xref: /freebsd/sys/kern/kern_umtx.c (revision 4072ae4e0291f97a2cc69fa4151988a99b06090c)
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_chain *uc;
666 	struct umtxq_queue *uh;
667 
668 	uc = umtxq_getchain(key);
669 	UMTXQ_LOCKED_ASSERT(uc);
670 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
671 	if (uh != NULL)
672 		return (uh->length);
673 	return (0);
674 }
675 
676 /*
677  * Check if there are multiple PI waiters and returns first
678  * waiter.
679  */
680 static int
681 umtxq_count_pi(struct umtx_key *key, struct umtx_q **first)
682 {
683 	struct umtxq_chain *uc;
684 	struct umtxq_queue *uh;
685 
686 	*first = NULL;
687 	uc = umtxq_getchain(key);
688 	UMTXQ_LOCKED_ASSERT(uc);
689 	uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
690 	if (uh != NULL) {
691 		*first = TAILQ_FIRST(&uh->head);
692 		return (uh->length);
693 	}
694 	return (0);
695 }
696 
697 static int
698 umtxq_check_susp(struct thread *td)
699 {
700 	struct proc *p;
701 	int error;
702 
703 	/*
704 	 * The check for TDF_NEEDSUSPCHK is racy, but it is enough to
705 	 * eventually break the lockstep loop.
706 	 */
707 	if ((td->td_flags & TDF_NEEDSUSPCHK) == 0)
708 		return (0);
709 	error = 0;
710 	p = td->td_proc;
711 	PROC_LOCK(p);
712 	if (P_SHOULDSTOP(p) ||
713 	    ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) {
714 		if (p->p_flag & P_SINGLE_EXIT)
715 			error = EINTR;
716 		else
717 			error = ERESTART;
718 	}
719 	PROC_UNLOCK(p);
720 	return (error);
721 }
722 
723 /*
724  * Wake up threads waiting on an userland object.
725  */
726 
727 static int
728 umtxq_signal_queue(struct umtx_key *key, int n_wake, int q)
729 {
730 	struct umtxq_chain *uc;
731 	struct umtxq_queue *uh;
732 	struct umtx_q *uq;
733 	int ret;
734 
735 	ret = 0;
736 	uc = umtxq_getchain(key);
737 	UMTXQ_LOCKED_ASSERT(uc);
738 	uh = umtxq_queue_lookup(key, q);
739 	if (uh != NULL) {
740 		while ((uq = TAILQ_FIRST(&uh->head)) != NULL) {
741 			umtxq_remove_queue(uq, q);
742 			wakeup(uq);
743 			if (++ret >= n_wake)
744 				return (ret);
745 		}
746 	}
747 	return (ret);
748 }
749 
750 
751 /*
752  * Wake up specified thread.
753  */
754 static inline void
755 umtxq_signal_thread(struct umtx_q *uq)
756 {
757 	struct umtxq_chain *uc;
758 
759 	uc = umtxq_getchain(&uq->uq_key);
760 	UMTXQ_LOCKED_ASSERT(uc);
761 	umtxq_remove(uq);
762 	wakeup(uq);
763 }
764 
765 static inline int
766 tstohz(const struct timespec *tsp)
767 {
768 	struct timeval tv;
769 
770 	TIMESPEC_TO_TIMEVAL(&tv, tsp);
771 	return tvtohz(&tv);
772 }
773 
774 static void
775 abs_timeout_init(struct abs_timeout *timo, int clockid, int absolute,
776 	const struct timespec *timeout)
777 {
778 
779 	timo->clockid = clockid;
780 	if (!absolute) {
781 		timo->is_abs_real = false;
782 		abs_timeout_update(timo);
783 		timo->end = timo->cur;
784 		timespecadd(&timo->end, timeout);
785 	} else {
786 		timo->end = *timeout;
787 		timo->is_abs_real = clockid == CLOCK_REALTIME ||
788 		    clockid == CLOCK_REALTIME_FAST ||
789 		    clockid == CLOCK_REALTIME_PRECISE;
790 		/*
791 		 * If is_abs_real, umtxq_sleep will read the clock
792 		 * after setting td_rtcgen; otherwise, read it here.
793 		 */
794 		if (!timo->is_abs_real) {
795 			abs_timeout_update(timo);
796 		}
797 	}
798 }
799 
800 static void
801 abs_timeout_init2(struct abs_timeout *timo, const struct _umtx_time *umtxtime)
802 {
803 
804 	abs_timeout_init(timo, umtxtime->_clockid,
805 	    (umtxtime->_flags & UMTX_ABSTIME) != 0, &umtxtime->_timeout);
806 }
807 
808 static inline void
809 abs_timeout_update(struct abs_timeout *timo)
810 {
811 
812 	kern_clock_gettime(curthread, timo->clockid, &timo->cur);
813 }
814 
815 static int
816 abs_timeout_gethz(struct abs_timeout *timo)
817 {
818 	struct timespec tts;
819 
820 	if (timespeccmp(&timo->end, &timo->cur, <=))
821 		return (-1);
822 	tts = timo->end;
823 	timespecsub(&tts, &timo->cur);
824 	return (tstohz(&tts));
825 }
826 
827 static uint32_t
828 umtx_unlock_val(uint32_t flags, bool rb)
829 {
830 
831 	if (rb)
832 		return (UMUTEX_RB_OWNERDEAD);
833 	else if ((flags & UMUTEX_NONCONSISTENT) != 0)
834 		return (UMUTEX_RB_NOTRECOV);
835 	else
836 		return (UMUTEX_UNOWNED);
837 
838 }
839 
840 /*
841  * Put thread into sleep state, before sleeping, check if
842  * thread was removed from umtx queue.
843  */
844 static inline int
845 umtxq_sleep(struct umtx_q *uq, const char *wmesg, struct abs_timeout *abstime)
846 {
847 	struct umtxq_chain *uc;
848 	int error, timo;
849 
850 	if (abstime != NULL && abstime->is_abs_real) {
851 		curthread->td_rtcgen = atomic_load_acq_int(&rtc_generation);
852 		abs_timeout_update(abstime);
853 	}
854 
855 	uc = umtxq_getchain(&uq->uq_key);
856 	UMTXQ_LOCKED_ASSERT(uc);
857 	for (;;) {
858 		if (!(uq->uq_flags & UQF_UMTXQ)) {
859 			error = 0;
860 			break;
861 		}
862 		if (abstime != NULL) {
863 			timo = abs_timeout_gethz(abstime);
864 			if (timo < 0) {
865 				error = ETIMEDOUT;
866 				break;
867 			}
868 		} else
869 			timo = 0;
870 		error = msleep(uq, &uc->uc_lock, PCATCH | PDROP, wmesg, timo);
871 		if (error == EINTR || error == ERESTART) {
872 			umtxq_lock(&uq->uq_key);
873 			break;
874 		}
875 		if (abstime != NULL) {
876 			if (abstime->is_abs_real)
877 				curthread->td_rtcgen =
878 				    atomic_load_acq_int(&rtc_generation);
879 			abs_timeout_update(abstime);
880 		}
881 		umtxq_lock(&uq->uq_key);
882 	}
883 
884 	curthread->td_rtcgen = 0;
885 	return (error);
886 }
887 
888 /*
889  * Convert userspace address into unique logical address.
890  */
891 int
892 umtx_key_get(const void *addr, int type, int share, struct umtx_key *key)
893 {
894 	struct thread *td = curthread;
895 	vm_map_t map;
896 	vm_map_entry_t entry;
897 	vm_pindex_t pindex;
898 	vm_prot_t prot;
899 	boolean_t wired;
900 
901 	key->type = type;
902 	if (share == THREAD_SHARE) {
903 		key->shared = 0;
904 		key->info.private.vs = td->td_proc->p_vmspace;
905 		key->info.private.addr = (uintptr_t)addr;
906 	} else {
907 		MPASS(share == PROCESS_SHARE || share == AUTO_SHARE);
908 		map = &td->td_proc->p_vmspace->vm_map;
909 		if (vm_map_lookup(&map, (vm_offset_t)addr, VM_PROT_WRITE,
910 		    &entry, &key->info.shared.object, &pindex, &prot,
911 		    &wired) != KERN_SUCCESS) {
912 			return (EFAULT);
913 		}
914 
915 		if ((share == PROCESS_SHARE) ||
916 		    (share == AUTO_SHARE &&
917 		     VM_INHERIT_SHARE == entry->inheritance)) {
918 			key->shared = 1;
919 			key->info.shared.offset = (vm_offset_t)addr -
920 			    entry->start + entry->offset;
921 			vm_object_reference(key->info.shared.object);
922 		} else {
923 			key->shared = 0;
924 			key->info.private.vs = td->td_proc->p_vmspace;
925 			key->info.private.addr = (uintptr_t)addr;
926 		}
927 		vm_map_lookup_done(map, entry);
928 	}
929 
930 	umtxq_hash(key);
931 	return (0);
932 }
933 
934 /*
935  * Release key.
936  */
937 void
938 umtx_key_release(struct umtx_key *key)
939 {
940 	if (key->shared)
941 		vm_object_deallocate(key->info.shared.object);
942 }
943 
944 /*
945  * Fetch and compare value, sleep on the address if value is not changed.
946  */
947 static int
948 do_wait(struct thread *td, void *addr, u_long id,
949     struct _umtx_time *timeout, int compat32, int is_private)
950 {
951 	struct abs_timeout timo;
952 	struct umtx_q *uq;
953 	u_long tmp;
954 	uint32_t tmp32;
955 	int error = 0;
956 
957 	uq = td->td_umtxq;
958 	if ((error = umtx_key_get(addr, TYPE_SIMPLE_WAIT,
959 		is_private ? THREAD_SHARE : AUTO_SHARE, &uq->uq_key)) != 0)
960 		return (error);
961 
962 	if (timeout != NULL)
963 		abs_timeout_init2(&timo, timeout);
964 
965 	umtxq_lock(&uq->uq_key);
966 	umtxq_insert(uq);
967 	umtxq_unlock(&uq->uq_key);
968 	if (compat32 == 0) {
969 		error = fueword(addr, &tmp);
970 		if (error != 0)
971 			error = EFAULT;
972 	} else {
973 		error = fueword32(addr, &tmp32);
974 		if (error == 0)
975 			tmp = tmp32;
976 		else
977 			error = EFAULT;
978 	}
979 	umtxq_lock(&uq->uq_key);
980 	if (error == 0) {
981 		if (tmp == id)
982 			error = umtxq_sleep(uq, "uwait", timeout == NULL ?
983 			    NULL : &timo);
984 		if ((uq->uq_flags & UQF_UMTXQ) == 0)
985 			error = 0;
986 		else
987 			umtxq_remove(uq);
988 	} else if ((uq->uq_flags & UQF_UMTXQ) != 0) {
989 		umtxq_remove(uq);
990 	}
991 	umtxq_unlock(&uq->uq_key);
992 	umtx_key_release(&uq->uq_key);
993 	if (error == ERESTART)
994 		error = EINTR;
995 	return (error);
996 }
997 
998 /*
999  * Wake up threads sleeping on the specified address.
1000  */
1001 int
1002 kern_umtx_wake(struct thread *td, void *uaddr, int n_wake, int is_private)
1003 {
1004 	struct umtx_key key;
1005 	int ret;
1006 
1007 	if ((ret = umtx_key_get(uaddr, TYPE_SIMPLE_WAIT,
1008 	    is_private ? THREAD_SHARE : AUTO_SHARE, &key)) != 0)
1009 		return (ret);
1010 	umtxq_lock(&key);
1011 	umtxq_signal(&key, n_wake);
1012 	umtxq_unlock(&key);
1013 	umtx_key_release(&key);
1014 	return (0);
1015 }
1016 
1017 /*
1018  * Lock PTHREAD_PRIO_NONE protocol POSIX mutex.
1019  */
1020 static int
1021 do_lock_normal(struct thread *td, struct umutex *m, uint32_t flags,
1022     struct _umtx_time *timeout, int mode)
1023 {
1024 	struct abs_timeout timo;
1025 	struct umtx_q *uq;
1026 	uint32_t owner, old, id;
1027 	int error, rv;
1028 
1029 	id = td->td_tid;
1030 	uq = td->td_umtxq;
1031 	error = 0;
1032 	if (timeout != NULL)
1033 		abs_timeout_init2(&timo, timeout);
1034 
1035 	/*
1036 	 * Care must be exercised when dealing with umtx structure. It
1037 	 * can fault on any access.
1038 	 */
1039 	for (;;) {
1040 		rv = fueword32(&m->m_owner, &owner);
1041 		if (rv == -1)
1042 			return (EFAULT);
1043 		if (mode == _UMUTEX_WAIT) {
1044 			if (owner == UMUTEX_UNOWNED ||
1045 			    owner == UMUTEX_CONTESTED ||
1046 			    owner == UMUTEX_RB_OWNERDEAD ||
1047 			    owner == UMUTEX_RB_NOTRECOV)
1048 				return (0);
1049 		} else {
1050 			/*
1051 			 * Robust mutex terminated.  Kernel duty is to
1052 			 * return EOWNERDEAD to the userspace.  The
1053 			 * umutex.m_flags UMUTEX_NONCONSISTENT is set
1054 			 * by the common userspace code.
1055 			 */
1056 			if (owner == UMUTEX_RB_OWNERDEAD) {
1057 				rv = casueword32(&m->m_owner,
1058 				    UMUTEX_RB_OWNERDEAD, &owner,
1059 				    id | UMUTEX_CONTESTED);
1060 				if (rv == -1)
1061 					return (EFAULT);
1062 				if (owner == UMUTEX_RB_OWNERDEAD)
1063 					return (EOWNERDEAD); /* success */
1064 				rv = umtxq_check_susp(td);
1065 				if (rv != 0)
1066 					return (rv);
1067 				continue;
1068 			}
1069 			if (owner == UMUTEX_RB_NOTRECOV)
1070 				return (ENOTRECOVERABLE);
1071 
1072 
1073 			/*
1074 			 * Try the uncontested case.  This should be
1075 			 * done in userland.
1076 			 */
1077 			rv = casueword32(&m->m_owner, UMUTEX_UNOWNED,
1078 			    &owner, id);
1079 			/* The address was invalid. */
1080 			if (rv == -1)
1081 				return (EFAULT);
1082 
1083 			/* The acquire succeeded. */
1084 			if (owner == UMUTEX_UNOWNED)
1085 				return (0);
1086 
1087 			/*
1088 			 * If no one owns it but it is contested try
1089 			 * to acquire it.
1090 			 */
1091 			if (owner == UMUTEX_CONTESTED) {
1092 				rv = casueword32(&m->m_owner,
1093 				    UMUTEX_CONTESTED, &owner,
1094 				    id | UMUTEX_CONTESTED);
1095 				/* The address was invalid. */
1096 				if (rv == -1)
1097 					return (EFAULT);
1098 
1099 				if (owner == UMUTEX_CONTESTED)
1100 					return (0);
1101 
1102 				rv = umtxq_check_susp(td);
1103 				if (rv != 0)
1104 					return (rv);
1105 
1106 				/*
1107 				 * If this failed the lock has
1108 				 * changed, restart.
1109 				 */
1110 				continue;
1111 			}
1112 		}
1113 
1114 		if (mode == _UMUTEX_TRY)
1115 			return (EBUSY);
1116 
1117 		/*
1118 		 * If we caught a signal, we have retried and now
1119 		 * exit immediately.
1120 		 */
1121 		if (error != 0)
1122 			return (error);
1123 
1124 		if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX,
1125 		    GET_SHARE(flags), &uq->uq_key)) != 0)
1126 			return (error);
1127 
1128 		umtxq_lock(&uq->uq_key);
1129 		umtxq_busy(&uq->uq_key);
1130 		umtxq_insert(uq);
1131 		umtxq_unlock(&uq->uq_key);
1132 
1133 		/*
1134 		 * Set the contested bit so that a release in user space
1135 		 * knows to use the system call for unlock.  If this fails
1136 		 * either some one else has acquired the lock or it has been
1137 		 * released.
1138 		 */
1139 		rv = casueword32(&m->m_owner, owner, &old,
1140 		    owner | UMUTEX_CONTESTED);
1141 
1142 		/* The address was invalid. */
1143 		if (rv == -1) {
1144 			umtxq_lock(&uq->uq_key);
1145 			umtxq_remove(uq);
1146 			umtxq_unbusy(&uq->uq_key);
1147 			umtxq_unlock(&uq->uq_key);
1148 			umtx_key_release(&uq->uq_key);
1149 			return (EFAULT);
1150 		}
1151 
1152 		/*
1153 		 * We set the contested bit, sleep. Otherwise the lock changed
1154 		 * and we need to retry or we lost a race to the thread
1155 		 * unlocking the umtx.
1156 		 */
1157 		umtxq_lock(&uq->uq_key);
1158 		umtxq_unbusy(&uq->uq_key);
1159 		if (old == owner)
1160 			error = umtxq_sleep(uq, "umtxn", timeout == NULL ?
1161 			    NULL : &timo);
1162 		umtxq_remove(uq);
1163 		umtxq_unlock(&uq->uq_key);
1164 		umtx_key_release(&uq->uq_key);
1165 
1166 		if (error == 0)
1167 			error = umtxq_check_susp(td);
1168 	}
1169 
1170 	return (0);
1171 }
1172 
1173 /*
1174  * Unlock PTHREAD_PRIO_NONE protocol POSIX mutex.
1175  */
1176 static int
1177 do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1178 {
1179 	struct umtx_key key;
1180 	uint32_t owner, old, id, newlock;
1181 	int error, count;
1182 
1183 	id = td->td_tid;
1184 	/*
1185 	 * Make sure we own this mtx.
1186 	 */
1187 	error = fueword32(&m->m_owner, &owner);
1188 	if (error == -1)
1189 		return (EFAULT);
1190 
1191 	if ((owner & ~UMUTEX_CONTESTED) != id)
1192 		return (EPERM);
1193 
1194 	newlock = umtx_unlock_val(flags, rb);
1195 	if ((owner & UMUTEX_CONTESTED) == 0) {
1196 		error = casueword32(&m->m_owner, owner, &old, newlock);
1197 		if (error == -1)
1198 			return (EFAULT);
1199 		if (old == owner)
1200 			return (0);
1201 		owner = old;
1202 	}
1203 
1204 	/* We should only ever be in here for contested locks */
1205 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1206 	    &key)) != 0)
1207 		return (error);
1208 
1209 	umtxq_lock(&key);
1210 	umtxq_busy(&key);
1211 	count = umtxq_count(&key);
1212 	umtxq_unlock(&key);
1213 
1214 	/*
1215 	 * When unlocking the umtx, it must be marked as unowned if
1216 	 * there is zero or one thread only waiting for it.
1217 	 * Otherwise, it must be marked as contested.
1218 	 */
1219 	if (count > 1)
1220 		newlock |= UMUTEX_CONTESTED;
1221 	error = casueword32(&m->m_owner, owner, &old, newlock);
1222 	umtxq_lock(&key);
1223 	umtxq_signal(&key, 1);
1224 	umtxq_unbusy(&key);
1225 	umtxq_unlock(&key);
1226 	umtx_key_release(&key);
1227 	if (error == -1)
1228 		return (EFAULT);
1229 	if (old != owner)
1230 		return (EINVAL);
1231 	return (0);
1232 }
1233 
1234 /*
1235  * Check if the mutex is available and wake up a waiter,
1236  * only for simple mutex.
1237  */
1238 static int
1239 do_wake_umutex(struct thread *td, struct umutex *m)
1240 {
1241 	struct umtx_key key;
1242 	uint32_t owner;
1243 	uint32_t flags;
1244 	int error;
1245 	int count;
1246 
1247 	error = fueword32(&m->m_owner, &owner);
1248 	if (error == -1)
1249 		return (EFAULT);
1250 
1251 	if ((owner & ~UMUTEX_CONTESTED) != 0 && owner != UMUTEX_RB_OWNERDEAD &&
1252 	    owner != UMUTEX_RB_NOTRECOV)
1253 		return (0);
1254 
1255 	error = fueword32(&m->m_flags, &flags);
1256 	if (error == -1)
1257 		return (EFAULT);
1258 
1259 	/* We should only ever be in here for contested locks */
1260 	if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1261 	    &key)) != 0)
1262 		return (error);
1263 
1264 	umtxq_lock(&key);
1265 	umtxq_busy(&key);
1266 	count = umtxq_count(&key);
1267 	umtxq_unlock(&key);
1268 
1269 	if (count <= 1 && owner != UMUTEX_RB_OWNERDEAD &&
1270 	    owner != UMUTEX_RB_NOTRECOV) {
1271 		error = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
1272 		    UMUTEX_UNOWNED);
1273 		if (error == -1)
1274 			error = EFAULT;
1275 	}
1276 
1277 	umtxq_lock(&key);
1278 	if (error == 0 && count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1279 	    owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1280 		umtxq_signal(&key, 1);
1281 	umtxq_unbusy(&key);
1282 	umtxq_unlock(&key);
1283 	umtx_key_release(&key);
1284 	return (error);
1285 }
1286 
1287 /*
1288  * Check if the mutex has waiters and tries to fix contention bit.
1289  */
1290 static int
1291 do_wake2_umutex(struct thread *td, struct umutex *m, uint32_t flags)
1292 {
1293 	struct umtx_key key;
1294 	uint32_t owner, old;
1295 	int type;
1296 	int error;
1297 	int count;
1298 
1299 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT |
1300 	    UMUTEX_ROBUST)) {
1301 	case 0:
1302 	case UMUTEX_ROBUST:
1303 		type = TYPE_NORMAL_UMUTEX;
1304 		break;
1305 	case UMUTEX_PRIO_INHERIT:
1306 		type = TYPE_PI_UMUTEX;
1307 		break;
1308 	case (UMUTEX_PRIO_INHERIT | UMUTEX_ROBUST):
1309 		type = TYPE_PI_ROBUST_UMUTEX;
1310 		break;
1311 	case UMUTEX_PRIO_PROTECT:
1312 		type = TYPE_PP_UMUTEX;
1313 		break;
1314 	case (UMUTEX_PRIO_PROTECT | UMUTEX_ROBUST):
1315 		type = TYPE_PP_ROBUST_UMUTEX;
1316 		break;
1317 	default:
1318 		return (EINVAL);
1319 	}
1320 	if ((error = umtx_key_get(m, type, GET_SHARE(flags), &key)) != 0)
1321 		return (error);
1322 
1323 	owner = 0;
1324 	umtxq_lock(&key);
1325 	umtxq_busy(&key);
1326 	count = umtxq_count(&key);
1327 	umtxq_unlock(&key);
1328 	/*
1329 	 * Only repair contention bit if there is a waiter, this means the mutex
1330 	 * is still being referenced by userland code, otherwise don't update
1331 	 * any memory.
1332 	 */
1333 	if (count > 1) {
1334 		error = fueword32(&m->m_owner, &owner);
1335 		if (error == -1)
1336 			error = EFAULT;
1337 		while (error == 0 && (owner & UMUTEX_CONTESTED) == 0) {
1338 			error = casueword32(&m->m_owner, owner, &old,
1339 			    owner | UMUTEX_CONTESTED);
1340 			if (error == -1) {
1341 				error = EFAULT;
1342 				break;
1343 			}
1344 			if (old == owner)
1345 				break;
1346 			owner = old;
1347 			error = umtxq_check_susp(td);
1348 			if (error != 0)
1349 				break;
1350 		}
1351 	} else if (count == 1) {
1352 		error = fueword32(&m->m_owner, &owner);
1353 		if (error == -1)
1354 			error = EFAULT;
1355 		while (error == 0 && (owner & ~UMUTEX_CONTESTED) != 0 &&
1356 		    (owner & UMUTEX_CONTESTED) == 0) {
1357 			error = casueword32(&m->m_owner, owner, &old,
1358 			    owner | UMUTEX_CONTESTED);
1359 			if (error == -1) {
1360 				error = EFAULT;
1361 				break;
1362 			}
1363 			if (old == owner)
1364 				break;
1365 			owner = old;
1366 			error = umtxq_check_susp(td);
1367 			if (error != 0)
1368 				break;
1369 		}
1370 	}
1371 	umtxq_lock(&key);
1372 	if (error == EFAULT) {
1373 		umtxq_signal(&key, INT_MAX);
1374 	} else if (count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1375 	    owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1376 		umtxq_signal(&key, 1);
1377 	umtxq_unbusy(&key);
1378 	umtxq_unlock(&key);
1379 	umtx_key_release(&key);
1380 	return (error);
1381 }
1382 
1383 static inline struct umtx_pi *
1384 umtx_pi_alloc(int flags)
1385 {
1386 	struct umtx_pi *pi;
1387 
1388 	pi = uma_zalloc(umtx_pi_zone, M_ZERO | flags);
1389 	TAILQ_INIT(&pi->pi_blocked);
1390 	atomic_add_int(&umtx_pi_allocated, 1);
1391 	return (pi);
1392 }
1393 
1394 static inline void
1395 umtx_pi_free(struct umtx_pi *pi)
1396 {
1397 	uma_zfree(umtx_pi_zone, pi);
1398 	atomic_add_int(&umtx_pi_allocated, -1);
1399 }
1400 
1401 /*
1402  * Adjust the thread's position on a pi_state after its priority has been
1403  * changed.
1404  */
1405 static int
1406 umtx_pi_adjust_thread(struct umtx_pi *pi, struct thread *td)
1407 {
1408 	struct umtx_q *uq, *uq1, *uq2;
1409 	struct thread *td1;
1410 
1411 	mtx_assert(&umtx_lock, MA_OWNED);
1412 	if (pi == NULL)
1413 		return (0);
1414 
1415 	uq = td->td_umtxq;
1416 
1417 	/*
1418 	 * Check if the thread needs to be moved on the blocked chain.
1419 	 * It needs to be moved if either its priority is lower than
1420 	 * the previous thread or higher than the next thread.
1421 	 */
1422 	uq1 = TAILQ_PREV(uq, umtxq_head, uq_lockq);
1423 	uq2 = TAILQ_NEXT(uq, uq_lockq);
1424 	if ((uq1 != NULL && UPRI(td) < UPRI(uq1->uq_thread)) ||
1425 	    (uq2 != NULL && UPRI(td) > UPRI(uq2->uq_thread))) {
1426 		/*
1427 		 * Remove thread from blocked chain and determine where
1428 		 * it should be moved to.
1429 		 */
1430 		TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1431 		TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1432 			td1 = uq1->uq_thread;
1433 			MPASS(td1->td_proc->p_magic == P_MAGIC);
1434 			if (UPRI(td1) > UPRI(td))
1435 				break;
1436 		}
1437 
1438 		if (uq1 == NULL)
1439 			TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1440 		else
1441 			TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1442 	}
1443 	return (1);
1444 }
1445 
1446 static struct umtx_pi *
1447 umtx_pi_next(struct umtx_pi *pi)
1448 {
1449 	struct umtx_q *uq_owner;
1450 
1451 	if (pi->pi_owner == NULL)
1452 		return (NULL);
1453 	uq_owner = pi->pi_owner->td_umtxq;
1454 	if (uq_owner == NULL)
1455 		return (NULL);
1456 	return (uq_owner->uq_pi_blocked);
1457 }
1458 
1459 /*
1460  * Floyd's Cycle-Finding Algorithm.
1461  */
1462 static bool
1463 umtx_pi_check_loop(struct umtx_pi *pi)
1464 {
1465 	struct umtx_pi *pi1;	/* fast iterator */
1466 
1467 	mtx_assert(&umtx_lock, MA_OWNED);
1468 	if (pi == NULL)
1469 		return (false);
1470 	pi1 = pi;
1471 	for (;;) {
1472 		pi = umtx_pi_next(pi);
1473 		if (pi == NULL)
1474 			break;
1475 		pi1 = umtx_pi_next(pi1);
1476 		if (pi1 == NULL)
1477 			break;
1478 		pi1 = umtx_pi_next(pi1);
1479 		if (pi1 == NULL)
1480 			break;
1481 		if (pi == pi1)
1482 			return (true);
1483 	}
1484 	return (false);
1485 }
1486 
1487 /*
1488  * Propagate priority when a thread is blocked on POSIX
1489  * PI mutex.
1490  */
1491 static void
1492 umtx_propagate_priority(struct thread *td)
1493 {
1494 	struct umtx_q *uq;
1495 	struct umtx_pi *pi;
1496 	int pri;
1497 
1498 	mtx_assert(&umtx_lock, MA_OWNED);
1499 	pri = UPRI(td);
1500 	uq = td->td_umtxq;
1501 	pi = uq->uq_pi_blocked;
1502 	if (pi == NULL)
1503 		return;
1504 	if (umtx_pi_check_loop(pi))
1505 		return;
1506 
1507 	for (;;) {
1508 		td = pi->pi_owner;
1509 		if (td == NULL || td == curthread)
1510 			return;
1511 
1512 		MPASS(td->td_proc != NULL);
1513 		MPASS(td->td_proc->p_magic == P_MAGIC);
1514 
1515 		thread_lock(td);
1516 		if (td->td_lend_user_pri > pri)
1517 			sched_lend_user_prio(td, pri);
1518 		else {
1519 			thread_unlock(td);
1520 			break;
1521 		}
1522 		thread_unlock(td);
1523 
1524 		/*
1525 		 * Pick up the lock that td is blocked on.
1526 		 */
1527 		uq = td->td_umtxq;
1528 		pi = uq->uq_pi_blocked;
1529 		if (pi == NULL)
1530 			break;
1531 		/* Resort td on the list if needed. */
1532 		umtx_pi_adjust_thread(pi, td);
1533 	}
1534 }
1535 
1536 /*
1537  * Unpropagate priority for a PI mutex when a thread blocked on
1538  * it is interrupted by signal or resumed by others.
1539  */
1540 static void
1541 umtx_repropagate_priority(struct umtx_pi *pi)
1542 {
1543 	struct umtx_q *uq, *uq_owner;
1544 	struct umtx_pi *pi2;
1545 	int pri;
1546 
1547 	mtx_assert(&umtx_lock, MA_OWNED);
1548 
1549 	if (umtx_pi_check_loop(pi))
1550 		return;
1551 	while (pi != NULL && pi->pi_owner != NULL) {
1552 		pri = PRI_MAX;
1553 		uq_owner = pi->pi_owner->td_umtxq;
1554 
1555 		TAILQ_FOREACH(pi2, &uq_owner->uq_pi_contested, pi_link) {
1556 			uq = TAILQ_FIRST(&pi2->pi_blocked);
1557 			if (uq != NULL) {
1558 				if (pri > UPRI(uq->uq_thread))
1559 					pri = UPRI(uq->uq_thread);
1560 			}
1561 		}
1562 
1563 		if (pri > uq_owner->uq_inherited_pri)
1564 			pri = uq_owner->uq_inherited_pri;
1565 		thread_lock(pi->pi_owner);
1566 		sched_lend_user_prio(pi->pi_owner, pri);
1567 		thread_unlock(pi->pi_owner);
1568 		if ((pi = uq_owner->uq_pi_blocked) != NULL)
1569 			umtx_pi_adjust_thread(pi, uq_owner->uq_thread);
1570 	}
1571 }
1572 
1573 /*
1574  * Insert a PI mutex into owned list.
1575  */
1576 static void
1577 umtx_pi_setowner(struct umtx_pi *pi, struct thread *owner)
1578 {
1579 	struct umtx_q *uq_owner;
1580 
1581 	uq_owner = owner->td_umtxq;
1582 	mtx_assert(&umtx_lock, MA_OWNED);
1583 	MPASS(pi->pi_owner == NULL);
1584 	pi->pi_owner = owner;
1585 	TAILQ_INSERT_TAIL(&uq_owner->uq_pi_contested, pi, pi_link);
1586 }
1587 
1588 
1589 /*
1590  * Disown a PI mutex, and remove it from the owned list.
1591  */
1592 static void
1593 umtx_pi_disown(struct umtx_pi *pi)
1594 {
1595 
1596 	mtx_assert(&umtx_lock, MA_OWNED);
1597 	TAILQ_REMOVE(&pi->pi_owner->td_umtxq->uq_pi_contested, pi, pi_link);
1598 	pi->pi_owner = NULL;
1599 }
1600 
1601 /*
1602  * Claim ownership of a PI mutex.
1603  */
1604 static int
1605 umtx_pi_claim(struct umtx_pi *pi, struct thread *owner)
1606 {
1607 	struct umtx_q *uq;
1608 	int pri;
1609 
1610 	mtx_lock(&umtx_lock);
1611 	if (pi->pi_owner == owner) {
1612 		mtx_unlock(&umtx_lock);
1613 		return (0);
1614 	}
1615 
1616 	if (pi->pi_owner != NULL) {
1617 		/*
1618 		 * userland may have already messed the mutex, sigh.
1619 		 */
1620 		mtx_unlock(&umtx_lock);
1621 		return (EPERM);
1622 	}
1623 	umtx_pi_setowner(pi, owner);
1624 	uq = TAILQ_FIRST(&pi->pi_blocked);
1625 	if (uq != NULL) {
1626 		pri = UPRI(uq->uq_thread);
1627 		thread_lock(owner);
1628 		if (pri < UPRI(owner))
1629 			sched_lend_user_prio(owner, pri);
1630 		thread_unlock(owner);
1631 	}
1632 	mtx_unlock(&umtx_lock);
1633 	return (0);
1634 }
1635 
1636 /*
1637  * Adjust a thread's order position in its blocked PI mutex,
1638  * this may result new priority propagating process.
1639  */
1640 void
1641 umtx_pi_adjust(struct thread *td, u_char oldpri)
1642 {
1643 	struct umtx_q *uq;
1644 	struct umtx_pi *pi;
1645 
1646 	uq = td->td_umtxq;
1647 	mtx_lock(&umtx_lock);
1648 	/*
1649 	 * Pick up the lock that td is blocked on.
1650 	 */
1651 	pi = uq->uq_pi_blocked;
1652 	if (pi != NULL) {
1653 		umtx_pi_adjust_thread(pi, td);
1654 		umtx_repropagate_priority(pi);
1655 	}
1656 	mtx_unlock(&umtx_lock);
1657 }
1658 
1659 /*
1660  * Sleep on a PI mutex.
1661  */
1662 static int
1663 umtxq_sleep_pi(struct umtx_q *uq, struct umtx_pi *pi, uint32_t owner,
1664     const char *wmesg, struct abs_timeout *timo, bool shared)
1665 {
1666 	struct umtxq_chain *uc;
1667 	struct thread *td, *td1;
1668 	struct umtx_q *uq1;
1669 	int error, pri;
1670 
1671 	error = 0;
1672 	td = uq->uq_thread;
1673 	KASSERT(td == curthread, ("inconsistent uq_thread"));
1674 	uc = umtxq_getchain(&uq->uq_key);
1675 	UMTXQ_LOCKED_ASSERT(uc);
1676 	KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));
1677 	umtxq_insert(uq);
1678 	mtx_lock(&umtx_lock);
1679 	if (pi->pi_owner == NULL) {
1680 		mtx_unlock(&umtx_lock);
1681 		td1 = tdfind(owner, shared ? -1 : td->td_proc->p_pid);
1682 		mtx_lock(&umtx_lock);
1683 		if (td1 != NULL) {
1684 			if (pi->pi_owner == NULL)
1685 				umtx_pi_setowner(pi, td1);
1686 			PROC_UNLOCK(td1->td_proc);
1687 		}
1688 	}
1689 
1690 	TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1691 		pri = UPRI(uq1->uq_thread);
1692 		if (pri > UPRI(td))
1693 			break;
1694 	}
1695 
1696 	if (uq1 != NULL)
1697 		TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1698 	else
1699 		TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1700 
1701 	uq->uq_pi_blocked = pi;
1702 	thread_lock(td);
1703 	td->td_flags |= TDF_UPIBLOCKED;
1704 	thread_unlock(td);
1705 	umtx_propagate_priority(td);
1706 	mtx_unlock(&umtx_lock);
1707 	umtxq_unbusy(&uq->uq_key);
1708 
1709 	error = umtxq_sleep(uq, wmesg, timo);
1710 	umtxq_remove(uq);
1711 
1712 	mtx_lock(&umtx_lock);
1713 	uq->uq_pi_blocked = NULL;
1714 	thread_lock(td);
1715 	td->td_flags &= ~TDF_UPIBLOCKED;
1716 	thread_unlock(td);
1717 	TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1718 	umtx_repropagate_priority(pi);
1719 	mtx_unlock(&umtx_lock);
1720 	umtxq_unlock(&uq->uq_key);
1721 
1722 	return (error);
1723 }
1724 
1725 /*
1726  * Add reference count for a PI mutex.
1727  */
1728 static void
1729 umtx_pi_ref(struct umtx_pi *pi)
1730 {
1731 	struct umtxq_chain *uc;
1732 
1733 	uc = umtxq_getchain(&pi->pi_key);
1734 	UMTXQ_LOCKED_ASSERT(uc);
1735 	pi->pi_refcount++;
1736 }
1737 
1738 /*
1739  * Decrease reference count for a PI mutex, if the counter
1740  * is decreased to zero, its memory space is freed.
1741  */
1742 static void
1743 umtx_pi_unref(struct umtx_pi *pi)
1744 {
1745 	struct umtxq_chain *uc;
1746 
1747 	uc = umtxq_getchain(&pi->pi_key);
1748 	UMTXQ_LOCKED_ASSERT(uc);
1749 	KASSERT(pi->pi_refcount > 0, ("invalid reference count"));
1750 	if (--pi->pi_refcount == 0) {
1751 		mtx_lock(&umtx_lock);
1752 		if (pi->pi_owner != NULL)
1753 			umtx_pi_disown(pi);
1754 		KASSERT(TAILQ_EMPTY(&pi->pi_blocked),
1755 			("blocked queue not empty"));
1756 		mtx_unlock(&umtx_lock);
1757 		TAILQ_REMOVE(&uc->uc_pi_list, pi, pi_hashlink);
1758 		umtx_pi_free(pi);
1759 	}
1760 }
1761 
1762 /*
1763  * Find a PI mutex in hash table.
1764  */
1765 static struct umtx_pi *
1766 umtx_pi_lookup(struct umtx_key *key)
1767 {
1768 	struct umtxq_chain *uc;
1769 	struct umtx_pi *pi;
1770 
1771 	uc = umtxq_getchain(key);
1772 	UMTXQ_LOCKED_ASSERT(uc);
1773 
1774 	TAILQ_FOREACH(pi, &uc->uc_pi_list, pi_hashlink) {
1775 		if (umtx_key_match(&pi->pi_key, key)) {
1776 			return (pi);
1777 		}
1778 	}
1779 	return (NULL);
1780 }
1781 
1782 /*
1783  * Insert a PI mutex into hash table.
1784  */
1785 static inline void
1786 umtx_pi_insert(struct umtx_pi *pi)
1787 {
1788 	struct umtxq_chain *uc;
1789 
1790 	uc = umtxq_getchain(&pi->pi_key);
1791 	UMTXQ_LOCKED_ASSERT(uc);
1792 	TAILQ_INSERT_TAIL(&uc->uc_pi_list, pi, pi_hashlink);
1793 }
1794 
1795 /*
1796  * Lock a PI mutex.
1797  */
1798 static int
1799 do_lock_pi(struct thread *td, struct umutex *m, uint32_t flags,
1800     struct _umtx_time *timeout, int try)
1801 {
1802 	struct abs_timeout timo;
1803 	struct umtx_q *uq;
1804 	struct umtx_pi *pi, *new_pi;
1805 	uint32_t id, old_owner, owner, old;
1806 	int error, rv;
1807 
1808 	id = td->td_tid;
1809 	uq = td->td_umtxq;
1810 
1811 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
1812 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
1813 	    &uq->uq_key)) != 0)
1814 		return (error);
1815 
1816 	if (timeout != NULL)
1817 		abs_timeout_init2(&timo, timeout);
1818 
1819 	umtxq_lock(&uq->uq_key);
1820 	pi = umtx_pi_lookup(&uq->uq_key);
1821 	if (pi == NULL) {
1822 		new_pi = umtx_pi_alloc(M_NOWAIT);
1823 		if (new_pi == NULL) {
1824 			umtxq_unlock(&uq->uq_key);
1825 			new_pi = umtx_pi_alloc(M_WAITOK);
1826 			umtxq_lock(&uq->uq_key);
1827 			pi = umtx_pi_lookup(&uq->uq_key);
1828 			if (pi != NULL) {
1829 				umtx_pi_free(new_pi);
1830 				new_pi = NULL;
1831 			}
1832 		}
1833 		if (new_pi != NULL) {
1834 			new_pi->pi_key = uq->uq_key;
1835 			umtx_pi_insert(new_pi);
1836 			pi = new_pi;
1837 		}
1838 	}
1839 	umtx_pi_ref(pi);
1840 	umtxq_unlock(&uq->uq_key);
1841 
1842 	/*
1843 	 * Care must be exercised when dealing with umtx structure.  It
1844 	 * can fault on any access.
1845 	 */
1846 	for (;;) {
1847 		/*
1848 		 * Try the uncontested case.  This should be done in userland.
1849 		 */
1850 		rv = casueword32(&m->m_owner, UMUTEX_UNOWNED, &owner, id);
1851 		/* The address was invalid. */
1852 		if (rv == -1) {
1853 			error = EFAULT;
1854 			break;
1855 		}
1856 
1857 		/* The acquire succeeded. */
1858 		if (owner == UMUTEX_UNOWNED) {
1859 			error = 0;
1860 			break;
1861 		}
1862 
1863 		if (owner == UMUTEX_RB_NOTRECOV) {
1864 			error = ENOTRECOVERABLE;
1865 			break;
1866 		}
1867 
1868 		/* If no one owns it but it is contested try to acquire it. */
1869 		if (owner == UMUTEX_CONTESTED || owner == UMUTEX_RB_OWNERDEAD) {
1870 			old_owner = owner;
1871 			rv = casueword32(&m->m_owner, owner, &owner,
1872 			    id | UMUTEX_CONTESTED);
1873 			/* The address was invalid. */
1874 			if (rv == -1) {
1875 				error = EFAULT;
1876 				break;
1877 			}
1878 
1879 			if (owner == old_owner) {
1880 				umtxq_lock(&uq->uq_key);
1881 				umtxq_busy(&uq->uq_key);
1882 				error = umtx_pi_claim(pi, td);
1883 				umtxq_unbusy(&uq->uq_key);
1884 				umtxq_unlock(&uq->uq_key);
1885 				if (error != 0) {
1886 					/*
1887 					 * Since we're going to return an
1888 					 * error, restore the m_owner to its
1889 					 * previous, unowned state to avoid
1890 					 * compounding the problem.
1891 					 */
1892 					(void)casuword32(&m->m_owner,
1893 					    id | UMUTEX_CONTESTED,
1894 					    old_owner);
1895 				}
1896 				if (error == 0 &&
1897 				    old_owner == UMUTEX_RB_OWNERDEAD)
1898 					error = EOWNERDEAD;
1899 				break;
1900 			}
1901 
1902 			error = umtxq_check_susp(td);
1903 			if (error != 0)
1904 				break;
1905 
1906 			/* If this failed the lock has changed, restart. */
1907 			continue;
1908 		}
1909 
1910 		if ((owner & ~UMUTEX_CONTESTED) == id) {
1911 			error = EDEADLK;
1912 			break;
1913 		}
1914 
1915 		if (try != 0) {
1916 			error = EBUSY;
1917 			break;
1918 		}
1919 
1920 		/*
1921 		 * If we caught a signal, we have retried and now
1922 		 * exit immediately.
1923 		 */
1924 		if (error != 0)
1925 			break;
1926 
1927 		umtxq_lock(&uq->uq_key);
1928 		umtxq_busy(&uq->uq_key);
1929 		umtxq_unlock(&uq->uq_key);
1930 
1931 		/*
1932 		 * Set the contested bit so that a release in user space
1933 		 * knows to use the system call for unlock.  If this fails
1934 		 * either some one else has acquired the lock or it has been
1935 		 * released.
1936 		 */
1937 		rv = casueword32(&m->m_owner, owner, &old, owner |
1938 		    UMUTEX_CONTESTED);
1939 
1940 		/* The address was invalid. */
1941 		if (rv == -1) {
1942 			umtxq_unbusy_unlocked(&uq->uq_key);
1943 			error = EFAULT;
1944 			break;
1945 		}
1946 
1947 		umtxq_lock(&uq->uq_key);
1948 		/*
1949 		 * We set the contested bit, sleep. Otherwise the lock changed
1950 		 * and we need to retry or we lost a race to the thread
1951 		 * unlocking the umtx.  Note that the UMUTEX_RB_OWNERDEAD
1952 		 * value for owner is impossible there.
1953 		 */
1954 		if (old == owner) {
1955 			error = umtxq_sleep_pi(uq, pi,
1956 			    owner & ~UMUTEX_CONTESTED,
1957 			    "umtxpi", timeout == NULL ? NULL : &timo,
1958 			    (flags & USYNC_PROCESS_SHARED) != 0);
1959 			if (error != 0)
1960 				continue;
1961 		} else {
1962 			umtxq_unbusy(&uq->uq_key);
1963 			umtxq_unlock(&uq->uq_key);
1964 		}
1965 
1966 		error = umtxq_check_susp(td);
1967 		if (error != 0)
1968 			break;
1969 	}
1970 
1971 	umtxq_lock(&uq->uq_key);
1972 	umtx_pi_unref(pi);
1973 	umtxq_unlock(&uq->uq_key);
1974 
1975 	umtx_key_release(&uq->uq_key);
1976 	return (error);
1977 }
1978 
1979 /*
1980  * Unlock a PI mutex.
1981  */
1982 static int
1983 do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1984 {
1985 	struct umtx_key key;
1986 	struct umtx_q *uq_first, *uq_first2, *uq_me;
1987 	struct umtx_pi *pi, *pi2;
1988 	uint32_t id, new_owner, old, owner;
1989 	int count, error, pri;
1990 
1991 	id = td->td_tid;
1992 	/*
1993 	 * Make sure we own this mtx.
1994 	 */
1995 	error = fueword32(&m->m_owner, &owner);
1996 	if (error == -1)
1997 		return (EFAULT);
1998 
1999 	if ((owner & ~UMUTEX_CONTESTED) != id)
2000 		return (EPERM);
2001 
2002 	new_owner = umtx_unlock_val(flags, rb);
2003 
2004 	/* This should be done in userland */
2005 	if ((owner & UMUTEX_CONTESTED) == 0) {
2006 		error = casueword32(&m->m_owner, owner, &old, new_owner);
2007 		if (error == -1)
2008 			return (EFAULT);
2009 		if (old == owner)
2010 			return (0);
2011 		owner = old;
2012 	}
2013 
2014 	/* We should only ever be in here for contested locks */
2015 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2016 	    TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2017 	    &key)) != 0)
2018 		return (error);
2019 
2020 	umtxq_lock(&key);
2021 	umtxq_busy(&key);
2022 	count = umtxq_count_pi(&key, &uq_first);
2023 	if (uq_first != NULL) {
2024 		mtx_lock(&umtx_lock);
2025 		pi = uq_first->uq_pi_blocked;
2026 		KASSERT(pi != NULL, ("pi == NULL?"));
2027 		if (pi->pi_owner != td && !(rb && pi->pi_owner == NULL)) {
2028 			mtx_unlock(&umtx_lock);
2029 			umtxq_unbusy(&key);
2030 			umtxq_unlock(&key);
2031 			umtx_key_release(&key);
2032 			/* userland messed the mutex */
2033 			return (EPERM);
2034 		}
2035 		uq_me = td->td_umtxq;
2036 		if (pi->pi_owner == td)
2037 			umtx_pi_disown(pi);
2038 		/* get highest priority thread which is still sleeping. */
2039 		uq_first = TAILQ_FIRST(&pi->pi_blocked);
2040 		while (uq_first != NULL &&
2041 		    (uq_first->uq_flags & UQF_UMTXQ) == 0) {
2042 			uq_first = TAILQ_NEXT(uq_first, uq_lockq);
2043 		}
2044 		pri = PRI_MAX;
2045 		TAILQ_FOREACH(pi2, &uq_me->uq_pi_contested, pi_link) {
2046 			uq_first2 = TAILQ_FIRST(&pi2->pi_blocked);
2047 			if (uq_first2 != NULL) {
2048 				if (pri > UPRI(uq_first2->uq_thread))
2049 					pri = UPRI(uq_first2->uq_thread);
2050 			}
2051 		}
2052 		thread_lock(td);
2053 		sched_lend_user_prio(td, pri);
2054 		thread_unlock(td);
2055 		mtx_unlock(&umtx_lock);
2056 		if (uq_first)
2057 			umtxq_signal_thread(uq_first);
2058 	} else {
2059 		pi = umtx_pi_lookup(&key);
2060 		/*
2061 		 * A umtx_pi can exist if a signal or timeout removed the
2062 		 * last waiter from the umtxq, but there is still
2063 		 * a thread in do_lock_pi() holding the umtx_pi.
2064 		 */
2065 		if (pi != NULL) {
2066 			/*
2067 			 * The umtx_pi can be unowned, such as when a thread
2068 			 * has just entered do_lock_pi(), allocated the
2069 			 * umtx_pi, and unlocked the umtxq.
2070 			 * If the current thread owns it, it must disown it.
2071 			 */
2072 			mtx_lock(&umtx_lock);
2073 			if (pi->pi_owner == td)
2074 				umtx_pi_disown(pi);
2075 			mtx_unlock(&umtx_lock);
2076 		}
2077 	}
2078 	umtxq_unlock(&key);
2079 
2080 	/*
2081 	 * When unlocking the umtx, it must be marked as unowned if
2082 	 * there is zero or one thread only waiting for it.
2083 	 * Otherwise, it must be marked as contested.
2084 	 */
2085 
2086 	if (count > 1)
2087 		new_owner |= UMUTEX_CONTESTED;
2088 	error = casueword32(&m->m_owner, owner, &old, new_owner);
2089 
2090 	umtxq_unbusy_unlocked(&key);
2091 	umtx_key_release(&key);
2092 	if (error == -1)
2093 		return (EFAULT);
2094 	if (old != owner)
2095 		return (EINVAL);
2096 	return (0);
2097 }
2098 
2099 /*
2100  * Lock a PP mutex.
2101  */
2102 static int
2103 do_lock_pp(struct thread *td, struct umutex *m, uint32_t flags,
2104     struct _umtx_time *timeout, int try)
2105 {
2106 	struct abs_timeout timo;
2107 	struct umtx_q *uq, *uq2;
2108 	struct umtx_pi *pi;
2109 	uint32_t ceiling;
2110 	uint32_t owner, id;
2111 	int error, pri, old_inherited_pri, su, rv;
2112 
2113 	id = td->td_tid;
2114 	uq = td->td_umtxq;
2115 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2116 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2117 	    &uq->uq_key)) != 0)
2118 		return (error);
2119 
2120 	if (timeout != NULL)
2121 		abs_timeout_init2(&timo, timeout);
2122 
2123 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2124 	for (;;) {
2125 		old_inherited_pri = uq->uq_inherited_pri;
2126 		umtxq_lock(&uq->uq_key);
2127 		umtxq_busy(&uq->uq_key);
2128 		umtxq_unlock(&uq->uq_key);
2129 
2130 		rv = fueword32(&m->m_ceilings[0], &ceiling);
2131 		if (rv == -1) {
2132 			error = EFAULT;
2133 			goto out;
2134 		}
2135 		ceiling = RTP_PRIO_MAX - ceiling;
2136 		if (ceiling > RTP_PRIO_MAX) {
2137 			error = EINVAL;
2138 			goto out;
2139 		}
2140 
2141 		mtx_lock(&umtx_lock);
2142 		if (UPRI(td) < PRI_MIN_REALTIME + ceiling) {
2143 			mtx_unlock(&umtx_lock);
2144 			error = EINVAL;
2145 			goto out;
2146 		}
2147 		if (su && PRI_MIN_REALTIME + ceiling < uq->uq_inherited_pri) {
2148 			uq->uq_inherited_pri = PRI_MIN_REALTIME + ceiling;
2149 			thread_lock(td);
2150 			if (uq->uq_inherited_pri < UPRI(td))
2151 				sched_lend_user_prio(td, uq->uq_inherited_pri);
2152 			thread_unlock(td);
2153 		}
2154 		mtx_unlock(&umtx_lock);
2155 
2156 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2157 		    id | UMUTEX_CONTESTED);
2158 		/* The address was invalid. */
2159 		if (rv == -1) {
2160 			error = EFAULT;
2161 			break;
2162 		}
2163 
2164 		if (owner == UMUTEX_CONTESTED) {
2165 			error = 0;
2166 			break;
2167 		} else if (owner == UMUTEX_RB_OWNERDEAD) {
2168 			rv = casueword32(&m->m_owner, UMUTEX_RB_OWNERDEAD,
2169 			    &owner, id | UMUTEX_CONTESTED);
2170 			if (rv == -1) {
2171 				error = EFAULT;
2172 				break;
2173 			}
2174 			if (owner == UMUTEX_RB_OWNERDEAD) {
2175 				error = EOWNERDEAD; /* success */
2176 				break;
2177 			}
2178 			error = 0;
2179 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2180 			error = ENOTRECOVERABLE;
2181 			break;
2182 		}
2183 
2184 		if (try != 0) {
2185 			error = EBUSY;
2186 			break;
2187 		}
2188 
2189 		/*
2190 		 * If we caught a signal, we have retried and now
2191 		 * exit immediately.
2192 		 */
2193 		if (error != 0)
2194 			break;
2195 
2196 		umtxq_lock(&uq->uq_key);
2197 		umtxq_insert(uq);
2198 		umtxq_unbusy(&uq->uq_key);
2199 		error = umtxq_sleep(uq, "umtxpp", timeout == NULL ?
2200 		    NULL : &timo);
2201 		umtxq_remove(uq);
2202 		umtxq_unlock(&uq->uq_key);
2203 
2204 		mtx_lock(&umtx_lock);
2205 		uq->uq_inherited_pri = old_inherited_pri;
2206 		pri = PRI_MAX;
2207 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2208 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2209 			if (uq2 != NULL) {
2210 				if (pri > UPRI(uq2->uq_thread))
2211 					pri = UPRI(uq2->uq_thread);
2212 			}
2213 		}
2214 		if (pri > uq->uq_inherited_pri)
2215 			pri = uq->uq_inherited_pri;
2216 		thread_lock(td);
2217 		sched_lend_user_prio(td, pri);
2218 		thread_unlock(td);
2219 		mtx_unlock(&umtx_lock);
2220 	}
2221 
2222 	if (error != 0 && error != EOWNERDEAD) {
2223 		mtx_lock(&umtx_lock);
2224 		uq->uq_inherited_pri = old_inherited_pri;
2225 		pri = PRI_MAX;
2226 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2227 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2228 			if (uq2 != NULL) {
2229 				if (pri > UPRI(uq2->uq_thread))
2230 					pri = UPRI(uq2->uq_thread);
2231 			}
2232 		}
2233 		if (pri > uq->uq_inherited_pri)
2234 			pri = uq->uq_inherited_pri;
2235 		thread_lock(td);
2236 		sched_lend_user_prio(td, pri);
2237 		thread_unlock(td);
2238 		mtx_unlock(&umtx_lock);
2239 	}
2240 
2241 out:
2242 	umtxq_unbusy_unlocked(&uq->uq_key);
2243 	umtx_key_release(&uq->uq_key);
2244 	return (error);
2245 }
2246 
2247 /*
2248  * Unlock a PP mutex.
2249  */
2250 static int
2251 do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2252 {
2253 	struct umtx_key key;
2254 	struct umtx_q *uq, *uq2;
2255 	struct umtx_pi *pi;
2256 	uint32_t id, owner, rceiling;
2257 	int error, pri, new_inherited_pri, su;
2258 
2259 	id = td->td_tid;
2260 	uq = td->td_umtxq;
2261 	su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2262 
2263 	/*
2264 	 * Make sure we own this mtx.
2265 	 */
2266 	error = fueword32(&m->m_owner, &owner);
2267 	if (error == -1)
2268 		return (EFAULT);
2269 
2270 	if ((owner & ~UMUTEX_CONTESTED) != id)
2271 		return (EPERM);
2272 
2273 	error = copyin(&m->m_ceilings[1], &rceiling, sizeof(uint32_t));
2274 	if (error != 0)
2275 		return (error);
2276 
2277 	if (rceiling == -1)
2278 		new_inherited_pri = PRI_MAX;
2279 	else {
2280 		rceiling = RTP_PRIO_MAX - rceiling;
2281 		if (rceiling > RTP_PRIO_MAX)
2282 			return (EINVAL);
2283 		new_inherited_pri = PRI_MIN_REALTIME + rceiling;
2284 	}
2285 
2286 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2287 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2288 	    &key)) != 0)
2289 		return (error);
2290 	umtxq_lock(&key);
2291 	umtxq_busy(&key);
2292 	umtxq_unlock(&key);
2293 	/*
2294 	 * For priority protected mutex, always set unlocked state
2295 	 * to UMUTEX_CONTESTED, so that userland always enters kernel
2296 	 * to lock the mutex, it is necessary because thread priority
2297 	 * has to be adjusted for such mutex.
2298 	 */
2299 	error = suword32(&m->m_owner, umtx_unlock_val(flags, rb) |
2300 	    UMUTEX_CONTESTED);
2301 
2302 	umtxq_lock(&key);
2303 	if (error == 0)
2304 		umtxq_signal(&key, 1);
2305 	umtxq_unbusy(&key);
2306 	umtxq_unlock(&key);
2307 
2308 	if (error == -1)
2309 		error = EFAULT;
2310 	else {
2311 		mtx_lock(&umtx_lock);
2312 		if (su != 0)
2313 			uq->uq_inherited_pri = new_inherited_pri;
2314 		pri = PRI_MAX;
2315 		TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2316 			uq2 = TAILQ_FIRST(&pi->pi_blocked);
2317 			if (uq2 != NULL) {
2318 				if (pri > UPRI(uq2->uq_thread))
2319 					pri = UPRI(uq2->uq_thread);
2320 			}
2321 		}
2322 		if (pri > uq->uq_inherited_pri)
2323 			pri = uq->uq_inherited_pri;
2324 		thread_lock(td);
2325 		sched_lend_user_prio(td, pri);
2326 		thread_unlock(td);
2327 		mtx_unlock(&umtx_lock);
2328 	}
2329 	umtx_key_release(&key);
2330 	return (error);
2331 }
2332 
2333 static int
2334 do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
2335     uint32_t *old_ceiling)
2336 {
2337 	struct umtx_q *uq;
2338 	uint32_t flags, id, owner, save_ceiling;
2339 	int error, rv, rv1;
2340 
2341 	error = fueword32(&m->m_flags, &flags);
2342 	if (error == -1)
2343 		return (EFAULT);
2344 	if ((flags & UMUTEX_PRIO_PROTECT) == 0)
2345 		return (EINVAL);
2346 	if (ceiling > RTP_PRIO_MAX)
2347 		return (EINVAL);
2348 	id = td->td_tid;
2349 	uq = td->td_umtxq;
2350 	if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2351 	    TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2352 	    &uq->uq_key)) != 0)
2353 		return (error);
2354 	for (;;) {
2355 		umtxq_lock(&uq->uq_key);
2356 		umtxq_busy(&uq->uq_key);
2357 		umtxq_unlock(&uq->uq_key);
2358 
2359 		rv = fueword32(&m->m_ceilings[0], &save_ceiling);
2360 		if (rv == -1) {
2361 			error = EFAULT;
2362 			break;
2363 		}
2364 
2365 		rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2366 		    id | UMUTEX_CONTESTED);
2367 		if (rv == -1) {
2368 			error = EFAULT;
2369 			break;
2370 		}
2371 
2372 		if (owner == UMUTEX_CONTESTED) {
2373 			rv = suword32(&m->m_ceilings[0], ceiling);
2374 			rv1 = suword32(&m->m_owner, UMUTEX_CONTESTED);
2375 			error = (rv == 0 && rv1 == 0) ? 0: EFAULT;
2376 			break;
2377 		}
2378 
2379 		if ((owner & ~UMUTEX_CONTESTED) == id) {
2380 			rv = suword32(&m->m_ceilings[0], ceiling);
2381 			error = rv == 0 ? 0 : EFAULT;
2382 			break;
2383 		}
2384 
2385 		if (owner == UMUTEX_RB_OWNERDEAD) {
2386 			error = EOWNERDEAD;
2387 			break;
2388 		} else if (owner == UMUTEX_RB_NOTRECOV) {
2389 			error = ENOTRECOVERABLE;
2390 			break;
2391 		}
2392 
2393 		/*
2394 		 * If we caught a signal, we have retried and now
2395 		 * exit immediately.
2396 		 */
2397 		if (error != 0)
2398 			break;
2399 
2400 		/*
2401 		 * We set the contested bit, sleep. Otherwise the lock changed
2402 		 * and we need to retry or we lost a race to the thread
2403 		 * unlocking the umtx.
2404 		 */
2405 		umtxq_lock(&uq->uq_key);
2406 		umtxq_insert(uq);
2407 		umtxq_unbusy(&uq->uq_key);
2408 		error = umtxq_sleep(uq, "umtxpp", NULL);
2409 		umtxq_remove(uq);
2410 		umtxq_unlock(&uq->uq_key);
2411 	}
2412 	umtxq_lock(&uq->uq_key);
2413 	if (error == 0)
2414 		umtxq_signal(&uq->uq_key, INT_MAX);
2415 	umtxq_unbusy(&uq->uq_key);
2416 	umtxq_unlock(&uq->uq_key);
2417 	umtx_key_release(&uq->uq_key);
2418 	if (error == 0 && old_ceiling != NULL) {
2419 		rv = suword32(old_ceiling, save_ceiling);
2420 		error = rv == 0 ? 0 : EFAULT;
2421 	}
2422 	return (error);
2423 }
2424 
2425 /*
2426  * Lock a userland POSIX mutex.
2427  */
2428 static int
2429 do_lock_umutex(struct thread *td, struct umutex *m,
2430     struct _umtx_time *timeout, int mode)
2431 {
2432 	uint32_t flags;
2433 	int error;
2434 
2435 	error = fueword32(&m->m_flags, &flags);
2436 	if (error == -1)
2437 		return (EFAULT);
2438 
2439 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2440 	case 0:
2441 		error = do_lock_normal(td, m, flags, timeout, mode);
2442 		break;
2443 	case UMUTEX_PRIO_INHERIT:
2444 		error = do_lock_pi(td, m, flags, timeout, mode);
2445 		break;
2446 	case UMUTEX_PRIO_PROTECT:
2447 		error = do_lock_pp(td, m, flags, timeout, mode);
2448 		break;
2449 	default:
2450 		return (EINVAL);
2451 	}
2452 	if (timeout == NULL) {
2453 		if (error == EINTR && mode != _UMUTEX_WAIT)
2454 			error = ERESTART;
2455 	} else {
2456 		/* Timed-locking is not restarted. */
2457 		if (error == ERESTART)
2458 			error = EINTR;
2459 	}
2460 	return (error);
2461 }
2462 
2463 /*
2464  * Unlock a userland POSIX mutex.
2465  */
2466 static int
2467 do_unlock_umutex(struct thread *td, struct umutex *m, bool rb)
2468 {
2469 	uint32_t flags;
2470 	int error;
2471 
2472 	error = fueword32(&m->m_flags, &flags);
2473 	if (error == -1)
2474 		return (EFAULT);
2475 
2476 	switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2477 	case 0:
2478 		return (do_unlock_normal(td, m, flags, rb));
2479 	case UMUTEX_PRIO_INHERIT:
2480 		return (do_unlock_pi(td, m, flags, rb));
2481 	case UMUTEX_PRIO_PROTECT:
2482 		return (do_unlock_pp(td, m, flags, rb));
2483 	}
2484 
2485 	return (EINVAL);
2486 }
2487 
2488 static int
2489 do_cv_wait(struct thread *td, struct ucond *cv, struct umutex *m,
2490     struct timespec *timeout, u_long wflags)
2491 {
2492 	struct abs_timeout timo;
2493 	struct umtx_q *uq;
2494 	uint32_t flags, clockid, hasw;
2495 	int error;
2496 
2497 	uq = td->td_umtxq;
2498 	error = fueword32(&cv->c_flags, &flags);
2499 	if (error == -1)
2500 		return (EFAULT);
2501 	error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &uq->uq_key);
2502 	if (error != 0)
2503 		return (error);
2504 
2505 	if ((wflags & CVWAIT_CLOCKID) != 0) {
2506 		error = fueword32(&cv->c_clockid, &clockid);
2507 		if (error == -1) {
2508 			umtx_key_release(&uq->uq_key);
2509 			return (EFAULT);
2510 		}
2511 		if (clockid < CLOCK_REALTIME ||
2512 		    clockid >= CLOCK_THREAD_CPUTIME_ID) {
2513 			/* hmm, only HW clock id will work. */
2514 			umtx_key_release(&uq->uq_key);
2515 			return (EINVAL);
2516 		}
2517 	} else {
2518 		clockid = CLOCK_REALTIME;
2519 	}
2520 
2521 	umtxq_lock(&uq->uq_key);
2522 	umtxq_busy(&uq->uq_key);
2523 	umtxq_insert(uq);
2524 	umtxq_unlock(&uq->uq_key);
2525 
2526 	/*
2527 	 * Set c_has_waiters to 1 before releasing user mutex, also
2528 	 * don't modify cache line when unnecessary.
2529 	 */
2530 	error = fueword32(&cv->c_has_waiters, &hasw);
2531 	if (error == 0 && hasw == 0)
2532 		suword32(&cv->c_has_waiters, 1);
2533 
2534 	umtxq_unbusy_unlocked(&uq->uq_key);
2535 
2536 	error = do_unlock_umutex(td, m, false);
2537 
2538 	if (timeout != NULL)
2539 		abs_timeout_init(&timo, clockid, (wflags & CVWAIT_ABSTIME) != 0,
2540 		    timeout);
2541 
2542 	umtxq_lock(&uq->uq_key);
2543 	if (error == 0) {
2544 		error = umtxq_sleep(uq, "ucond", timeout == NULL ?
2545 		    NULL : &timo);
2546 	}
2547 
2548 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
2549 		error = 0;
2550 	else {
2551 		/*
2552 		 * This must be timeout,interrupted by signal or
2553 		 * surprious wakeup, clear c_has_waiter flag when
2554 		 * necessary.
2555 		 */
2556 		umtxq_busy(&uq->uq_key);
2557 		if ((uq->uq_flags & UQF_UMTXQ) != 0) {
2558 			int oldlen = uq->uq_cur_queue->length;
2559 			umtxq_remove(uq);
2560 			if (oldlen == 1) {
2561 				umtxq_unlock(&uq->uq_key);
2562 				suword32(&cv->c_has_waiters, 0);
2563 				umtxq_lock(&uq->uq_key);
2564 			}
2565 		}
2566 		umtxq_unbusy(&uq->uq_key);
2567 		if (error == ERESTART)
2568 			error = EINTR;
2569 	}
2570 
2571 	umtxq_unlock(&uq->uq_key);
2572 	umtx_key_release(&uq->uq_key);
2573 	return (error);
2574 }
2575 
2576 /*
2577  * Signal a userland condition variable.
2578  */
2579 static int
2580 do_cv_signal(struct thread *td, struct ucond *cv)
2581 {
2582 	struct umtx_key key;
2583 	int error, cnt, nwake;
2584 	uint32_t flags;
2585 
2586 	error = fueword32(&cv->c_flags, &flags);
2587 	if (error == -1)
2588 		return (EFAULT);
2589 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
2590 		return (error);
2591 	umtxq_lock(&key);
2592 	umtxq_busy(&key);
2593 	cnt = umtxq_count(&key);
2594 	nwake = umtxq_signal(&key, 1);
2595 	if (cnt <= nwake) {
2596 		umtxq_unlock(&key);
2597 		error = suword32(&cv->c_has_waiters, 0);
2598 		if (error == -1)
2599 			error = EFAULT;
2600 		umtxq_lock(&key);
2601 	}
2602 	umtxq_unbusy(&key);
2603 	umtxq_unlock(&key);
2604 	umtx_key_release(&key);
2605 	return (error);
2606 }
2607 
2608 static int
2609 do_cv_broadcast(struct thread *td, struct ucond *cv)
2610 {
2611 	struct umtx_key key;
2612 	int error;
2613 	uint32_t flags;
2614 
2615 	error = fueword32(&cv->c_flags, &flags);
2616 	if (error == -1)
2617 		return (EFAULT);
2618 	if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
2619 		return (error);
2620 
2621 	umtxq_lock(&key);
2622 	umtxq_busy(&key);
2623 	umtxq_signal(&key, INT_MAX);
2624 	umtxq_unlock(&key);
2625 
2626 	error = suword32(&cv->c_has_waiters, 0);
2627 	if (error == -1)
2628 		error = EFAULT;
2629 
2630 	umtxq_unbusy_unlocked(&key);
2631 
2632 	umtx_key_release(&key);
2633 	return (error);
2634 }
2635 
2636 static int
2637 do_rw_rdlock(struct thread *td, struct urwlock *rwlock, long fflag, struct _umtx_time *timeout)
2638 {
2639 	struct abs_timeout timo;
2640 	struct umtx_q *uq;
2641 	uint32_t flags, wrflags;
2642 	int32_t state, oldstate;
2643 	int32_t blocked_readers;
2644 	int error, error1, rv;
2645 
2646 	uq = td->td_umtxq;
2647 	error = fueword32(&rwlock->rw_flags, &flags);
2648 	if (error == -1)
2649 		return (EFAULT);
2650 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
2651 	if (error != 0)
2652 		return (error);
2653 
2654 	if (timeout != NULL)
2655 		abs_timeout_init2(&timo, timeout);
2656 
2657 	wrflags = URWLOCK_WRITE_OWNER;
2658 	if (!(fflag & URWLOCK_PREFER_READER) && !(flags & URWLOCK_PREFER_READER))
2659 		wrflags |= URWLOCK_WRITE_WAITERS;
2660 
2661 	for (;;) {
2662 		rv = fueword32(&rwlock->rw_state, &state);
2663 		if (rv == -1) {
2664 			umtx_key_release(&uq->uq_key);
2665 			return (EFAULT);
2666 		}
2667 
2668 		/* try to lock it */
2669 		while (!(state & wrflags)) {
2670 			if (__predict_false(URWLOCK_READER_COUNT(state) == URWLOCK_MAX_READERS)) {
2671 				umtx_key_release(&uq->uq_key);
2672 				return (EAGAIN);
2673 			}
2674 			rv = casueword32(&rwlock->rw_state, state,
2675 			    &oldstate, state + 1);
2676 			if (rv == -1) {
2677 				umtx_key_release(&uq->uq_key);
2678 				return (EFAULT);
2679 			}
2680 			if (oldstate == state) {
2681 				umtx_key_release(&uq->uq_key);
2682 				return (0);
2683 			}
2684 			error = umtxq_check_susp(td);
2685 			if (error != 0)
2686 				break;
2687 			state = oldstate;
2688 		}
2689 
2690 		if (error)
2691 			break;
2692 
2693 		/* grab monitor lock */
2694 		umtxq_lock(&uq->uq_key);
2695 		umtxq_busy(&uq->uq_key);
2696 		umtxq_unlock(&uq->uq_key);
2697 
2698 		/*
2699 		 * re-read the state, in case it changed between the try-lock above
2700 		 * and the check below
2701 		 */
2702 		rv = fueword32(&rwlock->rw_state, &state);
2703 		if (rv == -1)
2704 			error = EFAULT;
2705 
2706 		/* set read contention bit */
2707 		while (error == 0 && (state & wrflags) &&
2708 		    !(state & URWLOCK_READ_WAITERS)) {
2709 			rv = casueword32(&rwlock->rw_state, state,
2710 			    &oldstate, state | URWLOCK_READ_WAITERS);
2711 			if (rv == -1) {
2712 				error = EFAULT;
2713 				break;
2714 			}
2715 			if (oldstate == state)
2716 				goto sleep;
2717 			state = oldstate;
2718 			error = umtxq_check_susp(td);
2719 			if (error != 0)
2720 				break;
2721 		}
2722 		if (error != 0) {
2723 			umtxq_unbusy_unlocked(&uq->uq_key);
2724 			break;
2725 		}
2726 
2727 		/* state is changed while setting flags, restart */
2728 		if (!(state & wrflags)) {
2729 			umtxq_unbusy_unlocked(&uq->uq_key);
2730 			error = umtxq_check_susp(td);
2731 			if (error != 0)
2732 				break;
2733 			continue;
2734 		}
2735 
2736 sleep:
2737 		/* contention bit is set, before sleeping, increase read waiter count */
2738 		rv = fueword32(&rwlock->rw_blocked_readers,
2739 		    &blocked_readers);
2740 		if (rv == -1) {
2741 			umtxq_unbusy_unlocked(&uq->uq_key);
2742 			error = EFAULT;
2743 			break;
2744 		}
2745 		suword32(&rwlock->rw_blocked_readers, blocked_readers+1);
2746 
2747 		while (state & wrflags) {
2748 			umtxq_lock(&uq->uq_key);
2749 			umtxq_insert(uq);
2750 			umtxq_unbusy(&uq->uq_key);
2751 
2752 			error = umtxq_sleep(uq, "urdlck", timeout == NULL ?
2753 			    NULL : &timo);
2754 
2755 			umtxq_busy(&uq->uq_key);
2756 			umtxq_remove(uq);
2757 			umtxq_unlock(&uq->uq_key);
2758 			if (error)
2759 				break;
2760 			rv = fueword32(&rwlock->rw_state, &state);
2761 			if (rv == -1) {
2762 				error = EFAULT;
2763 				break;
2764 			}
2765 		}
2766 
2767 		/* decrease read waiter count, and may clear read contention bit */
2768 		rv = fueword32(&rwlock->rw_blocked_readers,
2769 		    &blocked_readers);
2770 		if (rv == -1) {
2771 			umtxq_unbusy_unlocked(&uq->uq_key);
2772 			error = EFAULT;
2773 			break;
2774 		}
2775 		suword32(&rwlock->rw_blocked_readers, blocked_readers-1);
2776 		if (blocked_readers == 1) {
2777 			rv = fueword32(&rwlock->rw_state, &state);
2778 			if (rv == -1) {
2779 				umtxq_unbusy_unlocked(&uq->uq_key);
2780 				error = EFAULT;
2781 				break;
2782 			}
2783 			for (;;) {
2784 				rv = casueword32(&rwlock->rw_state, state,
2785 				    &oldstate, state & ~URWLOCK_READ_WAITERS);
2786 				if (rv == -1) {
2787 					error = EFAULT;
2788 					break;
2789 				}
2790 				if (oldstate == state)
2791 					break;
2792 				state = oldstate;
2793 				error1 = umtxq_check_susp(td);
2794 				if (error1 != 0) {
2795 					if (error == 0)
2796 						error = error1;
2797 					break;
2798 				}
2799 			}
2800 		}
2801 
2802 		umtxq_unbusy_unlocked(&uq->uq_key);
2803 		if (error != 0)
2804 			break;
2805 	}
2806 	umtx_key_release(&uq->uq_key);
2807 	if (error == ERESTART)
2808 		error = EINTR;
2809 	return (error);
2810 }
2811 
2812 static int
2813 do_rw_wrlock(struct thread *td, struct urwlock *rwlock, struct _umtx_time *timeout)
2814 {
2815 	struct abs_timeout timo;
2816 	struct umtx_q *uq;
2817 	uint32_t flags;
2818 	int32_t state, oldstate;
2819 	int32_t blocked_writers;
2820 	int32_t blocked_readers;
2821 	int error, error1, rv;
2822 
2823 	uq = td->td_umtxq;
2824 	error = fueword32(&rwlock->rw_flags, &flags);
2825 	if (error == -1)
2826 		return (EFAULT);
2827 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
2828 	if (error != 0)
2829 		return (error);
2830 
2831 	if (timeout != NULL)
2832 		abs_timeout_init2(&timo, timeout);
2833 
2834 	blocked_readers = 0;
2835 	for (;;) {
2836 		rv = fueword32(&rwlock->rw_state, &state);
2837 		if (rv == -1) {
2838 			umtx_key_release(&uq->uq_key);
2839 			return (EFAULT);
2840 		}
2841 		while (!(state & URWLOCK_WRITE_OWNER) && URWLOCK_READER_COUNT(state) == 0) {
2842 			rv = casueword32(&rwlock->rw_state, state,
2843 			    &oldstate, state | URWLOCK_WRITE_OWNER);
2844 			if (rv == -1) {
2845 				umtx_key_release(&uq->uq_key);
2846 				return (EFAULT);
2847 			}
2848 			if (oldstate == state) {
2849 				umtx_key_release(&uq->uq_key);
2850 				return (0);
2851 			}
2852 			state = oldstate;
2853 			error = umtxq_check_susp(td);
2854 			if (error != 0)
2855 				break;
2856 		}
2857 
2858 		if (error) {
2859 			if (!(state & (URWLOCK_WRITE_OWNER|URWLOCK_WRITE_WAITERS)) &&
2860 			    blocked_readers != 0) {
2861 				umtxq_lock(&uq->uq_key);
2862 				umtxq_busy(&uq->uq_key);
2863 				umtxq_signal_queue(&uq->uq_key, INT_MAX, UMTX_SHARED_QUEUE);
2864 				umtxq_unbusy(&uq->uq_key);
2865 				umtxq_unlock(&uq->uq_key);
2866 			}
2867 
2868 			break;
2869 		}
2870 
2871 		/* grab monitor lock */
2872 		umtxq_lock(&uq->uq_key);
2873 		umtxq_busy(&uq->uq_key);
2874 		umtxq_unlock(&uq->uq_key);
2875 
2876 		/*
2877 		 * re-read the state, in case it changed between the try-lock above
2878 		 * and the check below
2879 		 */
2880 		rv = fueword32(&rwlock->rw_state, &state);
2881 		if (rv == -1)
2882 			error = EFAULT;
2883 
2884 		while (error == 0 && ((state & URWLOCK_WRITE_OWNER) ||
2885 		    URWLOCK_READER_COUNT(state) != 0) &&
2886 		    (state & URWLOCK_WRITE_WAITERS) == 0) {
2887 			rv = casueword32(&rwlock->rw_state, state,
2888 			    &oldstate, state | URWLOCK_WRITE_WAITERS);
2889 			if (rv == -1) {
2890 				error = EFAULT;
2891 				break;
2892 			}
2893 			if (oldstate == state)
2894 				goto sleep;
2895 			state = oldstate;
2896 			error = umtxq_check_susp(td);
2897 			if (error != 0)
2898 				break;
2899 		}
2900 		if (error != 0) {
2901 			umtxq_unbusy_unlocked(&uq->uq_key);
2902 			break;
2903 		}
2904 
2905 		if (!(state & URWLOCK_WRITE_OWNER) && URWLOCK_READER_COUNT(state) == 0) {
2906 			umtxq_unbusy_unlocked(&uq->uq_key);
2907 			error = umtxq_check_susp(td);
2908 			if (error != 0)
2909 				break;
2910 			continue;
2911 		}
2912 sleep:
2913 		rv = fueword32(&rwlock->rw_blocked_writers,
2914 		    &blocked_writers);
2915 		if (rv == -1) {
2916 			umtxq_unbusy_unlocked(&uq->uq_key);
2917 			error = EFAULT;
2918 			break;
2919 		}
2920 		suword32(&rwlock->rw_blocked_writers, blocked_writers+1);
2921 
2922 		while ((state & URWLOCK_WRITE_OWNER) || URWLOCK_READER_COUNT(state) != 0) {
2923 			umtxq_lock(&uq->uq_key);
2924 			umtxq_insert_queue(uq, UMTX_EXCLUSIVE_QUEUE);
2925 			umtxq_unbusy(&uq->uq_key);
2926 
2927 			error = umtxq_sleep(uq, "uwrlck", timeout == NULL ?
2928 			    NULL : &timo);
2929 
2930 			umtxq_busy(&uq->uq_key);
2931 			umtxq_remove_queue(uq, UMTX_EXCLUSIVE_QUEUE);
2932 			umtxq_unlock(&uq->uq_key);
2933 			if (error)
2934 				break;
2935 			rv = fueword32(&rwlock->rw_state, &state);
2936 			if (rv == -1) {
2937 				error = EFAULT;
2938 				break;
2939 			}
2940 		}
2941 
2942 		rv = fueword32(&rwlock->rw_blocked_writers,
2943 		    &blocked_writers);
2944 		if (rv == -1) {
2945 			umtxq_unbusy_unlocked(&uq->uq_key);
2946 			error = EFAULT;
2947 			break;
2948 		}
2949 		suword32(&rwlock->rw_blocked_writers, blocked_writers-1);
2950 		if (blocked_writers == 1) {
2951 			rv = fueword32(&rwlock->rw_state, &state);
2952 			if (rv == -1) {
2953 				umtxq_unbusy_unlocked(&uq->uq_key);
2954 				error = EFAULT;
2955 				break;
2956 			}
2957 			for (;;) {
2958 				rv = casueword32(&rwlock->rw_state, state,
2959 				    &oldstate, state & ~URWLOCK_WRITE_WAITERS);
2960 				if (rv == -1) {
2961 					error = EFAULT;
2962 					break;
2963 				}
2964 				if (oldstate == state)
2965 					break;
2966 				state = oldstate;
2967 				error1 = umtxq_check_susp(td);
2968 				/*
2969 				 * We are leaving the URWLOCK_WRITE_WAITERS
2970 				 * behind, but this should not harm the
2971 				 * correctness.
2972 				 */
2973 				if (error1 != 0) {
2974 					if (error == 0)
2975 						error = error1;
2976 					break;
2977 				}
2978 			}
2979 			rv = fueword32(&rwlock->rw_blocked_readers,
2980 			    &blocked_readers);
2981 			if (rv == -1) {
2982 				umtxq_unbusy_unlocked(&uq->uq_key);
2983 				error = EFAULT;
2984 				break;
2985 			}
2986 		} else
2987 			blocked_readers = 0;
2988 
2989 		umtxq_unbusy_unlocked(&uq->uq_key);
2990 	}
2991 
2992 	umtx_key_release(&uq->uq_key);
2993 	if (error == ERESTART)
2994 		error = EINTR;
2995 	return (error);
2996 }
2997 
2998 static int
2999 do_rw_unlock(struct thread *td, struct urwlock *rwlock)
3000 {
3001 	struct umtx_q *uq;
3002 	uint32_t flags;
3003 	int32_t state, oldstate;
3004 	int error, rv, q, count;
3005 
3006 	uq = td->td_umtxq;
3007 	error = fueword32(&rwlock->rw_flags, &flags);
3008 	if (error == -1)
3009 		return (EFAULT);
3010 	error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3011 	if (error != 0)
3012 		return (error);
3013 
3014 	error = fueword32(&rwlock->rw_state, &state);
3015 	if (error == -1) {
3016 		error = EFAULT;
3017 		goto out;
3018 	}
3019 	if (state & URWLOCK_WRITE_OWNER) {
3020 		for (;;) {
3021 			rv = casueword32(&rwlock->rw_state, state,
3022 			    &oldstate, state & ~URWLOCK_WRITE_OWNER);
3023 			if (rv == -1) {
3024 				error = EFAULT;
3025 				goto out;
3026 			}
3027 			if (oldstate != state) {
3028 				state = oldstate;
3029 				if (!(oldstate & URWLOCK_WRITE_OWNER)) {
3030 					error = EPERM;
3031 					goto out;
3032 				}
3033 				error = umtxq_check_susp(td);
3034 				if (error != 0)
3035 					goto out;
3036 			} else
3037 				break;
3038 		}
3039 	} else if (URWLOCK_READER_COUNT(state) != 0) {
3040 		for (;;) {
3041 			rv = casueword32(&rwlock->rw_state, state,
3042 			    &oldstate, state - 1);
3043 			if (rv == -1) {
3044 				error = EFAULT;
3045 				goto out;
3046 			}
3047 			if (oldstate != state) {
3048 				state = oldstate;
3049 				if (URWLOCK_READER_COUNT(oldstate) == 0) {
3050 					error = EPERM;
3051 					goto out;
3052 				}
3053 				error = umtxq_check_susp(td);
3054 				if (error != 0)
3055 					goto out;
3056 			} else
3057 				break;
3058 		}
3059 	} else {
3060 		error = EPERM;
3061 		goto out;
3062 	}
3063 
3064 	count = 0;
3065 
3066 	if (!(flags & URWLOCK_PREFER_READER)) {
3067 		if (state & URWLOCK_WRITE_WAITERS) {
3068 			count = 1;
3069 			q = UMTX_EXCLUSIVE_QUEUE;
3070 		} else if (state & URWLOCK_READ_WAITERS) {
3071 			count = INT_MAX;
3072 			q = UMTX_SHARED_QUEUE;
3073 		}
3074 	} else {
3075 		if (state & URWLOCK_READ_WAITERS) {
3076 			count = INT_MAX;
3077 			q = UMTX_SHARED_QUEUE;
3078 		} else if (state & URWLOCK_WRITE_WAITERS) {
3079 			count = 1;
3080 			q = UMTX_EXCLUSIVE_QUEUE;
3081 		}
3082 	}
3083 
3084 	if (count) {
3085 		umtxq_lock(&uq->uq_key);
3086 		umtxq_busy(&uq->uq_key);
3087 		umtxq_signal_queue(&uq->uq_key, count, q);
3088 		umtxq_unbusy(&uq->uq_key);
3089 		umtxq_unlock(&uq->uq_key);
3090 	}
3091 out:
3092 	umtx_key_release(&uq->uq_key);
3093 	return (error);
3094 }
3095 
3096 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3097 static int
3098 do_sem_wait(struct thread *td, struct _usem *sem, struct _umtx_time *timeout)
3099 {
3100 	struct abs_timeout timo;
3101 	struct umtx_q *uq;
3102 	uint32_t flags, count, count1;
3103 	int error, rv;
3104 
3105 	uq = td->td_umtxq;
3106 	error = fueword32(&sem->_flags, &flags);
3107 	if (error == -1)
3108 		return (EFAULT);
3109 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3110 	if (error != 0)
3111 		return (error);
3112 
3113 	if (timeout != NULL)
3114 		abs_timeout_init2(&timo, timeout);
3115 
3116 	umtxq_lock(&uq->uq_key);
3117 	umtxq_busy(&uq->uq_key);
3118 	umtxq_insert(uq);
3119 	umtxq_unlock(&uq->uq_key);
3120 	rv = casueword32(&sem->_has_waiters, 0, &count1, 1);
3121 	if (rv == 0)
3122 		rv = fueword32(&sem->_count, &count);
3123 	if (rv == -1 || count != 0) {
3124 		umtxq_lock(&uq->uq_key);
3125 		umtxq_unbusy(&uq->uq_key);
3126 		umtxq_remove(uq);
3127 		umtxq_unlock(&uq->uq_key);
3128 		umtx_key_release(&uq->uq_key);
3129 		return (rv == -1 ? EFAULT : 0);
3130 	}
3131 	umtxq_lock(&uq->uq_key);
3132 	umtxq_unbusy(&uq->uq_key);
3133 
3134 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3135 
3136 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3137 		error = 0;
3138 	else {
3139 		umtxq_remove(uq);
3140 		/* A relative timeout cannot be restarted. */
3141 		if (error == ERESTART && timeout != NULL &&
3142 		    (timeout->_flags & UMTX_ABSTIME) == 0)
3143 			error = EINTR;
3144 	}
3145 	umtxq_unlock(&uq->uq_key);
3146 	umtx_key_release(&uq->uq_key);
3147 	return (error);
3148 }
3149 
3150 /*
3151  * Signal a userland semaphore.
3152  */
3153 static int
3154 do_sem_wake(struct thread *td, struct _usem *sem)
3155 {
3156 	struct umtx_key key;
3157 	int error, cnt;
3158 	uint32_t flags;
3159 
3160 	error = fueword32(&sem->_flags, &flags);
3161 	if (error == -1)
3162 		return (EFAULT);
3163 	if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3164 		return (error);
3165 	umtxq_lock(&key);
3166 	umtxq_busy(&key);
3167 	cnt = umtxq_count(&key);
3168 	if (cnt > 0) {
3169 		/*
3170 		 * Check if count is greater than 0, this means the memory is
3171 		 * still being referenced by user code, so we can safely
3172 		 * update _has_waiters flag.
3173 		 */
3174 		if (cnt == 1) {
3175 			umtxq_unlock(&key);
3176 			error = suword32(&sem->_has_waiters, 0);
3177 			umtxq_lock(&key);
3178 			if (error == -1)
3179 				error = EFAULT;
3180 		}
3181 		umtxq_signal(&key, 1);
3182 	}
3183 	umtxq_unbusy(&key);
3184 	umtxq_unlock(&key);
3185 	umtx_key_release(&key);
3186 	return (error);
3187 }
3188 #endif
3189 
3190 static int
3191 do_sem2_wait(struct thread *td, struct _usem2 *sem, struct _umtx_time *timeout)
3192 {
3193 	struct abs_timeout timo;
3194 	struct umtx_q *uq;
3195 	uint32_t count, flags;
3196 	int error, rv;
3197 
3198 	uq = td->td_umtxq;
3199 	flags = fuword32(&sem->_flags);
3200 	error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3201 	if (error != 0)
3202 		return (error);
3203 
3204 	if (timeout != NULL)
3205 		abs_timeout_init2(&timo, timeout);
3206 
3207 	umtxq_lock(&uq->uq_key);
3208 	umtxq_busy(&uq->uq_key);
3209 	umtxq_insert(uq);
3210 	umtxq_unlock(&uq->uq_key);
3211 	rv = fueword32(&sem->_count, &count);
3212 	if (rv == -1) {
3213 		umtxq_lock(&uq->uq_key);
3214 		umtxq_unbusy(&uq->uq_key);
3215 		umtxq_remove(uq);
3216 		umtxq_unlock(&uq->uq_key);
3217 		umtx_key_release(&uq->uq_key);
3218 		return (EFAULT);
3219 	}
3220 	for (;;) {
3221 		if (USEM_COUNT(count) != 0) {
3222 			umtxq_lock(&uq->uq_key);
3223 			umtxq_unbusy(&uq->uq_key);
3224 			umtxq_remove(uq);
3225 			umtxq_unlock(&uq->uq_key);
3226 			umtx_key_release(&uq->uq_key);
3227 			return (0);
3228 		}
3229 		if (count == USEM_HAS_WAITERS)
3230 			break;
3231 		rv = casueword32(&sem->_count, 0, &count, USEM_HAS_WAITERS);
3232 		if (rv == -1) {
3233 			umtxq_lock(&uq->uq_key);
3234 			umtxq_unbusy(&uq->uq_key);
3235 			umtxq_remove(uq);
3236 			umtxq_unlock(&uq->uq_key);
3237 			umtx_key_release(&uq->uq_key);
3238 			return (EFAULT);
3239 		}
3240 		if (count == 0)
3241 			break;
3242 	}
3243 	umtxq_lock(&uq->uq_key);
3244 	umtxq_unbusy(&uq->uq_key);
3245 
3246 	error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3247 
3248 	if ((uq->uq_flags & UQF_UMTXQ) == 0)
3249 		error = 0;
3250 	else {
3251 		umtxq_remove(uq);
3252 		if (timeout != NULL && (timeout->_flags & UMTX_ABSTIME) == 0) {
3253 			/* A relative timeout cannot be restarted. */
3254 			if (error == ERESTART)
3255 				error = EINTR;
3256 			if (error == EINTR) {
3257 				abs_timeout_update(&timo);
3258 				timeout->_timeout = timo.end;
3259 				timespecsub(&timeout->_timeout, &timo.cur);
3260 			}
3261 		}
3262 	}
3263 	umtxq_unlock(&uq->uq_key);
3264 	umtx_key_release(&uq->uq_key);
3265 	return (error);
3266 }
3267 
3268 /*
3269  * Signal a userland semaphore.
3270  */
3271 static int
3272 do_sem2_wake(struct thread *td, struct _usem2 *sem)
3273 {
3274 	struct umtx_key key;
3275 	int error, cnt, rv;
3276 	uint32_t count, flags;
3277 
3278 	rv = fueword32(&sem->_flags, &flags);
3279 	if (rv == -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 		 * If this was the last sleeping thread, clear the waiters
3289 		 * flag in _count.
3290 		 */
3291 		if (cnt == 1) {
3292 			umtxq_unlock(&key);
3293 			rv = fueword32(&sem->_count, &count);
3294 			while (rv != -1 && count & USEM_HAS_WAITERS)
3295 				rv = casueword32(&sem->_count, count, &count,
3296 				    count & ~USEM_HAS_WAITERS);
3297 			if (rv == -1)
3298 				error = EFAULT;
3299 			umtxq_lock(&key);
3300 		}
3301 
3302 		umtxq_signal(&key, 1);
3303 	}
3304 	umtxq_unbusy(&key);
3305 	umtxq_unlock(&key);
3306 	umtx_key_release(&key);
3307 	return (error);
3308 }
3309 
3310 inline int
3311 umtx_copyin_timeout(const void *addr, struct timespec *tsp)
3312 {
3313 	int error;
3314 
3315 	error = copyin(addr, tsp, sizeof(struct timespec));
3316 	if (error == 0) {
3317 		if (tsp->tv_sec < 0 ||
3318 		    tsp->tv_nsec >= 1000000000 ||
3319 		    tsp->tv_nsec < 0)
3320 			error = EINVAL;
3321 	}
3322 	return (error);
3323 }
3324 
3325 static inline int
3326 umtx_copyin_umtx_time(const void *addr, size_t size, struct _umtx_time *tp)
3327 {
3328 	int error;
3329 
3330 	if (size <= sizeof(struct timespec)) {
3331 		tp->_clockid = CLOCK_REALTIME;
3332 		tp->_flags = 0;
3333 		error = copyin(addr, &tp->_timeout, sizeof(struct timespec));
3334 	} else
3335 		error = copyin(addr, tp, sizeof(struct _umtx_time));
3336 	if (error != 0)
3337 		return (error);
3338 	if (tp->_timeout.tv_sec < 0 ||
3339 	    tp->_timeout.tv_nsec >= 1000000000 || tp->_timeout.tv_nsec < 0)
3340 		return (EINVAL);
3341 	return (0);
3342 }
3343 
3344 static int
3345 __umtx_op_unimpl(struct thread *td, struct _umtx_op_args *uap)
3346 {
3347 
3348 	return (EOPNOTSUPP);
3349 }
3350 
3351 static int
3352 __umtx_op_wait(struct thread *td, struct _umtx_op_args *uap)
3353 {
3354 	struct _umtx_time timeout, *tm_p;
3355 	int error;
3356 
3357 	if (uap->uaddr2 == NULL)
3358 		tm_p = NULL;
3359 	else {
3360 		error = umtx_copyin_umtx_time(
3361 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3362 		if (error != 0)
3363 			return (error);
3364 		tm_p = &timeout;
3365 	}
3366 	return (do_wait(td, uap->obj, uap->val, tm_p, 0, 0));
3367 }
3368 
3369 static int
3370 __umtx_op_wait_uint(struct thread *td, struct _umtx_op_args *uap)
3371 {
3372 	struct _umtx_time timeout, *tm_p;
3373 	int error;
3374 
3375 	if (uap->uaddr2 == NULL)
3376 		tm_p = NULL;
3377 	else {
3378 		error = umtx_copyin_umtx_time(
3379 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3380 		if (error != 0)
3381 			return (error);
3382 		tm_p = &timeout;
3383 	}
3384 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
3385 }
3386 
3387 static int
3388 __umtx_op_wait_uint_private(struct thread *td, struct _umtx_op_args *uap)
3389 {
3390 	struct _umtx_time *tm_p, timeout;
3391 	int error;
3392 
3393 	if (uap->uaddr2 == NULL)
3394 		tm_p = NULL;
3395 	else {
3396 		error = umtx_copyin_umtx_time(
3397 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3398 		if (error != 0)
3399 			return (error);
3400 		tm_p = &timeout;
3401 	}
3402 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
3403 }
3404 
3405 static int
3406 __umtx_op_wake(struct thread *td, struct _umtx_op_args *uap)
3407 {
3408 
3409 	return (kern_umtx_wake(td, uap->obj, uap->val, 0));
3410 }
3411 
3412 #define BATCH_SIZE	128
3413 static int
3414 __umtx_op_nwake_private(struct thread *td, struct _umtx_op_args *uap)
3415 {
3416 	char *uaddrs[BATCH_SIZE], **upp;
3417 	int count, error, i, pos, tocopy;
3418 
3419 	upp = (char **)uap->obj;
3420 	error = 0;
3421 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3422 	    pos += tocopy) {
3423 		tocopy = MIN(count, BATCH_SIZE);
3424 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(char *));
3425 		if (error != 0)
3426 			break;
3427 		for (i = 0; i < tocopy; ++i)
3428 			kern_umtx_wake(td, uaddrs[i], INT_MAX, 1);
3429 		maybe_yield();
3430 	}
3431 	return (error);
3432 }
3433 
3434 static int
3435 __umtx_op_wake_private(struct thread *td, struct _umtx_op_args *uap)
3436 {
3437 
3438 	return (kern_umtx_wake(td, uap->obj, uap->val, 1));
3439 }
3440 
3441 static int
3442 __umtx_op_lock_umutex(struct thread *td, struct _umtx_op_args *uap)
3443 {
3444 	struct _umtx_time *tm_p, timeout;
3445 	int error;
3446 
3447 	/* Allow a null timespec (wait forever). */
3448 	if (uap->uaddr2 == NULL)
3449 		tm_p = NULL;
3450 	else {
3451 		error = umtx_copyin_umtx_time(
3452 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3453 		if (error != 0)
3454 			return (error);
3455 		tm_p = &timeout;
3456 	}
3457 	return (do_lock_umutex(td, uap->obj, tm_p, 0));
3458 }
3459 
3460 static int
3461 __umtx_op_trylock_umutex(struct thread *td, struct _umtx_op_args *uap)
3462 {
3463 
3464 	return (do_lock_umutex(td, uap->obj, NULL, _UMUTEX_TRY));
3465 }
3466 
3467 static int
3468 __umtx_op_wait_umutex(struct thread *td, struct _umtx_op_args *uap)
3469 {
3470 	struct _umtx_time *tm_p, timeout;
3471 	int error;
3472 
3473 	/* Allow a null timespec (wait forever). */
3474 	if (uap->uaddr2 == NULL)
3475 		tm_p = NULL;
3476 	else {
3477 		error = umtx_copyin_umtx_time(
3478 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3479 		if (error != 0)
3480 			return (error);
3481 		tm_p = &timeout;
3482 	}
3483 	return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
3484 }
3485 
3486 static int
3487 __umtx_op_wake_umutex(struct thread *td, struct _umtx_op_args *uap)
3488 {
3489 
3490 	return (do_wake_umutex(td, uap->obj));
3491 }
3492 
3493 static int
3494 __umtx_op_unlock_umutex(struct thread *td, struct _umtx_op_args *uap)
3495 {
3496 
3497 	return (do_unlock_umutex(td, uap->obj, false));
3498 }
3499 
3500 static int
3501 __umtx_op_set_ceiling(struct thread *td, struct _umtx_op_args *uap)
3502 {
3503 
3504 	return (do_set_ceiling(td, uap->obj, uap->val, uap->uaddr1));
3505 }
3506 
3507 static int
3508 __umtx_op_cv_wait(struct thread *td, struct _umtx_op_args *uap)
3509 {
3510 	struct timespec *ts, timeout;
3511 	int error;
3512 
3513 	/* Allow a null timespec (wait forever). */
3514 	if (uap->uaddr2 == NULL)
3515 		ts = NULL;
3516 	else {
3517 		error = umtx_copyin_timeout(uap->uaddr2, &timeout);
3518 		if (error != 0)
3519 			return (error);
3520 		ts = &timeout;
3521 	}
3522 	return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
3523 }
3524 
3525 static int
3526 __umtx_op_cv_signal(struct thread *td, struct _umtx_op_args *uap)
3527 {
3528 
3529 	return (do_cv_signal(td, uap->obj));
3530 }
3531 
3532 static int
3533 __umtx_op_cv_broadcast(struct thread *td, struct _umtx_op_args *uap)
3534 {
3535 
3536 	return (do_cv_broadcast(td, uap->obj));
3537 }
3538 
3539 static int
3540 __umtx_op_rw_rdlock(struct thread *td, struct _umtx_op_args *uap)
3541 {
3542 	struct _umtx_time timeout;
3543 	int error;
3544 
3545 	/* Allow a null timespec (wait forever). */
3546 	if (uap->uaddr2 == NULL) {
3547 		error = do_rw_rdlock(td, uap->obj, uap->val, 0);
3548 	} else {
3549 		error = umtx_copyin_umtx_time(uap->uaddr2,
3550 		   (size_t)uap->uaddr1, &timeout);
3551 		if (error != 0)
3552 			return (error);
3553 		error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
3554 	}
3555 	return (error);
3556 }
3557 
3558 static int
3559 __umtx_op_rw_wrlock(struct thread *td, struct _umtx_op_args *uap)
3560 {
3561 	struct _umtx_time timeout;
3562 	int error;
3563 
3564 	/* Allow a null timespec (wait forever). */
3565 	if (uap->uaddr2 == NULL) {
3566 		error = do_rw_wrlock(td, uap->obj, 0);
3567 	} else {
3568 		error = umtx_copyin_umtx_time(uap->uaddr2,
3569 		   (size_t)uap->uaddr1, &timeout);
3570 		if (error != 0)
3571 			return (error);
3572 
3573 		error = do_rw_wrlock(td, uap->obj, &timeout);
3574 	}
3575 	return (error);
3576 }
3577 
3578 static int
3579 __umtx_op_rw_unlock(struct thread *td, struct _umtx_op_args *uap)
3580 {
3581 
3582 	return (do_rw_unlock(td, uap->obj));
3583 }
3584 
3585 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3586 static int
3587 __umtx_op_sem_wait(struct thread *td, struct _umtx_op_args *uap)
3588 {
3589 	struct _umtx_time *tm_p, timeout;
3590 	int error;
3591 
3592 	/* Allow a null timespec (wait forever). */
3593 	if (uap->uaddr2 == NULL)
3594 		tm_p = NULL;
3595 	else {
3596 		error = umtx_copyin_umtx_time(
3597 		    uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3598 		if (error != 0)
3599 			return (error);
3600 		tm_p = &timeout;
3601 	}
3602 	return (do_sem_wait(td, uap->obj, tm_p));
3603 }
3604 
3605 static int
3606 __umtx_op_sem_wake(struct thread *td, struct _umtx_op_args *uap)
3607 {
3608 
3609 	return (do_sem_wake(td, uap->obj));
3610 }
3611 #endif
3612 
3613 static int
3614 __umtx_op_wake2_umutex(struct thread *td, struct _umtx_op_args *uap)
3615 {
3616 
3617 	return (do_wake2_umutex(td, uap->obj, uap->val));
3618 }
3619 
3620 static int
3621 __umtx_op_sem2_wait(struct thread *td, struct _umtx_op_args *uap)
3622 {
3623 	struct _umtx_time *tm_p, timeout;
3624 	size_t uasize;
3625 	int error;
3626 
3627 	/* Allow a null timespec (wait forever). */
3628 	if (uap->uaddr2 == NULL) {
3629 		uasize = 0;
3630 		tm_p = NULL;
3631 	} else {
3632 		uasize = (size_t)uap->uaddr1;
3633 		error = umtx_copyin_umtx_time(uap->uaddr2, uasize, &timeout);
3634 		if (error != 0)
3635 			return (error);
3636 		tm_p = &timeout;
3637 	}
3638 	error = do_sem2_wait(td, uap->obj, tm_p);
3639 	if (error == EINTR && uap->uaddr2 != NULL &&
3640 	    (timeout._flags & UMTX_ABSTIME) == 0 &&
3641 	    uasize >= sizeof(struct _umtx_time) + sizeof(struct timespec)) {
3642 		error = copyout(&timeout._timeout,
3643 		    (struct _umtx_time *)uap->uaddr2 + 1,
3644 		    sizeof(struct timespec));
3645 		if (error == 0) {
3646 			error = EINTR;
3647 		}
3648 	}
3649 
3650 	return (error);
3651 }
3652 
3653 static int
3654 __umtx_op_sem2_wake(struct thread *td, struct _umtx_op_args *uap)
3655 {
3656 
3657 	return (do_sem2_wake(td, uap->obj));
3658 }
3659 
3660 #define	USHM_OBJ_UMTX(o)						\
3661     ((struct umtx_shm_obj_list *)(&(o)->umtx_data))
3662 
3663 #define	USHMF_REG_LINKED	0x0001
3664 #define	USHMF_OBJ_LINKED	0x0002
3665 struct umtx_shm_reg {
3666 	TAILQ_ENTRY(umtx_shm_reg) ushm_reg_link;
3667 	LIST_ENTRY(umtx_shm_reg) ushm_obj_link;
3668 	struct umtx_key		ushm_key;
3669 	struct ucred		*ushm_cred;
3670 	struct shmfd		*ushm_obj;
3671 	u_int			ushm_refcnt;
3672 	u_int			ushm_flags;
3673 };
3674 
3675 LIST_HEAD(umtx_shm_obj_list, umtx_shm_reg);
3676 TAILQ_HEAD(umtx_shm_reg_head, umtx_shm_reg);
3677 
3678 static uma_zone_t umtx_shm_reg_zone;
3679 static struct umtx_shm_reg_head umtx_shm_registry[UMTX_CHAINS];
3680 static struct mtx umtx_shm_lock;
3681 static struct umtx_shm_reg_head umtx_shm_reg_delfree =
3682     TAILQ_HEAD_INITIALIZER(umtx_shm_reg_delfree);
3683 
3684 static void umtx_shm_free_reg(struct umtx_shm_reg *reg);
3685 
3686 static void
3687 umtx_shm_reg_delfree_tq(void *context __unused, int pending __unused)
3688 {
3689 	struct umtx_shm_reg_head d;
3690 	struct umtx_shm_reg *reg, *reg1;
3691 
3692 	TAILQ_INIT(&d);
3693 	mtx_lock(&umtx_shm_lock);
3694 	TAILQ_CONCAT(&d, &umtx_shm_reg_delfree, ushm_reg_link);
3695 	mtx_unlock(&umtx_shm_lock);
3696 	TAILQ_FOREACH_SAFE(reg, &d, ushm_reg_link, reg1) {
3697 		TAILQ_REMOVE(&d, reg, ushm_reg_link);
3698 		umtx_shm_free_reg(reg);
3699 	}
3700 }
3701 
3702 static struct task umtx_shm_reg_delfree_task =
3703     TASK_INITIALIZER(0, umtx_shm_reg_delfree_tq, NULL);
3704 
3705 static struct umtx_shm_reg *
3706 umtx_shm_find_reg_locked(const struct umtx_key *key)
3707 {
3708 	struct umtx_shm_reg *reg;
3709 	struct umtx_shm_reg_head *reg_head;
3710 
3711 	KASSERT(key->shared, ("umtx_p_find_rg: private key"));
3712 	mtx_assert(&umtx_shm_lock, MA_OWNED);
3713 	reg_head = &umtx_shm_registry[key->hash];
3714 	TAILQ_FOREACH(reg, reg_head, ushm_reg_link) {
3715 		KASSERT(reg->ushm_key.shared,
3716 		    ("non-shared key on reg %p %d", reg, reg->ushm_key.shared));
3717 		if (reg->ushm_key.info.shared.object ==
3718 		    key->info.shared.object &&
3719 		    reg->ushm_key.info.shared.offset ==
3720 		    key->info.shared.offset) {
3721 			KASSERT(reg->ushm_key.type == TYPE_SHM, ("TYPE_USHM"));
3722 			KASSERT(reg->ushm_refcnt > 0,
3723 			    ("reg %p refcnt 0 onlist", reg));
3724 			KASSERT((reg->ushm_flags & USHMF_REG_LINKED) != 0,
3725 			    ("reg %p not linked", reg));
3726 			reg->ushm_refcnt++;
3727 			return (reg);
3728 		}
3729 	}
3730 	return (NULL);
3731 }
3732 
3733 static struct umtx_shm_reg *
3734 umtx_shm_find_reg(const struct umtx_key *key)
3735 {
3736 	struct umtx_shm_reg *reg;
3737 
3738 	mtx_lock(&umtx_shm_lock);
3739 	reg = umtx_shm_find_reg_locked(key);
3740 	mtx_unlock(&umtx_shm_lock);
3741 	return (reg);
3742 }
3743 
3744 static void
3745 umtx_shm_free_reg(struct umtx_shm_reg *reg)
3746 {
3747 
3748 	chgumtxcnt(reg->ushm_cred->cr_ruidinfo, -1, 0);
3749 	crfree(reg->ushm_cred);
3750 	shm_drop(reg->ushm_obj);
3751 	uma_zfree(umtx_shm_reg_zone, reg);
3752 }
3753 
3754 static bool
3755 umtx_shm_unref_reg_locked(struct umtx_shm_reg *reg, bool force)
3756 {
3757 	bool res;
3758 
3759 	mtx_assert(&umtx_shm_lock, MA_OWNED);
3760 	KASSERT(reg->ushm_refcnt > 0, ("ushm_reg %p refcnt 0", reg));
3761 	reg->ushm_refcnt--;
3762 	res = reg->ushm_refcnt == 0;
3763 	if (res || force) {
3764 		if ((reg->ushm_flags & USHMF_REG_LINKED) != 0) {
3765 			TAILQ_REMOVE(&umtx_shm_registry[reg->ushm_key.hash],
3766 			    reg, ushm_reg_link);
3767 			reg->ushm_flags &= ~USHMF_REG_LINKED;
3768 		}
3769 		if ((reg->ushm_flags & USHMF_OBJ_LINKED) != 0) {
3770 			LIST_REMOVE(reg, ushm_obj_link);
3771 			reg->ushm_flags &= ~USHMF_OBJ_LINKED;
3772 		}
3773 	}
3774 	return (res);
3775 }
3776 
3777 static void
3778 umtx_shm_unref_reg(struct umtx_shm_reg *reg, bool force)
3779 {
3780 	vm_object_t object;
3781 	bool dofree;
3782 
3783 	if (force) {
3784 		object = reg->ushm_obj->shm_object;
3785 		VM_OBJECT_WLOCK(object);
3786 		object->flags |= OBJ_UMTXDEAD;
3787 		VM_OBJECT_WUNLOCK(object);
3788 	}
3789 	mtx_lock(&umtx_shm_lock);
3790 	dofree = umtx_shm_unref_reg_locked(reg, force);
3791 	mtx_unlock(&umtx_shm_lock);
3792 	if (dofree)
3793 		umtx_shm_free_reg(reg);
3794 }
3795 
3796 void
3797 umtx_shm_object_init(vm_object_t object)
3798 {
3799 
3800 	LIST_INIT(USHM_OBJ_UMTX(object));
3801 }
3802 
3803 void
3804 umtx_shm_object_terminated(vm_object_t object)
3805 {
3806 	struct umtx_shm_reg *reg, *reg1;
3807 	bool dofree;
3808 
3809 	dofree = false;
3810 	mtx_lock(&umtx_shm_lock);
3811 	LIST_FOREACH_SAFE(reg, USHM_OBJ_UMTX(object), ushm_obj_link, reg1) {
3812 		if (umtx_shm_unref_reg_locked(reg, true)) {
3813 			TAILQ_INSERT_TAIL(&umtx_shm_reg_delfree, reg,
3814 			    ushm_reg_link);
3815 			dofree = true;
3816 		}
3817 	}
3818 	mtx_unlock(&umtx_shm_lock);
3819 	if (dofree)
3820 		taskqueue_enqueue(taskqueue_thread, &umtx_shm_reg_delfree_task);
3821 }
3822 
3823 static int
3824 umtx_shm_create_reg(struct thread *td, const struct umtx_key *key,
3825     struct umtx_shm_reg **res)
3826 {
3827 	struct umtx_shm_reg *reg, *reg1;
3828 	struct ucred *cred;
3829 	int error;
3830 
3831 	reg = umtx_shm_find_reg(key);
3832 	if (reg != NULL) {
3833 		*res = reg;
3834 		return (0);
3835 	}
3836 	cred = td->td_ucred;
3837 	if (!chgumtxcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_UMTXP)))
3838 		return (ENOMEM);
3839 	reg = uma_zalloc(umtx_shm_reg_zone, M_WAITOK | M_ZERO);
3840 	reg->ushm_refcnt = 1;
3841 	bcopy(key, &reg->ushm_key, sizeof(*key));
3842 	reg->ushm_obj = shm_alloc(td->td_ucred, O_RDWR);
3843 	reg->ushm_cred = crhold(cred);
3844 	error = shm_dotruncate(reg->ushm_obj, PAGE_SIZE);
3845 	if (error != 0) {
3846 		umtx_shm_free_reg(reg);
3847 		return (error);
3848 	}
3849 	mtx_lock(&umtx_shm_lock);
3850 	reg1 = umtx_shm_find_reg_locked(key);
3851 	if (reg1 != NULL) {
3852 		mtx_unlock(&umtx_shm_lock);
3853 		umtx_shm_free_reg(reg);
3854 		*res = reg1;
3855 		return (0);
3856 	}
3857 	reg->ushm_refcnt++;
3858 	TAILQ_INSERT_TAIL(&umtx_shm_registry[key->hash], reg, ushm_reg_link);
3859 	LIST_INSERT_HEAD(USHM_OBJ_UMTX(key->info.shared.object), reg,
3860 	    ushm_obj_link);
3861 	reg->ushm_flags = USHMF_REG_LINKED | USHMF_OBJ_LINKED;
3862 	mtx_unlock(&umtx_shm_lock);
3863 	*res = reg;
3864 	return (0);
3865 }
3866 
3867 static int
3868 umtx_shm_alive(struct thread *td, void *addr)
3869 {
3870 	vm_map_t map;
3871 	vm_map_entry_t entry;
3872 	vm_object_t object;
3873 	vm_pindex_t pindex;
3874 	vm_prot_t prot;
3875 	int res, ret;
3876 	boolean_t wired;
3877 
3878 	map = &td->td_proc->p_vmspace->vm_map;
3879 	res = vm_map_lookup(&map, (uintptr_t)addr, VM_PROT_READ, &entry,
3880 	    &object, &pindex, &prot, &wired);
3881 	if (res != KERN_SUCCESS)
3882 		return (EFAULT);
3883 	if (object == NULL)
3884 		ret = EINVAL;
3885 	else
3886 		ret = (object->flags & OBJ_UMTXDEAD) != 0 ? ENOTTY : 0;
3887 	vm_map_lookup_done(map, entry);
3888 	return (ret);
3889 }
3890 
3891 static void
3892 umtx_shm_init(void)
3893 {
3894 	int i;
3895 
3896 	umtx_shm_reg_zone = uma_zcreate("umtx_shm", sizeof(struct umtx_shm_reg),
3897 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
3898 	mtx_init(&umtx_shm_lock, "umtxshm", NULL, MTX_DEF);
3899 	for (i = 0; i < nitems(umtx_shm_registry); i++)
3900 		TAILQ_INIT(&umtx_shm_registry[i]);
3901 }
3902 
3903 static int
3904 umtx_shm(struct thread *td, void *addr, u_int flags)
3905 {
3906 	struct umtx_key key;
3907 	struct umtx_shm_reg *reg;
3908 	struct file *fp;
3909 	int error, fd;
3910 
3911 	if (__bitcount(flags & (UMTX_SHM_CREAT | UMTX_SHM_LOOKUP |
3912 	    UMTX_SHM_DESTROY| UMTX_SHM_ALIVE)) != 1)
3913 		return (EINVAL);
3914 	if ((flags & UMTX_SHM_ALIVE) != 0)
3915 		return (umtx_shm_alive(td, addr));
3916 	error = umtx_key_get(addr, TYPE_SHM, PROCESS_SHARE, &key);
3917 	if (error != 0)
3918 		return (error);
3919 	KASSERT(key.shared == 1, ("non-shared key"));
3920 	if ((flags & UMTX_SHM_CREAT) != 0) {
3921 		error = umtx_shm_create_reg(td, &key, &reg);
3922 	} else {
3923 		reg = umtx_shm_find_reg(&key);
3924 		if (reg == NULL)
3925 			error = ESRCH;
3926 	}
3927 	umtx_key_release(&key);
3928 	if (error != 0)
3929 		return (error);
3930 	KASSERT(reg != NULL, ("no reg"));
3931 	if ((flags & UMTX_SHM_DESTROY) != 0) {
3932 		umtx_shm_unref_reg(reg, true);
3933 	} else {
3934 #if 0
3935 #ifdef MAC
3936 		error = mac_posixshm_check_open(td->td_ucred,
3937 		    reg->ushm_obj, FFLAGS(O_RDWR));
3938 		if (error == 0)
3939 #endif
3940 			error = shm_access(reg->ushm_obj, td->td_ucred,
3941 			    FFLAGS(O_RDWR));
3942 		if (error == 0)
3943 #endif
3944 			error = falloc_caps(td, &fp, &fd, O_CLOEXEC, NULL);
3945 		if (error == 0) {
3946 			shm_hold(reg->ushm_obj);
3947 			finit(fp, FFLAGS(O_RDWR), DTYPE_SHM, reg->ushm_obj,
3948 			    &shm_ops);
3949 			td->td_retval[0] = fd;
3950 			fdrop(fp, td);
3951 		}
3952 	}
3953 	umtx_shm_unref_reg(reg, false);
3954 	return (error);
3955 }
3956 
3957 static int
3958 __umtx_op_shm(struct thread *td, struct _umtx_op_args *uap)
3959 {
3960 
3961 	return (umtx_shm(td, uap->uaddr1, uap->val));
3962 }
3963 
3964 static int
3965 umtx_robust_lists(struct thread *td, struct umtx_robust_lists_params *rbp)
3966 {
3967 
3968 	td->td_rb_list = rbp->robust_list_offset;
3969 	td->td_rbp_list = rbp->robust_priv_list_offset;
3970 	td->td_rb_inact = rbp->robust_inact_offset;
3971 	return (0);
3972 }
3973 
3974 static int
3975 __umtx_op_robust_lists(struct thread *td, struct _umtx_op_args *uap)
3976 {
3977 	struct umtx_robust_lists_params rb;
3978 	int error;
3979 
3980 	if (uap->val > sizeof(rb))
3981 		return (EINVAL);
3982 	bzero(&rb, sizeof(rb));
3983 	error = copyin(uap->uaddr1, &rb, uap->val);
3984 	if (error != 0)
3985 		return (error);
3986 	return (umtx_robust_lists(td, &rb));
3987 }
3988 
3989 typedef int (*_umtx_op_func)(struct thread *td, struct _umtx_op_args *uap);
3990 
3991 static const _umtx_op_func op_table[] = {
3992 	[UMTX_OP_RESERVED0]	= __umtx_op_unimpl,
3993 	[UMTX_OP_RESERVED1]	= __umtx_op_unimpl,
3994 	[UMTX_OP_WAIT]		= __umtx_op_wait,
3995 	[UMTX_OP_WAKE]		= __umtx_op_wake,
3996 	[UMTX_OP_MUTEX_TRYLOCK]	= __umtx_op_trylock_umutex,
3997 	[UMTX_OP_MUTEX_LOCK]	= __umtx_op_lock_umutex,
3998 	[UMTX_OP_MUTEX_UNLOCK]	= __umtx_op_unlock_umutex,
3999 	[UMTX_OP_SET_CEILING]	= __umtx_op_set_ceiling,
4000 	[UMTX_OP_CV_WAIT]	= __umtx_op_cv_wait,
4001 	[UMTX_OP_CV_SIGNAL]	= __umtx_op_cv_signal,
4002 	[UMTX_OP_CV_BROADCAST]	= __umtx_op_cv_broadcast,
4003 	[UMTX_OP_WAIT_UINT]	= __umtx_op_wait_uint,
4004 	[UMTX_OP_RW_RDLOCK]	= __umtx_op_rw_rdlock,
4005 	[UMTX_OP_RW_WRLOCK]	= __umtx_op_rw_wrlock,
4006 	[UMTX_OP_RW_UNLOCK]	= __umtx_op_rw_unlock,
4007 	[UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private,
4008 	[UMTX_OP_WAKE_PRIVATE]	= __umtx_op_wake_private,
4009 	[UMTX_OP_MUTEX_WAIT]	= __umtx_op_wait_umutex,
4010 	[UMTX_OP_MUTEX_WAKE]	= __umtx_op_wake_umutex,
4011 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4012 	[UMTX_OP_SEM_WAIT]	= __umtx_op_sem_wait,
4013 	[UMTX_OP_SEM_WAKE]	= __umtx_op_sem_wake,
4014 #else
4015 	[UMTX_OP_SEM_WAIT]	= __umtx_op_unimpl,
4016 	[UMTX_OP_SEM_WAKE]	= __umtx_op_unimpl,
4017 #endif
4018 	[UMTX_OP_NWAKE_PRIVATE]	= __umtx_op_nwake_private,
4019 	[UMTX_OP_MUTEX_WAKE2]	= __umtx_op_wake2_umutex,
4020 	[UMTX_OP_SEM2_WAIT]	= __umtx_op_sem2_wait,
4021 	[UMTX_OP_SEM2_WAKE]	= __umtx_op_sem2_wake,
4022 	[UMTX_OP_SHM]		= __umtx_op_shm,
4023 	[UMTX_OP_ROBUST_LISTS]	= __umtx_op_robust_lists,
4024 };
4025 
4026 int
4027 sys__umtx_op(struct thread *td, struct _umtx_op_args *uap)
4028 {
4029 
4030 	if ((unsigned)uap->op < nitems(op_table))
4031 		return (*op_table[uap->op])(td, uap);
4032 	return (EINVAL);
4033 }
4034 
4035 #ifdef COMPAT_FREEBSD32
4036 
4037 struct timespec32 {
4038 	int32_t tv_sec;
4039 	int32_t tv_nsec;
4040 };
4041 
4042 struct umtx_time32 {
4043 	struct	timespec32	timeout;
4044 	uint32_t		flags;
4045 	uint32_t		clockid;
4046 };
4047 
4048 static inline int
4049 umtx_copyin_timeout32(void *addr, struct timespec *tsp)
4050 {
4051 	struct timespec32 ts32;
4052 	int error;
4053 
4054 	error = copyin(addr, &ts32, sizeof(struct timespec32));
4055 	if (error == 0) {
4056 		if (ts32.tv_sec < 0 ||
4057 		    ts32.tv_nsec >= 1000000000 ||
4058 		    ts32.tv_nsec < 0)
4059 			error = EINVAL;
4060 		else {
4061 			tsp->tv_sec = ts32.tv_sec;
4062 			tsp->tv_nsec = ts32.tv_nsec;
4063 		}
4064 	}
4065 	return (error);
4066 }
4067 
4068 static inline int
4069 umtx_copyin_umtx_time32(const void *addr, size_t size, struct _umtx_time *tp)
4070 {
4071 	struct umtx_time32 t32;
4072 	int error;
4073 
4074 	t32.clockid = CLOCK_REALTIME;
4075 	t32.flags   = 0;
4076 	if (size <= sizeof(struct timespec32))
4077 		error = copyin(addr, &t32.timeout, sizeof(struct timespec32));
4078 	else
4079 		error = copyin(addr, &t32, sizeof(struct umtx_time32));
4080 	if (error != 0)
4081 		return (error);
4082 	if (t32.timeout.tv_sec < 0 ||
4083 	    t32.timeout.tv_nsec >= 1000000000 || t32.timeout.tv_nsec < 0)
4084 		return (EINVAL);
4085 	tp->_timeout.tv_sec = t32.timeout.tv_sec;
4086 	tp->_timeout.tv_nsec = t32.timeout.tv_nsec;
4087 	tp->_flags = t32.flags;
4088 	tp->_clockid = t32.clockid;
4089 	return (0);
4090 }
4091 
4092 static int
4093 __umtx_op_wait_compat32(struct thread *td, struct _umtx_op_args *uap)
4094 {
4095 	struct _umtx_time *tm_p, timeout;
4096 	int error;
4097 
4098 	if (uap->uaddr2 == NULL)
4099 		tm_p = NULL;
4100 	else {
4101 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4102 			(size_t)uap->uaddr1, &timeout);
4103 		if (error != 0)
4104 			return (error);
4105 		tm_p = &timeout;
4106 	}
4107 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
4108 }
4109 
4110 static int
4111 __umtx_op_lock_umutex_compat32(struct thread *td, struct _umtx_op_args *uap)
4112 {
4113 	struct _umtx_time *tm_p, timeout;
4114 	int error;
4115 
4116 	/* Allow a null timespec (wait forever). */
4117 	if (uap->uaddr2 == NULL)
4118 		tm_p = NULL;
4119 	else {
4120 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4121 			    (size_t)uap->uaddr1, &timeout);
4122 		if (error != 0)
4123 			return (error);
4124 		tm_p = &timeout;
4125 	}
4126 	return (do_lock_umutex(td, uap->obj, tm_p, 0));
4127 }
4128 
4129 static int
4130 __umtx_op_wait_umutex_compat32(struct thread *td, struct _umtx_op_args *uap)
4131 {
4132 	struct _umtx_time *tm_p, timeout;
4133 	int error;
4134 
4135 	/* Allow a null timespec (wait forever). */
4136 	if (uap->uaddr2 == NULL)
4137 		tm_p = NULL;
4138 	else {
4139 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4140 		    (size_t)uap->uaddr1, &timeout);
4141 		if (error != 0)
4142 			return (error);
4143 		tm_p = &timeout;
4144 	}
4145 	return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
4146 }
4147 
4148 static int
4149 __umtx_op_cv_wait_compat32(struct thread *td, struct _umtx_op_args *uap)
4150 {
4151 	struct timespec *ts, timeout;
4152 	int error;
4153 
4154 	/* Allow a null timespec (wait forever). */
4155 	if (uap->uaddr2 == NULL)
4156 		ts = NULL;
4157 	else {
4158 		error = umtx_copyin_timeout32(uap->uaddr2, &timeout);
4159 		if (error != 0)
4160 			return (error);
4161 		ts = &timeout;
4162 	}
4163 	return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
4164 }
4165 
4166 static int
4167 __umtx_op_rw_rdlock_compat32(struct thread *td, struct _umtx_op_args *uap)
4168 {
4169 	struct _umtx_time timeout;
4170 	int error;
4171 
4172 	/* Allow a null timespec (wait forever). */
4173 	if (uap->uaddr2 == NULL) {
4174 		error = do_rw_rdlock(td, uap->obj, uap->val, 0);
4175 	} else {
4176 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4177 		    (size_t)uap->uaddr1, &timeout);
4178 		if (error != 0)
4179 			return (error);
4180 		error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
4181 	}
4182 	return (error);
4183 }
4184 
4185 static int
4186 __umtx_op_rw_wrlock_compat32(struct thread *td, struct _umtx_op_args *uap)
4187 {
4188 	struct _umtx_time timeout;
4189 	int error;
4190 
4191 	/* Allow a null timespec (wait forever). */
4192 	if (uap->uaddr2 == NULL) {
4193 		error = do_rw_wrlock(td, uap->obj, 0);
4194 	} else {
4195 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4196 		    (size_t)uap->uaddr1, &timeout);
4197 		if (error != 0)
4198 			return (error);
4199 		error = do_rw_wrlock(td, uap->obj, &timeout);
4200 	}
4201 	return (error);
4202 }
4203 
4204 static int
4205 __umtx_op_wait_uint_private_compat32(struct thread *td, struct _umtx_op_args *uap)
4206 {
4207 	struct _umtx_time *tm_p, timeout;
4208 	int error;
4209 
4210 	if (uap->uaddr2 == NULL)
4211 		tm_p = NULL;
4212 	else {
4213 		error = umtx_copyin_umtx_time32(
4214 		    uap->uaddr2, (size_t)uap->uaddr1,&timeout);
4215 		if (error != 0)
4216 			return (error);
4217 		tm_p = &timeout;
4218 	}
4219 	return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
4220 }
4221 
4222 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4223 static int
4224 __umtx_op_sem_wait_compat32(struct thread *td, struct _umtx_op_args *uap)
4225 {
4226 	struct _umtx_time *tm_p, timeout;
4227 	int error;
4228 
4229 	/* Allow a null timespec (wait forever). */
4230 	if (uap->uaddr2 == NULL)
4231 		tm_p = NULL;
4232 	else {
4233 		error = umtx_copyin_umtx_time32(uap->uaddr2,
4234 		    (size_t)uap->uaddr1, &timeout);
4235 		if (error != 0)
4236 			return (error);
4237 		tm_p = &timeout;
4238 	}
4239 	return (do_sem_wait(td, uap->obj, tm_p));
4240 }
4241 #endif
4242 
4243 static int
4244 __umtx_op_sem2_wait_compat32(struct thread *td, struct _umtx_op_args *uap)
4245 {
4246 	struct _umtx_time *tm_p, timeout;
4247 	size_t uasize;
4248 	int error;
4249 
4250 	/* Allow a null timespec (wait forever). */
4251 	if (uap->uaddr2 == NULL) {
4252 		uasize = 0;
4253 		tm_p = NULL;
4254 	} else {
4255 		uasize = (size_t)uap->uaddr1;
4256 		error = umtx_copyin_umtx_time32(uap->uaddr2, uasize, &timeout);
4257 		if (error != 0)
4258 			return (error);
4259 		tm_p = &timeout;
4260 	}
4261 	error = do_sem2_wait(td, uap->obj, tm_p);
4262 	if (error == EINTR && uap->uaddr2 != NULL &&
4263 	    (timeout._flags & UMTX_ABSTIME) == 0 &&
4264 	    uasize >= sizeof(struct umtx_time32) + sizeof(struct timespec32)) {
4265 		struct timespec32 remain32 = {
4266 			.tv_sec = timeout._timeout.tv_sec,
4267 			.tv_nsec = timeout._timeout.tv_nsec
4268 		};
4269 		error = copyout(&remain32,
4270 		    (struct umtx_time32 *)uap->uaddr2 + 1,
4271 		    sizeof(struct timespec32));
4272 		if (error == 0) {
4273 			error = EINTR;
4274 		}
4275 	}
4276 
4277 	return (error);
4278 }
4279 
4280 static int
4281 __umtx_op_nwake_private32(struct thread *td, struct _umtx_op_args *uap)
4282 {
4283 	uint32_t uaddrs[BATCH_SIZE], **upp;
4284 	int count, error, i, pos, tocopy;
4285 
4286 	upp = (uint32_t **)uap->obj;
4287 	error = 0;
4288 	for (count = uap->val, pos = 0; count > 0; count -= tocopy,
4289 	    pos += tocopy) {
4290 		tocopy = MIN(count, BATCH_SIZE);
4291 		error = copyin(upp + pos, uaddrs, tocopy * sizeof(uint32_t));
4292 		if (error != 0)
4293 			break;
4294 		for (i = 0; i < tocopy; ++i)
4295 			kern_umtx_wake(td, (void *)(intptr_t)uaddrs[i],
4296 			    INT_MAX, 1);
4297 		maybe_yield();
4298 	}
4299 	return (error);
4300 }
4301 
4302 struct umtx_robust_lists_params_compat32 {
4303 	uint32_t	robust_list_offset;
4304 	uint32_t	robust_priv_list_offset;
4305 	uint32_t	robust_inact_offset;
4306 };
4307 
4308 static int
4309 __umtx_op_robust_lists_compat32(struct thread *td, struct _umtx_op_args *uap)
4310 {
4311 	struct umtx_robust_lists_params rb;
4312 	struct umtx_robust_lists_params_compat32 rb32;
4313 	int error;
4314 
4315 	if (uap->val > sizeof(rb32))
4316 		return (EINVAL);
4317 	bzero(&rb, sizeof(rb));
4318 	bzero(&rb32, sizeof(rb32));
4319 	error = copyin(uap->uaddr1, &rb32, uap->val);
4320 	if (error != 0)
4321 		return (error);
4322 	rb.robust_list_offset = rb32.robust_list_offset;
4323 	rb.robust_priv_list_offset = rb32.robust_priv_list_offset;
4324 	rb.robust_inact_offset = rb32.robust_inact_offset;
4325 	return (umtx_robust_lists(td, &rb));
4326 }
4327 
4328 static const _umtx_op_func op_table_compat32[] = {
4329 	[UMTX_OP_RESERVED0]	= __umtx_op_unimpl,
4330 	[UMTX_OP_RESERVED1]	= __umtx_op_unimpl,
4331 	[UMTX_OP_WAIT]		= __umtx_op_wait_compat32,
4332 	[UMTX_OP_WAKE]		= __umtx_op_wake,
4333 	[UMTX_OP_MUTEX_TRYLOCK]	= __umtx_op_trylock_umutex,
4334 	[UMTX_OP_MUTEX_LOCK]	= __umtx_op_lock_umutex_compat32,
4335 	[UMTX_OP_MUTEX_UNLOCK]	= __umtx_op_unlock_umutex,
4336 	[UMTX_OP_SET_CEILING]	= __umtx_op_set_ceiling,
4337 	[UMTX_OP_CV_WAIT]	= __umtx_op_cv_wait_compat32,
4338 	[UMTX_OP_CV_SIGNAL]	= __umtx_op_cv_signal,
4339 	[UMTX_OP_CV_BROADCAST]	= __umtx_op_cv_broadcast,
4340 	[UMTX_OP_WAIT_UINT]	= __umtx_op_wait_compat32,
4341 	[UMTX_OP_RW_RDLOCK]	= __umtx_op_rw_rdlock_compat32,
4342 	[UMTX_OP_RW_WRLOCK]	= __umtx_op_rw_wrlock_compat32,
4343 	[UMTX_OP_RW_UNLOCK]	= __umtx_op_rw_unlock,
4344 	[UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private_compat32,
4345 	[UMTX_OP_WAKE_PRIVATE]	= __umtx_op_wake_private,
4346 	[UMTX_OP_MUTEX_WAIT]	= __umtx_op_wait_umutex_compat32,
4347 	[UMTX_OP_MUTEX_WAKE]	= __umtx_op_wake_umutex,
4348 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4349 	[UMTX_OP_SEM_WAIT]	= __umtx_op_sem_wait_compat32,
4350 	[UMTX_OP_SEM_WAKE]	= __umtx_op_sem_wake,
4351 #else
4352 	[UMTX_OP_SEM_WAIT]	= __umtx_op_unimpl,
4353 	[UMTX_OP_SEM_WAKE]	= __umtx_op_unimpl,
4354 #endif
4355 	[UMTX_OP_NWAKE_PRIVATE]	= __umtx_op_nwake_private32,
4356 	[UMTX_OP_MUTEX_WAKE2]	= __umtx_op_wake2_umutex,
4357 	[UMTX_OP_SEM2_WAIT]	= __umtx_op_sem2_wait_compat32,
4358 	[UMTX_OP_SEM2_WAKE]	= __umtx_op_sem2_wake,
4359 	[UMTX_OP_SHM]		= __umtx_op_shm,
4360 	[UMTX_OP_ROBUST_LISTS]	= __umtx_op_robust_lists_compat32,
4361 };
4362 
4363 int
4364 freebsd32_umtx_op(struct thread *td, struct freebsd32_umtx_op_args *uap)
4365 {
4366 
4367 	if ((unsigned)uap->op < nitems(op_table_compat32)) {
4368 		return (*op_table_compat32[uap->op])(td,
4369 		    (struct _umtx_op_args *)uap);
4370 	}
4371 	return (EINVAL);
4372 }
4373 #endif
4374 
4375 void
4376 umtx_thread_init(struct thread *td)
4377 {
4378 
4379 	td->td_umtxq = umtxq_alloc();
4380 	td->td_umtxq->uq_thread = td;
4381 }
4382 
4383 void
4384 umtx_thread_fini(struct thread *td)
4385 {
4386 
4387 	umtxq_free(td->td_umtxq);
4388 }
4389 
4390 /*
4391  * It will be called when new thread is created, e.g fork().
4392  */
4393 void
4394 umtx_thread_alloc(struct thread *td)
4395 {
4396 	struct umtx_q *uq;
4397 
4398 	uq = td->td_umtxq;
4399 	uq->uq_inherited_pri = PRI_MAX;
4400 
4401 	KASSERT(uq->uq_flags == 0, ("uq_flags != 0"));
4402 	KASSERT(uq->uq_thread == td, ("uq_thread != td"));
4403 	KASSERT(uq->uq_pi_blocked == NULL, ("uq_pi_blocked != NULL"));
4404 	KASSERT(TAILQ_EMPTY(&uq->uq_pi_contested), ("uq_pi_contested is not empty"));
4405 }
4406 
4407 /*
4408  * exec() hook.
4409  *
4410  * Clear robust lists for all process' threads, not delaying the
4411  * cleanup to thread_exit hook, since the relevant address space is
4412  * destroyed right now.
4413  */
4414 static void
4415 umtx_exec_hook(void *arg __unused, struct proc *p,
4416     struct image_params *imgp __unused)
4417 {
4418 	struct thread *td;
4419 
4420 	KASSERT(p == curproc, ("need curproc"));
4421 	PROC_LOCK(p);
4422 	KASSERT((p->p_flag & P_HADTHREADS) == 0 ||
4423 	    (p->p_flag & P_STOPPED_SINGLE) != 0,
4424 	    ("curproc must be single-threaded"));
4425 	FOREACH_THREAD_IN_PROC(p, td) {
4426 		KASSERT(td == curthread ||
4427 		    ((td->td_flags & TDF_BOUNDARY) != 0 && TD_IS_SUSPENDED(td)),
4428 		    ("running thread %p %p", p, td));
4429 		PROC_UNLOCK(p);
4430 		umtx_thread_cleanup(td);
4431 		PROC_LOCK(p);
4432 		td->td_rb_list = td->td_rbp_list = td->td_rb_inact = 0;
4433 	}
4434 	PROC_UNLOCK(p);
4435 }
4436 
4437 /*
4438  * thread_exit() hook.
4439  */
4440 void
4441 umtx_thread_exit(struct thread *td)
4442 {
4443 
4444 	umtx_thread_cleanup(td);
4445 }
4446 
4447 static int
4448 umtx_read_uptr(struct thread *td, uintptr_t ptr, uintptr_t *res)
4449 {
4450 	u_long res1;
4451 #ifdef COMPAT_FREEBSD32
4452 	uint32_t res32;
4453 #endif
4454 	int error;
4455 
4456 #ifdef COMPAT_FREEBSD32
4457 	if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
4458 		error = fueword32((void *)ptr, &res32);
4459 		if (error == 0)
4460 			res1 = res32;
4461 	} else
4462 #endif
4463 	{
4464 		error = fueword((void *)ptr, &res1);
4465 	}
4466 	if (error == 0)
4467 		*res = res1;
4468 	else
4469 		error = EFAULT;
4470 	return (error);
4471 }
4472 
4473 static void
4474 umtx_read_rb_list(struct thread *td, struct umutex *m, uintptr_t *rb_list)
4475 {
4476 #ifdef COMPAT_FREEBSD32
4477 	struct umutex32 m32;
4478 
4479 	if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
4480 		memcpy(&m32, m, sizeof(m32));
4481 		*rb_list = m32.m_rb_lnk;
4482 	} else
4483 #endif
4484 		*rb_list = m->m_rb_lnk;
4485 }
4486 
4487 static int
4488 umtx_handle_rb(struct thread *td, uintptr_t rbp, uintptr_t *rb_list, bool inact)
4489 {
4490 	struct umutex m;
4491 	int error;
4492 
4493 	KASSERT(td->td_proc == curproc, ("need current vmspace"));
4494 	error = copyin((void *)rbp, &m, sizeof(m));
4495 	if (error != 0)
4496 		return (error);
4497 	if (rb_list != NULL)
4498 		umtx_read_rb_list(td, &m, rb_list);
4499 	if ((m.m_flags & UMUTEX_ROBUST) == 0)
4500 		return (EINVAL);
4501 	if ((m.m_owner & ~UMUTEX_CONTESTED) != td->td_tid)
4502 		/* inact is cleared after unlock, allow the inconsistency */
4503 		return (inact ? 0 : EINVAL);
4504 	return (do_unlock_umutex(td, (struct umutex *)rbp, true));
4505 }
4506 
4507 static void
4508 umtx_cleanup_rb_list(struct thread *td, uintptr_t rb_list, uintptr_t *rb_inact,
4509     const char *name)
4510 {
4511 	int error, i;
4512 	uintptr_t rbp;
4513 	bool inact;
4514 
4515 	if (rb_list == 0)
4516 		return;
4517 	error = umtx_read_uptr(td, rb_list, &rbp);
4518 	for (i = 0; error == 0 && rbp != 0 && i < umtx_max_rb; i++) {
4519 		if (rbp == *rb_inact) {
4520 			inact = true;
4521 			*rb_inact = 0;
4522 		} else
4523 			inact = false;
4524 		error = umtx_handle_rb(td, rbp, &rbp, inact);
4525 	}
4526 	if (i == umtx_max_rb && umtx_verbose_rb) {
4527 		uprintf("comm %s pid %d: reached umtx %smax rb %d\n",
4528 		    td->td_proc->p_comm, td->td_proc->p_pid, name, umtx_max_rb);
4529 	}
4530 	if (error != 0 && umtx_verbose_rb) {
4531 		uprintf("comm %s pid %d: handling %srb error %d\n",
4532 		    td->td_proc->p_comm, td->td_proc->p_pid, name, error);
4533 	}
4534 }
4535 
4536 /*
4537  * Clean up umtx data.
4538  */
4539 static void
4540 umtx_thread_cleanup(struct thread *td)
4541 {
4542 	struct umtx_q *uq;
4543 	struct umtx_pi *pi;
4544 	uintptr_t rb_inact;
4545 
4546 	/*
4547 	 * Disown pi mutexes.
4548 	 */
4549 	uq = td->td_umtxq;
4550 	if (uq != NULL) {
4551 		mtx_lock(&umtx_lock);
4552 		uq->uq_inherited_pri = PRI_MAX;
4553 		while ((pi = TAILQ_FIRST(&uq->uq_pi_contested)) != NULL) {
4554 			pi->pi_owner = NULL;
4555 			TAILQ_REMOVE(&uq->uq_pi_contested, pi, pi_link);
4556 		}
4557 		mtx_unlock(&umtx_lock);
4558 		thread_lock(td);
4559 		sched_lend_user_prio(td, PRI_MAX);
4560 		thread_unlock(td);
4561 	}
4562 
4563 	/*
4564 	 * Handle terminated robust mutexes.  Must be done after
4565 	 * robust pi disown, otherwise unlock could see unowned
4566 	 * entries.
4567 	 */
4568 	rb_inact = td->td_rb_inact;
4569 	if (rb_inact != 0)
4570 		(void)umtx_read_uptr(td, rb_inact, &rb_inact);
4571 	umtx_cleanup_rb_list(td, td->td_rb_list, &rb_inact, "");
4572 	umtx_cleanup_rb_list(td, td->td_rbp_list, &rb_inact, "priv ");
4573 	if (rb_inact != 0)
4574 		(void)umtx_handle_rb(td, rb_inact, NULL, true);
4575 }
4576