1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2015, 2016 The FreeBSD Foundation
5 * Copyright (c) 2004, David Xu <[email protected]>
6 * Copyright (c) 2002, Jeffrey Roberson <[email protected]>
7 * All rights reserved.
8 *
9 * Portions of this software were developed by Konstantin Belousov
10 * under sponsorship from the FreeBSD Foundation.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 * 1. Redistributions of source code must retain the above copyright
16 * notice unmodified, this list of conditions, and the following
17 * disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
23 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
24 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
25 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
27 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
31 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32 */
33
34 #include <sys/cdefs.h>
35 #include "opt_umtx_profiling.h"
36
37 #include <sys/param.h>
38 #include <sys/kernel.h>
39 #include <sys/fcntl.h>
40 #include <sys/file.h>
41 #include <sys/filedesc.h>
42 #include <sys/limits.h>
43 #include <sys/lock.h>
44 #include <sys/malloc.h>
45 #include <sys/mman.h>
46 #include <sys/mutex.h>
47 #include <sys/priv.h>
48 #include <sys/proc.h>
49 #include <sys/resource.h>
50 #include <sys/resourcevar.h>
51 #include <sys/rwlock.h>
52 #include <sys/sbuf.h>
53 #include <sys/sched.h>
54 #include <sys/smp.h>
55 #include <sys/sysctl.h>
56 #include <sys/systm.h>
57 #include <sys/sysproto.h>
58 #include <sys/syscallsubr.h>
59 #include <sys/taskqueue.h>
60 #include <sys/time.h>
61 #include <sys/eventhandler.h>
62 #include <sys/umtx.h>
63 #include <sys/umtxvar.h>
64
65 #include <security/mac/mac_framework.h>
66
67 #include <vm/vm.h>
68 #include <vm/vm_param.h>
69 #include <vm/pmap.h>
70 #include <vm/vm_map.h>
71 #include <vm/vm_object.h>
72
73 #include <machine/atomic.h>
74 #include <machine/cpu.h>
75
76 #include <compat/freebsd32/freebsd32.h>
77 #ifdef COMPAT_FREEBSD32
78 #include <compat/freebsd32/freebsd32_proto.h>
79 #endif
80
81 #define _UMUTEX_TRY 1
82 #define _UMUTEX_WAIT 2
83
84 #ifdef UMTX_PROFILING
85 #define UPROF_PERC_BIGGER(w, f, sw, sf) \
86 (((w) > (sw)) || ((w) == (sw) && (f) > (sf)))
87 #endif
88
89 #define UMTXQ_LOCKED_ASSERT(uc) mtx_assert(&(uc)->uc_lock, MA_OWNED)
90 #ifdef INVARIANTS
91 #define UMTXQ_ASSERT_LOCKED_BUSY(key) do { \
92 struct umtxq_chain *uc; \
93 \
94 uc = umtxq_getchain(key); \
95 mtx_assert(&uc->uc_lock, MA_OWNED); \
96 KASSERT(uc->uc_busy != 0, ("umtx chain is not busy")); \
97 } while (0)
98 #else
99 #define UMTXQ_ASSERT_LOCKED_BUSY(key) do {} while (0)
100 #endif
101
102 /*
103 * Don't propagate time-sharing priority, there is a security reason,
104 * a user can simply introduce PI-mutex, let thread A lock the mutex,
105 * and let another thread B block on the mutex, because B is
106 * sleeping, its priority will be boosted, this causes A's priority to
107 * be boosted via priority propagating too and will never be lowered even
108 * if it is using 100%CPU, this is unfair to other processes.
109 */
110
111 #define UPRI(td) (((td)->td_user_pri >= PRI_MIN_TIMESHARE &&\
112 (td)->td_user_pri <= PRI_MAX_TIMESHARE) ?\
113 PRI_MAX_TIMESHARE : (td)->td_user_pri)
114
115 #define GOLDEN_RATIO_PRIME 2654404609U
116 #ifndef UMTX_CHAINS
117 #define UMTX_CHAINS 512
118 #endif
119 #define UMTX_SHIFTS (__WORD_BIT - 9)
120
121 #define GET_SHARE(flags) \
122 (((flags) & USYNC_PROCESS_SHARED) == 0 ? THREAD_SHARE : PROCESS_SHARE)
123
124 #define BUSY_SPINS 200
125
126 struct umtx_copyops {
127 int (*copyin_timeout)(const void *uaddr, struct timespec *tsp);
128 int (*copyin_umtx_time)(const void *uaddr, size_t size,
129 struct _umtx_time *tp);
130 int (*copyin_robust_lists)(const void *uaddr, size_t size,
131 struct umtx_robust_lists_params *rbp);
132 int (*copyout_timeout)(void *uaddr, size_t size,
133 struct timespec *tsp);
134 const size_t timespec_sz;
135 const size_t umtx_time_sz;
136 const bool compat32;
137 };
138
139 _Static_assert(sizeof(struct umutex) == sizeof(struct umutex32), "umutex32");
140 _Static_assert(__offsetof(struct umutex, m_spare[0]) ==
141 __offsetof(struct umutex32, m_spare[0]), "m_spare32");
142
143 int umtx_shm_vnobj_persistent = 0;
144 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_vnode_persistent, CTLFLAG_RWTUN,
145 &umtx_shm_vnobj_persistent, 0,
146 "False forces destruction of umtx attached to file, on last close");
147 static int umtx_max_rb = 1000;
148 SYSCTL_INT(_kern_ipc, OID_AUTO, umtx_max_robust, CTLFLAG_RWTUN,
149 &umtx_max_rb, 0,
150 "Maximum number of robust mutexes allowed for each thread");
151
152 static uma_zone_t umtx_pi_zone;
153 static struct umtxq_chain umtxq_chains[2][UMTX_CHAINS];
154 static MALLOC_DEFINE(M_UMTX, "umtx", "UMTX queue memory");
155 static int umtx_pi_allocated;
156
157 static SYSCTL_NODE(_debug, OID_AUTO, umtx, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
158 "umtx debug");
159 SYSCTL_INT(_debug_umtx, OID_AUTO, umtx_pi_allocated, CTLFLAG_RD,
160 &umtx_pi_allocated, 0, "Allocated umtx_pi");
161 static int umtx_verbose_rb = 1;
162 SYSCTL_INT(_debug_umtx, OID_AUTO, robust_faults_verbose, CTLFLAG_RWTUN,
163 &umtx_verbose_rb, 0,
164 "");
165
166 #ifdef UMTX_PROFILING
167 static long max_length;
168 SYSCTL_LONG(_debug_umtx, OID_AUTO, max_length, CTLFLAG_RD, &max_length, 0, "max_length");
169 static SYSCTL_NODE(_debug_umtx, OID_AUTO, chains, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
170 "umtx chain stats");
171 #endif
172
173 static inline void umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
174 const struct _umtx_time *umtxtime);
175
176 static void umtx_shm_init(void);
177 static void umtxq_sysinit(void *);
178 static void umtxq_hash(struct umtx_key *key);
179 static int do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags,
180 bool rb);
181 static void umtx_thread_cleanup(struct thread *td);
182 SYSINIT(umtx, SI_SUB_EVENTHANDLER+1, SI_ORDER_MIDDLE, umtxq_sysinit, NULL);
183
184 #define umtxq_signal(key, nwake) umtxq_signal_queue((key), (nwake), UMTX_SHARED_QUEUE)
185
186 static struct mtx umtx_lock;
187
188 #ifdef UMTX_PROFILING
189 static void
umtx_init_profiling(void)190 umtx_init_profiling(void)
191 {
192 struct sysctl_oid *chain_oid;
193 char chain_name[10];
194 int i;
195
196 for (i = 0; i < UMTX_CHAINS; ++i) {
197 snprintf(chain_name, sizeof(chain_name), "%d", i);
198 chain_oid = SYSCTL_ADD_NODE(NULL,
199 SYSCTL_STATIC_CHILDREN(_debug_umtx_chains), OID_AUTO,
200 chain_name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
201 "umtx hash stats");
202 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
203 "max_length0", CTLFLAG_RD, &umtxq_chains[0][i].max_length, 0, NULL);
204 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO,
205 "max_length1", CTLFLAG_RD, &umtxq_chains[1][i].max_length, 0, NULL);
206 }
207 }
208
209 static int
sysctl_debug_umtx_chains_peaks(SYSCTL_HANDLER_ARGS)210 sysctl_debug_umtx_chains_peaks(SYSCTL_HANDLER_ARGS)
211 {
212 char buf[512];
213 struct sbuf sb;
214 struct umtxq_chain *uc;
215 u_int fract, i, j, tot, whole;
216 u_int sf0, sf1, sf2, sf3, sf4;
217 u_int si0, si1, si2, si3, si4;
218 u_int sw0, sw1, sw2, sw3, sw4;
219
220 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
221 for (i = 0; i < 2; i++) {
222 tot = 0;
223 for (j = 0; j < UMTX_CHAINS; ++j) {
224 uc = &umtxq_chains[i][j];
225 mtx_lock(&uc->uc_lock);
226 tot += uc->max_length;
227 mtx_unlock(&uc->uc_lock);
228 }
229 if (tot == 0)
230 sbuf_printf(&sb, "%u) Empty ", i);
231 else {
232 sf0 = sf1 = sf2 = sf3 = sf4 = 0;
233 si0 = si1 = si2 = si3 = si4 = 0;
234 sw0 = sw1 = sw2 = sw3 = sw4 = 0;
235 for (j = 0; j < UMTX_CHAINS; j++) {
236 uc = &umtxq_chains[i][j];
237 mtx_lock(&uc->uc_lock);
238 whole = uc->max_length * 100;
239 mtx_unlock(&uc->uc_lock);
240 fract = (whole % tot) * 100;
241 if (UPROF_PERC_BIGGER(whole, fract, sw0, sf0)) {
242 sf0 = fract;
243 si0 = j;
244 sw0 = whole;
245 } else if (UPROF_PERC_BIGGER(whole, fract, sw1,
246 sf1)) {
247 sf1 = fract;
248 si1 = j;
249 sw1 = whole;
250 } else if (UPROF_PERC_BIGGER(whole, fract, sw2,
251 sf2)) {
252 sf2 = fract;
253 si2 = j;
254 sw2 = whole;
255 } else if (UPROF_PERC_BIGGER(whole, fract, sw3,
256 sf3)) {
257 sf3 = fract;
258 si3 = j;
259 sw3 = whole;
260 } else if (UPROF_PERC_BIGGER(whole, fract, sw4,
261 sf4)) {
262 sf4 = fract;
263 si4 = j;
264 sw4 = whole;
265 }
266 }
267 sbuf_printf(&sb, "queue %u:\n", i);
268 sbuf_printf(&sb, "1st: %u.%u%% idx: %u\n", sw0 / tot,
269 sf0 / tot, si0);
270 sbuf_printf(&sb, "2nd: %u.%u%% idx: %u\n", sw1 / tot,
271 sf1 / tot, si1);
272 sbuf_printf(&sb, "3rd: %u.%u%% idx: %u\n", sw2 / tot,
273 sf2 / tot, si2);
274 sbuf_printf(&sb, "4th: %u.%u%% idx: %u\n", sw3 / tot,
275 sf3 / tot, si3);
276 sbuf_printf(&sb, "5th: %u.%u%% idx: %u\n", sw4 / tot,
277 sf4 / tot, si4);
278 }
279 }
280 sbuf_trim(&sb);
281 sbuf_finish(&sb);
282 sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
283 sbuf_delete(&sb);
284 return (0);
285 }
286
287 static int
sysctl_debug_umtx_chains_clear(SYSCTL_HANDLER_ARGS)288 sysctl_debug_umtx_chains_clear(SYSCTL_HANDLER_ARGS)
289 {
290 struct umtxq_chain *uc;
291 u_int i, j;
292 int clear, error;
293
294 clear = 0;
295 error = sysctl_handle_int(oidp, &clear, 0, req);
296 if (error != 0 || req->newptr == NULL)
297 return (error);
298
299 if (clear != 0) {
300 for (i = 0; i < 2; ++i) {
301 for (j = 0; j < UMTX_CHAINS; ++j) {
302 uc = &umtxq_chains[i][j];
303 mtx_lock(&uc->uc_lock);
304 uc->length = 0;
305 uc->max_length = 0;
306 mtx_unlock(&uc->uc_lock);
307 }
308 }
309 }
310 return (0);
311 }
312
313 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, clear,
314 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
315 sysctl_debug_umtx_chains_clear, "I",
316 "Clear umtx chains statistics");
317 SYSCTL_PROC(_debug_umtx_chains, OID_AUTO, peaks,
318 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
319 sysctl_debug_umtx_chains_peaks, "A",
320 "Highest peaks in chains max length");
321 #endif
322
323 static void
umtxq_sysinit(void * arg __unused)324 umtxq_sysinit(void *arg __unused)
325 {
326 int i, j;
327
328 umtx_pi_zone = uma_zcreate("umtx pi", sizeof(struct umtx_pi),
329 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
330 for (i = 0; i < 2; ++i) {
331 for (j = 0; j < UMTX_CHAINS; ++j) {
332 mtx_init(&umtxq_chains[i][j].uc_lock, "umtxql", NULL,
333 MTX_DEF | MTX_DUPOK);
334 LIST_INIT(&umtxq_chains[i][j].uc_queue[0]);
335 LIST_INIT(&umtxq_chains[i][j].uc_queue[1]);
336 LIST_INIT(&umtxq_chains[i][j].uc_spare_queue);
337 TAILQ_INIT(&umtxq_chains[i][j].uc_pi_list);
338 umtxq_chains[i][j].uc_busy = 0;
339 umtxq_chains[i][j].uc_waiters = 0;
340 #ifdef UMTX_PROFILING
341 umtxq_chains[i][j].length = 0;
342 umtxq_chains[i][j].max_length = 0;
343 #endif
344 }
345 }
346 #ifdef UMTX_PROFILING
347 umtx_init_profiling();
348 #endif
349 mtx_init(&umtx_lock, "umtx lock", NULL, MTX_DEF);
350 umtx_shm_init();
351 }
352
353 struct umtx_q *
umtxq_alloc(void)354 umtxq_alloc(void)
355 {
356 struct umtx_q *uq;
357
358 uq = malloc(sizeof(struct umtx_q), M_UMTX, M_WAITOK | M_ZERO);
359 uq->uq_spare_queue = malloc(sizeof(struct umtxq_queue), M_UMTX,
360 M_WAITOK | M_ZERO);
361 TAILQ_INIT(&uq->uq_spare_queue->head);
362 TAILQ_INIT(&uq->uq_pi_contested);
363 uq->uq_inherited_pri = PRI_MAX;
364 return (uq);
365 }
366
367 void
umtxq_free(struct umtx_q * uq)368 umtxq_free(struct umtx_q *uq)
369 {
370
371 MPASS(uq->uq_spare_queue != NULL);
372 free(uq->uq_spare_queue, M_UMTX);
373 free(uq, M_UMTX);
374 }
375
376 static inline void
umtxq_hash(struct umtx_key * key)377 umtxq_hash(struct umtx_key *key)
378 {
379 unsigned n;
380
381 n = (uintptr_t)key->info.both.a + key->info.both.b;
382 key->hash = ((n * GOLDEN_RATIO_PRIME) >> UMTX_SHIFTS) % UMTX_CHAINS;
383 }
384
385 struct umtxq_chain *
umtxq_getchain(struct umtx_key * key)386 umtxq_getchain(struct umtx_key *key)
387 {
388
389 if (key->type <= TYPE_SEM)
390 return (&umtxq_chains[1][key->hash]);
391 return (&umtxq_chains[0][key->hash]);
392 }
393
394 /*
395 * Set chain to busy state when following operation
396 * may be blocked (kernel mutex can not be used).
397 */
398 void
umtxq_busy(struct umtx_key * key)399 umtxq_busy(struct umtx_key *key)
400 {
401 struct umtxq_chain *uc;
402
403 uc = umtxq_getchain(key);
404 mtx_assert(&uc->uc_lock, MA_OWNED);
405 if (uc->uc_busy) {
406 #ifdef SMP
407 if (smp_cpus > 1) {
408 int count = BUSY_SPINS;
409 if (count > 0) {
410 umtxq_unlock(key);
411 while (uc->uc_busy && --count > 0)
412 cpu_spinwait();
413 umtxq_lock(key);
414 }
415 }
416 #endif
417 while (uc->uc_busy) {
418 uc->uc_waiters++;
419 msleep(uc, &uc->uc_lock, 0, "umtxqb", 0);
420 uc->uc_waiters--;
421 }
422 }
423 uc->uc_busy = 1;
424 }
425
426 /*
427 * Unbusy a chain.
428 */
429 void
umtxq_unbusy(struct umtx_key * key)430 umtxq_unbusy(struct umtx_key *key)
431 {
432 struct umtxq_chain *uc;
433
434 uc = umtxq_getchain(key);
435 mtx_assert(&uc->uc_lock, MA_OWNED);
436 KASSERT(uc->uc_busy != 0, ("not busy"));
437 uc->uc_busy = 0;
438 if (uc->uc_waiters)
439 wakeup_one(uc);
440 }
441
442 void
umtxq_unbusy_unlocked(struct umtx_key * key)443 umtxq_unbusy_unlocked(struct umtx_key *key)
444 {
445
446 umtxq_lock(key);
447 umtxq_unbusy(key);
448 umtxq_unlock(key);
449 }
450
451 static struct umtxq_queue *
umtxq_queue_lookup(struct umtx_key * key,int q)452 umtxq_queue_lookup(struct umtx_key *key, int q)
453 {
454 struct umtxq_queue *uh;
455 struct umtxq_chain *uc;
456
457 uc = umtxq_getchain(key);
458 UMTXQ_LOCKED_ASSERT(uc);
459 LIST_FOREACH(uh, &uc->uc_queue[q], link) {
460 if (umtx_key_match(&uh->key, key))
461 return (uh);
462 }
463
464 return (NULL);
465 }
466
467 void
umtxq_insert_queue(struct umtx_q * uq,int q)468 umtxq_insert_queue(struct umtx_q *uq, int q)
469 {
470 struct umtxq_queue *uh;
471 struct umtxq_chain *uc;
472
473 uc = umtxq_getchain(&uq->uq_key);
474 UMTXQ_LOCKED_ASSERT(uc);
475 KASSERT((uq->uq_flags & UQF_UMTXQ) == 0, ("umtx_q is already on queue"));
476 uh = umtxq_queue_lookup(&uq->uq_key, q);
477 if (uh != NULL) {
478 LIST_INSERT_HEAD(&uc->uc_spare_queue, uq->uq_spare_queue, link);
479 } else {
480 uh = uq->uq_spare_queue;
481 uh->key = uq->uq_key;
482 LIST_INSERT_HEAD(&uc->uc_queue[q], uh, link);
483 #ifdef UMTX_PROFILING
484 uc->length++;
485 if (uc->length > uc->max_length) {
486 uc->max_length = uc->length;
487 if (uc->max_length > max_length)
488 max_length = uc->max_length;
489 }
490 #endif
491 }
492 uq->uq_spare_queue = NULL;
493
494 TAILQ_INSERT_TAIL(&uh->head, uq, uq_link);
495 uh->length++;
496 uq->uq_flags |= UQF_UMTXQ;
497 uq->uq_cur_queue = uh;
498 return;
499 }
500
501 void
umtxq_remove_queue(struct umtx_q * uq,int q)502 umtxq_remove_queue(struct umtx_q *uq, int q)
503 {
504 struct umtxq_chain *uc;
505 struct umtxq_queue *uh;
506
507 uc = umtxq_getchain(&uq->uq_key);
508 UMTXQ_LOCKED_ASSERT(uc);
509 if (uq->uq_flags & UQF_UMTXQ) {
510 uh = uq->uq_cur_queue;
511 TAILQ_REMOVE(&uh->head, uq, uq_link);
512 uh->length--;
513 uq->uq_flags &= ~UQF_UMTXQ;
514 if (TAILQ_EMPTY(&uh->head)) {
515 KASSERT(uh->length == 0,
516 ("inconsistent umtxq_queue length"));
517 #ifdef UMTX_PROFILING
518 uc->length--;
519 #endif
520 LIST_REMOVE(uh, link);
521 } else {
522 uh = LIST_FIRST(&uc->uc_spare_queue);
523 KASSERT(uh != NULL, ("uc_spare_queue is empty"));
524 LIST_REMOVE(uh, link);
525 }
526 uq->uq_spare_queue = uh;
527 uq->uq_cur_queue = NULL;
528 }
529 }
530
531 /*
532 * Check if there are multiple waiters
533 */
534 int
umtxq_count(struct umtx_key * key)535 umtxq_count(struct umtx_key *key)
536 {
537 struct umtxq_queue *uh;
538
539 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
540 uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
541 if (uh != NULL)
542 return (uh->length);
543 return (0);
544 }
545
546 /*
547 * Check if there are multiple PI waiters and returns first
548 * waiter.
549 */
550 static int
umtxq_count_pi(struct umtx_key * key,struct umtx_q ** first)551 umtxq_count_pi(struct umtx_key *key, struct umtx_q **first)
552 {
553 struct umtxq_queue *uh;
554
555 *first = NULL;
556 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
557 uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
558 if (uh != NULL) {
559 *first = TAILQ_FIRST(&uh->head);
560 return (uh->length);
561 }
562 return (0);
563 }
564
565 /*
566 * Wake up threads waiting on an userland object by a bit mask.
567 */
568 int
umtxq_signal_mask(struct umtx_key * key,int n_wake,u_int bitset)569 umtxq_signal_mask(struct umtx_key *key, int n_wake, u_int bitset)
570 {
571 struct umtxq_queue *uh;
572 struct umtx_q *uq, *uq_temp;
573 int ret;
574
575 ret = 0;
576 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
577 uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
578 if (uh == NULL)
579 return (0);
580 TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
581 if ((uq->uq_bitset & bitset) == 0)
582 continue;
583 umtxq_remove_queue(uq, UMTX_SHARED_QUEUE);
584 wakeup_one(uq);
585 if (++ret >= n_wake)
586 break;
587 }
588 return (ret);
589 }
590
591 /*
592 * Wake up threads waiting on an userland object.
593 */
594
595 static int
umtxq_signal_queue(struct umtx_key * key,int n_wake,int q)596 umtxq_signal_queue(struct umtx_key *key, int n_wake, int q)
597 {
598 struct umtxq_queue *uh;
599 struct umtx_q *uq;
600 int ret;
601
602 ret = 0;
603 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
604 uh = umtxq_queue_lookup(key, q);
605 if (uh != NULL) {
606 while ((uq = TAILQ_FIRST(&uh->head)) != NULL) {
607 umtxq_remove_queue(uq, q);
608 wakeup(uq);
609 if (++ret >= n_wake)
610 return (ret);
611 }
612 }
613 return (ret);
614 }
615
616 /*
617 * Wake up specified thread.
618 */
619 static inline void
umtxq_signal_thread(struct umtx_q * uq)620 umtxq_signal_thread(struct umtx_q *uq)
621 {
622
623 UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
624 umtxq_remove(uq);
625 wakeup(uq);
626 }
627
628 /*
629 * Wake up a maximum of n_wake threads that are waiting on an userland
630 * object identified by key. The remaining threads are removed from queue
631 * identified by key and added to the queue identified by key2 (requeued).
632 * The n_requeue specifies an upper limit on the number of threads that
633 * are requeued to the second queue.
634 */
635 int
umtxq_requeue(struct umtx_key * key,int n_wake,struct umtx_key * key2,int n_requeue)636 umtxq_requeue(struct umtx_key *key, int n_wake, struct umtx_key *key2,
637 int n_requeue)
638 {
639 struct umtxq_queue *uh;
640 struct umtx_q *uq, *uq_temp;
641 int ret;
642
643 ret = 0;
644 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key));
645 UMTXQ_LOCKED_ASSERT(umtxq_getchain(key2));
646 uh = umtxq_queue_lookup(key, UMTX_SHARED_QUEUE);
647 if (uh == NULL)
648 return (0);
649 TAILQ_FOREACH_SAFE(uq, &uh->head, uq_link, uq_temp) {
650 if (++ret <= n_wake) {
651 umtxq_remove(uq);
652 wakeup_one(uq);
653 } else {
654 umtxq_remove(uq);
655 uq->uq_key = *key2;
656 umtxq_insert(uq);
657 if (ret - n_wake == n_requeue)
658 break;
659 }
660 }
661 return (ret);
662 }
663
664 static inline int
tstohz(const struct timespec * tsp)665 tstohz(const struct timespec *tsp)
666 {
667 struct timeval tv;
668
669 TIMESPEC_TO_TIMEVAL(&tv, tsp);
670 return tvtohz(&tv);
671 }
672
673 void
umtx_abs_timeout_init(struct umtx_abs_timeout * timo,int clockid,int absolute,const struct timespec * timeout)674 umtx_abs_timeout_init(struct umtx_abs_timeout *timo, int clockid,
675 int absolute, const struct timespec *timeout)
676 {
677
678 timo->clockid = clockid;
679 if (!absolute) {
680 timo->is_abs_real = false;
681 kern_clock_gettime(curthread, timo->clockid, &timo->cur);
682 timespecadd(&timo->cur, timeout, &timo->end);
683 } else {
684 timo->end = *timeout;
685 timo->is_abs_real = clockid == CLOCK_REALTIME ||
686 clockid == CLOCK_REALTIME_FAST ||
687 clockid == CLOCK_REALTIME_PRECISE ||
688 clockid == CLOCK_SECOND;
689 }
690 }
691
692 static void
umtx_abs_timeout_init2(struct umtx_abs_timeout * timo,const struct _umtx_time * umtxtime)693 umtx_abs_timeout_init2(struct umtx_abs_timeout *timo,
694 const struct _umtx_time *umtxtime)
695 {
696
697 umtx_abs_timeout_init(timo, umtxtime->_clockid,
698 (umtxtime->_flags & UMTX_ABSTIME) != 0, &umtxtime->_timeout);
699 }
700
701 static void
umtx_abs_timeout_enforce_min(sbintime_t * sbt)702 umtx_abs_timeout_enforce_min(sbintime_t *sbt)
703 {
704 sbintime_t when, mint;
705
706 mint = curproc->p_umtx_min_timeout;
707 if (__predict_false(mint != 0)) {
708 when = sbinuptime() + mint;
709 if (*sbt < when)
710 *sbt = when;
711 }
712 }
713
714 static int
umtx_abs_timeout_getsbt(struct umtx_abs_timeout * timo,sbintime_t * sbt,int * flags)715 umtx_abs_timeout_getsbt(struct umtx_abs_timeout *timo, sbintime_t *sbt,
716 int *flags)
717 {
718 struct bintime bt, bbt;
719 struct timespec tts;
720 sbintime_t rem;
721
722 switch (timo->clockid) {
723
724 /* Clocks that can be converted into absolute time. */
725 case CLOCK_REALTIME:
726 case CLOCK_REALTIME_PRECISE:
727 case CLOCK_REALTIME_FAST:
728 case CLOCK_MONOTONIC:
729 case CLOCK_MONOTONIC_PRECISE:
730 case CLOCK_MONOTONIC_FAST:
731 case CLOCK_UPTIME:
732 case CLOCK_UPTIME_PRECISE:
733 case CLOCK_UPTIME_FAST:
734 case CLOCK_SECOND:
735 timespec2bintime(&timo->end, &bt);
736 switch (timo->clockid) {
737 case CLOCK_REALTIME:
738 case CLOCK_REALTIME_PRECISE:
739 case CLOCK_REALTIME_FAST:
740 case CLOCK_SECOND:
741 getboottimebin(&bbt);
742 bintime_sub(&bt, &bbt);
743 break;
744 }
745 if (bt.sec < 0)
746 return (ETIMEDOUT);
747 if (bt.sec >= (SBT_MAX >> 32)) {
748 *sbt = 0;
749 *flags = 0;
750 return (0);
751 }
752 *sbt = bttosbt(bt);
753 umtx_abs_timeout_enforce_min(sbt);
754
755 /*
756 * Check if the absolute time should be aligned to
757 * avoid firing multiple timer events in non-periodic
758 * timer mode.
759 */
760 switch (timo->clockid) {
761 case CLOCK_REALTIME_FAST:
762 case CLOCK_MONOTONIC_FAST:
763 case CLOCK_UPTIME_FAST:
764 rem = *sbt % tc_tick_sbt;
765 if (__predict_true(rem != 0))
766 *sbt += tc_tick_sbt - rem;
767 break;
768 case CLOCK_SECOND:
769 rem = *sbt % SBT_1S;
770 if (__predict_true(rem != 0))
771 *sbt += SBT_1S - rem;
772 break;
773 }
774 *flags = C_ABSOLUTE;
775 return (0);
776
777 /* Clocks that has to be periodically polled. */
778 case CLOCK_VIRTUAL:
779 case CLOCK_PROF:
780 case CLOCK_THREAD_CPUTIME_ID:
781 case CLOCK_PROCESS_CPUTIME_ID:
782 default:
783 kern_clock_gettime(curthread, timo->clockid, &timo->cur);
784 if (timespeccmp(&timo->end, &timo->cur, <=))
785 return (ETIMEDOUT);
786 timespecsub(&timo->end, &timo->cur, &tts);
787 *sbt = tick_sbt * tstohz(&tts);
788 *flags = C_HARDCLOCK;
789 return (0);
790 }
791 }
792
793 static uint32_t
umtx_unlock_val(uint32_t flags,bool rb)794 umtx_unlock_val(uint32_t flags, bool rb)
795 {
796
797 if (rb)
798 return (UMUTEX_RB_OWNERDEAD);
799 else if ((flags & UMUTEX_NONCONSISTENT) != 0)
800 return (UMUTEX_RB_NOTRECOV);
801 else
802 return (UMUTEX_UNOWNED);
803
804 }
805
806 /*
807 * Put thread into sleep state, before sleeping, check if
808 * thread was removed from umtx queue.
809 */
810 int
umtxq_sleep(struct umtx_q * uq,const char * wmesg,struct umtx_abs_timeout * timo)811 umtxq_sleep(struct umtx_q *uq, const char *wmesg,
812 struct umtx_abs_timeout *timo)
813 {
814 struct umtxq_chain *uc;
815 sbintime_t sbt = 0;
816 int error, flags = 0;
817
818 uc = umtxq_getchain(&uq->uq_key);
819 UMTXQ_LOCKED_ASSERT(uc);
820 for (;;) {
821 if (!(uq->uq_flags & UQF_UMTXQ)) {
822 error = 0;
823 break;
824 }
825 if (timo != NULL) {
826 if (timo->is_abs_real)
827 curthread->td_rtcgen =
828 atomic_load_acq_int(&rtc_generation);
829 error = umtx_abs_timeout_getsbt(timo, &sbt, &flags);
830 if (error != 0)
831 break;
832 }
833 error = msleep_sbt(uq, &uc->uc_lock, PCATCH | PDROP, wmesg,
834 sbt, 0, flags);
835 uc = umtxq_getchain(&uq->uq_key);
836 mtx_lock(&uc->uc_lock);
837 if (error == EINTR || error == ERESTART)
838 break;
839 if (error == EWOULDBLOCK && (flags & C_ABSOLUTE) != 0) {
840 error = ETIMEDOUT;
841 break;
842 }
843 }
844
845 curthread->td_rtcgen = 0;
846 return (error);
847 }
848
849 /*
850 * Convert userspace address into unique logical address.
851 */
852 int
umtx_key_get(const void * addr,int type,int share,struct umtx_key * key)853 umtx_key_get(const void *addr, int type, int share, struct umtx_key *key)
854 {
855 struct thread *td = curthread;
856 vm_map_t map;
857 vm_map_entry_t entry;
858 vm_pindex_t pindex;
859 vm_prot_t prot;
860 boolean_t wired;
861
862 key->type = type;
863 if (share == THREAD_SHARE) {
864 key->shared = 0;
865 key->info.private.vs = td->td_proc->p_vmspace;
866 key->info.private.addr = (uintptr_t)addr;
867 } else {
868 MPASS(share == PROCESS_SHARE || share == AUTO_SHARE);
869 map = &td->td_proc->p_vmspace->vm_map;
870 if (vm_map_lookup(&map, (vm_offset_t)addr, VM_PROT_WRITE,
871 &entry, &key->info.shared.object, &pindex, &prot,
872 &wired) != KERN_SUCCESS) {
873 return (EFAULT);
874 }
875
876 if ((share == PROCESS_SHARE) ||
877 (share == AUTO_SHARE &&
878 VM_INHERIT_SHARE == entry->inheritance)) {
879 key->shared = 1;
880 key->info.shared.offset = (vm_offset_t)addr -
881 entry->start + entry->offset;
882 vm_object_reference(key->info.shared.object);
883 } else {
884 key->shared = 0;
885 key->info.private.vs = td->td_proc->p_vmspace;
886 key->info.private.addr = (uintptr_t)addr;
887 }
888 vm_map_lookup_done(map, entry);
889 }
890
891 umtxq_hash(key);
892 return (0);
893 }
894
895 /*
896 * Release key.
897 */
898 void
umtx_key_release(struct umtx_key * key)899 umtx_key_release(struct umtx_key *key)
900 {
901 if (key->shared)
902 vm_object_deallocate(key->info.shared.object);
903 }
904
905 #ifdef COMPAT_FREEBSD10
906 /*
907 * Lock a umtx object.
908 */
909 static int
do_lock_umtx(struct thread * td,struct umtx * umtx,u_long id,const struct timespec * timeout)910 do_lock_umtx(struct thread *td, struct umtx *umtx, u_long id,
911 const struct timespec *timeout)
912 {
913 struct umtx_abs_timeout timo;
914 struct umtx_q *uq;
915 u_long owner;
916 u_long old;
917 int error = 0;
918
919 uq = td->td_umtxq;
920 if (timeout != NULL)
921 umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
922
923 /*
924 * Care must be exercised when dealing with umtx structure. It
925 * can fault on any access.
926 */
927 for (;;) {
928 /*
929 * Try the uncontested case. This should be done in userland.
930 */
931 owner = casuword(&umtx->u_owner, UMTX_UNOWNED, id);
932
933 /* The acquire succeeded. */
934 if (owner == UMTX_UNOWNED)
935 return (0);
936
937 /* The address was invalid. */
938 if (owner == -1)
939 return (EFAULT);
940
941 /* If no one owns it but it is contested try to acquire it. */
942 if (owner == UMTX_CONTESTED) {
943 owner = casuword(&umtx->u_owner,
944 UMTX_CONTESTED, id | UMTX_CONTESTED);
945
946 if (owner == UMTX_CONTESTED)
947 return (0);
948
949 /* The address was invalid. */
950 if (owner == -1)
951 return (EFAULT);
952
953 error = thread_check_susp(td, false);
954 if (error != 0)
955 break;
956
957 /* If this failed the lock has changed, restart. */
958 continue;
959 }
960
961 /*
962 * If we caught a signal, we have retried and now
963 * exit immediately.
964 */
965 if (error != 0)
966 break;
967
968 if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK,
969 AUTO_SHARE, &uq->uq_key)) != 0)
970 return (error);
971
972 umtxq_lock(&uq->uq_key);
973 umtxq_busy(&uq->uq_key);
974 umtxq_insert(uq);
975 umtxq_unbusy(&uq->uq_key);
976 umtxq_unlock(&uq->uq_key);
977
978 /*
979 * Set the contested bit so that a release in user space
980 * knows to use the system call for unlock. If this fails
981 * either some one else has acquired the lock or it has been
982 * released.
983 */
984 old = casuword(&umtx->u_owner, owner, owner | UMTX_CONTESTED);
985
986 /* The address was invalid. */
987 if (old == -1) {
988 umtxq_lock(&uq->uq_key);
989 umtxq_remove(uq);
990 umtxq_unlock(&uq->uq_key);
991 umtx_key_release(&uq->uq_key);
992 return (EFAULT);
993 }
994
995 /*
996 * We set the contested bit, sleep. Otherwise the lock changed
997 * and we need to retry or we lost a race to the thread
998 * unlocking the umtx.
999 */
1000 umtxq_lock(&uq->uq_key);
1001 if (old == owner)
1002 error = umtxq_sleep(uq, "umtx", timeout == NULL ? NULL :
1003 &timo);
1004 umtxq_remove(uq);
1005 umtxq_unlock(&uq->uq_key);
1006 umtx_key_release(&uq->uq_key);
1007
1008 if (error == 0)
1009 error = thread_check_susp(td, false);
1010 }
1011
1012 if (timeout == NULL) {
1013 /* Mutex locking is restarted if it is interrupted. */
1014 if (error == EINTR)
1015 error = ERESTART;
1016 } else {
1017 /* Timed-locking is not restarted. */
1018 if (error == ERESTART)
1019 error = EINTR;
1020 }
1021 return (error);
1022 }
1023
1024 /*
1025 * Unlock a umtx object.
1026 */
1027 static int
do_unlock_umtx(struct thread * td,struct umtx * umtx,u_long id)1028 do_unlock_umtx(struct thread *td, struct umtx *umtx, u_long id)
1029 {
1030 struct umtx_key key;
1031 u_long owner;
1032 u_long old;
1033 int error;
1034 int count;
1035
1036 /*
1037 * Make sure we own this mtx.
1038 */
1039 owner = fuword(__DEVOLATILE(u_long *, &umtx->u_owner));
1040 if (owner == -1)
1041 return (EFAULT);
1042
1043 if ((owner & ~UMTX_CONTESTED) != id)
1044 return (EPERM);
1045
1046 /* This should be done in userland */
1047 if ((owner & UMTX_CONTESTED) == 0) {
1048 old = casuword(&umtx->u_owner, owner, UMTX_UNOWNED);
1049 if (old == -1)
1050 return (EFAULT);
1051 if (old == owner)
1052 return (0);
1053 owner = old;
1054 }
1055
1056 /* We should only ever be in here for contested locks */
1057 if ((error = umtx_key_get(umtx, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1058 &key)) != 0)
1059 return (error);
1060
1061 umtxq_lock(&key);
1062 umtxq_busy(&key);
1063 count = umtxq_count(&key);
1064 umtxq_unlock(&key);
1065
1066 /*
1067 * When unlocking the umtx, it must be marked as unowned if
1068 * there is zero or one thread only waiting for it.
1069 * Otherwise, it must be marked as contested.
1070 */
1071 old = casuword(&umtx->u_owner, owner,
1072 count <= 1 ? UMTX_UNOWNED : UMTX_CONTESTED);
1073 umtxq_lock(&key);
1074 umtxq_signal(&key,1);
1075 umtxq_unbusy(&key);
1076 umtxq_unlock(&key);
1077 umtx_key_release(&key);
1078 if (old == -1)
1079 return (EFAULT);
1080 if (old != owner)
1081 return (EINVAL);
1082 return (0);
1083 }
1084
1085 #ifdef COMPAT_FREEBSD32
1086
1087 /*
1088 * Lock a umtx object.
1089 */
1090 static int
do_lock_umtx32(struct thread * td,uint32_t * m,uint32_t id,const struct timespec * timeout)1091 do_lock_umtx32(struct thread *td, uint32_t *m, uint32_t id,
1092 const struct timespec *timeout)
1093 {
1094 struct umtx_abs_timeout timo;
1095 struct umtx_q *uq;
1096 uint32_t owner;
1097 uint32_t old;
1098 int error = 0;
1099
1100 uq = td->td_umtxq;
1101
1102 if (timeout != NULL)
1103 umtx_abs_timeout_init(&timo, CLOCK_REALTIME, 0, timeout);
1104
1105 /*
1106 * Care must be exercised when dealing with umtx structure. It
1107 * can fault on any access.
1108 */
1109 for (;;) {
1110 /*
1111 * Try the uncontested case. This should be done in userland.
1112 */
1113 owner = casuword32(m, UMUTEX_UNOWNED, id);
1114
1115 /* The acquire succeeded. */
1116 if (owner == UMUTEX_UNOWNED)
1117 return (0);
1118
1119 /* The address was invalid. */
1120 if (owner == -1)
1121 return (EFAULT);
1122
1123 /* If no one owns it but it is contested try to acquire it. */
1124 if (owner == UMUTEX_CONTESTED) {
1125 owner = casuword32(m,
1126 UMUTEX_CONTESTED, id | UMUTEX_CONTESTED);
1127 if (owner == UMUTEX_CONTESTED)
1128 return (0);
1129
1130 /* The address was invalid. */
1131 if (owner == -1)
1132 return (EFAULT);
1133
1134 error = thread_check_susp(td, false);
1135 if (error != 0)
1136 break;
1137
1138 /* If this failed the lock has changed, restart. */
1139 continue;
1140 }
1141
1142 /*
1143 * If we caught a signal, we have retried and now
1144 * exit immediately.
1145 */
1146 if (error != 0)
1147 return (error);
1148
1149 if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK,
1150 AUTO_SHARE, &uq->uq_key)) != 0)
1151 return (error);
1152
1153 umtxq_lock(&uq->uq_key);
1154 umtxq_busy(&uq->uq_key);
1155 umtxq_insert(uq);
1156 umtxq_unbusy(&uq->uq_key);
1157 umtxq_unlock(&uq->uq_key);
1158
1159 /*
1160 * Set the contested bit so that a release in user space
1161 * knows to use the system call for unlock. If this fails
1162 * either some one else has acquired the lock or it has been
1163 * released.
1164 */
1165 old = casuword32(m, owner, owner | UMUTEX_CONTESTED);
1166
1167 /* The address was invalid. */
1168 if (old == -1) {
1169 umtxq_lock(&uq->uq_key);
1170 umtxq_remove(uq);
1171 umtxq_unlock(&uq->uq_key);
1172 umtx_key_release(&uq->uq_key);
1173 return (EFAULT);
1174 }
1175
1176 /*
1177 * We set the contested bit, sleep. Otherwise the lock changed
1178 * and we need to retry or we lost a race to the thread
1179 * unlocking the umtx.
1180 */
1181 umtxq_lock(&uq->uq_key);
1182 if (old == owner)
1183 error = umtxq_sleep(uq, "umtx", timeout == NULL ?
1184 NULL : &timo);
1185 umtxq_remove(uq);
1186 umtxq_unlock(&uq->uq_key);
1187 umtx_key_release(&uq->uq_key);
1188
1189 if (error == 0)
1190 error = thread_check_susp(td, false);
1191 }
1192
1193 if (timeout == NULL) {
1194 /* Mutex locking is restarted if it is interrupted. */
1195 if (error == EINTR)
1196 error = ERESTART;
1197 } else {
1198 /* Timed-locking is not restarted. */
1199 if (error == ERESTART)
1200 error = EINTR;
1201 }
1202 return (error);
1203 }
1204
1205 /*
1206 * Unlock a umtx object.
1207 */
1208 static int
do_unlock_umtx32(struct thread * td,uint32_t * m,uint32_t id)1209 do_unlock_umtx32(struct thread *td, uint32_t *m, uint32_t id)
1210 {
1211 struct umtx_key key;
1212 uint32_t owner;
1213 uint32_t old;
1214 int error;
1215 int count;
1216
1217 /*
1218 * Make sure we own this mtx.
1219 */
1220 owner = fuword32(m);
1221 if (owner == -1)
1222 return (EFAULT);
1223
1224 if ((owner & ~UMUTEX_CONTESTED) != id)
1225 return (EPERM);
1226
1227 /* This should be done in userland */
1228 if ((owner & UMUTEX_CONTESTED) == 0) {
1229 old = casuword32(m, owner, UMUTEX_UNOWNED);
1230 if (old == -1)
1231 return (EFAULT);
1232 if (old == owner)
1233 return (0);
1234 owner = old;
1235 }
1236
1237 /* We should only ever be in here for contested locks */
1238 if ((error = umtx_key_get(m, TYPE_SIMPLE_LOCK, AUTO_SHARE,
1239 &key)) != 0)
1240 return (error);
1241
1242 umtxq_lock(&key);
1243 umtxq_busy(&key);
1244 count = umtxq_count(&key);
1245 umtxq_unlock(&key);
1246
1247 /*
1248 * When unlocking the umtx, it must be marked as unowned if
1249 * there is zero or one thread only waiting for it.
1250 * Otherwise, it must be marked as contested.
1251 */
1252 old = casuword32(m, owner,
1253 count <= 1 ? UMUTEX_UNOWNED : UMUTEX_CONTESTED);
1254 umtxq_lock(&key);
1255 umtxq_signal(&key,1);
1256 umtxq_unbusy(&key);
1257 umtxq_unlock(&key);
1258 umtx_key_release(&key);
1259 if (old == -1)
1260 return (EFAULT);
1261 if (old != owner)
1262 return (EINVAL);
1263 return (0);
1264 }
1265 #endif /* COMPAT_FREEBSD32 */
1266 #endif /* COMPAT_FREEBSD10 */
1267
1268 /*
1269 * Fetch and compare value, sleep on the address if value is not changed.
1270 */
1271 static int
do_wait(struct thread * td,void * addr,u_long id,struct _umtx_time * timeout,int compat32,int is_private)1272 do_wait(struct thread *td, void *addr, u_long id,
1273 struct _umtx_time *timeout, int compat32, int is_private)
1274 {
1275 struct umtx_abs_timeout timo;
1276 struct umtx_q *uq;
1277 u_long tmp;
1278 uint32_t tmp32;
1279 int error = 0;
1280
1281 uq = td->td_umtxq;
1282 if ((error = umtx_key_get(addr, TYPE_SIMPLE_WAIT,
1283 is_private ? THREAD_SHARE : AUTO_SHARE, &uq->uq_key)) != 0)
1284 return (error);
1285
1286 if (timeout != NULL)
1287 umtx_abs_timeout_init2(&timo, timeout);
1288
1289 umtxq_lock(&uq->uq_key);
1290 umtxq_insert(uq);
1291 umtxq_unlock(&uq->uq_key);
1292 if (compat32 == 0) {
1293 error = fueword(addr, &tmp);
1294 if (error != 0)
1295 error = EFAULT;
1296 } else {
1297 error = fueword32(addr, &tmp32);
1298 if (error == 0)
1299 tmp = tmp32;
1300 else
1301 error = EFAULT;
1302 }
1303 umtxq_lock(&uq->uq_key);
1304 if (error == 0) {
1305 if (tmp == id)
1306 error = umtxq_sleep(uq, "uwait", timeout == NULL ?
1307 NULL : &timo);
1308 if ((uq->uq_flags & UQF_UMTXQ) == 0)
1309 error = 0;
1310 else
1311 umtxq_remove(uq);
1312 } else if ((uq->uq_flags & UQF_UMTXQ) != 0) {
1313 umtxq_remove(uq);
1314 }
1315 umtxq_unlock(&uq->uq_key);
1316 umtx_key_release(&uq->uq_key);
1317 if (error == ERESTART)
1318 error = EINTR;
1319 return (error);
1320 }
1321
1322 /*
1323 * Wake up threads sleeping on the specified address.
1324 */
1325 int
kern_umtx_wake(struct thread * td,void * uaddr,int n_wake,int is_private)1326 kern_umtx_wake(struct thread *td, void *uaddr, int n_wake, int is_private)
1327 {
1328 struct umtx_key key;
1329 int ret;
1330
1331 if ((ret = umtx_key_get(uaddr, TYPE_SIMPLE_WAIT,
1332 is_private ? THREAD_SHARE : AUTO_SHARE, &key)) != 0)
1333 return (ret);
1334 umtxq_lock(&key);
1335 umtxq_signal(&key, n_wake);
1336 umtxq_unlock(&key);
1337 umtx_key_release(&key);
1338 return (0);
1339 }
1340
1341 /*
1342 * Lock PTHREAD_PRIO_NONE protocol POSIX mutex.
1343 */
1344 static int
do_lock_normal(struct thread * td,struct umutex * m,uint32_t flags,struct _umtx_time * timeout,int mode)1345 do_lock_normal(struct thread *td, struct umutex *m, uint32_t flags,
1346 struct _umtx_time *timeout, int mode)
1347 {
1348 struct umtx_abs_timeout timo;
1349 struct umtx_q *uq;
1350 uint32_t owner, old, id;
1351 int error, rv;
1352
1353 id = td->td_tid;
1354 uq = td->td_umtxq;
1355 error = 0;
1356 if (timeout != NULL)
1357 umtx_abs_timeout_init2(&timo, timeout);
1358
1359 /*
1360 * Care must be exercised when dealing with umtx structure. It
1361 * can fault on any access.
1362 */
1363 for (;;) {
1364 rv = fueword32(&m->m_owner, &owner);
1365 if (rv == -1)
1366 return (EFAULT);
1367 if (mode == _UMUTEX_WAIT) {
1368 if (owner == UMUTEX_UNOWNED ||
1369 owner == UMUTEX_CONTESTED ||
1370 owner == UMUTEX_RB_OWNERDEAD ||
1371 owner == UMUTEX_RB_NOTRECOV)
1372 return (0);
1373 } else {
1374 /*
1375 * Robust mutex terminated. Kernel duty is to
1376 * return EOWNERDEAD to the userspace. The
1377 * umutex.m_flags UMUTEX_NONCONSISTENT is set
1378 * by the common userspace code.
1379 */
1380 if (owner == UMUTEX_RB_OWNERDEAD) {
1381 rv = casueword32(&m->m_owner,
1382 UMUTEX_RB_OWNERDEAD, &owner,
1383 id | UMUTEX_CONTESTED);
1384 if (rv == -1)
1385 return (EFAULT);
1386 if (rv == 0) {
1387 MPASS(owner == UMUTEX_RB_OWNERDEAD);
1388 return (EOWNERDEAD); /* success */
1389 }
1390 MPASS(rv == 1);
1391 rv = thread_check_susp(td, false);
1392 if (rv != 0)
1393 return (rv);
1394 continue;
1395 }
1396 if (owner == UMUTEX_RB_NOTRECOV)
1397 return (ENOTRECOVERABLE);
1398
1399 /*
1400 * Try the uncontested case. This should be
1401 * done in userland.
1402 */
1403 rv = casueword32(&m->m_owner, UMUTEX_UNOWNED,
1404 &owner, id);
1405 /* The address was invalid. */
1406 if (rv == -1)
1407 return (EFAULT);
1408
1409 /* The acquire succeeded. */
1410 if (rv == 0) {
1411 MPASS(owner == UMUTEX_UNOWNED);
1412 return (0);
1413 }
1414
1415 /*
1416 * If no one owns it but it is contested try
1417 * to acquire it.
1418 */
1419 MPASS(rv == 1);
1420 if (owner == UMUTEX_CONTESTED) {
1421 rv = casueword32(&m->m_owner,
1422 UMUTEX_CONTESTED, &owner,
1423 id | UMUTEX_CONTESTED);
1424 /* The address was invalid. */
1425 if (rv == -1)
1426 return (EFAULT);
1427 if (rv == 0) {
1428 MPASS(owner == UMUTEX_CONTESTED);
1429 return (0);
1430 }
1431 if (rv == 1) {
1432 rv = thread_check_susp(td, false);
1433 if (rv != 0)
1434 return (rv);
1435 }
1436
1437 /*
1438 * If this failed the lock has
1439 * changed, restart.
1440 */
1441 continue;
1442 }
1443
1444 /* rv == 1 but not contested, likely store failure */
1445 rv = thread_check_susp(td, false);
1446 if (rv != 0)
1447 return (rv);
1448 }
1449
1450 if (mode == _UMUTEX_TRY)
1451 return (EBUSY);
1452
1453 /*
1454 * If we caught a signal, we have retried and now
1455 * exit immediately.
1456 */
1457 if (error != 0)
1458 return (error);
1459
1460 if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX,
1461 GET_SHARE(flags), &uq->uq_key)) != 0)
1462 return (error);
1463
1464 umtxq_lock(&uq->uq_key);
1465 umtxq_busy(&uq->uq_key);
1466 umtxq_insert(uq);
1467 umtxq_unlock(&uq->uq_key);
1468
1469 /*
1470 * Set the contested bit so that a release in user space
1471 * knows to use the system call for unlock. If this fails
1472 * either some one else has acquired the lock or it has been
1473 * released.
1474 */
1475 rv = casueword32(&m->m_owner, owner, &old,
1476 owner | UMUTEX_CONTESTED);
1477
1478 /* The address was invalid or casueword failed to store. */
1479 if (rv == -1 || rv == 1) {
1480 umtxq_lock(&uq->uq_key);
1481 umtxq_remove(uq);
1482 umtxq_unbusy(&uq->uq_key);
1483 umtxq_unlock(&uq->uq_key);
1484 umtx_key_release(&uq->uq_key);
1485 if (rv == -1)
1486 return (EFAULT);
1487 if (rv == 1) {
1488 rv = thread_check_susp(td, false);
1489 if (rv != 0)
1490 return (rv);
1491 }
1492 continue;
1493 }
1494
1495 /*
1496 * We set the contested bit, sleep. Otherwise the lock changed
1497 * and we need to retry or we lost a race to the thread
1498 * unlocking the umtx.
1499 */
1500 umtxq_lock(&uq->uq_key);
1501 umtxq_unbusy(&uq->uq_key);
1502 MPASS(old == owner);
1503 error = umtxq_sleep(uq, "umtxn", timeout == NULL ?
1504 NULL : &timo);
1505 umtxq_remove(uq);
1506 umtxq_unlock(&uq->uq_key);
1507 umtx_key_release(&uq->uq_key);
1508
1509 if (error == 0)
1510 error = thread_check_susp(td, false);
1511 }
1512
1513 return (0);
1514 }
1515
1516 /*
1517 * Unlock PTHREAD_PRIO_NONE protocol POSIX mutex.
1518 */
1519 static int
do_unlock_normal(struct thread * td,struct umutex * m,uint32_t flags,bool rb)1520 do_unlock_normal(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
1521 {
1522 struct umtx_key key;
1523 uint32_t owner, old, id, newlock;
1524 int error, count;
1525
1526 id = td->td_tid;
1527
1528 again:
1529 /*
1530 * Make sure we own this mtx.
1531 */
1532 error = fueword32(&m->m_owner, &owner);
1533 if (error == -1)
1534 return (EFAULT);
1535
1536 if ((owner & ~UMUTEX_CONTESTED) != id)
1537 return (EPERM);
1538
1539 newlock = umtx_unlock_val(flags, rb);
1540 if ((owner & UMUTEX_CONTESTED) == 0) {
1541 error = casueword32(&m->m_owner, owner, &old, newlock);
1542 if (error == -1)
1543 return (EFAULT);
1544 if (error == 1) {
1545 error = thread_check_susp(td, false);
1546 if (error != 0)
1547 return (error);
1548 goto again;
1549 }
1550 MPASS(old == owner);
1551 return (0);
1552 }
1553
1554 /* We should only ever be in here for contested locks */
1555 if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1556 &key)) != 0)
1557 return (error);
1558
1559 umtxq_lock(&key);
1560 umtxq_busy(&key);
1561 count = umtxq_count(&key);
1562 umtxq_unlock(&key);
1563
1564 /*
1565 * When unlocking the umtx, it must be marked as unowned if
1566 * there is zero or one thread only waiting for it.
1567 * Otherwise, it must be marked as contested.
1568 */
1569 if (count > 1)
1570 newlock |= UMUTEX_CONTESTED;
1571 error = casueword32(&m->m_owner, owner, &old, newlock);
1572 umtxq_lock(&key);
1573 umtxq_signal(&key, 1);
1574 umtxq_unbusy(&key);
1575 umtxq_unlock(&key);
1576 umtx_key_release(&key);
1577 if (error == -1)
1578 return (EFAULT);
1579 if (error == 1) {
1580 if (old != owner)
1581 return (EINVAL);
1582 error = thread_check_susp(td, false);
1583 if (error != 0)
1584 return (error);
1585 goto again;
1586 }
1587 return (0);
1588 }
1589
1590 /*
1591 * Check if the mutex is available and wake up a waiter,
1592 * only for simple mutex.
1593 */
1594 static int
do_wake_umutex(struct thread * td,struct umutex * m)1595 do_wake_umutex(struct thread *td, struct umutex *m)
1596 {
1597 struct umtx_key key;
1598 uint32_t owner;
1599 uint32_t flags;
1600 int error;
1601 int count;
1602
1603 again:
1604 error = fueword32(&m->m_owner, &owner);
1605 if (error == -1)
1606 return (EFAULT);
1607
1608 if ((owner & ~UMUTEX_CONTESTED) != 0 && owner != UMUTEX_RB_OWNERDEAD &&
1609 owner != UMUTEX_RB_NOTRECOV)
1610 return (0);
1611
1612 error = fueword32(&m->m_flags, &flags);
1613 if (error == -1)
1614 return (EFAULT);
1615
1616 /* We should only ever be in here for contested locks */
1617 if ((error = umtx_key_get(m, TYPE_NORMAL_UMUTEX, GET_SHARE(flags),
1618 &key)) != 0)
1619 return (error);
1620
1621 umtxq_lock(&key);
1622 umtxq_busy(&key);
1623 count = umtxq_count(&key);
1624 umtxq_unlock(&key);
1625
1626 if (count <= 1 && owner != UMUTEX_RB_OWNERDEAD &&
1627 owner != UMUTEX_RB_NOTRECOV) {
1628 error = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
1629 UMUTEX_UNOWNED);
1630 if (error == -1) {
1631 error = EFAULT;
1632 } else if (error == 1) {
1633 umtxq_lock(&key);
1634 umtxq_unbusy(&key);
1635 umtxq_unlock(&key);
1636 umtx_key_release(&key);
1637 error = thread_check_susp(td, false);
1638 if (error != 0)
1639 return (error);
1640 goto again;
1641 }
1642 }
1643
1644 umtxq_lock(&key);
1645 if (error == 0 && count != 0) {
1646 MPASS((owner & ~UMUTEX_CONTESTED) == 0 ||
1647 owner == UMUTEX_RB_OWNERDEAD ||
1648 owner == UMUTEX_RB_NOTRECOV);
1649 umtxq_signal(&key, 1);
1650 }
1651 umtxq_unbusy(&key);
1652 umtxq_unlock(&key);
1653 umtx_key_release(&key);
1654 return (error);
1655 }
1656
1657 /*
1658 * Check if the mutex has waiters and tries to fix contention bit.
1659 */
1660 static int
do_wake2_umutex(struct thread * td,struct umutex * m,uint32_t flags)1661 do_wake2_umutex(struct thread *td, struct umutex *m, uint32_t flags)
1662 {
1663 struct umtx_key key;
1664 uint32_t owner, old;
1665 int type;
1666 int error;
1667 int count;
1668
1669 switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT |
1670 UMUTEX_ROBUST)) {
1671 case 0:
1672 case UMUTEX_ROBUST:
1673 type = TYPE_NORMAL_UMUTEX;
1674 break;
1675 case UMUTEX_PRIO_INHERIT:
1676 type = TYPE_PI_UMUTEX;
1677 break;
1678 case (UMUTEX_PRIO_INHERIT | UMUTEX_ROBUST):
1679 type = TYPE_PI_ROBUST_UMUTEX;
1680 break;
1681 case UMUTEX_PRIO_PROTECT:
1682 type = TYPE_PP_UMUTEX;
1683 break;
1684 case (UMUTEX_PRIO_PROTECT | UMUTEX_ROBUST):
1685 type = TYPE_PP_ROBUST_UMUTEX;
1686 break;
1687 default:
1688 return (EINVAL);
1689 }
1690 if ((error = umtx_key_get(m, type, GET_SHARE(flags), &key)) != 0)
1691 return (error);
1692
1693 owner = 0;
1694 umtxq_lock(&key);
1695 umtxq_busy(&key);
1696 count = umtxq_count(&key);
1697 umtxq_unlock(&key);
1698
1699 error = fueword32(&m->m_owner, &owner);
1700 if (error == -1)
1701 error = EFAULT;
1702
1703 /*
1704 * Only repair contention bit if there is a waiter, this means
1705 * the mutex is still being referenced by userland code,
1706 * otherwise don't update any memory.
1707 */
1708 while (error == 0 && (owner & UMUTEX_CONTESTED) == 0 &&
1709 (count > 1 || (count == 1 && (owner & ~UMUTEX_CONTESTED) != 0))) {
1710 error = casueword32(&m->m_owner, owner, &old,
1711 owner | UMUTEX_CONTESTED);
1712 if (error == -1) {
1713 error = EFAULT;
1714 break;
1715 }
1716 if (error == 0) {
1717 MPASS(old == owner);
1718 break;
1719 }
1720 owner = old;
1721 error = thread_check_susp(td, false);
1722 }
1723
1724 umtxq_lock(&key);
1725 if (error == EFAULT) {
1726 umtxq_signal(&key, INT_MAX);
1727 } else if (count != 0 && ((owner & ~UMUTEX_CONTESTED) == 0 ||
1728 owner == UMUTEX_RB_OWNERDEAD || owner == UMUTEX_RB_NOTRECOV))
1729 umtxq_signal(&key, 1);
1730 umtxq_unbusy(&key);
1731 umtxq_unlock(&key);
1732 umtx_key_release(&key);
1733 return (error);
1734 }
1735
1736 struct umtx_pi *
umtx_pi_alloc(int flags)1737 umtx_pi_alloc(int flags)
1738 {
1739 struct umtx_pi *pi;
1740
1741 pi = uma_zalloc(umtx_pi_zone, M_ZERO | flags);
1742 TAILQ_INIT(&pi->pi_blocked);
1743 atomic_add_int(&umtx_pi_allocated, 1);
1744 return (pi);
1745 }
1746
1747 void
umtx_pi_free(struct umtx_pi * pi)1748 umtx_pi_free(struct umtx_pi *pi)
1749 {
1750 uma_zfree(umtx_pi_zone, pi);
1751 atomic_add_int(&umtx_pi_allocated, -1);
1752 }
1753
1754 /*
1755 * Adjust the thread's position on a pi_state after its priority has been
1756 * changed.
1757 */
1758 static int
umtx_pi_adjust_thread(struct umtx_pi * pi,struct thread * td)1759 umtx_pi_adjust_thread(struct umtx_pi *pi, struct thread *td)
1760 {
1761 struct umtx_q *uq, *uq1, *uq2;
1762 struct thread *td1;
1763
1764 mtx_assert(&umtx_lock, MA_OWNED);
1765 if (pi == NULL)
1766 return (0);
1767
1768 uq = td->td_umtxq;
1769
1770 /*
1771 * Check if the thread needs to be moved on the blocked chain.
1772 * It needs to be moved if either its priority is lower than
1773 * the previous thread or higher than the next thread.
1774 */
1775 uq1 = TAILQ_PREV(uq, umtxq_head, uq_lockq);
1776 uq2 = TAILQ_NEXT(uq, uq_lockq);
1777 if ((uq1 != NULL && UPRI(td) < UPRI(uq1->uq_thread)) ||
1778 (uq2 != NULL && UPRI(td) > UPRI(uq2->uq_thread))) {
1779 /*
1780 * Remove thread from blocked chain and determine where
1781 * it should be moved to.
1782 */
1783 TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
1784 TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
1785 td1 = uq1->uq_thread;
1786 MPASS(td1->td_proc->p_magic == P_MAGIC);
1787 if (UPRI(td1) > UPRI(td))
1788 break;
1789 }
1790
1791 if (uq1 == NULL)
1792 TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
1793 else
1794 TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
1795 }
1796 return (1);
1797 }
1798
1799 static struct umtx_pi *
umtx_pi_next(struct umtx_pi * pi)1800 umtx_pi_next(struct umtx_pi *pi)
1801 {
1802 struct umtx_q *uq_owner;
1803
1804 if (pi->pi_owner == NULL)
1805 return (NULL);
1806 uq_owner = pi->pi_owner->td_umtxq;
1807 if (uq_owner == NULL)
1808 return (NULL);
1809 return (uq_owner->uq_pi_blocked);
1810 }
1811
1812 /*
1813 * Floyd's Cycle-Finding Algorithm.
1814 */
1815 static bool
umtx_pi_check_loop(struct umtx_pi * pi)1816 umtx_pi_check_loop(struct umtx_pi *pi)
1817 {
1818 struct umtx_pi *pi1; /* fast iterator */
1819
1820 mtx_assert(&umtx_lock, MA_OWNED);
1821 if (pi == NULL)
1822 return (false);
1823 pi1 = pi;
1824 for (;;) {
1825 pi = umtx_pi_next(pi);
1826 if (pi == NULL)
1827 break;
1828 pi1 = umtx_pi_next(pi1);
1829 if (pi1 == NULL)
1830 break;
1831 pi1 = umtx_pi_next(pi1);
1832 if (pi1 == NULL)
1833 break;
1834 if (pi == pi1)
1835 return (true);
1836 }
1837 return (false);
1838 }
1839
1840 /*
1841 * Propagate priority when a thread is blocked on POSIX
1842 * PI mutex.
1843 */
1844 static void
umtx_propagate_priority(struct thread * td)1845 umtx_propagate_priority(struct thread *td)
1846 {
1847 struct umtx_q *uq;
1848 struct umtx_pi *pi;
1849 int pri;
1850
1851 mtx_assert(&umtx_lock, MA_OWNED);
1852 pri = UPRI(td);
1853 uq = td->td_umtxq;
1854 pi = uq->uq_pi_blocked;
1855 if (pi == NULL)
1856 return;
1857 if (umtx_pi_check_loop(pi))
1858 return;
1859
1860 for (;;) {
1861 td = pi->pi_owner;
1862 if (td == NULL || td == curthread)
1863 return;
1864
1865 MPASS(td->td_proc != NULL);
1866 MPASS(td->td_proc->p_magic == P_MAGIC);
1867
1868 thread_lock(td);
1869 if (td->td_lend_user_pri > pri)
1870 sched_lend_user_prio(td, pri);
1871 else {
1872 thread_unlock(td);
1873 break;
1874 }
1875 thread_unlock(td);
1876
1877 /*
1878 * Pick up the lock that td is blocked on.
1879 */
1880 uq = td->td_umtxq;
1881 pi = uq->uq_pi_blocked;
1882 if (pi == NULL)
1883 break;
1884 /* Resort td on the list if needed. */
1885 umtx_pi_adjust_thread(pi, td);
1886 }
1887 }
1888
1889 /*
1890 * Unpropagate priority for a PI mutex when a thread blocked on
1891 * it is interrupted by signal or resumed by others.
1892 */
1893 static void
umtx_repropagate_priority(struct umtx_pi * pi)1894 umtx_repropagate_priority(struct umtx_pi *pi)
1895 {
1896 struct umtx_q *uq, *uq_owner;
1897 struct umtx_pi *pi2;
1898 int pri;
1899
1900 mtx_assert(&umtx_lock, MA_OWNED);
1901
1902 if (umtx_pi_check_loop(pi))
1903 return;
1904 while (pi != NULL && pi->pi_owner != NULL) {
1905 pri = PRI_MAX;
1906 uq_owner = pi->pi_owner->td_umtxq;
1907
1908 TAILQ_FOREACH(pi2, &uq_owner->uq_pi_contested, pi_link) {
1909 uq = TAILQ_FIRST(&pi2->pi_blocked);
1910 if (uq != NULL) {
1911 if (pri > UPRI(uq->uq_thread))
1912 pri = UPRI(uq->uq_thread);
1913 }
1914 }
1915
1916 if (pri > uq_owner->uq_inherited_pri)
1917 pri = uq_owner->uq_inherited_pri;
1918 thread_lock(pi->pi_owner);
1919 sched_lend_user_prio(pi->pi_owner, pri);
1920 thread_unlock(pi->pi_owner);
1921 if ((pi = uq_owner->uq_pi_blocked) != NULL)
1922 umtx_pi_adjust_thread(pi, uq_owner->uq_thread);
1923 }
1924 }
1925
1926 /*
1927 * Insert a PI mutex into owned list.
1928 */
1929 static void
umtx_pi_setowner(struct umtx_pi * pi,struct thread * owner)1930 umtx_pi_setowner(struct umtx_pi *pi, struct thread *owner)
1931 {
1932 struct umtx_q *uq_owner;
1933
1934 uq_owner = owner->td_umtxq;
1935 mtx_assert(&umtx_lock, MA_OWNED);
1936 MPASS(pi->pi_owner == NULL);
1937 pi->pi_owner = owner;
1938 TAILQ_INSERT_TAIL(&uq_owner->uq_pi_contested, pi, pi_link);
1939 }
1940
1941 /*
1942 * Disown a PI mutex, and remove it from the owned list.
1943 */
1944 static void
umtx_pi_disown(struct umtx_pi * pi)1945 umtx_pi_disown(struct umtx_pi *pi)
1946 {
1947
1948 mtx_assert(&umtx_lock, MA_OWNED);
1949 TAILQ_REMOVE(&pi->pi_owner->td_umtxq->uq_pi_contested, pi, pi_link);
1950 pi->pi_owner = NULL;
1951 }
1952
1953 /*
1954 * Claim ownership of a PI mutex.
1955 */
1956 int
umtx_pi_claim(struct umtx_pi * pi,struct thread * owner)1957 umtx_pi_claim(struct umtx_pi *pi, struct thread *owner)
1958 {
1959 struct umtx_q *uq;
1960 int pri;
1961
1962 mtx_lock(&umtx_lock);
1963 if (pi->pi_owner == owner) {
1964 mtx_unlock(&umtx_lock);
1965 return (0);
1966 }
1967
1968 if (pi->pi_owner != NULL) {
1969 /*
1970 * userland may have already messed the mutex, sigh.
1971 */
1972 mtx_unlock(&umtx_lock);
1973 return (EPERM);
1974 }
1975 umtx_pi_setowner(pi, owner);
1976 uq = TAILQ_FIRST(&pi->pi_blocked);
1977 if (uq != NULL) {
1978 pri = UPRI(uq->uq_thread);
1979 thread_lock(owner);
1980 if (pri < UPRI(owner))
1981 sched_lend_user_prio(owner, pri);
1982 thread_unlock(owner);
1983 }
1984 mtx_unlock(&umtx_lock);
1985 return (0);
1986 }
1987
1988 /*
1989 * Adjust a thread's order position in its blocked PI mutex,
1990 * this may result new priority propagating process.
1991 */
1992 void
umtx_pi_adjust(struct thread * td,u_char oldpri)1993 umtx_pi_adjust(struct thread *td, u_char oldpri)
1994 {
1995 struct umtx_q *uq;
1996 struct umtx_pi *pi;
1997
1998 uq = td->td_umtxq;
1999 mtx_lock(&umtx_lock);
2000 /*
2001 * Pick up the lock that td is blocked on.
2002 */
2003 pi = uq->uq_pi_blocked;
2004 if (pi != NULL) {
2005 umtx_pi_adjust_thread(pi, td);
2006 umtx_repropagate_priority(pi);
2007 }
2008 mtx_unlock(&umtx_lock);
2009 }
2010
2011 /*
2012 * Sleep on a PI mutex.
2013 */
2014 int
umtxq_sleep_pi(struct umtx_q * uq,struct umtx_pi * pi,uint32_t owner,const char * wmesg,struct umtx_abs_timeout * timo,bool shared)2015 umtxq_sleep_pi(struct umtx_q *uq, struct umtx_pi *pi, uint32_t owner,
2016 const char *wmesg, struct umtx_abs_timeout *timo, bool shared)
2017 {
2018 struct thread *td, *td1;
2019 struct umtx_q *uq1;
2020 int error, pri;
2021 #ifdef INVARIANTS
2022 struct umtxq_chain *uc;
2023
2024 uc = umtxq_getchain(&pi->pi_key);
2025 #endif
2026 error = 0;
2027 td = uq->uq_thread;
2028 KASSERT(td == curthread, ("inconsistent uq_thread"));
2029 UMTXQ_LOCKED_ASSERT(umtxq_getchain(&uq->uq_key));
2030 KASSERT(uc->uc_busy != 0, ("umtx chain is not busy"));
2031 umtxq_insert(uq);
2032 mtx_lock(&umtx_lock);
2033 if (pi->pi_owner == NULL) {
2034 mtx_unlock(&umtx_lock);
2035 td1 = tdfind(owner, shared ? -1 : td->td_proc->p_pid);
2036 mtx_lock(&umtx_lock);
2037 if (td1 != NULL) {
2038 if (pi->pi_owner == NULL)
2039 umtx_pi_setowner(pi, td1);
2040 PROC_UNLOCK(td1->td_proc);
2041 }
2042 }
2043
2044 TAILQ_FOREACH(uq1, &pi->pi_blocked, uq_lockq) {
2045 pri = UPRI(uq1->uq_thread);
2046 if (pri > UPRI(td))
2047 break;
2048 }
2049
2050 if (uq1 != NULL)
2051 TAILQ_INSERT_BEFORE(uq1, uq, uq_lockq);
2052 else
2053 TAILQ_INSERT_TAIL(&pi->pi_blocked, uq, uq_lockq);
2054
2055 uq->uq_pi_blocked = pi;
2056 thread_lock(td);
2057 td->td_flags |= TDF_UPIBLOCKED;
2058 thread_unlock(td);
2059 umtx_propagate_priority(td);
2060 mtx_unlock(&umtx_lock);
2061 umtxq_unbusy(&uq->uq_key);
2062
2063 error = umtxq_sleep(uq, wmesg, timo);
2064 umtxq_remove(uq);
2065
2066 mtx_lock(&umtx_lock);
2067 uq->uq_pi_blocked = NULL;
2068 thread_lock(td);
2069 td->td_flags &= ~TDF_UPIBLOCKED;
2070 thread_unlock(td);
2071 TAILQ_REMOVE(&pi->pi_blocked, uq, uq_lockq);
2072 umtx_repropagate_priority(pi);
2073 mtx_unlock(&umtx_lock);
2074 umtxq_unlock(&uq->uq_key);
2075
2076 return (error);
2077 }
2078
2079 /*
2080 * Add reference count for a PI mutex.
2081 */
2082 void
umtx_pi_ref(struct umtx_pi * pi)2083 umtx_pi_ref(struct umtx_pi *pi)
2084 {
2085
2086 UMTXQ_LOCKED_ASSERT(umtxq_getchain(&pi->pi_key));
2087 pi->pi_refcount++;
2088 }
2089
2090 /*
2091 * Decrease reference count for a PI mutex, if the counter
2092 * is decreased to zero, its memory space is freed.
2093 */
2094 void
umtx_pi_unref(struct umtx_pi * pi)2095 umtx_pi_unref(struct umtx_pi *pi)
2096 {
2097 struct umtxq_chain *uc;
2098
2099 uc = umtxq_getchain(&pi->pi_key);
2100 UMTXQ_LOCKED_ASSERT(uc);
2101 KASSERT(pi->pi_refcount > 0, ("invalid reference count"));
2102 if (--pi->pi_refcount == 0) {
2103 mtx_lock(&umtx_lock);
2104 if (pi->pi_owner != NULL)
2105 umtx_pi_disown(pi);
2106 KASSERT(TAILQ_EMPTY(&pi->pi_blocked),
2107 ("blocked queue not empty"));
2108 mtx_unlock(&umtx_lock);
2109 TAILQ_REMOVE(&uc->uc_pi_list, pi, pi_hashlink);
2110 umtx_pi_free(pi);
2111 }
2112 }
2113
2114 /*
2115 * Find a PI mutex in hash table.
2116 */
2117 struct umtx_pi *
umtx_pi_lookup(struct umtx_key * key)2118 umtx_pi_lookup(struct umtx_key *key)
2119 {
2120 struct umtxq_chain *uc;
2121 struct umtx_pi *pi;
2122
2123 uc = umtxq_getchain(key);
2124 UMTXQ_LOCKED_ASSERT(uc);
2125
2126 TAILQ_FOREACH(pi, &uc->uc_pi_list, pi_hashlink) {
2127 if (umtx_key_match(&pi->pi_key, key)) {
2128 return (pi);
2129 }
2130 }
2131 return (NULL);
2132 }
2133
2134 /*
2135 * Insert a PI mutex into hash table.
2136 */
2137 void
umtx_pi_insert(struct umtx_pi * pi)2138 umtx_pi_insert(struct umtx_pi *pi)
2139 {
2140 struct umtxq_chain *uc;
2141
2142 uc = umtxq_getchain(&pi->pi_key);
2143 UMTXQ_LOCKED_ASSERT(uc);
2144 TAILQ_INSERT_TAIL(&uc->uc_pi_list, pi, pi_hashlink);
2145 }
2146
2147 /*
2148 * Drop a PI mutex and wakeup a top waiter.
2149 */
2150 int
umtx_pi_drop(struct thread * td,struct umtx_key * key,bool rb,int * count)2151 umtx_pi_drop(struct thread *td, struct umtx_key *key, bool rb, int *count)
2152 {
2153 struct umtx_q *uq_first, *uq_first2, *uq_me;
2154 struct umtx_pi *pi, *pi2;
2155 int pri;
2156
2157 UMTXQ_ASSERT_LOCKED_BUSY(key);
2158 *count = umtxq_count_pi(key, &uq_first);
2159 if (uq_first != NULL) {
2160 mtx_lock(&umtx_lock);
2161 pi = uq_first->uq_pi_blocked;
2162 KASSERT(pi != NULL, ("pi == NULL?"));
2163 if (pi->pi_owner != td && !(rb && pi->pi_owner == NULL)) {
2164 mtx_unlock(&umtx_lock);
2165 /* userland messed the mutex */
2166 return (EPERM);
2167 }
2168 uq_me = td->td_umtxq;
2169 if (pi->pi_owner == td)
2170 umtx_pi_disown(pi);
2171 /* get highest priority thread which is still sleeping. */
2172 uq_first = TAILQ_FIRST(&pi->pi_blocked);
2173 while (uq_first != NULL &&
2174 (uq_first->uq_flags & UQF_UMTXQ) == 0) {
2175 uq_first = TAILQ_NEXT(uq_first, uq_lockq);
2176 }
2177 pri = PRI_MAX;
2178 TAILQ_FOREACH(pi2, &uq_me->uq_pi_contested, pi_link) {
2179 uq_first2 = TAILQ_FIRST(&pi2->pi_blocked);
2180 if (uq_first2 != NULL) {
2181 if (pri > UPRI(uq_first2->uq_thread))
2182 pri = UPRI(uq_first2->uq_thread);
2183 }
2184 }
2185 thread_lock(td);
2186 sched_lend_user_prio(td, pri);
2187 thread_unlock(td);
2188 mtx_unlock(&umtx_lock);
2189 if (uq_first)
2190 umtxq_signal_thread(uq_first);
2191 } else {
2192 pi = umtx_pi_lookup(key);
2193 /*
2194 * A umtx_pi can exist if a signal or timeout removed the
2195 * last waiter from the umtxq, but there is still
2196 * a thread in do_lock_pi() holding the umtx_pi.
2197 */
2198 if (pi != NULL) {
2199 /*
2200 * The umtx_pi can be unowned, such as when a thread
2201 * has just entered do_lock_pi(), allocated the
2202 * umtx_pi, and unlocked the umtxq.
2203 * If the current thread owns it, it must disown it.
2204 */
2205 mtx_lock(&umtx_lock);
2206 if (pi->pi_owner == td)
2207 umtx_pi_disown(pi);
2208 mtx_unlock(&umtx_lock);
2209 }
2210 }
2211 return (0);
2212 }
2213
2214 /*
2215 * Lock a PI mutex.
2216 */
2217 static int
do_lock_pi(struct thread * td,struct umutex * m,uint32_t flags,struct _umtx_time * timeout,int try)2218 do_lock_pi(struct thread *td, struct umutex *m, uint32_t flags,
2219 struct _umtx_time *timeout, int try)
2220 {
2221 struct umtx_abs_timeout timo;
2222 struct umtx_q *uq;
2223 struct umtx_pi *pi, *new_pi;
2224 uint32_t id, old_owner, owner, old;
2225 int error, rv;
2226
2227 id = td->td_tid;
2228 uq = td->td_umtxq;
2229
2230 if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2231 TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2232 &uq->uq_key)) != 0)
2233 return (error);
2234
2235 if (timeout != NULL)
2236 umtx_abs_timeout_init2(&timo, timeout);
2237
2238 umtxq_lock(&uq->uq_key);
2239 pi = umtx_pi_lookup(&uq->uq_key);
2240 if (pi == NULL) {
2241 new_pi = umtx_pi_alloc(M_NOWAIT);
2242 if (new_pi == NULL) {
2243 umtxq_unlock(&uq->uq_key);
2244 new_pi = umtx_pi_alloc(M_WAITOK);
2245 umtxq_lock(&uq->uq_key);
2246 pi = umtx_pi_lookup(&uq->uq_key);
2247 if (pi != NULL) {
2248 umtx_pi_free(new_pi);
2249 new_pi = NULL;
2250 }
2251 }
2252 if (new_pi != NULL) {
2253 new_pi->pi_key = uq->uq_key;
2254 umtx_pi_insert(new_pi);
2255 pi = new_pi;
2256 }
2257 }
2258 umtx_pi_ref(pi);
2259 umtxq_unlock(&uq->uq_key);
2260
2261 /*
2262 * Care must be exercised when dealing with umtx structure. It
2263 * can fault on any access.
2264 */
2265 for (;;) {
2266 /*
2267 * Try the uncontested case. This should be done in userland.
2268 */
2269 rv = casueword32(&m->m_owner, UMUTEX_UNOWNED, &owner, id);
2270 /* The address was invalid. */
2271 if (rv == -1) {
2272 error = EFAULT;
2273 break;
2274 }
2275 /* The acquire succeeded. */
2276 if (rv == 0) {
2277 MPASS(owner == UMUTEX_UNOWNED);
2278 error = 0;
2279 break;
2280 }
2281
2282 if (owner == UMUTEX_RB_NOTRECOV) {
2283 error = ENOTRECOVERABLE;
2284 break;
2285 }
2286
2287 /*
2288 * Nobody owns it, but the acquire failed. This can happen
2289 * with ll/sc atomics.
2290 */
2291 if (owner == UMUTEX_UNOWNED) {
2292 error = thread_check_susp(td, true);
2293 if (error != 0)
2294 break;
2295 continue;
2296 }
2297
2298 /*
2299 * Avoid overwriting a possible error from sleep due
2300 * to the pending signal with suspension check result.
2301 */
2302 if (error == 0) {
2303 error = thread_check_susp(td, true);
2304 if (error != 0)
2305 break;
2306 }
2307
2308 /* If no one owns it but it is contested try to acquire it. */
2309 if (owner == UMUTEX_CONTESTED || owner == UMUTEX_RB_OWNERDEAD) {
2310 old_owner = owner;
2311 rv = casueword32(&m->m_owner, owner, &owner,
2312 id | UMUTEX_CONTESTED);
2313 /* The address was invalid. */
2314 if (rv == -1) {
2315 error = EFAULT;
2316 break;
2317 }
2318 if (rv == 1) {
2319 if (error == 0) {
2320 error = thread_check_susp(td, true);
2321 if (error != 0)
2322 break;
2323 }
2324
2325 /*
2326 * If this failed the lock could
2327 * changed, restart.
2328 */
2329 continue;
2330 }
2331
2332 MPASS(rv == 0);
2333 MPASS(owner == old_owner);
2334 umtxq_lock(&uq->uq_key);
2335 umtxq_busy(&uq->uq_key);
2336 error = umtx_pi_claim(pi, td);
2337 umtxq_unbusy(&uq->uq_key);
2338 umtxq_unlock(&uq->uq_key);
2339 if (error != 0) {
2340 /*
2341 * Since we're going to return an
2342 * error, restore the m_owner to its
2343 * previous, unowned state to avoid
2344 * compounding the problem.
2345 */
2346 (void)casuword32(&m->m_owner,
2347 id | UMUTEX_CONTESTED, old_owner);
2348 }
2349 if (error == 0 && old_owner == UMUTEX_RB_OWNERDEAD)
2350 error = EOWNERDEAD;
2351 break;
2352 }
2353
2354 if ((owner & ~UMUTEX_CONTESTED) == id) {
2355 error = EDEADLK;
2356 break;
2357 }
2358
2359 if (try != 0) {
2360 error = EBUSY;
2361 break;
2362 }
2363
2364 /*
2365 * If we caught a signal, we have retried and now
2366 * exit immediately.
2367 */
2368 if (error != 0)
2369 break;
2370
2371 umtxq_lock(&uq->uq_key);
2372 umtxq_busy(&uq->uq_key);
2373 umtxq_unlock(&uq->uq_key);
2374
2375 /*
2376 * Set the contested bit so that a release in user space
2377 * knows to use the system call for unlock. If this fails
2378 * either some one else has acquired the lock or it has been
2379 * released.
2380 */
2381 rv = casueword32(&m->m_owner, owner, &old, owner |
2382 UMUTEX_CONTESTED);
2383
2384 /* The address was invalid. */
2385 if (rv == -1) {
2386 umtxq_unbusy_unlocked(&uq->uq_key);
2387 error = EFAULT;
2388 break;
2389 }
2390 if (rv == 1) {
2391 umtxq_unbusy_unlocked(&uq->uq_key);
2392 error = thread_check_susp(td, true);
2393 if (error != 0)
2394 break;
2395
2396 /*
2397 * The lock changed and we need to retry or we
2398 * lost a race to the thread unlocking the
2399 * umtx. Note that the UMUTEX_RB_OWNERDEAD
2400 * value for owner is impossible there.
2401 */
2402 continue;
2403 }
2404
2405 umtxq_lock(&uq->uq_key);
2406
2407 /* We set the contested bit, sleep. */
2408 MPASS(old == owner);
2409 error = umtxq_sleep_pi(uq, pi, owner & ~UMUTEX_CONTESTED,
2410 "umtxpi", timeout == NULL ? NULL : &timo,
2411 (flags & USYNC_PROCESS_SHARED) != 0);
2412 if (error != 0)
2413 continue;
2414
2415 error = thread_check_susp(td, false);
2416 if (error != 0)
2417 break;
2418 }
2419
2420 umtxq_lock(&uq->uq_key);
2421 umtx_pi_unref(pi);
2422 umtxq_unlock(&uq->uq_key);
2423
2424 umtx_key_release(&uq->uq_key);
2425 return (error);
2426 }
2427
2428 /*
2429 * Unlock a PI mutex.
2430 */
2431 static int
do_unlock_pi(struct thread * td,struct umutex * m,uint32_t flags,bool rb)2432 do_unlock_pi(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2433 {
2434 struct umtx_key key;
2435 uint32_t id, new_owner, old, owner;
2436 int count, error;
2437
2438 id = td->td_tid;
2439
2440 usrloop:
2441 /*
2442 * Make sure we own this mtx.
2443 */
2444 error = fueword32(&m->m_owner, &owner);
2445 if (error == -1)
2446 return (EFAULT);
2447
2448 if ((owner & ~UMUTEX_CONTESTED) != id)
2449 return (EPERM);
2450
2451 new_owner = umtx_unlock_val(flags, rb);
2452
2453 /* This should be done in userland */
2454 if ((owner & UMUTEX_CONTESTED) == 0) {
2455 error = casueword32(&m->m_owner, owner, &old, new_owner);
2456 if (error == -1)
2457 return (EFAULT);
2458 if (error == 1) {
2459 error = thread_check_susp(td, true);
2460 if (error != 0)
2461 return (error);
2462 goto usrloop;
2463 }
2464 if (old == owner)
2465 return (0);
2466 owner = old;
2467 }
2468
2469 /* We should only ever be in here for contested locks */
2470 if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2471 TYPE_PI_ROBUST_UMUTEX : TYPE_PI_UMUTEX, GET_SHARE(flags),
2472 &key)) != 0)
2473 return (error);
2474
2475 umtxq_lock(&key);
2476 umtxq_busy(&key);
2477 error = umtx_pi_drop(td, &key, rb, &count);
2478 if (error != 0) {
2479 umtxq_unbusy(&key);
2480 umtxq_unlock(&key);
2481 umtx_key_release(&key);
2482 /* userland messed the mutex */
2483 return (error);
2484 }
2485 umtxq_unlock(&key);
2486
2487 /*
2488 * When unlocking the umtx, it must be marked as unowned if
2489 * there is zero or one thread only waiting for it.
2490 * Otherwise, it must be marked as contested.
2491 */
2492
2493 if (count > 1)
2494 new_owner |= UMUTEX_CONTESTED;
2495 again:
2496 error = casueword32(&m->m_owner, owner, &old, new_owner);
2497 if (error == 1) {
2498 error = thread_check_susp(td, false);
2499 if (error == 0)
2500 goto again;
2501 }
2502 umtxq_unbusy_unlocked(&key);
2503 umtx_key_release(&key);
2504 if (error == -1)
2505 return (EFAULT);
2506 if (error == 0 && old != owner)
2507 return (EINVAL);
2508 return (error);
2509 }
2510
2511 /*
2512 * Lock a PP mutex.
2513 */
2514 static int
do_lock_pp(struct thread * td,struct umutex * m,uint32_t flags,struct _umtx_time * timeout,int try)2515 do_lock_pp(struct thread *td, struct umutex *m, uint32_t flags,
2516 struct _umtx_time *timeout, int try)
2517 {
2518 struct umtx_abs_timeout timo;
2519 struct umtx_q *uq, *uq2;
2520 struct umtx_pi *pi;
2521 uint32_t ceiling;
2522 uint32_t owner, id;
2523 int error, pri, old_inherited_pri, new_pri, rv;
2524 bool su;
2525
2526 id = td->td_tid;
2527 uq = td->td_umtxq;
2528 if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2529 TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2530 &uq->uq_key)) != 0)
2531 return (error);
2532
2533 if (timeout != NULL)
2534 umtx_abs_timeout_init2(&timo, timeout);
2535
2536 su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2537 for (;;) {
2538 old_inherited_pri = uq->uq_inherited_pri;
2539 umtxq_lock(&uq->uq_key);
2540 umtxq_busy(&uq->uq_key);
2541 umtxq_unlock(&uq->uq_key);
2542
2543 rv = fueword32(&m->m_ceilings[0], &ceiling);
2544 if (rv == -1) {
2545 error = EFAULT;
2546 goto out;
2547 }
2548 ceiling = RTP_PRIO_MAX - ceiling;
2549 if (ceiling > RTP_PRIO_MAX) {
2550 error = EINVAL;
2551 goto out;
2552 }
2553 new_pri = PRI_MIN_REALTIME + ceiling;
2554
2555 if (td->td_base_user_pri < new_pri) {
2556 error = EINVAL;
2557 goto out;
2558 }
2559 if (su) {
2560 mtx_lock(&umtx_lock);
2561 if (new_pri < uq->uq_inherited_pri) {
2562 uq->uq_inherited_pri = new_pri;
2563 thread_lock(td);
2564 if (new_pri < UPRI(td))
2565 sched_lend_user_prio(td, new_pri);
2566 thread_unlock(td);
2567 }
2568 mtx_unlock(&umtx_lock);
2569 }
2570
2571 rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2572 id | UMUTEX_CONTESTED);
2573 /* The address was invalid. */
2574 if (rv == -1) {
2575 error = EFAULT;
2576 break;
2577 }
2578 if (rv == 0) {
2579 MPASS(owner == UMUTEX_CONTESTED);
2580 error = 0;
2581 break;
2582 }
2583 /* rv == 1 */
2584 if (owner == UMUTEX_RB_OWNERDEAD) {
2585 rv = casueword32(&m->m_owner, UMUTEX_RB_OWNERDEAD,
2586 &owner, id | UMUTEX_CONTESTED);
2587 if (rv == -1) {
2588 error = EFAULT;
2589 break;
2590 }
2591 if (rv == 0) {
2592 MPASS(owner == UMUTEX_RB_OWNERDEAD);
2593 error = EOWNERDEAD; /* success */
2594 break;
2595 }
2596
2597 /*
2598 * rv == 1, only check for suspension if we
2599 * did not already catched a signal. If we
2600 * get an error from the check, the same
2601 * condition is checked by the umtxq_sleep()
2602 * call below, so we should obliterate the
2603 * error to not skip the last loop iteration.
2604 */
2605 if (error == 0) {
2606 error = thread_check_susp(td, false);
2607 if (error == 0) {
2608 if (try != 0)
2609 error = EBUSY;
2610 else
2611 continue;
2612 }
2613 error = 0;
2614 }
2615 } else if (owner == UMUTEX_RB_NOTRECOV) {
2616 error = ENOTRECOVERABLE;
2617 }
2618
2619 if (try != 0)
2620 error = EBUSY;
2621
2622 /*
2623 * If we caught a signal, we have retried and now
2624 * exit immediately.
2625 */
2626 if (error != 0)
2627 break;
2628
2629 umtxq_lock(&uq->uq_key);
2630 umtxq_insert(uq);
2631 umtxq_unbusy(&uq->uq_key);
2632 error = umtxq_sleep(uq, "umtxpp", timeout == NULL ?
2633 NULL : &timo);
2634 umtxq_remove(uq);
2635 umtxq_unlock(&uq->uq_key);
2636
2637 mtx_lock(&umtx_lock);
2638 uq->uq_inherited_pri = old_inherited_pri;
2639 pri = PRI_MAX;
2640 TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2641 uq2 = TAILQ_FIRST(&pi->pi_blocked);
2642 if (uq2 != NULL) {
2643 if (pri > UPRI(uq2->uq_thread))
2644 pri = UPRI(uq2->uq_thread);
2645 }
2646 }
2647 if (pri > uq->uq_inherited_pri)
2648 pri = uq->uq_inherited_pri;
2649 thread_lock(td);
2650 sched_lend_user_prio(td, pri);
2651 thread_unlock(td);
2652 mtx_unlock(&umtx_lock);
2653 }
2654
2655 if (error != 0 && error != EOWNERDEAD) {
2656 mtx_lock(&umtx_lock);
2657 uq->uq_inherited_pri = old_inherited_pri;
2658 pri = PRI_MAX;
2659 TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2660 uq2 = TAILQ_FIRST(&pi->pi_blocked);
2661 if (uq2 != NULL) {
2662 if (pri > UPRI(uq2->uq_thread))
2663 pri = UPRI(uq2->uq_thread);
2664 }
2665 }
2666 if (pri > uq->uq_inherited_pri)
2667 pri = uq->uq_inherited_pri;
2668 thread_lock(td);
2669 sched_lend_user_prio(td, pri);
2670 thread_unlock(td);
2671 mtx_unlock(&umtx_lock);
2672 }
2673
2674 out:
2675 umtxq_unbusy_unlocked(&uq->uq_key);
2676 umtx_key_release(&uq->uq_key);
2677 return (error);
2678 }
2679
2680 /*
2681 * Unlock a PP mutex.
2682 */
2683 static int
do_unlock_pp(struct thread * td,struct umutex * m,uint32_t flags,bool rb)2684 do_unlock_pp(struct thread *td, struct umutex *m, uint32_t flags, bool rb)
2685 {
2686 struct umtx_key key;
2687 struct umtx_q *uq, *uq2;
2688 struct umtx_pi *pi;
2689 uint32_t id, owner, rceiling;
2690 int error, pri, new_inherited_pri;
2691 bool su;
2692
2693 id = td->td_tid;
2694 uq = td->td_umtxq;
2695 su = (priv_check(td, PRIV_SCHED_RTPRIO) == 0);
2696
2697 /*
2698 * Make sure we own this mtx.
2699 */
2700 error = fueword32(&m->m_owner, &owner);
2701 if (error == -1)
2702 return (EFAULT);
2703
2704 if ((owner & ~UMUTEX_CONTESTED) != id)
2705 return (EPERM);
2706
2707 error = copyin(&m->m_ceilings[1], &rceiling, sizeof(uint32_t));
2708 if (error != 0)
2709 return (error);
2710
2711 if (rceiling == -1)
2712 new_inherited_pri = PRI_MAX;
2713 else {
2714 rceiling = RTP_PRIO_MAX - rceiling;
2715 if (rceiling > RTP_PRIO_MAX)
2716 return (EINVAL);
2717 new_inherited_pri = PRI_MIN_REALTIME + rceiling;
2718 }
2719
2720 if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2721 TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2722 &key)) != 0)
2723 return (error);
2724 umtxq_lock(&key);
2725 umtxq_busy(&key);
2726 umtxq_unlock(&key);
2727 /*
2728 * For priority protected mutex, always set unlocked state
2729 * to UMUTEX_CONTESTED, so that userland always enters kernel
2730 * to lock the mutex, it is necessary because thread priority
2731 * has to be adjusted for such mutex.
2732 */
2733 error = suword32(&m->m_owner, umtx_unlock_val(flags, rb) |
2734 UMUTEX_CONTESTED);
2735
2736 umtxq_lock(&key);
2737 if (error == 0)
2738 umtxq_signal(&key, 1);
2739 umtxq_unbusy(&key);
2740 umtxq_unlock(&key);
2741
2742 if (error == -1)
2743 error = EFAULT;
2744 else {
2745 mtx_lock(&umtx_lock);
2746 if (su || new_inherited_pri == PRI_MAX)
2747 uq->uq_inherited_pri = new_inherited_pri;
2748 pri = PRI_MAX;
2749 TAILQ_FOREACH(pi, &uq->uq_pi_contested, pi_link) {
2750 uq2 = TAILQ_FIRST(&pi->pi_blocked);
2751 if (uq2 != NULL) {
2752 if (pri > UPRI(uq2->uq_thread))
2753 pri = UPRI(uq2->uq_thread);
2754 }
2755 }
2756 if (pri > uq->uq_inherited_pri)
2757 pri = uq->uq_inherited_pri;
2758 thread_lock(td);
2759 sched_lend_user_prio(td, pri);
2760 thread_unlock(td);
2761 mtx_unlock(&umtx_lock);
2762 }
2763 umtx_key_release(&key);
2764 return (error);
2765 }
2766
2767 static int
do_set_ceiling(struct thread * td,struct umutex * m,uint32_t ceiling,uint32_t * old_ceiling)2768 do_set_ceiling(struct thread *td, struct umutex *m, uint32_t ceiling,
2769 uint32_t *old_ceiling)
2770 {
2771 struct umtx_q *uq;
2772 uint32_t flags, id, owner, save_ceiling;
2773 int error, rv, rv1;
2774
2775 error = fueword32(&m->m_flags, &flags);
2776 if (error == -1)
2777 return (EFAULT);
2778 if ((flags & UMUTEX_PRIO_PROTECT) == 0)
2779 return (EINVAL);
2780 if (ceiling > RTP_PRIO_MAX)
2781 return (EINVAL);
2782 id = td->td_tid;
2783 uq = td->td_umtxq;
2784 if ((error = umtx_key_get(m, (flags & UMUTEX_ROBUST) != 0 ?
2785 TYPE_PP_ROBUST_UMUTEX : TYPE_PP_UMUTEX, GET_SHARE(flags),
2786 &uq->uq_key)) != 0)
2787 return (error);
2788 for (;;) {
2789 umtxq_lock(&uq->uq_key);
2790 umtxq_busy(&uq->uq_key);
2791 umtxq_unlock(&uq->uq_key);
2792
2793 rv = fueword32(&m->m_ceilings[0], &save_ceiling);
2794 if (rv == -1) {
2795 error = EFAULT;
2796 break;
2797 }
2798
2799 rv = casueword32(&m->m_owner, UMUTEX_CONTESTED, &owner,
2800 id | UMUTEX_CONTESTED);
2801 if (rv == -1) {
2802 error = EFAULT;
2803 break;
2804 }
2805
2806 if (rv == 0) {
2807 MPASS(owner == UMUTEX_CONTESTED);
2808 rv = suword32(&m->m_ceilings[0], ceiling);
2809 rv1 = suword32(&m->m_owner, UMUTEX_CONTESTED);
2810 error = (rv == 0 && rv1 == 0) ? 0: EFAULT;
2811 break;
2812 }
2813
2814 if ((owner & ~UMUTEX_CONTESTED) == id) {
2815 rv = suword32(&m->m_ceilings[0], ceiling);
2816 error = rv == 0 ? 0 : EFAULT;
2817 break;
2818 }
2819
2820 if (owner == UMUTEX_RB_OWNERDEAD) {
2821 error = EOWNERDEAD;
2822 break;
2823 } else if (owner == UMUTEX_RB_NOTRECOV) {
2824 error = ENOTRECOVERABLE;
2825 break;
2826 }
2827
2828 /*
2829 * If we caught a signal, we have retried and now
2830 * exit immediately.
2831 */
2832 if (error != 0)
2833 break;
2834
2835 /*
2836 * We set the contested bit, sleep. Otherwise the lock changed
2837 * and we need to retry or we lost a race to the thread
2838 * unlocking the umtx.
2839 */
2840 umtxq_lock(&uq->uq_key);
2841 umtxq_insert(uq);
2842 umtxq_unbusy(&uq->uq_key);
2843 error = umtxq_sleep(uq, "umtxpp", NULL);
2844 umtxq_remove(uq);
2845 umtxq_unlock(&uq->uq_key);
2846 }
2847 umtxq_lock(&uq->uq_key);
2848 if (error == 0)
2849 umtxq_signal(&uq->uq_key, INT_MAX);
2850 umtxq_unbusy(&uq->uq_key);
2851 umtxq_unlock(&uq->uq_key);
2852 umtx_key_release(&uq->uq_key);
2853 if (error == 0 && old_ceiling != NULL) {
2854 rv = suword32(old_ceiling, save_ceiling);
2855 error = rv == 0 ? 0 : EFAULT;
2856 }
2857 return (error);
2858 }
2859
2860 /*
2861 * Lock a userland POSIX mutex.
2862 */
2863 static int
do_lock_umutex(struct thread * td,struct umutex * m,struct _umtx_time * timeout,int mode)2864 do_lock_umutex(struct thread *td, struct umutex *m,
2865 struct _umtx_time *timeout, int mode)
2866 {
2867 uint32_t flags;
2868 int error;
2869
2870 error = fueword32(&m->m_flags, &flags);
2871 if (error == -1)
2872 return (EFAULT);
2873
2874 switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2875 case 0:
2876 error = do_lock_normal(td, m, flags, timeout, mode);
2877 break;
2878 case UMUTEX_PRIO_INHERIT:
2879 error = do_lock_pi(td, m, flags, timeout, mode);
2880 break;
2881 case UMUTEX_PRIO_PROTECT:
2882 error = do_lock_pp(td, m, flags, timeout, mode);
2883 break;
2884 default:
2885 return (EINVAL);
2886 }
2887 if (timeout == NULL) {
2888 if (error == EINTR && mode != _UMUTEX_WAIT)
2889 error = ERESTART;
2890 } else {
2891 /* Timed-locking is not restarted. */
2892 if (error == ERESTART)
2893 error = EINTR;
2894 }
2895 return (error);
2896 }
2897
2898 /*
2899 * Unlock a userland POSIX mutex.
2900 */
2901 static int
do_unlock_umutex(struct thread * td,struct umutex * m,bool rb)2902 do_unlock_umutex(struct thread *td, struct umutex *m, bool rb)
2903 {
2904 uint32_t flags;
2905 int error;
2906
2907 error = fueword32(&m->m_flags, &flags);
2908 if (error == -1)
2909 return (EFAULT);
2910
2911 switch (flags & (UMUTEX_PRIO_INHERIT | UMUTEX_PRIO_PROTECT)) {
2912 case 0:
2913 return (do_unlock_normal(td, m, flags, rb));
2914 case UMUTEX_PRIO_INHERIT:
2915 return (do_unlock_pi(td, m, flags, rb));
2916 case UMUTEX_PRIO_PROTECT:
2917 return (do_unlock_pp(td, m, flags, rb));
2918 }
2919
2920 return (EINVAL);
2921 }
2922
2923 static int
do_cv_wait(struct thread * td,struct ucond * cv,struct umutex * m,struct timespec * timeout,u_long wflags)2924 do_cv_wait(struct thread *td, struct ucond *cv, struct umutex *m,
2925 struct timespec *timeout, u_long wflags)
2926 {
2927 struct umtx_abs_timeout timo;
2928 struct umtx_q *uq;
2929 uint32_t flags, clockid, hasw;
2930 int error;
2931
2932 uq = td->td_umtxq;
2933 error = fueword32(&cv->c_flags, &flags);
2934 if (error == -1)
2935 return (EFAULT);
2936 error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &uq->uq_key);
2937 if (error != 0)
2938 return (error);
2939
2940 if ((wflags & CVWAIT_CLOCKID) != 0) {
2941 error = fueword32(&cv->c_clockid, &clockid);
2942 if (error == -1) {
2943 umtx_key_release(&uq->uq_key);
2944 return (EFAULT);
2945 }
2946 if (clockid < CLOCK_REALTIME ||
2947 clockid >= CLOCK_THREAD_CPUTIME_ID) {
2948 /* hmm, only HW clock id will work. */
2949 umtx_key_release(&uq->uq_key);
2950 return (EINVAL);
2951 }
2952 } else {
2953 clockid = CLOCK_REALTIME;
2954 }
2955
2956 umtxq_lock(&uq->uq_key);
2957 umtxq_busy(&uq->uq_key);
2958 umtxq_insert(uq);
2959 umtxq_unlock(&uq->uq_key);
2960
2961 /*
2962 * Set c_has_waiters to 1 before releasing user mutex, also
2963 * don't modify cache line when unnecessary.
2964 */
2965 error = fueword32(&cv->c_has_waiters, &hasw);
2966 if (error == 0 && hasw == 0)
2967 error = suword32(&cv->c_has_waiters, 1);
2968 if (error != 0) {
2969 umtxq_lock(&uq->uq_key);
2970 umtxq_remove(uq);
2971 umtxq_unbusy(&uq->uq_key);
2972 error = EFAULT;
2973 goto out;
2974 }
2975
2976 umtxq_unbusy_unlocked(&uq->uq_key);
2977
2978 error = do_unlock_umutex(td, m, false);
2979
2980 if (timeout != NULL)
2981 umtx_abs_timeout_init(&timo, clockid,
2982 (wflags & CVWAIT_ABSTIME) != 0, timeout);
2983
2984 umtxq_lock(&uq->uq_key);
2985 if (error == 0) {
2986 error = umtxq_sleep(uq, "ucond", timeout == NULL ?
2987 NULL : &timo);
2988 }
2989
2990 if ((uq->uq_flags & UQF_UMTXQ) == 0)
2991 error = 0;
2992 else {
2993 /*
2994 * This must be timeout,interrupted by signal or
2995 * surprious wakeup, clear c_has_waiter flag when
2996 * necessary.
2997 */
2998 umtxq_busy(&uq->uq_key);
2999 if ((uq->uq_flags & UQF_UMTXQ) != 0) {
3000 int oldlen = uq->uq_cur_queue->length;
3001 umtxq_remove(uq);
3002 if (oldlen == 1) {
3003 umtxq_unlock(&uq->uq_key);
3004 if (suword32(&cv->c_has_waiters, 0) != 0 &&
3005 error == 0)
3006 error = EFAULT;
3007 umtxq_lock(&uq->uq_key);
3008 }
3009 }
3010 umtxq_unbusy(&uq->uq_key);
3011 if (error == ERESTART)
3012 error = EINTR;
3013 }
3014 out:
3015 umtxq_unlock(&uq->uq_key);
3016 umtx_key_release(&uq->uq_key);
3017 return (error);
3018 }
3019
3020 /*
3021 * Signal a userland condition variable.
3022 */
3023 static int
do_cv_signal(struct thread * td,struct ucond * cv)3024 do_cv_signal(struct thread *td, struct ucond *cv)
3025 {
3026 struct umtx_key key;
3027 int error, cnt, nwake;
3028 uint32_t flags;
3029
3030 error = fueword32(&cv->c_flags, &flags);
3031 if (error == -1)
3032 return (EFAULT);
3033 if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
3034 return (error);
3035 umtxq_lock(&key);
3036 umtxq_busy(&key);
3037 cnt = umtxq_count(&key);
3038 nwake = umtxq_signal(&key, 1);
3039 if (cnt <= nwake) {
3040 umtxq_unlock(&key);
3041 error = suword32(&cv->c_has_waiters, 0);
3042 if (error == -1)
3043 error = EFAULT;
3044 umtxq_lock(&key);
3045 }
3046 umtxq_unbusy(&key);
3047 umtxq_unlock(&key);
3048 umtx_key_release(&key);
3049 return (error);
3050 }
3051
3052 static int
do_cv_broadcast(struct thread * td,struct ucond * cv)3053 do_cv_broadcast(struct thread *td, struct ucond *cv)
3054 {
3055 struct umtx_key key;
3056 int error;
3057 uint32_t flags;
3058
3059 error = fueword32(&cv->c_flags, &flags);
3060 if (error == -1)
3061 return (EFAULT);
3062 if ((error = umtx_key_get(cv, TYPE_CV, GET_SHARE(flags), &key)) != 0)
3063 return (error);
3064
3065 umtxq_lock(&key);
3066 umtxq_busy(&key);
3067 umtxq_signal(&key, INT_MAX);
3068 umtxq_unlock(&key);
3069
3070 error = suword32(&cv->c_has_waiters, 0);
3071 if (error == -1)
3072 error = EFAULT;
3073
3074 umtxq_unbusy_unlocked(&key);
3075
3076 umtx_key_release(&key);
3077 return (error);
3078 }
3079
3080 static int
do_rw_rdlock(struct thread * td,struct urwlock * rwlock,long fflag,struct _umtx_time * timeout)3081 do_rw_rdlock(struct thread *td, struct urwlock *rwlock, long fflag,
3082 struct _umtx_time *timeout)
3083 {
3084 struct umtx_abs_timeout timo;
3085 struct umtx_q *uq;
3086 uint32_t flags, wrflags;
3087 int32_t state, oldstate;
3088 int32_t blocked_readers;
3089 int error, error1, rv;
3090
3091 uq = td->td_umtxq;
3092 error = fueword32(&rwlock->rw_flags, &flags);
3093 if (error == -1)
3094 return (EFAULT);
3095 error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3096 if (error != 0)
3097 return (error);
3098
3099 if (timeout != NULL)
3100 umtx_abs_timeout_init2(&timo, timeout);
3101
3102 wrflags = URWLOCK_WRITE_OWNER;
3103 if (!(fflag & URWLOCK_PREFER_READER) && !(flags & URWLOCK_PREFER_READER))
3104 wrflags |= URWLOCK_WRITE_WAITERS;
3105
3106 for (;;) {
3107 rv = fueword32(&rwlock->rw_state, &state);
3108 if (rv == -1) {
3109 umtx_key_release(&uq->uq_key);
3110 return (EFAULT);
3111 }
3112
3113 /* try to lock it */
3114 while (!(state & wrflags)) {
3115 if (__predict_false(URWLOCK_READER_COUNT(state) ==
3116 URWLOCK_MAX_READERS)) {
3117 umtx_key_release(&uq->uq_key);
3118 return (EAGAIN);
3119 }
3120 rv = casueword32(&rwlock->rw_state, state,
3121 &oldstate, state + 1);
3122 if (rv == -1) {
3123 umtx_key_release(&uq->uq_key);
3124 return (EFAULT);
3125 }
3126 if (rv == 0) {
3127 MPASS(oldstate == state);
3128 umtx_key_release(&uq->uq_key);
3129 return (0);
3130 }
3131 error = thread_check_susp(td, true);
3132 if (error != 0)
3133 break;
3134 state = oldstate;
3135 }
3136
3137 if (error)
3138 break;
3139
3140 /* grab monitor lock */
3141 umtxq_lock(&uq->uq_key);
3142 umtxq_busy(&uq->uq_key);
3143 umtxq_unlock(&uq->uq_key);
3144
3145 /*
3146 * re-read the state, in case it changed between the try-lock above
3147 * and the check below
3148 */
3149 rv = fueword32(&rwlock->rw_state, &state);
3150 if (rv == -1)
3151 error = EFAULT;
3152
3153 /* set read contention bit */
3154 while (error == 0 && (state & wrflags) &&
3155 !(state & URWLOCK_READ_WAITERS)) {
3156 rv = casueword32(&rwlock->rw_state, state,
3157 &oldstate, state | URWLOCK_READ_WAITERS);
3158 if (rv == -1) {
3159 error = EFAULT;
3160 break;
3161 }
3162 if (rv == 0) {
3163 MPASS(oldstate == state);
3164 goto sleep;
3165 }
3166 state = oldstate;
3167 error = thread_check_susp(td, false);
3168 if (error != 0)
3169 break;
3170 }
3171 if (error != 0) {
3172 umtxq_unbusy_unlocked(&uq->uq_key);
3173 break;
3174 }
3175
3176 /* state is changed while setting flags, restart */
3177 if (!(state & wrflags)) {
3178 umtxq_unbusy_unlocked(&uq->uq_key);
3179 error = thread_check_susp(td, true);
3180 if (error != 0)
3181 break;
3182 continue;
3183 }
3184
3185 sleep:
3186 /*
3187 * Contention bit is set, before sleeping, increase
3188 * read waiter count.
3189 */
3190 rv = fueword32(&rwlock->rw_blocked_readers,
3191 &blocked_readers);
3192 if (rv == 0)
3193 rv = suword32(&rwlock->rw_blocked_readers,
3194 blocked_readers + 1);
3195 if (rv == -1) {
3196 umtxq_unbusy_unlocked(&uq->uq_key);
3197 error = EFAULT;
3198 break;
3199 }
3200
3201 while (state & wrflags) {
3202 umtxq_lock(&uq->uq_key);
3203 umtxq_insert(uq);
3204 umtxq_unbusy(&uq->uq_key);
3205
3206 error = umtxq_sleep(uq, "urdlck", timeout == NULL ?
3207 NULL : &timo);
3208
3209 umtxq_busy(&uq->uq_key);
3210 umtxq_remove(uq);
3211 umtxq_unlock(&uq->uq_key);
3212 if (error)
3213 break;
3214 rv = fueword32(&rwlock->rw_state, &state);
3215 if (rv == -1) {
3216 error = EFAULT;
3217 break;
3218 }
3219 }
3220
3221 /* decrease read waiter count, and may clear read contention bit */
3222 rv = fueword32(&rwlock->rw_blocked_readers,
3223 &blocked_readers);
3224 if (rv == 0)
3225 rv = suword32(&rwlock->rw_blocked_readers,
3226 blocked_readers - 1);
3227 if (rv == -1) {
3228 umtxq_unbusy_unlocked(&uq->uq_key);
3229 error = EFAULT;
3230 break;
3231 }
3232 if (blocked_readers == 1) {
3233 rv = fueword32(&rwlock->rw_state, &state);
3234 if (rv == -1) {
3235 umtxq_unbusy_unlocked(&uq->uq_key);
3236 error = EFAULT;
3237 break;
3238 }
3239 for (;;) {
3240 rv = casueword32(&rwlock->rw_state, state,
3241 &oldstate, state & ~URWLOCK_READ_WAITERS);
3242 if (rv == -1) {
3243 error = EFAULT;
3244 break;
3245 }
3246 if (rv == 0) {
3247 MPASS(oldstate == state);
3248 break;
3249 }
3250 state = oldstate;
3251 error1 = thread_check_susp(td, false);
3252 if (error1 != 0) {
3253 if (error == 0)
3254 error = error1;
3255 break;
3256 }
3257 }
3258 }
3259
3260 umtxq_unbusy_unlocked(&uq->uq_key);
3261 if (error != 0)
3262 break;
3263 }
3264 umtx_key_release(&uq->uq_key);
3265 if (error == ERESTART)
3266 error = EINTR;
3267 return (error);
3268 }
3269
3270 static int
do_rw_wrlock(struct thread * td,struct urwlock * rwlock,struct _umtx_time * timeout)3271 do_rw_wrlock(struct thread *td, struct urwlock *rwlock, struct _umtx_time *timeout)
3272 {
3273 struct umtx_abs_timeout timo;
3274 struct umtx_q *uq;
3275 uint32_t flags;
3276 int32_t state, oldstate;
3277 int32_t blocked_writers;
3278 int32_t blocked_readers;
3279 int error, error1, rv;
3280
3281 uq = td->td_umtxq;
3282 error = fueword32(&rwlock->rw_flags, &flags);
3283 if (error == -1)
3284 return (EFAULT);
3285 error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3286 if (error != 0)
3287 return (error);
3288
3289 if (timeout != NULL)
3290 umtx_abs_timeout_init2(&timo, timeout);
3291
3292 blocked_readers = 0;
3293 for (;;) {
3294 rv = fueword32(&rwlock->rw_state, &state);
3295 if (rv == -1) {
3296 umtx_key_release(&uq->uq_key);
3297 return (EFAULT);
3298 }
3299 while ((state & URWLOCK_WRITE_OWNER) == 0 &&
3300 URWLOCK_READER_COUNT(state) == 0) {
3301 rv = casueword32(&rwlock->rw_state, state,
3302 &oldstate, state | URWLOCK_WRITE_OWNER);
3303 if (rv == -1) {
3304 umtx_key_release(&uq->uq_key);
3305 return (EFAULT);
3306 }
3307 if (rv == 0) {
3308 MPASS(oldstate == state);
3309 umtx_key_release(&uq->uq_key);
3310 return (0);
3311 }
3312 state = oldstate;
3313 error = thread_check_susp(td, true);
3314 if (error != 0)
3315 break;
3316 }
3317
3318 if (error) {
3319 if ((state & (URWLOCK_WRITE_OWNER |
3320 URWLOCK_WRITE_WAITERS)) == 0 &&
3321 blocked_readers != 0) {
3322 umtxq_lock(&uq->uq_key);
3323 umtxq_busy(&uq->uq_key);
3324 umtxq_signal_queue(&uq->uq_key, INT_MAX,
3325 UMTX_SHARED_QUEUE);
3326 umtxq_unbusy(&uq->uq_key);
3327 umtxq_unlock(&uq->uq_key);
3328 }
3329
3330 break;
3331 }
3332
3333 /* grab monitor lock */
3334 umtxq_lock(&uq->uq_key);
3335 umtxq_busy(&uq->uq_key);
3336 umtxq_unlock(&uq->uq_key);
3337
3338 /*
3339 * Re-read the state, in case it changed between the
3340 * try-lock above and the check below.
3341 */
3342 rv = fueword32(&rwlock->rw_state, &state);
3343 if (rv == -1)
3344 error = EFAULT;
3345
3346 while (error == 0 && ((state & URWLOCK_WRITE_OWNER) ||
3347 URWLOCK_READER_COUNT(state) != 0) &&
3348 (state & URWLOCK_WRITE_WAITERS) == 0) {
3349 rv = casueword32(&rwlock->rw_state, state,
3350 &oldstate, state | URWLOCK_WRITE_WAITERS);
3351 if (rv == -1) {
3352 error = EFAULT;
3353 break;
3354 }
3355 if (rv == 0) {
3356 MPASS(oldstate == state);
3357 goto sleep;
3358 }
3359 state = oldstate;
3360 error = thread_check_susp(td, false);
3361 if (error != 0)
3362 break;
3363 }
3364 if (error != 0) {
3365 umtxq_unbusy_unlocked(&uq->uq_key);
3366 break;
3367 }
3368
3369 if ((state & URWLOCK_WRITE_OWNER) == 0 &&
3370 URWLOCK_READER_COUNT(state) == 0) {
3371 umtxq_unbusy_unlocked(&uq->uq_key);
3372 error = thread_check_susp(td, false);
3373 if (error != 0)
3374 break;
3375 continue;
3376 }
3377 sleep:
3378 rv = fueword32(&rwlock->rw_blocked_writers,
3379 &blocked_writers);
3380 if (rv == 0)
3381 rv = suword32(&rwlock->rw_blocked_writers,
3382 blocked_writers + 1);
3383 if (rv == -1) {
3384 umtxq_unbusy_unlocked(&uq->uq_key);
3385 error = EFAULT;
3386 break;
3387 }
3388
3389 while ((state & URWLOCK_WRITE_OWNER) ||
3390 URWLOCK_READER_COUNT(state) != 0) {
3391 umtxq_lock(&uq->uq_key);
3392 umtxq_insert_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3393 umtxq_unbusy(&uq->uq_key);
3394
3395 error = umtxq_sleep(uq, "uwrlck", timeout == NULL ?
3396 NULL : &timo);
3397
3398 umtxq_busy(&uq->uq_key);
3399 umtxq_remove_queue(uq, UMTX_EXCLUSIVE_QUEUE);
3400 umtxq_unlock(&uq->uq_key);
3401 if (error)
3402 break;
3403 rv = fueword32(&rwlock->rw_state, &state);
3404 if (rv == -1) {
3405 error = EFAULT;
3406 break;
3407 }
3408 }
3409
3410 rv = fueword32(&rwlock->rw_blocked_writers,
3411 &blocked_writers);
3412 if (rv == 0)
3413 rv = suword32(&rwlock->rw_blocked_writers,
3414 blocked_writers - 1);
3415 if (rv == -1) {
3416 umtxq_unbusy_unlocked(&uq->uq_key);
3417 error = EFAULT;
3418 break;
3419 }
3420 if (blocked_writers == 1) {
3421 rv = fueword32(&rwlock->rw_state, &state);
3422 if (rv == -1) {
3423 umtxq_unbusy_unlocked(&uq->uq_key);
3424 error = EFAULT;
3425 break;
3426 }
3427 for (;;) {
3428 rv = casueword32(&rwlock->rw_state, state,
3429 &oldstate, state & ~URWLOCK_WRITE_WAITERS);
3430 if (rv == -1) {
3431 error = EFAULT;
3432 break;
3433 }
3434 if (rv == 0) {
3435 MPASS(oldstate == state);
3436 break;
3437 }
3438 state = oldstate;
3439 error1 = thread_check_susp(td, false);
3440 /*
3441 * We are leaving the URWLOCK_WRITE_WAITERS
3442 * behind, but this should not harm the
3443 * correctness.
3444 */
3445 if (error1 != 0) {
3446 if (error == 0)
3447 error = error1;
3448 break;
3449 }
3450 }
3451 rv = fueword32(&rwlock->rw_blocked_readers,
3452 &blocked_readers);
3453 if (rv == -1) {
3454 umtxq_unbusy_unlocked(&uq->uq_key);
3455 error = EFAULT;
3456 break;
3457 }
3458 } else
3459 blocked_readers = 0;
3460
3461 umtxq_unbusy_unlocked(&uq->uq_key);
3462 }
3463
3464 umtx_key_release(&uq->uq_key);
3465 if (error == ERESTART)
3466 error = EINTR;
3467 return (error);
3468 }
3469
3470 static int
do_rw_unlock(struct thread * td,struct urwlock * rwlock)3471 do_rw_unlock(struct thread *td, struct urwlock *rwlock)
3472 {
3473 struct umtx_q *uq;
3474 uint32_t flags;
3475 int32_t state, oldstate;
3476 int error, rv, q, count;
3477
3478 uq = td->td_umtxq;
3479 error = fueword32(&rwlock->rw_flags, &flags);
3480 if (error == -1)
3481 return (EFAULT);
3482 error = umtx_key_get(rwlock, TYPE_RWLOCK, GET_SHARE(flags), &uq->uq_key);
3483 if (error != 0)
3484 return (error);
3485
3486 error = fueword32(&rwlock->rw_state, &state);
3487 if (error == -1) {
3488 error = EFAULT;
3489 goto out;
3490 }
3491 if (state & URWLOCK_WRITE_OWNER) {
3492 for (;;) {
3493 rv = casueword32(&rwlock->rw_state, state,
3494 &oldstate, state & ~URWLOCK_WRITE_OWNER);
3495 if (rv == -1) {
3496 error = EFAULT;
3497 goto out;
3498 }
3499 if (rv == 1) {
3500 state = oldstate;
3501 if (!(oldstate & URWLOCK_WRITE_OWNER)) {
3502 error = EPERM;
3503 goto out;
3504 }
3505 error = thread_check_susp(td, true);
3506 if (error != 0)
3507 goto out;
3508 } else
3509 break;
3510 }
3511 } else if (URWLOCK_READER_COUNT(state) != 0) {
3512 for (;;) {
3513 rv = casueword32(&rwlock->rw_state, state,
3514 &oldstate, state - 1);
3515 if (rv == -1) {
3516 error = EFAULT;
3517 goto out;
3518 }
3519 if (rv == 1) {
3520 state = oldstate;
3521 if (URWLOCK_READER_COUNT(oldstate) == 0) {
3522 error = EPERM;
3523 goto out;
3524 }
3525 error = thread_check_susp(td, true);
3526 if (error != 0)
3527 goto out;
3528 } else
3529 break;
3530 }
3531 } else {
3532 error = EPERM;
3533 goto out;
3534 }
3535
3536 count = 0;
3537
3538 if (!(flags & URWLOCK_PREFER_READER)) {
3539 if (state & URWLOCK_WRITE_WAITERS) {
3540 count = 1;
3541 q = UMTX_EXCLUSIVE_QUEUE;
3542 } else if (state & URWLOCK_READ_WAITERS) {
3543 count = INT_MAX;
3544 q = UMTX_SHARED_QUEUE;
3545 }
3546 } else {
3547 if (state & URWLOCK_READ_WAITERS) {
3548 count = INT_MAX;
3549 q = UMTX_SHARED_QUEUE;
3550 } else if (state & URWLOCK_WRITE_WAITERS) {
3551 count = 1;
3552 q = UMTX_EXCLUSIVE_QUEUE;
3553 }
3554 }
3555
3556 if (count) {
3557 umtxq_lock(&uq->uq_key);
3558 umtxq_busy(&uq->uq_key);
3559 umtxq_signal_queue(&uq->uq_key, count, q);
3560 umtxq_unbusy(&uq->uq_key);
3561 umtxq_unlock(&uq->uq_key);
3562 }
3563 out:
3564 umtx_key_release(&uq->uq_key);
3565 return (error);
3566 }
3567
3568 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
3569 static int
do_sem_wait(struct thread * td,struct _usem * sem,struct _umtx_time * timeout)3570 do_sem_wait(struct thread *td, struct _usem *sem, struct _umtx_time *timeout)
3571 {
3572 struct umtx_abs_timeout timo;
3573 struct umtx_q *uq;
3574 uint32_t flags, count, count1;
3575 int error, rv, rv1;
3576
3577 uq = td->td_umtxq;
3578 error = fueword32(&sem->_flags, &flags);
3579 if (error == -1)
3580 return (EFAULT);
3581 error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3582 if (error != 0)
3583 return (error);
3584
3585 if (timeout != NULL)
3586 umtx_abs_timeout_init2(&timo, timeout);
3587
3588 again:
3589 umtxq_lock(&uq->uq_key);
3590 umtxq_busy(&uq->uq_key);
3591 umtxq_insert(uq);
3592 umtxq_unlock(&uq->uq_key);
3593 rv = casueword32(&sem->_has_waiters, 0, &count1, 1);
3594 if (rv != -1)
3595 rv1 = fueword32(&sem->_count, &count);
3596 if (rv == -1 || rv1 == -1 || count != 0 || (rv == 1 && count1 == 0)) {
3597 if (rv == 0)
3598 rv = suword32(&sem->_has_waiters, 0);
3599 umtxq_lock(&uq->uq_key);
3600 umtxq_unbusy(&uq->uq_key);
3601 umtxq_remove(uq);
3602 umtxq_unlock(&uq->uq_key);
3603 if (rv == -1 || rv1 == -1) {
3604 error = EFAULT;
3605 goto out;
3606 }
3607 if (count != 0) {
3608 error = 0;
3609 goto out;
3610 }
3611 MPASS(rv == 1 && count1 == 0);
3612 rv = thread_check_susp(td, true);
3613 if (rv == 0)
3614 goto again;
3615 error = rv;
3616 goto out;
3617 }
3618 umtxq_lock(&uq->uq_key);
3619 umtxq_unbusy(&uq->uq_key);
3620
3621 error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3622
3623 if ((uq->uq_flags & UQF_UMTXQ) == 0)
3624 error = 0;
3625 else {
3626 umtxq_remove(uq);
3627 /* A relative timeout cannot be restarted. */
3628 if (error == ERESTART && timeout != NULL &&
3629 (timeout->_flags & UMTX_ABSTIME) == 0)
3630 error = EINTR;
3631 }
3632 umtxq_unlock(&uq->uq_key);
3633 out:
3634 umtx_key_release(&uq->uq_key);
3635 return (error);
3636 }
3637
3638 /*
3639 * Signal a userland semaphore.
3640 */
3641 static int
do_sem_wake(struct thread * td,struct _usem * sem)3642 do_sem_wake(struct thread *td, struct _usem *sem)
3643 {
3644 struct umtx_key key;
3645 int error, cnt;
3646 uint32_t flags;
3647
3648 error = fueword32(&sem->_flags, &flags);
3649 if (error == -1)
3650 return (EFAULT);
3651 if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3652 return (error);
3653 umtxq_lock(&key);
3654 umtxq_busy(&key);
3655 cnt = umtxq_count(&key);
3656 if (cnt > 0) {
3657 /*
3658 * Check if count is greater than 0, this means the memory is
3659 * still being referenced by user code, so we can safely
3660 * update _has_waiters flag.
3661 */
3662 if (cnt == 1) {
3663 umtxq_unlock(&key);
3664 error = suword32(&sem->_has_waiters, 0);
3665 umtxq_lock(&key);
3666 if (error == -1)
3667 error = EFAULT;
3668 }
3669 umtxq_signal(&key, 1);
3670 }
3671 umtxq_unbusy(&key);
3672 umtxq_unlock(&key);
3673 umtx_key_release(&key);
3674 return (error);
3675 }
3676 #endif
3677
3678 static int
do_sem2_wait(struct thread * td,struct _usem2 * sem,struct _umtx_time * timeout)3679 do_sem2_wait(struct thread *td, struct _usem2 *sem, struct _umtx_time *timeout)
3680 {
3681 struct umtx_abs_timeout timo;
3682 struct umtx_q *uq;
3683 uint32_t count, flags;
3684 int error, rv;
3685
3686 uq = td->td_umtxq;
3687 flags = fuword32(&sem->_flags);
3688 if (timeout != NULL)
3689 umtx_abs_timeout_init2(&timo, timeout);
3690
3691 again:
3692 error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &uq->uq_key);
3693 if (error != 0)
3694 return (error);
3695 umtxq_lock(&uq->uq_key);
3696 umtxq_busy(&uq->uq_key);
3697 umtxq_insert(uq);
3698 umtxq_unlock(&uq->uq_key);
3699 rv = fueword32(&sem->_count, &count);
3700 if (rv == -1) {
3701 umtxq_lock(&uq->uq_key);
3702 umtxq_unbusy(&uq->uq_key);
3703 umtxq_remove(uq);
3704 umtxq_unlock(&uq->uq_key);
3705 umtx_key_release(&uq->uq_key);
3706 return (EFAULT);
3707 }
3708 for (;;) {
3709 if (USEM_COUNT(count) != 0) {
3710 umtxq_lock(&uq->uq_key);
3711 umtxq_unbusy(&uq->uq_key);
3712 umtxq_remove(uq);
3713 umtxq_unlock(&uq->uq_key);
3714 umtx_key_release(&uq->uq_key);
3715 return (0);
3716 }
3717 if (count == USEM_HAS_WAITERS)
3718 break;
3719 rv = casueword32(&sem->_count, 0, &count, USEM_HAS_WAITERS);
3720 if (rv == 0)
3721 break;
3722 umtxq_lock(&uq->uq_key);
3723 umtxq_unbusy(&uq->uq_key);
3724 umtxq_remove(uq);
3725 umtxq_unlock(&uq->uq_key);
3726 umtx_key_release(&uq->uq_key);
3727 if (rv == -1)
3728 return (EFAULT);
3729 rv = thread_check_susp(td, true);
3730 if (rv != 0)
3731 return (rv);
3732 goto again;
3733 }
3734 umtxq_lock(&uq->uq_key);
3735 umtxq_unbusy(&uq->uq_key);
3736
3737 error = umtxq_sleep(uq, "usem", timeout == NULL ? NULL : &timo);
3738
3739 if ((uq->uq_flags & UQF_UMTXQ) == 0)
3740 error = 0;
3741 else {
3742 umtxq_remove(uq);
3743 if (timeout != NULL && (timeout->_flags & UMTX_ABSTIME) == 0) {
3744 /* A relative timeout cannot be restarted. */
3745 if (error == ERESTART)
3746 error = EINTR;
3747 if (error == EINTR) {
3748 kern_clock_gettime(curthread, timo.clockid,
3749 &timo.cur);
3750 timespecsub(&timo.end, &timo.cur,
3751 &timeout->_timeout);
3752 }
3753 }
3754 }
3755 umtxq_unlock(&uq->uq_key);
3756 umtx_key_release(&uq->uq_key);
3757 return (error);
3758 }
3759
3760 /*
3761 * Signal a userland semaphore.
3762 */
3763 static int
do_sem2_wake(struct thread * td,struct _usem2 * sem)3764 do_sem2_wake(struct thread *td, struct _usem2 *sem)
3765 {
3766 struct umtx_key key;
3767 int error, cnt, rv;
3768 uint32_t count, flags;
3769
3770 rv = fueword32(&sem->_flags, &flags);
3771 if (rv == -1)
3772 return (EFAULT);
3773 if ((error = umtx_key_get(sem, TYPE_SEM, GET_SHARE(flags), &key)) != 0)
3774 return (error);
3775 umtxq_lock(&key);
3776 umtxq_busy(&key);
3777 cnt = umtxq_count(&key);
3778 if (cnt > 0) {
3779 /*
3780 * If this was the last sleeping thread, clear the waiters
3781 * flag in _count.
3782 */
3783 if (cnt == 1) {
3784 umtxq_unlock(&key);
3785 rv = fueword32(&sem->_count, &count);
3786 while (rv != -1 && count & USEM_HAS_WAITERS) {
3787 rv = casueword32(&sem->_count, count, &count,
3788 count & ~USEM_HAS_WAITERS);
3789 if (rv == 1) {
3790 rv = thread_check_susp(td, true);
3791 if (rv != 0)
3792 break;
3793 }
3794 }
3795 if (rv == -1)
3796 error = EFAULT;
3797 else if (rv > 0) {
3798 error = rv;
3799 }
3800 umtxq_lock(&key);
3801 }
3802
3803 umtxq_signal(&key, 1);
3804 }
3805 umtxq_unbusy(&key);
3806 umtxq_unlock(&key);
3807 umtx_key_release(&key);
3808 return (error);
3809 }
3810
3811 #ifdef COMPAT_FREEBSD10
3812 int
freebsd10__umtx_lock(struct thread * td,struct freebsd10__umtx_lock_args * uap)3813 freebsd10__umtx_lock(struct thread *td, struct freebsd10__umtx_lock_args *uap)
3814 {
3815 return (do_lock_umtx(td, uap->umtx, td->td_tid, 0));
3816 }
3817
3818 int
freebsd10__umtx_unlock(struct thread * td,struct freebsd10__umtx_unlock_args * uap)3819 freebsd10__umtx_unlock(struct thread *td,
3820 struct freebsd10__umtx_unlock_args *uap)
3821 {
3822 return (do_unlock_umtx(td, uap->umtx, td->td_tid));
3823 }
3824 #endif
3825
3826 inline int
umtx_copyin_timeout(const void * uaddr,struct timespec * tsp)3827 umtx_copyin_timeout(const void *uaddr, struct timespec *tsp)
3828 {
3829 int error;
3830
3831 error = copyin(uaddr, tsp, sizeof(*tsp));
3832 if (error == 0) {
3833 if (!timespecvalid_interval(tsp))
3834 error = EINVAL;
3835 }
3836 return (error);
3837 }
3838
3839 static inline int
umtx_copyin_umtx_time(const void * uaddr,size_t size,struct _umtx_time * tp)3840 umtx_copyin_umtx_time(const void *uaddr, size_t size, struct _umtx_time *tp)
3841 {
3842 int error;
3843
3844 if (size <= sizeof(tp->_timeout)) {
3845 tp->_clockid = CLOCK_REALTIME;
3846 tp->_flags = 0;
3847 error = copyin(uaddr, &tp->_timeout, sizeof(tp->_timeout));
3848 } else
3849 error = copyin(uaddr, tp, sizeof(*tp));
3850 if (error != 0)
3851 return (error);
3852 if (!timespecvalid_interval(&tp->_timeout))
3853 return (EINVAL);
3854 return (0);
3855 }
3856
3857 static int
umtx_copyin_robust_lists(const void * uaddr,size_t size,struct umtx_robust_lists_params * rb)3858 umtx_copyin_robust_lists(const void *uaddr, size_t size,
3859 struct umtx_robust_lists_params *rb)
3860 {
3861
3862 if (size > sizeof(*rb))
3863 return (EINVAL);
3864 return (copyin(uaddr, rb, size));
3865 }
3866
3867 static int
umtx_copyout_timeout(void * uaddr,size_t sz,struct timespec * tsp)3868 umtx_copyout_timeout(void *uaddr, size_t sz, struct timespec *tsp)
3869 {
3870
3871 /*
3872 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
3873 * and we're only called if sz >= sizeof(timespec) as supplied in the
3874 * copyops.
3875 */
3876 KASSERT(sz >= sizeof(*tsp),
3877 ("umtx_copyops specifies incorrect sizes"));
3878
3879 return (copyout(tsp, uaddr, sizeof(*tsp)));
3880 }
3881
3882 #ifdef COMPAT_FREEBSD10
3883 static int
__umtx_op_lock_umtx(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)3884 __umtx_op_lock_umtx(struct thread *td, struct _umtx_op_args *uap,
3885 const struct umtx_copyops *ops)
3886 {
3887 struct timespec *ts, timeout;
3888 int error;
3889
3890 /* Allow a null timespec (wait forever). */
3891 if (uap->uaddr2 == NULL)
3892 ts = NULL;
3893 else {
3894 error = ops->copyin_timeout(uap->uaddr2, &timeout);
3895 if (error != 0)
3896 return (error);
3897 ts = &timeout;
3898 }
3899 #ifdef COMPAT_FREEBSD32
3900 if (ops->compat32)
3901 return (do_lock_umtx32(td, uap->obj, uap->val, ts));
3902 #endif
3903 return (do_lock_umtx(td, uap->obj, uap->val, ts));
3904 }
3905
3906 static int
__umtx_op_unlock_umtx(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)3907 __umtx_op_unlock_umtx(struct thread *td, struct _umtx_op_args *uap,
3908 const struct umtx_copyops *ops)
3909 {
3910 #ifdef COMPAT_FREEBSD32
3911 if (ops->compat32)
3912 return (do_unlock_umtx32(td, uap->obj, uap->val));
3913 #endif
3914 return (do_unlock_umtx(td, uap->obj, uap->val));
3915 }
3916 #endif /* COMPAT_FREEBSD10 */
3917
3918 #if !defined(COMPAT_FREEBSD10)
3919 static int
__umtx_op_unimpl(struct thread * td __unused,struct _umtx_op_args * uap __unused,const struct umtx_copyops * ops __unused)3920 __umtx_op_unimpl(struct thread *td __unused, struct _umtx_op_args *uap __unused,
3921 const struct umtx_copyops *ops __unused)
3922 {
3923 return (EOPNOTSUPP);
3924 }
3925 #endif /* COMPAT_FREEBSD10 */
3926
3927 static int
__umtx_op_wait(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)3928 __umtx_op_wait(struct thread *td, struct _umtx_op_args *uap,
3929 const struct umtx_copyops *ops)
3930 {
3931 struct _umtx_time timeout, *tm_p;
3932 int error;
3933
3934 if (uap->uaddr2 == NULL)
3935 tm_p = NULL;
3936 else {
3937 error = ops->copyin_umtx_time(
3938 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3939 if (error != 0)
3940 return (error);
3941 tm_p = &timeout;
3942 }
3943 return (do_wait(td, uap->obj, uap->val, tm_p, ops->compat32, 0));
3944 }
3945
3946 static int
__umtx_op_wait_uint(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)3947 __umtx_op_wait_uint(struct thread *td, struct _umtx_op_args *uap,
3948 const struct umtx_copyops *ops)
3949 {
3950 struct _umtx_time timeout, *tm_p;
3951 int error;
3952
3953 if (uap->uaddr2 == NULL)
3954 tm_p = NULL;
3955 else {
3956 error = ops->copyin_umtx_time(
3957 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3958 if (error != 0)
3959 return (error);
3960 tm_p = &timeout;
3961 }
3962 return (do_wait(td, uap->obj, uap->val, tm_p, 1, 0));
3963 }
3964
3965 static int
__umtx_op_wait_uint_private(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)3966 __umtx_op_wait_uint_private(struct thread *td, struct _umtx_op_args *uap,
3967 const struct umtx_copyops *ops)
3968 {
3969 struct _umtx_time *tm_p, timeout;
3970 int error;
3971
3972 if (uap->uaddr2 == NULL)
3973 tm_p = NULL;
3974 else {
3975 error = ops->copyin_umtx_time(
3976 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
3977 if (error != 0)
3978 return (error);
3979 tm_p = &timeout;
3980 }
3981 return (do_wait(td, uap->obj, uap->val, tm_p, 1, 1));
3982 }
3983
3984 static int
__umtx_op_wake(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)3985 __umtx_op_wake(struct thread *td, struct _umtx_op_args *uap,
3986 const struct umtx_copyops *ops __unused)
3987 {
3988
3989 return (kern_umtx_wake(td, uap->obj, uap->val, 0));
3990 }
3991
3992 #define BATCH_SIZE 128
3993 static int
__umtx_op_nwake_private_native(struct thread * td,struct _umtx_op_args * uap)3994 __umtx_op_nwake_private_native(struct thread *td, struct _umtx_op_args *uap)
3995 {
3996 char *uaddrs[BATCH_SIZE], **upp;
3997 int count, error, i, pos, tocopy;
3998
3999 upp = (char **)uap->obj;
4000 error = 0;
4001 for (count = uap->val, pos = 0; count > 0; count -= tocopy,
4002 pos += tocopy) {
4003 tocopy = MIN(count, BATCH_SIZE);
4004 error = copyin(upp + pos, uaddrs, tocopy * sizeof(char *));
4005 if (error != 0)
4006 break;
4007 for (i = 0; i < tocopy; ++i) {
4008 kern_umtx_wake(td, uaddrs[i], INT_MAX, 1);
4009 }
4010 maybe_yield();
4011 }
4012 return (error);
4013 }
4014
4015 static int
__umtx_op_nwake_private_compat32(struct thread * td,struct _umtx_op_args * uap)4016 __umtx_op_nwake_private_compat32(struct thread *td, struct _umtx_op_args *uap)
4017 {
4018 uint32_t uaddrs[BATCH_SIZE], *upp;
4019 int count, error, i, pos, tocopy;
4020
4021 upp = (uint32_t *)uap->obj;
4022 error = 0;
4023 for (count = uap->val, pos = 0; count > 0; count -= tocopy,
4024 pos += tocopy) {
4025 tocopy = MIN(count, BATCH_SIZE);
4026 error = copyin(upp + pos, uaddrs, tocopy * sizeof(uint32_t));
4027 if (error != 0)
4028 break;
4029 for (i = 0; i < tocopy; ++i) {
4030 kern_umtx_wake(td, (void *)(uintptr_t)uaddrs[i],
4031 INT_MAX, 1);
4032 }
4033 maybe_yield();
4034 }
4035 return (error);
4036 }
4037
4038 static int
__umtx_op_nwake_private(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4039 __umtx_op_nwake_private(struct thread *td, struct _umtx_op_args *uap,
4040 const struct umtx_copyops *ops)
4041 {
4042
4043 if (ops->compat32)
4044 return (__umtx_op_nwake_private_compat32(td, uap));
4045 return (__umtx_op_nwake_private_native(td, uap));
4046 }
4047
4048 static int
__umtx_op_wake_private(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4049 __umtx_op_wake_private(struct thread *td, struct _umtx_op_args *uap,
4050 const struct umtx_copyops *ops __unused)
4051 {
4052
4053 return (kern_umtx_wake(td, uap->obj, uap->val, 1));
4054 }
4055
4056 static int
__umtx_op_lock_umutex(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4057 __umtx_op_lock_umutex(struct thread *td, struct _umtx_op_args *uap,
4058 const struct umtx_copyops *ops)
4059 {
4060 struct _umtx_time *tm_p, timeout;
4061 int error;
4062
4063 /* Allow a null timespec (wait forever). */
4064 if (uap->uaddr2 == NULL)
4065 tm_p = NULL;
4066 else {
4067 error = ops->copyin_umtx_time(
4068 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4069 if (error != 0)
4070 return (error);
4071 tm_p = &timeout;
4072 }
4073 return (do_lock_umutex(td, uap->obj, tm_p, 0));
4074 }
4075
4076 static int
__umtx_op_trylock_umutex(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4077 __umtx_op_trylock_umutex(struct thread *td, struct _umtx_op_args *uap,
4078 const struct umtx_copyops *ops __unused)
4079 {
4080
4081 return (do_lock_umutex(td, uap->obj, NULL, _UMUTEX_TRY));
4082 }
4083
4084 static int
__umtx_op_wait_umutex(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4085 __umtx_op_wait_umutex(struct thread *td, struct _umtx_op_args *uap,
4086 const struct umtx_copyops *ops)
4087 {
4088 struct _umtx_time *tm_p, timeout;
4089 int error;
4090
4091 /* Allow a null timespec (wait forever). */
4092 if (uap->uaddr2 == NULL)
4093 tm_p = NULL;
4094 else {
4095 error = ops->copyin_umtx_time(
4096 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4097 if (error != 0)
4098 return (error);
4099 tm_p = &timeout;
4100 }
4101 return (do_lock_umutex(td, uap->obj, tm_p, _UMUTEX_WAIT));
4102 }
4103
4104 static int
__umtx_op_wake_umutex(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4105 __umtx_op_wake_umutex(struct thread *td, struct _umtx_op_args *uap,
4106 const struct umtx_copyops *ops __unused)
4107 {
4108
4109 return (do_wake_umutex(td, uap->obj));
4110 }
4111
4112 static int
__umtx_op_unlock_umutex(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4113 __umtx_op_unlock_umutex(struct thread *td, struct _umtx_op_args *uap,
4114 const struct umtx_copyops *ops __unused)
4115 {
4116
4117 return (do_unlock_umutex(td, uap->obj, false));
4118 }
4119
4120 static int
__umtx_op_set_ceiling(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4121 __umtx_op_set_ceiling(struct thread *td, struct _umtx_op_args *uap,
4122 const struct umtx_copyops *ops __unused)
4123 {
4124
4125 return (do_set_ceiling(td, uap->obj, uap->val, uap->uaddr1));
4126 }
4127
4128 static int
__umtx_op_cv_wait(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4129 __umtx_op_cv_wait(struct thread *td, struct _umtx_op_args *uap,
4130 const struct umtx_copyops *ops)
4131 {
4132 struct timespec *ts, timeout;
4133 int error;
4134
4135 /* Allow a null timespec (wait forever). */
4136 if (uap->uaddr2 == NULL)
4137 ts = NULL;
4138 else {
4139 error = ops->copyin_timeout(uap->uaddr2, &timeout);
4140 if (error != 0)
4141 return (error);
4142 ts = &timeout;
4143 }
4144 return (do_cv_wait(td, uap->obj, uap->uaddr1, ts, uap->val));
4145 }
4146
4147 static int
__umtx_op_cv_signal(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4148 __umtx_op_cv_signal(struct thread *td, struct _umtx_op_args *uap,
4149 const struct umtx_copyops *ops __unused)
4150 {
4151
4152 return (do_cv_signal(td, uap->obj));
4153 }
4154
4155 static int
__umtx_op_cv_broadcast(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4156 __umtx_op_cv_broadcast(struct thread *td, struct _umtx_op_args *uap,
4157 const struct umtx_copyops *ops __unused)
4158 {
4159
4160 return (do_cv_broadcast(td, uap->obj));
4161 }
4162
4163 static int
__umtx_op_rw_rdlock(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4164 __umtx_op_rw_rdlock(struct thread *td, struct _umtx_op_args *uap,
4165 const struct umtx_copyops *ops)
4166 {
4167 struct _umtx_time timeout;
4168 int error;
4169
4170 /* Allow a null timespec (wait forever). */
4171 if (uap->uaddr2 == NULL) {
4172 error = do_rw_rdlock(td, uap->obj, uap->val, 0);
4173 } else {
4174 error = ops->copyin_umtx_time(uap->uaddr2,
4175 (size_t)uap->uaddr1, &timeout);
4176 if (error != 0)
4177 return (error);
4178 error = do_rw_rdlock(td, uap->obj, uap->val, &timeout);
4179 }
4180 return (error);
4181 }
4182
4183 static int
__umtx_op_rw_wrlock(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4184 __umtx_op_rw_wrlock(struct thread *td, struct _umtx_op_args *uap,
4185 const struct umtx_copyops *ops)
4186 {
4187 struct _umtx_time timeout;
4188 int error;
4189
4190 /* Allow a null timespec (wait forever). */
4191 if (uap->uaddr2 == NULL) {
4192 error = do_rw_wrlock(td, uap->obj, 0);
4193 } else {
4194 error = ops->copyin_umtx_time(uap->uaddr2,
4195 (size_t)uap->uaddr1, &timeout);
4196 if (error != 0)
4197 return (error);
4198
4199 error = do_rw_wrlock(td, uap->obj, &timeout);
4200 }
4201 return (error);
4202 }
4203
4204 static int
__umtx_op_rw_unlock(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4205 __umtx_op_rw_unlock(struct thread *td, struct _umtx_op_args *uap,
4206 const struct umtx_copyops *ops __unused)
4207 {
4208
4209 return (do_rw_unlock(td, uap->obj));
4210 }
4211
4212 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4213 static int
__umtx_op_sem_wait(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4214 __umtx_op_sem_wait(struct thread *td, struct _umtx_op_args *uap,
4215 const struct umtx_copyops *ops)
4216 {
4217 struct _umtx_time *tm_p, timeout;
4218 int error;
4219
4220 /* Allow a null timespec (wait forever). */
4221 if (uap->uaddr2 == NULL)
4222 tm_p = NULL;
4223 else {
4224 error = ops->copyin_umtx_time(
4225 uap->uaddr2, (size_t)uap->uaddr1, &timeout);
4226 if (error != 0)
4227 return (error);
4228 tm_p = &timeout;
4229 }
4230 return (do_sem_wait(td, uap->obj, tm_p));
4231 }
4232
4233 static int
__umtx_op_sem_wake(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4234 __umtx_op_sem_wake(struct thread *td, struct _umtx_op_args *uap,
4235 const struct umtx_copyops *ops __unused)
4236 {
4237
4238 return (do_sem_wake(td, uap->obj));
4239 }
4240 #endif
4241
4242 static int
__umtx_op_wake2_umutex(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4243 __umtx_op_wake2_umutex(struct thread *td, struct _umtx_op_args *uap,
4244 const struct umtx_copyops *ops __unused)
4245 {
4246
4247 return (do_wake2_umutex(td, uap->obj, uap->val));
4248 }
4249
4250 static int
__umtx_op_sem2_wait(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4251 __umtx_op_sem2_wait(struct thread *td, struct _umtx_op_args *uap,
4252 const struct umtx_copyops *ops)
4253 {
4254 struct _umtx_time *tm_p, timeout;
4255 size_t uasize;
4256 int error;
4257
4258 /* Allow a null timespec (wait forever). */
4259 if (uap->uaddr2 == NULL) {
4260 uasize = 0;
4261 tm_p = NULL;
4262 } else {
4263 uasize = (size_t)uap->uaddr1;
4264 error = ops->copyin_umtx_time(uap->uaddr2, uasize, &timeout);
4265 if (error != 0)
4266 return (error);
4267 tm_p = &timeout;
4268 }
4269 error = do_sem2_wait(td, uap->obj, tm_p);
4270 if (error == EINTR && uap->uaddr2 != NULL &&
4271 (timeout._flags & UMTX_ABSTIME) == 0 &&
4272 uasize >= ops->umtx_time_sz + ops->timespec_sz) {
4273 error = ops->copyout_timeout(
4274 (void *)((uintptr_t)uap->uaddr2 + ops->umtx_time_sz),
4275 uasize - ops->umtx_time_sz, &timeout._timeout);
4276 if (error == 0) {
4277 error = EINTR;
4278 }
4279 }
4280
4281 return (error);
4282 }
4283
4284 static int
__umtx_op_sem2_wake(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4285 __umtx_op_sem2_wake(struct thread *td, struct _umtx_op_args *uap,
4286 const struct umtx_copyops *ops __unused)
4287 {
4288
4289 return (do_sem2_wake(td, uap->obj));
4290 }
4291
4292 #define USHM_OBJ_UMTX(o) \
4293 ((struct umtx_shm_obj_list *)(&(o)->umtx_data))
4294
4295 #define USHMF_LINKED 0x0001
4296 struct umtx_shm_reg {
4297 TAILQ_ENTRY(umtx_shm_reg) ushm_reg_link;
4298 LIST_ENTRY(umtx_shm_reg) ushm_obj_link;
4299 struct umtx_key ushm_key;
4300 struct ucred *ushm_cred;
4301 struct shmfd *ushm_obj;
4302 u_int ushm_refcnt;
4303 u_int ushm_flags;
4304 };
4305
4306 LIST_HEAD(umtx_shm_obj_list, umtx_shm_reg);
4307 TAILQ_HEAD(umtx_shm_reg_head, umtx_shm_reg);
4308
4309 static uma_zone_t umtx_shm_reg_zone;
4310 static struct umtx_shm_reg_head umtx_shm_registry[UMTX_CHAINS];
4311 static struct mtx umtx_shm_lock;
4312 static struct umtx_shm_reg_head umtx_shm_reg_delfree =
4313 TAILQ_HEAD_INITIALIZER(umtx_shm_reg_delfree);
4314
4315 static void umtx_shm_free_reg(struct umtx_shm_reg *reg);
4316
4317 static void
umtx_shm_reg_delfree_tq(void * context __unused,int pending __unused)4318 umtx_shm_reg_delfree_tq(void *context __unused, int pending __unused)
4319 {
4320 struct umtx_shm_reg_head d;
4321 struct umtx_shm_reg *reg, *reg1;
4322
4323 TAILQ_INIT(&d);
4324 mtx_lock(&umtx_shm_lock);
4325 TAILQ_CONCAT(&d, &umtx_shm_reg_delfree, ushm_reg_link);
4326 mtx_unlock(&umtx_shm_lock);
4327 TAILQ_FOREACH_SAFE(reg, &d, ushm_reg_link, reg1) {
4328 TAILQ_REMOVE(&d, reg, ushm_reg_link);
4329 umtx_shm_free_reg(reg);
4330 }
4331 }
4332
4333 static struct task umtx_shm_reg_delfree_task =
4334 TASK_INITIALIZER(0, umtx_shm_reg_delfree_tq, NULL);
4335
4336 /*
4337 * Returns 0 if a SHM with the passed key is found in the registry, in which
4338 * case it is returned through 'oreg'. Otherwise, returns an error among ESRCH
4339 * (no corresponding SHM; ESRCH was chosen for compatibility, ENOENT would have
4340 * been preferable) or EOVERFLOW (there is a corresponding SHM, but reference
4341 * count would overflow, so can't return it), in which case '*oreg' is left
4342 * unchanged.
4343 */
4344 static int
umtx_shm_find_reg_locked(const struct umtx_key * key,struct umtx_shm_reg ** const oreg)4345 umtx_shm_find_reg_locked(const struct umtx_key *key,
4346 struct umtx_shm_reg **const oreg)
4347 {
4348 struct umtx_shm_reg *reg;
4349 struct umtx_shm_reg_head *reg_head;
4350
4351 KASSERT(key->shared, ("umtx_p_find_rg: private key"));
4352 mtx_assert(&umtx_shm_lock, MA_OWNED);
4353 reg_head = &umtx_shm_registry[key->hash];
4354 TAILQ_FOREACH(reg, reg_head, ushm_reg_link) {
4355 KASSERT(reg->ushm_key.shared,
4356 ("non-shared key on reg %p %d", reg, reg->ushm_key.shared));
4357 if (reg->ushm_key.info.shared.object ==
4358 key->info.shared.object &&
4359 reg->ushm_key.info.shared.offset ==
4360 key->info.shared.offset) {
4361 KASSERT(reg->ushm_key.type == TYPE_SHM, ("TYPE_USHM"));
4362 KASSERT(reg->ushm_refcnt != 0,
4363 ("reg %p refcnt 0 onlist", reg));
4364 KASSERT((reg->ushm_flags & USHMF_LINKED) != 0,
4365 ("reg %p not linked", reg));
4366 /*
4367 * Don't let overflow happen, just deny a new reference
4368 * (this is additional protection against some reference
4369 * count leak, which is known not to be the case at the
4370 * time of this writing).
4371 */
4372 if (__predict_false(reg->ushm_refcnt == UINT_MAX))
4373 return (EOVERFLOW);
4374 reg->ushm_refcnt++;
4375 *oreg = reg;
4376 return (0);
4377 }
4378 }
4379 return (ESRCH);
4380 }
4381
4382 /*
4383 * Calls umtx_shm_find_reg_unlocked() under the 'umtx_shm_lock'.
4384 */
4385 static int
umtx_shm_find_reg(const struct umtx_key * key,struct umtx_shm_reg ** const oreg)4386 umtx_shm_find_reg(const struct umtx_key *key, struct umtx_shm_reg **const oreg)
4387 {
4388 int error;
4389
4390 mtx_lock(&umtx_shm_lock);
4391 error = umtx_shm_find_reg_locked(key, oreg);
4392 mtx_unlock(&umtx_shm_lock);
4393 return (error);
4394 }
4395
4396 static void
umtx_shm_free_reg(struct umtx_shm_reg * reg)4397 umtx_shm_free_reg(struct umtx_shm_reg *reg)
4398 {
4399
4400 chgumtxcnt(reg->ushm_cred->cr_ruidinfo, -1, 0);
4401 crfree(reg->ushm_cred);
4402 shm_drop(reg->ushm_obj);
4403 uma_zfree(umtx_shm_reg_zone, reg);
4404 }
4405
4406 static bool
umtx_shm_unref_reg_locked(struct umtx_shm_reg * reg,bool linked_ref)4407 umtx_shm_unref_reg_locked(struct umtx_shm_reg *reg, bool linked_ref)
4408 {
4409 mtx_assert(&umtx_shm_lock, MA_OWNED);
4410 KASSERT(reg->ushm_refcnt != 0, ("ushm_reg %p refcnt 0", reg));
4411
4412 if (linked_ref) {
4413 if ((reg->ushm_flags & USHMF_LINKED) == 0)
4414 /*
4415 * The reference tied to USHMF_LINKED has already been
4416 * released concurrently.
4417 */
4418 return (false);
4419
4420 TAILQ_REMOVE(&umtx_shm_registry[reg->ushm_key.hash], reg,
4421 ushm_reg_link);
4422 LIST_REMOVE(reg, ushm_obj_link);
4423 reg->ushm_flags &= ~USHMF_LINKED;
4424 }
4425
4426 reg->ushm_refcnt--;
4427 return (reg->ushm_refcnt == 0);
4428 }
4429
4430 static void
umtx_shm_unref_reg(struct umtx_shm_reg * reg,bool linked_ref)4431 umtx_shm_unref_reg(struct umtx_shm_reg *reg, bool linked_ref)
4432 {
4433 vm_object_t object;
4434 bool dofree;
4435
4436 if (linked_ref) {
4437 /*
4438 * Note: This may be executed multiple times on the same
4439 * shared-memory VM object in presence of concurrent callers
4440 * because 'umtx_shm_lock' is not held all along in umtx_shm()
4441 * and here.
4442 */
4443 object = reg->ushm_obj->shm_object;
4444 VM_OBJECT_WLOCK(object);
4445 vm_object_set_flag(object, OBJ_UMTXDEAD);
4446 VM_OBJECT_WUNLOCK(object);
4447 }
4448 mtx_lock(&umtx_shm_lock);
4449 dofree = umtx_shm_unref_reg_locked(reg, linked_ref);
4450 mtx_unlock(&umtx_shm_lock);
4451 if (dofree)
4452 umtx_shm_free_reg(reg);
4453 }
4454
4455 void
umtx_shm_object_init(vm_object_t object)4456 umtx_shm_object_init(vm_object_t object)
4457 {
4458
4459 LIST_INIT(USHM_OBJ_UMTX(object));
4460 }
4461
4462 void
umtx_shm_object_terminated(vm_object_t object)4463 umtx_shm_object_terminated(vm_object_t object)
4464 {
4465 struct umtx_shm_reg *reg, *reg1;
4466 bool dofree;
4467
4468 if (LIST_EMPTY(USHM_OBJ_UMTX(object)))
4469 return;
4470
4471 dofree = false;
4472 mtx_lock(&umtx_shm_lock);
4473 LIST_FOREACH_SAFE(reg, USHM_OBJ_UMTX(object), ushm_obj_link, reg1) {
4474 if (umtx_shm_unref_reg_locked(reg, true)) {
4475 TAILQ_INSERT_TAIL(&umtx_shm_reg_delfree, reg,
4476 ushm_reg_link);
4477 dofree = true;
4478 }
4479 }
4480 mtx_unlock(&umtx_shm_lock);
4481 if (dofree)
4482 taskqueue_enqueue(taskqueue_thread, &umtx_shm_reg_delfree_task);
4483 }
4484
4485 static int
umtx_shm_create_reg(struct thread * td,const struct umtx_key * key,struct umtx_shm_reg ** res)4486 umtx_shm_create_reg(struct thread *td, const struct umtx_key *key,
4487 struct umtx_shm_reg **res)
4488 {
4489 struct umtx_shm_reg *reg, *reg1;
4490 struct ucred *cred;
4491 int error;
4492
4493 error = umtx_shm_find_reg(key, res);
4494 if (error != ESRCH) {
4495 /*
4496 * Either no error occured, and '*res' was filled, or EOVERFLOW
4497 * was returned, indicating a reference count limit, and we
4498 * won't create a duplicate registration. In both cases, we are
4499 * done.
4500 */
4501 return (error);
4502 }
4503 /* No entry, we will create one. */
4504
4505 cred = td->td_ucred;
4506 if (!chgumtxcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_UMTXP)))
4507 return (ENOMEM);
4508 reg = uma_zalloc(umtx_shm_reg_zone, M_WAITOK | M_ZERO);
4509 bcopy(key, ®->ushm_key, sizeof(*key));
4510 reg->ushm_obj = shm_alloc(td->td_ucred, O_RDWR, false);
4511 reg->ushm_cred = crhold(cred);
4512 error = shm_dotruncate(reg->ushm_obj, PAGE_SIZE);
4513 if (error != 0) {
4514 umtx_shm_free_reg(reg);
4515 return (error);
4516 }
4517 mtx_lock(&umtx_shm_lock);
4518 /* Re-lookup as 'umtx_shm_lock' has been temporarily released. */
4519 error = umtx_shm_find_reg_locked(key, ®1);
4520 switch (error) {
4521 case 0:
4522 mtx_unlock(&umtx_shm_lock);
4523 umtx_shm_free_reg(reg);
4524 *res = reg1;
4525 return (0);
4526 case ESRCH:
4527 break;
4528 default:
4529 mtx_unlock(&umtx_shm_lock);
4530 umtx_shm_free_reg(reg);
4531 return (error);
4532 }
4533 TAILQ_INSERT_TAIL(&umtx_shm_registry[key->hash], reg, ushm_reg_link);
4534 LIST_INSERT_HEAD(USHM_OBJ_UMTX(key->info.shared.object), reg,
4535 ushm_obj_link);
4536 reg->ushm_flags = USHMF_LINKED;
4537 /*
4538 * This is one reference for the registry and the list of shared
4539 * mutexes referenced by the VM object containing the lock pointer, and
4540 * another for the caller, which it will free after use. So, one of
4541 * these is tied to the presence of USHMF_LINKED.
4542 */
4543 reg->ushm_refcnt = 2;
4544 mtx_unlock(&umtx_shm_lock);
4545 *res = reg;
4546 return (0);
4547 }
4548
4549 static int
umtx_shm_alive(struct thread * td,void * addr)4550 umtx_shm_alive(struct thread *td, void *addr)
4551 {
4552 vm_map_t map;
4553 vm_map_entry_t entry;
4554 vm_object_t object;
4555 vm_pindex_t pindex;
4556 vm_prot_t prot;
4557 int res, ret;
4558 boolean_t wired;
4559
4560 map = &td->td_proc->p_vmspace->vm_map;
4561 res = vm_map_lookup(&map, (uintptr_t)addr, VM_PROT_READ, &entry,
4562 &object, &pindex, &prot, &wired);
4563 if (res != KERN_SUCCESS)
4564 return (EFAULT);
4565 if (object == NULL)
4566 ret = EINVAL;
4567 else
4568 ret = (object->flags & OBJ_UMTXDEAD) != 0 ? ENOTTY : 0;
4569 vm_map_lookup_done(map, entry);
4570 return (ret);
4571 }
4572
4573 static void
umtx_shm_init(void)4574 umtx_shm_init(void)
4575 {
4576 int i;
4577
4578 umtx_shm_reg_zone = uma_zcreate("umtx_shm", sizeof(struct umtx_shm_reg),
4579 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
4580 mtx_init(&umtx_shm_lock, "umtxshm", NULL, MTX_DEF);
4581 for (i = 0; i < nitems(umtx_shm_registry); i++)
4582 TAILQ_INIT(&umtx_shm_registry[i]);
4583 }
4584
4585 static int
umtx_shm(struct thread * td,void * addr,u_int flags)4586 umtx_shm(struct thread *td, void *addr, u_int flags)
4587 {
4588 struct umtx_key key;
4589 struct umtx_shm_reg *reg;
4590 struct file *fp;
4591 int error, fd;
4592
4593 if (__bitcount(flags & (UMTX_SHM_CREAT | UMTX_SHM_LOOKUP |
4594 UMTX_SHM_DESTROY| UMTX_SHM_ALIVE)) != 1)
4595 return (EINVAL);
4596 if ((flags & UMTX_SHM_ALIVE) != 0)
4597 return (umtx_shm_alive(td, addr));
4598 error = umtx_key_get(addr, TYPE_SHM, PROCESS_SHARE, &key);
4599 if (error != 0)
4600 return (error);
4601 KASSERT(key.shared == 1, ("non-shared key"));
4602 error = (flags & UMTX_SHM_CREAT) != 0 ?
4603 umtx_shm_create_reg(td, &key, ®) :
4604 umtx_shm_find_reg(&key, ®);
4605 umtx_key_release(&key);
4606 if (error != 0)
4607 return (error);
4608 KASSERT(reg != NULL, ("no reg"));
4609 if ((flags & UMTX_SHM_DESTROY) != 0) {
4610 umtx_shm_unref_reg(reg, true);
4611 } else {
4612 #if 0
4613 #ifdef MAC
4614 error = mac_posixshm_check_open(td->td_ucred,
4615 reg->ushm_obj, FFLAGS(O_RDWR));
4616 if (error == 0)
4617 #endif
4618 error = shm_access(reg->ushm_obj, td->td_ucred,
4619 FFLAGS(O_RDWR));
4620 if (error == 0)
4621 #endif
4622 error = falloc_caps(td, &fp, &fd, O_CLOEXEC, NULL);
4623 if (error == 0) {
4624 shm_hold(reg->ushm_obj);
4625 finit(fp, FFLAGS(O_RDWR), DTYPE_SHM, reg->ushm_obj,
4626 &shm_ops);
4627 td->td_retval[0] = fd;
4628 fdrop(fp, td);
4629 }
4630 }
4631 umtx_shm_unref_reg(reg, false);
4632 return (error);
4633 }
4634
4635 static int
__umtx_op_shm(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops __unused)4636 __umtx_op_shm(struct thread *td, struct _umtx_op_args *uap,
4637 const struct umtx_copyops *ops __unused)
4638 {
4639
4640 return (umtx_shm(td, uap->uaddr1, uap->val));
4641 }
4642
4643 static int
__umtx_op_robust_lists(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4644 __umtx_op_robust_lists(struct thread *td, struct _umtx_op_args *uap,
4645 const struct umtx_copyops *ops)
4646 {
4647 struct umtx_robust_lists_params rb;
4648 int error;
4649
4650 if (ops->compat32) {
4651 if ((td->td_pflags2 & TDP2_COMPAT32RB) == 0 &&
4652 (td->td_rb_list != 0 || td->td_rbp_list != 0 ||
4653 td->td_rb_inact != 0))
4654 return (EBUSY);
4655 } else if ((td->td_pflags2 & TDP2_COMPAT32RB) != 0) {
4656 return (EBUSY);
4657 }
4658
4659 bzero(&rb, sizeof(rb));
4660 error = ops->copyin_robust_lists(uap->uaddr1, uap->val, &rb);
4661 if (error != 0)
4662 return (error);
4663
4664 if (ops->compat32)
4665 td->td_pflags2 |= TDP2_COMPAT32RB;
4666
4667 td->td_rb_list = rb.robust_list_offset;
4668 td->td_rbp_list = rb.robust_priv_list_offset;
4669 td->td_rb_inact = rb.robust_inact_offset;
4670 return (0);
4671 }
4672
4673 static int
__umtx_op_get_min_timeout(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4674 __umtx_op_get_min_timeout(struct thread *td, struct _umtx_op_args *uap,
4675 const struct umtx_copyops *ops)
4676 {
4677 long val;
4678 int error, val1;
4679
4680 val = sbttons(td->td_proc->p_umtx_min_timeout);
4681 if (ops->compat32) {
4682 val1 = (int)val;
4683 error = copyout(&val1, uap->uaddr1, sizeof(val1));
4684 } else {
4685 error = copyout(&val, uap->uaddr1, sizeof(val));
4686 }
4687 return (error);
4688 }
4689
4690 static int
__umtx_op_set_min_timeout(struct thread * td,struct _umtx_op_args * uap,const struct umtx_copyops * ops)4691 __umtx_op_set_min_timeout(struct thread *td, struct _umtx_op_args *uap,
4692 const struct umtx_copyops *ops)
4693 {
4694 if (uap->val < 0)
4695 return (EINVAL);
4696 td->td_proc->p_umtx_min_timeout = nstosbt(uap->val);
4697 return (0);
4698 }
4699
4700 #if defined(__i386__) || defined(__amd64__)
4701 /*
4702 * Provide the standard 32-bit definitions for x86, since native/compat32 use a
4703 * 32-bit time_t there. Other architectures just need the i386 definitions
4704 * along with their standard compat32.
4705 */
4706 struct timespecx32 {
4707 int64_t tv_sec;
4708 int32_t tv_nsec;
4709 };
4710
4711 struct umtx_timex32 {
4712 struct timespecx32 _timeout;
4713 uint32_t _flags;
4714 uint32_t _clockid;
4715 };
4716
4717 #ifndef __i386__
4718 #define timespeci386 timespec32
4719 #define umtx_timei386 umtx_time32
4720 #endif
4721 #else /* !__i386__ && !__amd64__ */
4722 /* 32-bit architectures can emulate i386, so define these almost everywhere. */
4723 struct timespeci386 {
4724 int32_t tv_sec;
4725 int32_t tv_nsec;
4726 };
4727
4728 struct umtx_timei386 {
4729 struct timespeci386 _timeout;
4730 uint32_t _flags;
4731 uint32_t _clockid;
4732 };
4733
4734 #if defined(__LP64__)
4735 #define timespecx32 timespec32
4736 #define umtx_timex32 umtx_time32
4737 #endif
4738 #endif
4739
4740 static int
umtx_copyin_robust_lists32(const void * uaddr,size_t size,struct umtx_robust_lists_params * rbp)4741 umtx_copyin_robust_lists32(const void *uaddr, size_t size,
4742 struct umtx_robust_lists_params *rbp)
4743 {
4744 struct umtx_robust_lists_params_compat32 rb32;
4745 int error;
4746
4747 if (size > sizeof(rb32))
4748 return (EINVAL);
4749 bzero(&rb32, sizeof(rb32));
4750 error = copyin(uaddr, &rb32, size);
4751 if (error != 0)
4752 return (error);
4753 CP(rb32, *rbp, robust_list_offset);
4754 CP(rb32, *rbp, robust_priv_list_offset);
4755 CP(rb32, *rbp, robust_inact_offset);
4756 return (0);
4757 }
4758
4759 #ifndef __i386__
4760 static inline int
umtx_copyin_timeouti386(const void * uaddr,struct timespec * tsp)4761 umtx_copyin_timeouti386(const void *uaddr, struct timespec *tsp)
4762 {
4763 struct timespeci386 ts32;
4764 int error;
4765
4766 error = copyin(uaddr, &ts32, sizeof(ts32));
4767 if (error == 0) {
4768 if (!timespecvalid_interval(&ts32))
4769 error = EINVAL;
4770 else {
4771 CP(ts32, *tsp, tv_sec);
4772 CP(ts32, *tsp, tv_nsec);
4773 }
4774 }
4775 return (error);
4776 }
4777
4778 static inline int
umtx_copyin_umtx_timei386(const void * uaddr,size_t size,struct _umtx_time * tp)4779 umtx_copyin_umtx_timei386(const void *uaddr, size_t size, struct _umtx_time *tp)
4780 {
4781 struct umtx_timei386 t32;
4782 int error;
4783
4784 t32._clockid = CLOCK_REALTIME;
4785 t32._flags = 0;
4786 if (size <= sizeof(t32._timeout))
4787 error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4788 else
4789 error = copyin(uaddr, &t32, sizeof(t32));
4790 if (error != 0)
4791 return (error);
4792 if (!timespecvalid_interval(&t32._timeout))
4793 return (EINVAL);
4794 TS_CP(t32, *tp, _timeout);
4795 CP(t32, *tp, _flags);
4796 CP(t32, *tp, _clockid);
4797 return (0);
4798 }
4799
4800 static int
umtx_copyout_timeouti386(void * uaddr,size_t sz,struct timespec * tsp)4801 umtx_copyout_timeouti386(void *uaddr, size_t sz, struct timespec *tsp)
4802 {
4803 struct timespeci386 remain32 = {
4804 .tv_sec = tsp->tv_sec,
4805 .tv_nsec = tsp->tv_nsec,
4806 };
4807
4808 /*
4809 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4810 * and we're only called if sz >= sizeof(timespec) as supplied in the
4811 * copyops.
4812 */
4813 KASSERT(sz >= sizeof(remain32),
4814 ("umtx_copyops specifies incorrect sizes"));
4815
4816 return (copyout(&remain32, uaddr, sizeof(remain32)));
4817 }
4818 #endif /* !__i386__ */
4819
4820 #if defined(__i386__) || defined(__LP64__)
4821 static inline int
umtx_copyin_timeoutx32(const void * uaddr,struct timespec * tsp)4822 umtx_copyin_timeoutx32(const void *uaddr, struct timespec *tsp)
4823 {
4824 struct timespecx32 ts32;
4825 int error;
4826
4827 error = copyin(uaddr, &ts32, sizeof(ts32));
4828 if (error == 0) {
4829 if (!timespecvalid_interval(&ts32))
4830 error = EINVAL;
4831 else {
4832 CP(ts32, *tsp, tv_sec);
4833 CP(ts32, *tsp, tv_nsec);
4834 }
4835 }
4836 return (error);
4837 }
4838
4839 static inline int
umtx_copyin_umtx_timex32(const void * uaddr,size_t size,struct _umtx_time * tp)4840 umtx_copyin_umtx_timex32(const void *uaddr, size_t size, struct _umtx_time *tp)
4841 {
4842 struct umtx_timex32 t32;
4843 int error;
4844
4845 t32._clockid = CLOCK_REALTIME;
4846 t32._flags = 0;
4847 if (size <= sizeof(t32._timeout))
4848 error = copyin(uaddr, &t32._timeout, sizeof(t32._timeout));
4849 else
4850 error = copyin(uaddr, &t32, sizeof(t32));
4851 if (error != 0)
4852 return (error);
4853 if (!timespecvalid_interval(&t32._timeout))
4854 return (EINVAL);
4855 TS_CP(t32, *tp, _timeout);
4856 CP(t32, *tp, _flags);
4857 CP(t32, *tp, _clockid);
4858 return (0);
4859 }
4860
4861 static int
umtx_copyout_timeoutx32(void * uaddr,size_t sz,struct timespec * tsp)4862 umtx_copyout_timeoutx32(void *uaddr, size_t sz, struct timespec *tsp)
4863 {
4864 struct timespecx32 remain32 = {
4865 .tv_sec = tsp->tv_sec,
4866 .tv_nsec = tsp->tv_nsec,
4867 };
4868
4869 /*
4870 * Should be guaranteed by the caller, sz == uaddr1 - sizeof(_umtx_time)
4871 * and we're only called if sz >= sizeof(timespec) as supplied in the
4872 * copyops.
4873 */
4874 KASSERT(sz >= sizeof(remain32),
4875 ("umtx_copyops specifies incorrect sizes"));
4876
4877 return (copyout(&remain32, uaddr, sizeof(remain32)));
4878 }
4879 #endif /* __i386__ || __LP64__ */
4880
4881 typedef int (*_umtx_op_func)(struct thread *td, struct _umtx_op_args *uap,
4882 const struct umtx_copyops *umtx_ops);
4883
4884 static const _umtx_op_func op_table[] = {
4885 #ifdef COMPAT_FREEBSD10
4886 [UMTX_OP_LOCK] = __umtx_op_lock_umtx,
4887 [UMTX_OP_UNLOCK] = __umtx_op_unlock_umtx,
4888 #else
4889 [UMTX_OP_LOCK] = __umtx_op_unimpl,
4890 [UMTX_OP_UNLOCK] = __umtx_op_unimpl,
4891 #endif
4892 [UMTX_OP_WAIT] = __umtx_op_wait,
4893 [UMTX_OP_WAKE] = __umtx_op_wake,
4894 [UMTX_OP_MUTEX_TRYLOCK] = __umtx_op_trylock_umutex,
4895 [UMTX_OP_MUTEX_LOCK] = __umtx_op_lock_umutex,
4896 [UMTX_OP_MUTEX_UNLOCK] = __umtx_op_unlock_umutex,
4897 [UMTX_OP_SET_CEILING] = __umtx_op_set_ceiling,
4898 [UMTX_OP_CV_WAIT] = __umtx_op_cv_wait,
4899 [UMTX_OP_CV_SIGNAL] = __umtx_op_cv_signal,
4900 [UMTX_OP_CV_BROADCAST] = __umtx_op_cv_broadcast,
4901 [UMTX_OP_WAIT_UINT] = __umtx_op_wait_uint,
4902 [UMTX_OP_RW_RDLOCK] = __umtx_op_rw_rdlock,
4903 [UMTX_OP_RW_WRLOCK] = __umtx_op_rw_wrlock,
4904 [UMTX_OP_RW_UNLOCK] = __umtx_op_rw_unlock,
4905 [UMTX_OP_WAIT_UINT_PRIVATE] = __umtx_op_wait_uint_private,
4906 [UMTX_OP_WAKE_PRIVATE] = __umtx_op_wake_private,
4907 [UMTX_OP_MUTEX_WAIT] = __umtx_op_wait_umutex,
4908 [UMTX_OP_MUTEX_WAKE] = __umtx_op_wake_umutex,
4909 #if defined(COMPAT_FREEBSD9) || defined(COMPAT_FREEBSD10)
4910 [UMTX_OP_SEM_WAIT] = __umtx_op_sem_wait,
4911 [UMTX_OP_SEM_WAKE] = __umtx_op_sem_wake,
4912 #else
4913 [UMTX_OP_SEM_WAIT] = __umtx_op_unimpl,
4914 [UMTX_OP_SEM_WAKE] = __umtx_op_unimpl,
4915 #endif
4916 [UMTX_OP_NWAKE_PRIVATE] = __umtx_op_nwake_private,
4917 [UMTX_OP_MUTEX_WAKE2] = __umtx_op_wake2_umutex,
4918 [UMTX_OP_SEM2_WAIT] = __umtx_op_sem2_wait,
4919 [UMTX_OP_SEM2_WAKE] = __umtx_op_sem2_wake,
4920 [UMTX_OP_SHM] = __umtx_op_shm,
4921 [UMTX_OP_ROBUST_LISTS] = __umtx_op_robust_lists,
4922 [UMTX_OP_GET_MIN_TIMEOUT] = __umtx_op_get_min_timeout,
4923 [UMTX_OP_SET_MIN_TIMEOUT] = __umtx_op_set_min_timeout,
4924 };
4925
4926 static const struct umtx_copyops umtx_native_ops = {
4927 .copyin_timeout = umtx_copyin_timeout,
4928 .copyin_umtx_time = umtx_copyin_umtx_time,
4929 .copyin_robust_lists = umtx_copyin_robust_lists,
4930 .copyout_timeout = umtx_copyout_timeout,
4931 .timespec_sz = sizeof(struct timespec),
4932 .umtx_time_sz = sizeof(struct _umtx_time),
4933 };
4934
4935 #ifndef __i386__
4936 static const struct umtx_copyops umtx_native_opsi386 = {
4937 .copyin_timeout = umtx_copyin_timeouti386,
4938 .copyin_umtx_time = umtx_copyin_umtx_timei386,
4939 .copyin_robust_lists = umtx_copyin_robust_lists32,
4940 .copyout_timeout = umtx_copyout_timeouti386,
4941 .timespec_sz = sizeof(struct timespeci386),
4942 .umtx_time_sz = sizeof(struct umtx_timei386),
4943 .compat32 = true,
4944 };
4945 #endif
4946
4947 #if defined(__i386__) || defined(__LP64__)
4948 /* i386 can emulate other 32-bit archs, too! */
4949 static const struct umtx_copyops umtx_native_opsx32 = {
4950 .copyin_timeout = umtx_copyin_timeoutx32,
4951 .copyin_umtx_time = umtx_copyin_umtx_timex32,
4952 .copyin_robust_lists = umtx_copyin_robust_lists32,
4953 .copyout_timeout = umtx_copyout_timeoutx32,
4954 .timespec_sz = sizeof(struct timespecx32),
4955 .umtx_time_sz = sizeof(struct umtx_timex32),
4956 .compat32 = true,
4957 };
4958
4959 #ifdef COMPAT_FREEBSD32
4960 #ifdef __amd64__
4961 #define umtx_native_ops32 umtx_native_opsi386
4962 #else
4963 #define umtx_native_ops32 umtx_native_opsx32
4964 #endif
4965 #endif /* COMPAT_FREEBSD32 */
4966 #endif /* __i386__ || __LP64__ */
4967
4968 #define UMTX_OP__FLAGS (UMTX_OP__32BIT | UMTX_OP__I386)
4969
4970 static int
kern__umtx_op(struct thread * td,void * obj,int op,unsigned long val,void * uaddr1,void * uaddr2,const struct umtx_copyops * ops)4971 kern__umtx_op(struct thread *td, void *obj, int op, unsigned long val,
4972 void *uaddr1, void *uaddr2, const struct umtx_copyops *ops)
4973 {
4974 struct _umtx_op_args uap = {
4975 .obj = obj,
4976 .op = op & ~UMTX_OP__FLAGS,
4977 .val = val,
4978 .uaddr1 = uaddr1,
4979 .uaddr2 = uaddr2
4980 };
4981
4982 if ((uap.op >= nitems(op_table)))
4983 return (EINVAL);
4984 return ((*op_table[uap.op])(td, &uap, ops));
4985 }
4986
4987 int
sys__umtx_op(struct thread * td,struct _umtx_op_args * uap)4988 sys__umtx_op(struct thread *td, struct _umtx_op_args *uap)
4989 {
4990 static const struct umtx_copyops *umtx_ops;
4991
4992 umtx_ops = &umtx_native_ops;
4993 #ifdef __LP64__
4994 if ((uap->op & (UMTX_OP__32BIT | UMTX_OP__I386)) != 0) {
4995 if ((uap->op & UMTX_OP__I386) != 0)
4996 umtx_ops = &umtx_native_opsi386;
4997 else
4998 umtx_ops = &umtx_native_opsx32;
4999 }
5000 #elif !defined(__i386__)
5001 /* We consider UMTX_OP__32BIT a nop on !i386 ILP32. */
5002 if ((uap->op & UMTX_OP__I386) != 0)
5003 umtx_ops = &umtx_native_opsi386;
5004 #else
5005 /* Likewise, UMTX_OP__I386 is a nop on i386. */
5006 if ((uap->op & UMTX_OP__32BIT) != 0)
5007 umtx_ops = &umtx_native_opsx32;
5008 #endif
5009 return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
5010 uap->uaddr2, umtx_ops));
5011 }
5012
5013 #ifdef COMPAT_FREEBSD32
5014 #ifdef COMPAT_FREEBSD10
5015 int
freebsd10_freebsd32__umtx_lock(struct thread * td,struct freebsd10_freebsd32__umtx_lock_args * uap)5016 freebsd10_freebsd32__umtx_lock(struct thread *td,
5017 struct freebsd10_freebsd32__umtx_lock_args *uap)
5018 {
5019 return (do_lock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid, NULL));
5020 }
5021
5022 int
freebsd10_freebsd32__umtx_unlock(struct thread * td,struct freebsd10_freebsd32__umtx_unlock_args * uap)5023 freebsd10_freebsd32__umtx_unlock(struct thread *td,
5024 struct freebsd10_freebsd32__umtx_unlock_args *uap)
5025 {
5026 return (do_unlock_umtx32(td, (uint32_t *)uap->umtx, td->td_tid));
5027 }
5028 #endif /* COMPAT_FREEBSD10 */
5029
5030 int
freebsd32__umtx_op(struct thread * td,struct freebsd32__umtx_op_args * uap)5031 freebsd32__umtx_op(struct thread *td, struct freebsd32__umtx_op_args *uap)
5032 {
5033
5034 return (kern__umtx_op(td, uap->obj, uap->op, uap->val, uap->uaddr1,
5035 uap->uaddr2, &umtx_native_ops32));
5036 }
5037 #endif /* COMPAT_FREEBSD32 */
5038
5039 void
umtx_thread_init(struct thread * td)5040 umtx_thread_init(struct thread *td)
5041 {
5042
5043 td->td_umtxq = umtxq_alloc();
5044 td->td_umtxq->uq_thread = td;
5045 }
5046
5047 void
umtx_thread_fini(struct thread * td)5048 umtx_thread_fini(struct thread *td)
5049 {
5050
5051 umtxq_free(td->td_umtxq);
5052 }
5053
5054 /*
5055 * It will be called when new thread is created, e.g fork().
5056 */
5057 void
umtx_thread_alloc(struct thread * td)5058 umtx_thread_alloc(struct thread *td)
5059 {
5060 struct umtx_q *uq;
5061
5062 uq = td->td_umtxq;
5063 uq->uq_inherited_pri = PRI_MAX;
5064
5065 KASSERT(uq->uq_flags == 0, ("uq_flags != 0"));
5066 KASSERT(uq->uq_thread == td, ("uq_thread != td"));
5067 KASSERT(uq->uq_pi_blocked == NULL, ("uq_pi_blocked != NULL"));
5068 KASSERT(TAILQ_EMPTY(&uq->uq_pi_contested), ("uq_pi_contested is not empty"));
5069 }
5070
5071 /*
5072 * exec() hook.
5073 *
5074 * Clear robust lists for all process' threads, not delaying the
5075 * cleanup to thread exit, since the relevant address space is
5076 * destroyed right now.
5077 */
5078 void
umtx_exec(struct proc * p)5079 umtx_exec(struct proc *p)
5080 {
5081 struct thread *td;
5082
5083 KASSERT(p == curproc, ("need curproc"));
5084 KASSERT((p->p_flag & P_HADTHREADS) == 0 ||
5085 (p->p_flag & P_STOPPED_SINGLE) != 0,
5086 ("curproc must be single-threaded"));
5087 /*
5088 * There is no need to lock the list as only this thread can be
5089 * running.
5090 */
5091 FOREACH_THREAD_IN_PROC(p, td) {
5092 KASSERT(td == curthread ||
5093 ((td->td_flags & TDF_BOUNDARY) != 0 && TD_IS_SUSPENDED(td)),
5094 ("running thread %p %p", p, td));
5095 umtx_thread_cleanup(td);
5096 td->td_rb_list = td->td_rbp_list = td->td_rb_inact = 0;
5097 }
5098
5099 p->p_umtx_min_timeout = 0;
5100 }
5101
5102 /*
5103 * thread exit hook.
5104 */
5105 void
umtx_thread_exit(struct thread * td)5106 umtx_thread_exit(struct thread *td)
5107 {
5108
5109 umtx_thread_cleanup(td);
5110 }
5111
5112 static int
umtx_read_uptr(struct thread * td,uintptr_t ptr,uintptr_t * res,bool compat32)5113 umtx_read_uptr(struct thread *td, uintptr_t ptr, uintptr_t *res, bool compat32)
5114 {
5115 u_long res1;
5116 uint32_t res32;
5117 int error;
5118
5119 if (compat32) {
5120 error = fueword32((void *)ptr, &res32);
5121 if (error == 0)
5122 res1 = res32;
5123 } else {
5124 error = fueword((void *)ptr, &res1);
5125 }
5126 if (error == 0)
5127 *res = res1;
5128 else
5129 error = EFAULT;
5130 return (error);
5131 }
5132
5133 static void
umtx_read_rb_list(struct thread * td,struct umutex * m,uintptr_t * rb_list,bool compat32)5134 umtx_read_rb_list(struct thread *td, struct umutex *m, uintptr_t *rb_list,
5135 bool compat32)
5136 {
5137 struct umutex32 m32;
5138
5139 if (compat32) {
5140 memcpy(&m32, m, sizeof(m32));
5141 *rb_list = m32.m_rb_lnk;
5142 } else {
5143 *rb_list = m->m_rb_lnk;
5144 }
5145 }
5146
5147 static int
umtx_handle_rb(struct thread * td,uintptr_t rbp,uintptr_t * rb_list,bool inact,bool compat32)5148 umtx_handle_rb(struct thread *td, uintptr_t rbp, uintptr_t *rb_list, bool inact,
5149 bool compat32)
5150 {
5151 struct umutex m;
5152 int error;
5153
5154 KASSERT(td->td_proc == curproc, ("need current vmspace"));
5155 error = copyin((void *)rbp, &m, sizeof(m));
5156 if (error != 0)
5157 return (error);
5158 if (rb_list != NULL)
5159 umtx_read_rb_list(td, &m, rb_list, compat32);
5160 if ((m.m_flags & UMUTEX_ROBUST) == 0)
5161 return (EINVAL);
5162 if ((m.m_owner & ~UMUTEX_CONTESTED) != td->td_tid)
5163 /* inact is cleared after unlock, allow the inconsistency */
5164 return (inact ? 0 : EINVAL);
5165 return (do_unlock_umutex(td, (struct umutex *)rbp, true));
5166 }
5167
5168 static void
umtx_cleanup_rb_list(struct thread * td,uintptr_t rb_list,uintptr_t * rb_inact,const char * name,bool compat32)5169 umtx_cleanup_rb_list(struct thread *td, uintptr_t rb_list, uintptr_t *rb_inact,
5170 const char *name, bool compat32)
5171 {
5172 int error, i;
5173 uintptr_t rbp;
5174 bool inact;
5175
5176 if (rb_list == 0)
5177 return;
5178 error = umtx_read_uptr(td, rb_list, &rbp, compat32);
5179 for (i = 0; error == 0 && rbp != 0 && i < umtx_max_rb; i++) {
5180 if (rbp == *rb_inact) {
5181 inact = true;
5182 *rb_inact = 0;
5183 } else
5184 inact = false;
5185 error = umtx_handle_rb(td, rbp, &rbp, inact, compat32);
5186 }
5187 if (i == umtx_max_rb && umtx_verbose_rb) {
5188 uprintf("comm %s pid %d: reached umtx %smax rb %d\n",
5189 td->td_proc->p_comm, td->td_proc->p_pid, name, umtx_max_rb);
5190 }
5191 if (error != 0 && umtx_verbose_rb) {
5192 uprintf("comm %s pid %d: handling %srb error %d\n",
5193 td->td_proc->p_comm, td->td_proc->p_pid, name, error);
5194 }
5195 }
5196
5197 /*
5198 * Clean up umtx data.
5199 */
5200 static void
umtx_thread_cleanup(struct thread * td)5201 umtx_thread_cleanup(struct thread *td)
5202 {
5203 struct umtx_q *uq;
5204 struct umtx_pi *pi;
5205 uintptr_t rb_inact;
5206 bool compat32;
5207
5208 /*
5209 * Disown pi mutexes.
5210 */
5211 uq = td->td_umtxq;
5212 if (uq != NULL) {
5213 if (uq->uq_inherited_pri != PRI_MAX ||
5214 !TAILQ_EMPTY(&uq->uq_pi_contested)) {
5215 mtx_lock(&umtx_lock);
5216 uq->uq_inherited_pri = PRI_MAX;
5217 while ((pi = TAILQ_FIRST(&uq->uq_pi_contested)) != NULL) {
5218 pi->pi_owner = NULL;
5219 TAILQ_REMOVE(&uq->uq_pi_contested, pi, pi_link);
5220 }
5221 mtx_unlock(&umtx_lock);
5222 }
5223 sched_lend_user_prio_cond(td, PRI_MAX);
5224 }
5225
5226 compat32 = (td->td_pflags2 & TDP2_COMPAT32RB) != 0;
5227 td->td_pflags2 &= ~TDP2_COMPAT32RB;
5228
5229 if (td->td_rb_inact == 0 && td->td_rb_list == 0 && td->td_rbp_list == 0)
5230 return;
5231
5232 /*
5233 * Handle terminated robust mutexes. Must be done after
5234 * robust pi disown, otherwise unlock could see unowned
5235 * entries.
5236 */
5237 rb_inact = td->td_rb_inact;
5238 if (rb_inact != 0)
5239 (void)umtx_read_uptr(td, rb_inact, &rb_inact, compat32);
5240 umtx_cleanup_rb_list(td, td->td_rb_list, &rb_inact, "", compat32);
5241 umtx_cleanup_rb_list(td, td->td_rbp_list, &rb_inact, "priv ", compat32);
5242 if (rb_inact != 0)
5243 (void)umtx_handle_rb(td, rb_inact, NULL, true, compat32);
5244 }
5245