	if (dofree)
4334 		umtx_shm_free_reg(reg);
4335 }
4336 
4337 void
4338 umtx_shm_object_init(vm_object_t object)
4339 {
4340 
4341 	LIST_INIT(USHM_OBJ_UMTX(object));
4342 }
4343 
4344 void
4345 umtx_shm_object_terminated(vm_object_t object)
4346 {
4347 	struct umtx_shm_reg *reg, *reg1;
4348 	bool dofree;
4349 
4350 	if (LIST_EMPTY(USHM_OBJ_UMTX(object)))
4351 		return;
4352 
4353 	dofree = false;
4354 	mtx_lock(&umtx_shm_lock);
4355 	LIST_FOREACH_SAFE(reg, USHM_OBJ_UMTX(object), ushm_obj_link, reg1) {
4356 		if (umtx_shm_unref_reg_locked(reg, true)) {
4357 			TAILQ_INSERT_TAIL(&umtx_shm_reg_delfree, reg,
4358 			    ushm_reg_link);
4359 			dofree = true;
4360 		}
4361 	}
4362 	mtx_unlock(&umtx_shm_lock);
4363 	if (dofree)
4364 		taskqueue_enqueue(taskqueue_thread, &umtx_shm_reg_delfree_task);
4365 }
4366 
4367 static int
4368 umtx_shm_create_reg(struct thread *td, const struct umtx_key *key,
4369     struct umtx_shm_reg **res)
4370 {
4371 	struct umtx_shm_reg *reg, *reg1;
4372 	struct ucred *cred;
4373 	int error;
4374 
4375 	reg = umtx_shm_find_reg(key);
4376 	if (reg != NULL) {
4377 		*res = reg;
4378 		return (0);
4379 	}
4380 	cred = td->td_ucred;
4381 	if (!chgumtxcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_UMTXP)))
4382 		return (ENOMEM);
4383 	reg = uma_zalloc(umtx_shm_reg_zone, M_WAITOK | M_ZERO);
4384 	reg->ushm_refcnt = 1;
4385 	bcopy(key, &reg->ushm_key, sizeof(*key));
4386 	reg->ushm_obj = shm_alloc(td->td_ucred, O_RDWR, false);
4387 	reg->ushm_cred = crhold(cred);
4388 	error = shm_dotruncate(reg->ushm_obj, PAGE_SIZE);
4389 	if (error != 0) {
4390 		umtx_shm_free_reg(reg);
4391 		return (error);
4392 	}
4393 	mtx_lock(&umtx_shm_lock);
4394 	reg1 = umtx_shm_find_reg_locked(key);
4395 	if (reg1 != NULL) {
4396 		mtx_unlock(&umtx_shm_lock);
4397 		umtx_shm_free_reg(reg);
4398 		*res = reg1;
4399 		return (0);
4400 	}
4401 	reg->ushm_refcnt++;
4402 	TAILQ_INSERT_TAIL(&umtx_shm_registry[key->hash], reg, ushm_reg_link);
4403 	LIST_INSERT_HEAD(USHM_OBJ_UMTX(key->info.shared.object), reg,
4404 	    ushm_obj_link);
4405 	reg->ushm_flags = USHMF_REG_LINKED | USHMF_OBJ_LINKED;
4406 	mtx_unlock(&umtx_shm_lock);
4407 	*res = reg;
4408 	return (0);
4409 }
4410 
4411 static int
4412 umtx_shm_alive(struct thread *td, void *addr)
4413 {
4414 	vm_map_t map;
4415 	vm_map_entry_t entry;
4416 	vm_object_t object;
4417 	vm_pindex_t pindex;
4418 	vm_prot_t prot;
4419 	int res, ret;
4420 	boolean_t wired;
4421 
4422 	map = &td->td_proc->p_vmspace->vm_map;
4423 	res = vm_map_lookup(&map, (uintptr_t)addr, VM_PROT_READ, &entry,
4424 	    &object, &pindex, &prot, &wired);
4425 	if (res != KERN_SUCCESS)
4426 		return (EFAULT);
4427 	if (object == NULL)
4428 		ret = EINVAL;
4429 	else
4430 		ret = (object->flags & OBJ_UMTXDEAD) != 0 ? ENOTTY : 0;
4431 	vm_map_lookup_done(map, entry);
4432 	return (ret);
4433 }
4434 
4435 static void
4436 umtx_shm_init(void)
4437 {
4438 	int i;
4439 
4440 	umtx_shm_reg_zone = uma_zcreate("umtx_shm", sizeof(struct umtx_shm_reg),
4441 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
4442 	mtx_init(&umtx_shm_lock, "umtxshm", NULL, MTX_DEF);
4443 	for (i = 0; i < nitems(umtx_shm_registry); i++)
4444 		TAILQ_INIT(&umtx_shm_registry[i]);
4445 }
4446 
4447 static int
4448 umtx_shm(struct thread *td, void *addr, u_int flags)
4449 {
4450 	struct umtx_key key;
4451 	struct umtx_shm_reg *reg;
4452 	struct file *fp;
4453 	int error, fd;
4454 
4455 	if (__bitcount(flags & (UMTX_SHM_CREAT | UMTX_SHM_LOOKUP |
4456 	    UMTX_SHM_DESTROY| UMTX_SHM_ALIVE)) != 1)
4457 		return (EINVAL);
4458 	if ((flags & UMTX_SHM_ALIVE) != 0)
4459 		return (umtx_shm_alive(td, addr));
4460 	error = umtx_key_get(addr, TYPE_SHM, PROCESS_SHARE, &key);
4461 	if (error != 0)
4462 		return (error);
4463 	KASSERT(key.shared == 1, ("non-shared key"));
4464 	if ((flags & UMTX_SHM_CREAT) != 0) {
4465 		error = umtx_shm_create_reg(td, &key, &reg);
4466 	} else {
4467 		reg = umtx_shm_find_reg(&key);
4468 		if (reg == NULL)
4469 			error = ESRCH;
4470 	}
4471 	umtx_key_release(&key);
4472 	if (error != 0)
4473 		return (error);
4474 	KASSERT(reg != NULL, ("no reg"));
4475 	if ((flags & UMTX_SHM_DESTROY) != 0) {
4476 		umtx_shm_unref_reg(reg, true);
4477 	} else {
4478 #if 0
4479 #ifdef MAC
4480 		error = mac_posixshm_check_open(td->td_ucred,
4481 		    reg->ushm_obj, FFLAGS(O_RDWR));
4482 		if (error == 0)
4483 #endif
4484 			error = shm_access(reg->ushm_obj, td->td_ucred,
4485 			    FFLAGS(O_RDWR));
4486 		if (error == 0)
4487 #endif
4488 			error = falloc_caps(td, &fp, &fd, O_CLOEXEC, NULL);
4489 		if (error == 0) {
4490 			shm_hold(reg->ushm_obj);
4491 			finit(fp, FFLAGS(O_RDWR), DTYPE_SHM, reg->ushm_obj,
4492 			    &shm_ops);
4493 			td->td_retval[0] = fd;
4494 			fdrop(fp, td);
4495 		}
4496 	}
4497 	umtx_shm_unref_reg(reg, false);
4498 	return (error);
4499 }
4500 
4501 static int
4502 __umtx_op_shm(struct thread *td, struct _umtx_op_args *uap,
4503     const struct umtx_copyops *ops __unused)
4504 {
4505 
4506 	return (umtx_shm(td, uap->uaddr1, uap->val));
4507 }
4508 
4509 static int
4510 __umtx_op_robust_lists(struct thread *td, struct _umtx_op_args *uap,
4511     const struct umtx_copyops *ops)
4512 {
4513 	struct umtx_robust_lists_params rb;
4514 	int error;
4515 
4516 	if (ops->compat32) {
4517 		if ((td->td_pflags2 & TDP2_COMPAT32RB) == 0 &&
4518 		    (td->td_rb_list != 0 || td->td_rbp_list != 0 ||
4519 		    td->td_rb_inact != 0))
4520 			return (EBUSY);
4521 	} else if ((td->td_pflags2 & TDP2_COMPAT32RB) != 0) {
4522 		return (EBUSY);
4523 	}
4524 
4525 	bzero(&rb, sizeof(rb));
4526 	error = ops->copyin_robust_lists(uap->uaddr1, uap->val, &rb);
4527 	if (error != 0)
4528 		return (error);
4529 
4530 	if (ops->compat32)
4531 		td->td_pflags2 |= TDP2_COMPAT32RB;
4532 
4533 	td->td_rb_list = rb.robust_list_offset;
4534 	td->td_rbp_list = rb.robust_priv_list_offset;
4535 	td->td_rb_inact = rb.robust_inact_offset;
4536 	return (0);
4537 }
4538 
4539 #if defined(__i386__) || defined(__amd64__)
4540 /*
4541  * Provide the standard 32-bit definitions for x86, since native/compat32 use a
4542  * 32-bit time_t there.  Other architectures just need the i386 definitions
4543  * along with their standard compat32.
4544  */
4545 struct timespecx32 {
4546 	int64_t			tv_sec;
4547 	int32_t			tv_nsec;
4548 };
4549 
4550 struct umtx_timex32 {
4551 	struct	timespecx32	_timeout;
4552 	uint32_t		_flags;
4553 	uint32_t		_clockid;
4554 };
4555 
4556 #ifndef __i386__
4557 #define	timespeci386	timespec32
4558 #define	umtx_timei386	umtx_time32
4559 #endif
4560 #else /* !__i386__ && !__amd64__ */
4561 /* 32-bit architectures can emulate i386, so define these almost everywhere. */
4562 struct timespeci386 {
4563 	int32_t			tv_sec;
4564 	int32_t			tv_nsec;
4565 };
4566 
4567 struct umtx_timei386 {
4568 	struct	timespeci386	_timeout;
4569 	uint32_t		_flags;
4570 	uint32_t		_clockid;
4571 };
4572 
4573 #if defined(__LP64__)
4574 #define	timespecx32	timespec32
4575 #define	umtx_timex32	umtx_time32
4576 #endif
4577 #endif
4578 
4579 static int
4580 umtx_copyin_robust_lists32(const void *uaddr, size_t size,
4581     struct umtx_robust_lists_params *rbp)
4582 {
4583 	struct umtx_robust_lists_params_compat32 rb32;
4584 	int error;
4585 
4586 	if (size > sizeof(rb32))
4587 		return (EINVAL);
4588 	bzero(&rb32, sizeof(rb32));
4589 	error = copyin(uaddr, &rb32, size);
4590 	if (error != 0)
4591 		return (error);
4592 	CP(rb32, *rbp, robust_list_offset);
4593 	CP(rb32, *rbp, robust_priv_list_offset);
4594 	CP(rb32, *rbp, robust_inact_offset);
4595 	return (0);
4596 }
4597 
4598 #ifndef __i386__
4599 static inline int
4600 umtx_copyin_timeouti386(const void *uaddr, struct timespec *tsp)
4601 {
4602 	struct timespeci386 ts32;
4603 	int error;
4604 
4605 	error = copyin(uaddr, &ts32, sizeof(ts32));
4606 	if (error == 0) {
4607 		if (ts32.tv_sec < 0 ||
4608 		    ts32.tv_nsec >= 1000000000 ||
4609 		    ts32.tv_nsec < 0)
4610 			error = EINVAL;
4611 		else {
4612 			CP(ts32, *tsp, tv_sec);
4613 			CP(ts32, *tsp, tv_nsec);
4614 		}
4615 	}
4616 	return (error);
4617 }
4618 
4619 static inline int
4620 umtx_copyin_umtx_timei386(const void *uaddr, size_t size, struct _umtx_time *tp)
4621 {
4622 	struct umtx_timei386 t32;
4623 	int error;
4624 
4625 	t32._clockid = CLOCK_REALTIME;
4626 	t32._flags   = 0;
4627 	if (size <= sizeof(t32._timeout))
4628 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4629 	else
4630 		error = copyin(uaddr, &t32, sizeof(t32));
4631 	if (error != 0)
4632 		return (error);
4633 	if (t32._timeout.tv_sec < 0 ||
4634 	    t32._timeout.tv_nsec >= 1000000000 || t32._timeout.tv_nsec < 0)
4635 		return (EINVAL);
4636 	TS_CP(t32, *tp, _timeout);
4637 	CP(t32, *tp, _flags);
4638 	CP(t32, *tp, _clockid);
4639 	return (0);
4640 }
4641 
4642 static int
4643 umtx_copyout_timeouti386(void *uaddr, size_t sz, struct timespec *tsp)
4644 {
4645 	struct timespeci386 remain32 = {
4646 		.tv_sec = tsp->tv_sec,
4647 		.tv_nsec = tsp->tv_nsec,
4648 	};
4649 
4650 	/*
4651 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4652 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4653 	 * copyops.
4654 	 */
4655 	KASSERT(sz >= sizeof(remain32),
4656 	    ("umtx_copyops specifies incorrect sizes"));
4657 
4658 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4659 }
4660 #endif /* !__i386__ */
4661 
4662 #if defined(__i386__) || defined(__LP64__)
4663 static inline int
4664 umtx_copyin_timeoutx32(const void *uaddr, struct timespec *tsp)
4665 {
4666 	struct timespecx32 ts32;
4667 	int error;
4668 
4669 	error = copyin(uaddr, &ts32, sizeof(ts32));
4670 	if (error == 0) {
4671 		if (ts32.tv_sec < 0 ||
4672 		    ts32.tv_nsec >= 1000000000 ||
4673 		    ts32.tv_nsec < 0)
4674 			error = EINVAL;
4675 		else {
4676 			CP(ts32, *tsp, tv_sec);
4677 			CP(ts32, *tsp, tv_nsec);
4678 		}
4679 	}
4680 	return (error);
4681 }
4682 
4683 static inline int
4684 umtx_copyin_umtx_timex32(const void *uaddr, size_t size, struct _umtx_time *tp)
4685 {
4686 	struct umtx_timex32 t32;
4687 	int error;
4688 
4689 	t32._clockid = CLOCK_REALTIME;
4690 	t32._flags   = 0;
4691 	if (size <= sizeof(t32._timeout))
4692 		error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4693 	else
4694 		error = copyin(uaddr, &t32, sizeof(t32));
4695 	if (error != 0)
4696 		return (error);
4697 	if (t32._timeout.tv_sec < 0 ||
4698 	    t32._timeout.tv_nsec >= 1000000000 || t32._timeout.tv_nsec < 0)
4699 		return (EINVAL);
4700 	TS_CP(t32, *tp, _timeout);
4701 	CP(t32, *tp, _flags);
4702 	CP(t32, *tp, _clockid);
4703 	return (0);
4704 }
4705 
4706 static int
4707 umtx_copyout_timeoutx32(void *uaddr, size_t sz, struct timespec *tsp)
4708 {
4709 	struct timespecx32 remain32 = {
4710 		.tv_sec = tsp->tv_sec,
4711 		.tv_nsec = tsp->tv_nsec,
4712 	};
4713 
4714 	/*
4715 	 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4716 	 * and we're only called if sz >= sizeof(timespec) as supplied in the
4717 	 * copyops.
4718 	 */
4719 	KASSERT(sz >= sizeof(remain32),
4720 	    ("umtx_copyops specifies incorrect sizes"));
4721 
4722 	return (copyout(&remain32, uaddr, sizeof(remain32)));
4723 }
4724 #endif /* __i386__ || __LP64__ */
4725 
4726 typedef int (*_umtx_op_func)(struct thread *td, struct _umtx_op_args *uap,
4727     const struct umtx_copyops *umtx_ops);
4728 
4729 static const _umtx_op_func op_table[] = {
4730 #ifdef COMPAT_FREEBSD10
4731 	[UMTX_OP_LOCK]		= __umtx_op_lock_umtx,
4732 	[UMTX_OP_UNLOCK]	= __umtx_op_unlock_umtx,
4733 #else
4734 	[UMTX_OP_LOCK]		= __umtx_op_unimpl,
4735 	[UMTX_OP_UNLOCK]	= __umtx_op_unimpl,
4736 #endif
4737 	[UMTX_OP_WAIT]		= __umtx_op_wait,
4738 	[UMTX_OP_WAKE]		= __umtx_op_wake,
4739 	[UMTX_OP_MUTEX_TRYLOCK]	= __umtx_op_trylock_umutex,
4740 	[UMTX_OP_MUTEX_LOCK]	= __umtx_op_lock_umutex,
4741 	[UMTX_OP_MUTEX_UNLOCK]	= __umtx_op_unlock_umutex,
4742 	[UMTX_OP_SET_CEILING]	= __umtx_op_set_ceiling,
4743 	[UMTX_OP_CV_WAIT]	= __umtx_op_cv_wait,
4744 	[UMTX_OP_CV_SIGNAL]	= __umtx_op_cv_signal,
4745 	[UMTX_OP_CV_BROADCAST]	= __umtx_op_cv_broadcast,
4746 	[UMTX_OP_WAIT_UINT]	= __umtx_op_wait_uint,
4747 	[UMTX_OP_RW_RDLOCK]	= __umtx_op_rw_rdlock,
4748 	[UMTX_OP_RW_WRLOCK]	= __umtx_op_rw_wrlock,
4749 	[UMTX_OP_RW_UNLOCK]	= __umtx_op_rw_unlock,
4750 	[UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private,
4751 	[UMTX_OP_WAKE_PRIVATE]	= __umtx_op_wake_private,
4752 	[UMTX_OP_MUTEX_WAIT]	= __umtx_op_wait_umutex,
4753 	[UMTX_OP_MUTEX_WAKE]	= __umtx_op_wake_umutex,
4754 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4755 	[UMTX_OP_SEM_WAIT]	= __umtx_op_sem_wait,
4756 	[UMTX_OP_SEM_WAKE]	= __umtx_op_sem_wake,
4757 #else
4758 	[UMTX_OP_SEM_WAIT]	= __umtx_op_unimpl,
4759 	[UMTX_OP_SEM_WAKE]	= __umtx_op_unimpl,
4760 #endif
4761 	[UMTX_OP_NWAKE_PRIVATE]	= __umtx_op_nwake_private,
4762 	[UMTX_OP_MUTEX_WAKE2]	= __umtx_op_wake2_umutex,
4763 	[UMTX_OP_SEM2_WAIT]	= __umtx_op_sem2_wait,
4764 	[UMTX_OP_SEM2_WAKE]	= __umtx_op_sem2_wake,
4765 	[UMTX_OP_SHM]		= __umtx_op_shm,
4766 	[UMTX_OP_ROBUST_LISTS]	= __umtx_op_robust_lists,
4767 };
4768 
4769 static const struct umtx_copyops umtx_native_ops = {
4770 	.copyin_timeout = umtx_copyin_timeout,
4771 	.copyin_umtx_time = umtx_copyin_umtx_time,
4772 	.copyin_robust_lists = umtx_copyin_robust_lists,
4773 	.copyout_timeout = umtx_copyout_timeout,
4774 	.timespec_sz = sizeof(struct timespec),
4775 	.umtx_time_sz = sizeof(struct _umtx_time),
4776 };
4777 
4778 #ifndef __i386__
4779 static const struct umtx_copyops umtx_native_opsi386 = {
4780 	.copyin_timeout = umtx_copyin_timeouti386,
4781 	.copyin_umtx_time = umtx_copyin_umtx_timei386,
4782 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4783 	.copyout_timeout = umtx_copyout_timeouti386,
4784 	.timespec_sz = sizeof(struct timespeci386),
4785 	.umtx_time_sz = sizeof(struct umtx_timei386),
4786 	.compat32 = true,
4787 };
4788 #endif
4789 
4790 #if defined(__i386__) || defined(__LP64__)
4791 /* i386 can emulate other 32-bit archs, too! */
4792 static const struct umtx_copyops umtx_native_opsx32 = {
4793 	.copyin_timeout = umtx_copyin_timeoutx32,
4794 	.copyin_umtx_time = umtx_copyin_umtx_timex32,
4795 	.copyin_robust_lists = umtx_copyin_robust_lists32,
4796 	.copyout_timeout = umtx_copyout_timeoutx32,
4797 	.timespec_sz = sizeof(struct timespecx32),
4798 	.umtx_time_sz = sizeof(struct umtx_timex32),
4799 	.compat32 = true,
4800 };
4801 
4802 #ifdef COMPAT_FREEBSD32
4803 #ifdef __amd64__
4804 #define	umtx_native_ops32	umtx_native_opsi386
4805 #else
4806 #define	umtx_native_ops32	umtx_native_opsx32
4807 #endif
4808 #endif /* COMPAT_FREEBSD32 */
4809 #endif /* __i386__ || __LP64__ */
4810 
4811 #define	UMTX_OP__FLAGS	(UMTX_OP__32BIT | UMTX_OP__I386)
4812 
4813 static int
4814 kern__umtx_op(struct thread *td, void *obj, int op, unsigned long val,
4815     void *uaddr1, void *uaddr2, const struct umtx_copyops *ops)
4816 {
4817 	struct _umtx_op_args uap = {
4818 		.obj = obj,
4819 		.op = op & ~UMTX_OP__FLAGS,
4820 		.val = val,
4821 		.uaddr1 = uaddr1,
4822 		.uaddr2 = uaddr2
4823 	};
4824 
4825 	if ((uap.op >= nitems(op_table)))
4826 		return (EINVAL);
4827 	return ((*op_table[uap.op])(td, &uap, ops));
4828 }
4829 
4830 int
4831 sys__umtx_op(struct thread *td, struct _umtx_op_args *uap)
4832 {
4833 	static const struct umtx_copyops *umtx_ops;
4834 
4835 	umtx_ops = &umtx_native_ops;
4836 #ifdef __LP64__
4837 	if ((uap->op & (UMTX_OP__32BIT | UMTX_OP__I386)) != 0) {
4838 		if ((uap->op & UMTX_OP__I386) != 0)
4839 			umtx_ops = &umtx_native_opsi386;
4840 		else
4841 			umtx_ops = &umtx_native_opsx32;
4842 	}
4843 #elif !defined(__i386__)
4844 	/* We consider UMTX_OP__32BIT a nop on !i386 ILP32. */
4845 	if ((uap->op & UMTX_OP__I386) != 0)
4846 		umtx_ops = &umtx_native_opsi386;
4847 #else
4848 	/* Likewise, UMTX_OP__I386 is a nop on i386. */
4849 	if ((uap->op & UMTX_OP__32BIT) != 0)
4850 		umtx_ops = &umtx_native_opsx32;
4851 #endif
4852 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4853 	    uap->uaddr2, umtx_ops));
4854 }
4855 
4856 #ifdef COMPAT_FREEBSD32
4857 #ifdef COMPAT_FREEBSD10
4858 int
4859 freebsd10_freebsd32_umtx_lock(struct thread *td,
4860     struct freebsd10_freebsd32_umtx_lock_args *uap)
4861 {
4862 	return (do_lock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid, NULL));
4863 }
4864 
4865 int
4866 freebsd10_freebsd32_umtx_unlock(struct thread *td,
4867     struct freebsd10_freebsd32_umtx_unlock_args *uap)
4868 {
4869 	return (do_unlock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid));
4870 }
4871 #endif /* COMPAT_FREEBSD10 */
4872 
4873 int
4874 freebsd32__umtx_op(struct thread *td, struct freebsd32__umtx_op_args *uap)
4875 {
4876 
4877 	return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr,
4878 	    uap->uaddr2, &umtx_native_ops32));
4879 }
4880 #endif /* COMPAT_FREEBSD32 */
4881 
4882 void
4883 umtx_thread_init(struct thread *td)
4884 {
4885 
4886 	td->td_umtxq = umtxq_alloc();
4887 	td->td_umtxq->uq_thread = td;
4888 }
4889 
4890 void
4891 umtx_thread_fini(struct thread *td)
4892 {
4893 
4894 	umtxq_free(td->td_umtxq);
4895 }
4896 
4897 /*
4898  * It will be called when new thread is created, e.g fork().
4899  */
4900 void
4901 umtx_thread_alloc(struct thread *td)
4902 {
4903 	struct umtx_q *uq;
4904 
4905 	uq = td->td_umtxq;
4906 	uq->uq_inherited_pri = PRI_MAX;
4907 
4908 	KASSERT(uq->uq_flags == 0, ("uq_flags != 0"));
4909 	KASSERT(uq->uq_thread == td, ("uq_thread != td"));
4910 	KASSERT(uq->uq_pi_blocked == NULL, ("uq_pi_blocked != NULL"));
4911 	KASSERT(TAILQ_EMPTY(&uq->uq_pi_contested), ("uq_pi_contested is not empty"));
4912 }
4913 
4914 /*
4915  * exec() hook.
4916  *
4917  * Clear robust lists for all process' threads, not delaying the
4918  * cleanup to thread exit, since the relevant address space is
4919  * destroyed right now.
4920  */
4921 void
4922 umtx_exec(struct proc *p)
4923 {
4924 	struct thread *td;
4925 
4926 	KASSERT(p == curproc, ("need curproc"));
4927 	KASSERT((p->p_flag & P_HADTHREADS) == 0 ||
4928 	    (p->p_flag & P_STOPPED_SINGLE) != 0,
4929 	    ("curproc must be single-threaded"));
4930 	/*
4931 	 * There is no need to lock the list as only this thread can be
4932 	 * running.
4933 	 */
4934 	FOREACH_THREAD_IN_PROC(p, td) {
4935 		KASSERT(td == curthread ||
4936 		    ((td->td_flags & TDF_BOUNDARY) != 0 && TD_IS_SUSPENDED(td)),
4937 		    ("running thread %p %p", p, td));
4938 		umtx_thread_cleanup(td);
4939 		td->td_rb_list = td->td_rbp_list = td->td_rb_inact = 0;
4940 	}
4941 }
4942 
4943 /*
4944  * thread exit hook.
4945  */
4946 void
4947 umtx_thread_exit(struct thread *td)
4948 {
4949 
4950 	umtx_thread_cleanup(td);
4951 }
4952 
4953 static int
4954 umtx_read_uptr(struct thread *td, uintptr_t ptr, uintptr_t *res, bool compat32)
4955 {
4956 	u_long res1;
4957 	uint32_t res32;
4958 	int error;
4959 
4960 	if (compat32) {
4961 		error = fueword32((void *)ptr, &res32);
4962 		if (error == 0)
4963 			res1 = res32;
4964 	} else {
4965 		error = fueword((void *)ptr, &res1);
4966 	}
4967 	if (error == 0)
4968 		*res = res1;
4969 	else
4970 		error = EFAULT;
4971 	return (error);
4972 }
4973 
4974 static void
4975 umtx_read_rb_list(struct thread *td, struct umutex *m, uintptr_t *rb_list,
4976     bool compat32)
4977 {
4978 	struct umutex32 m32;
4979 
4980 	if (compat32) {
4981 		memcpy(&m32, m, sizeof(m32));
4982 		*rb_list = m32.m_rb_lnk;
4983 	} else {
4984 		*rb_list = m->m_rb_lnk;
4985 	}
4986 }
4987 
4988 static int
4989 umtx_handle_rb(struct thread *td, uintptr_t rbp, uintptr_t *rb_list, bool inact,
4990     bool compat32)
4991 {
4992 	struct umutex m;
4993 	int error;
4994 
4995 	KASSERT(td->td_proc == curproc, ("need current vmspace"));
4996 	error = copyin((void *)rbp, &m, sizeof(m));
4997 	if (error != 0)
4998 		return (error);
4999 	if (rb_list != NULL)
5000 		umtx_read_rb_list(td, &m, rb_list, compat32);
5001 	if ((m.m_flags & UMUTEX_ROBUST) == 0)
5002 		return (EINVAL);
5003 	if ((m.m_owner & ~UMUTEX_CONTESTED) != td->td_tid)
5004 		/* inact is cleared after unlock, allow the inconsistency */
5005 		return (inact ? 0 : EINVAL);
5006 	return (do_unlock_umutex(td, (struct umutex *)rbp, true));
5007 }
5008 
5009 static void
5010 umtx_cleanup_rb_list(struct thread *td, uintptr_t rb_list, uintptr_t *rb_inact,
5011     const char *name, bool compat32)
5012 {
5013 	int error, i;
5014 	uintptr_t rbp;
5015 	bool inact;
5016 
5017 	if (rb_list == 0)
5018 		return;
5019 	error = umtx_read_uptr(td, rb_list, &rbp, compat32);
5020 	for (i = 0; error == 0 && rbp != 0 && i < umtx_max_rb; i++) {
5021 		if (rbp == *rb_inact) {
5022 			inact = true;
5023 			*rb_inact = 0;
5024 		} else
5025 			inact = false;
5026 		error = umtx_handle_rb(td, rbp, &rbp, inact, compat32);
5027 	}
5028 	if (i == umtx_max_rb && umtx_verbose_rb) {
5029 		uprintf("comm %s pid %d: reached umtx %smax rb %d\n",
5030 		    td->td_proc->p_comm, td->td_proc->p_pid, name, umtx_max_rb);
5031 	}
5032 	if (error != 0 && umtx_verbose_rb) {
5033 		uprintf("comm %s pid %d: handling %srb error %d\n",
5034 		    td->td_proc->p_comm, td->td_proc->p_pid, name, error);
5035 	}
5036 }
5037 
5038 /*
5039  * Clean up umtx data.
5040  */
5041 static void
5042 umtx_thread_cleanup(struct thread *td)
5043 {
5044 	struct umtx_q *uq;
5045 	struct umtx_pi *pi;
5046 	uintptr_t rb_inact;
5047 	bool compat32;
5048 
5049 	/*
5050 	 * Disown pi mutexes.
5051 	 */
5052 	uq = td->td_umtxq;
5053 	if (uq != NULL) {
5054 		if (uq->uq_inherited_pri != PRI_MAX ||
5055 		    !TAILQ_EMPTY(&uq->uq_pi_contested)) {
5056 			mtx_lock(&umtx_lock);
5057 			uq->uq_inherited_pri = PRI_MAX;
5058 			while ((pi = TAILQ_FIRST(&uq->uq_pi_contested)) != NULL) {
5059 				pi->pi_owner = NULL;
5060 				TAILQ_REMOVE(&uq->uq_pi_contested, pi, pi_link);
5061 			}
5062 			mtx_unlock(&umtx_lock);
5063 		}
5064 		sched_lend_user_prio_cond(td, PRI_MAX);
5065 	}
5066 
5067 	compat32 = (td->td_pflags2 & TDP2_COMPAT32RB) != 0;
5068 	td->td_pflags2 &= ~TDP2_COMPAT32RB;
5069 
5070 	if (td->td_rb_inact == 0 && td->td_rb_list == 0 && td->td_rbp_list == 0)
5071 		return;
5072 
5073 	/*
5074 	 * Handle terminated robust mutexes.  Must be done after
5075 	 * robust pi disown, otherwise unlock could see unowned
5076 	 * entries.
5077 	 */
5078 	rb_inact = td->td_rb_inact;
5079 	if (rb_inact != 0)
5080 		(void)umtx_read_uptr(td, rb_inact, &rb_inact, compat32);
5081 	umtx_cleanup_rb_list(td, td->td_rb_list, &rb_inact, "", compat32);
5082 	umtx_cleanup_rb_list(td, td->td_rbp_list, &rb_inact, "priv ", compat32);
5083 	if (rb_inact != 0)
5084 		(void)umtx_handle_rb(td, rb_inact, NULL, true, compat32);
5085 }
5086