xref: /freebsd/sys/kern/kern_umtx.c (revision 51f61fc0c7ece7a30c737341e65455841bc3f04e)
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 
67 #include <security/mac/mac_framework.h>
68 
69 #include <vm/vm.h>
70 #include <vm/vm_param.h>
71 #include <vm/pmap.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_object.h>
74 
75 #include <machine/atomic.h>
76 #include <machine/cpu.h>
77 
78 #ifdef COMPAT_FREEBSD32
79 #include <compat/freebsd32/freebsd32_proto.h>
80 #endif
81 
82 #define _UMUTEX_TRY		1
83 #define _UMUTEX_WAIT		2
84 
85 #ifdef UMTX_PROFILING
86 #define	UPROF_PERC_BIGGER(w, f, sw, sf)					\
87 	(((w) > (sw)) || ((w) == (sw) && (f) > (sf)))
88 #endif
89 
90 /* Priority inheritance mutex info. */
91 struct umtx_pi {
92 	/* Owner thread */
93 	struct thread		*pi_owner;
94 
95 	/* Reference count */
96 	int			pi_refcount;
97 
98  	/* List entry to link umtx holding by thread */
99 	TAILQ_ENTRY(umtx_pi)	pi_link;
100 
101 	/* List entry in hash */
102 	TAILQ_ENTRY(umtx_pi)	pi_hashlink;
103 
104 	/* List for waiters */
105 	TAILQ_HEAD(,umtx_q)	pi_blocked;
106 
107 	/* Identify a userland lock object */
108 	struct umtx_key		pi_key;
109 };
110 
111 /* A userland synchronous object user. */
112 struct umtx_q {
113 	/* Linked list for the hash. */
114 	TAILQ_ENTRY(umtx_q)	uq_link;
115 
116 	/* Umtx key. */
117 	struct umtx_key		uq_key;
118 
119 	/* Umtx flags. */
120 	int			uq_flags;
121 #define UQF_UMTXQ	0x0001
122 
123 	/* The thread waits on. */
124 	struct thread		*uq_thread;
125 
126 	/*
127 	 * Blocked on PI mutex. read can use chain lock
128 	 * or umtx_lock, write must have both chain lock and
129 	 * umtx_lock being hold.
130 	 */
131 	struct umtx_pi		*uq_pi_blocked;
132 
133 	/* On blocked list */
134 	TAILQ_ENTRY(umtx_q)	uq_lockq;
135 
136 	/* Thread contending with us */
137 	TAILQ_HEAD(,umtx_pi)	uq_pi_contested;
138 
139 	/* Inherited priority from PP mutex */
140 	u_char			uq_inherited_pri;
141 
142 	/* Spare queue ready to be reused */
143 	struct umtxq_queue	*uq_spare_queue;
144 
145 	/* The queue we on */
146 	struct umtxq_queue	*uq_cur_queue;
147 };
148 
149 TAILQ_HEAD(umtxq_head, umtx_q);
150 
151 /* Per-key wait-queue */
152 struct umtxq_queue {
153 	struct umtxq_head	head;
154 	struct umtx_key		key;
155 	LIST_ENTRY(umtxq_queue)	link;
156 	int			length;
157 };
158 
159 LIST_HEAD(umtxq_list, umtxq_queue);
160 
161 /* Userland lock object's wait-queue chain */
162 struct umtxq_chain {
163 	/* Lock for this chain. */
164 	struct mtx		uc_lock;
165 
166 	/* List of sleep queues. */
167 	struct umtxq_list	uc_queue[2];
168 #define UMTX_SHARED_QUEUE	0
169 #define UMTX_EXCLUSIVE_QUEUE	1
170 
171 	LIST_HEAD(, umtxq_queue) uc_spare_queue;
172 
173 	/* Busy flag */
174 	char			uc_busy;
175 
176 	/* Chain lock waiters */
177 	int			uc_waiters;
178 
179 	/* All PI in the list */
180 	TAILQ_HEAD(,umtx_pi)	uc_pi_list;
181 
182 #ifdef UMTX_PROFILING
183 	u_int 			length;
184 	u_int			max_length;
185 #endif
186 };
187 
188 #define	UMTXQ_LOCKED_ASSERT(uc)		mtx_assert(&(uc)->uc_lock, MA_OWNED)
189 
190 /*
191  * Don't propagate time-sharing priority, there is a security reason,
192  * a user can simply introduce PI-mutex, let thread A lock the mutex,
193  * and let another thread B block on the mutex, because B is
194  * sleeping, its priority will be boosted, this causes A's priority to
195  * be boosted via priority propagating too and will never be lowered even
196  * if it is using 100%CPU, this is unfair to other processes.
197  */
198 
199 #define UPRI(td)	(((td)->td_user_pri >= PRI_MIN_TIMESHARE &&\
200 			  (td)->td_user_pri <= PRI_MAX_TIMESHARE) ?\
201 			 PRI_MAX_TIMESHARE : (td)->td_user_pri)
202 
203 #define	GOLDEN_RATIO_PRIME	2654404609U
204 #ifndef	UMTX_CHAINS
205 #define	UMTX_CHAINS		512
206 #endif
207 #define	UMTX_SHIFTS		(__WORD_BIT - 9)
208 
209 #define	GET_SHARE(flags)	\
210     (((flags) & USYNC_PROCESS_SHARED) == 0 ? THREAD_SHARE : PROCESS_SHARE)
211 
212 #define BUSY_SPINS		200
213 
214 struct abs_timeout {
215 	int clockid;
216 	bool is_abs_real;	/* TIMER_ABSTIME && CLOCK_REALTIME* */
217 	struct timespec cur;
218 	struct timespec end;
219 };
220 
221 #ifdef COMPAT_FREEBSD32
222 struct umutex32 {
223 	volatile __lwpid_t	m_owner;	/* Owner of the mutex */
224 	__uint32_t		m_flags;	/* Flags of the mutex */
225 	__uint32_t		m_ceilings[2];	/* Priority protect ceiling */
226 	__uint32_t		m_rb_lnk;	/* Robust linkage */
227 	__uint32_t		m_pad;
228 	__uint32_t		m_spare[2];
229 };
230 
231 _Static_assert(sizeof(struct umutex) == sizeof(struct umutex32), "umutex32");
232 _Static_assert(__offsetof(struct umutex, m_spare[0]) ==
233     __offsetof(struct umutex32, m_spare[0]), "m_spare32");
234 #endif
235 
236 int umtx_shm_vnobj_persistent = 0;
237 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_vnode_persistent, CTLFLAG_RWTUN,
238     &umtx_shm_vnobj_persistent, 0,
239     "False forces destruction of umtx attached to file, on last close");
240 static int umtx_max_rb = 1000;
241 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_max_robust, CTLFLAG_RWTUN,
242     &umtx_max_rb, 0,
243     "");
244 
245 static uma_zone_t		umtx_pi_zone;
246 static struct umtxq_chain	umtxq_chains[2][UMTX_CHAINS];
247 static MALLOC_DEFINE(M_UMTX, "umtx", "UMTX queue memory");
248 static int			umtx_pi_allocated;
249 
250 static SYSCTL_NODE(_debug, OID_AUTO, umtx, CTLFLAG_RW, 0, "umtx debug");
251 SYSCTL_INT(_debug_umtx, OID_AUTO, umtx_pi_allocated, CTLFLAG_RD,
252     &umtx_pi_allocated, 0, "Allocated umtx_pi");
253 static int umtx_verbose_rb = 1;
254 SYSCTL_INT(_debug_umtx, OID_AUTO, robust_faults_verbose, CTLFLAG_RWTUN,
255     &umtx_verbose_rb, 0,
256     "");
257 
258 #ifdef UMTX_PROFILING
259 static long max_length;
260 SYSCTL_LONG(_debug_umtx, OID_AUTO, max_length, CTLFLAG_RD, &max_length, 0, "max_length");
261 static SYSCTL_NODE(_debug_umtx, OID_AUTO, chains, CTLFLAG_RD, 0, "umtx chain stats");
262 #endif
263 
264 static void abs_timeout_update(struct abs_timeout *timo);
265 
266 static void umtx_shm_init(void);
267 static void umtxq_sysinit(void *);
268 static void umtxq_hash(struct umtx_key *key);
269 static struct umtxq_chain *umtxq_getchain(struct umtx_key *key);
270 static void umtxq_lock(struct umtx_key *key);
271 static void umtxq_unlock(struct umtx_key *key);
272 static void umtxq_busy(struct umtx_key *key);
273 static void umtxq_unbusy(struct umtx_key *key);
274 static void umtxq_insert_queue(struct umtx_q *uq, int q);
275 static void umtxq_remove_queue(struct umtx_q *uq, int q);
276 static int umtxq_sleep(struct umtx_q *uq, const char *wmesg, struct abs_timeout *);
277 static int umtxq_count(struct umtx_key *key);
278 static struct umtx_pi *umtx_pi_alloc(int);
279 static void umtx_pi_free(struct umtx_pi *pi);
280 static int do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags,
281     bool rb);
282 static void umtx_thread_cleanup(struct thread *td);
283 static void umtx_exec_hook(void *arg __unused, struct proc *p __unused,
284     struct image_params *imgp __unused);
285 SYSINIT(umtx, SI_SUB_EVENTHANDLER+1, SI_ORDER_MIDDLE, umtxq_sysinit, NULL);
286 
287 #define umtxq_signal(key, nwake)	umtxq_signal_queue((key), (nwake), UMTX_SHARED_QUEUE)
288 #define umtxq_insert(uq)	umtxq_insert_queue((uq), UMTX_SHARED_QUEUE)
289 #define umtxq_remove(uq)	umtxq_remove_queue((uq), UMTX_SHARED_QUEUE)
290 
291 static struct mtx umtx_lock;
292 
293 #ifdef UMTX_PROFILING
294 static void
295 umtx_init_profiling(void)
296 {
297 	struct sysctl_oid *chain_oid;
298 	char chain_name[10];
299 	int i;
300 
301 	for (i = 0; i < UMTX_CHAINS; ++i) {
302 		snprintf(chain_name, sizeof(chain_name), "%d", i);
303 		chain_oid = SYSCTL_ADD_NODE(NULL,
304 		    SYSCTL_STATIC_CHILDREN(_debug_umtx_chains), OID_AUTO,
305 		    chain_name, CTLFLAG_RD, NULL, "umtx hash stats");
306 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
307 		    "max_length0", CTLFLAG_RD, &umtxq_chains[0][i].max_length, 0, NULL);
308 		SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
309 		    "max_length1", CTLFLAG_RD, &umtxq_chains[1][i].max_length, 0, NULL);
310 	}
311 }
312 
313 static int
314 sysctl_debug_umtx_chains_peaks(SYSCTL_HANDLER_ARGS)
315 {
316 	char buf[512];
317 	struct sbuf sb;
318 	struct umtxq_chain *uc;
319 	u_int fract, i, j, tot, whole;
320 	u_int sf0, sf1, sf2, sf3, sf4;
321 	u_int si0, si1, si2, si3, si4;
322 	u_int sw0, sw1, sw2, sw3, sw4;
323 
324 	sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
325 	for (i = 0; i < 2; i++) {
326 		tot = 0;
327 		for (j = 0; j < UMTX_CHAINS; ++j) {
328 			uc = &umtxq_chains[i][j];
329 			mtx_lock(&uc->uc_lock);
330 			tot += uc->max_length;
331 			mtx_unlock(&uc->uc_lock);
332 		}
333 		if (tot == 0)
334 			sbuf_printf(&sb, "%u) Empty ", i);
335 		else {
336 			sf0 = sf1 = sf2 = sf3 = sf4 = 0;
337 			si0 = si1 = si2 = si3 = si4 = 0;
338 			sw0 = sw1 = sw2 = sw3 = sw4 = 0;
339 			for (j = 0; j < UMTX_CHAINS; j++) {
340 				uc = &umtxq_chains[i][j];
341 				mtx_lock(&uc->uc_lock);
342 				whole = uc->max_length * 100;
343 				mtx_unlock(&uc->uc_lock);
344 				fract = (whole % tot) * 100;
345 				if (UPROF_PERC_BIGGER(whole, fract, sw0, sf0)) {
346 					sf0 = fract;
347 					si0 = j;
348 					sw0 = whole;
349 				} else if (UPROF_PERC_BIGGER(whole, fract, sw1,
350 				    sf1)) {
351 					sf1 = fract;
352 					si1 = j;
353 					sw1 = whole;
354 				} else if (UPROF_PERC_BIGGER(whole, fract, sw2,
355 				    sf2)) {
356 					sf2 = fract;
357 					si2 = j;
358 					sw2 = whole;
359 				} else if (UPROF_PERC_BIGGER(whole, fract, sw3,
360 				    sf3)) {
361 					sf3 = fract;
362 					si3 = j;
363 					sw3 = whole;
364 				} else if (UPROF_PERC_BIGGER(whole, fract, sw4,
365 				    sf4)) {
366 					sf4 = fract;
367 					si4 = j;
368 					sw4 = whole;
369 				}
370 			}
371 			sbuf_printf(&sb, "queue %u:\n", i);
372 			sbuf_printf(&sb, "1st: %u.%u%% idx: %u\n", sw0 / tot,
373 			    sf0 / tot, si0);
374 			sbuf_printf(&sb, "2nd: %u.%u%% idx: %u\n", sw1 / tot,
375 			    sf1 / tot, si1);
376 			sbuf_printf(&sb, "3rd: %u.%u%% idx: %u\n", sw2 / tot,
377 			    sf2 / tot, si2);
378 			sbuf_printf(&sb, "4th: %u.%u%% idx: %u\n", sw3 / tot,
379 			    sf3 / tot, si3);
380 			sbuf_printf(&sb, "5th: %u.%u%% idx: %u\n", sw4 / tot,
381 			    sf4 / tot, si4);
382 		}
383 	}
384 	sbuf_trim(&sb);
385 	sbuf_finish(&sb);
386 	sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
387 	sbuf_delete(&sb);
388 	return (0);
389 }
390 
391 static int
392 sysctl_debug_umtx_chains_clear(SYSCTL_HANDLER_ARGS)
393 {
394 	struct umtxq_chain *uc;
395 	u_int i, j;
396 	int clear, error;
397 
398 	clear = 0;
399 	error = sysctl_handle_int(oidp, &clear, 0, req);
400 	if (error != 0 || req->newptr == NULL)
401 		return (error);
402 
403 	if (clear != 0) {
404 		for (i = 0; i < 2; ++i) {
405 			for (j = 0; j < UMTX_CHAINS; ++j) {
406 				uc = &umtxq_chains[i][j];
407 				mtx_lock(&uc->uc_lock);
408 				uc->length = 0;
409 				uc->max_length = 0;
410 				mtx_unlock(&uc->uc_lock);
411 			}
412 		}
413 	}
414 	return (0);
415 }
416 
417 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, clear,
418     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
419     sysctl_debug_umtx_chains_clear, "I", "Clear umtx chains statistics");
420 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, peaks,
421     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
422     sysctl_debug_umtx_chains_peaks, "A", "Highest peaks in chains max length");
423 #endif
424 
425 static void
426 umtxq_sysinit(void *arg __unused)
427 {
428 	int i, j;
429 
430 	umtx_pi_zone = uma_zcreate("umtx pi", sizeof(struct umtx_pi),
431 		NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
432 	for (i = 0; i < 2; ++i) {
433 		for (j = 0; j < UMTX_CHAINS; ++j) {
434 			mtx_init(&umtxq_chains[i][j].uc_lock, "umtxql", NULL,
435 				 MTX_DEF | MTX_DUPOK);
436 			LIST_INIT(&umtxq_chains[i][j].uc_queue[0]);
437 			LIST_INIT(&umtxq_chains[i][j].uc_queue[1]);
438 			LIST_INIT(&umtxq_chains[i][j].uc_spare_queue);
439 			TAILQ_INIT(&umtxq_chains[i][j].uc_pi_list);
440 			umtxq_chains[i][j].uc_busy = 0;
441 			umtxq_chains[i][j].uc_waiters = 0;
442 #ifdef UMTX_PROFILING
443 			umtxq_chains[i][j].length = 0;
444 			umtxq_chains[i][j].max_length = 0;
445 #endif
446 		}
447 	}
448 #ifdef UMTX_PROFILING
449 	umtx_init_profiling();
450 #endif
451 	mtx_init(&umtx_lock, "umtx lock", NULL, MTX_DEF);
452 	EVENTHANDLER_REGISTER(process_exec, umtx_exec_hook, NULL,
453 	    EVENTHANDLER_PRI_ANY);
454 	umtx_shm_init();
455 }
456 
457 struct umtx_q *
458 umtxq_alloc(void)
459 {
460 	struct umtx_q *uq;
461 
462 	uq = malloc(sizeof(struct umtx_q), M_UMTX, M_WAITOK | M_ZERO);
463 	uq->uq_spare_queue = malloc(sizeof(struct umtxq_queue), M_UMTX,
464 	    M_WAITOK | M_ZERO);
465 	TAILQ_INIT(&uq->uq_spare_queue->head);
466 	TAILQ_INIT(&uq->uq_pi_contested);
467 	uq->uq_inherited_pri = PRI_MAX;
468 	return (uq);
469 }
470 
471 void
472 umtxq_free(struct umtx_q *uq)
473 {
474 
475 	MPASS(uq->uq_spare_queue != NULL);
476 	free(uq->uq_spare_queue, M_UMTX);
477 	free(uq, M_UMTX);
478 }
479 
480 static inline void
481 umtxq_hash(struct umtx_key *key)
482 {
483 	unsigned n;
484 
485 	n = (uintptr_t)key->info.both.a + key->info.both.b;
486 	key->hash = ((n * GOLDEN_RATIO_PRIME) >> UMTX_SHIFTS) % UMTX_CHAINS;
487 }
488 
489 static inline struct umtxq_chain *
490 umtxq_getchain(struct umtx_key *key)
491 {
492 
493 	if (key->type <= TYPE_SEM)
494 		return (&umtxq_chains[1][key->hash]);
495 	return (&umtxq_chains[0][key->hash]);
496 }
497 
498 /*
499  * Lock a chain.
500  */
501 static inline void
502 umtxq_lock(struct umtx_key *key)
503 {
504 	struct umtxq_chain *uc;
505 
506 	uc = umtxq_getchain(key);
507 	mtx_lock(&uc->uc_lock);
508 }
509 
510 /*
511  * Unlock a chain.
512  */
513 static inline void
514 umtxq_unlock(struct umtx_key *key)
515 {
516 	struct umtxq_chain *uc;
517 
518 	uc = umtxq_getchain(key);
519 	mtx_unlock(&uc->uc_lock);
520 }
521 
522 /*
523  * Set chain to busy state when following operation
524  * may be blocked (kernel mutex can not be used).
525  */
526 static inline void
527 umtxq_busy(struct umtx_key *key)
528 {
529 	struct umtxq_chain *uc;
530 
531 	uc = umtxq_getchain(key);
532 	mtx_assert(&uc->uc_lock, MA_OWNED);
533 	if (uc->uc_busy) {
534 #ifdef SMP
535 		if (smp_cpus > 1) {
536 			int count = BUSY_SPINS;
537 			if (count > 0) {
538 				umtxq_unlock(key);
539 				while (uc->uc_busy && --count > 0)
540 					cpu_spinwait();
541 				umtxq_lock(key);
542 			}
543 		}
544 #endif
545 		while (uc->uc_busy) {
546 			uc->uc_waiters++;
547 			msleep(uc, &uc->uc_lock, 0, "umtxqb", 0);
548 			uc->uc_waiters--;
549 		}
550 	}
551 	uc->uc_busy = 1;
552 }
553 
554 /*
555  * Unbusy a chain.
556  */
557 static inline void
558 umtxq_unbusy(struct umtx_key *key)
559 {
560 	struct umtxq_chain *uc;
561 
562 	uc = umtxq_getchain(key);
563 	mtx_assert(&uc->uc_lock, MA_OWNED);
564 	KASSERT(uc->uc_busy != 0, ("not busy"));
565 	uc->uc_busy = 0;
566 	if (uc->uc_waiters)
567 		wakeup_one(uc);
568 }
569 
570 static inline void
571 umtxq_unbusy_unlocked(struct umtx_key *key)
572 {
573 
574 	umtxq_lock(key);
575 	umtxq_unbusy(key);
576 	umtxq_unlock(key);
577 }
578 
579 static struct umtxq_queue *
580 umtxq_queue_lookup(struct umtx_key *key, int q)
581 {
582 	struct umtxq_queue *uh;
583 	struct umtxq_chain *uc;
584 
585 	uc = umtxq_getchain(key);
586 	UMTXQ_LOCKED_ASSERT(uc);
587 	LIST_FOREACH(uh, &uc->uc_queue[q], link) {
588 		if (umtx_key_match(&uh->key, key))
589 			return (uh);
590 	}
591 
592 	return (NULL);
593 }
594 
595 static inline void
596 umtxq_insert_queue(struct umtx_q *uq, int q)
597 {
598 	struct umtxq_queue *uh;
599 	struct umtxq_chain *uc;
600 
601 	uc = umtxq_getchain(&uq->uq_key);
602 	UMTXQ_LOCKED_ASSERT(uc);
603 	KASSERT((uq->uq_flags & UQF_UMTXQ) == 0, ("umtx_q is already on queue"));
604 	uh = umtxq_queue_lookup(&uq->uq_key, q);
605 	if (uh != NULL) {
606 		LIST_INSERT_HEAD(&uc->uc_spare_queue, uq->uq_spare_queue, link);
607 	} else {
608 		uh = uq->uq_spare_queue;
609 		uh->key = uq->uq_key;
610 		LIST_INSERT_HEAD(&uc->uc_queue[q], uh, link);
611 #ifdef UMTX_PROFILING
612 		uc->length++;
613 		if (uc->length > uc->max_length) {
614 			uc->max_length = uc->length;
615 			if (uc->max_length > max_length)
616 				max_length = uc->max_length;
617 		}
618 #endif
619 	}
620 	uq->uq_spare_queue = NULL;
621 
622 	TAILQ_INSERT_TAIL(&uh->head, uq, uq_link);
623 	uh->length++;
624 	uq->uq_flags |= UQF_UMTXQ;
625 	uq->uq_cur_queue = uh;
626 	return;
627 }
628 
629 static inline void
630 umtxq_remove_queue(struct umtx_q *uq, int q)
631 {
632 	struct umtxq_chain *uc;
633 	struct umtxq_queue *uh;
634 
635 	uc = umtxq_getchain(&uq->uq_key);
636 	UMTXQ_LOCKED_ASSERT(uc);
637 	if (uq->uq_flags & UQF_UMTXQ) {
638 		uh = uq->uq_cur_queue;
639 		TAILQ_REMOVE(&uh->head, uq, uq_link);
640 		uh->length--;
641 		uq->uq_flags &= ~UQF_UMTXQ;
642 		if (TAILQ_EMPTY(&uh->head)) {
643 			KASSERT(uh->length == 0,
644 			    ("inconsistent umtxq_queue length"));
645 #ifdef UMTX_PROFILING
646 			uc->length--;
647 #endif
648 			LIST_REMOVE(uh, link);
649 		} else {
650 			uh = LIST_FIRST(&uc->uc_spare_queue);
651 			KASSERT(uh != NULL, ("uc_spare_queue is empty"));
652 			LIST_REMOVE(uh, link);
653 		}
654 		uq->uq_spare_queue = uh;
655 		uq->uq_cur_queue = NULL;
656 	}
657 }
658 
659 /*
660  * Check if there are multiple waiters
661  */
662 static int
663 umtxq_count(struct umtx_key *key)
664 {
665 	struct umtxq_queue *uh;
666 
667 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
668 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
669 	if (uh != NULL)
670 		return (uh->length);
671 	return (0);
672 }
673 
674 /*
675  * Check if there are multiple PI waiters and returns first
676  * waiter.
677  */
678 static int
679 umtxq_count_pi(struct umtx_key *key, struct umtx_q **first)
680 {
681 	struct umtxq_queue *uh;
682 
683 	*first = NULL;
684 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
685 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
686 	if (uh != NULL) {
687 		*first = TAILQ_FIRST(&uh->head);
688 		return (uh->length);
689 	}
690 	return (0);
691 }
692 
693 /*
694  * Check for possible stops and suspensions while executing a umtx
695  * locking operation.
696  *
697  * The sleep argument controls whether the function can handle a stop
698  * request itself or it should return ERESTART and the request is
699  * proceed at the kernel/user boundary in ast.
700  *
701  * Typically, when retrying due to casueword(9) failure (rv == 1), we
702  * should handle the stop requests there, with exception of cases when
703  * the thread busied the umtx key, or when functions return
704  * immediately if umtxq_check_susp() returned non-zero.  On the other
705  * hand, retrying the whole lock operation, we better not stop there
706  * but delegate the handling to ast.
707  *
708  * If the request is for thread termination P_SINGLE_EXIT, we cannot
709  * handle it at all, and simply return EINTR.
710  */
711 static int
712 umtxq_check_susp(struct thread *td, bool sleep)
713 {
714 	struct proc *p;
715 	int error;
716 
717 	/*
718 	 * The check for TDF_NEEDSUSPCHK is racy, but it is enough to
719 	 * eventually break the lockstep loop.
720 	 */
721 	if ((td->td_flags & TDF_NEEDSUSPCHK) == 0)
722 		return (0);
723 	error = 0;
724 	p = td->td_proc;
725 	PROC_LOCK(p);
726 	if (p->p_flag & P_SINGLE_EXIT)
727 		error = EINTR;
728 	else if (P_SHOULDSTOP(p) ||
729 	    ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND)))
730 		error = sleep ? thread_suspend_check(0) : ERESTART;
731 	PROC_UNLOCK(p);
732 	return (error);
733 }
734 
735 /*
736  * Wake up threads waiting on an userland object.
737  */
738 
739 static int
740 umtxq_signal_queue(struct umtx_key *key, int n_wake, int q)
741 {
742 	struct umtxq_queue *uh;
743 	struct umtx_q *uq;
744 	int ret;
745 
746 	ret = 0;
747 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
748 	uh = umtxq_queue_lookup(key, q);
749 	if (uh != NULL) {
750 		while ((uq = TAILQ_FIRST(&uh->head)) != NULL) {
751 			umtxq_remove_queue(uq, q);
752 			wakeup(uq);
753 			if (++ret >= n_wake)
754 				return (ret);
755 		}
756 	}
757 	return (ret);
758 }
759 
760 
761 /*
762  * Wake up specified thread.
763  */
764 static inline void
765 umtxq_signal_thread(struct umtx_q *uq)
766 {
767 
768 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
769 	umtxq_remove(uq);
770 	wakeup(uq);
771 }
772 
773 static inline int
774 tstohz(const struct timespec *tsp)
775 {
776 	struct timeval tv;
777 
778 	TIMESPEC_TO_TIMEVAL(&tv, tsp);
779 	return tvtohz(&tv);
780 }
781 
782 static void
783 abs_timeout_init(struct abs_timeout *timo, int clockid, int absolute,
784 	const struct timespec *timeout)
785 {
786 
787 	timo->clockid = clockid;
788 	if (!absolute) {
789 		timo->is_abs_real = false;
790 		abs_timeout_update(timo);
791 		timespecadd(&timo->cur, timeout, &timo->end);
792 	} else {
793 		timo->end = *timeout;
794 		timo->is_abs_real = clockid == CLOCK_REALTIME ||
795 		    clockid == CLOCK_REALTIME_FAST ||
796 		    clockid == CLOCK_REALTIME_PRECISE;
797 		/*
798 		 * If is_abs_real, umtxq_sleep will read the clock
799 		 * after setting td_rtcgen; otherwise, read it here.
800 		 */
801 		if (!timo->is_abs_real) {
802 			abs_timeout_update(timo);
803 		}
804 	}
805 }
806 
807 static void
808 abs_timeout_init2(struct abs_timeout *timo, const struct _umtx_time *umtxtime)
809 {
810 
811 	abs_timeout_init(timo, umtxtime->_clockid,
812 	    (umtxtime->_flags & UMTX_ABSTIME) != 0, &umtxtime->_timeout);
813 }
814 
815 static inline void
816 abs_timeout_update(struct abs_timeout *timo)
817 {
818 
819 	kern_clock_gettime(curthread, timo->clockid, &timo->cur);
820 }
821 
822 static int
823 abs_timeout_gethz(struct abs_timeout *timo)
824 {
825 	struct timespec tts;
826 
827 	if (timespeccmp(&timo->end, &timo->cur, <=))
828 		return (-1);
829 	timespecsub(&timo->end, &timo->cur, &tts);
830 	return (tstohz(&tts));
831 }
832 
833 static uint32_t
834 umtx_unlock_val(uint32_t flags, bool rb)
835 {
836 
837 	if (rb)
838 		return (UMUTEX_RB_OWNERDEAD);
839 	else if ((flags & UMUTEX_NONCONSISTENT) != 0)
840 		return (UMUTEX_RB_NOTRECOV);
841 	else
842 		return (UMUTEX_UNOWNED);
843 
844 }
845 
846 /*
847  * Put thread into sleep state, before sleeping, check if
848  * thread was removed from umtx queue.
849  */
850 static inline int
851 umtxq_sleep(struct umtx_q *uq, const char *wmesg, struct abs_timeout *abstime)
852 {
853 	struct umtxq_chain *uc;
854 	int error, timo;
855 
856 	if (abstime != NULL && abstime->is_abs_real) {
857 		curthread->td_rtcgen = atomic_load_acq_int(&rtc_generation);
858 		abs_timeout_update(abstime);
859 	}
860 
861 	uc = umtxq_getchain(&uq->uq_key);
862 	UMTXQ_LOCKED_ASSERT(uc);
863 	for (;;) {
864 		if (!(uq->uq_flags & UQF_UMTXQ)) {
865 			error = 0;
866 			break;
867 		}
868 		if (abstime != NULL) {
869 			timo = abs_timeout_gethz(abstime);
870 			if (timo < 0) {
871 				error = ETIMEDOUT;
872 				break;
873 			}
874 		} else
875 			timo = 0;
876 		error = msleep(uq, &uc->uc_lock, PCATCH | PDROP, wmesg, timo);
877 		if (error == EINTR || error == ERESTART) {
878 			umtxq_lock(&uq->uq_key);
879 			break;
880 		}
881 		if (abstime != NULL) {
882 			if (abstime->is_abs_real)
883 				curthread->td_rtcgen =
884 				    atomic_load_acq_int(&rtc_generation);
885 			abs_timeout_update(abstime);
886 		}
887 		umtxq_lock(&uq->uq_key);
888 	}
889 
890 	curthread->td_rtcgen = 0;
891 	return (error);
892 }
893 
894 /*
895  * Convert userspace address into unique logical address.
896  */
897 int
898 umtx_key_get(const void *addr, int type, int share, struct umtx_key *key)
899 {
900 	struct thread *td = curthread;
901 	vm_map_t map;
902 	vm_map_entry_t entry;
903 	vm_pindex_t pindex;
904 	vm_prot_t prot;
905 	boolean_t wired;
906 
907 	key->type = type;
908 	if (share == THREAD_SHARE) {
909 		key->shared = 0;
910 		key->info.private.vs = td->td_proc->p_vmspace;
911 		key->info.private.addr = (uintptr_t)addr;
912 	} else {
913 		MPASS(share == PROCESS_SHARE || share == AUTO_SHARE);
914 		map = &td->td_proc->p_vmspace->vm_map;
915 		if (vm_map_lookup(&map, (vm_offset_t)addr, VM_PROT_WRITE,
916 		    &entry, &key->info.shared.object, &pindex, &prot,
917 		    &wired) != KERN_SUCCESS) {
918 			return (EFAULT);
919 		}
920 
921 		if ((share == PROCESS_SHARE) ||
922 		    (share == AUTO_SHARE &&
923 		     VM_INHERIT_SHARE == entry->inheritance)) {
924 			key->shared = 1;
925 			key->info.shared.offset = (vm_offset_t)addr -
926 			    entry->start + entry->offset;
927 			vm_object_reference(key->info.shared.object);
928 		} else {
929 			key->shared = 0;
930 			key->info.private.vs = td->td_proc->p_vmspace;
931 			key->info.private.addr = (uintptr_t)addr;
932 		}
933 		vm_map_lookup_done(map, entry);
934 	}
935 
936 	umtxq_hash(key);
937 	return (0);
938 }
939 
940 /*
941  * Release key.
942  */
943 void
944 umtx_key_release(struct umtx_key *key)
945 {
946 	if (key->shared)
947 		vm_object_deallocate(key->info.shared.object);
948 }
949 
950 /*
951  * Fetch and compare value, sleep on the address if value is not changed.
952  */
953 static int
954 do_wait(struct thread *td, void *addr, u_long id,
955     struct _umtx_time *timeout, int compat32, int is_private)
956 {
957 	struct abs_timeout timo;
958 	struct umtx_q *uq;
959 	u_long tmp;
960 	uint32_t tmp32;
961 	int error = 0;
962 
963 	uq = td->td_umtxq;
964 	if ((error = umtx_key_get(addr, TYPE_SIMPLE_WAIT,
965 		is_private ? THREAD_SHARE : AUTO_SHARE, &uq->uq_key)) != 0)
966 		return (error);
967 
968 	if (timeout != NULL)
969 		abs_timeout_init2(&timo, timeout);
970 
971 	umtxq_lock(&uq->uq_key);
972 	umtxq_insert(uq);
973 	umtxq_unlock(&uq->uq_key);
974 	if (compat32 == 0) {
975 		error = fueword(addr, &tmp);
976 		if (error != 0)
977 			error = EFAULT;
978 	} else {
979 		error = fueword32(addr, &tmp32);
980 		if (error == 0)
981 			tmp = tmp32;
982 		else
983 			error = EFAULT;
984 	}
985 	umtxq_lock(&uq->uq_key);
986 	if (error == 0) {
987 		if (tmp == id)
988 			error = umtxq_sleep(uq, "uwait", timeout == NULL ?
989 			    NULL : &timo);
990 		if ((uq->uq_flags & UQF_UMTXQ) == 0)
991 			error = 0;
992 		else
993 			umtxq_remove(uq);
994 	} else if ((uq->uq_flags & UQF_UMTXQ) != 0) {
995 		umtxq_remove(uq);
996 	}
997 	umtxq_unlock(&uq->uq_key);
998 	umtx_key_release(&uq->uq_key);
999 	if (error == ERESTART)
1000 		error = EINTR;
1001 	return (error);
1002 }
1003 
1004 /*
1005  * Wake up threads sleeping on the specified address.
1006  */
1007 int
1008 kern_umtx_wake(struct thread *td, void *uaddr, int n_wake, int is_private)
1009 {
1010 	struct umtx_key key;
1011 	int ret;
1012 
1013 	if ((ret = umtx_key_get(uaddr, TYPE_SIMPLE_WAIT,
1014 	    is_private ? THREAD_SHARE : AUTO_SHARE, &key)) != 0)
1015 		return (ret);
1016 	umtxq_lock(&key);
1017 	umtxq_signal(&key, n_wake);
1018 	umtxq_unlock(&key);
1019 	umtx_key_release(&key);
1020 	return (0);
1021 }
1022 
1023 /*
1024  * Lock PTHREAD_PRIO_NONE protocol POSIX mutex.
1025  */
1026 static int
1027 do_lock_normal(struct thread *td, struct umutex *m, uint32_t flags,
1028     struct _umtx_time *timeout, int mode)
1029 {
1030 	struct abs_timeout timo;
1031 	struct umtx_q *uq;
1032 	uint32_t owner, old, id;
1033 	int error, rv;
1034 
1035 	id = td->td_tid;
1036 	uq = td->td_umtxq;
1037 	error = 0;
1038 	if (timeout != NULL)
1039 		abs_timeout_init2(&timo, timeout);
1040 
1041 	/*
1042 	 * Care must be exercised when dealing with umtx structure. It
1043 	 * can fault on any access.
1044 	 */
1045 	for (;;) {
1046 		rv = fueword32(&m->m_owner, &owner);
1047 		if (rv == -1)
1048 			return (EFAULT);
1049 		if (mode == _UMUTEX_WAIT) {
1050 			if (owner == UMUTEX_UNOWNED ||
1051 			    owner == UMUTEX_CONTESTED ||
1052 			    owner == UMUTEX_RB_OWNERDEAD ||
1053 			    owner == UMUTEX_RB_NOTRECOV)
1054 				return (0);
1055 		} else {
1056 			/*
1057 			 * Robust mutex terminated.  Kernel duty is to
1058 			 * return EOWNERDEAD to the userspace.  The
1059 			 * umutex.m_flags UMUTEX_NONCONSISTENT is set
1060 			 * by the common userspace code.
1061 			 */
1062 			if (owner == UMUTEX_RB_OWNERDEAD) {
1063 				rv = casueword32(&m->m_owner,
1064 				    UMUTEX_RB_OWNERDEAD, &owner,
1065 				    id | UMUTEX_CONTESTED);
1066 				if (rv == -1)
1067 					return (EFAULT);
1068 				if (rv == 0) {
1069 					MPASS(owner == UMUTEX_RB_OWNERDEAD);
1070 					return (EOWNERDEAD); /* success */
1071 				}
1072 				MPASS(rv == 1);
1073 				rv = umtxq_check_susp(td, false);
1074 				if (rv != 0)
1075 					return (rv);
1076 				continue;
1077 			}
1078 			if (owner == UMUTEX_RB_NOTRECOV)
1079 				return (ENOTRECOVERABLE);
1080 
1081 			/*
1082 			 * Try the uncontested case.  This should be
1083 			 * done in userland.
1084 			 */
1085 			rv = casueword32(&m->m_owner, UMUTEX_UNOWNED,
1086 			    &owner, id);
1087 			/* The address was invalid. */
1088 			if (rv == -1)
1089 				return (EFAULT);
1090 
1091 			/* The acquire succeeded. */
1092 			if (rv == 0) {
1093 				MPASS(owner == UMUTEX_UNOWNED);
1094 				return (0);
1095 			}
1096 
1097 			/*
1098 			 * If no one owns it but it is contested try
1099 			 * to acquire it.
1100 			 */
1101 			MPASS(rv == 1);
1102 			if (owner == UMUTEX_CONTESTED) {
1103 				rv = casueword32(&m->m_owner,
1104 				    UMUTEX_CONTESTED, &owner,
1105 				    id | UMUTEX_CONTESTED);
1106 				/* The address was invalid. */
1107 				if (rv == -1)
1108 					return (EFAULT);
1109 				if (rv == 0) {
1110 					MPASS(owner == UMUTEX_CONTESTED);
1111 					return (0);
1112 				}
1113 				if (rv == 1) {
1114 					rv = umtxq_check_susp(td, false);
1115 					if (rv != 0)
1116 						return (rv);
1117 				}
1118 
1119 				/*
1120 				 * If this failed the lock has
1121 				 * changed, restart.
1122 				 */
1123 				continue;
1124 			}
1125 
1126 			/* rv == 1 but not contested, likely store failure */
1127 			rv = umtxq_check_susp(td, false);
1128 			if (rv != 0)
1129 				return (rv);
1130 		}
1131 
1132 		if (mode == _UMUTEX_TRY)
1133 			return (EBUSY);
1134 
1135 		/*
1136 		 * If we caught a signal, we have retried and now
1137 		 * exit immediately.
1138 		 */
1139 		if (error != 0)
1140 			return (error);
1141 
1142 		if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX,
1143 		    GET_SHARE(flags), &uq->uq_key)) != 0)
1144 			return (error);
1145 
1146 		umtxq_lock(&uq->uq_key);
1147 		umtxq_busy(&uq->uq_key);
1148 		umtxq_insert(uq);
1149 		umtxq_unlock(&uq->uq_key);
1150 
1151 		/*
1152 		 * Set the contested bit so that a release in user space
1153 		 * knows to use the system call for unlock.  If this fails
1154 		 * either some one else has acquired the lock or it has been
1155 		 * released.
1156 		 */
1157 		rv = casueword32(&m->m_owner, owner, &old,
1158 		    owner | UMUTEX_CONTESTED);
1159 
1160 		/* The address was invalid or casueword failed to store. */
1161 		if (rv == -1 || rv == 1) {
1162 			umtxq_lock(&uq->uq_key);
1163 			umtxq_remove(uq);
1164 			umtxq_unbusy(&uq->uq_key);
1165 			umtxq_unlock(&uq->uq_key);
1166 			umtx_key_release(&uq->uq_key);
1167 			if (rv == -1)
1168 				return (EFAULT);
1169 			if (rv == 1) {
1170 				rv = umtxq_check_susp(td, false);
1171 				if (rv != 0)
1172 					return (rv);
1173 			}
1174 			continue;
1175 		}
1176 
1177 		/*
1178 		 * We set the contested bit, sleep. Otherwise the lock changed
1179 		 * and we need to retry or we lost a race to the thread
1180 		 * unlocking the umtx.
1181 		 */
1182 		umtxq_lock(&uq->uq_key);
1183 		umtxq_unbusy(&uq->uq_key);
1184 		MPASS(old == owner);
1185 		error = umtxq_sleep(uq, "umtxn", timeout == NULL ?
1186 		    NULL : &timo);
1187 		umtxq_remove(uq);
1188 		umtxq_unlock(&uq->uq_key);
1189 		umtx_key_release(&uq->uq_key);
1190 
1191 		if (error == 0)
1192 			error = umtxq_check_susp(td, false);
1193 	}
1194 
1195 	return (0);
1196 }
1197 
1198 /*
1199  * Unlock PTHREAD_PRIO_NONE protocol POSIX mutex.
1200  */
1201 static int
1202 do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1203 {
1204 	struct umtx_key key;
1205 	uint32_t owner, old, id, newlock;
1206 	int error, count;
1207 
1208 	id = td->td_tid;
1209 
1210 again:
1211 	/*
1212 	 * Make sure we own this mtx.
1213 	 */
1214 	error = fueword32(&m->m_owner, &owner);
1215 	if (error == -1)
1216 		return (EFAULT);
1217 
1218 	if ((owner & ~UMUTEX_CONTESTED) != id)
1219 		return (EPERM);
1220 
1221 	newlock = umtx_unlock_val(flags, rb);
1222 	if ((owner & UMUTEX_CONTESTED) == 0) {
1223 		error = casueword32(&m->m_owner, owner, &old, newlock);
1224 		if (error == -1)
1225 			return (EFAULT);
1226 		if (error == 1) {
1227 			error = umtxq_check_susp(td, false);
1228 			if (error != 0)
1229 				return (error);
1230 			goto again;
1231 		}
1232 		MPASS(old == owner);
1233 		return (0);
1234 	}
1235 
1236 	/* We should only ever be in here for contested locks */
1237 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1238 	    &key)) != 0)
1239 		return (error);
1240 
1241 	umtxq_lock(&key);
1242 	umtxq_busy(&key);
1243 	count = umtxq_count(&key);
1244 	umtxq_unlock(&key);
1245 
1246 	/*
1247 	 * When unlocking the umtx, it must be marked as unowned if
1248 	 * there is zero or one thread only waiting for it.
1249 	 * Otherwise, it must be marked as contested.
1250 	 */
1251 	if (count > 1)
1252 		newlock |= UMUTEX_CONTESTED;
1253 	error = casueword32(&m->m_owner, owner, &old, newlock);
1254 	umtxq_lock(&key);
1255 	umtxq_signal(&key, 1);
1256 	umtxq_unbusy(&key);
1257 	umtxq_unlock(&key);
1258 	umtx_key_release(&key);
1259 	if (error == -1)
1260 		return (EFAULT);
1261 	if (error == 1) {
1262 		if (old != owner)
1263 			return (EINVAL);
1264 		error = umtxq_check_susp(td, false);
1265 		if (error != 0)
1266 			return (error);
1267 		goto again;
1268 	}
1269 	return (0);
1270 }
1271 
1272 /*
1273  * Check if the mutex is available and wake up a waiter,
1274  * only for simple mutex.
1275  */
1276 static int
1277 do_wake_umutex(struct thread *td, struct umutex *m)
1278 {
1279 	struct umtx_key key;
1280 	uint32_t owner;
1281 	uint32_t flags;
1282 	int error;
1283 	int count;
1284 
1285 again:
1286 	error = fueword32(&m->m_owner, &owner);
1287 	if (error == -1)
1288 		return (EFAULT);
1289 
1290 	if ((owner & ~UMUTEX_CONTESTED) != 0 && owner != UMUTEX_RB_OWNERDEAD &&
1291 	    owner != UMUTEX_RB_NOTRECOV)
1292 		return (0);
1293 
1294 	error = fueword32(&m->m_flags, &flags);
1295 	if (error == -1)
1296 		return (EFAULT);
1297 
1298 	/* We should only ever be in here for contested locks */
1299 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1300 	    &key)) != 0)
1301 		return (error);
1302 
1303 	umtxq_lock(&key);
1304 	umtxq_busy(&key);
1305 	count = umtxq_count(&key);
1306 	umtxq_unlock(&key);
1307 
1308 	if (count <= 1 && owner != UMUTEX_RB_OWNERDEAD &&
1309 	    owner != UMUTEX_RB_NOTRECOV) {
1310 		error = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
1311 		    UMUTEX_UNOWNED);
1312 		if (error == -1) {
1313 			error = EFAULT;
1314 		} else if (error == 1) {
1315 			umtxq_lock(&key);
1316 			umtxq_unbusy(&key);
1317 			umtxq_unlock(&key);
1318 			umtx_key_release(&key);
1319 			error = umtxq_check_susp(td, false);
1320 			if (error != 0)
1321 				return (error);
1322 			goto again;
1323 		}
1324 	}
1325 
1326 	umtxq_lock(&key);
1327 	if (error == 0 && count != 0) {
1328 		MPASS((owner & ~UMUTEX_CONTESTED) == 0 ||
1329 		    owner == UMUTEX_RB_OWNERDEAD ||
1330 		    owner == UMUTEX_RB_NOTRECOV);
1331 		umtxq_signal(&key, 1);
1332 	}
1333 	umtxq_unbusy(&key);
1334 	umtxq_unlock(&key);
1335 	umtx_key_release(&key);
1336 	return (error);
1337 }
1338 
1339 /*
1340  * Check if the mutex has waiters and tries to fix contention bit.
1341  */
1342 static int
1343 do_wake2_umutex(struct thread *td, struct umutex *m, uint32_t flags)
1344 {
1345 	struct umtx_key key;
1346 	uint32_t owner, old;
1347 	int type;
1348 	int error;
1349 	int count;
1350 
1351 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT |
1352 	    UMUTEX_ROBUST)) {
1353 	case 0:
1354 	case UMUTEX_ROBUST:
1355 		type = TYPE_NORMAL_UMUTEX;
1356 		break;
1357 	case UMUTEX_PRIO_INHERIT:
1358 		type = TYPE_PI_UMUTEX;
1359 		break;
1360 	case (UMUTEX_PRIO_INHERIT | UMUTEX_ROBUST):
1361 		type = TYPE_PI_ROBUST_UMUTEX;
1362 		break;
1363 	case UMUTEX_PRIO_PROTECT:
1364 		type = TYPE_PP_UMUTEX;
1365 		break;
1366 	case (UMUTEX_PRIO_PROTECT | UMUTEX_ROBUST):
1367 		type = TYPE_PP_ROBUST_UMUTEX;
1368 		break;
1369 	default:
1370 		return (EINVAL);
1371 	}
1372 	if ((error = umtx_key_get(m, type, GET_SHARE(flags), &key)) != 0)
1373 		return (error);
1374 
1375 	owner = 0;
1376 	umtxq_lock(&key);
1377 	umtxq_busy(&key);
1378 	count = umtxq_count(&key);
1379 	umtxq_unlock(&key);
1380 
1381 	error = fueword32(&m->m_owner, &owner);
1382 	if (error == -1)
1383 		error = EFAULT;
1384 
1385 	/*
1386 	 * Only repair contention bit if there is a waiter, this means
1387 	 * the mutex is still being referenced by userland code,
1388 	 * otherwise don't update any memory.
1389 	 */
1390 	while (error == 0 && (owner & UMUTEX_CONTESTED) == 0 &&
1391 	    (count > 1 || (count == 1 && (owner & ~UMUTEX_CONTESTED) != 0))) {
1392 		error = casueword32(&m->m_owner, owner, &old,
1393 		    owner | UMUTEX_CONTESTED);
1394 		if (error == -1) {
1395 			error = EFAULT;
1396 			break;
1397 		}
1398 		if (error == 0) {
1399 			MPASS(old == owner);
1400 			break;
1401 		}
1402 		owner = old;
1403 		error = umtxq_check_susp(td, false);
1404 	}
1405 
1406 	umtxq_lock(&key);
1407 	if (error == EFAULT) {
1408 		umtxq_signal(&key, INT_MAX);
1409 	} else if (count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1410 	    owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1411 		umtxq_signal(&key, 1);
1412 	umtxq_unbusy(&key);
1413 	umtxq_unlock(&key);
1414 	umtx_key_release(&key);
1415 	return (error);
1416 }
1417 
1418 static inline struct umtx_pi *
1419 umtx_pi_alloc(int flags)
1420 {
1421 	struct umtx_pi *pi;
1422 
1423 	pi = uma_zalloc(umtx_pi_zone, M_ZERO | flags);
1424 	TAILQ_INIT(&pi->pi_blocked);
1425 	atomic_add_int(&umtx_pi_allocated, 1);
1426 	return (pi);
1427 }
1428 
1429 static inline void
1430 umtx_pi_free(struct umtx_pi *pi)
1431 {
1432 	uma_zfree(umtx_pi_zone, pi);
1433 	atomic_add_int(&umtx_pi_allocated, -1);
1434 }
1435 
1436 /*
1437  * Adjust the thread's position on a pi_state after its priority has been
1438  * changed.
1439  */
1440 static int
1441 umtx_pi_adjust_thread(struct umtx_pi *pi, struct thread *td)
1442 {
1443 	struct umtx_q *uq, *uq1, *uq2;
1444 	struct thread *td1;
1445 
1446 	mtx_assert(&umtx_lock, MA_OWNED);
1447 	if (pi == NULL)
1448 		return (0);
1449 
1450 	uq = td->td_umtxq;
1451 
1452 	/*
1453 	 * Check if the thread needs to be moved on the blocked chain.
1454 	 * It needs to be moved if either its priority is lower than
1455 	 * the previous thread or higher than the next thread.
1456 	 */
1457 	uq1 = TAILQ_PREV(uq, umtxq_head, uq_lockq);
1458 	uq2 = TAILQ_NEXT(uq, uq_lockq);
1459 	if ((uq1 != NULL && UPRI(td) < UPRI(uq1->uq_thread)) ||
1460 	    (uq2 != NULL && UPRI(td) > UPRI(uq2->uq_thread))) {
1461 		/*
1462 		 * Remove thread from blocked chain and determine where
1463 		 * it should be moved to.
1464 		 */
1465 		TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1466 		TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1467 			td1 = uq1->uq_thread;
1468 			MPASS(td1->td_proc->p_magic == P_MAGIC);
1469 			if (UPRI(td1) > UPRI(td))
1470 				break;
1471 		}
1472 
1473 		if (uq1 == NULL)
1474 			TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1475 		else
1476 			TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1477 	}
1478 	return (1);
1479 }
1480 
1481 static struct umtx_pi *
1482 umtx_pi_next(struct umtx_pi *pi)
1483 {
1484 	struct umtx_q *uq_owner;
1485 
1486 	if (pi->pi_owner == NULL)
1487 		return (NULL);
1488 	uq_owner = pi->pi_owner->td_umtxq;
1489 	if (uq_owner == NULL)
1490 		return (NULL);
1491 	return (uq_owner->uq_pi_blocked);
1492 }
1493 
1494 /*
1495  * Floyd's Cycle-Finding Algorithm.
1496  */
1497 static bool
1498 umtx_pi_check_loop(struct umtx_pi *pi)
1499 {
1500 	struct umtx_pi *pi1;	/* fast iterator */
1501 
1502 	mtx_assert(&umtx_lock, MA_OWNED);
1503 	if (pi == NULL)
1504 		return (false);
1505 	pi1 = pi;
1506 	for (;;) {
1507 		pi = umtx_pi_next(pi);
1508 		if (pi == NULL)
1509 			break;
1510 		pi1 = umtx_pi_next(pi1);
1511 		if (pi1 == NULL)
1512 			break;
1513 		pi1 = umtx_pi_next(pi1);
1514 		if (pi1 == NULL)
1515 			break;
1516 		if (pi == pi1)
1517 			return (true);
1518 	}
1519 	return (false);
1520 }
1521 
1522 /*
1523  * Propagate priority when a thread is blocked on POSIX
1524  * PI mutex.
1525  */
1526 static void
1527 umtx_propagate_priority(struct thread *td)
1528 {
1529 	struct umtx_q *uq;
1530 	struct umtx_pi *pi;
1531 	int pri;
1532 
1533 	mtx_assert(&umtx_lock, MA_OWNED);
1534 	pri = UPRI(td);
1535 	uq = td->td_umtxq;
1536 	pi = uq->uq_pi_blocked;
1537 	if (pi == NULL)
1538 		return;
1539 	if (umtx_pi_check_loop(pi))
1540 		return;
1541 
1542 	for (;;) {
1543 		td = pi->pi_owner;
1544 		if (td == NULL || td == curthread)
1545 			return;
1546 
1547 		MPASS(td->td_proc != NULL);
1548 		MPASS(td->td_proc->p_magic == P_MAGIC);
1549 
1550 		thread_lock(td);
1551 		if (td->td_lend_user_pri > pri)
1552 			sched_lend_user_prio(td, pri);
1553 		else {
1554 			thread_unlock(td);
1555 			break;
1556 		}
1557 		thread_unlock(td);
1558 
1559 		/*
1560 		 * Pick up the lock that td is blocked on.
1561 		 */
1562 		uq = td->td_umtxq;
1563 		pi = uq->uq_pi_blocked;
1564 		if (pi == NULL)
1565 			break;
1566 		/* Resort td on the list if needed. */
1567 		umtx_pi_adjust_thread(pi, td);
1568 	}
1569 }
1570 
1571 /*
1572  * Unpropagate priority for a PI mutex when a thread blocked on
1573  * it is interrupted by signal or resumed by others.
1574  */
1575 static void
1576 umtx_repropagate_priority(struct umtx_pi *pi)
1577 {
1578 	struct umtx_q *uq, *uq_owner;
1579 	struct umtx_pi *pi2;
1580 	int pri;
1581 
1582 	mtx_assert(&umtx_lock, MA_OWNED);
1583 
1584 	if (umtx_pi_check_loop(pi))
1585 		return;
1586 	while (pi != NULL && pi->pi_owner != NULL) {
1587 		pri = PRI_MAX;
1588 		uq_owner = pi->pi_owner->td_umtxq;
1589 
1590 		TAILQ_FOREACH(pi2, &uq_owner->uq_pi_contested, pi_link) {
1591 			uq = TAILQ_FIRST(&pi2->pi_blocked);
1592 			if (uq != NULL) {
1593 				if (pri > UPRI(uq->uq_thread))
1594 					pri = UPRI(uq->uq_thread);
1595 			}
1596 		}
1597 
1598 		if (pri > uq_owner->uq_inherited_pri)
1599 			pri = uq_owner->uq_inherited_pri;
1600 		thread_lock(pi->pi_owner);
1601 		sched_lend_user_prio(pi->pi_owner, pri);
1602 		thread_unlock(pi->pi_owner);
1603 		if ((pi = uq_owner->uq_pi_blocked) != NULL)
1604 			umtx_pi_adjust_thread(pi, uq_owner->uq_thread);
1605 	}
1606 }
1607 
1608 /*
1609  * Insert a PI mutex into owned list.
1610  */
1611 static void
1612 umtx_pi_setowner(struct umtx_pi *pi, struct thread *owner)
1613 {
1614 	struct umtx_q *uq_owner;
1615 
1616 	uq_owner = owner->td_umtxq;
1617 	mtx_assert(&umtx_lock, MA_OWNED);
1618 	MPASS(pi->pi_owner == NULL);
1619 	pi->pi_owner = owner;
1620 	TAILQ_INSERT_TAIL(&uq_owner->uq_pi_contested, pi, pi_link);
1621 }
1622 
1623 
1624 /*
1625  * Disown a PI mutex, and remove it from the owned list.
1626  */
1627 static void
1628 umtx_pi_disown(struct umtx_pi *pi)
1629 {
1630 
1631 	mtx_assert(&umtx_lock, MA_OWNED);
1632 	TAILQ_REMOVE(&pi->pi_owner->td_umtxq->uq_pi_contested, pi, pi_link);
1633 	pi->pi_owner = NULL;
1634 }
1635 
1636 /*
1637  * Claim ownership of a PI mutex.
1638  */
1639 static int
1640 umtx_pi_claim(struct umtx_pi *pi, struct thread *owner)
1641 {
1642 	struct umtx_q *uq;
1643 	int pri;
1644 
1645 	mtx_lock(&umtx_lock);
1646 	if (pi->pi_owner == owner) {
1647 		mtx_unlock(&umtx_lock);
1648 		return (0);
1649 	}
1650 
1651 	if (pi->pi_owner != NULL) {
1652 		/*
1653 		 * userland may have already messed the mutex, sigh.
1654 		 */
1655 		mtx_unlock(&umtx_lock);
1656 		return (EPERM);
1657 	}
1658 	umtx_pi_setowner(pi, owner);
1659 	uq = TAILQ_FIRST(&pi->pi_blocked);
1660 	if (uq != NULL) {
1661 		pri = UPRI(uq->uq_thread);
1662 		thread_lock(owner);
1663 		if (pri < UPRI(owner))
1664 			sched_lend_user_prio(owner, pri);
1665 		thread_unlock(owner);
1666 	}
1667 	mtx_unlock(&umtx_lock);
1668 	return (0);
1669 }
1670 
1671 /*
1672  * Adjust a thread's order position in its blocked PI mutex,
1673  * this may result new priority propagating process.
1674  */
1675 void
1676 umtx_pi_adjust(struct thread *td, u_char oldpri)
1677 {
1678 	struct umtx_q *uq;
1679 	struct umtx_pi *pi;
1680 
1681 	uq = td->td_umtxq;
1682 	mtx_lock(&umtx_lock);
1683 	/*
1684 	 * Pick up the lock that td is blocked on.
1685 	 */
1686 	pi = uq->uq_pi_blocked;
1687 	if (pi != NULL) {
1688 		umtx_pi_adjust_thread(pi, td);
1689 		umtx_repropagate_priority(pi);
1690 	}
1691 	mtx_unlock(&umtx_lock);
1692 }
1693 
1694 /*
1695  * Sleep on a PI mutex.
1696  */
1697 static int
1698 umtxq_sleep_pi(struct umtx_q *uq, struct umtx_pi *pi, uint32_t owner,
1699     const char *wmesg, struct abs_timeout *timo, bool shared)
1700 {
1701 	struct thread *td, *td1;
1702 	struct umtx_q *uq1;
1703 	int error, pri;
1704 #ifdef INVARIANTS
1705 	struct umtxq_chain *uc;
1706 
1707 	uc = umtxq_getchain(&pi->pi_key);
1708 #endif
1709 	error = 0;
1710 	td = uq->uq_thread;
1711 	KASSERT(td == curthread, ("inconsistent uq_thread"));
1712 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
1713 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));
1714 	umtxq_insert(uq);
1715 	mtx_lock(&umtx_lock);
1716 	if (pi->pi_owner == NULL) {
1717 		mtx_unlock(&umtx_lock);
1718 		td1 = tdfind(owner, shared ? -1 : td->td_proc->p_pid);
1719 		mtx_lock(&umtx_lock);
1720 		if (td1 != NULL) {
1721 			if (pi->pi_owner == NULL)
1722 				umtx_pi_setowner(pi, td1);
1723 			PROC_UNLOCK(td1->td_proc);
1724 		}
1725 	}
1726 
1727 	TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1728 		pri = UPRI(uq1->uq_thread);
1729 		if (pri > UPRI(td))
1730 			break;
1731 	}
1732 
1733 	if (uq1 != NULL)
1734 		TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1735 	else
1736 		TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1737 
1738 	uq->uq_pi_blocked = pi;
1739 	thread_lock(td);
1740 	td->td_flags |= TDF_UPIBLOCKED;
1741 	thread_unlock(td);
1742 	umtx_propagate_priority(td);
1743 	mtx_unlock(&umtx_lock);
1744 	umtxq_unbusy(&uq->uq_key);
1745 
1746 	error = umtxq_sleep(uq, wmesg, timo);
1747 	umtxq_remove(uq);
1748 
1749 	mtx_lock(&umtx_lock);
1750 	uq->uq_pi_blocked = NULL;
1751 	thread_lock(td);
1752 	td->td_flags &= ~TDF_UPIBLOCKED;
1753 	thread_unlock(td);
1754 	TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1755 	umtx_repropagate_priority(pi);
1756 	mtx_unlock(&umtx_lock);
1757 	umtxq_unlock(&uq->uq_key);
1758 
1759 	return (error);
1760 }
1761 
1762 /*
1763  * Add reference count for a PI mutex.
1764  */
1765 static void
1766 umtx_pi_ref(struct umtx_pi *pi)
1767 {
1768 
1769 	UMTXQ_LOCKED_ASSERT(umtxq_getchain(&pi->pi_key));
1770 	pi->pi_refcount++;
1771 }
1772 
1773 /*
1774  * Decrease reference count for a PI mutex, if the counter
1775  * is decreased to zero, its memory space is freed.
1776  */
1777 static void
1778 umtx_pi_unref(struct umtx_pi *pi)
1779 {
1780 	struct umtxq_chain *uc;
1781 
1782 	uc = umtxq_getchain(&pi->pi_key);
1783 	UMTXQ_LOCKED_ASSERT(uc);
1784 	KASSERT(pi->pi_refcount > 0, ("invalid reference count"));
1785 	if (--pi->pi_refcount == 0) {
1786 		mtx_lock(&umtx_lock);
1787 		if (pi->pi_owner != NULL)
1788 			umtx_pi_disown(pi);
1789 		KASSERT(TAILQ_EMPTY(&pi->pi_blocked),
1790 			("blocked queue not empty"));
1791 		mtx_unlock(&umtx_lock);
1792 		TAILQ_REMOVE(&uc->uc_pi_list, pi, pi_hashlink);
1793 		umtx_pi_free(pi);
1794 	}
1795 }
1796 
1797 /*
1798  * Find a PI mutex in hash table.
1799  */
1800 static struct umtx_pi *
1801 umtx_pi_lookup(struct umtx_key *key)
1802 {
1803 	struct umtxq_chain *uc;
1804 	struct umtx_pi *pi;
1805 
1806 	uc = umtxq_getchain(key);
1807 	UMTXQ_LOCKED_ASSERT(uc);
1808 
1809 	TAILQ_FOREACH(pi, &uc->uc_pi_list, pi_hashlink) {
1810 		if (umtx_key_match(&pi->pi_key, key)) {
1811 			return (pi);
1812 		}
1813 	}
1814 	return (NULL);
1815 }
1816 
1817 /*
1818  * Insert a PI mutex into hash table.
1819  */
1820 static inline void
1821 umtx_pi_insert(struct umtx_pi *pi)
1822 {
1823 	struct umtxq_chain *uc;
1824 
1825 	uc = umtxq_getchain(&pi->pi_key);
1826 	UMTXQ_LOCKED_ASSERT(uc);
1827 	TAILQ_INSERT_TAIL(&uc->uc_pi_list, pi, pi_hashlink);
1828 }
1829 
1830 /*
1831  * Lock a PI mutex.
1832  */
1833 static int
1834 do_lock_pi(struct thread *td, struct umutex *m, uint32_t flags,
1835     struct _umtx_time *timeout, int try)
1836 {
1837 	struct abs_timeout timo;
1838 	struct umtx_q *uq;
1839 	struct umtx_pi *pi, *new_pi;
1840 	uint32_t id, old_owner, owner, old;
1841 	int error, rv;
1842 
1843 	id = td->td_tid;
1844 	uq = td->td_umtxq;
1845 
1846 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
1847 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
1848 	    &uq->uq_key)) != 0)
1849 		return (error);
1850 
1851 	if (timeout != NULL)
1852 		abs_timeout_init2(&timo, timeout);
1853 
1854 	umtxq_lock(&uq->uq_key);
1855 	pi = umtx_pi_lookup(&uq->uq_key);
1856 	if (pi == NULL) {
1857 		new_pi = umtx_pi_alloc(M_NOWAIT);
1858 		if (new_pi == NULL) {
1859 			umtxq_unlock(&uq->uq_key);
1860 			new_pi = umtx_pi_alloc(M_WAITOK);
1861 			umtxq_lock(&uq->uq_key);
1862 			pi = umtx_pi_lookup(&uq->uq_key);
1863 			if (pi != NULL) {
1864 				umtx_pi_free(new_pi);
1865 				new_pi = NULL;
1866 			}
1867 		}
1868 		if (new_pi != NULL) {
1869 			new_pi->pi_key = uq->uq_key;
1870 			umtx_pi_insert(new_pi);
1871 			pi = new_pi;
1872 		}
1873 	}
1874 	umtx_pi_ref(pi);
1875 	umtxq_unlock(&uq->uq_key);
1876 
1877 	/*
1878 	 * Care must be exercised when dealing with umtx structure.  It
1879 	 * can fault on any access.
1880 	 */
1881 	for (;;) {
1882 		/*
1883 		 * Try the uncontested case.  This should be done in userland.
1884 		 */
1885 		rv = casueword32(&m->m_owner, UMUTEX_UNOWNED, &owner, id);
1886 		/* The address was invalid. */
1887 		if (rv == -1) {
1888 			error = EFAULT;
1889 			break;
1890 		}
1891 		/* The acquire succeeded. */
1892 		if (rv == 0) {
1893 			MPASS(owner == UMUTEX_UNOWNED);
1894 			error = 0;
1895 			break;
1896 		}
1897 
1898 		if (owner == UMUTEX_RB_NOTRECOV) {
1899 			error = ENOTRECOVERABLE;
1900 			break;
1901 		}
1902 
1903 		/*
1904 		 * Avoid overwriting a possible error from sleep due
1905 		 * to the pending signal with suspension check result.
1906 		 */
1907 		if (error == 0) {
1908 			error = umtxq_check_susp(td, true);
1909 			if (error != 0)
1910 				break;
1911 		}
1912 
1913 		/* If no one owns it but it is contested try to acquire it. */
1914 		if (owner == UMUTEX_CONTESTED || owner == UMUTEX_RB_OWNERDEAD) {
1915 			old_owner = owner;
1916 			rv = casueword32(&m->m_owner, owner, &owner,
1917 			    id | UMUTEX_CONTESTED);
1918 			/* The address was invalid. */
1919 			if (rv == -1) {
1920 				error = EFAULT;
1921 				break;
1922 			}
1923 			if (rv == 1) {
1924 				if (error == 0) {
1925 					error = umtxq_check_susp(td, true);
1926 					if (error != 0)
1927 						break;
1928 				}
1929 
1930 				/*
1931 				 * If this failed the lock could
1932 				 * changed, restart.
1933 				 */
1934 				continue;
1935 			}
1936 
1937 			MPASS(rv == 0);
1938 			MPASS(owner == old_owner);
1939 			umtxq_lock(&uq->uq_key);
1940 			umtxq_busy(&uq->uq_key);
1941 			error = umtx_pi_claim(pi, td);
1942 			umtxq_unbusy(&uq->uq_key);
1943 			umtxq_unlock(&uq->uq_key);
1944 			if (error != 0) {
1945 				/*
1946 				 * Since we're going to return an
1947 				 * error, restore the m_owner to its
1948 				 * previous, unowned state to avoid
1949 				 * compounding the problem.
1950 				 */
1951 				(void)casuword32(&m->m_owner,
1952 				    id | UMUTEX_CONTESTED, old_owner);
1953 			}
1954 			if (error == 0 && old_owner == UMUTEX_RB_OWNERDEAD)
1955 				error = EOWNERDEAD;
1956 			break;
1957 		}
1958 
1959 		if ((owner & ~UMUTEX_CONTESTED) == id) {
1960 			error = EDEADLK;
1961 			break;
1962 		}
1963 
1964 		if (try != 0) {
1965 			error = EBUSY;
1966 			break;
1967 		}
1968 
1969 		/*
1970 		 * If we caught a signal, we have retried and now
1971 		 * exit immediately.
1972 		 */
1973 		if (error != 0)
1974 			break;
1975 
1976 		umtxq_lock(&uq->uq_key);
1977 		umtxq_busy(&uq->uq_key);
1978 		umtxq_unlock(&uq->uq_key);
1979 
1980 		/*
1981 		 * Set the contested bit so that a release in user space
1982 		 * knows to use the system call for unlock.  If this fails
1983 		 * either some one else has acquired the lock or it has been
1984 		 * released.
1985 		 */
1986 		rv = casueword32(&m->m_owner, owner, &old, owner |
1987 		    UMUTEX_CONTESTED);
1988 
1989 		/* The address was invalid. */
1990 		if (rv == -1) {
1991 			umtxq_unbusy_unlocked(&uq->uq_key);
1992 			error = EFAULT;
1993 			break;
1994 		}
1995 		if (rv == 1) {
1996 			umtxq_unbusy_unlocked(&uq->uq_key);
1997 			error = umtxq_check_susp(td, true);
1998 			if (error != 0)
1999 				break;
2000 
2001 			/*
2002 			 * The lock changed and we need to retry or we
2003 			 * lost a race to the thread unlocking the
2004 			 * umtx.  Note that the UMUTEX_RB_OWNERDEAD
2005 			 * value for owner is impossible there.
2006 			 */
2007 			continue;
2008 		}
2009 
2010 		umtxq_lock(&uq->uq_key);
2011 
2012 		/* We set the contested bit, sleep. */
2013 		MPASS(old == owner);
2014 		error = umtxq_sleep_pi(uq, pi, owner & ~UMUTEX_CONTESTED,
2015 		    "umtxpi", timeout == NULL ? NULL : &timo,
2016 		    (flags & USYNC_PROCESS_SHARED) != 0);
2017 		if (error != 0)
2018 			continue;
2019 
2020 		error = umtxq_check_susp(td, false);
2021 		if (error != 0)
2022 			break;
2023 	}
2024 
2025 	umtxq_lock(&uq->uq_key);
2026 	umtx_pi_unref(pi);
2027 	umtxq_unlock(&uq->uq_key);
2028 
2029 	umtx_key_release(&uq->uq_key);
2030 	return (error);
2031 }
2032 
2033 /*
2034  * Unlock a PI mutex.
2035  */
2036 static int
2037 do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2038 {
2039 	struct umtx_key key;
2040 	struct umtx_q *uq_first, *uq_first2, *uq_me;
2041 	struct umtx_pi *pi, *pi2;
2042 	uint32_t id, new_owner, old, owner;
2043 	int count, error, pri;
2044 
2045 	id = td->td_tid;
2046 
2047 usrloop:
2048 	/*
2049 	 * Make sure we own this mtx.
2050 	 */
2051 	error = fueword32(&m->m_owner, &owner);
2052 	if (error == -1)
2053 		return (EFAULT);
2054 
2055 	if ((owner & ~UMUTEX_CONTESTED) != id)
2056 		return (EPERM);
2057 
2058 	new_owner = umtx_unlock_val(flags, rb);
2059 
2060 	/* This should be done in userland */
2061 	if ((owner & UMUTEX_CONTESTED) == 0) {
2062 		error = casueword32(&m->m_owner, owner, &old, new_owner);
2063 		if (error == -1)
2064 			return (EFAULT);
2065 		if (error == 1) {
2066 			error = umtxq_check_susp(td, true);
2067 			if (error != 0)
2068 				return (error);
2069 			goto usrloop;
2070 		}
2071 		if (old == owner)
2072 			return (0);
2073 		owner = old;
2074 	}
2075 
2076 	/* We should only ever be in here for contested locks */
2077 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2078 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2079 	    &key)) != 0)
2080 		return (error);
2081 
2082 	umtxq_lock(&key);
2083 	umtxq_busy(&key);
2084 	count = umtxq_count_pi(&key, &uq_first);
2085 	if (uq_first != NULL) {
2086 		mtx_lock(&umtx_lock);
2087 		pi = uq_first->uq_pi_blocked;
2088 		KASSERT(pi != NULL, ("pi == NULL?"));
2089 		if (pi->pi_owner != td && !(rb && pi->pi_owner == NULL)) {
2090 			mtx_unlock(&umtx_lock);
2091 			umtxq_unbusy(&key);
2092 			umtxq_unlock(&key);
2093 			umtx_key_release(&key);
2094 			/* userland messed the mutex */
2095 			return (EPERM);
2096 		}
2097 		uq_me = td->td_umtxq;
2098 		if (pi->pi_owner == td)
2099 			umtx_pi_disown(pi);
2100 		/* get highest priority thread which is still sleeping. */
2101 		uq_first = TAILQ_FIRST(&pi->pi_blocked);
2102 		while (uq_first != NULL &&
2103 		    (uq_first->uq_flags & UQF_UMTXQ) == 0) {
2104 			uq_first = TAILQ_NEXT(uq_first, uq_lockq);
2105 		}
2106 		pri = PRI_MAX;
2107 		TAILQ_FOREACH(pi2, &uq_me->uq_pi_contested, pi_link) {
2108 			uq_first2 = TAILQ_FIRST(&pi2->pi_blocked);
2109 			if (uq_first2 != NULL) {
2110 				if (pri > UPRI(uq_first2->uq_thread))
2111 					pri = UPRI(uq_first2->uq_thread);
2112 			}
2113 		}
2114 		thread_lock(td);
2115 		sched_lend_user_prio(td, pri);
2116 		thread_unlock(td);
2117 		mtx_unlock(&umtx_lock);
2118 		if (uq_first)
2119 			umtxq_signal_thread(uq_first);
2120 	} else {
2121 		pi = umtx_pi_lookup(&key);
2122 		/*
2123 		 * A umtx_pi can exist if a signal or timeout removed the
2124 		 * last waiter from the umtxq, but there is still
2125 		 * a thread in do_lock_pi() holding the umtx_pi.
2126 		 */
2127 		if (pi != NULL) {
2128 			/*
2129 			 * The umtx_pi can be unowned, such as when a thread
2130 			 * has just entered do_lock_pi(), allocated the
2131 			 * umtx_pi, and unlocked the umtxq.
2132 			 * If the current thread owns it, it must disown it.
2133 			 */
2134 			mtx_lock(&umtx_lock);
2135 			if (pi->pi_owner == td)
2136 				umtx_pi_disown(pi);
2137 			mtx_unlock(&umtx_lock);
2138 		}
2139 	}
2140 	umtxq_unlock(&key);
2141 
2142 	/*
2143 	 * When unlocking the umtx, it must be marked as unowned if
2144 	 * there is zero or one thread only waiting for it.
2145 	 * Otherwise, it must be marked as contested.
2146 	 */
2147 
2148 	if (count > 1)
2149 		new_owner |= UMUTEX_CONTESTED;
2150 again:
2151 	error = casueword32(&m->m_owner, owner, &old, new_owner);
2152 	if (error == 1) {
2153 		error = umtxq_check_susp(td, false);
2154 		if (error == 0)
2155 			goto again;
2156 	}
2157 	umtxq_unbusy_unlocked(&key);
2158 	umtx_key_release(&key);
2159 	if (error == -1)
2160 		return (EFAULT);
2161 	if (error == 0 && old != owner)
2162 		return (EINVAL);
2163 	return (error);
2164 }
2165 
2166 /*
2167  * Lock a PP mutex.
2168  */
2169 static int
2170 do_lock_pp(struct thread *td, struct umutex *m, uint32_t flags,
2171     struct _umtx_time *timeout, int try)
2172 {
2173 	struct abs_timeout timo;
2174 	struct umtx_q *uq, *uq2;
2175 	struct umtx_pi *pi;
2176 	uint32_t ceiling;
2177 	uint32_t owner, id;
2178 	int error, pri, old_inherited_pri, su, rv;
2179 
2180 	id = td->td_tid;
2181 	uq = td->td_umtxq;
2182 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2183 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2184 	    &uq->uq_key)) != 0)
2185 		return (error);
2186 
2187 	if (timeout != NULL)
2188 		abs_timeout_init2(&timo, timeout);
2189 
2190 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2191 	for (;;) {
2192 		old_inherited_pri = uq->uq_inherited_pri;
2193 		umtxq_lock(&uq->uq_key);
2194 		umtxq_busy(&uq->uq_key);
2195 		umtxq_unlock(&uq->uq_key);
2196 
2197 		rv = fueword32(&m->m_ceilings[0], &ceiling);
2198 		if (rv == -1) {
2199 			error = EFAULT;
2200 			goto out;
2201 		}
2202 		ceiling = RTP_PRIO_MAX - ceiling;
2203 		if (ceiling > RTP_PRIO_MAX) {
2204 			error = EINVAL;
2205 			goto out;
2206 		}
2207 
2208 		mtx_lock(&umtx_lock);
2209 		if (UPRI(td) < PRI_MIN_REALTIME + ceiling) {
2210 			mtx_unlock(&umtx_lock);
2211 			error = EINVAL;
2212 			goto out;
2213 		}
2214 		if (su && PRI_MIN_REALTIME + ceiling < uq->uq_inherited_pri) {
2215 			uq->uq_inherited_pri = PRI_MIN_REALTIME + ceiling;
2216 			thread_lock(td);
2217 			if (uq->uq_inherited_pri < UPRI(td))
2218 				sched_lend_user_prio(td, uq->uq_inherited_pri);
2219 			thread_unlock(td);
2220 		}
2221 		mtx_unlock(&umtx_lock);
2222 
2223 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2224 		    id | UMUTEX_CONTESTED);
2225 		/* The address was invalid. */
2226 		if (rv == -1) {
2227 			error = EFAULT;
2228 			break;
2229 		}
2230 		if (rv == 0) {
2231 			MPASS(owner == UMUTEX_CONTESTED);
2232 			error = 0;
2233 			break;
2234 		}
2235 		/* rv == 1 */
2236 		if (owner == UMUTEX_RB_OWNERDEAD) {
2237 			rv = casueword32(&m->m_owner, UMUTEX_RB_OWNERDEAD,
2238 			    &owner, id | UMUTEX_CONTESTED);
2239 			if (rv == -1) {
2240 				error = EFAULT;
2241 				break;
2242 			}
2243 			if (rv == 0) {
2244 				MPASS(owner == UMUTEX_RB_OWNERDEAD);
2245 				error = EOWNERDEAD; /* success */
2246 				break;
2247 			}
2248 
2249 			/*
2250 			 *  rv == 1, only check for suspension if we
2251 			 *  did not already catched a signal.  If we
2252 			 *  get an error from the check, the same
2253 			 *  condition is checked by the umtxq_sleep()
2254 			 *  call below, so we should obliterate the
2255 			 *  error to not skip the last loop iteration.
2256 			 */
2257 			if (error == 0) {
2258 				error = umtxq_check_susp(td, false);
2259 				if (error == 0) {
2260 					if (try != 0)
2261 						error = EBUSY;
2262 					else
2263 						continue;
2264 				}
2265 				error = 0;
2266 			}
2267 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2268 			error = ENOTRECOVERABLE;
2269 		}
2270 
2271 		if (try != 0)
2272 			error = EBUSY;
2273 
2274 		/*
2275 		 * If we caught a signal, we have retried and now
2276 		 * exit immediately.
2277 		 */
2278 		if (error != 0)
2279 			break;
2280 
2281 		umtxq_lock(&uq->uq_key);
2282 		umtxq_insert(uq);
2283 		umtxq_unbusy(&uq->uq_key);
2284 		error = umtxq_sleep(uq, "umtxpp", timeout == NULL ?
2285 		    NULL : &timo);
2286 		umtxq_remove(uq);
2287 		umtxq_unlock(&uq->uq_key);
2288 
2289 		mtx_lock(&umtx_lock);
2290 		uq->uq_inherited_pri = old_inherited_pri;
2291 		pri = PRI_MAX;
2292 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2293 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2294 			if (uq2 != NULL) {
2295 				if (pri > UPRI(uq2->uq_thread))
2296 					pri = UPRI(uq2->uq_thread);
2297 			}
2298 		}
2299 		if (pri > uq->uq_inherited_pri)
2300 			pri = uq->uq_inherited_pri;
2301 		thread_lock(td);
2302 		sched_lend_user_prio(td, pri);
2303 		thread_unlock(td);
2304 		mtx_unlock(&umtx_lock);
2305 	}
2306 
2307 	if (error != 0 && error != EOWNERDEAD) {
2308 		mtx_lock(&umtx_lock);
2309 		uq->uq_inherited_pri = old_inherited_pri;
2310 		pri = PRI_MAX;
2311 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2312 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2313 			if (uq2 != NULL) {
2314 				if (pri > UPRI(uq2->uq_thread))
2315 					pri = UPRI(uq2->uq_thread);
2316 			}
2317 		}
2318 		if (pri > uq->uq_inherited_pri)
2319 			pri = uq->uq_inherited_pri;
2320 		thread_lock(td);
2321 		sched_lend_user_prio(td, pri);
2322 		thread_unlock(td);
2323 		mtx_unlock(&umtx_lock);
2324 	}
2325 
2326 out:
2327 	umtxq_unbusy_unlocked(&uq->uq_key);
2328 	umtx_key_release(&uq->uq_key);
2329 	return (error);
2330 }
2331 
2332 /*
2333  * Unlock a PP mutex.
2334  */
2335 static int
2336 do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2337 {
2338 	struct umtx_key key;
2339 	struct umtx_q *uq, *uq2;
2340 	struct umtx_pi *pi;
2341 	uint32_t id, owner, rceiling;
2342 	int error, pri, new_inherited_pri, su;
2343 
2344 	id = td->td_tid;
2345 	uq = td->td_umtxq;
2346 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2347 
2348 	/*
2349 	 * Make sure we own this mtx.
2350 	 */
2351 	error = fueword32(&m->m_owner, &owner);
2352 	if (error == -1)
2353 		return (EFAULT);
2354 
2355 	if ((owner & ~UMUTEX_CONTESTED) != id)
2356 		return (EPERM);
2357 
2358 	error = copyin(&m->m_ceilings[1], &rceiling, sizeof(uint32_t));
2359 	if (error != 0)
2360 		return (error);
2361 
2362 	if (rceiling == -1)
2363 		new_inherited_pri = PRI_MAX;
2364 	else {
2365 		rceiling = RTP_PRIO_MAX - rceiling;
2366 		if (rceiling > RTP_PRIO_MAX)
2367 			return (EINVAL);
2368 		new_inherited_pri = PRI_MIN_REALTIME + rceiling;
2369 	}
2370 
2371 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2372 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2373 	    &key)) != 0)
2374 		return (error);
2375 	umtxq_lock(&key);
2376 	umtxq_busy(&key);
2377 	umtxq_unlock(&key);
2378 	/*
2379 	 * For priority protected mutex, always set unlocked state
2380 	 * to UMUTEX_CONTESTED, so that userland always enters kernel
2381 	 * to lock the mutex, it is necessary because thread priority
2382 	 * has to be adjusted for such mutex.
2383 	 */
2384 	error = suword32(&m->m_owner, umtx_unlock_val(flags, rb) |
2385 	    UMUTEX_CONTESTED);
2386 
2387 	umtxq_lock(&key);
2388 	if (error == 0)
2389 		umtxq_signal(&key, 1);
2390 	umtxq_unbusy(&key);
2391 	umtxq_unlock(&key);
2392 
2393 	if (error == -1)
2394 		error = EFAULT;
2395 	else {
2396 		mtx_lock(&umtx_lock);
2397 		if (su != 0)
2398 			uq->uq_inherited_pri = new_inherited_pri;
2399 		pri = PRI_MAX;
2400 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2401 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2402 			if (uq2 != NULL) {
2403 				if (pri > UPRI(uq2->uq_thread))
2404 					pri = UPRI(uq2->uq_thread);
2405 			}
2406 		}
2407 		if (pri > uq->uq_inherited_pri)
2408 			pri = uq->uq_inherited_pri;
2409 		thread_lock(td);
2410 		sched_lend_user_prio(td, pri);
2411 		thread_unlock(td);
2412 		mtx_unlock(&umtx_lock);
2413 	}
2414 	umtx_key_release(&key);
2415 	return (error);
2416 }
2417 
2418 static int
2419 do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
2420     uint32_t *old_ceiling)
2421 {
2422 	struct umtx_q *uq;
2423 	uint32_t flags, id, owner, save_ceiling;
2424 	int error, rv, rv1;
2425 
2426 	error = fueword32(&m->m_flags, &flags);
2427 	if (error == -1)
2428 		return (EFAULT);
2429 	if ((flags & UMUTEX_PRIO_PROTECT) == 0)
2430 		return (EINVAL);
2431 	if (ceiling > RTP_PRIO_MAX)
2432 		return (EINVAL);
2433 	id = td->td_tid;
2434 	uq = td->td_umtxq;
2435 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2436 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2437 	    &uq->uq_key)) != 0)
2438 		return (error);
2439 	for (;;) {
2440 		umtxq_lock(&uq->uq_key);
2441 		umtxq_busy(&uq->uq_key);
2442 		umtxq_unlock(&uq->uq_key);
2443 
2444 		rv = fueword32(&m->m_ceilings[0], &save_ceiling);
2445 		if (rv == -1) {
2446 			error = EFAULT;
2447 			break;
2448 		}
2449 
2450 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2451 		    id | UMUTEX_CONTESTED);
2452 		if (rv == -1) {
2453 			error = EFAULT;
2454 			break;
2455 		}
2456 
2457 		if (rv == 0) {
2458 			MPASS(owner == UMUTEX_CONTESTED);
2459 			rv = suword32(&m->m_ceilings[0], ceiling);
2460 			rv1 = suword32(&m->m_owner, UMUTEX_CONTESTED);
2461 			error = (rv == 0 && rv1 == 0) ? 0: EFAULT;
2462 			break;
2463 		}
2464 
2465 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2466 			rv = suword32(&m->m_ceilings[0], ceiling);
2467 			error = rv == 0 ? 0 : EFAULT;
2468 			break;
2469 		}
2470 
2471 		if (owner == UMUTEX_RB_OWNERDEAD) {
2472 			error = EOWNERDEAD;
2473 			break;
2474 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2475 			error = ENOTRECOVERABLE;
2476 			break;
2477 		}
2478 
2479 		/*
2480 		 * If we caught a signal, we have retried and now
2481 		 * exit immediately.
2482 		 */
2483 		if (error != 0)
2484 			break;
2485 
2486 		/*
2487 		 * We set the contested bit, sleep. Otherwise the lock changed
2488 		 * and we need to retry or we lost a race to the thread
2489 		 * unlocking the umtx.
2490 		 */
2491 		umtxq_lock(&uq->uq_key);
2492 		umtxq_insert(uq);
2493 		umtxq_unbusy(&uq->uq_key);
2494 		error = umtxq_sleep(uq, "umtxpp", NULL);
2495 		umtxq_remove(uq);
2496 		umtxq_unlock(&uq->uq_key);
2497 	}
2498 	umtxq_lock(&uq->uq_key);
2499 	if (error == 0)
2500 		umtxq_signal(&uq->uq_key, INT_MAX);
2501 	umtxq_unbusy(&uq->uq_key);
2502 	umtxq_unlock(&uq->uq_key);
2503 	umtx_key_release(&uq->uq_key);
2504 	if (error == 0 && old_ceiling != NULL) {
2505 		rv = suword32(old_ceiling, save_ceiling);
2506 		error = rv == 0 ? 0 : EFAULT;
2507 	}
2508 	return (error);
2509 }
2510 
2511 /*
2512  * Lock a userland POSIX mutex.
2513  */
2514 static int
2515 do_lock_umutex(struct thread *td, struct umutex *m,
2516     struct _umtx_time *timeout, int mode)
2517 {
2518 	uint32_t flags;
2519 	int error;
2520 
2521 	error = fueword32(&m->m_flags, &flags);
2522 	if (error == -1)
2523 		return (EFAULT);
2524 
2525 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2526 	case 0:
2527 		error = do_lock_normal(td, m, flags, timeout, mode);
2528 		break;
2529 	case UMUTEX_PRIO_INHERIT:
2530 		error = do_lock_pi(td, m, flags, timeout, mode);
2531 		break;
2532 	case UMUTEX_PRIO_PROTECT:
2533 		error = do_lock_pp(td, m, flags, timeout, mode);
2534 		break;
2535 	default:
2536 		return (EINVAL);
2537 	}
2538 	if (timeout == NULL) {
2539 		if (error == EINTR && mode != _UMUTEX_WAIT)
2540 			error = ERESTART;
2541 	} else {
2542 		/* Timed-locking is not restarted. */
2543 		if (error == ERESTART)
2544 			error = EINTR;
2545 	}
2546 	return (error);
2547 }
2548 
2549 /*
2550  * Unlock a userland POSIX mutex.
2551  */
2552 static int
2553 do_unlock_umutex(struct thread *td, struct umutex *m, bool rb)
2554 {
2555 	uint32_t flags;
2556 	int error;
2557 
2558 	error = fueword32(&m->m_flags, &flags);
2559 	if (error == -1)
2560 		return (EFAULT);
2561 
2562 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2563 	case 0:
2564 		return (do_unlock_normal(td, m, flags, rb));
2565 	case UMUTEX_PRIO_INHERIT:
2566 		return (do_unlock_pi(td, m, flags, rb));
2567 	case UMUTEX_PRIO_PROTECT:
2568 		return (do_unlock_pp(td, m, flags, rb));
2569 	}
2570 
2571 	return (EINVAL);
2572 }
2573 
2574 static int
2575 do_cv_wait(struct thread *td, struct ucond *cv, struct umutex *m,
2576     struct timespec *timeout, u_long wflags)
2577 {
2578 	struct abs_timeout timo;
2579 	struct umtx_q *uq;
2580 	uint32_t flags, clockid, hasw;
2581 	int error;
2582 
2583 	uq = td->td_umtxq;
2584 	error = fueword32(&cv->c_flags, &flags);
2585 	if (error == -1)
2586 		return (EFAULT);
2587 	error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &uq->uq_key);
2588 	if (error != 0)
2589 		return (error);
2590 
2591 	if ((wflags & CVWAIT_CLOCKID) != 0) {
2592 		error = fueword32(&cv->c_clockid, &clockid);
2593 		if (error == -1) {
2594 			umtx_key_release(&uq->uq_key);
2595 			return (EFAULT);
2596 		}
2597 		if (clockid < CLOCK_REALTIME ||
2598 		    clockid >= CLOCK_THREAD_CPUTIME_ID) {
2599 			/* hmm, only HW clock id will work. */
2600 			umtx_key_release(&uq->uq_key);
2601 			return (EINVAL);
2602 		}
2603 	} else {
2604 		clockid = CLOCK_REALTIME;
2605 	}
2606 
2607 	umtxq_lock(&uq->uq_key);
2608 	umtxq_busy(&uq->uq_key);
2609 	umtxq_insert(uq);
2610 	umtxq_unlock(&uq->uq_key);
2611 
2612 	/*
2613 	 * Set c_has_waiters to 1 before releasing user mutex, also
2614 	 * don't modify cache line when unnecessary.
2615 	 */
2616 	error = fueword32(&cv->c_has_waiters, &hasw);
2617 	if (error == 0 && hasw == 0)
2618 		suword32(&cv->c_has_waiters, 1);
2619 
2620 	umtxq_unbusy_unlocked(&uq->uq_key);
2621 
2622 	error = do_unlock_umutex(td, m, false);
2623 
2624 	if (timeout != NULL)
2625 		abs_timeout_init(&timo, clockid, (wflags & CVWAIT_ABSTIME) != 0,
2626 		    timeout);
2627 
2628 	umtxq_lock(&uq->uq_key);
2629 	if (error == 0) {
2630 		error = umtxq_sleep(uq, "ucond", timeout == NULL ?
2631 		    NULL : &timo);
2632 	}
2633 
2634 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
2635 		error = 0;
2636 	else {
2637 		/*
2638 		 * This must be timeout,interrupted by signal or
2639 		 * surprious wakeup, clear c_has_waiter flag when
2640 		 * necessary.
2641 		 */
2642 		umtxq_busy(&uq->uq_key);
2643 		if ((uq->uq_flags & UQF_UMTXQ) != 0) {
2644 			int oldlen = uq->uq_cur_queue->length;
2645 			umtxq_remove(uq);
2646 			if (oldlen == 1) {
2647 				umtxq_unlock(&uq->uq_key);
2648 				suword32(&cv->c_has_waiters, 0);
2649 				umtxq_lock(&uq->uq_key);
2650 			}
2651 		}
2652 		umtxq_unbusy(&uq->uq_key);
2653 		if (error == ERESTART)
2654 			error = EINTR;
2655 	}
2656 
2657 	umtxq_unlock(&uq->uq_key);
2658 	umtx_key_release(&uq->uq_key);
2659 	return (error);
2660 }
2661 
2662 /*
2663  * Signal a userland condition variable.
2664  */
2665 static int
2666 do_cv_signal(struct thread *td, struct ucond *cv)
2667 {
2668 	struct umtx_key key;
2669 	int error, cnt, nwake;
2670 	uint32_t flags;
2671 
2672 	error = fueword32(&cv->c_flags, &flags);
2673 	if (error == -1)
2674 		return (EFAULT);
2675 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
2676 		return (error);
2677 	umtxq_lock(&key);
2678 	umtxq_busy(&key);
2679 	cnt = umtxq_count(&key);
2680 	nwake = umtxq_signal(&key, 1);
2681 	if (cnt <= nwake) {
2682 		umtxq_unlock(&key);
2683 		error = suword32(&cv->c_has_waiters, 0);
2684 		if (error == -1)
2685 			error = EFAULT;
2686 		umtxq_lock(&key);
2687 	}
2688 	umtxq_unbusy(&key);
2689 	umtxq_unlock(&key);
2690 	umtx_key_release(&key);
2691 	return (error);
2692 }
2693 
2694 static int
2695 do_cv_broadcast(struct thread *td, struct ucond *cv)
2696 {
2697 	struct umtx_key key;
2698 	int error;
2699 	uint32_t flags;
2700 
2701 	error = fueword32(&cv->c_flags, &flags);
2702 	if (error == -1)
2703 		return (EFAULT);
2704 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
2705 		return (error);
2706 
2707 	umtxq_lock(&key);
2708 	umtxq_busy(&key);
2709 	umtxq_signal(&key, INT_MAX);
2710 	umtxq_unlock(&key);
2711 
2712 	error = suword32(&cv->c_has_waiters, 0);
2713 	if (error == -1)
2714 		error = EFAULT;
2715 
2716 	umtxq_unbusy_unlocked(&key);
2717 
2718 	umtx_key_release(&key);
2719 	return (error);
2720 }
2721 
2722 static int
2723 do_rw_rdlock(struct thread *td, struct urwlock *rwlock, long fflag,
2724     struct _umtx_time *timeout)
2725 {
2726 	struct abs_timeout timo;
2727 	struct umtx_q *uq;
2728 	uint32_t flags, wrflags;
2729 	int32_t state, oldstate;
2730 	int32_t blocked_readers;
2731 	int error, error1, rv;
2732 
2733 	uq = td->td_umtxq;
2734 	error = fueword32(&rwlock->rw_flags, &flags);
2735 	if (error == -1)
2736 		return (EFAULT);
2737 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
2738 	if (error != 0)
2739 		return (error);
2740 
2741 	if (timeout != NULL)
2742 		abs_timeout_init2(&timo, timeout);
2743 
2744 	wrflags = URWLOCK_WRITE_OWNER;
2745 	if (!(fflag & URWLOCK_PREFER_READER) && !(flags & URWLOCK_PREFER_READER))
2746 		wrflags |= URWLOCK_WRITE_WAITERS;
2747 
2748 	for (;;) {
2749 		rv = fueword32(&rwlock->rw_state, &state);
2750 		if (rv == -1) {
2751 			umtx_key_release(&uq->uq_key);
2752 			return (EFAULT);
2753 		}
2754 
2755 		/* try to lock it */
2756 		while (!(state & wrflags)) {
2757 			if (__predict_false(URWLOCK_READER_COUNT(state) ==
2758 			    URWLOCK_MAX_READERS)) {
2759 				umtx_key_release(&uq->uq_key);
2760 				return (EAGAIN);
2761 			}
2762 			rv = casueword32(&rwlock->rw_state, state,
2763 			    &oldstate, state + 1);
2764 			if (rv == -1) {
2765 				umtx_key_release(&uq->uq_key);
2766 				return (EFAULT);
2767 			}
2768 			if (rv == 0) {
2769 				MPASS(oldstate == state);
2770 				umtx_key_release(&uq->uq_key);
2771 				return (0);
2772 			}
2773 			error = umtxq_check_susp(td, true);
2774 			if (error != 0)
2775 				break;
2776 			state = oldstate;
2777 		}
2778 
2779 		if (error)
2780 			break;
2781 
2782 		/* grab monitor lock */
2783 		umtxq_lock(&uq->uq_key);
2784 		umtxq_busy(&uq->uq_key);
2785 		umtxq_unlock(&uq->uq_key);
2786 
2787 		/*
2788 		 * re-read the state, in case it changed between the try-lock above
2789 		 * and the check below
2790 		 */
2791 		rv = fueword32(&rwlock->rw_state, &state);
2792 		if (rv == -1)
2793 			error = EFAULT;
2794 
2795 		/* set read contention bit */
2796 		while (error == 0 && (state & wrflags) &&
2797 		    !(state & URWLOCK_READ_WAITERS)) {
2798 			rv = casueword32(&rwlock->rw_state, state,
2799 			    &oldstate, state | URWLOCK_READ_WAITERS);
2800 			if (rv == -1) {
2801 				error = EFAULT;
2802 				break;
2803 			}
2804 			if (rv == 0) {
2805 				MPASS(oldstate == state);
2806 				goto sleep;
2807 			}
2808 			state = oldstate;
2809 			error = umtxq_check_susp(td, false);
2810 			if (error != 0)
2811 				break;
2812 		}
2813 		if (error != 0) {
2814 			umtxq_unbusy_unlocked(&uq->uq_key);
2815 			break;
2816 		}
2817 
2818 		/* state is changed while setting flags, restart */
2819 		if (!(state & wrflags)) {
2820 			umtxq_unbusy_unlocked(&uq->uq_key);
2821 			error = umtxq_check_susp(td, true);
2822 			if (error != 0)
2823 				break;
2824 			continue;
2825 		}
2826 
2827 sleep:
2828 		/*
2829 		 * Contention bit is set, before sleeping, increase
2830 		 * read waiter count.
2831 		 */
2832 		rv = fueword32(&rwlock->rw_blocked_readers,
2833 		    &blocked_readers);
2834 		if (rv == -1) {
2835 			umtxq_unbusy_unlocked(&uq->uq_key);
2836 			error = EFAULT;
2837 			break;
2838 		}
2839 		suword32(&rwlock->rw_blocked_readers, blocked_readers+1);
2840 
2841 		while (state & wrflags) {
2842 			umtxq_lock(&uq->uq_key);
2843 			umtxq_insert(uq);
2844 			umtxq_unbusy(&uq->uq_key);
2845 
2846 			error = umtxq_sleep(uq, "urdlck", timeout == NULL ?
2847 			    NULL : &timo);
2848 
2849 			umtxq_busy(&uq->uq_key);
2850 			umtxq_remove(uq);
2851 			umtxq_unlock(&uq->uq_key);
2852 			if (error)
2853 				break;
2854 			rv = fueword32(&rwlock->rw_state, &state);
2855 			if (rv == -1) {
2856 				error = EFAULT;
2857 				break;
2858 			}
2859 		}
2860 
2861 		/* decrease read waiter count, and may clear read contention bit */
2862 		rv = fueword32(&rwlock->rw_blocked_readers,
2863 		    &blocked_readers);
2864 		if (rv == -1) {
2865 			umtxq_unbusy_unlocked(&uq->uq_key);
2866 			error = EFAULT;
2867 			break;
2868 		}
2869 		suword32(&rwlock->rw_blocked_readers, blocked_readers-1);
2870 		if (blocked_readers == 1) {
2871 			rv = fueword32(&rwlock->rw_state, &state);
2872 			if (rv == -1) {
2873 				umtxq_unbusy_unlocked(&uq->uq_key);
2874 				error = EFAULT;
2875 				break;
2876 			}
2877 			for (;;) {
2878 				rv = casueword32(&rwlock->rw_state, state,
2879 				    &oldstate, state & ~URWLOCK_READ_WAITERS);
2880 				if (rv == -1) {
2881 					error = EFAULT;
2882 					break;
2883 				}
2884 				if (rv == 0) {
2885 					MPASS(oldstate == state);
2886 					break;
2887 				}
2888 				state = oldstate;
2889 				error1 = umtxq_check_susp(td, false);
2890 				if (error1 != 0) {
2891 					if (error == 0)
2892 						error = error1;
2893 					break;
2894 				}
2895 			}
2896 		}
2897 
2898 		umtxq_unbusy_unlocked(&uq->uq_key);
2899 		if (error != 0)
2900 			break;
2901 	}
2902 	umtx_key_release(&uq->uq_key);
2903 	if (error == ERESTART)
2904 		error = EINTR;
2905 	return (error);
2906 }
2907 
2908 static int
2909 do_rw_wrlock(struct thread *td, struct urwlock *rwlock, struct _umtx_time *timeout)
2910 {
2911 	struct abs_timeout timo;
2912 	struct umtx_q *uq;
2913 	uint32_t flags;
2914 	int32_t state, oldstate;
2915 	int32_t blocked_writers;
2916 	int32_t blocked_readers;
2917 	int error, error1, rv;
2918 
2919 	uq = td->td_umtxq;
2920 	error = fueword32(&rwlock->rw_flags, &flags);
2921 	if (error == -1)
2922 		return (EFAULT);
2923 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
2924 	if (error != 0)
2925 		return (error);
2926 
2927 	if (timeout != NULL)
2928 		abs_timeout_init2(&timo, timeout);
2929 
2930 	blocked_readers = 0;
2931 	for (;;) {
2932 		rv = fueword32(&rwlock->rw_state, &state);
2933 		if (rv == -1) {
2934 			umtx_key_release(&uq->uq_key);
2935 			return (EFAULT);
2936 		}
2937 		while ((state & URWLOCK_WRITE_OWNER) == 0 &&
2938 		    URWLOCK_READER_COUNT(state) == 0) {
2939 			rv = casueword32(&rwlock->rw_state, state,
2940 			    &oldstate, state | URWLOCK_WRITE_OWNER);
2941 			if (rv == -1) {
2942 				umtx_key_release(&uq->uq_key);
2943 				return (EFAULT);
2944 			}
2945 			if (rv == 0) {
2946 				MPASS(oldstate == state);
2947 				umtx_key_release(&uq->uq_key);
2948 				return (0);
2949 			}
2950 			state = oldstate;
2951 			error = umtxq_check_susp(td, true);
2952 			if (error != 0)
2953 				break;
2954 		}
2955 
2956 		if (error) {
2957 			if ((state & (URWLOCK_WRITE_OWNER |
2958 			    URWLOCK_WRITE_WAITERS)) == 0 &&
2959 			    blocked_readers != 0) {
2960 				umtxq_lock(&uq->uq_key);
2961 				umtxq_busy(&uq->uq_key);
2962 				umtxq_signal_queue(&uq->uq_key, INT_MAX,
2963 				    UMTX_SHARED_QUEUE);
2964 				umtxq_unbusy(&uq->uq_key);
2965 				umtxq_unlock(&uq->uq_key);
2966 			}
2967 
2968 			break;
2969 		}
2970 
2971 		/* grab monitor lock */
2972 		umtxq_lock(&uq->uq_key);
2973 		umtxq_busy(&uq->uq_key);
2974 		umtxq_unlock(&uq->uq_key);
2975 
2976 		/*
2977 		 * Re-read the state, in case it changed between the
2978 		 * try-lock above and the check below.
2979 		 */
2980 		rv = fueword32(&rwlock->rw_state, &state);
2981 		if (rv == -1)
2982 			error = EFAULT;
2983 
2984 		while (error == 0 && ((state & URWLOCK_WRITE_OWNER) ||
2985 		    URWLOCK_READER_COUNT(state) != 0) &&
2986 		    (state & URWLOCK_WRITE_WAITERS) == 0) {
2987 			rv = casueword32(&rwlock->rw_state, state,
2988 			    &oldstate, state | URWLOCK_WRITE_WAITERS);
2989 			if (rv == -1) {
2990 				error = EFAULT;
2991 				break;
2992 			}
2993 			if (rv == 0) {
2994 				MPASS(oldstate == state);
2995 				goto sleep;
2996 			}
2997 			state = oldstate;
2998 			error = umtxq_check_susp(td, false);
2999 			if (error != 0)
3000 				break;
3001 		}
3002 		if (error != 0) {
3003 			umtxq_unbusy_unlocked(&uq->uq_key);
3004 			break;
3005 		}
3006 
3007 		if ((state & URWLOCK_WRITE_OWNER) == 0 &&
3008 		    URWLOCK_READER_COUNT(state) == 0) {
3009 			umtxq_unbusy_unlocked(&uq->uq_key);
3010 			error = umtxq_check_susp(td, false);
3011 			if (error != 0)
3012 				break;
3013 			continue;
3014 		}
3015 sleep:
3016 		rv = fueword32(&rwlock->rw_blocked_writers,
3017 		    &blocked_writers);
3018 		if (rv == -1) {
3019 			umtxq_unbusy_unlocked(&uq->uq_key);
3020 			error = EFAULT;
3021 			break;
3022 		}
3023 		suword32(&rwlock->rw_blocked_writers, blocked_writers + 1);
3024 
3025 		while ((state & URWLOCK_WRITE_OWNER) ||
3026 		    URWLOCK_READER_COUNT(state) != 0) {
3027 			umtxq_lock(&uq->uq_key);
3028 			umtxq_insert_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3029 			umtxq_unbusy(&uq->uq_key);
3030 
3031 			error = umtxq_sleep(uq, "uwrlck", timeout == NULL ?
3032 			    NULL : &timo);
3033 
3034 			umtxq_busy(&uq->uq_key);
3035 			umtxq_remove_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3036 			umtxq_unlock(&uq->uq_key);
3037 			if (error)
3038 				break;
3039 			rv = fueword32(&rwlock->rw_state, &state);
3040 			if (rv == -1) {
3041 				error = EFAULT;
3042 				break;
3043 			}
3044 		}
3045 
3046 		rv = fueword32(&rwlock->rw_blocked_writers,
3047 		    &blocked_writers);
3048 		if (rv == -1) {
3049 			umtxq_unbusy_unlocked(&uq->uq_key);
3050 			error = EFAULT;
3051 			break;
3052 		}
3053 		suword32(&rwlock->rw_blocked_writers, blocked_writers-1);
3054 		if (blocked_writers == 1) {
3055 			rv = fueword32(&rwlock->rw_state, &state);
3056 			if (rv == -1) {
3057 				umtxq_unbusy_unlocked(&uq->uq_key);
3058 				error = EFAULT;
3059 				break;
3060 			}
3061 			for (;;) {
3062 				rv = casueword32(&rwlock->rw_state, state,
3063 				    &oldstate, state & ~URWLOCK_WRITE_WAITERS);
3064 				if (rv == -1) {
3065 					error = EFAULT;
3066 					break;
3067 				}
3068 				if (rv == 0) {
3069 					MPASS(oldstate == state);
3070 					break;
3071 				}
3072 				state = oldstate;
3073 				error1 = umtxq_check_susp(td, false);
3074 				/*
3075 				 * We are leaving the URWLOCK_WRITE_WAITERS
3076 				 * behind, but this should not harm the
3077 				 * correctness.
3078 				 */
3079 				if (error1 != 0) {
3080 					if (error == 0)
3081 						error = error1;
3082 					break;
3083 				}
3084 			}
3085 			rv = fueword32(&rwlock->rw_blocked_readers,
3086 			    &blocked_readers);
3087 			if (rv == -1) {
3088 				umtxq_unbusy_unlocked(&uq->uq_key);
3089 				error = EFAULT;
3090 				break;
3091 			}
3092 		} else
3093 			blocked_readers = 0;
3094 
3095 		umtxq_unbusy_unlocked(&uq->uq_key);
3096 	}
3097 
3098 	umtx_key_release(&uq->uq_key);
3099 	if (error == ERESTART)
3100 		error = EINTR;
3101 	return (error);
3102 }
3103 
3104 static int
3105 do_rw_unlock(struct thread *td, struct urwlock *rwlock)
3106 {
3107 	struct umtx_q *uq;
3108 	uint32_t flags;
3109 	int32_t state, oldstate;
3110 	int error, rv, q, count;
3111 
3112 	uq = td->td_umtxq;
3113 	error = fueword32(&rwlock->rw_flags, &flags);
3114 	if (error == -1)
3115 		return (EFAULT);
3116 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3117 	if (error != 0)
3118 		return (error);
3119 
3120 	error = fueword32(&rwlock->rw_state, &state);
3121 	if (error == -1) {
3122 		error = EFAULT;
3123 		goto out;
3124 	}
3125 	if (state & URWLOCK_WRITE_OWNER) {
3126 		for (;;) {
3127 			rv = casueword32(&rwlock->rw_state, state,
3128 			    &oldstate, state & ~URWLOCK_WRITE_OWNER);
3129 			if (rv == -1) {
3130 				error = EFAULT;
3131 				goto out;
3132 			}
3133 			if (rv == 1) {
3134 				state = oldstate;
3135 				if (!(oldstate & URWLOCK_WRITE_OWNER)) {
3136 					error = EPERM;
3137 					goto out;
3138 				}
3139 				error = umtxq_check_susp(td, true);
3140 				if (error != 0)
3141 					goto out;
3142 			} else
3143 				break;
3144 		}
3145 	} else if (URWLOCK_READER_COUNT(state) != 0) {
3146 		for (;;) {
3147 			rv = casueword32(&rwlock->rw_state, state,
3148 			    &oldstate, state - 1);
3149 			if (rv == -1) {
3150 				error = EFAULT;
3151 				goto out;
3152 			}
3153 			if (rv == 1) {
3154 				state = oldstate;
3155 				if (URWLOCK_READER_COUNT(oldstate) == 0) {
3156 					error = EPERM;
3157 					goto out;
3158 				}
3159 				error = umtxq_check_susp(td, true);
3160 				if (error != 0)
3161 					goto out;
3162 			} else
3163 				break;
3164 		}
3165 	} else {
3166 		error = EPERM;
3167 		goto out;
3168 	}
3169 
3170 	count = 0;
3171 
3172 	if (!(flags & URWLOCK_PREFER_READER)) {
3173 		if (state & URWLOCK_WRITE_WAITERS) {
3174 			count = 1;
3175 			q = UMTX_EXCLUSIVE_QUEUE;
3176 		} else if (state & URWLOCK_READ_WAITERS) {
3177 			count = INT_MAX;
3178 			q = UMTX_SHARED_QUEUE;
3179 		}
3180 	} else {
3181 		if (state & URWLOCK_READ_WAITERS) {
3182 			count = INT_MAX;
3183 			q = UMTX_SHARED_QUEUE;
3184 		} else if (state & URWLOCK_WRITE_WAITERS) {
3185 			count = 1;
3186 			q = UMTX_EXCLUSIVE_QUEUE;
3187 		}
3188 	}
3189 
3190 	if (count) {
3191 		umtxq_lock(&uq->uq_key);
3192 		umtxq_busy(&uq->uq_key);
3193 		umtxq_signal_queue(&uq->uq_key, count, q);
3194 		umtxq_unbusy(&uq->uq_key);
3195 		umtxq_unlock(&uq->uq_key);
3196 	}
3197 out:
3198 	umtx_key_release(&uq->uq_key);
3199 	return (error);
3200 }
3201 
3202 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3203 static int
3204 do_sem_wait(struct thread *td, struct _usem *sem, struct _umtx_time *timeout)
3205 {
3206 	struct abs_timeout timo;
3207 	struct umtx_q *uq;
3208 	uint32_t flags, count, count1;
3209 	int error, rv, rv1;
3210 
3211 	uq = td->td_umtxq;
3212 	error = fueword32(&sem->_flags, &flags);
3213 	if (error == -1)
3214 		return (EFAULT);
3215 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3216 	if (error != 0)
3217 		return (error);
3218 
3219 	if (timeout != NULL)
3220 		abs_timeout_init2(&timo, timeout);
3221 
3222 again:
3223 	umtxq_lock(&uq->uq_key);
3224 	umtxq_busy(&uq->uq_key);
3225 	umtxq_insert(uq);
3226 	umtxq_unlock(&uq->uq_key);
3227 	rv = casueword32(&sem->_has_waiters, 0, &count1, 1);
3228 	if (rv == 0)
3229 		rv1 = fueword32(&sem->_count, &count);
3230 	if (rv == -1 || (rv == 0 && (rv1 == -1 || count != 0)) ||
3231 	    (rv == 1 && count1 == 0)) {
3232 		umtxq_lock(&uq->uq_key);
3233 		umtxq_unbusy(&uq->uq_key);
3234 		umtxq_remove(uq);
3235 		umtxq_unlock(&uq->uq_key);
3236 		if (rv == 1) {
3237 			rv = umtxq_check_susp(td, true);
3238 			if (rv == 0)
3239 				goto again;
3240 			error = rv;
3241 			goto out;
3242 		}
3243 		if (rv == 0)
3244 			rv = rv1;
3245 		error = rv == -1 ? EFAULT : 0;
3246 		goto out;
3247 	}
3248 	umtxq_lock(&uq->uq_key);
3249 	umtxq_unbusy(&uq->uq_key);
3250 
3251 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3252 
3253 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3254 		error = 0;
3255 	else {
3256 		umtxq_remove(uq);
3257 		/* A relative timeout cannot be restarted. */
3258 		if (error == ERESTART && timeout != NULL &&
3259 		    (timeout->_flags & UMTX_ABSTIME) == 0)
3260 			error = EINTR;
3261 	}
3262 	umtxq_unlock(&uq->uq_key);
3263 out:
3264 	umtx_key_release(&uq->uq_key);
3265 	return (error);
3266 }
3267 
3268 /*
3269  * Signal a userland semaphore.
3270  */
3271 static int
3272 do_sem_wake(struct thread *td, struct _usem *sem)
3273 {
3274 	struct umtx_key key;
3275 	int error, cnt;
3276 	uint32_t flags;
3277 
3278 	error = fueword32(&sem->_flags, &flags);
3279 	if (error == -1)
3280 		return (EFAULT);
3281 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3282 		return (error);
3283 	umtxq_lock(&key);
3284 	umtxq_busy(&key);
3285 	cnt = umtxq_count(&key);
3286 	if (cnt > 0) {
3287 		/*
3288 		 * Check if count is greater than 0, this means the memory is
3289 		 * still being referenced by user code, so we can safely
3290 		 * update _has_waiters flag.
3291 		 */
3292 		if (cnt == 1) {
3293 			umtxq_unlock(&key);
3294 			error = suword32(&sem->_has_waiters, 0);
3295 			umtxq_lock(&key);
3296 			if (error == -1)
3297 				error = EFAULT;
3298 		}
3299 		umtxq_signal(&key, 1);
3300 	}
3301 	umtxq_unbusy(&key);
3302 	umtxq_unlock(&key);
3303 	umtx_key_release(&key);
3304 	return (error);
3305 }
3306 #endif
3307 
3308 static int
3309 do_sem2_wait(struct thread *td, struct _usem2 *sem, struct _umtx_time *timeout)
3310 {
3311 	struct abs_timeout timo;
3312 	struct umtx_q *uq;
3313 	uint32_t count, flags;
3314 	int error, rv;
3315 
3316 	uq = td->td_umtxq;
3317 	flags = fuword32(&sem->_flags);
3318 	if (timeout != NULL)
3319 		abs_timeout_init2(&timo, timeout);
3320 
3321 again:
3322 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3323 	if (error != 0)
3324 		return (error);
3325 	umtxq_lock(&uq->uq_key);
3326 	umtxq_busy(&uq->uq_key);
3327 	umtxq_insert(uq);
3328 	umtxq_unlock(&uq->uq_key);
3329 	rv = fueword32(&sem->_count, &count);
3330 	if (rv == -1) {
3331 		umtxq_lock(&uq->uq_key);
3332 		umtxq_unbusy(&uq->uq_key);
3333 		umtxq_remove(uq);
3334 		umtxq_unlock(&uq->uq_key);
3335 		umtx_key_release(&uq->uq_key);
3336 		return (EFAULT);
3337 	}
3338 	for (;;) {
3339 		if (USEM_COUNT(count) != 0) {
3340 			umtxq_lock(&uq->uq_key);
3341 			umtxq_unbusy(&uq->uq_key);
3342 			umtxq_remove(uq);
3343 			umtxq_unlock(&uq->uq_key);
3344 			umtx_key_release(&uq->uq_key);
3345 			return (0);
3346 		}
3347 		if (count == USEM_HAS_WAITERS)
3348 			break;
3349 		rv = casueword32(&sem->_count, 0, &count, USEM_HAS_WAITERS);
3350 		if (rv == 0)
3351 			break;
3352 		umtxq_lock(&uq->uq_key);
3353 		umtxq_unbusy(&uq->uq_key);
3354 		umtxq_remove(uq);
3355 		umtxq_unlock(&uq->uq_key);
3356 		umtx_key_release(&uq->uq_key);
3357 		if (rv == -1)
3358 			return (EFAULT);
3359 		rv = umtxq_check_susp(td, true);
3360 		if (rv != 0)
3361 			return (rv);
3362 		goto again;
3363 	}
3364 	umtxq_lock(&uq->uq_key);
3365 	umtxq_unbusy(&uq->uq_key);
3366 
3367 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3368 
3369 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3370 		error = 0;
3371 	else {
3372 		umtxq_remove(uq);
3373 		if (timeout != NULL && (timeout->_flags & UMTX_ABSTIME) == 0) {
3374 			/* A relative timeout cannot be restarted. */
3375 			if (error == ERESTART)
3376 				error = EINTR;
3377 			if (error == EINTR) {
3378 				abs_timeout_update(&timo);
3379 				timespecsub(&timo.end, &timo.cur,
3380 				    &timeout->_timeout);
3381 			}
3382 		}
3383 	}
3384 	umtxq_unlock(&uq->uq_key);
3385 	umtx_key_release(&uq->uq_key);
3386 	return (error);
3387 }
3388 
3389 /*
3390  * Signal a userland semaphore.
3391  */
3392 static int
3393 do_sem2_wake(struct thread *td, struct _usem2 *sem)
3394 {
3395 	struct umtx_key key;
3396 	int error, cnt, rv;
3397 	uint32_t count, flags;
3398 
3399 	rv = fueword32(&sem->_flags, &flags);
3400 	if (rv == -1)
3401 		return (EFAULT);
3402 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3403 		return (error);
3404 	umtxq_lock(&key);
3405 	umtxq_busy(&key);
3406 	cnt = umtxq_count(&key);
3407 	if (cnt > 0) {
3408 		/*
3409 		 * If this was the last sleeping thread, clear the waiters
3410 		 * flag in _count.
3411 		 */
3412 		if (cnt == 1) {
3413 			umtxq_unlock(&key);
3414 			rv = fueword32(&sem->_count, &count);
3415 			while (rv != -1 && count & USEM_HAS_WAITERS) {
3416 				rv = casueword32(&sem->_count, count, &count,
3417 				    count & ~USEM_HAS_WAITERS);
3418 				if (rv == 1) {
3419 					rv = umtxq_check_susp(td, true);
3420 					if (rv != 0)
3421 						break;
3422 				}
3423 			}
3424 			if (rv == -1)
3425 				error = EFAULT;
3426 			else if (rv > 0) {
3427 				error = rv;
3428 			}
3429 			umtxq_lock(&key);
3430 		}
3431 
3432 		umtxq_signal(&key, 1);
3433 	}
3434 	umtxq_unbusy(&key);
3435 	umtxq_unlock(&key);
3436 	umtx_key_release(&key);
3437 	return (error);
3438 }
3439 
3440 inline int
3441 umtx_copyin_timeout(const void *addr, struct timespec *tsp)
3442 {
3443 	int error;
3444 
3445 	error = copyin(addr, tsp, sizeof(struct timespec));
3446 	if (error == 0) {
3447 		if (tsp->tv_sec < 0 ||
3448 		    tsp->tv_nsec >= 1000000000 ||
3449 		    tsp->tv_nsec < 0)
3450 			error = EINVAL;
3451 	}
3452 	return (error);
3453 }
3454 
3455 static inline int
3456 umtx_copyin_umtx_time(const void *addr, size_t size, struct _umtx_time *tp)
3457 {
3458 	int error;
3459 
3460 	if (size <= sizeof(struct timespec)) {
3461 		tp->_clockid = CLOCK_REALTIME;
3462 		tp->_flags = 0;
3463 		error = copyin(addr, &tp->_timeout, sizeof(struct timespec));
3464 	} else
3465 		error = copyin(addr, tp, sizeof(struct _umtx_time));
3466 	if (error != 0)
3467 		return (error);
3468 	if (tp->_timeout.tv_sec < 0 ||
3469 	    tp->_timeout.tv_nsec >= 1000000000 || tp->_timeout.tv_nsec < 0)
3470 		return (EINVAL);
3471 	return (0);
3472 }
3473 
3474 static int
3475 __umtx_op_unimpl(struct thread *td, struct _umtx_op_args *uap)
3476 {
3477 
3478 	return (EOPNOTSUPP);
3479 }
3480 
3481 static int
3482 __umtx_op_wait(struct thread *td, struct _umtx_op_args *uap)
3483 {
3484 	struct _umtx_time timeout, *tm_p;
3485 	int error;
3486 
3487 	if (uap->uaddr2 == NULL)
3488 		tm_p = NULL;
3489 	else {
3490 		error = umtx_copyin_umtx_time(
3491 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3492 		if (error != 0)
3493 			return (error);
3494 		tm_p = &timeout;
3495 	}
3496 	return (do_wait(td, uap->obj, uap->val, tm_p, 0, 0));
3497 }
3498 
3499 static int
3500 __umtx_op_wait_uint(struct thread *td, struct _umtx_op_args *uap)
3501 {
3502 	struct _umtx_time timeout, *tm_p;
3503 	int error;
3504 
3505 	if (uap->uaddr2 == NULL)
3506 		tm_p = NULL;
3507 	else {
3508 		error = umtx_copyin_umtx_time(
3509 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3510 		if (error != 0)
3511 			return (error);
3512 		tm_p = &timeout;
3513 	}
3514 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
3515 }
3516 
3517 static int
3518 __umtx_op_wait_uint_private(struct thread *td, struct _umtx_op_args *uap)
3519 {
3520 	struct _umtx_time *tm_p, timeout;
3521 	int error;
3522 
3523 	if (uap->uaddr2 == NULL)
3524 		tm_p = NULL;
3525 	else {
3526 		error = umtx_copyin_umtx_time(
3527 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3528 		if (error != 0)
3529 			return (error);
3530 		tm_p = &timeout;
3531 	}
3532 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
3533 }
3534 
3535 static int
3536 __umtx_op_wake(struct thread *td, struct _umtx_op_args *uap)
3537 {
3538 
3539 	return (kern_umtx_wake(td, uap->obj, uap->val, 0));
3540 }
3541 
3542 #define BATCH_SIZE	128
3543 static int
3544 __umtx_op_nwake_private(struct thread *td, struct _umtx_op_args *uap)
3545 {
3546 	char *uaddrs[BATCH_SIZE], **upp;
3547 	int count, error, i, pos, tocopy;
3548 
3549 	upp = (char **)uap->obj;
3550 	error = 0;
3551 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3552 	    pos += tocopy) {
3553 		tocopy = MIN(count, BATCH_SIZE);
3554 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(char *));
3555 		if (error != 0)
3556 			break;
3557 		for (i = 0; i < tocopy; ++i)
3558 			kern_umtx_wake(td, uaddrs[i], INT_MAX, 1);
3559 		maybe_yield();
3560 	}
3561 	return (error);
3562 }
3563 
3564 static int
3565 __umtx_op_wake_private(struct thread *td, struct _umtx_op_args *uap)
3566 {
3567 
3568 	return (kern_umtx_wake(td, uap->obj, uap->val, 1));
3569 }
3570 
3571 static int
3572 __umtx_op_lock_umutex(struct thread *td, struct _umtx_op_args *uap)
3573 {
3574 	struct _umtx_time *tm_p, timeout;
3575 	int error;
3576 
3577 	/* Allow a null timespec (wait forever). */
3578 	if (uap->uaddr2 == NULL)
3579 		tm_p = NULL;
3580 	else {
3581 		error = umtx_copyin_umtx_time(
3582 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3583 		if (error != 0)
3584 			return (error);
3585 		tm_p = &timeout;
3586 	}
3587 	return (do_lock_umutex(td, uap->obj, tm_p, 0));
3588 }
3589 
3590 static int
3591 __umtx_op_trylock_umutex(struct thread *td, struct _umtx_op_args *uap)
3592 {
3593 
3594 	return (do_lock_umutex(td, uap->obj, NULL, _UMUTEX_TRY));
3595 }
3596 
3597 static int
3598 __umtx_op_wait_umutex(struct thread *td, struct _umtx_op_args *uap)
3599 {
3600 	struct _umtx_time *tm_p, timeout;
3601 	int error;
3602 
3603 	/* Allow a null timespec (wait forever). */
3604 	if (uap->uaddr2 == NULL)
3605 		tm_p = NULL;
3606 	else {
3607 		error = umtx_copyin_umtx_time(
3608 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3609 		if (error != 0)
3610 			return (error);
3611 		tm_p = &timeout;
3612 	}
3613 	return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
3614 }
3615 
3616 static int
3617 __umtx_op_wake_umutex(struct thread *td, struct _umtx_op_args *uap)
3618 {
3619 
3620 	return (do_wake_umutex(td, uap->obj));
3621 }
3622 
3623 static int
3624 __umtx_op_unlock_umutex(struct thread *td, struct _umtx_op_args *uap)
3625 {
3626 
3627 	return (do_unlock_umutex(td, uap->obj, false));
3628 }
3629 
3630 static int
3631 __umtx_op_set_ceiling(struct thread *td, struct _umtx_op_args *uap)
3632 {
3633 
3634 	return (do_set_ceiling(td, uap->obj, uap->val, uap->uaddr1));
3635 }
3636 
3637 static int
3638 __umtx_op_cv_wait(struct thread *td, struct _umtx_op_args *uap)
3639 {
3640 	struct timespec *ts, timeout;
3641 	int error;
3642 
3643 	/* Allow a null timespec (wait forever). */
3644 	if (uap->uaddr2 == NULL)
3645 		ts = NULL;
3646 	else {
3647 		error = umtx_copyin_timeout(uap->uaddr2, &timeout);
3648 		if (error != 0)
3649 			return (error);
3650 		ts = &timeout;
3651 	}
3652 	return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
3653 }
3654 
3655 static int
3656 __umtx_op_cv_signal(struct thread *td, struct _umtx_op_args *uap)
3657 {
3658 
3659 	return (do_cv_signal(td, uap->obj));
3660 }
3661 
3662 static int
3663 __umtx_op_cv_broadcast(struct thread *td, struct _umtx_op_args *uap)
3664 {
3665 
3666 	return (do_cv_broadcast(td, uap->obj));
3667 }
3668 
3669 static int
3670 __umtx_op_rw_rdlock(struct thread *td, struct _umtx_op_args *uap)
3671 {
3672 	struct _umtx_time timeout;
3673 	int error;
3674 
3675 	/* Allow a null timespec (wait forever). */
3676 	if (uap->uaddr2 == NULL) {
3677 		error = do_rw_rdlock(td, uap->obj, uap->val, 0);
3678 	} else {
3679 		error = umtx_copyin_umtx_time(uap->uaddr2,
3680 		   (size_t)uap->uaddr1, &timeout);
3681 		if (error != 0)
3682 			return (error);
3683 		error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
3684 	}
3685 	return (error);
3686 }
3687 
3688 static int
3689 __umtx_op_rw_wrlock(struct thread *td, struct _umtx_op_args *uap)
3690 {
3691 	struct _umtx_time timeout;
3692 	int error;
3693 
3694 	/* Allow a null timespec (wait forever). */
3695 	if (uap->uaddr2 == NULL) {
3696 		error = do_rw_wrlock(td, uap->obj, 0);
3697 	} else {
3698 		error = umtx_copyin_umtx_time(uap->uaddr2,
3699 		   (size_t)uap->uaddr1, &timeout);
3700 		if (error != 0)
3701 			return (error);
3702 
3703 		error = do_rw_wrlock(td, uap->obj, &timeout);
3704 	}
3705 	return (error);
3706 }
3707 
3708 static int
3709 __umtx_op_rw_unlock(struct thread *td, struct _umtx_op_args *uap)
3710 {
3711 
3712 	return (do_rw_unlock(td, uap->obj));
3713 }
3714 
3715 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3716 static int
3717 __umtx_op_sem_wait(struct thread *td, struct _umtx_op_args *uap)
3718 {
3719 	struct _umtx_time *tm_p, timeout;
3720 	int error;
3721 
3722 	/* Allow a null timespec (wait forever). */
3723 	if (uap->uaddr2 == NULL)
3724 		tm_p = NULL;
3725 	else {
3726 		error = umtx_copyin_umtx_time(
3727 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3728 		if (error != 0)
3729 			return (error);
3730 		tm_p = &timeout;
3731 	}
3732 	return (do_sem_wait(td, uap->obj, tm_p));
3733 }
3734 
3735 static int
3736 __umtx_op_sem_wake(struct thread *td, struct _umtx_op_args *uap)
3737 {
3738 
3739 	return (do_sem_wake(td, uap->obj));
3740 }
3741 #endif
3742 
3743 static int
3744 __umtx_op_wake2_umutex(struct thread *td, struct _umtx_op_args *uap)
3745 {
3746 
3747 	return (do_wake2_umutex(td, uap->obj, uap->val));
3748 }
3749 
3750 static int
3751 __umtx_op_sem2_wait(struct thread *td, struct _umtx_op_args *uap)
3752 {
3753 	struct _umtx_time *tm_p, timeout;
3754 	size_t uasize;
3755 	int error;
3756 
3757 	/* Allow a null timespec (wait forever). */
3758 	if (uap->uaddr2 == NULL) {
3759 		uasize = 0;
3760 		tm_p = NULL;
3761 	} else {
3762 		uasize = (size_t)uap->uaddr1;
3763 		error = umtx_copyin_umtx_time(uap->uaddr2, uasize, &timeout);
3764 		if (error != 0)
3765 			return (error);
3766 		tm_p = &timeout;
3767 	}
3768 	error = do_sem2_wait(td, uap->obj, tm_p);
3769 	if (error == EINTR && uap->uaddr2 != NULL &&
3770 	    (timeout._flags & UMTX_ABSTIME) == 0 &&
3771 	    uasize >= sizeof(struct _umtx_time) + sizeof(struct timespec)) {
3772 		error = copyout(&timeout._timeout,
3773 		    (struct _umtx_time *)uap->uaddr2 + 1,
3774 		    sizeof(struct timespec));
3775 		if (error == 0) {
3776 			error = EINTR;
3777 		}
3778 	}
3779 
3780 	return (error);
3781 }
3782 
3783 static int
3784 __umtx_op_sem2_wake(struct thread *td, struct _umtx_op_args *uap)
3785 {
3786 
3787 	return (do_sem2_wake(td, uap->obj));
3788 }
3789 
3790 #define	USHM_OBJ_UMTX(o)						\
3791     ((struct umtx_shm_obj_list *)(&(o)->umtx_data))
3792 
3793 #define	USHMF_REG_LINKED	0x0001
3794 #define	USHMF_OBJ_LINKED	0x0002
3795 struct umtx_shm_reg {
3796 	TAILQ_ENTRY(umtx_shm_reg) ushm_reg_link;
3797 	LIST_ENTRY(umtx_shm_reg) ushm_obj_link;
3798 	struct umtx_key		ushm_key;
3799 	struct ucred		*ushm_cred;
3800 	struct shmfd		*ushm_obj;
3801 	u_int			ushm_refcnt;
3802 	u_int			ushm_flags;
3803 };
3804 
3805 LIST_HEAD(umtx_shm_obj_list, umtx_shm_reg);
3806 TAILQ_HEAD(umtx_shm_reg_head, umtx_shm_reg);
3807 
3808 static uma_zone_t umtx_shm_reg_zone;
3809 static struct umtx_shm_reg_head umtx_shm_registry[UMTX_CHAINS];
3810 static struct mtx umtx_shm_lock;
3811 static struct umtx_shm_reg_head umtx_shm_reg_delfree =
3812     TAILQ_HEAD_INITIALIZER(umtx_shm_reg_delfree);
3813 
3814 static void umtx_shm_free_reg(struct umtx_shm_reg *reg);
3815 
3816 static void
3817 umtx_shm_reg_delfree_tq(void *context __unused, int pending __unused)
3818 {
3819 	struct umtx_shm_reg_head d;
3820 	struct umtx_shm_reg *reg, *reg1;
3821 
3822 	TAILQ_INIT(&d);
3823 	mtx_lock(&umtx_shm_lock);
3824 	TAILQ_CONCAT(&d, &umtx_shm_reg_delfree, ushm_reg_link);
3825 	mtx_unlock(&umtx_shm_lock);
3826 	TAILQ_FOREACH_SAFE(reg, &d, ushm_reg_link, reg1) {
3827 		TAILQ_REMOVE(&d, reg, ushm_reg_link);
3828 		umtx_shm_free_reg(reg);
3829 	}
3830 }
3831 
3832 static struct task umtx_shm_reg_delfree_task =
3833     TASK_INITIALIZER(0, umtx_shm_reg_delfree_tq, NULL);
3834 
3835 static struct umtx_shm_reg *
3836 umtx_shm_find_reg_locked(const struct umtx_key *key)
3837 {
3838 	struct umtx_shm_reg *reg;
3839 	struct umtx_shm_reg_head *reg_head;
3840 
3841 	KASSERT(key->shared, ("umtx_p_find_rg: private key"));
3842 	mtx_assert(&umtx_shm_lock, MA_OWNED);
3843 	reg_head = &umtx_shm_registry[key->hash];
3844 	TAILQ_FOREACH(reg, reg_head, ushm_reg_link) {
3845 		KASSERT(reg->ushm_key.shared,
3846 		    ("non-shared key on reg %p %d", reg, reg->ushm_key.shared));
3847 		if (reg->ushm_key.info.shared.object ==
3848 		    key->info.shared.object &&
3849 		    reg->ushm_key.info.shared.offset ==
3850 		    key->info.shared.offset) {
3851 			KASSERT(reg->ushm_key.type == TYPE_SHM, ("TYPE_USHM"));
3852 			KASSERT(reg->ushm_refcnt > 0,
3853 			    ("reg %p refcnt 0 onlist", reg));
3854 			KASSERT((reg->ushm_flags & USHMF_REG_LINKED) != 0,
3855 			    ("reg %p not linked", reg));
3856 			reg->ushm_refcnt++;
3857 			return (reg);
3858 		}
3859 	}
3860 	return (NULL);
3861 }
3862 
3863 static struct umtx_shm_reg *
3864 umtx_shm_find_reg(const struct umtx_key *key)
3865 {
3866 	struct umtx_shm_reg *reg;
3867 
3868 	mtx_lock(&umtx_shm_lock);
3869 	reg = umtx_shm_find_reg_locked(key);
3870 	mtx_unlock(&umtx_shm_lock);
3871 	return (reg);
3872 }
3873 
3874 static void
3875 umtx_shm_free_reg(struct umtx_shm_reg *reg)
3876 {
3877 
3878 	chgumtxcnt(reg->ushm_cred->cr_ruidinfo, -1, 0);
3879 	crfree(reg->ushm_cred);
3880 	shm_drop(reg->ushm_obj);
3881 	uma_zfree(umtx_shm_reg_zone, reg);
3882 }
3883 
3884 static bool
3885 umtx_shm_unref_reg_locked(struct umtx_shm_reg *reg, bool force)
3886 {
3887 	bool res;
3888 
3889 	mtx_assert(&umtx_shm_lock, MA_OWNED);
3890 	KASSERT(reg->ushm_refcnt > 0, ("ushm_reg %p refcnt 0", reg));
3891 	reg->ushm_refcnt--;
3892 	res = reg->ushm_refcnt == 0;
3893 	if (res || force) {
3894 		if ((reg->ushm_flags & USHMF_REG_LINKED) != 0) {
3895 			TAILQ_REMOVE(&umtx_shm_registry[reg->ushm_key.hash],
3896 			    reg, ushm_reg_link);
3897 			reg->ushm_flags &= ~USHMF_REG_LINKED;
3898 		}
3899 		if ((reg->ushm_flags & USHMF_OBJ_LINKED) != 0) {
3900 			LIST_REMOVE(reg, ushm_obj_link);
3901 			reg->ushm_flags &= ~USHMF_OBJ_LINKED;
3902 		}
3903 	}
3904 	return (res);
3905 }
3906 
3907 static void
3908 umtx_shm_unref_reg(struct umtx_shm_reg *reg, bool force)
3909 {
3910 	vm_object_t object;
3911 	bool dofree;
3912 
3913 	if (force) {
3914 		object = reg->ushm_obj->shm_object;
3915 		VM_OBJECT_WLOCK(object);
3916 		object->flags |= OBJ_UMTXDEAD;
3917 		VM_OBJECT_WUNLOCK(object);
3918 	}
3919 	mtx_lock(&umtx_shm_lock);
3920 	dofree = umtx_shm_unref_reg_locked(reg, force);
3921 	mtx_unlock(&umtx_shm_lock);
3922 	if (dofree)
3923 		umtx_shm_free_reg(reg);
3924 }
3925 
3926 void
3927 umtx_shm_object_init(vm_object_t object)
3928 {
3929 
3930 	LIST_INIT(USHM_OBJ_UMTX(object));
3931 }
3932 
3933 void
3934 umtx_shm_object_terminated(vm_object_t object)
3935 {
3936 	struct umtx_shm_reg *reg, *reg1;
3937 	bool dofree;
3938 
3939 	if (LIST_EMPTY(USHM_OBJ_UMTX(object)))
3940 		return;
3941 
3942 	dofree = false;
3943 	mtx_lock(&umtx_shm_lock);
3944 	LIST_FOREACH_SAFE(reg, USHM_OBJ_UMTX(object), ushm_obj_link, reg1) {
3945 		if (umtx_shm_unref_reg_locked(reg, true)) {
3946 			TAILQ_INSERT_TAIL(&umtx_shm_reg_delfree, reg,
3947 			    ushm_reg_link);
3948 			dofree = true;
3949 		}
3950 	}
3951 	mtx_unlock(&umtx_shm_lock);
3952 	if (dofree)
3953 		taskqueue_enqueue(taskqueue_thread, &umtx_shm_reg_delfree_task);
3954 }
3955 
3956 static int
3957 umtx_shm_create_reg(struct thread *td, const struct umtx_key *key,
3958     struct umtx_shm_reg **res)
3959 {
3960 	struct umtx_shm_reg *reg, *reg1;
3961 	struct ucred *cred;
3962 	int error;
3963 
3964 	reg = umtx_shm_find_reg(key);
3965 	if (reg != NULL) {
3966 		*res = reg;
3967 		return (0);
3968 	}
3969 	cred = td->td_ucred;
3970 	if (!chgumtxcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_UMTXP)))
3971 		return (ENOMEM);
3972 	reg = uma_zalloc(umtx_shm_reg_zone, M_WAITOK | M_ZERO);
3973 	reg->ushm_refcnt = 1;
3974 	bcopy(key, &reg->ushm_key, sizeof(*key));
3975 	reg->ushm_obj = shm_alloc(td->td_ucred, O_RDWR);
3976 	reg->ushm_cred = crhold(cred);
3977 	error = shm_dotruncate(reg->ushm_obj, PAGE_SIZE);
3978 	if (error != 0) {
3979 		umtx_shm_free_reg(reg);
3980 		return (error);
3981 	}
3982 	mtx_lock(&umtx_shm_lock);
3983 	reg1 = umtx_shm_find_reg_locked(key);
3984 	if (reg1 != NULL) {
3985 		mtx_unlock(&umtx_shm_lock);
3986 		umtx_shm_free_reg(reg);
3987 		*res = reg1;
3988 		return (0);
3989 	}
3990 	reg->ushm_refcnt++;
3991 	TAILQ_INSERT_TAIL(&umtx_shm_registry[key->hash], reg, ushm_reg_link);
3992 	LIST_INSERT_HEAD(USHM_OBJ_UMTX(key->info.shared.object), reg,
3993 	    ushm_obj_link);
3994 	reg->ushm_flags = USHMF_REG_LINKED | USHMF_OBJ_LINKED;
3995 	mtx_unlock(&umtx_shm_lock);
3996 	*res = reg;
3997 	return (0);
3998 }
3999 
4000 static int
4001 umtx_shm_alive(struct thread *td, void *addr)
4002 {
4003 	vm_map_t map;
4004 	vm_map_entry_t entry;
4005 	vm_object_t object;
4006 	vm_pindex_t pindex;
4007 	vm_prot_t prot;
4008 	int res, ret;
4009 	boolean_t wired;
4010 
4011 	map = &td->td_proc->p_vmspace->vm_map;
4012 	res = vm_map_lookup(&map, (uintptr_t)addr, VM_PROT_READ, &entry,
4013 	    &object, &pindex, &prot, &wired);
4014 	if (res != KERN_SUCCESS)
4015 		return (EFAULT);
4016 	if (object == NULL)
4017 		ret = EINVAL;
4018 	else
4019 		ret = (object->flags & OBJ_UMTXDEAD) != 0 ? ENOTTY : 0;
4020 	vm_map_lookup_done(map, entry);
4021 	return (ret);
4022 }
4023 
4024 static void
4025 umtx_shm_init(void)
4026 {
4027 	int i;
4028 
4029 	umtx_shm_reg_zone = uma_zcreate("umtx_shm", sizeof(struct umtx_shm_reg),
4030 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
4031 	mtx_init(&umtx_shm_lock, "umtxshm", NULL, MTX_DEF);
4032 	for (i = 0; i < nitems(umtx_shm_registry); i++)
4033 		TAILQ_INIT(&umtx_shm_registry[i]);
4034 }
4035 
4036 static int
4037 umtx_shm(struct thread *td, void *addr, u_int flags)
4038 {
4039 	struct umtx_key key;
4040 	struct umtx_shm_reg *reg;
4041 	struct file *fp;
4042 	int error, fd;
4043 
4044 	if (__bitcount(flags & (UMTX_SHM_CREAT | UMTX_SHM_LOOKUP |
4045 	    UMTX_SHM_DESTROY| UMTX_SHM_ALIVE)) != 1)
4046 		return (EINVAL);
4047 	if ((flags & UMTX_SHM_ALIVE) != 0)
4048 		return (umtx_shm_alive(td, addr));
4049 	error = umtx_key_get(addr, TYPE_SHM, PROCESS_SHARE, &key);
4050 	if (error != 0)
4051 		return (error);
4052 	KASSERT(key.shared == 1, ("non-shared key"));
4053 	if ((flags & UMTX_SHM_CREAT) != 0) {
4054 		error = umtx_shm_create_reg(td, &key, &reg);
4055 	} else {
4056 		reg = umtx_shm_find_reg(&key);
4057 		if (reg == NULL)
4058 			error = ESRCH;
4059 	}
4060 	umtx_key_release(&key);
4061 	if (error != 0)
4062 		return (error);
4063 	KASSERT(reg != NULL, ("no reg"));
4064 	if ((flags & UMTX_SHM_DESTROY) != 0) {
4065 		umtx_shm_unref_reg(reg, true);
4066 	} else {
4067 #if 0
4068 #ifdef MAC
4069 		error = mac_posixshm_check_open(td->td_ucred,
4070 		    reg->ushm_obj, FFLAGS(O_RDWR));
4071 		if (error == 0)
4072 #endif
4073 			error = shm_access(reg->ushm_obj, td->td_ucred,
4074 			    FFLAGS(O_RDWR));
4075 		if (error == 0)
4076 #endif
4077 			error = falloc_caps(td, &fp, &fd, O_CLOEXEC, NULL);
4078 		if (error == 0) {
4079 			shm_hold(reg->ushm_obj);
4080 			finit(fp, FFLAGS(O_RDWR), DTYPE_SHM, reg->ushm_obj,
4081 			    &shm_ops);
4082 			td->td_retval[0] = fd;
4083 			fdrop(fp, td);
4084 		}
4085 	}
4086 	umtx_shm_unref_reg(reg, false);
4087 	return (error);
4088 }
4089 
4090 static int
4091 __umtx_op_shm(struct thread *td, struct _umtx_op_args *uap)
4092 {
4093 
4094 	return (umtx_shm(td, uap->uaddr1, uap->val));
4095 }
4096 
4097 static int
4098 umtx_robust_lists(struct thread *td, struct umtx_robust_lists_params *rbp)
4099 {
4100 
4101 	td->td_rb_list = rbp->robust_list_offset;
4102 	td->td_rbp_list = rbp->robust_priv_list_offset;
4103 	td->td_rb_inact = rbp->robust_inact_offset;
4104 	return (0);
4105 }
4106 
4107 static int
4108 __umtx_op_robust_lists(struct thread *td, struct _umtx_op_args *uap)
4109 {
4110 	struct umtx_robust_lists_params rb;
4111 	int error;
4112 
4113 	if (uap->val > sizeof(rb))
4114 		return (EINVAL);
4115 	bzero(&rb, sizeof(rb));
4116 	error = copyin(uap->uaddr1, &rb, uap->val);
4117 	if (error != 0)
4118 		return (error);
4119 	return (umtx_robust_lists(td, &rb));
4120 }
4121 
4122 typedef int (*_umtx_op_func)(struct thread *td, struct _umtx_op_args *uap);
4123 
4124 static const _umtx_op_func op_table[] = {
4125 	[UMTX_OP_RESERVED0]	= __umtx_op_unimpl,
4126 	[UMTX_OP_RESERVED1]	= __umtx_op_unimpl,
4127 	[UMTX_OP_WAIT]		= __umtx_op_wait,
4128 	[UMTX_OP_WAKE]		= __umtx_op_wake,
4129 	[UMTX_OP_MUTEX_TRYLOCK]	= __umtx_op_trylock_umutex,
4130 	[UMTX_OP_MUTEX_LOCK]	= __umtx_op_lock_umutex,
4131 	[UMTX_OP_MUTEX_UNLOCK]	= __umtx_op_unlock_umutex,
4132 	[UMTX_OP_SET_CEILING]	= __umtx_op_set_ceiling,
4133 	[UMTX_OP_CV_WAIT]	= __umtx_op_cv_wait,
4134 	[UMTX_OP_CV_SIGNAL]	= __umtx_op_cv_signal,
4135 	[UMTX_OP_CV_BROADCAST]	= __umtx_op_cv_broadcast,
4136 	[UMTX_OP_WAIT_UINT]	= __umtx_op_wait_uint,
4137 	[UMTX_OP_RW_RDLOCK]	= __umtx_op_rw_rdlock,
4138 	[UMTX_OP_RW_WRLOCK]	= __umtx_op_rw_wrlock,
4139 	[UMTX_OP_RW_UNLOCK]	= __umtx_op_rw_unlock,
4140 	[UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private,
4141 	[UMTX_OP_WAKE_PRIVATE]	= __umtx_op_wake_private,
4142 	[UMTX_OP_MUTEX_WAIT]	= __umtx_op_wait_umutex,
4143 	[UMTX_OP_MUTEX_WAKE]	= __umtx_op_wake_umutex,
4144 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4145 	[UMTX_OP_SEM_WAIT]	= __umtx_op_sem_wait,
4146 	[UMTX_OP_SEM_WAKE]	= __umtx_op_sem_wake,
4147 #else
4148 	[UMTX_OP_SEM_WAIT]	= __umtx_op_unimpl,
4149 	[UMTX_OP_SEM_WAKE]	= __umtx_op_unimpl,
4150 #endif
4151 	[UMTX_OP_NWAKE_PRIVATE]	= __umtx_op_nwake_private,
4152 	[UMTX_OP_MUTEX_WAKE2]	= __umtx_op_wake2_umutex,
4153 	[UMTX_OP_SEM2_WAIT]	= __umtx_op_sem2_wait,
4154 	[UMTX_OP_SEM2_WAKE]	= __umtx_op_sem2_wake,
4155 	[UMTX_OP_SHM]		= __umtx_op_shm,
4156 	[UMTX_OP_ROBUST_LISTS]	= __umtx_op_robust_lists,
4157 };
4158 
4159 int
4160 sys__umtx_op(struct thread *td, struct _umtx_op_args *uap)
4161 {
4162 
4163 	if ((unsigned)uap->op < nitems(op_table))
4164 		return (*op_table[uap->op])(td, uap);
4165 	return (EINVAL);
4166 }
4167 
4168 #ifdef COMPAT_FREEBSD32
4169 
4170 struct timespec32 {
4171 	int32_t tv_sec;
4172 	int32_t tv_nsec;
4173 };
4174 
4175 struct umtx_time32 {
4176 	struct	timespec32	timeout;
4177 	uint32_t		flags;
4178 	uint32_t		clockid;
4179 };
4180 
4181 static inline int
4182 umtx_copyin_timeout32(void *addr, struct timespec *tsp)
4183 {
4184 	struct timespec32 ts32;
4185 	int error;
4186 
4187 	error = copyin(addr, &ts32, sizeof(struct timespec32));
4188 	if (error == 0) {
4189 		if (ts32.tv_sec < 0 ||
4190 		    ts32.tv_nsec >= 1000000000 ||
4191 		    ts32.tv_nsec < 0)
4192 			error = EINVAL;
4193 		else {
4194 			tsp->tv_sec = ts32.tv_sec;
4195 			tsp->tv_nsec = ts32.tv_nsec;
4196 		}
4197 	}
4198 	return (error);
4199 }
4200 
4201 static inline int
4202 umtx_copyin_umtx_time32(const void *addr, size_t size, struct _umtx_time *tp)
4203 {
4204 	struct umtx_time32 t32;
4205 	int error;
4206 
4207 	t32.clockid = CLOCK_REALTIME;
4208 	t32.flags   = 0;
4209 	if (size <= sizeof(struct timespec32))
4210 		error = copyin(addr, &t32.timeout, sizeof(struct timespec32));
4211 	else
4212 		error = copyin(addr, &t32, sizeof(struct umtx_time32));
4213 	if (error != 0)
4214 		return (error);
4215 	if (t32.timeout.tv_sec < 0 ||
4216 	    t32.timeout.tv_nsec >= 1000000000 || t32.timeout.tv_nsec < 0)
4217 		return (EINVAL);
4218 	tp->_timeout.tv_sec = t32.timeout.tv_sec;
4219 	tp->_timeout.tv_nsec = t32.timeout.tv_nsec;
4220 	tp->_flags = t32.flags;
4221 	tp->_clockid = t32.clockid;
4222 	return (0);
4223 }
4224 
4225 static int
4226 __umtx_op_wait_compat32(struct thread *td, struct _umtx_op_args *uap)
4227 {
4228 	struct _umtx_time *tm_p, timeout;
4229 	int error;
4230 
4231 	if (uap->uaddr2 == NULL)
4232 		tm_p = NULL;
4233 	else {
4234 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4235 			(size_t)uap->uaddr1, &timeout);
4236 		if (error != 0)
4237 			return (error);
4238 		tm_p = &timeout;
4239 	}
4240 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
4241 }
4242 
4243 static int
4244 __umtx_op_lock_umutex_compat32(struct thread *td, struct _umtx_op_args *uap)
4245 {
4246 	struct _umtx_time *tm_p, timeout;
4247 	int error;
4248 
4249 	/* Allow a null timespec (wait forever). */
4250 	if (uap->uaddr2 == NULL)
4251 		tm_p = NULL;
4252 	else {
4253 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4254 			    (size_t)uap->uaddr1, &timeout);
4255 		if (error != 0)
4256 			return (error);
4257 		tm_p = &timeout;
4258 	}
4259 	return (do_lock_umutex(td, uap->obj, tm_p, 0));
4260 }
4261 
4262 static int
4263 __umtx_op_wait_umutex_compat32(struct thread *td, struct _umtx_op_args *uap)
4264 {
4265 	struct _umtx_time *tm_p, timeout;
4266 	int error;
4267 
4268 	/* Allow a null timespec (wait forever). */
4269 	if (uap->uaddr2 == NULL)
4270 		tm_p = NULL;
4271 	else {
4272 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4273 		    (size_t)uap->uaddr1, &timeout);
4274 		if (error != 0)
4275 			return (error);
4276 		tm_p = &timeout;
4277 	}
4278 	return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
4279 }
4280 
4281 static int
4282 __umtx_op_cv_wait_compat32(struct thread *td, struct _umtx_op_args *uap)
4283 {
4284 	struct timespec *ts, timeout;
4285 	int error;
4286 
4287 	/* Allow a null timespec (wait forever). */
4288 	if (uap->uaddr2 == NULL)
4289 		ts = NULL;
4290 	else {
4291 		error = umtx_copyin_timeout32(uap->uaddr2, &timeout);
4292 		if (error != 0)
4293 			return (error);
4294 		ts = &timeout;
4295 	}
4296 	return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
4297 }
4298 
4299 static int
4300 __umtx_op_rw_rdlock_compat32(struct thread *td, struct _umtx_op_args *uap)
4301 {
4302 	struct _umtx_time timeout;
4303 	int error;
4304 
4305 	/* Allow a null timespec (wait forever). */
4306 	if (uap->uaddr2 == NULL) {
4307 		error = do_rw_rdlock(td, uap->obj, uap->val, 0);
4308 	} else {
4309 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4310 		    (size_t)uap->uaddr1, &timeout);
4311 		if (error != 0)
4312 			return (error);
4313 		error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
4314 	}
4315 	return (error);
4316 }
4317 
4318 static int
4319 __umtx_op_rw_wrlock_compat32(struct thread *td, struct _umtx_op_args *uap)
4320 {
4321 	struct _umtx_time timeout;
4322 	int error;
4323 
4324 	/* Allow a null timespec (wait forever). */
4325 	if (uap->uaddr2 == NULL) {
4326 		error = do_rw_wrlock(td, uap->obj, 0);
4327 	} else {
4328 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4329 		    (size_t)uap->uaddr1, &timeout);
4330 		if (error != 0)
4331 			return (error);
4332 		error = do_rw_wrlock(td, uap->obj, &timeout);
4333 	}
4334 	return (error);
4335 }
4336 
4337 static int
4338 __umtx_op_wait_uint_private_compat32(struct thread *td, struct _umtx_op_args *uap)
4339 {
4340 	struct _umtx_time *tm_p, timeout;
4341 	int error;
4342 
4343 	if (uap->uaddr2 == NULL)
4344 		tm_p = NULL;
4345 	else {
4346 		error = umtx_copyin_umtx_time32(
4347 		    uap->uaddr2, (size_t)uap->uaddr1,&timeout);
4348 		if (error != 0)
4349 			return (error);
4350 		tm_p = &timeout;
4351 	}
4352 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
4353 }
4354 
4355 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4356 static int
4357 __umtx_op_sem_wait_compat32(struct thread *td, struct _umtx_op_args *uap)
4358 {
4359 	struct _umtx_time *tm_p, timeout;
4360 	int error;
4361 
4362 	/* Allow a null timespec (wait forever). */
4363 	if (uap->uaddr2 == NULL)
4364 		tm_p = NULL;
4365 	else {
4366 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4367 		    (size_t)uap->uaddr1, &timeout);
4368 		if (error != 0)
4369 			return (error);
4370 		tm_p = &timeout;
4371 	}
4372 	return (do_sem_wait(td, uap->obj, tm_p));
4373 }
4374 #endif
4375 
4376 static int
4377 __umtx_op_sem2_wait_compat32(struct thread *td, struct _umtx_op_args *uap)
4378 {
4379 	struct _umtx_time *tm_p, timeout;
4380 	size_t uasize;
4381 	int error;
4382 
4383 	/* Allow a null timespec (wait forever). */
4384 	if (uap->uaddr2 == NULL) {
4385 		uasize = 0;
4386 		tm_p = NULL;
4387 	} else {
4388 		uasize = (size_t)uap->uaddr1;
4389 		error = umtx_copyin_umtx_time32(uap->uaddr2, uasize, &timeout);
4390 		if (error != 0)
4391 			return (error);
4392 		tm_p = &timeout;
4393 	}
4394 	error = do_sem2_wait(td, uap->obj, tm_p);
4395 	if (error == EINTR && uap->uaddr2 != NULL &&
4396 	    (timeout._flags & UMTX_ABSTIME) == 0 &&
4397 	    uasize >= sizeof(struct umtx_time32) + sizeof(struct timespec32)) {
4398 		struct timespec32 remain32 = {
4399 			.tv_sec = timeout._timeout.tv_sec,
4400 			.tv_nsec = timeout._timeout.tv_nsec
4401 		};
4402 		error = copyout(&remain32,
4403 		    (struct umtx_time32 *)uap->uaddr2 + 1,
4404 		    sizeof(struct timespec32));
4405 		if (error == 0) {
4406 			error = EINTR;
4407 		}
4408 	}
4409 
4410 	return (error);
4411 }
4412 
4413 static int
4414 __umtx_op_nwake_private32(struct thread *td, struct _umtx_op_args *uap)
4415 {
4416 	uint32_t uaddrs[BATCH_SIZE], **upp;
4417 	int count, error, i, pos, tocopy;
4418 
4419 	upp = (uint32_t **)uap->obj;
4420 	error = 0;
4421 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
4422 	    pos += tocopy) {
4423 		tocopy = MIN(count, BATCH_SIZE);
4424 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(uint32_t));
4425 		if (error != 0)
4426 			break;
4427 		for (i = 0; i < tocopy; ++i)
4428 			kern_umtx_wake(td, (void *)(intptr_t)uaddrs[i],
4429 			    INT_MAX, 1);
4430 		maybe_yield();
4431 	}
4432 	return (error);
4433 }
4434 
4435 struct umtx_robust_lists_params_compat32 {
4436 	uint32_t	robust_list_offset;
4437 	uint32_t	robust_priv_list_offset;
4438 	uint32_t	robust_inact_offset;
4439 };
4440 
4441 static int
4442 __umtx_op_robust_lists_compat32(struct thread *td, struct _umtx_op_args *uap)
4443 {
4444 	struct umtx_robust_lists_params rb;
4445 	struct umtx_robust_lists_params_compat32 rb32;
4446 	int error;
4447 
4448 	if (uap->val > sizeof(rb32))
4449 		return (EINVAL);
4450 	bzero(&rb, sizeof(rb));
4451 	bzero(&rb32, sizeof(rb32));
4452 	error = copyin(uap->uaddr1, &rb32, uap->val);
4453 	if (error != 0)
4454 		return (error);
4455 	rb.robust_list_offset = rb32.robust_list_offset;
4456 	rb.robust_priv_list_offset = rb32.robust_priv_list_offset;
4457 	rb.robust_inact_offset = rb32.robust_inact_offset;
4458 	return (umtx_robust_lists(td, &rb));
4459 }
4460 
4461 static const _umtx_op_func op_table_compat32[] = {
4462 	[UMTX_OP_RESERVED0]	= __umtx_op_unimpl,
4463 	[UMTX_OP_RESERVED1]	= __umtx_op_unimpl,
4464 	[UMTX_OP_WAIT]		= __umtx_op_wait_compat32,
4465 	[UMTX_OP_WAKE]		= __umtx_op_wake,
4466 	[UMTX_OP_MUTEX_TRYLOCK]	= __umtx_op_trylock_umutex,
4467 	[UMTX_OP_MUTEX_LOCK]	= __umtx_op_lock_umutex_compat32,
4468 	[UMTX_OP_MUTEX_UNLOCK]	= __umtx_op_unlock_umutex,
4469 	[UMTX_OP_SET_CEILING]	= __umtx_op_set_ceiling,
4470 	[UMTX_OP_CV_WAIT]	= __umtx_op_cv_wait_compat32,
4471 	[UMTX_OP_CV_SIGNAL]	= __umtx_op_cv_signal,
4472 	[UMTX_OP_CV_BROADCAST]	= __umtx_op_cv_broadcast,
4473 	[UMTX_OP_WAIT_UINT]	= __umtx_op_wait_compat32,
4474 	[UMTX_OP_RW_RDLOCK]	= __umtx_op_rw_rdlock_compat32,
4475 	[UMTX_OP_RW_WRLOCK]	= __umtx_op_rw_wrlock_compat32,
4476 	[UMTX_OP_RW_UNLOCK]	= __umtx_op_rw_unlock,
4477 	[UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private_compat32,
4478 	[UMTX_OP_WAKE_PRIVATE]	= __umtx_op_wake_private,
4479 	[UMTX_OP_MUTEX_WAIT]	= __umtx_op_wait_umutex_compat32,
4480 	[UMTX_OP_MUTEX_WAKE]	= __umtx_op_wake_umutex,
4481 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4482 	[UMTX_OP_SEM_WAIT]	= __umtx_op_sem_wait_compat32,
4483 	[UMTX_OP_SEM_WAKE]	= __umtx_op_sem_wake,
4484 #else
4485 	[UMTX_OP_SEM_WAIT]	= __umtx_op_unimpl,
4486 	[UMTX_OP_SEM_WAKE]	= __umtx_op_unimpl,
4487 #endif
4488 	[UMTX_OP_NWAKE_PRIVATE]	= __umtx_op_nwake_private32,
4489 	[UMTX_OP_MUTEX_WAKE2]	= __umtx_op_wake2_umutex,
4490 	[UMTX_OP_SEM2_WAIT]	= __umtx_op_sem2_wait_compat32,
4491 	[UMTX_OP_SEM2_WAKE]	= __umtx_op_sem2_wake,
4492 	[UMTX_OP_SHM]		= __umtx_op_shm,
4493 	[UMTX_OP_ROBUST_LISTS]	= __umtx_op_robust_lists_compat32,
4494 };
4495 
4496 int
4497 freebsd32__umtx_op(struct thread *td, struct freebsd32__umtx_op_args *uap)
4498 {
4499 
4500 	if ((unsigned)uap->op < nitems(op_table_compat32)) {
4501 		return (*op_table_compat32[uap->op])(td,
4502 		    (struct _umtx_op_args *)uap);
4503 	}
4504 	return (EINVAL);
4505 }
4506 #endif
4507 
4508 void
4509 umtx_thread_init(struct thread *td)
4510 {
4511 
4512 	td->td_umtxq = umtxq_alloc();
4513 	td->td_umtxq->uq_thread = td;
4514 }
4515 
4516 void
4517 umtx_thread_fini(struct thread *td)
4518 {
4519 
4520 	umtxq_free(td->td_umtxq);
4521 }
4522 
4523 /*
4524  * It will be called when new thread is created, e.g fork().
4525  */
4526 void
4527 umtx_thread_alloc(struct thread *td)
4528 {
4529 	struct umtx_q *uq;
4530 
4531 	uq = td->td_umtxq;
4532 	uq->uq_inherited_pri = PRI_MAX;
4533 
4534 	KASSERT(uq->uq_flags == 0, ("uq_flags != 0"));
4535 	KASSERT(uq->uq_thread == td, ("uq_thread != td"));
4536 	KASSERT(uq->uq_pi_blocked == NULL, ("uq_pi_blocked != NULL"));
4537 	KASSERT(TAILQ_EMPTY(&uq->uq_pi_contested), ("uq_pi_contested is not empty"));
4538 }
4539 
4540 /*
4541  * exec() hook.
4542  *
4543  * Clear robust lists for all process' threads, not delaying the
4544  * cleanup to thread_exit hook, since the relevant address space is
4545  * destroyed right now.
4546  */
4547 static void
4548 umtx_exec_hook(void *arg __unused, struct proc *p,
4549     struct image_params *imgp __unused)
4550 {
4551 	struct thread *td;
4552 
4553 	KASSERT(p == curproc, ("need curproc"));
4554 	KASSERT((p->p_flag & P_HADTHREADS) == 0 ||
4555 	    (p->p_flag & P_STOPPED_SINGLE) != 0,
4556 	    ("curproc must be single-threaded"));
4557 	/*
4558 	 * There is no need to lock the list as only this thread can be
4559 	 * running.
4560 	 */
4561 	FOREACH_THREAD_IN_PROC(p, td) {
4562 		KASSERT(td == curthread ||
4563 		    ((td->td_flags & TDF_BOUNDARY) != 0 && TD_IS_SUSPENDED(td)),
4564 		    ("running thread %p %p", p, td));
4565 		umtx_thread_cleanup(td);
4566 		td->td_rb_list = td->td_rbp_list = td->td_rb_inact = 0;
4567 	}
4568 }
4569 
4570 /*
4571  * thread_exit() hook.
4572  */
4573 void
4574 umtx_thread_exit(struct thread *td)
4575 {
4576 
4577 	umtx_thread_cleanup(td);
4578 }
4579 
4580 static int
4581 umtx_read_uptr(struct thread *td, uintptr_t ptr, uintptr_t *res)
4582 {
4583 	u_long res1;
4584 #ifdef COMPAT_FREEBSD32
4585 	uint32_t res32;
4586 #endif
4587 	int error;
4588 
4589 #ifdef COMPAT_FREEBSD32
4590 	if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
4591 		error = fueword32((void *)ptr, &res32);
4592 		if (error == 0)
4593 			res1 = res32;
4594 	} else
4595 #endif
4596 	{
4597 		error = fueword((void *)ptr, &res1);
4598 	}
4599 	if (error == 0)
4600 		*res = res1;
4601 	else
4602 		error = EFAULT;
4603 	return (error);
4604 }
4605 
4606 static void
4607 umtx_read_rb_list(struct thread *td, struct umutex *m, uintptr_t *rb_list)
4608 {
4609 #ifdef COMPAT_FREEBSD32
4610 	struct umutex32 m32;
4611 
4612 	if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
4613 		memcpy(&m32, m, sizeof(m32));
4614 		*rb_list = m32.m_rb_lnk;
4615 	} else
4616 #endif
4617 		*rb_list = m->m_rb_lnk;
4618 }
4619 
4620 static int
4621 umtx_handle_rb(struct thread *td, uintptr_t rbp, uintptr_t *rb_list, bool inact)
4622 {
4623 	struct umutex m;
4624 	int error;
4625 
4626 	KASSERT(td->td_proc == curproc, ("need current vmspace"));
4627 	error = copyin((void *)rbp, &m, sizeof(m));
4628 	if (error != 0)
4629 		return (error);
4630 	if (rb_list != NULL)
4631 		umtx_read_rb_list(td, &m, rb_list);
4632 	if ((m.m_flags & UMUTEX_ROBUST) == 0)
4633 		return (EINVAL);
4634 	if ((m.m_owner & ~UMUTEX_CONTESTED) != td->td_tid)
4635 		/* inact is cleared after unlock, allow the inconsistency */
4636 		return (inact ? 0 : EINVAL);
4637 	return (do_unlock_umutex(td, (struct umutex *)rbp, true));
4638 }
4639 
4640 static void
4641 umtx_cleanup_rb_list(struct thread *td, uintptr_t rb_list, uintptr_t *rb_inact,
4642     const char *name)
4643 {
4644 	int error, i;
4645 	uintptr_t rbp;
4646 	bool inact;
4647 
4648 	if (rb_list == 0)
4649 		return;
4650 	error = umtx_read_uptr(td, rb_list, &rbp);
4651 	for (i = 0; error == 0 && rbp != 0 && i < umtx_max_rb; i++) {
4652 		if (rbp == *rb_inact) {
4653 			inact = true;
4654 			*rb_inact = 0;
4655 		} else
4656 			inact = false;
4657 		error = umtx_handle_rb(td, rbp, &rbp, inact);
4658 	}
4659 	if (i == umtx_max_rb && umtx_verbose_rb) {
4660 		uprintf("comm %s pid %d: reached umtx %smax rb %d\n",
4661 		    td->td_proc->p_comm, td->td_proc->p_pid, name, umtx_max_rb);
4662 	}
4663 	if (error != 0 && umtx_verbose_rb) {
4664 		uprintf("comm %s pid %d: handling %srb error %d\n",
4665 		    td->td_proc->p_comm, td->td_proc->p_pid, name, error);
4666 	}
4667 }
4668 
4669 /*
4670  * Clean up umtx data.
4671  */
4672 static void
4673 umtx_thread_cleanup(struct thread *td)
4674 {
4675 	struct umtx_q *uq;
4676 	struct umtx_pi *pi;
4677 	uintptr_t rb_inact;
4678 
4679 	/*
4680 	 * Disown pi mutexes.
4681 	 */
4682 	uq = td->td_umtxq;
4683 	if (uq != NULL) {
4684 		if (uq->uq_inherited_pri != PRI_MAX ||
4685 		    !TAILQ_EMPTY(&uq->uq_pi_contested)) {
4686 			mtx_lock(&umtx_lock);
4687 			uq->uq_inherited_pri = PRI_MAX;
4688 			while ((pi = TAILQ_FIRST(&uq->uq_pi_contested)) != NULL) {
4689 				pi->pi_owner = NULL;
4690 				TAILQ_REMOVE(&uq->uq_pi_contested, pi, pi_link);
4691 			}
4692 			mtx_unlock(&umtx_lock);
4693 		}
4694 		sched_lend_user_prio_cond(td, PRI_MAX);
4695 	}
4696 
4697 	if (td->td_rb_inact == 0 && td->td_rb_list == 0 && td->td_rbp_list == 0)
4698 		return;
4699 
4700 	/*
4701 	 * Handle terminated robust mutexes.  Must be done after
4702 	 * robust pi disown, otherwise unlock could see unowned
4703 	 * entries.
4704 	 */
4705 	rb_inact = td->td_rb_inact;
4706 	if (rb_inact != 0)
4707 		(void)umtx_read_uptr(td, rb_inact, &rb_inact);
4708 	umtx_cleanup_rb_list(td, td->td_rb_list, &rb_inact, "");
4709 	umtx_cleanup_rb_list(td, td->td_rbp_list, &rb_inact, "priv ");
4710 	if (rb_inact != 0)
4711 		(void)umtx_handle_rb(td, rb_inact, NULL, true);
4712 }
4713