2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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>
9 * Portions of this software were developed by Konstantin Belousov
10 * under sponsorship from the FreeBSD Foundation.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice unmodified, this list of conditions, and the following
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.
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.
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include "opt_umtx_profiling.h"
39 #include <sys/param.h>
40 #include <sys/kernel.h>
41 #include <sys/fcntl.h>
43 #include <sys/filedesc.h>
44 #include <sys/limits.h>
46 #include <sys/malloc.h>
48 #include <sys/mutex.h>
51 #include <sys/resource.h>
52 #include <sys/resourcevar.h>
53 #include <sys/rwlock.h>
55 #include <sys/sched.h>
57 #include <sys/sysctl.h>
58 #include <sys/systm.h>
59 #include <sys/sysproto.h>
60 #include <sys/syscallsubr.h>
61 #include <sys/taskqueue.h>
63 #include <sys/eventhandler.h>
65 #include <sys/umtxvar.h>
67 #include <security/mac/mac_framework.h>
70 #include <vm/vm_param.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_object.h>
75 #include <machine/atomic.h>
76 #include <machine/cpu.h>
78 #include <compat/freebsd32/freebsd32.h>
79 #ifdef COMPAT_FREEBSD32
80 #include <compat/freebsd32/freebsd32_proto.h>
84 #define _UMUTEX_WAIT 2
87 #define UPROF_PERC_BIGGER(w, f, sw, sf) \
88 (((w) > (sw)) || ((w) == (sw) && (f) > (sf)))
91 #define UMTXQ_LOCKED_ASSERT(uc) mtx_assert(&(uc)->uc_lock, MA_OWNED)
93 #define UMTXQ_ASSERT_LOCKED_BUSY(key) do { \
94 struct umtxq_chain *uc; \
96 uc = umtxq_getchain(key); \
97 mtx_assert(&uc->uc_lock, MA_OWNED); \
98 KASSERT(uc->uc_busy != 0, ("umtx chain is not busy")); \
101 #define UMTXQ_ASSERT_LOCKED_BUSY(key) do {} while (0)
105 * Don't propagate time-sharing priority, there is a security reason,
106 * a user can simply introduce PI-mutex, let thread A lock the mutex,
107 * and let another thread B block on the mutex, because B is
108 * sleeping, its priority will be boosted, this causes A's priority to
109 * be boosted via priority propagating too and will never be lowered even
110 * if it is using 100%CPU, this is unfair to other processes.
113 #define UPRI(td) (((td)->td_user_pri >= PRI_MIN_TIMESHARE &&\
114 (td)->td_user_pri <= PRI_MAX_TIMESHARE) ?\
115 PRI_MAX_TIMESHARE : (td)->td_user_pri)
117 #define GOLDEN_RATIO_PRIME 2654404609U
119 #define UMTX_CHAINS 512
121 #define UMTX_SHIFTS (__WORD_BIT - 9)
123 #define GET_SHARE(flags) \
124 (((flags) & USYNC_PROCESS_SHARED) == 0 ? THREAD_SHARE : PROCESS_SHARE)
126 #define BUSY_SPINS 200
128 struct umtx_copyops {
129 int (*copyin_timeout)(const void *uaddr, struct timespec *tsp);
130 int (*copyin_umtx_time)(const void *uaddr, size_t size,
131 struct _umtx_time *tp);
132 int (*copyin_robust_lists)(const void *uaddr, size_t size,
133 struct umtx_robust_lists_params *rbp);
134 int (*copyout_timeout)(void *uaddr, size_t size,
135 struct timespec *tsp);
136 const size_t timespec_sz;
137 const size_t umtx_time_sz;
141 _Static_assert(sizeof(struct umutex) == sizeof(struct umutex32), "umutex32");
142 _Static_assert(__offsetof(struct umutex, m_spare[0]) ==
143 __offsetof(struct umutex32, m_spare[0]), "m_spare32");
145 int umtx_shm_vnobj_persistent = 0;
146 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_vnode_persistent, CTLFLAG_RWTUN,
147 &umtx_shm_vnobj_persistent, 0,
148 "False forces destruction of umtx attached to file, on last close");
149 static int umtx_max_rb = 1000;
150 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_max_robust, CTLFLAG_RWTUN,
152 "Maximum number of robust mutexes allowed for each thread");
154 static uma_zone_t umtx_pi_zone;
155 static struct umtxq_chain umtxq_chains[2][UMTX_CHAINS];
156 static MALLOC_DEFINE(M_UMTX, "umtx", "UMTX queue memory");
157 static int umtx_pi_allocated;
159 static SYSCTL_NODE(_debug, OID_AUTO, umtx, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
161 SYSCTL_INT(_debug_umtx, OID_AUTO, umtx_pi_allocated, CTLFLAG_RD,
162 &umtx_pi_allocated, 0, "Allocated umtx_pi");
163 static int umtx_verbose_rb = 1;
164 SYSCTL_INT(_debug_umtx, OID_AUTO, robust_faults_verbose, CTLFLAG_RWTUN,
168 #ifdef UMTX_PROFILING
169 static long max_length;
170 SYSCTL_LONG(_debug_umtx, OID_AUTO, max_length, CTLFLAG_RD, &max_length, 0, "max_length");
171 static SYSCTL_NODE(_debug_umtx, OID_AUTO, chains, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
175 static inline void umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
176 const struct _umtx_time *umtxtime);
178 static void umtx_shm_init(void);
179 static void umtxq_sysinit(void *);
180 static void umtxq_hash(struct umtx_key *key);
181 static int do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags,
183 static void umtx_thread_cleanup(struct thread *td);
184 SYSINIT(umtx, SI_SUB_EVENTHANDLER+1, SI_ORDER_MIDDLE, umtxq_sysinit, NULL);
186 #define umtxq_signal(key, nwake) umtxq_signal_queue((key), (nwake), UMTX_SHARED_QUEUE)
188 static struct mtx umtx_lock;
190 #ifdef UMTX_PROFILING
192 umtx_init_profiling(void)
194 struct sysctl_oid *chain_oid;
198 for (i = 0; i < UMTX_CHAINS; ++i) {
199 snprintf(chain_name, sizeof(chain_name), "%d", i);
200 chain_oid = SYSCTL_ADD_NODE(NULL,
201 SYSCTL_STATIC_CHILDREN(_debug_umtx_chains), OID_AUTO,
202 chain_name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
204 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
205 "max_length0", CTLFLAG_RD, &umtxq_chains[0][i].max_length, 0, NULL);
206 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
207 "max_length1", CTLFLAG_RD, &umtxq_chains[1][i].max_length, 0, NULL);
212 sysctl_debug_umtx_chains_peaks(SYSCTL_HANDLER_ARGS)
216 struct umtxq_chain *uc;
217 u_int fract, i, j, tot, whole;
218 u_int sf0, sf1, sf2, sf3, sf4;
219 u_int si0, si1, si2, si3, si4;
220 u_int sw0, sw1, sw2, sw3, sw4;
222 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
223 for (i = 0; i < 2; i++) {
225 for (j = 0; j < UMTX_CHAINS; ++j) {
226 uc = &umtxq_chains[i][j];
227 mtx_lock(&uc->uc_lock);
228 tot += uc->max_length;
229 mtx_unlock(&uc->uc_lock);
232 sbuf_printf(&sb, "%u) Empty ", i);
234 sf0 = sf1 = sf2 = sf3 = sf4 = 0;
235 si0 = si1 = si2 = si3 = si4 = 0;
236 sw0 = sw1 = sw2 = sw3 = sw4 = 0;
237 for (j = 0; j < UMTX_CHAINS; j++) {
238 uc = &umtxq_chains[i][j];
239 mtx_lock(&uc->uc_lock);
240 whole = uc->max_length * 100;
241 mtx_unlock(&uc->uc_lock);
242 fract = (whole % tot) * 100;
243 if (UPROF_PERC_BIGGER(whole, fract, sw0, sf0)) {
247 } else if (UPROF_PERC_BIGGER(whole, fract, sw1,
252 } else if (UPROF_PERC_BIGGER(whole, fract, sw2,
257 } else if (UPROF_PERC_BIGGER(whole, fract, sw3,
262 } else if (UPROF_PERC_BIGGER(whole, fract, sw4,
269 sbuf_printf(&sb, "queue %u:\n", i);
270 sbuf_printf(&sb, "1st: %u.%u%% idx: %u\n", sw0 / tot,
272 sbuf_printf(&sb, "2nd: %u.%u%% idx: %u\n", sw1 / tot,
274 sbuf_printf(&sb, "3rd: %u.%u%% idx: %u\n", sw2 / tot,
276 sbuf_printf(&sb, "4th: %u.%u%% idx: %u\n", sw3 / tot,
278 sbuf_printf(&sb, "5th: %u.%u%% idx: %u\n", sw4 / tot,
284 sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
290 sysctl_debug_umtx_chains_clear(SYSCTL_HANDLER_ARGS)
292 struct umtxq_chain *uc;
297 error = sysctl_handle_int(oidp, &clear, 0, req);
298 if (error != 0 || req->newptr == NULL)
302 for (i = 0; i < 2; ++i) {
303 for (j = 0; j < UMTX_CHAINS; ++j) {
304 uc = &umtxq_chains[i][j];
305 mtx_lock(&uc->uc_lock);
308 mtx_unlock(&uc->uc_lock);
315 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, clear,
316 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
317 sysctl_debug_umtx_chains_clear, "I",
318 "Clear umtx chains statistics");
319 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, peaks,
320 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
321 sysctl_debug_umtx_chains_peaks, "A",
322 "Highest peaks in chains max length");
326 umtxq_sysinit(void *arg __unused)
330 umtx_pi_zone = uma_zcreate("umtx pi", sizeof(struct umtx_pi),
331 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
332 for (i = 0; i < 2; ++i) {
333 for (j = 0; j < UMTX_CHAINS; ++j) {
334 mtx_init(&umtxq_chains[i][j].uc_lock, "umtxql", NULL,
335 MTX_DEF | MTX_DUPOK);
336 LIST_INIT(&umtxq_chains[i][j].uc_queue[0]);
337 LIST_INIT(&umtxq_chains[i][j].uc_queue[1]);
338 LIST_INIT(&umtxq_chains[i][j].uc_spare_queue);
339 TAILQ_INIT(&umtxq_chains[i][j].uc_pi_list);
340 umtxq_chains[i][j].uc_busy = 0;
341 umtxq_chains[i][j].uc_waiters = 0;
342 #ifdef UMTX_PROFILING
343 umtxq_chains[i][j].length = 0;
344 umtxq_chains[i][j].max_length = 0;
348 #ifdef UMTX_PROFILING
349 umtx_init_profiling();
351 mtx_init(&umtx_lock, "umtx lock", NULL, MTX_DEF);
360 uq = malloc(sizeof(struct umtx_q), M_UMTX, M_WAITOK | M_ZERO);
361 uq->uq_spare_queue = malloc(sizeof(struct umtxq_queue), M_UMTX,
363 TAILQ_INIT(&uq->uq_spare_queue->head);
364 TAILQ_INIT(&uq->uq_pi_contested);
365 uq->uq_inherited_pri = PRI_MAX;
370 umtxq_free(struct umtx_q *uq)
373 MPASS(uq->uq_spare_queue != NULL);
374 free(uq->uq_spare_queue, M_UMTX);
379 umtxq_hash(struct umtx_key *key)
383 n = (uintptr_t)key->info.both.a + key->info.both.b;
384 key->hash = ((n * GOLDEN_RATIO_PRIME) >> UMTX_SHIFTS) % UMTX_CHAINS;
388 umtxq_getchain(struct umtx_key *key)
391 if (key->type <= TYPE_SEM)
392 return (&umtxq_chains[1][key->hash]);
393 return (&umtxq_chains[0][key->hash]);
397 * Set chain to busy state when following operation
398 * may be blocked (kernel mutex can not be used).
401 umtxq_busy(struct umtx_key *key)
403 struct umtxq_chain *uc;
405 uc = umtxq_getchain(key);
406 mtx_assert(&uc->uc_lock, MA_OWNED);
410 int count = BUSY_SPINS;
413 while (uc->uc_busy && --count > 0)
419 while (uc->uc_busy) {
421 msleep(uc, &uc->uc_lock, 0, "umtxqb", 0);
432 umtxq_unbusy(struct umtx_key *key)
434 struct umtxq_chain *uc;
436 uc = umtxq_getchain(key);
437 mtx_assert(&uc->uc_lock, MA_OWNED);
438 KASSERT(uc->uc_busy != 0, ("not busy"));
445 umtxq_unbusy_unlocked(struct umtx_key *key)
453 static struct umtxq_queue *
454 umtxq_queue_lookup(struct umtx_key *key, int q)
456 struct umtxq_queue *uh;
457 struct umtxq_chain *uc;
459 uc = umtxq_getchain(key);
460 UMTXQ_LOCKED_ASSERT(uc);
461 LIST_FOREACH(uh, &uc->uc_queue[q], link) {
462 if (umtx_key_match(&uh->key, key))
470 umtxq_insert_queue(struct umtx_q *uq, int q)
472 struct umtxq_queue *uh;
473 struct umtxq_chain *uc;
475 uc = umtxq_getchain(&uq->uq_key);
476 UMTXQ_LOCKED_ASSERT(uc);
477 KASSERT((uq->uq_flags & UQF_UMTXQ) == 0, ("umtx_q is already on queue"));
478 uh = umtxq_queue_lookup(&uq->uq_key, q);
480 LIST_INSERT_HEAD(&uc->uc_spare_queue, uq->uq_spare_queue, link);
482 uh = uq->uq_spare_queue;
483 uh->key = uq->uq_key;
484 LIST_INSERT_HEAD(&uc->uc_queue[q], uh, link);
485 #ifdef UMTX_PROFILING
487 if (uc->length > uc->max_length) {
488 uc->max_length = uc->length;
489 if (uc->max_length > max_length)
490 max_length = uc->max_length;
494 uq->uq_spare_queue = NULL;
496 TAILQ_INSERT_TAIL(&uh->head, uq, uq_link);
498 uq->uq_flags |= UQF_UMTXQ;
499 uq->uq_cur_queue = uh;
504 umtxq_remove_queue(struct umtx_q *uq, int q)
506 struct umtxq_chain *uc;
507 struct umtxq_queue *uh;
509 uc = umtxq_getchain(&uq->uq_key);
510 UMTXQ_LOCKED_ASSERT(uc);
511 if (uq->uq_flags & UQF_UMTXQ) {
512 uh = uq->uq_cur_queue;
513 TAILQ_REMOVE(&uh->head, uq, uq_link);
515 uq->uq_flags &= ~UQF_UMTXQ;
516 if (TAILQ_EMPTY(&uh->head)) {
517 KASSERT(uh->length == 0,
518 ("inconsistent umtxq_queue length"));
519 #ifdef UMTX_PROFILING
522 LIST_REMOVE(uh, link);
524 uh = LIST_FIRST(&uc->uc_spare_queue);
525 KASSERT(uh != NULL, ("uc_spare_queue is empty"));
526 LIST_REMOVE(uh, link);
528 uq->uq_spare_queue = uh;
529 uq->uq_cur_queue = NULL;
534 * Check if there are multiple waiters
537 umtxq_count(struct umtx_key *key)
539 struct umtxq_queue *uh;
541 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
542 uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
549 * Check if there are multiple PI waiters and returns first
553 umtxq_count_pi(struct umtx_key *key, struct umtx_q **first)
555 struct umtxq_queue *uh;
558 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
559 uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
561 *first = TAILQ_FIRST(&uh->head);
568 * Wake up threads waiting on an userland object by a bit mask.
571 umtxq_signal_mask(struct umtx_key *key, int n_wake, u_int bitset)
573 struct umtxq_queue *uh;
574 struct umtx_q *uq, *uq_temp;
578 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
579 uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
582 TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
583 if ((uq->uq_bitset & bitset) == 0)
585 umtxq_remove_queue(uq, UMTX_SHARED_QUEUE);
594 * Wake up threads waiting on an userland object.
598 umtxq_signal_queue(struct umtx_key *key, int n_wake, int q)
600 struct umtxq_queue *uh;
605 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
606 uh = umtxq_queue_lookup(key, q);
608 while ((uq = TAILQ_FIRST(&uh->head)) != NULL) {
609 umtxq_remove_queue(uq, q);
619 * Wake up specified thread.
622 umtxq_signal_thread(struct umtx_q *uq)
625 UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
631 * Wake up a maximum of n_wake threads that are waiting on an userland
632 * object identified by key. The remaining threads are removed from queue
633 * identified by key and added to the queue identified by key2 (requeued).
634 * The n_requeue specifies an upper limit on the number of threads that
635 * are requeued to the second queue.
638 umtxq_requeue(struct umtx_key *key, int n_wake, struct umtx_key *key2,
641 struct umtxq_queue *uh;
642 struct umtx_q *uq, *uq_temp;
646 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
647 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key2));
648 uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
651 TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
652 if (++ret <= n_wake) {
659 if (ret - n_wake == n_requeue)
667 tstohz(const struct timespec *tsp)
671 TIMESPEC_TO_TIMEVAL(&tv, tsp);
676 umtx_abs_timeout_init(struct umtx_abs_timeout *timo, int clockid,
677 int absolute, const struct timespec *timeout)
680 timo->clockid = clockid;
682 timo->is_abs_real = false;
683 kern_clock_gettime(curthread, timo->clockid, &timo->cur);
684 timespecadd(&timo->cur, timeout, &timo->end);
686 timo->end = *timeout;
687 timo->is_abs_real = clockid == CLOCK_REALTIME ||
688 clockid == CLOCK_REALTIME_FAST ||
689 clockid == CLOCK_REALTIME_PRECISE ||
690 clockid == CLOCK_SECOND;
695 umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
696 const struct _umtx_time *umtxtime)
699 umtx_abs_timeout_init(timo, umtxtime->_clockid,
700 (umtxtime->_flags & UMTX_ABSTIME) != 0, &umtxtime->_timeout);
704 umtx_abs_timeout_getsbt(struct umtx_abs_timeout *timo, sbintime_t *sbt,
707 struct bintime bt, bbt;
711 switch (timo->clockid) {
713 /* Clocks that can be converted into absolute time. */
715 case CLOCK_REALTIME_PRECISE:
716 case CLOCK_REALTIME_FAST:
717 case CLOCK_MONOTONIC:
718 case CLOCK_MONOTONIC_PRECISE:
719 case CLOCK_MONOTONIC_FAST:
721 case CLOCK_UPTIME_PRECISE:
722 case CLOCK_UPTIME_FAST:
724 timespec2bintime(&timo->end, &bt);
725 switch (timo->clockid) {
727 case CLOCK_REALTIME_PRECISE:
728 case CLOCK_REALTIME_FAST:
730 getboottimebin(&bbt);
731 bintime_sub(&bt, &bbt);
736 if (bt.sec >= (SBT_MAX >> 32)) {
744 * Check if the absolute time should be aligned to
745 * avoid firing multiple timer events in non-periodic
748 switch (timo->clockid) {
749 case CLOCK_REALTIME_FAST:
750 case CLOCK_MONOTONIC_FAST:
751 case CLOCK_UPTIME_FAST:
752 rem = *sbt % tc_tick_sbt;
753 if (__predict_true(rem != 0))
754 *sbt += tc_tick_sbt - rem;
758 if (__predict_true(rem != 0))
759 *sbt += SBT_1S - rem;
765 /* Clocks that has to be periodically polled. */
768 case CLOCK_THREAD_CPUTIME_ID:
769 case CLOCK_PROCESS_CPUTIME_ID:
771 kern_clock_gettime(curthread, timo->clockid, &timo->cur);
772 if (timespeccmp(&timo->end, &timo->cur, <=))
774 timespecsub(&timo->end, &timo->cur, &tts);
775 *sbt = tick_sbt * tstohz(&tts);
776 *flags = C_HARDCLOCK;
782 umtx_unlock_val(uint32_t flags, bool rb)
786 return (UMUTEX_RB_OWNERDEAD);
787 else if ((flags & UMUTEX_NONCONSISTENT) != 0)
788 return (UMUTEX_RB_NOTRECOV);
790 return (UMUTEX_UNOWNED);
795 * Put thread into sleep state, before sleeping, check if
796 * thread was removed from umtx queue.
799 umtxq_sleep(struct umtx_q *uq, const char *wmesg,
800 struct umtx_abs_timeout *timo)
802 struct umtxq_chain *uc;
804 int error, flags = 0;
806 uc = umtxq_getchain(&uq->uq_key);
807 UMTXQ_LOCKED_ASSERT(uc);
809 if (!(uq->uq_flags & UQF_UMTXQ)) {
814 if (timo->is_abs_real)
815 curthread->td_rtcgen =
816 atomic_load_acq_int(&rtc_generation);
817 error = umtx_abs_timeout_getsbt(timo, &sbt, &flags);
821 error = msleep_sbt(uq, &uc->uc_lock, PCATCH | PDROP, wmesg,
823 uc = umtxq_getchain(&uq->uq_key);
824 mtx_lock(&uc->uc_lock);
825 if (error == EINTR || error == ERESTART)
827 if (error == EWOULDBLOCK && (flags & C_ABSOLUTE) != 0) {
833 curthread->td_rtcgen = 0;
838 * Convert userspace address into unique logical address.
841 umtx_key_get(const void *addr, int type, int share, struct umtx_key *key)
843 struct thread *td = curthread;
845 vm_map_entry_t entry;
851 if (share == THREAD_SHARE) {
853 key->info.private.vs = td->td_proc->p_vmspace;
854 key->info.private.addr = (uintptr_t)addr;
856 MPASS(share == PROCESS_SHARE || share == AUTO_SHARE);
857 map = &td->td_proc->p_vmspace->vm_map;
858 if (vm_map_lookup(&map, (vm_offset_t)addr, VM_PROT_WRITE,
859 &entry, &key->info.shared.object, &pindex, &prot,
860 &wired) != KERN_SUCCESS) {
864 if ((share == PROCESS_SHARE) ||
865 (share == AUTO_SHARE &&
866 VM_INHERIT_SHARE == entry->inheritance)) {
868 key->info.shared.offset = (vm_offset_t)addr -
869 entry->start + entry->offset;
870 vm_object_reference(key->info.shared.object);
873 key->info.private.vs = td->td_proc->p_vmspace;
874 key->info.private.addr = (uintptr_t)addr;
876 vm_map_lookup_done(map, entry);
887 umtx_key_release(struct umtx_key *key)
890 vm_object_deallocate(key->info.shared.object);
893 #ifdef COMPAT_FREEBSD10
895 * Lock a umtx object.
898 do_lock_umtx(struct thread *td, struct umtx *umtx, u_long id,
899 const struct timespec *timeout)
901 struct umtx_abs_timeout timo;
909 umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
912 * Care must be exercised when dealing with umtx structure. It
913 * can fault on any access.
917 * Try the uncontested case. This should be done in userland.
919 owner = casuword(&umtx->u_owner, UMTX_UNOWNED, id);
921 /* The acquire succeeded. */
922 if (owner == UMTX_UNOWNED)
925 /* The address was invalid. */
929 /* If no one owns it but it is contested try to acquire it. */
930 if (owner == UMTX_CONTESTED) {
931 owner = casuword(&umtx->u_owner,
932 UMTX_CONTESTED, id | UMTX_CONTESTED);
934 if (owner == UMTX_CONTESTED)
937 /* The address was invalid. */
941 error = thread_check_susp(td, false);
945 /* If this failed the lock has changed, restart. */
950 * If we caught a signal, we have retried and now
956 if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK,
957 AUTO_SHARE, &uq->uq_key)) != 0)
960 umtxq_lock(&uq->uq_key);
961 umtxq_busy(&uq->uq_key);
963 umtxq_unbusy(&uq->uq_key);
964 umtxq_unlock(&uq->uq_key);
967 * Set the contested bit so that a release in user space
968 * knows to use the system call for unlock. If this fails
969 * either some one else has acquired the lock or it has been
972 old = casuword(&umtx->u_owner, owner, owner | UMTX_CONTESTED);
974 /* The address was invalid. */
976 umtxq_lock(&uq->uq_key);
978 umtxq_unlock(&uq->uq_key);
979 umtx_key_release(&uq->uq_key);
984 * We set the contested bit, sleep. Otherwise the lock changed
985 * and we need to retry or we lost a race to the thread
986 * unlocking the umtx.
988 umtxq_lock(&uq->uq_key);
990 error = umtxq_sleep(uq, "umtx", timeout == NULL ? NULL :
993 umtxq_unlock(&uq->uq_key);
994 umtx_key_release(&uq->uq_key);
997 error = thread_check_susp(td, false);
1000 if (timeout == NULL) {
1001 /* Mutex locking is restarted if it is interrupted. */
1005 /* Timed-locking is not restarted. */
1006 if (error == ERESTART)
1013 * Unlock a umtx object.
1016 do_unlock_umtx(struct thread *td, struct umtx *umtx, u_long id)
1018 struct umtx_key key;
1025 * Make sure we own this mtx.
1027 owner = fuword(__DEVOLATILE(u_long *, &umtx->u_owner));
1031 if ((owner & ~UMTX_CONTESTED) != id)
1034 /* This should be done in userland */
1035 if ((owner & UMTX_CONTESTED) == 0) {
1036 old = casuword(&umtx->u_owner, owner, UMTX_UNOWNED);
1044 /* We should only ever be in here for contested locks */
1045 if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1051 count = umtxq_count(&key);
1055 * When unlocking the umtx, it must be marked as unowned if
1056 * there is zero or one thread only waiting for it.
1057 * Otherwise, it must be marked as contested.
1059 old = casuword(&umtx->u_owner, owner,
1060 count <= 1 ? UMTX_UNOWNED : UMTX_CONTESTED);
1062 umtxq_signal(&key,1);
1065 umtx_key_release(&key);
1073 #ifdef COMPAT_FREEBSD32
1076 * Lock a umtx object.
1079 do_lock_umtx32(struct thread *td, uint32_t *m, uint32_t id,
1080 const struct timespec *timeout)
1082 struct umtx_abs_timeout timo;
1090 if (timeout != NULL)
1091 umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
1094 * Care must be exercised when dealing with umtx structure. It
1095 * can fault on any access.
1099 * Try the uncontested case. This should be done in userland.
1101 owner = casuword32(m, UMUTEX_UNOWNED, id);
1103 /* The acquire succeeded. */
1104 if (owner == UMUTEX_UNOWNED)
1107 /* The address was invalid. */
1111 /* If no one owns it but it is contested try to acquire it. */
1112 if (owner == UMUTEX_CONTESTED) {
1113 owner = casuword32(m,
1114 UMUTEX_CONTESTED, id | UMUTEX_CONTESTED);
1115 if (owner == UMUTEX_CONTESTED)
1118 /* The address was invalid. */
1122 error = thread_check_susp(td, false);
1126 /* If this failed the lock has changed, restart. */
1131 * If we caught a signal, we have retried and now
1137 if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK,
1138 AUTO_SHARE, &uq->uq_key)) != 0)
1141 umtxq_lock(&uq->uq_key);
1142 umtxq_busy(&uq->uq_key);
1144 umtxq_unbusy(&uq->uq_key);
1145 umtxq_unlock(&uq->uq_key);
1148 * Set the contested bit so that a release in user space
1149 * knows to use the system call for unlock. If this fails
1150 * either some one else has acquired the lock or it has been
1153 old = casuword32(m, owner, owner | UMUTEX_CONTESTED);
1155 /* The address was invalid. */
1157 umtxq_lock(&uq->uq_key);
1159 umtxq_unlock(&uq->uq_key);
1160 umtx_key_release(&uq->uq_key);
1165 * We set the contested bit, sleep. Otherwise the lock changed
1166 * and we need to retry or we lost a race to the thread
1167 * unlocking the umtx.
1169 umtxq_lock(&uq->uq_key);
1171 error = umtxq_sleep(uq, "umtx", timeout == NULL ?
1174 umtxq_unlock(&uq->uq_key);
1175 umtx_key_release(&uq->uq_key);
1178 error = thread_check_susp(td, false);
1181 if (timeout == NULL) {
1182 /* Mutex locking is restarted if it is interrupted. */
1186 /* Timed-locking is not restarted. */
1187 if (error == ERESTART)
1194 * Unlock a umtx object.
1197 do_unlock_umtx32(struct thread *td, uint32_t *m, uint32_t id)
1199 struct umtx_key key;
1206 * Make sure we own this mtx.
1208 owner = fuword32(m);
1212 if ((owner & ~UMUTEX_CONTESTED) != id)
1215 /* This should be done in userland */
1216 if ((owner & UMUTEX_CONTESTED) == 0) {
1217 old = casuword32(m, owner, UMUTEX_UNOWNED);
1225 /* We should only ever be in here for contested locks */
1226 if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1232 count = umtxq_count(&key);
1236 * When unlocking the umtx, it must be marked as unowned if
1237 * there is zero or one thread only waiting for it.
1238 * Otherwise, it must be marked as contested.
1240 old = casuword32(m, owner,
1241 count <= 1 ? UMUTEX_UNOWNED : UMUTEX_CONTESTED);
1243 umtxq_signal(&key,1);
1246 umtx_key_release(&key);
1253 #endif /* COMPAT_FREEBSD32 */
1254 #endif /* COMPAT_FREEBSD10 */
1257 * Fetch and compare value, sleep on the address if value is not changed.
1260 do_wait(struct thread *td, void *addr, u_long id,
1261 struct _umtx_time *timeout, int compat32, int is_private)
1263 struct umtx_abs_timeout timo;
1270 if ((error = umtx_key_get(addr, TYPE_SIMPLE_WAIT,
1271 is_private ? THREAD_SHARE : AUTO_SHARE, &uq->uq_key)) != 0)
1274 if (timeout != NULL)
1275 umtx_abs_timeout_init2(&timo, timeout);
1277 umtxq_lock(&uq->uq_key);
1279 umtxq_unlock(&uq->uq_key);
1280 if (compat32 == 0) {
1281 error = fueword(addr, &tmp);
1285 error = fueword32(addr, &tmp32);
1291 umtxq_lock(&uq->uq_key);
1294 error = umtxq_sleep(uq, "uwait", timeout == NULL ?
1296 if ((uq->uq_flags & UQF_UMTXQ) == 0)
1300 } else if ((uq->uq_flags & UQF_UMTXQ) != 0) {
1303 umtxq_unlock(&uq->uq_key);
1304 umtx_key_release(&uq->uq_key);
1305 if (error == ERESTART)
1311 * Wake up threads sleeping on the specified address.
1314 kern_umtx_wake(struct thread *td, void *uaddr, int n_wake, int is_private)
1316 struct umtx_key key;
1319 if ((ret = umtx_key_get(uaddr, TYPE_SIMPLE_WAIT,
1320 is_private ? THREAD_SHARE : AUTO_SHARE, &key)) != 0)
1323 umtxq_signal(&key, n_wake);
1325 umtx_key_release(&key);
1330 * Lock PTHREAD_PRIO_NONE protocol POSIX mutex.
1333 do_lock_normal(struct thread *td, struct umutex *m, uint32_t flags,
1334 struct _umtx_time *timeout, int mode)
1336 struct umtx_abs_timeout timo;
1338 uint32_t owner, old, id;
1344 if (timeout != NULL)
1345 umtx_abs_timeout_init2(&timo, timeout);
1348 * Care must be exercised when dealing with umtx structure. It
1349 * can fault on any access.
1352 rv = fueword32(&m->m_owner, &owner);
1355 if (mode == _UMUTEX_WAIT) {
1356 if (owner == UMUTEX_UNOWNED ||
1357 owner == UMUTEX_CONTESTED ||
1358 owner == UMUTEX_RB_OWNERDEAD ||
1359 owner == UMUTEX_RB_NOTRECOV)
1363 * Robust mutex terminated. Kernel duty is to
1364 * return EOWNERDEAD to the userspace. The
1365 * umutex.m_flags UMUTEX_NONCONSISTENT is set
1366 * by the common userspace code.
1368 if (owner == UMUTEX_RB_OWNERDEAD) {
1369 rv = casueword32(&m->m_owner,
1370 UMUTEX_RB_OWNERDEAD, &owner,
1371 id | UMUTEX_CONTESTED);
1375 MPASS(owner == UMUTEX_RB_OWNERDEAD);
1376 return (EOWNERDEAD); /* success */
1379 rv = thread_check_susp(td, false);
1384 if (owner == UMUTEX_RB_NOTRECOV)
1385 return (ENOTRECOVERABLE);
1388 * Try the uncontested case. This should be
1391 rv = casueword32(&m->m_owner, UMUTEX_UNOWNED,
1393 /* The address was invalid. */
1397 /* The acquire succeeded. */
1399 MPASS(owner == UMUTEX_UNOWNED);
1404 * If no one owns it but it is contested try
1408 if (owner == UMUTEX_CONTESTED) {
1409 rv = casueword32(&m->m_owner,
1410 UMUTEX_CONTESTED, &owner,
1411 id | UMUTEX_CONTESTED);
1412 /* The address was invalid. */
1416 MPASS(owner == UMUTEX_CONTESTED);
1420 rv = thread_check_susp(td, false);
1426 * If this failed the lock has
1432 /* rv == 1 but not contested, likely store failure */
1433 rv = thread_check_susp(td, false);
1438 if (mode == _UMUTEX_TRY)
1442 * If we caught a signal, we have retried and now
1448 if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX,
1449 GET_SHARE(flags), &uq->uq_key)) != 0)
1452 umtxq_lock(&uq->uq_key);
1453 umtxq_busy(&uq->uq_key);
1455 umtxq_unlock(&uq->uq_key);
1458 * Set the contested bit so that a release in user space
1459 * knows to use the system call for unlock. If this fails
1460 * either some one else has acquired the lock or it has been
1463 rv = casueword32(&m->m_owner, owner, &old,
1464 owner | UMUTEX_CONTESTED);
1466 /* The address was invalid or casueword failed to store. */
1467 if (rv == -1 || rv == 1) {
1468 umtxq_lock(&uq->uq_key);
1470 umtxq_unbusy(&uq->uq_key);
1471 umtxq_unlock(&uq->uq_key);
1472 umtx_key_release(&uq->uq_key);
1476 rv = thread_check_susp(td, false);
1484 * We set the contested bit, sleep. Otherwise the lock changed
1485 * and we need to retry or we lost a race to the thread
1486 * unlocking the umtx.
1488 umtxq_lock(&uq->uq_key);
1489 umtxq_unbusy(&uq->uq_key);
1490 MPASS(old == owner);
1491 error = umtxq_sleep(uq, "umtxn", timeout == NULL ?
1494 umtxq_unlock(&uq->uq_key);
1495 umtx_key_release(&uq->uq_key);
1498 error = thread_check_susp(td, false);
1505 * Unlock PTHREAD_PRIO_NONE protocol POSIX mutex.
1508 do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1510 struct umtx_key key;
1511 uint32_t owner, old, id, newlock;
1518 * Make sure we own this mtx.
1520 error = fueword32(&m->m_owner, &owner);
1524 if ((owner & ~UMUTEX_CONTESTED) != id)
1527 newlock = umtx_unlock_val(flags, rb);
1528 if ((owner & UMUTEX_CONTESTED) == 0) {
1529 error = casueword32(&m->m_owner, owner, &old, newlock);
1533 error = thread_check_susp(td, false);
1538 MPASS(old == owner);
1542 /* We should only ever be in here for contested locks */
1543 if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1549 count = umtxq_count(&key);
1553 * When unlocking the umtx, it must be marked as unowned if
1554 * there is zero or one thread only waiting for it.
1555 * Otherwise, it must be marked as contested.
1558 newlock |= UMUTEX_CONTESTED;
1559 error = casueword32(&m->m_owner, owner, &old, newlock);
1561 umtxq_signal(&key, 1);
1564 umtx_key_release(&key);
1570 error = thread_check_susp(td, false);
1579 * Check if the mutex is available and wake up a waiter,
1580 * only for simple mutex.
1583 do_wake_umutex(struct thread *td, struct umutex *m)
1585 struct umtx_key key;
1592 error = fueword32(&m->m_owner, &owner);
1596 if ((owner & ~UMUTEX_CONTESTED) != 0 && owner != UMUTEX_RB_OWNERDEAD &&
1597 owner != UMUTEX_RB_NOTRECOV)
1600 error = fueword32(&m->m_flags, &flags);
1604 /* We should only ever be in here for contested locks */
1605 if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1611 count = umtxq_count(&key);
1614 if (count <= 1 && owner != UMUTEX_RB_OWNERDEAD &&
1615 owner != UMUTEX_RB_NOTRECOV) {
1616 error = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
1620 } else if (error == 1) {
1624 umtx_key_release(&key);
1625 error = thread_check_susp(td, false);
1633 if (error == 0 && count != 0) {
1634 MPASS((owner & ~UMUTEX_CONTESTED) == 0 ||
1635 owner == UMUTEX_RB_OWNERDEAD ||
1636 owner == UMUTEX_RB_NOTRECOV);
1637 umtxq_signal(&key, 1);
1641 umtx_key_release(&key);
1646 * Check if the mutex has waiters and tries to fix contention bit.
1649 do_wake2_umutex(struct thread *td, struct umutex *m, uint32_t flags)
1651 struct umtx_key key;
1652 uint32_t owner, old;
1657 switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT |
1661 type = TYPE_NORMAL_UMUTEX;
1663 case UMUTEX_PRIO_INHERIT:
1664 type = TYPE_PI_UMUTEX;
1666 case (UMUTEX_PRIO_INHERIT | UMUTEX_ROBUST):
1667 type = TYPE_PI_ROBUST_UMUTEX;
1669 case UMUTEX_PRIO_PROTECT:
1670 type = TYPE_PP_UMUTEX;
1672 case (UMUTEX_PRIO_PROTECT | UMUTEX_ROBUST):
1673 type = TYPE_PP_ROBUST_UMUTEX;
1678 if ((error = umtx_key_get(m, type, GET_SHARE(flags), &key)) != 0)
1684 count = umtxq_count(&key);
1687 error = fueword32(&m->m_owner, &owner);
1692 * Only repair contention bit if there is a waiter, this means
1693 * the mutex is still being referenced by userland code,
1694 * otherwise don't update any memory.
1696 while (error == 0 && (owner & UMUTEX_CONTESTED) == 0 &&
1697 (count > 1 || (count == 1 && (owner & ~UMUTEX_CONTESTED) != 0))) {
1698 error = casueword32(&m->m_owner, owner, &old,
1699 owner | UMUTEX_CONTESTED);
1705 MPASS(old == owner);
1709 error = thread_check_susp(td, false);
1713 if (error == EFAULT) {
1714 umtxq_signal(&key, INT_MAX);
1715 } else if (count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1716 owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1717 umtxq_signal(&key, 1);
1720 umtx_key_release(&key);
1725 umtx_pi_alloc(int flags)
1729 pi = uma_zalloc(umtx_pi_zone, M_ZERO | flags);
1730 TAILQ_INIT(&pi->pi_blocked);
1731 atomic_add_int(&umtx_pi_allocated, 1);
1736 umtx_pi_free(struct umtx_pi *pi)
1738 uma_zfree(umtx_pi_zone, pi);
1739 atomic_add_int(&umtx_pi_allocated, -1);
1743 * Adjust the thread's position on a pi_state after its priority has been
1747 umtx_pi_adjust_thread(struct umtx_pi *pi, struct thread *td)
1749 struct umtx_q *uq, *uq1, *uq2;
1752 mtx_assert(&umtx_lock, MA_OWNED);
1759 * Check if the thread needs to be moved on the blocked chain.
1760 * It needs to be moved if either its priority is lower than
1761 * the previous thread or higher than the next thread.
1763 uq1 = TAILQ_PREV(uq, umtxq_head, uq_lockq);
1764 uq2 = TAILQ_NEXT(uq, uq_lockq);
1765 if ((uq1 != NULL && UPRI(td) < UPRI(uq1->uq_thread)) ||
1766 (uq2 != NULL && UPRI(td) > UPRI(uq2->uq_thread))) {
1768 * Remove thread from blocked chain and determine where
1769 * it should be moved to.
1771 TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1772 TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1773 td1 = uq1->uq_thread;
1774 MPASS(td1->td_proc->p_magic == P_MAGIC);
1775 if (UPRI(td1) > UPRI(td))
1780 TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1782 TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1787 static struct umtx_pi *
1788 umtx_pi_next(struct umtx_pi *pi)
1790 struct umtx_q *uq_owner;
1792 if (pi->pi_owner == NULL)
1794 uq_owner = pi->pi_owner->td_umtxq;
1795 if (uq_owner == NULL)
1797 return (uq_owner->uq_pi_blocked);
1801 * Floyd's Cycle-Finding Algorithm.
1804 umtx_pi_check_loop(struct umtx_pi *pi)
1806 struct umtx_pi *pi1; /* fast iterator */
1808 mtx_assert(&umtx_lock, MA_OWNED);
1813 pi = umtx_pi_next(pi);
1816 pi1 = umtx_pi_next(pi1);
1819 pi1 = umtx_pi_next(pi1);
1829 * Propagate priority when a thread is blocked on POSIX
1833 umtx_propagate_priority(struct thread *td)
1839 mtx_assert(&umtx_lock, MA_OWNED);
1842 pi = uq->uq_pi_blocked;
1845 if (umtx_pi_check_loop(pi))
1850 if (td == NULL || td == curthread)
1853 MPASS(td->td_proc != NULL);
1854 MPASS(td->td_proc->p_magic == P_MAGIC);
1857 if (td->td_lend_user_pri > pri)
1858 sched_lend_user_prio(td, pri);
1866 * Pick up the lock that td is blocked on.
1869 pi = uq->uq_pi_blocked;
1872 /* Resort td on the list if needed. */
1873 umtx_pi_adjust_thread(pi, td);
1878 * Unpropagate priority for a PI mutex when a thread blocked on
1879 * it is interrupted by signal or resumed by others.
1882 umtx_repropagate_priority(struct umtx_pi *pi)
1884 struct umtx_q *uq, *uq_owner;
1885 struct umtx_pi *pi2;
1888 mtx_assert(&umtx_lock, MA_OWNED);
1890 if (umtx_pi_check_loop(pi))
1892 while (pi != NULL && pi->pi_owner != NULL) {
1894 uq_owner = pi->pi_owner->td_umtxq;
1896 TAILQ_FOREACH(pi2, &uq_owner->uq_pi_contested, pi_link) {
1897 uq = TAILQ_FIRST(&pi2->pi_blocked);
1899 if (pri > UPRI(uq->uq_thread))
1900 pri = UPRI(uq->uq_thread);
1904 if (pri > uq_owner->uq_inherited_pri)
1905 pri = uq_owner->uq_inherited_pri;
1906 thread_lock(pi->pi_owner);
1907 sched_lend_user_prio(pi->pi_owner, pri);
1908 thread_unlock(pi->pi_owner);
1909 if ((pi = uq_owner->uq_pi_blocked) != NULL)
1910 umtx_pi_adjust_thread(pi, uq_owner->uq_thread);
1915 * Insert a PI mutex into owned list.
1918 umtx_pi_setowner(struct umtx_pi *pi, struct thread *owner)
1920 struct umtx_q *uq_owner;
1922 uq_owner = owner->td_umtxq;
1923 mtx_assert(&umtx_lock, MA_OWNED);
1924 MPASS(pi->pi_owner == NULL);
1925 pi->pi_owner = owner;
1926 TAILQ_INSERT_TAIL(&uq_owner->uq_pi_contested, pi, pi_link);
1930 * Disown a PI mutex, and remove it from the owned list.
1933 umtx_pi_disown(struct umtx_pi *pi)
1936 mtx_assert(&umtx_lock, MA_OWNED);
1937 TAILQ_REMOVE(&pi->pi_owner->td_umtxq->uq_pi_contested, pi, pi_link);
1938 pi->pi_owner = NULL;
1942 * Claim ownership of a PI mutex.
1945 umtx_pi_claim(struct umtx_pi *pi, struct thread *owner)
1950 mtx_lock(&umtx_lock);
1951 if (pi->pi_owner == owner) {
1952 mtx_unlock(&umtx_lock);
1956 if (pi->pi_owner != NULL) {
1958 * userland may have already messed the mutex, sigh.
1960 mtx_unlock(&umtx_lock);
1963 umtx_pi_setowner(pi, owner);
1964 uq = TAILQ_FIRST(&pi->pi_blocked);
1966 pri = UPRI(uq->uq_thread);
1968 if (pri < UPRI(owner))
1969 sched_lend_user_prio(owner, pri);
1970 thread_unlock(owner);
1972 mtx_unlock(&umtx_lock);
1977 * Adjust a thread's order position in its blocked PI mutex,
1978 * this may result new priority propagating process.
1981 umtx_pi_adjust(struct thread *td, u_char oldpri)
1987 mtx_lock(&umtx_lock);
1989 * Pick up the lock that td is blocked on.
1991 pi = uq->uq_pi_blocked;
1993 umtx_pi_adjust_thread(pi, td);
1994 umtx_repropagate_priority(pi);
1996 mtx_unlock(&umtx_lock);
2000 * Sleep on a PI mutex.
2003 umtxq_sleep_pi(struct umtx_q *uq, struct umtx_pi *pi, uint32_t owner,
2004 const char *wmesg, struct umtx_abs_timeout *timo, bool shared)
2006 struct thread *td, *td1;
2010 struct umtxq_chain *uc;
2012 uc = umtxq_getchain(&pi->pi_key);
2016 KASSERT(td == curthread, ("inconsistent uq_thread"));
2017 UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
2018 KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));
2020 mtx_lock(&umtx_lock);
2021 if (pi->pi_owner == NULL) {
2022 mtx_unlock(&umtx_lock);
2023 td1 = tdfind(owner, shared ? -1 : td->td_proc->p_pid);
2024 mtx_lock(&umtx_lock);
2026 if (pi->pi_owner == NULL)
2027 umtx_pi_setowner(pi, td1);
2028 PROC_UNLOCK(td1->td_proc);
2032 TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
2033 pri = UPRI(uq1->uq_thread);
2039 TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
2041 TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
2043 uq->uq_pi_blocked = pi;
2045 td->td_flags |= TDF_UPIBLOCKED;
2047 umtx_propagate_priority(td);
2048 mtx_unlock(&umtx_lock);
2049 umtxq_unbusy(&uq->uq_key);
2051 error = umtxq_sleep(uq, wmesg, timo);
2054 mtx_lock(&umtx_lock);
2055 uq->uq_pi_blocked = NULL;
2057 td->td_flags &= ~TDF_UPIBLOCKED;
2059 TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
2060 umtx_repropagate_priority(pi);
2061 mtx_unlock(&umtx_lock);
2062 umtxq_unlock(&uq->uq_key);
2068 * Add reference count for a PI mutex.
2071 umtx_pi_ref(struct umtx_pi *pi)
2074 UMTXQ_LOCKED_ASSERT(umtxq_getchain(&pi->pi_key));
2079 * Decrease reference count for a PI mutex, if the counter
2080 * is decreased to zero, its memory space is freed.
2083 umtx_pi_unref(struct umtx_pi *pi)
2085 struct umtxq_chain *uc;
2087 uc = umtxq_getchain(&pi->pi_key);
2088 UMTXQ_LOCKED_ASSERT(uc);
2089 KASSERT(pi->pi_refcount > 0, ("invalid reference count"));
2090 if (--pi->pi_refcount == 0) {
2091 mtx_lock(&umtx_lock);
2092 if (pi->pi_owner != NULL)
2094 KASSERT(TAILQ_EMPTY(&pi->pi_blocked),
2095 ("blocked queue not empty"));
2096 mtx_unlock(&umtx_lock);
2097 TAILQ_REMOVE(&uc->uc_pi_list, pi, pi_hashlink);
2103 * Find a PI mutex in hash table.
2106 umtx_pi_lookup(struct umtx_key *key)
2108 struct umtxq_chain *uc;
2111 uc = umtxq_getchain(key);
2112 UMTXQ_LOCKED_ASSERT(uc);
2114 TAILQ_FOREACH(pi, &uc->uc_pi_list, pi_hashlink) {
2115 if (umtx_key_match(&pi->pi_key, key)) {
2123 * Insert a PI mutex into hash table.
2126 umtx_pi_insert(struct umtx_pi *pi)
2128 struct umtxq_chain *uc;
2130 uc = umtxq_getchain(&pi->pi_key);
2131 UMTXQ_LOCKED_ASSERT(uc);
2132 TAILQ_INSERT_TAIL(&uc->uc_pi_list, pi, pi_hashlink);
2136 * Drop a PI mutex and wakeup a top waiter.
2139 umtx_pi_drop(struct thread *td, struct umtx_key *key, bool rb, int *count)
2141 struct umtx_q *uq_first, *uq_first2, *uq_me;
2142 struct umtx_pi *pi, *pi2;
2145 UMTXQ_ASSERT_LOCKED_BUSY(key);
2146 *count = umtxq_count_pi(key, &uq_first);
2147 if (uq_first != NULL) {
2148 mtx_lock(&umtx_lock);
2149 pi = uq_first->uq_pi_blocked;
2150 KASSERT(pi != NULL, ("pi == NULL?"));
2151 if (pi->pi_owner != td && !(rb && pi->pi_owner == NULL)) {
2152 mtx_unlock(&umtx_lock);
2153 /* userland messed the mutex */
2156 uq_me = td->td_umtxq;
2157 if (pi->pi_owner == td)
2159 /* get highest priority thread which is still sleeping. */
2160 uq_first = TAILQ_FIRST(&pi->pi_blocked);
2161 while (uq_first != NULL &&
2162 (uq_first->uq_flags & UQF_UMTXQ) == 0) {
2163 uq_first = TAILQ_NEXT(uq_first, uq_lockq);
2166 TAILQ_FOREACH(pi2, &uq_me->uq_pi_contested, pi_link) {
2167 uq_first2 = TAILQ_FIRST(&pi2->pi_blocked);
2168 if (uq_first2 != NULL) {
2169 if (pri > UPRI(uq_first2->uq_thread))
2170 pri = UPRI(uq_first2->uq_thread);
2174 sched_lend_user_prio(td, pri);
2176 mtx_unlock(&umtx_lock);
2178 umtxq_signal_thread(uq_first);
2180 pi = umtx_pi_lookup(key);
2182 * A umtx_pi can exist if a signal or timeout removed the
2183 * last waiter from the umtxq, but there is still
2184 * a thread in do_lock_pi() holding the umtx_pi.
2188 * The umtx_pi can be unowned, such as when a thread
2189 * has just entered do_lock_pi(), allocated the
2190 * umtx_pi, and unlocked the umtxq.
2191 * If the current thread owns it, it must disown it.
2193 mtx_lock(&umtx_lock);
2194 if (pi->pi_owner == td)
2196 mtx_unlock(&umtx_lock);
2206 do_lock_pi(struct thread *td, struct umutex *m, uint32_t flags,
2207 struct _umtx_time *timeout, int try)
2209 struct umtx_abs_timeout timo;
2211 struct umtx_pi *pi, *new_pi;
2212 uint32_t id, old_owner, owner, old;
2218 if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2219 TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2223 if (timeout != NULL)
2224 umtx_abs_timeout_init2(&timo, timeout);
2226 umtxq_lock(&uq->uq_key);
2227 pi = umtx_pi_lookup(&uq->uq_key);
2229 new_pi = umtx_pi_alloc(M_NOWAIT);
2230 if (new_pi == NULL) {
2231 umtxq_unlock(&uq->uq_key);
2232 new_pi = umtx_pi_alloc(M_WAITOK);
2233 umtxq_lock(&uq->uq_key);
2234 pi = umtx_pi_lookup(&uq->uq_key);
2236 umtx_pi_free(new_pi);
2240 if (new_pi != NULL) {
2241 new_pi->pi_key = uq->uq_key;
2242 umtx_pi_insert(new_pi);
2247 umtxq_unlock(&uq->uq_key);
2250 * Care must be exercised when dealing with umtx structure. It
2251 * can fault on any access.
2255 * Try the uncontested case. This should be done in userland.
2257 rv = casueword32(&m->m_owner, UMUTEX_UNOWNED, &owner, id);
2258 /* The address was invalid. */
2263 /* The acquire succeeded. */
2265 MPASS(owner == UMUTEX_UNOWNED);
2270 if (owner == UMUTEX_RB_NOTRECOV) {
2271 error = ENOTRECOVERABLE;
2276 * Nobody owns it, but the acquire failed. This can happen
2277 * with ll/sc atomics.
2279 if (owner == UMUTEX_UNOWNED) {
2280 error = thread_check_susp(td, true);
2287 * Avoid overwriting a possible error from sleep due
2288 * to the pending signal with suspension check result.
2291 error = thread_check_susp(td, true);
2296 /* If no one owns it but it is contested try to acquire it. */
2297 if (owner == UMUTEX_CONTESTED || owner == UMUTEX_RB_OWNERDEAD) {
2299 rv = casueword32(&m->m_owner, owner, &owner,
2300 id | UMUTEX_CONTESTED);
2301 /* The address was invalid. */
2308 error = thread_check_susp(td, true);
2314 * If this failed the lock could
2321 MPASS(owner == old_owner);
2322 umtxq_lock(&uq->uq_key);
2323 umtxq_busy(&uq->uq_key);
2324 error = umtx_pi_claim(pi, td);
2325 umtxq_unbusy(&uq->uq_key);
2326 umtxq_unlock(&uq->uq_key);
2329 * Since we're going to return an
2330 * error, restore the m_owner to its
2331 * previous, unowned state to avoid
2332 * compounding the problem.
2334 (void)casuword32(&m->m_owner,
2335 id | UMUTEX_CONTESTED, old_owner);
2337 if (error == 0 && old_owner == UMUTEX_RB_OWNERDEAD)
2342 if ((owner & ~UMUTEX_CONTESTED) == id) {
2353 * If we caught a signal, we have retried and now
2359 umtxq_lock(&uq->uq_key);
2360 umtxq_busy(&uq->uq_key);
2361 umtxq_unlock(&uq->uq_key);
2364 * Set the contested bit so that a release in user space
2365 * knows to use the system call for unlock. If this fails
2366 * either some one else has acquired the lock or it has been
2369 rv = casueword32(&m->m_owner, owner, &old, owner |
2372 /* The address was invalid. */
2374 umtxq_unbusy_unlocked(&uq->uq_key);
2379 umtxq_unbusy_unlocked(&uq->uq_key);
2380 error = thread_check_susp(td, true);
2385 * The lock changed and we need to retry or we
2386 * lost a race to the thread unlocking the
2387 * umtx. Note that the UMUTEX_RB_OWNERDEAD
2388 * value for owner is impossible there.
2393 umtxq_lock(&uq->uq_key);
2395 /* We set the contested bit, sleep. */
2396 MPASS(old == owner);
2397 error = umtxq_sleep_pi(uq, pi, owner & ~UMUTEX_CONTESTED,
2398 "umtxpi", timeout == NULL ? NULL : &timo,
2399 (flags & USYNC_PROCESS_SHARED) != 0);
2403 error = thread_check_susp(td, false);
2408 umtxq_lock(&uq->uq_key);
2410 umtxq_unlock(&uq->uq_key);
2412 umtx_key_release(&uq->uq_key);
2417 * Unlock a PI mutex.
2420 do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2422 struct umtx_key key;
2423 uint32_t id, new_owner, old, owner;
2430 * Make sure we own this mtx.
2432 error = fueword32(&m->m_owner, &owner);
2436 if ((owner & ~UMUTEX_CONTESTED) != id)
2439 new_owner = umtx_unlock_val(flags, rb);
2441 /* This should be done in userland */
2442 if ((owner & UMUTEX_CONTESTED) == 0) {
2443 error = casueword32(&m->m_owner, owner, &old, new_owner);
2447 error = thread_check_susp(td, true);
2457 /* We should only ever be in here for contested locks */
2458 if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2459 TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2465 error = umtx_pi_drop(td, &key, rb, &count);
2469 umtx_key_release(&key);
2470 /* userland messed the mutex */
2476 * When unlocking the umtx, it must be marked as unowned if
2477 * there is zero or one thread only waiting for it.
2478 * Otherwise, it must be marked as contested.
2482 new_owner |= UMUTEX_CONTESTED;
2484 error = casueword32(&m->m_owner, owner, &old, new_owner);
2486 error = thread_check_susp(td, false);
2490 umtxq_unbusy_unlocked(&key);
2491 umtx_key_release(&key);
2494 if (error == 0 && old != owner)
2503 do_lock_pp(struct thread *td, struct umutex *m, uint32_t flags,
2504 struct _umtx_time *timeout, int try)
2506 struct umtx_abs_timeout timo;
2507 struct umtx_q *uq, *uq2;
2511 int error, pri, old_inherited_pri, su, rv;
2515 if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2516 TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2520 if (timeout != NULL)
2521 umtx_abs_timeout_init2(&timo, timeout);
2523 su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2525 old_inherited_pri = uq->uq_inherited_pri;
2526 umtxq_lock(&uq->uq_key);
2527 umtxq_busy(&uq->uq_key);
2528 umtxq_unlock(&uq->uq_key);
2530 rv = fueword32(&m->m_ceilings[0], &ceiling);
2535 ceiling = RTP_PRIO_MAX - ceiling;
2536 if (ceiling > RTP_PRIO_MAX) {
2541 mtx_lock(&umtx_lock);
2542 if (UPRI(td) < PRI_MIN_REALTIME + ceiling) {
2543 mtx_unlock(&umtx_lock);
2547 if (su && PRI_MIN_REALTIME + ceiling < uq->uq_inherited_pri) {
2548 uq->uq_inherited_pri = PRI_MIN_REALTIME + ceiling;
2550 if (uq->uq_inherited_pri < UPRI(td))
2551 sched_lend_user_prio(td, uq->uq_inherited_pri);
2554 mtx_unlock(&umtx_lock);
2556 rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2557 id | UMUTEX_CONTESTED);
2558 /* The address was invalid. */
2564 MPASS(owner == UMUTEX_CONTESTED);
2569 if (owner == UMUTEX_RB_OWNERDEAD) {
2570 rv = casueword32(&m->m_owner, UMUTEX_RB_OWNERDEAD,
2571 &owner, id | UMUTEX_CONTESTED);
2577 MPASS(owner == UMUTEX_RB_OWNERDEAD);
2578 error = EOWNERDEAD; /* success */
2583 * rv == 1, only check for suspension if we
2584 * did not already catched a signal. If we
2585 * get an error from the check, the same
2586 * condition is checked by the umtxq_sleep()
2587 * call below, so we should obliterate the
2588 * error to not skip the last loop iteration.
2591 error = thread_check_susp(td, false);
2600 } else if (owner == UMUTEX_RB_NOTRECOV) {
2601 error = ENOTRECOVERABLE;
2608 * If we caught a signal, we have retried and now
2614 umtxq_lock(&uq->uq_key);
2616 umtxq_unbusy(&uq->uq_key);
2617 error = umtxq_sleep(uq, "umtxpp", timeout == NULL ?
2620 umtxq_unlock(&uq->uq_key);
2622 mtx_lock(&umtx_lock);
2623 uq->uq_inherited_pri = old_inherited_pri;
2625 TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2626 uq2 = TAILQ_FIRST(&pi->pi_blocked);
2628 if (pri > UPRI(uq2->uq_thread))
2629 pri = UPRI(uq2->uq_thread);
2632 if (pri > uq->uq_inherited_pri)
2633 pri = uq->uq_inherited_pri;
2635 sched_lend_user_prio(td, pri);
2637 mtx_unlock(&umtx_lock);
2640 if (error != 0 && error != EOWNERDEAD) {
2641 mtx_lock(&umtx_lock);
2642 uq->uq_inherited_pri = old_inherited_pri;
2644 TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2645 uq2 = TAILQ_FIRST(&pi->pi_blocked);
2647 if (pri > UPRI(uq2->uq_thread))
2648 pri = UPRI(uq2->uq_thread);
2651 if (pri > uq->uq_inherited_pri)
2652 pri = uq->uq_inherited_pri;
2654 sched_lend_user_prio(td, pri);
2656 mtx_unlock(&umtx_lock);
2660 umtxq_unbusy_unlocked(&uq->uq_key);
2661 umtx_key_release(&uq->uq_key);
2666 * Unlock a PP mutex.
2669 do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2671 struct umtx_key key;
2672 struct umtx_q *uq, *uq2;
2674 uint32_t id, owner, rceiling;
2675 int error, pri, new_inherited_pri, su;
2679 su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2682 * Make sure we own this mtx.
2684 error = fueword32(&m->m_owner, &owner);
2688 if ((owner & ~UMUTEX_CONTESTED) != id)
2691 error = copyin(&m->m_ceilings[1], &rceiling, sizeof(uint32_t));
2696 new_inherited_pri = PRI_MAX;
2698 rceiling = RTP_PRIO_MAX - rceiling;
2699 if (rceiling > RTP_PRIO_MAX)
2701 new_inherited_pri = PRI_MIN_REALTIME + rceiling;
2704 if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2705 TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2712 * For priority protected mutex, always set unlocked state
2713 * to UMUTEX_CONTESTED, so that userland always enters kernel
2714 * to lock the mutex, it is necessary because thread priority
2715 * has to be adjusted for such mutex.
2717 error = suword32(&m->m_owner, umtx_unlock_val(flags, rb) |
2722 umtxq_signal(&key, 1);
2729 mtx_lock(&umtx_lock);
2731 uq->uq_inherited_pri = new_inherited_pri;
2733 TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2734 uq2 = TAILQ_FIRST(&pi->pi_blocked);
2736 if (pri > UPRI(uq2->uq_thread))
2737 pri = UPRI(uq2->uq_thread);
2740 if (pri > uq->uq_inherited_pri)
2741 pri = uq->uq_inherited_pri;
2743 sched_lend_user_prio(td, pri);
2745 mtx_unlock(&umtx_lock);
2747 umtx_key_release(&key);
2752 do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
2753 uint32_t *old_ceiling)
2756 uint32_t flags, id, owner, save_ceiling;
2759 error = fueword32(&m->m_flags, &flags);
2762 if ((flags & UMUTEX_PRIO_PROTECT) == 0)
2764 if (ceiling > RTP_PRIO_MAX)
2768 if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2769 TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2773 umtxq_lock(&uq->uq_key);
2774 umtxq_busy(&uq->uq_key);
2775 umtxq_unlock(&uq->uq_key);
2777 rv = fueword32(&m->m_ceilings[0], &save_ceiling);
2783 rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2784 id | UMUTEX_CONTESTED);
2791 MPASS(owner == UMUTEX_CONTESTED);
2792 rv = suword32(&m->m_ceilings[0], ceiling);
2793 rv1 = suword32(&m->m_owner, UMUTEX_CONTESTED);
2794 error = (rv == 0 && rv1 == 0) ? 0: EFAULT;
2798 if ((owner & ~UMUTEX_CONTESTED) == id) {
2799 rv = suword32(&m->m_ceilings[0], ceiling);
2800 error = rv == 0 ? 0 : EFAULT;
2804 if (owner == UMUTEX_RB_OWNERDEAD) {
2807 } else if (owner == UMUTEX_RB_NOTRECOV) {
2808 error = ENOTRECOVERABLE;
2813 * If we caught a signal, we have retried and now
2820 * We set the contested bit, sleep. Otherwise the lock changed
2821 * and we need to retry or we lost a race to the thread
2822 * unlocking the umtx.
2824 umtxq_lock(&uq->uq_key);
2826 umtxq_unbusy(&uq->uq_key);
2827 error = umtxq_sleep(uq, "umtxpp", NULL);
2829 umtxq_unlock(&uq->uq_key);
2831 umtxq_lock(&uq->uq_key);
2833 umtxq_signal(&uq->uq_key, INT_MAX);
2834 umtxq_unbusy(&uq->uq_key);
2835 umtxq_unlock(&uq->uq_key);
2836 umtx_key_release(&uq->uq_key);
2837 if (error == 0 && old_ceiling != NULL) {
2838 rv = suword32(old_ceiling, save_ceiling);
2839 error = rv == 0 ? 0 : EFAULT;
2845 * Lock a userland POSIX mutex.
2848 do_lock_umutex(struct thread *td, struct umutex *m,
2849 struct _umtx_time *timeout, int mode)
2854 error = fueword32(&m->m_flags, &flags);
2858 switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2860 error = do_lock_normal(td, m, flags, timeout, mode);
2862 case UMUTEX_PRIO_INHERIT:
2863 error = do_lock_pi(td, m, flags, timeout, mode);
2865 case UMUTEX_PRIO_PROTECT:
2866 error = do_lock_pp(td, m, flags, timeout, mode);
2871 if (timeout == NULL) {
2872 if (error == EINTR && mode != _UMUTEX_WAIT)
2875 /* Timed-locking is not restarted. */
2876 if (error == ERESTART)
2883 * Unlock a userland POSIX mutex.
2886 do_unlock_umutex(struct thread *td, struct umutex *m, bool rb)
2891 error = fueword32(&m->m_flags, &flags);
2895 switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2897 return (do_unlock_normal(td, m, flags, rb));
2898 case UMUTEX_PRIO_INHERIT:
2899 return (do_unlock_pi(td, m, flags, rb));
2900 case UMUTEX_PRIO_PROTECT:
2901 return (do_unlock_pp(td, m, flags, rb));
2908 do_cv_wait(struct thread *td, struct ucond *cv, struct umutex *m,
2909 struct timespec *timeout, u_long wflags)
2911 struct umtx_abs_timeout timo;
2913 uint32_t flags, clockid, hasw;
2917 error = fueword32(&cv->c_flags, &flags);
2920 error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &uq->uq_key);
2924 if ((wflags & CVWAIT_CLOCKID) != 0) {
2925 error = fueword32(&cv->c_clockid, &clockid);
2927 umtx_key_release(&uq->uq_key);
2930 if (clockid < CLOCK_REALTIME ||
2931 clockid >= CLOCK_THREAD_CPUTIME_ID) {
2932 /* hmm, only HW clock id will work. */
2933 umtx_key_release(&uq->uq_key);
2937 clockid = CLOCK_REALTIME;
2940 umtxq_lock(&uq->uq_key);
2941 umtxq_busy(&uq->uq_key);
2943 umtxq_unlock(&uq->uq_key);
2946 * Set c_has_waiters to 1 before releasing user mutex, also
2947 * don't modify cache line when unnecessary.
2949 error = fueword32(&cv->c_has_waiters, &hasw);
2950 if (error == 0 && hasw == 0)
2951 suword32(&cv->c_has_waiters, 1);
2953 umtxq_unbusy_unlocked(&uq->uq_key);
2955 error = do_unlock_umutex(td, m, false);
2957 if (timeout != NULL)
2958 umtx_abs_timeout_init(&timo, clockid,
2959 (wflags & CVWAIT_ABSTIME) != 0, timeout);
2961 umtxq_lock(&uq->uq_key);
2963 error = umtxq_sleep(uq, "ucond", timeout == NULL ?
2967 if ((uq->uq_flags & UQF_UMTXQ) == 0)
2971 * This must be timeout,interrupted by signal or
2972 * surprious wakeup, clear c_has_waiter flag when
2975 umtxq_busy(&uq->uq_key);
2976 if ((uq->uq_flags & UQF_UMTXQ) != 0) {
2977 int oldlen = uq->uq_cur_queue->length;
2980 umtxq_unlock(&uq->uq_key);
2981 suword32(&cv->c_has_waiters, 0);
2982 umtxq_lock(&uq->uq_key);
2985 umtxq_unbusy(&uq->uq_key);
2986 if (error == ERESTART)
2990 umtxq_unlock(&uq->uq_key);
2991 umtx_key_release(&uq->uq_key);
2996 * Signal a userland condition variable.
2999 do_cv_signal(struct thread *td, struct ucond *cv)
3001 struct umtx_key key;
3002 int error, cnt, nwake;
3005 error = fueword32(&cv->c_flags, &flags);
3008 if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
3012 cnt = umtxq_count(&key);
3013 nwake = umtxq_signal(&key, 1);
3016 error = suword32(&cv->c_has_waiters, 0);
3023 umtx_key_release(&key);
3028 do_cv_broadcast(struct thread *td, struct ucond *cv)
3030 struct umtx_key key;
3034 error = fueword32(&cv->c_flags, &flags);
3037 if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
3042 umtxq_signal(&key, INT_MAX);
3045 error = suword32(&cv->c_has_waiters, 0);
3049 umtxq_unbusy_unlocked(&key);
3051 umtx_key_release(&key);
3056 do_rw_rdlock(struct thread *td, struct urwlock *rwlock, long fflag,
3057 struct _umtx_time *timeout)
3059 struct umtx_abs_timeout timo;
3061 uint32_t flags, wrflags;
3062 int32_t state, oldstate;
3063 int32_t blocked_readers;
3064 int error, error1, rv;
3067 error = fueword32(&rwlock->rw_flags, &flags);
3070 error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3074 if (timeout != NULL)
3075 umtx_abs_timeout_init2(&timo, timeout);
3077 wrflags = URWLOCK_WRITE_OWNER;
3078 if (!(fflag & URWLOCK_PREFER_READER) && !(flags & URWLOCK_PREFER_READER))
3079 wrflags |= URWLOCK_WRITE_WAITERS;
3082 rv = fueword32(&rwlock->rw_state, &state);
3084 umtx_key_release(&uq->uq_key);
3088 /* try to lock it */
3089 while (!(state & wrflags)) {
3090 if (__predict_false(URWLOCK_READER_COUNT(state) ==
3091 URWLOCK_MAX_READERS)) {
3092 umtx_key_release(&uq->uq_key);
3095 rv = casueword32(&rwlock->rw_state, state,
3096 &oldstate, state + 1);
3098 umtx_key_release(&uq->uq_key);
3102 MPASS(oldstate == state);
3103 umtx_key_release(&uq->uq_key);
3106 error = thread_check_susp(td, true);
3115 /* grab monitor lock */
3116 umtxq_lock(&uq->uq_key);
3117 umtxq_busy(&uq->uq_key);
3118 umtxq_unlock(&uq->uq_key);
3121 * re-read the state, in case it changed between the try-lock above
3122 * and the check below
3124 rv = fueword32(&rwlock->rw_state, &state);
3128 /* set read contention bit */
3129 while (error == 0 && (state & wrflags) &&
3130 !(state & URWLOCK_READ_WAITERS)) {
3131 rv = casueword32(&rwlock->rw_state, state,
3132 &oldstate, state | URWLOCK_READ_WAITERS);
3138 MPASS(oldstate == state);
3142 error = thread_check_susp(td, false);
3147 umtxq_unbusy_unlocked(&uq->uq_key);
3151 /* state is changed while setting flags, restart */
3152 if (!(state & wrflags)) {
3153 umtxq_unbusy_unlocked(&uq->uq_key);
3154 error = thread_check_susp(td, true);
3162 * Contention bit is set, before sleeping, increase
3163 * read waiter count.
3165 rv = fueword32(&rwlock->rw_blocked_readers,
3168 umtxq_unbusy_unlocked(&uq->uq_key);
3172 suword32(&rwlock->rw_blocked_readers, blocked_readers+1);
3174 while (state & wrflags) {
3175 umtxq_lock(&uq->uq_key);
3177 umtxq_unbusy(&uq->uq_key);
3179 error = umtxq_sleep(uq, "urdlck", timeout == NULL ?
3182 umtxq_busy(&uq->uq_key);
3184 umtxq_unlock(&uq->uq_key);
3187 rv = fueword32(&rwlock->rw_state, &state);
3194 /* decrease read waiter count, and may clear read contention bit */
3195 rv = fueword32(&rwlock->rw_blocked_readers,
3198 umtxq_unbusy_unlocked(&uq->uq_key);
3202 suword32(&rwlock->rw_blocked_readers, blocked_readers-1);
3203 if (blocked_readers == 1) {
3204 rv = fueword32(&rwlock->rw_state, &state);
3206 umtxq_unbusy_unlocked(&uq->uq_key);
3211 rv = casueword32(&rwlock->rw_state, state,
3212 &oldstate, state & ~URWLOCK_READ_WAITERS);
3218 MPASS(oldstate == state);
3222 error1 = thread_check_susp(td, false);
3231 umtxq_unbusy_unlocked(&uq->uq_key);
3235 umtx_key_release(&uq->uq_key);
3236 if (error == ERESTART)
3242 do_rw_wrlock(struct thread *td, struct urwlock *rwlock, struct _umtx_time *timeout)
3244 struct umtx_abs_timeout timo;
3247 int32_t state, oldstate;
3248 int32_t blocked_writers;
3249 int32_t blocked_readers;
3250 int error, error1, rv;
3253 error = fueword32(&rwlock->rw_flags, &flags);
3256 error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3260 if (timeout != NULL)
3261 umtx_abs_timeout_init2(&timo, timeout);
3263 blocked_readers = 0;
3265 rv = fueword32(&rwlock->rw_state, &state);
3267 umtx_key_release(&uq->uq_key);
3270 while ((state & URWLOCK_WRITE_OWNER) == 0 &&
3271 URWLOCK_READER_COUNT(state) == 0) {
3272 rv = casueword32(&rwlock->rw_state, state,
3273 &oldstate, state | URWLOCK_WRITE_OWNER);
3275 umtx_key_release(&uq->uq_key);
3279 MPASS(oldstate == state);
3280 umtx_key_release(&uq->uq_key);
3284 error = thread_check_susp(td, true);
3290 if ((state & (URWLOCK_WRITE_OWNER |
3291 URWLOCK_WRITE_WAITERS)) == 0 &&
3292 blocked_readers != 0) {
3293 umtxq_lock(&uq->uq_key);
3294 umtxq_busy(&uq->uq_key);
3295 umtxq_signal_queue(&uq->uq_key, INT_MAX,
3297 umtxq_unbusy(&uq->uq_key);
3298 umtxq_unlock(&uq->uq_key);
3304 /* grab monitor lock */
3305 umtxq_lock(&uq->uq_key);
3306 umtxq_busy(&uq->uq_key);
3307 umtxq_unlock(&uq->uq_key);
3310 * Re-read the state, in case it changed between the
3311 * try-lock above and the check below.
3313 rv = fueword32(&rwlock->rw_state, &state);
3317 while (error == 0 && ((state & URWLOCK_WRITE_OWNER) ||
3318 URWLOCK_READER_COUNT(state) != 0) &&
3319 (state & URWLOCK_WRITE_WAITERS) == 0) {
3320 rv = casueword32(&rwlock->rw_state, state,
3321 &oldstate, state | URWLOCK_WRITE_WAITERS);
3327 MPASS(oldstate == state);
3331 error = thread_check_susp(td, false);
3336 umtxq_unbusy_unlocked(&uq->uq_key);
3340 if ((state & URWLOCK_WRITE_OWNER) == 0 &&
3341 URWLOCK_READER_COUNT(state) == 0) {
3342 umtxq_unbusy_unlocked(&uq->uq_key);
3343 error = thread_check_susp(td, false);
3349 rv = fueword32(&rwlock->rw_blocked_writers,
3352 umtxq_unbusy_unlocked(&uq->uq_key);
3356 suword32(&rwlock->rw_blocked_writers, blocked_writers + 1);
3358 while ((state & URWLOCK_WRITE_OWNER) ||
3359 URWLOCK_READER_COUNT(state) != 0) {
3360 umtxq_lock(&uq->uq_key);
3361 umtxq_insert_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3362 umtxq_unbusy(&uq->uq_key);
3364 error = umtxq_sleep(uq, "uwrlck", timeout == NULL ?
3367 umtxq_busy(&uq->uq_key);
3368 umtxq_remove_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3369 umtxq_unlock(&uq->uq_key);
3372 rv = fueword32(&rwlock->rw_state, &state);
3379 rv = fueword32(&rwlock->rw_blocked_writers,
3382 umtxq_unbusy_unlocked(&uq->uq_key);
3386 suword32(&rwlock->rw_blocked_writers, blocked_writers-1);
3387 if (blocked_writers == 1) {
3388 rv = fueword32(&rwlock->rw_state, &state);
3390 umtxq_unbusy_unlocked(&uq->uq_key);
3395 rv = casueword32(&rwlock->rw_state, state,
3396 &oldstate, state & ~URWLOCK_WRITE_WAITERS);
3402 MPASS(oldstate == state);
3406 error1 = thread_check_susp(td, false);
3408 * We are leaving the URWLOCK_WRITE_WAITERS
3409 * behind, but this should not harm the
3418 rv = fueword32(&rwlock->rw_blocked_readers,
3421 umtxq_unbusy_unlocked(&uq->uq_key);
3426 blocked_readers = 0;
3428 umtxq_unbusy_unlocked(&uq->uq_key);
3431 umtx_key_release(&uq->uq_key);
3432 if (error == ERESTART)
3438 do_rw_unlock(struct thread *td, struct urwlock *rwlock)
3442 int32_t state, oldstate;
3443 int error, rv, q, count;
3446 error = fueword32(&rwlock->rw_flags, &flags);
3449 error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3453 error = fueword32(&rwlock->rw_state, &state);
3458 if (state & URWLOCK_WRITE_OWNER) {
3460 rv = casueword32(&rwlock->rw_state, state,
3461 &oldstate, state & ~URWLOCK_WRITE_OWNER);
3468 if (!(oldstate & URWLOCK_WRITE_OWNER)) {
3472 error = thread_check_susp(td, true);
3478 } else if (URWLOCK_READER_COUNT(state) != 0) {
3480 rv = casueword32(&rwlock->rw_state, state,
3481 &oldstate, state - 1);
3488 if (URWLOCK_READER_COUNT(oldstate) == 0) {
3492 error = thread_check_susp(td, true);
3505 if (!(flags & URWLOCK_PREFER_READER)) {
3506 if (state & URWLOCK_WRITE_WAITERS) {
3508 q = UMTX_EXCLUSIVE_QUEUE;
3509 } else if (state & URWLOCK_READ_WAITERS) {
3511 q = UMTX_SHARED_QUEUE;
3514 if (state & URWLOCK_READ_WAITERS) {
3516 q = UMTX_SHARED_QUEUE;
3517 } else if (state & URWLOCK_WRITE_WAITERS) {
3519 q = UMTX_EXCLUSIVE_QUEUE;
3524 umtxq_lock(&uq->uq_key);
3525 umtxq_busy(&uq->uq_key);
3526 umtxq_signal_queue(&uq->uq_key, count, q);
3527 umtxq_unbusy(&uq->uq_key);
3528 umtxq_unlock(&uq->uq_key);
3531 umtx_key_release(&uq->uq_key);
3535 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3537 do_sem_wait(struct thread *td, struct _usem *sem, struct _umtx_time *timeout)
3539 struct umtx_abs_timeout timo;
3541 uint32_t flags, count, count1;
3545 error = fueword32(&sem->_flags, &flags);
3548 error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3552 if (timeout != NULL)
3553 umtx_abs_timeout_init2(&timo, timeout);
3556 umtxq_lock(&uq->uq_key);
3557 umtxq_busy(&uq->uq_key);
3559 umtxq_unlock(&uq->uq_key);
3560 rv = casueword32(&sem->_has_waiters, 0, &count1, 1);
3562 rv1 = fueword32(&sem->_count, &count);
3563 if (rv == -1 || rv1 == -1 || count != 0 || (rv == 1 && count1 == 0)) {
3565 suword32(&sem->_has_waiters, 0);
3566 umtxq_lock(&uq->uq_key);
3567 umtxq_unbusy(&uq->uq_key);
3569 umtxq_unlock(&uq->uq_key);
3570 if (rv == -1 || rv1 == -1) {
3578 MPASS(rv == 1 && count1 == 0);
3579 rv = thread_check_susp(td, true);
3585 umtxq_lock(&uq->uq_key);
3586 umtxq_unbusy(&uq->uq_key);
3588 error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3590 if ((uq->uq_flags & UQF_UMTXQ) == 0)
3594 /* A relative timeout cannot be restarted. */
3595 if (error == ERESTART && timeout != NULL &&
3596 (timeout->_flags & UMTX_ABSTIME) == 0)
3599 umtxq_unlock(&uq->uq_key);
3601 umtx_key_release(&uq->uq_key);
3606 * Signal a userland semaphore.
3609 do_sem_wake(struct thread *td, struct _usem *sem)
3611 struct umtx_key key;
3615 error = fueword32(&sem->_flags, &flags);
3618 if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3622 cnt = umtxq_count(&key);
3625 * Check if count is greater than 0, this means the memory is
3626 * still being referenced by user code, so we can safely
3627 * update _has_waiters flag.
3631 error = suword32(&sem->_has_waiters, 0);
3636 umtxq_signal(&key, 1);
3640 umtx_key_release(&key);
3646 do_sem2_wait(struct thread *td, struct _usem2 *sem, struct _umtx_time *timeout)
3648 struct umtx_abs_timeout timo;
3650 uint32_t count, flags;
3654 flags = fuword32(&sem->_flags);
3655 if (timeout != NULL)
3656 umtx_abs_timeout_init2(&timo, timeout);
3659 error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3662 umtxq_lock(&uq->uq_key);
3663 umtxq_busy(&uq->uq_key);
3665 umtxq_unlock(&uq->uq_key);
3666 rv = fueword32(&sem->_count, &count);
3668 umtxq_lock(&uq->uq_key);
3669 umtxq_unbusy(&uq->uq_key);
3671 umtxq_unlock(&uq->uq_key);
3672 umtx_key_release(&uq->uq_key);
3676 if (USEM_COUNT(count) != 0) {
3677 umtxq_lock(&uq->uq_key);
3678 umtxq_unbusy(&uq->uq_key);
3680 umtxq_unlock(&uq->uq_key);
3681 umtx_key_release(&uq->uq_key);
3684 if (count == USEM_HAS_WAITERS)
3686 rv = casueword32(&sem->_count, 0, &count, USEM_HAS_WAITERS);
3689 umtxq_lock(&uq->uq_key);
3690 umtxq_unbusy(&uq->uq_key);
3692 umtxq_unlock(&uq->uq_key);
3693 umtx_key_release(&uq->uq_key);
3696 rv = thread_check_susp(td, true);
3701 umtxq_lock(&uq->uq_key);
3702 umtxq_unbusy(&uq->uq_key);
3704 error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3706 if ((uq->uq_flags & UQF_UMTXQ) == 0)
3710 if (timeout != NULL && (timeout->_flags & UMTX_ABSTIME) == 0) {
3711 /* A relative timeout cannot be restarted. */
3712 if (error == ERESTART)
3714 if (error == EINTR) {
3715 kern_clock_gettime(curthread, timo.clockid,
3717 timespecsub(&timo.end, &timo.cur,
3718 &timeout->_timeout);
3722 umtxq_unlock(&uq->uq_key);
3723 umtx_key_release(&uq->uq_key);
3728 * Signal a userland semaphore.
3731 do_sem2_wake(struct thread *td, struct _usem2 *sem)
3733 struct umtx_key key;
3735 uint32_t count, flags;
3737 rv = fueword32(&sem->_flags, &flags);
3740 if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3744 cnt = umtxq_count(&key);
3747 * If this was the last sleeping thread, clear the waiters
3752 rv = fueword32(&sem->_count, &count);
3753 while (rv != -1 && count & USEM_HAS_WAITERS) {
3754 rv = casueword32(&sem->_count, count, &count,
3755 count & ~USEM_HAS_WAITERS);
3757 rv = thread_check_susp(td, true);
3770 umtxq_signal(&key, 1);
3774 umtx_key_release(&key);
3778 #ifdef COMPAT_FREEBSD10
3780 freebsd10__umtx_lock(struct thread *td, struct freebsd10__umtx_lock_args *uap)
3782 return (do_lock_umtx(td, uap->umtx, td->td_tid, 0));
3786 freebsd10__umtx_unlock(struct thread *td,
3787 struct freebsd10__umtx_unlock_args *uap)
3789 return (do_unlock_umtx(td, uap->umtx, td->td_tid));
3794 umtx_copyin_timeout(const void *uaddr, struct timespec *tsp)
3798 error = copyin(uaddr, tsp, sizeof(*tsp));
3800 if (!timespecvalid_interval(tsp))
3807 umtx_copyin_umtx_time(const void *uaddr, size_t size, struct _umtx_time *tp)
3811 if (size <= sizeof(tp->_timeout)) {
3812 tp->_clockid = CLOCK_REALTIME;
3814 error = copyin(uaddr, &tp->_timeout, sizeof(tp->_timeout));
3816 error = copyin(uaddr, tp, sizeof(*tp));
3819 if (!timespecvalid_interval(&tp->_timeout))
3825 umtx_copyin_robust_lists(const void *uaddr, size_t size,
3826 struct umtx_robust_lists_params *rb)
3829 if (size > sizeof(*rb))
3831 return (copyin(uaddr, rb, size));
3835 umtx_copyout_timeout(void *uaddr, size_t sz, struct timespec *tsp)
3839 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
3840 * and we're only called if sz >= sizeof(timespec) as supplied in the
3843 KASSERT(sz >= sizeof(*tsp),
3844 ("umtx_copyops specifies incorrect sizes"));
3846 return (copyout(tsp, uaddr, sizeof(*tsp)));
3849 #ifdef COMPAT_FREEBSD10
3851 __umtx_op_lock_umtx(struct thread *td, struct _umtx_op_args *uap,
3852 const struct umtx_copyops *ops)
3854 struct timespec *ts, timeout;
3857 /* Allow a null timespec (wait forever). */
3858 if (uap->uaddr2 == NULL)
3861 error = ops->copyin_timeout(uap->uaddr2, &timeout);
3866 #ifdef COMPAT_FREEBSD32
3868 return (do_lock_umtx32(td, uap->obj, uap->val, ts));
3870 return (do_lock_umtx(td, uap->obj, uap->val, ts));
3874 __umtx_op_unlock_umtx(struct thread *td, struct _umtx_op_args *uap,
3875 const struct umtx_copyops *ops)
3877 #ifdef COMPAT_FREEBSD32
3879 return (do_unlock_umtx32(td, uap->obj, uap->val));
3881 return (do_unlock_umtx(td, uap->obj, uap->val));
3883 #endif /* COMPAT_FREEBSD10 */
3885 #if !defined(COMPAT_FREEBSD10)
3887 __umtx_op_unimpl(struct thread *td __unused, struct _umtx_op_args *uap __unused,
3888 const struct umtx_copyops *ops __unused)
3890 return (EOPNOTSUPP);
3892 #endif /* COMPAT_FREEBSD10 */
3895 __umtx_op_wait(struct thread *td, struct _umtx_op_args *uap,
3896 const struct umtx_copyops *ops)
3898 struct _umtx_time timeout, *tm_p;
3901 if (uap->uaddr2 == NULL)
3904 error = ops->copyin_umtx_time(
3905 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3910 return (do_wait(td, uap->obj, uap->val, tm_p, ops->compat32, 0));
3914 __umtx_op_wait_uint(struct thread *td, struct _umtx_op_args *uap,
3915 const struct umtx_copyops *ops)
3917 struct _umtx_time timeout, *tm_p;
3920 if (uap->uaddr2 == NULL)
3923 error = ops->copyin_umtx_time(
3924 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3929 return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
3933 __umtx_op_wait_uint_private(struct thread *td, struct _umtx_op_args *uap,
3934 const struct umtx_copyops *ops)
3936 struct _umtx_time *tm_p, timeout;
3939 if (uap->uaddr2 == NULL)
3942 error = ops->copyin_umtx_time(
3943 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3948 return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
3952 __umtx_op_wake(struct thread *td, struct _umtx_op_args *uap,
3953 const struct umtx_copyops *ops __unused)
3956 return (kern_umtx_wake(td, uap->obj, uap->val, 0));
3959 #define BATCH_SIZE 128
3961 __umtx_op_nwake_private_native(struct thread *td, struct _umtx_op_args *uap)
3963 char *uaddrs[BATCH_SIZE], **upp;
3964 int count, error, i, pos, tocopy;
3966 upp = (char **)uap->obj;
3968 for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3970 tocopy = MIN(count, BATCH_SIZE);
3971 error = copyin(upp + pos, uaddrs, tocopy * sizeof(char *));
3974 for (i = 0; i < tocopy; ++i) {
3975 kern_umtx_wake(td, uaddrs[i], INT_MAX, 1);
3983 __umtx_op_nwake_private_compat32(struct thread *td, struct _umtx_op_args *uap)
3985 uint32_t uaddrs[BATCH_SIZE], *upp;
3986 int count, error, i, pos, tocopy;
3988 upp = (uint32_t *)uap->obj;
3990 for (count = uap->val, pos = 0; count > 0; count -= tocopy,
3992 tocopy = MIN(count, BATCH_SIZE);
3993 error = copyin(upp + pos, uaddrs, tocopy * sizeof(uint32_t));
3996 for (i = 0; i < tocopy; ++i) {
3997 kern_umtx_wake(td, (void *)(uintptr_t)uaddrs[i],
4006 __umtx_op_nwake_private(struct thread *td, struct _umtx_op_args *uap,
4007 const struct umtx_copyops *ops)
4011 return (__umtx_op_nwake_private_compat32(td, uap));
4012 return (__umtx_op_nwake_private_native(td, uap));
4016 __umtx_op_wake_private(struct thread *td, struct _umtx_op_args *uap,
4017 const struct umtx_copyops *ops __unused)
4020 return (kern_umtx_wake(td, uap->obj, uap->val, 1));
4024 __umtx_op_lock_umutex(struct thread *td, struct _umtx_op_args *uap,
4025 const struct umtx_copyops *ops)
4027 struct _umtx_time *tm_p, timeout;
4030 /* Allow a null timespec (wait forever). */
4031 if (uap->uaddr2 == NULL)
4034 error = ops->copyin_umtx_time(
4035 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4040 return (do_lock_umutex(td, uap->obj, tm_p, 0));
4044 __umtx_op_trylock_umutex(struct thread *td, struct _umtx_op_args *uap,
4045 const struct umtx_copyops *ops __unused)
4048 return (do_lock_umutex(td, uap->obj, NULL, _UMUTEX_TRY));
4052 __umtx_op_wait_umutex(struct thread *td, struct _umtx_op_args *uap,
4053 const struct umtx_copyops *ops)
4055 struct _umtx_time *tm_p, timeout;
4058 /* Allow a null timespec (wait forever). */
4059 if (uap->uaddr2 == NULL)
4062 error = ops->copyin_umtx_time(
4063 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4068 return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
4072 __umtx_op_wake_umutex(struct thread *td, struct _umtx_op_args *uap,
4073 const struct umtx_copyops *ops __unused)
4076 return (do_wake_umutex(td, uap->obj));
4080 __umtx_op_unlock_umutex(struct thread *td, struct _umtx_op_args *uap,
4081 const struct umtx_copyops *ops __unused)
4084 return (do_unlock_umutex(td, uap->obj, false));
4088 __umtx_op_set_ceiling(struct thread *td, struct _umtx_op_args *uap,
4089 const struct umtx_copyops *ops __unused)
4092 return (do_set_ceiling(td, uap->obj, uap->val, uap->uaddr1));
4096 __umtx_op_cv_wait(struct thread *td, struct _umtx_op_args *uap,
4097 const struct umtx_copyops *ops)
4099 struct timespec *ts, timeout;
4102 /* Allow a null timespec (wait forever). */
4103 if (uap->uaddr2 == NULL)
4106 error = ops->copyin_timeout(uap->uaddr2, &timeout);
4111 return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
4115 __umtx_op_cv_signal(struct thread *td, struct _umtx_op_args *uap,
4116 const struct umtx_copyops *ops __unused)
4119 return (do_cv_signal(td, uap->obj));
4123 __umtx_op_cv_broadcast(struct thread *td, struct _umtx_op_args *uap,
4124 const struct umtx_copyops *ops __unused)
4127 return (do_cv_broadcast(td, uap->obj));
4131 __umtx_op_rw_rdlock(struct thread *td, struct _umtx_op_args *uap,
4132 const struct umtx_copyops *ops)
4134 struct _umtx_time timeout;
4137 /* Allow a null timespec (wait forever). */
4138 if (uap->uaddr2 == NULL) {
4139 error = do_rw_rdlock(td, uap->obj, uap->val, 0);
4141 error = ops->copyin_umtx_time(uap->uaddr2,
4142 (size_t)uap->uaddr1, &timeout);
4145 error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
4151 __umtx_op_rw_wrlock(struct thread *td, struct _umtx_op_args *uap,
4152 const struct umtx_copyops *ops)
4154 struct _umtx_time timeout;
4157 /* Allow a null timespec (wait forever). */
4158 if (uap->uaddr2 == NULL) {
4159 error = do_rw_wrlock(td, uap->obj, 0);
4161 error = ops->copyin_umtx_time(uap->uaddr2,
4162 (size_t)uap->uaddr1, &timeout);
4166 error = do_rw_wrlock(td, uap->obj, &timeout);
4172 __umtx_op_rw_unlock(struct thread *td, struct _umtx_op_args *uap,
4173 const struct umtx_copyops *ops __unused)
4176 return (do_rw_unlock(td, uap->obj));
4179 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4181 __umtx_op_sem_wait(struct thread *td, struct _umtx_op_args *uap,
4182 const struct umtx_copyops *ops)
4184 struct _umtx_time *tm_p, timeout;
4187 /* Allow a null timespec (wait forever). */
4188 if (uap->uaddr2 == NULL)
4191 error = ops->copyin_umtx_time(
4192 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4197 return (do_sem_wait(td, uap->obj, tm_p));
4201 __umtx_op_sem_wake(struct thread *td, struct _umtx_op_args *uap,
4202 const struct umtx_copyops *ops __unused)
4205 return (do_sem_wake(td, uap->obj));
4210 __umtx_op_wake2_umutex(struct thread *td, struct _umtx_op_args *uap,
4211 const struct umtx_copyops *ops __unused)
4214 return (do_wake2_umutex(td, uap->obj, uap->val));
4218 __umtx_op_sem2_wait(struct thread *td, struct _umtx_op_args *uap,
4219 const struct umtx_copyops *ops)
4221 struct _umtx_time *tm_p, timeout;
4225 /* Allow a null timespec (wait forever). */
4226 if (uap->uaddr2 == NULL) {
4230 uasize = (size_t)uap->uaddr1;
4231 error = ops->copyin_umtx_time(uap->uaddr2, uasize, &timeout);
4236 error = do_sem2_wait(td, uap->obj, tm_p);
4237 if (error == EINTR && uap->uaddr2 != NULL &&
4238 (timeout._flags & UMTX_ABSTIME) == 0 &&
4239 uasize >= ops->umtx_time_sz + ops->timespec_sz) {
4240 error = ops->copyout_timeout(
4241 (void *)((uintptr_t)uap->uaddr2 + ops->umtx_time_sz),
4242 uasize - ops->umtx_time_sz, &timeout._timeout);
4252 __umtx_op_sem2_wake(struct thread *td, struct _umtx_op_args *uap,
4253 const struct umtx_copyops *ops __unused)
4256 return (do_sem2_wake(td, uap->obj));
4259 #define USHM_OBJ_UMTX(o) \
4260 ((struct umtx_shm_obj_list *)(&(o)->umtx_data))
4262 #define USHMF_REG_LINKED 0x0001
4263 #define USHMF_OBJ_LINKED 0x0002
4264 struct umtx_shm_reg {
4265 TAILQ_ENTRY(umtx_shm_reg) ushm_reg_link;
4266 LIST_ENTRY(umtx_shm_reg) ushm_obj_link;
4267 struct umtx_key ushm_key;
4268 struct ucred *ushm_cred;
4269 struct shmfd *ushm_obj;
4274 LIST_HEAD(umtx_shm_obj_list, umtx_shm_reg);
4275 TAILQ_HEAD(umtx_shm_reg_head, umtx_shm_reg);
4277 static uma_zone_t umtx_shm_reg_zone;
4278 static struct umtx_shm_reg_head umtx_shm_registry[UMTX_CHAINS];
4279 static struct mtx umtx_shm_lock;
4280 static struct umtx_shm_reg_head umtx_shm_reg_delfree =
4281 TAILQ_HEAD_INITIALIZER(umtx_shm_reg_delfree);
4283 static void umtx_shm_free_reg(struct umtx_shm_reg *reg);
4286 umtx_shm_reg_delfree_tq(void *context __unused, int pending __unused)
4288 struct umtx_shm_reg_head d;
4289 struct umtx_shm_reg *reg, *reg1;
4292 mtx_lock(&umtx_shm_lock);
4293 TAILQ_CONCAT(&d, &umtx_shm_reg_delfree, ushm_reg_link);
4294 mtx_unlock(&umtx_shm_lock);
4295 TAILQ_FOREACH_SAFE(reg, &d, ushm_reg_link, reg1) {
4296 TAILQ_REMOVE(&d, reg, ushm_reg_link);
4297 umtx_shm_free_reg(reg);
4301 static struct task umtx_shm_reg_delfree_task =
4302 TASK_INITIALIZER(0, umtx_shm_reg_delfree_tq, NULL);
4304 static struct umtx_shm_reg *
4305 umtx_shm_find_reg_locked(const struct umtx_key *key)
4307 struct umtx_shm_reg *reg;
4308 struct umtx_shm_reg_head *reg_head;
4310 KASSERT(key->shared, ("umtx_p_find_rg: private key"));
4311 mtx_assert(&umtx_shm_lock, MA_OWNED);
4312 reg_head = &umtx_shm_registry[key->hash];
4313 TAILQ_FOREACH(reg, reg_head, ushm_reg_link) {
4314 KASSERT(reg->ushm_key.shared,
4315 ("non-shared key on reg %p %d", reg, reg->ushm_key.shared));
4316 if (reg->ushm_key.info.shared.object ==
4317 key->info.shared.object &&
4318 reg->ushm_key.info.shared.offset ==
4319 key->info.shared.offset) {
4320 KASSERT(reg->ushm_key.type == TYPE_SHM, ("TYPE_USHM"));
4321 KASSERT(reg->ushm_refcnt > 0,
4322 ("reg %p refcnt 0 onlist", reg));
4323 KASSERT((reg->ushm_flags & USHMF_REG_LINKED) != 0,
4324 ("reg %p not linked", reg));
4332 static struct umtx_shm_reg *
4333 umtx_shm_find_reg(const struct umtx_key *key)
4335 struct umtx_shm_reg *reg;
4337 mtx_lock(&umtx_shm_lock);
4338 reg = umtx_shm_find_reg_locked(key);
4339 mtx_unlock(&umtx_shm_lock);
4344 umtx_shm_free_reg(struct umtx_shm_reg *reg)
4347 chgumtxcnt(reg->ushm_cred->cr_ruidinfo, -1, 0);
4348 crfree(reg->ushm_cred);
4349 shm_drop(reg->ushm_obj);
4350 uma_zfree(umtx_shm_reg_zone, reg);
4354 umtx_shm_unref_reg_locked(struct umtx_shm_reg *reg, bool force)
4358 mtx_assert(&umtx_shm_lock, MA_OWNED);
4359 KASSERT(reg->ushm_refcnt > 0, ("ushm_reg %p refcnt 0", reg));
4361 res = reg->ushm_refcnt == 0;
4363 if ((reg->ushm_flags & USHMF_REG_LINKED) != 0) {
4364 TAILQ_REMOVE(&umtx_shm_registry[reg->ushm_key.hash],
4365 reg, ushm_reg_link);
4366 reg->ushm_flags &= ~USHMF_REG_LINKED;
4368 if ((reg->ushm_flags & USHMF_OBJ_LINKED) != 0) {
4369 LIST_REMOVE(reg, ushm_obj_link);
4370 reg->ushm_flags &= ~USHMF_OBJ_LINKED;
4377 umtx_shm_unref_reg(struct umtx_shm_reg *reg, bool force)
4383 object = reg->ushm_obj->shm_object;
4384 VM_OBJECT_WLOCK(object);
4385 vm_object_set_flag(object, OBJ_UMTXDEAD);
4386 VM_OBJECT_WUNLOCK(object);
4388 mtx_lock(&umtx_shm_lock);
4389 dofree = umtx_shm_unref_reg_locked(reg, force);
4390 mtx_unlock(&umtx_shm_lock);
4392 umtx_shm_free_reg(reg);
4396 umtx_shm_object_init(vm_object_t object)
4399 LIST_INIT(USHM_OBJ_UMTX(object));
4403 umtx_shm_object_terminated(vm_object_t object)
4405 struct umtx_shm_reg *reg, *reg1;
4408 if (LIST_EMPTY(USHM_OBJ_UMTX(object)))
4412 mtx_lock(&umtx_shm_lock);
4413 LIST_FOREACH_SAFE(reg, USHM_OBJ_UMTX(object), ushm_obj_link, reg1) {
4414 if (umtx_shm_unref_reg_locked(reg, true)) {
4415 TAILQ_INSERT_TAIL(&umtx_shm_reg_delfree, reg,
4420 mtx_unlock(&umtx_shm_lock);
4422 taskqueue_enqueue(taskqueue_thread, &umtx_shm_reg_delfree_task);
4426 umtx_shm_create_reg(struct thread *td, const struct umtx_key *key,
4427 struct umtx_shm_reg **res)
4429 struct umtx_shm_reg *reg, *reg1;
4433 reg = umtx_shm_find_reg(key);
4438 cred = td->td_ucred;
4439 if (!chgumtxcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_UMTXP)))
4441 reg = uma_zalloc(umtx_shm_reg_zone, M_WAITOK | M_ZERO);
4442 reg->ushm_refcnt = 1;
4443 bcopy(key, ®->ushm_key, sizeof(*key));
4444 reg->ushm_obj = shm_alloc(td->td_ucred, O_RDWR, false);
4445 reg->ushm_cred = crhold(cred);
4446 error = shm_dotruncate(reg->ushm_obj, PAGE_SIZE);
4448 umtx_shm_free_reg(reg);
4451 mtx_lock(&umtx_shm_lock);
4452 reg1 = umtx_shm_find_reg_locked(key);
4454 mtx_unlock(&umtx_shm_lock);
4455 umtx_shm_free_reg(reg);
4460 TAILQ_INSERT_TAIL(&umtx_shm_registry[key->hash], reg, ushm_reg_link);
4461 LIST_INSERT_HEAD(USHM_OBJ_UMTX(key->info.shared.object), reg,
4463 reg->ushm_flags = USHMF_REG_LINKED | USHMF_OBJ_LINKED;
4464 mtx_unlock(&umtx_shm_lock);
4470 umtx_shm_alive(struct thread *td, void *addr)
4473 vm_map_entry_t entry;
4480 map = &td->td_proc->p_vmspace->vm_map;
4481 res = vm_map_lookup(&map, (uintptr_t)addr, VM_PROT_READ, &entry,
4482 &object, &pindex, &prot, &wired);
4483 if (res != KERN_SUCCESS)
4488 ret = (object->flags & OBJ_UMTXDEAD) != 0 ? ENOTTY : 0;
4489 vm_map_lookup_done(map, entry);
4498 umtx_shm_reg_zone = uma_zcreate("umtx_shm", sizeof(struct umtx_shm_reg),
4499 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
4500 mtx_init(&umtx_shm_lock, "umtxshm", NULL, MTX_DEF);
4501 for (i = 0; i < nitems(umtx_shm_registry); i++)
4502 TAILQ_INIT(&umtx_shm_registry[i]);
4506 umtx_shm(struct thread *td, void *addr, u_int flags)
4508 struct umtx_key key;
4509 struct umtx_shm_reg *reg;
4513 if (__bitcount(flags & (UMTX_SHM_CREAT | UMTX_SHM_LOOKUP |
4514 UMTX_SHM_DESTROY| UMTX_SHM_ALIVE)) != 1)
4516 if ((flags & UMTX_SHM_ALIVE) != 0)
4517 return (umtx_shm_alive(td, addr));
4518 error = umtx_key_get(addr, TYPE_SHM, PROCESS_SHARE, &key);
4521 KASSERT(key.shared == 1, ("non-shared key"));
4522 if ((flags & UMTX_SHM_CREAT) != 0) {
4523 error = umtx_shm_create_reg(td, &key, ®);
4525 reg = umtx_shm_find_reg(&key);
4529 umtx_key_release(&key);
4532 KASSERT(reg != NULL, ("no reg"));
4533 if ((flags & UMTX_SHM_DESTROY) != 0) {
4534 umtx_shm_unref_reg(reg, true);
4538 error = mac_posixshm_check_open(td->td_ucred,
4539 reg->ushm_obj, FFLAGS(O_RDWR));
4542 error = shm_access(reg->ushm_obj, td->td_ucred,
4546 error = falloc_caps(td, &fp, &fd, O_CLOEXEC, NULL);
4548 shm_hold(reg->ushm_obj);
4549 finit(fp, FFLAGS(O_RDWR), DTYPE_SHM, reg->ushm_obj,
4551 td->td_retval[0] = fd;
4555 umtx_shm_unref_reg(reg, false);
4560 __umtx_op_shm(struct thread *td, struct _umtx_op_args *uap,
4561 const struct umtx_copyops *ops __unused)
4564 return (umtx_shm(td, uap->uaddr1, uap->val));
4568 __umtx_op_robust_lists(struct thread *td, struct _umtx_op_args *uap,
4569 const struct umtx_copyops *ops)
4571 struct umtx_robust_lists_params rb;
4574 if (ops->compat32) {
4575 if ((td->td_pflags2 & TDP2_COMPAT32RB) == 0 &&
4576 (td->td_rb_list != 0 || td->td_rbp_list != 0 ||
4577 td->td_rb_inact != 0))
4579 } else if ((td->td_pflags2 & TDP2_COMPAT32RB) != 0) {
4583 bzero(&rb, sizeof(rb));
4584 error = ops->copyin_robust_lists(uap->uaddr1, uap->val, &rb);
4589 td->td_pflags2 |= TDP2_COMPAT32RB;
4591 td->td_rb_list = rb.robust_list_offset;
4592 td->td_rbp_list = rb.robust_priv_list_offset;
4593 td->td_rb_inact = rb.robust_inact_offset;
4597 #if defined(__i386__) || defined(__amd64__)
4599 * Provide the standard 32-bit definitions for x86, since native/compat32 use a
4600 * 32-bit time_t there. Other architectures just need the i386 definitions
4601 * along with their standard compat32.
4603 struct timespecx32 {
4608 struct umtx_timex32 {
4609 struct timespecx32 _timeout;
4615 #define timespeci386 timespec32
4616 #define umtx_timei386 umtx_time32
4618 #else /* !__i386__ && !__amd64__ */
4619 /* 32-bit architectures can emulate i386, so define these almost everywhere. */
4620 struct timespeci386 {
4625 struct umtx_timei386 {
4626 struct timespeci386 _timeout;
4631 #if defined(__LP64__)
4632 #define timespecx32 timespec32
4633 #define umtx_timex32 umtx_time32
4638 umtx_copyin_robust_lists32(const void *uaddr, size_t size,
4639 struct umtx_robust_lists_params *rbp)
4641 struct umtx_robust_lists_params_compat32 rb32;
4644 if (size > sizeof(rb32))
4646 bzero(&rb32, sizeof(rb32));
4647 error = copyin(uaddr, &rb32, size);
4650 CP(rb32, *rbp, robust_list_offset);
4651 CP(rb32, *rbp, robust_priv_list_offset);
4652 CP(rb32, *rbp, robust_inact_offset);
4658 umtx_copyin_timeouti386(const void *uaddr, struct timespec *tsp)
4660 struct timespeci386 ts32;
4663 error = copyin(uaddr, &ts32, sizeof(ts32));
4665 if (!timespecvalid_interval(&ts32))
4668 CP(ts32, *tsp, tv_sec);
4669 CP(ts32, *tsp, tv_nsec);
4676 umtx_copyin_umtx_timei386(const void *uaddr, size_t size, struct _umtx_time *tp)
4678 struct umtx_timei386 t32;
4681 t32._clockid = CLOCK_REALTIME;
4683 if (size <= sizeof(t32._timeout))
4684 error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4686 error = copyin(uaddr, &t32, sizeof(t32));
4689 if (!timespecvalid_interval(&t32._timeout))
4691 TS_CP(t32, *tp, _timeout);
4692 CP(t32, *tp, _flags);
4693 CP(t32, *tp, _clockid);
4698 umtx_copyout_timeouti386(void *uaddr, size_t sz, struct timespec *tsp)
4700 struct timespeci386 remain32 = {
4701 .tv_sec = tsp->tv_sec,
4702 .tv_nsec = tsp->tv_nsec,
4706 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4707 * and we're only called if sz >= sizeof(timespec) as supplied in the
4710 KASSERT(sz >= sizeof(remain32),
4711 ("umtx_copyops specifies incorrect sizes"));
4713 return (copyout(&remain32, uaddr, sizeof(remain32)));
4715 #endif /* !__i386__ */
4717 #if defined(__i386__) || defined(__LP64__)
4719 umtx_copyin_timeoutx32(const void *uaddr, struct timespec *tsp)
4721 struct timespecx32 ts32;
4724 error = copyin(uaddr, &ts32, sizeof(ts32));
4726 if (!timespecvalid_interval(&ts32))
4729 CP(ts32, *tsp, tv_sec);
4730 CP(ts32, *tsp, tv_nsec);
4737 umtx_copyin_umtx_timex32(const void *uaddr, size_t size, struct _umtx_time *tp)
4739 struct umtx_timex32 t32;
4742 t32._clockid = CLOCK_REALTIME;
4744 if (size <= sizeof(t32._timeout))
4745 error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4747 error = copyin(uaddr, &t32, sizeof(t32));
4750 if (!timespecvalid_interval(&t32._timeout))
4752 TS_CP(t32, *tp, _timeout);
4753 CP(t32, *tp, _flags);
4754 CP(t32, *tp, _clockid);
4759 umtx_copyout_timeoutx32(void *uaddr, size_t sz, struct timespec *tsp)
4761 struct timespecx32 remain32 = {
4762 .tv_sec = tsp->tv_sec,
4763 .tv_nsec = tsp->tv_nsec,
4767 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4768 * and we're only called if sz >= sizeof(timespec) as supplied in the
4771 KASSERT(sz >= sizeof(remain32),
4772 ("umtx_copyops specifies incorrect sizes"));
4774 return (copyout(&remain32, uaddr, sizeof(remain32)));
4776 #endif /* __i386__ || __LP64__ */
4778 typedef int (*_umtx_op_func)(struct thread *td, struct _umtx_op_args *uap,
4779 const struct umtx_copyops *umtx_ops);
4781 static const _umtx_op_func op_table[] = {
4782 #ifdef COMPAT_FREEBSD10
4783 [UMTX_OP_LOCK] = __umtx_op_lock_umtx,
4784 [UMTX_OP_UNLOCK] = __umtx_op_unlock_umtx,
4786 [UMTX_OP_LOCK] = __umtx_op_unimpl,
4787 [UMTX_OP_UNLOCK] = __umtx_op_unimpl,
4789 [UMTX_OP_WAIT] = __umtx_op_wait,
4790 [UMTX_OP_WAKE] = __umtx_op_wake,
4791 [UMTX_OP_MUTEX_TRYLOCK] = __umtx_op_trylock_umutex,
4792 [UMTX_OP_MUTEX_LOCK] = __umtx_op_lock_umutex,
4793 [UMTX_OP_MUTEX_UNLOCK] = __umtx_op_unlock_umutex,
4794 [UMTX_OP_SET_CEILING] = __umtx_op_set_ceiling,
4795 [UMTX_OP_CV_WAIT] = __umtx_op_cv_wait,
4796 [UMTX_OP_CV_SIGNAL] = __umtx_op_cv_signal,
4797 [UMTX_OP_CV_BROADCAST] = __umtx_op_cv_broadcast,
4798 [UMTX_OP_WAIT_UINT] = __umtx_op_wait_uint,
4799 [UMTX_OP_RW_RDLOCK] = __umtx_op_rw_rdlock,
4800 [UMTX_OP_RW_WRLOCK] = __umtx_op_rw_wrlock,
4801 [UMTX_OP_RW_UNLOCK] = __umtx_op_rw_unlock,
4802 [UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private,
4803 [UMTX_OP_WAKE_PRIVATE] = __umtx_op_wake_private,
4804 [UMTX_OP_MUTEX_WAIT] = __umtx_op_wait_umutex,
4805 [UMTX_OP_MUTEX_WAKE] = __umtx_op_wake_umutex,
4806 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4807 [UMTX_OP_SEM_WAIT] = __umtx_op_sem_wait,
4808 [UMTX_OP_SEM_WAKE] = __umtx_op_sem_wake,
4810 [UMTX_OP_SEM_WAIT] = __umtx_op_unimpl,
4811 [UMTX_OP_SEM_WAKE] = __umtx_op_unimpl,
4813 [UMTX_OP_NWAKE_PRIVATE] = __umtx_op_nwake_private,
4814 [UMTX_OP_MUTEX_WAKE2] = __umtx_op_wake2_umutex,
4815 [UMTX_OP_SEM2_WAIT] = __umtx_op_sem2_wait,
4816 [UMTX_OP_SEM2_WAKE] = __umtx_op_sem2_wake,
4817 [UMTX_OP_SHM] = __umtx_op_shm,
4818 [UMTX_OP_ROBUST_LISTS] = __umtx_op_robust_lists,
4821 static const struct umtx_copyops umtx_native_ops = {
4822 .copyin_timeout = umtx_copyin_timeout,
4823 .copyin_umtx_time = umtx_copyin_umtx_time,
4824 .copyin_robust_lists = umtx_copyin_robust_lists,
4825 .copyout_timeout = umtx_copyout_timeout,
4826 .timespec_sz = sizeof(struct timespec),
4827 .umtx_time_sz = sizeof(struct _umtx_time),
4831 static const struct umtx_copyops umtx_native_opsi386 = {
4832 .copyin_timeout = umtx_copyin_timeouti386,
4833 .copyin_umtx_time = umtx_copyin_umtx_timei386,
4834 .copyin_robust_lists = umtx_copyin_robust_lists32,
4835 .copyout_timeout = umtx_copyout_timeouti386,
4836 .timespec_sz = sizeof(struct timespeci386),
4837 .umtx_time_sz = sizeof(struct umtx_timei386),
4842 #if defined(__i386__) || defined(__LP64__)
4843 /* i386 can emulate other 32-bit archs, too! */
4844 static const struct umtx_copyops umtx_native_opsx32 = {
4845 .copyin_timeout = umtx_copyin_timeoutx32,
4846 .copyin_umtx_time = umtx_copyin_umtx_timex32,
4847 .copyin_robust_lists = umtx_copyin_robust_lists32,
4848 .copyout_timeout = umtx_copyout_timeoutx32,
4849 .timespec_sz = sizeof(struct timespecx32),
4850 .umtx_time_sz = sizeof(struct umtx_timex32),
4854 #ifdef COMPAT_FREEBSD32
4856 #define umtx_native_ops32 umtx_native_opsi386
4858 #define umtx_native_ops32 umtx_native_opsx32
4860 #endif /* COMPAT_FREEBSD32 */
4861 #endif /* __i386__ || __LP64__ */
4863 #define UMTX_OP__FLAGS (UMTX_OP__32BIT | UMTX_OP__I386)
4866 kern__umtx_op(struct thread *td, void *obj, int op, unsigned long val,
4867 void *uaddr1, void *uaddr2, const struct umtx_copyops *ops)
4869 struct _umtx_op_args uap = {
4871 .op = op & ~UMTX_OP__FLAGS,
4877 if ((uap.op >= nitems(op_table)))
4879 return ((*op_table[uap.op])(td, &uap, ops));
4883 sys__umtx_op(struct thread *td, struct _umtx_op_args *uap)
4885 static const struct umtx_copyops *umtx_ops;
4887 umtx_ops = &umtx_native_ops;
4889 if ((uap->op & (UMTX_OP__32BIT | UMTX_OP__I386)) != 0) {
4890 if ((uap->op & UMTX_OP__I386) != 0)
4891 umtx_ops = &umtx_native_opsi386;
4893 umtx_ops = &umtx_native_opsx32;
4895 #elif !defined(__i386__)
4896 /* We consider UMTX_OP__32BIT a nop on !i386 ILP32. */
4897 if ((uap->op & UMTX_OP__I386) != 0)
4898 umtx_ops = &umtx_native_opsi386;
4900 /* Likewise, UMTX_OP__I386 is a nop on i386. */
4901 if ((uap->op & UMTX_OP__32BIT) != 0)
4902 umtx_ops = &umtx_native_opsx32;
4904 return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4905 uap->uaddr2, umtx_ops));
4908 #ifdef COMPAT_FREEBSD32
4909 #ifdef COMPAT_FREEBSD10
4911 freebsd10_freebsd32__umtx_lock(struct thread *td,
4912 struct freebsd10_freebsd32__umtx_lock_args *uap)
4914 return (do_lock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid, NULL));
4918 freebsd10_freebsd32__umtx_unlock(struct thread *td,
4919 struct freebsd10_freebsd32__umtx_unlock_args *uap)
4921 return (do_unlock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid));
4923 #endif /* COMPAT_FREEBSD10 */
4926 freebsd32__umtx_op(struct thread *td, struct freebsd32__umtx_op_args *uap)
4929 return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
4930 uap->uaddr2, &umtx_native_ops32));
4932 #endif /* COMPAT_FREEBSD32 */
4935 umtx_thread_init(struct thread *td)
4938 td->td_umtxq = umtxq_alloc();
4939 td->td_umtxq->uq_thread = td;
4943 umtx_thread_fini(struct thread *td)
4946 umtxq_free(td->td_umtxq);
4950 * It will be called when new thread is created, e.g fork().
4953 umtx_thread_alloc(struct thread *td)
4958 uq->uq_inherited_pri = PRI_MAX;
4960 KASSERT(uq->uq_flags == 0, ("uq_flags != 0"));
4961 KASSERT(uq->uq_thread == td, ("uq_thread != td"));
4962 KASSERT(uq->uq_pi_blocked == NULL, ("uq_pi_blocked != NULL"));
4963 KASSERT(TAILQ_EMPTY(&uq->uq_pi_contested), ("uq_pi_contested is not empty"));
4969 * Clear robust lists for all process' threads, not delaying the
4970 * cleanup to thread exit, since the relevant address space is
4971 * destroyed right now.
4974 umtx_exec(struct proc *p)
4978 KASSERT(p == curproc, ("need curproc"));
4979 KASSERT((p->p_flag & P_HADTHREADS) == 0 ||
4980 (p->p_flag & P_STOPPED_SINGLE) != 0,
4981 ("curproc must be single-threaded"));
4983 * There is no need to lock the list as only this thread can be
4986 FOREACH_THREAD_IN_PROC(p, td) {
4987 KASSERT(td == curthread ||
4988 ((td->td_flags & TDF_BOUNDARY) != 0 && TD_IS_SUSPENDED(td)),
4989 ("running thread %p %p", p, td));
4990 umtx_thread_cleanup(td);
4991 td->td_rb_list = td->td_rbp_list = td->td_rb_inact = 0;
4999 umtx_thread_exit(struct thread *td)
5002 umtx_thread_cleanup(td);
5006 umtx_read_uptr(struct thread *td, uintptr_t ptr, uintptr_t *res, bool compat32)
5013 error = fueword32((void *)ptr, &res32);
5017 error = fueword((void *)ptr, &res1);
5027 umtx_read_rb_list(struct thread *td, struct umutex *m, uintptr_t *rb_list,
5030 struct umutex32 m32;
5033 memcpy(&m32, m, sizeof(m32));
5034 *rb_list = m32.m_rb_lnk;
5036 *rb_list = m->m_rb_lnk;
5041 umtx_handle_rb(struct thread *td, uintptr_t rbp, uintptr_t *rb_list, bool inact,
5047 KASSERT(td->td_proc == curproc, ("need current vmspace"));
5048 error = copyin((void *)rbp, &m, sizeof(m));
5051 if (rb_list != NULL)
5052 umtx_read_rb_list(td, &m, rb_list, compat32);
5053 if ((m.m_flags & UMUTEX_ROBUST) == 0)
5055 if ((m.m_owner & ~UMUTEX_CONTESTED) != td->td_tid)
5056 /* inact is cleared after unlock, allow the inconsistency */
5057 return (inact ? 0 : EINVAL);
5058 return (do_unlock_umutex(td, (struct umutex *)rbp, true));
5062 umtx_cleanup_rb_list(struct thread *td, uintptr_t rb_list, uintptr_t *rb_inact,
5063 const char *name, bool compat32)
5071 error = umtx_read_uptr(td, rb_list, &rbp, compat32);
5072 for (i = 0; error == 0 && rbp != 0 && i < umtx_max_rb; i++) {
5073 if (rbp == *rb_inact) {
5078 error = umtx_handle_rb(td, rbp, &rbp, inact, compat32);
5080 if (i == umtx_max_rb && umtx_verbose_rb) {
5081 uprintf("comm %s pid %d: reached umtx %smax rb %d\n",
5082 td->td_proc->p_comm, td->td_proc->p_pid, name, umtx_max_rb);
5084 if (error != 0 && umtx_verbose_rb) {
5085 uprintf("comm %s pid %d: handling %srb error %d\n",
5086 td->td_proc->p_comm, td->td_proc->p_pid, name, error);
5091 * Clean up umtx data.
5094 umtx_thread_cleanup(struct thread *td)
5102 * Disown pi mutexes.
5106 if (uq->uq_inherited_pri != PRI_MAX ||
5107 !TAILQ_EMPTY(&uq->uq_pi_contested)) {
5108 mtx_lock(&umtx_lock);
5109 uq->uq_inherited_pri = PRI_MAX;
5110 while ((pi = TAILQ_FIRST(&uq->uq_pi_contested)) != NULL) {
5111 pi->pi_owner = NULL;
5112 TAILQ_REMOVE(&uq->uq_pi_contested, pi, pi_link);
5114 mtx_unlock(&umtx_lock);
5116 sched_lend_user_prio_cond(td, PRI_MAX);
5119 compat32 = (td->td_pflags2 & TDP2_COMPAT32RB) != 0;
5120 td->td_pflags2 &= ~TDP2_COMPAT32RB;
5122 if (td->td_rb_inact == 0 && td->td_rb_list == 0 && td->td_rbp_list == 0)
5126 * Handle terminated robust mutexes. Must be done after
5127 * robust pi disown, otherwise unlock could see unowned
5130 rb_inact = td->td_rb_inact;
5132 (void)umtx_read_uptr(td, rb_inact, &rb_inact, compat32);
5133 umtx_cleanup_rb_list(td, td->td_rb_list, &rb_inact, "", compat32);
5134 umtx_cleanup_rb_list(td, td->td_rbp_list, &rb_inact, "priv ", compat32);
5136 (void)umtx_handle_rb(td, rb_inact, NULL, true, compat32);