59f2e8f18fdddc5a781e25d2478ca86e18a9fee5
[urcu.git] / src / urcu.c
1 /*
2 * urcu.c
3 *
4 * Userspace RCU library
5 *
6 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
8 *
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 *
23 * IBM's contributions to this file may be relicensed under LGPLv2 or later.
24 */
25
26 #define URCU_NO_COMPAT_IDENTIFIERS
27 #define _BSD_SOURCE
28 #define _LGPL_SOURCE
29 #define _DEFAULT_SOURCE
30 #include <stdio.h>
31 #include <pthread.h>
32 #include <signal.h>
33 #include <stdlib.h>
34 #include <stdint.h>
35 #include <string.h>
36 #include <errno.h>
37 #include <stdbool.h>
38 #include <poll.h>
39
40 #include <urcu/config.h>
41 #include <urcu/assert.h>
42 #include <urcu/arch.h>
43 #include <urcu/wfcqueue.h>
44 #include <urcu/map/urcu.h>
45 #include <urcu/static/urcu.h>
46 #include <urcu/pointer.h>
47 #include <urcu/tls-compat.h>
48
49 #include "urcu-die.h"
50 #include "urcu-wait.h"
51 #include "urcu-utils.h"
52
53 #define URCU_API_MAP
54 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
55 #undef _LGPL_SOURCE
56 #include <urcu/urcu.h>
57 #define _LGPL_SOURCE
58
59 /*
60 * If a reader is really non-cooperative and refuses to commit its
61 * rcu_active_readers count to memory (there is no barrier in the reader
62 * per-se), kick it after 10 loops waiting for it.
63 */
64 #define KICK_READER_LOOPS 10
65
66 /*
67 * Active attempts to check for reader Q.S. before calling futex().
68 */
69 #define RCU_QS_ACTIVE_ATTEMPTS 100
70
71 /* If the headers do not support membarrier system call, fall back on RCU_MB */
72 #ifdef __NR_membarrier
73 # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__)
74 #else
75 # define membarrier(...) -ENOSYS
76 #endif
77
78 enum membarrier_cmd {
79 MEMBARRIER_CMD_QUERY = 0,
80 MEMBARRIER_CMD_SHARED = (1 << 0),
81 /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
82 /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
83 MEMBARRIER_CMD_PRIVATE_EXPEDITED = (1 << 3),
84 MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED = (1 << 4),
85 };
86
87 #ifdef RCU_MEMBARRIER
88 static int init_done;
89 static int urcu_memb_has_sys_membarrier_private_expedited;
90
91 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
92 /*
93 * Explicitly initialize to zero because we can't alias a non-static
94 * uninitialized variable.
95 */
96 int urcu_memb_has_sys_membarrier = 0;
97 #endif
98
99 void __attribute__((constructor)) rcu_init(void);
100 #endif
101
102 #ifdef RCU_MB
103 void rcu_init(void)
104 {
105 }
106 #endif
107
108 #ifdef RCU_SIGNAL
109 static int init_done;
110
111 void __attribute__((constructor)) rcu_init(void);
112 void __attribute__((destructor)) rcu_exit(void);
113 #endif
114
115 /*
116 * rcu_gp_lock ensures mutual exclusion between threads calling
117 * synchronize_rcu().
118 */
119 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
120 /*
121 * rcu_registry_lock ensures mutual exclusion between threads
122 * registering and unregistering themselves to/from the registry, and
123 * with threads reading that registry from synchronize_rcu(). However,
124 * this lock is not held all the way through the completion of awaiting
125 * for the grace period. It is sporadically released between iterations
126 * on the registry.
127 * rcu_registry_lock may nest inside rcu_gp_lock.
128 */
129 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
130 struct urcu_gp rcu_gp = { .ctr = URCU_GP_COUNT };
131
132 /*
133 * Written to only by each individual reader. Read by both the reader and the
134 * writers.
135 */
136 DEFINE_URCU_TLS(struct urcu_reader, rcu_reader);
137
138 static CDS_LIST_HEAD(registry);
139
140 /*
141 * Queue keeping threads awaiting to wait for a grace period. Contains
142 * struct gp_waiters_thread objects.
143 */
144 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
145
146 static void mutex_lock(pthread_mutex_t *mutex)
147 {
148 int ret;
149
150 #ifndef DISTRUST_SIGNALS_EXTREME
151 ret = pthread_mutex_lock(mutex);
152 if (ret)
153 urcu_die(ret);
154 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
155 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
156 if (ret != EBUSY && ret != EINTR)
157 urcu_die(ret);
158 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
159 cmm_smp_mb();
160 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
161 cmm_smp_mb();
162 }
163 (void) poll(NULL, 0, 10);
164 }
165 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
166 }
167
168 static void mutex_unlock(pthread_mutex_t *mutex)
169 {
170 int ret;
171
172 ret = pthread_mutex_unlock(mutex);
173 if (ret)
174 urcu_die(ret);
175 }
176
177 #ifdef RCU_MEMBARRIER
178 static void smp_mb_master(void)
179 {
180 if (caa_likely(urcu_memb_has_sys_membarrier)) {
181 if (membarrier(urcu_memb_has_sys_membarrier_private_expedited ?
182 MEMBARRIER_CMD_PRIVATE_EXPEDITED :
183 MEMBARRIER_CMD_SHARED, 0))
184 urcu_die(errno);
185 } else {
186 cmm_smp_mb();
187 }
188 }
189 #endif
190
191 #ifdef RCU_MB
192 static void smp_mb_master(void)
193 {
194 cmm_smp_mb();
195 }
196 #endif
197
198 #ifdef RCU_SIGNAL
199 static void force_mb_all_readers(void)
200 {
201 struct urcu_reader *index;
202
203 /*
204 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
205 * compiler barriers around rcu read lock as real memory barriers.
206 */
207 if (cds_list_empty(&registry))
208 return;
209 /*
210 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
211 * a cache flush on architectures with non-coherent cache. Let's play
212 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
213 * cache flush is enforced.
214 */
215 cds_list_for_each_entry(index, &registry, node) {
216 CMM_STORE_SHARED(index->need_mb, 1);
217 pthread_kill(index->tid, SIGRCU);
218 }
219 /*
220 * Wait for sighandler (and thus mb()) to execute on every thread.
221 *
222 * Note that the pthread_kill() will never be executed on systems
223 * that correctly deliver signals in a timely manner. However, it
224 * is not uncommon for kernels to have bugs that can result in
225 * lost or unduly delayed signals.
226 *
227 * If you are seeing the below pthread_kill() executing much at
228 * all, we suggest testing the underlying kernel and filing the
229 * relevant bug report. For Linux kernels, we recommend getting
230 * the Linux Test Project (LTP).
231 */
232 cds_list_for_each_entry(index, &registry, node) {
233 while (CMM_LOAD_SHARED(index->need_mb)) {
234 pthread_kill(index->tid, SIGRCU);
235 (void) poll(NULL, 0, 1);
236 }
237 }
238 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
239 }
240
241 static void smp_mb_master(void)
242 {
243 force_mb_all_readers();
244 }
245 #endif /* #ifdef RCU_SIGNAL */
246
247 /*
248 * synchronize_rcu() waiting. Single thread.
249 * Always called with rcu_registry lock held. Releases this lock and
250 * grabs it again. Holds the lock when it returns.
251 */
252 static void wait_gp(void)
253 {
254 /*
255 * Read reader_gp before read futex. smp_mb_master() needs to
256 * be called with the rcu registry lock held in RCU_SIGNAL
257 * flavor.
258 */
259 smp_mb_master();
260 /* Temporarily unlock the registry lock. */
261 mutex_unlock(&rcu_registry_lock);
262 while (uatomic_read(&rcu_gp.futex) == -1) {
263 if (!futex_async(&rcu_gp.futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
264 /*
265 * Prior queued wakeups queued by unrelated code
266 * using the same address can cause futex wait to
267 * return 0 even through the futex value is still
268 * -1 (spurious wakeups). Check the value again
269 * in user-space to validate whether it really
270 * differs from -1.
271 */
272 continue;
273 }
274 switch (errno) {
275 case EAGAIN:
276 /* Value already changed. */
277 goto end;
278 case EINTR:
279 /* Retry if interrupted by signal. */
280 break; /* Get out of switch. Check again. */
281 default:
282 /* Unexpected error. */
283 urcu_die(errno);
284 }
285 }
286 end:
287 /*
288 * Re-lock the registry lock before the next loop.
289 */
290 mutex_lock(&rcu_registry_lock);
291 }
292
293 /*
294 * Always called with rcu_registry lock held. Releases this lock between
295 * iterations and grabs it again. Holds the lock when it returns.
296 */
297 static void wait_for_readers(struct cds_list_head *input_readers,
298 struct cds_list_head *cur_snap_readers,
299 struct cds_list_head *qsreaders)
300 {
301 unsigned int wait_loops = 0;
302 struct urcu_reader *index, *tmp;
303 #ifdef HAS_INCOHERENT_CACHES
304 unsigned int wait_gp_loops = 0;
305 #endif /* HAS_INCOHERENT_CACHES */
306
307 /*
308 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
309 * indicate quiescence (not nested), or observe the current
310 * rcu_gp.ctr value.
311 */
312 for (;;) {
313 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
314 wait_loops++;
315 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
316 uatomic_dec(&rcu_gp.futex);
317 /* Write futex before read reader_gp */
318 smp_mb_master();
319 }
320
321 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
322 switch (urcu_common_reader_state(&rcu_gp, &index->ctr)) {
323 case URCU_READER_ACTIVE_CURRENT:
324 if (cur_snap_readers) {
325 cds_list_move(&index->node,
326 cur_snap_readers);
327 break;
328 }
329 /* Fall-through */
330 case URCU_READER_INACTIVE:
331 cds_list_move(&index->node, qsreaders);
332 break;
333 case URCU_READER_ACTIVE_OLD:
334 /*
335 * Old snapshot. Leaving node in
336 * input_readers will make us busy-loop
337 * until the snapshot becomes current or
338 * the reader becomes inactive.
339 */
340 break;
341 }
342 }
343
344 #ifndef HAS_INCOHERENT_CACHES
345 if (cds_list_empty(input_readers)) {
346 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
347 /* Read reader_gp before write futex */
348 smp_mb_master();
349 uatomic_set(&rcu_gp.futex, 0);
350 }
351 break;
352 } else {
353 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
354 /* wait_gp unlocks/locks registry lock. */
355 wait_gp();
356 } else {
357 /* Temporarily unlock the registry lock. */
358 mutex_unlock(&rcu_registry_lock);
359 caa_cpu_relax();
360 /*
361 * Re-lock the registry lock before the
362 * next loop.
363 */
364 mutex_lock(&rcu_registry_lock);
365 }
366 }
367 #else /* #ifndef HAS_INCOHERENT_CACHES */
368 /*
369 * BUSY-LOOP. Force the reader thread to commit its
370 * URCU_TLS(rcu_reader).ctr update to memory if we wait
371 * for too long.
372 */
373 if (cds_list_empty(input_readers)) {
374 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
375 /* Read reader_gp before write futex */
376 smp_mb_master();
377 uatomic_set(&rcu_gp.futex, 0);
378 }
379 break;
380 } else {
381 if (wait_gp_loops == KICK_READER_LOOPS) {
382 smp_mb_master();
383 wait_gp_loops = 0;
384 }
385 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
386 /* wait_gp unlocks/locks registry lock. */
387 wait_gp();
388 wait_gp_loops++;
389 } else {
390 /* Temporarily unlock the registry lock. */
391 mutex_unlock(&rcu_registry_lock);
392 caa_cpu_relax();
393 /*
394 * Re-lock the registry lock before the
395 * next loop.
396 */
397 mutex_lock(&rcu_registry_lock);
398 }
399 }
400 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
401 }
402 }
403
404 void synchronize_rcu(void)
405 {
406 CDS_LIST_HEAD(cur_snap_readers);
407 CDS_LIST_HEAD(qsreaders);
408 DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
409 struct urcu_waiters waiters;
410
411 /*
412 * Add ourself to gp_waiters queue of threads awaiting to wait
413 * for a grace period. Proceed to perform the grace period only
414 * if we are the first thread added into the queue.
415 * The implicit memory barrier before urcu_wait_add()
416 * orders prior memory accesses of threads put into the wait
417 * queue before their insertion into the wait queue.
418 */
419 if (urcu_wait_add(&gp_waiters, &wait) != 0) {
420 /* Not first in queue: will be awakened by another thread. */
421 urcu_adaptative_busy_wait(&wait);
422 /* Order following memory accesses after grace period. */
423 cmm_smp_mb();
424 return;
425 }
426 /* We won't need to wake ourself up */
427 urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
428
429 mutex_lock(&rcu_gp_lock);
430
431 /*
432 * Move all waiters into our local queue.
433 */
434 urcu_move_waiters(&waiters, &gp_waiters);
435
436 mutex_lock(&rcu_registry_lock);
437
438 if (cds_list_empty(&registry))
439 goto out;
440
441 /*
442 * All threads should read qparity before accessing data structure
443 * where new ptr points to. Must be done within rcu_registry_lock
444 * because it iterates on reader threads.
445 */
446 /* Write new ptr before changing the qparity */
447 smp_mb_master();
448
449 /*
450 * Wait for readers to observe original parity or be quiescent.
451 * wait_for_readers() can release and grab again rcu_registry_lock
452 * internally.
453 */
454 wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
455
456 /*
457 * Must finish waiting for quiescent state for original parity before
458 * committing next rcu_gp.ctr update to memory. Failure to do so could
459 * result in the writer waiting forever while new readers are always
460 * accessing data (no progress). Enforce compiler-order of load
461 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
462 */
463 cmm_barrier();
464
465 /*
466 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
467 * model easier to understand. It does not have a big performance impact
468 * anyway, given this is the write-side.
469 */
470 cmm_smp_mb();
471
472 /* Switch parity: 0 -> 1, 1 -> 0 */
473 CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_GP_CTR_PHASE);
474
475 /*
476 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
477 * state. Failure to do so could result in the writer waiting forever
478 * while new readers are always accessing data (no progress). Enforce
479 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
480 */
481 cmm_barrier();
482
483 /*
484 *
485 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
486 * model easier to understand. It does not have a big performance impact
487 * anyway, given this is the write-side.
488 */
489 cmm_smp_mb();
490
491 /*
492 * Wait for readers to observe new parity or be quiescent.
493 * wait_for_readers() can release and grab again rcu_registry_lock
494 * internally.
495 */
496 wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
497
498 /*
499 * Put quiescent reader list back into registry.
500 */
501 cds_list_splice(&qsreaders, &registry);
502
503 /*
504 * Finish waiting for reader threads before letting the old ptr
505 * being freed. Must be done within rcu_registry_lock because it
506 * iterates on reader threads.
507 */
508 smp_mb_master();
509 out:
510 mutex_unlock(&rcu_registry_lock);
511 mutex_unlock(&rcu_gp_lock);
512
513 /*
514 * Wakeup waiters only after we have completed the grace period
515 * and have ensured the memory barriers at the end of the grace
516 * period have been issued.
517 */
518 urcu_wake_all_waiters(&waiters);
519 }
520
521 /*
522 * library wrappers to be used by non-LGPL compatible source code.
523 */
524
525 void rcu_read_lock(void)
526 {
527 _rcu_read_lock();
528 }
529
530 void rcu_read_unlock(void)
531 {
532 _rcu_read_unlock();
533 }
534
535 int rcu_read_ongoing(void)
536 {
537 return _rcu_read_ongoing();
538 }
539
540 void rcu_register_thread(void)
541 {
542 URCU_TLS(rcu_reader).tid = pthread_self();
543 urcu_posix_assert(URCU_TLS(rcu_reader).need_mb == 0);
544 urcu_posix_assert(!(URCU_TLS(rcu_reader).ctr & URCU_GP_CTR_NEST_MASK));
545
546 mutex_lock(&rcu_registry_lock);
547 urcu_posix_assert(!URCU_TLS(rcu_reader).registered);
548 URCU_TLS(rcu_reader).registered = 1;
549 rcu_init(); /* In case gcc does not support constructor attribute */
550 cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
551 mutex_unlock(&rcu_registry_lock);
552 }
553
554 void rcu_unregister_thread(void)
555 {
556 mutex_lock(&rcu_registry_lock);
557 urcu_posix_assert(URCU_TLS(rcu_reader).registered);
558 URCU_TLS(rcu_reader).registered = 0;
559 cds_list_del(&URCU_TLS(rcu_reader).node);
560 mutex_unlock(&rcu_registry_lock);
561 }
562
563 #ifdef RCU_MEMBARRIER
564
565 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
566 static
567 void rcu_sys_membarrier_status(bool available)
568 {
569 if (!available)
570 abort();
571 }
572 #else
573 static
574 void rcu_sys_membarrier_status(bool available)
575 {
576 if (!available)
577 return;
578 urcu_memb_has_sys_membarrier = 1;
579 }
580 #endif
581
582 static
583 void rcu_sys_membarrier_init(void)
584 {
585 bool available = false;
586 int mask;
587
588 mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
589 if (mask >= 0) {
590 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
591 if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
592 urcu_die(errno);
593 urcu_memb_has_sys_membarrier_private_expedited = 1;
594 available = true;
595 } else if (mask & MEMBARRIER_CMD_SHARED) {
596 available = true;
597 }
598 }
599 rcu_sys_membarrier_status(available);
600 }
601
602 void rcu_init(void)
603 {
604 if (init_done)
605 return;
606 init_done = 1;
607 rcu_sys_membarrier_init();
608 }
609 #endif
610
611 #ifdef RCU_SIGNAL
612 static void sigrcu_handler(int signo __attribute__((unused)),
613 siginfo_t *siginfo __attribute__((unused)),
614 void *context __attribute__((unused)))
615 {
616 /*
617 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
618 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
619 * executed on.
620 */
621 cmm_smp_mb();
622 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
623 cmm_smp_mb();
624 }
625
626 /*
627 * rcu_init constructor. Called when the library is linked, but also when
628 * reader threads are calling rcu_register_thread().
629 * Should only be called by a single thread at a given time. This is ensured by
630 * holing the rcu_registry_lock from rcu_register_thread() or by running
631 * at library load time, which should not be executed by multiple
632 * threads nor concurrently with rcu_register_thread() anyway.
633 */
634 void rcu_init(void)
635 {
636 struct sigaction act;
637 int ret;
638
639 if (init_done)
640 return;
641 init_done = 1;
642
643 act.sa_sigaction = sigrcu_handler;
644 act.sa_flags = SA_SIGINFO | SA_RESTART;
645 sigemptyset(&act.sa_mask);
646 ret = sigaction(SIGRCU, &act, NULL);
647 if (ret)
648 urcu_die(errno);
649 }
650
651 void rcu_exit(void)
652 {
653 /*
654 * Don't unregister the SIGRCU signal handler anymore, because
655 * call_rcu threads could still be using it shortly before the
656 * application exits.
657 * Assertion disabled because call_rcu threads are now rcu
658 * readers, and left running at exit.
659 * urcu_posix_assert(cds_list_empty(&registry));
660 */
661 }
662
663 #endif /* #ifdef RCU_SIGNAL */
664
665 DEFINE_RCU_FLAVOR(rcu_flavor);
666
667 #include "urcu-call-rcu-impl.h"
668 #include "urcu-defer-impl.h"
This page took 0.041909 seconds and 3 git commands to generate.