e77306584d77f0fc012f6bfbae466452b3556b74
[urcu.git] / 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 _BSD_SOURCE
27 #define _GNU_SOURCE
28 #define _LGPL_SOURCE
29 #define _DEFAULT_SOURCE
30 #include <stdio.h>
31 #include <pthread.h>
32 #include <signal.h>
33 #include <assert.h>
34 #include <stdlib.h>
35 #include <stdint.h>
36 #include <string.h>
37 #include <errno.h>
38 #include <poll.h>
39
40 #include "urcu/wfcqueue.h"
41 #include "urcu/map/urcu.h"
42 #include "urcu/static/urcu.h"
43 #include "urcu-pointer.h"
44 #include "urcu/tls-compat.h"
45
46 #include "urcu-die.h"
47 #include "urcu-wait.h"
48
49 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
50 #undef _LGPL_SOURCE
51 #include "urcu.h"
52 #define _LGPL_SOURCE
53
54 /*
55 * If a reader is really non-cooperative and refuses to commit its
56 * rcu_active_readers count to memory (there is no barrier in the reader
57 * per-se), kick it after 10 loops waiting for it.
58 */
59 #define KICK_READER_LOOPS 10
60
61 /*
62 * Active attempts to check for reader Q.S. before calling futex().
63 */
64 #define RCU_QS_ACTIVE_ATTEMPTS 100
65
66 /*
67 * RCU_MEMBARRIER is only possibly available on Linux.
68 */
69 #if defined(RCU_MEMBARRIER) && defined(__linux__)
70 #include <urcu/syscall-compat.h>
71 #endif
72
73 /* If the headers do not support SYS_membarrier, fall back on RCU_MB */
74 #ifdef SYS_membarrier
75 # define membarrier(...) syscall(SYS_membarrier, __VA_ARGS__)
76 #else
77 # define membarrier(...) -ENOSYS
78 #endif
79
80 #define MEMBARRIER_EXPEDITED (1 << 0)
81 #define MEMBARRIER_DELAYED (1 << 1)
82 #define MEMBARRIER_QUERY (1 << 16)
83
84 #ifdef RCU_MEMBARRIER
85 static int init_done;
86 int rcu_has_sys_membarrier;
87
88 void __attribute__((constructor)) rcu_init(void);
89 #endif
90
91 #ifdef RCU_MB
92 void rcu_init(void)
93 {
94 }
95 #endif
96
97 #ifdef RCU_SIGNAL
98 static int init_done;
99
100 void __attribute__((constructor)) rcu_init(void);
101 void __attribute__((destructor)) rcu_exit(void);
102 #endif
103
104 /*
105 * rcu_gp_lock ensures mutual exclusion between threads calling
106 * synchronize_rcu().
107 */
108 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
109 /*
110 * rcu_registry_lock ensures mutual exclusion between threads
111 * registering and unregistering themselves to/from the registry, and
112 * with threads reading that registry from synchronize_rcu(). However,
113 * this lock is not held all the way through the completion of awaiting
114 * for the grace period. It is sporadically released between iterations
115 * on the registry.
116 * rcu_registry_lock may nest inside rcu_gp_lock.
117 */
118 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
119 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
120
121 /*
122 * Written to only by each individual reader. Read by both the reader and the
123 * writers.
124 */
125 DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
126
127 static CDS_LIST_HEAD(registry);
128
129 /*
130 * Queue keeping threads awaiting to wait for a grace period. Contains
131 * struct gp_waiters_thread objects.
132 */
133 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
134
135 static void mutex_lock(pthread_mutex_t *mutex)
136 {
137 int ret;
138
139 #ifndef DISTRUST_SIGNALS_EXTREME
140 ret = pthread_mutex_lock(mutex);
141 if (ret)
142 urcu_die(ret);
143 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
144 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
145 if (ret != EBUSY && ret != EINTR)
146 urcu_die(ret);
147 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
148 cmm_smp_mb();
149 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
150 cmm_smp_mb();
151 }
152 (void) poll(NULL, 0, 10);
153 }
154 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
155 }
156
157 static void mutex_unlock(pthread_mutex_t *mutex)
158 {
159 int ret;
160
161 ret = pthread_mutex_unlock(mutex);
162 if (ret)
163 urcu_die(ret);
164 }
165
166 #ifdef RCU_MEMBARRIER
167 static void smp_mb_master(int group)
168 {
169 if (caa_likely(rcu_has_sys_membarrier))
170 (void) membarrier(MEMBARRIER_EXPEDITED);
171 else
172 cmm_smp_mb();
173 }
174 #endif
175
176 #ifdef RCU_MB
177 static void smp_mb_master(int group)
178 {
179 cmm_smp_mb();
180 }
181 #endif
182
183 #ifdef RCU_SIGNAL
184 static void force_mb_all_readers(void)
185 {
186 struct rcu_reader *index;
187
188 /*
189 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
190 * compiler barriers around rcu read lock as real memory barriers.
191 */
192 if (cds_list_empty(&registry))
193 return;
194 /*
195 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
196 * a cache flush on architectures with non-coherent cache. Let's play
197 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
198 * cache flush is enforced.
199 */
200 cds_list_for_each_entry(index, &registry, node) {
201 CMM_STORE_SHARED(index->need_mb, 1);
202 pthread_kill(index->tid, SIGRCU);
203 }
204 /*
205 * Wait for sighandler (and thus mb()) to execute on every thread.
206 *
207 * Note that the pthread_kill() will never be executed on systems
208 * that correctly deliver signals in a timely manner. However, it
209 * is not uncommon for kernels to have bugs that can result in
210 * lost or unduly delayed signals.
211 *
212 * If you are seeing the below pthread_kill() executing much at
213 * all, we suggest testing the underlying kernel and filing the
214 * relevant bug report. For Linux kernels, we recommend getting
215 * the Linux Test Project (LTP).
216 */
217 cds_list_for_each_entry(index, &registry, node) {
218 while (CMM_LOAD_SHARED(index->need_mb)) {
219 pthread_kill(index->tid, SIGRCU);
220 (void) poll(NULL, 0, 1);
221 }
222 }
223 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
224 }
225
226 static void smp_mb_master(int group)
227 {
228 force_mb_all_readers();
229 }
230 #endif /* #ifdef RCU_SIGNAL */
231
232 /*
233 * synchronize_rcu() waiting. Single thread.
234 */
235 static void wait_gp(void)
236 {
237 /* Read reader_gp before read futex */
238 smp_mb_master(RCU_MB_GROUP);
239 if (uatomic_read(&rcu_gp.futex) != -1)
240 return;
241 while (futex_async(&rcu_gp.futex, FUTEX_WAIT, -1,
242 NULL, NULL, 0)) {
243 switch (errno) {
244 case EWOULDBLOCK:
245 /* Value already changed. */
246 return;
247 case EINTR:
248 /* Retry if interrupted by signal. */
249 break; /* Get out of switch. */
250 default:
251 /* Unexpected error. */
252 urcu_die(errno);
253 }
254 }
255 }
256
257 /*
258 * Always called with rcu_registry lock held. Releases this lock between
259 * iterations and grabs it again. Holds the lock when it returns.
260 */
261 static void wait_for_readers(struct cds_list_head *input_readers,
262 struct cds_list_head *cur_snap_readers,
263 struct cds_list_head *qsreaders)
264 {
265 unsigned int wait_loops = 0;
266 struct rcu_reader *index, *tmp;
267 #ifdef HAS_INCOHERENT_CACHES
268 unsigned int wait_gp_loops = 0;
269 #endif /* HAS_INCOHERENT_CACHES */
270
271 /*
272 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
273 * indicate quiescence (not nested), or observe the current
274 * rcu_gp.ctr value.
275 */
276 for (;;) {
277 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
278 wait_loops++;
279 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
280 uatomic_dec(&rcu_gp.futex);
281 /* Write futex before read reader_gp */
282 smp_mb_master(RCU_MB_GROUP);
283 }
284
285 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
286 switch (rcu_reader_state(&index->ctr)) {
287 case RCU_READER_ACTIVE_CURRENT:
288 if (cur_snap_readers) {
289 cds_list_move(&index->node,
290 cur_snap_readers);
291 break;
292 }
293 /* Fall-through */
294 case RCU_READER_INACTIVE:
295 cds_list_move(&index->node, qsreaders);
296 break;
297 case RCU_READER_ACTIVE_OLD:
298 /*
299 * Old snapshot. Leaving node in
300 * input_readers will make us busy-loop
301 * until the snapshot becomes current or
302 * the reader becomes inactive.
303 */
304 break;
305 }
306 }
307
308 #ifndef HAS_INCOHERENT_CACHES
309 if (cds_list_empty(input_readers)) {
310 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
311 /* Read reader_gp before write futex */
312 smp_mb_master(RCU_MB_GROUP);
313 uatomic_set(&rcu_gp.futex, 0);
314 }
315 break;
316 } else {
317 /* Temporarily unlock the registry lock. */
318 mutex_unlock(&rcu_registry_lock);
319 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
320 wait_gp();
321 else
322 caa_cpu_relax();
323 /* Re-lock the registry lock before the next loop. */
324 mutex_lock(&rcu_registry_lock);
325 }
326 #else /* #ifndef HAS_INCOHERENT_CACHES */
327 /*
328 * BUSY-LOOP. Force the reader thread to commit its
329 * URCU_TLS(rcu_reader).ctr update to memory if we wait
330 * for too long.
331 */
332 if (cds_list_empty(input_readers)) {
333 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
334 /* Read reader_gp before write futex */
335 smp_mb_master(RCU_MB_GROUP);
336 uatomic_set(&rcu_gp.futex, 0);
337 }
338 break;
339 } else {
340 if (wait_gp_loops == KICK_READER_LOOPS) {
341 smp_mb_master(RCU_MB_GROUP);
342 wait_gp_loops = 0;
343 }
344 /* Temporarily unlock the registry lock. */
345 mutex_unlock(&rcu_registry_lock);
346 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
347 wait_gp();
348 wait_gp_loops++;
349 } else {
350 caa_cpu_relax();
351 }
352 /* Re-lock the registry lock before the next loop. */
353 mutex_lock(&rcu_registry_lock);
354 }
355 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
356 }
357 }
358
359 void synchronize_rcu(void)
360 {
361 CDS_LIST_HEAD(cur_snap_readers);
362 CDS_LIST_HEAD(qsreaders);
363 DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
364 struct urcu_waiters waiters;
365
366 /*
367 * Add ourself to gp_waiters queue of threads awaiting to wait
368 * for a grace period. Proceed to perform the grace period only
369 * if we are the first thread added into the queue.
370 * The implicit memory barrier before urcu_wait_add()
371 * orders prior memory accesses of threads put into the wait
372 * queue before their insertion into the wait queue.
373 */
374 if (urcu_wait_add(&gp_waiters, &wait) != 0) {
375 /* Not first in queue: will be awakened by another thread. */
376 urcu_adaptative_busy_wait(&wait);
377 /* Order following memory accesses after grace period. */
378 cmm_smp_mb();
379 return;
380 }
381 /* We won't need to wake ourself up */
382 urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
383
384 mutex_lock(&rcu_gp_lock);
385
386 /*
387 * Move all waiters into our local queue.
388 */
389 urcu_move_waiters(&waiters, &gp_waiters);
390
391 mutex_lock(&rcu_registry_lock);
392
393 if (cds_list_empty(&registry))
394 goto out;
395
396 /*
397 * All threads should read qparity before accessing data structure
398 * where new ptr points to. Must be done within rcu_registry_lock
399 * because it iterates on reader threads.
400 */
401 /* Write new ptr before changing the qparity */
402 smp_mb_master(RCU_MB_GROUP);
403
404 /*
405 * Wait for readers to observe original parity or be quiescent.
406 * wait_for_readers() can release and grab again rcu_registry_lock
407 * interally.
408 */
409 wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
410
411 /*
412 * Must finish waiting for quiescent state for original parity before
413 * committing next rcu_gp.ctr update to memory. Failure to do so could
414 * result in the writer waiting forever while new readers are always
415 * accessing data (no progress). Enforce compiler-order of load
416 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
417 */
418 cmm_barrier();
419
420 /*
421 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
422 * model easier to understand. It does not have a big performance impact
423 * anyway, given this is the write-side.
424 */
425 cmm_smp_mb();
426
427 /* Switch parity: 0 -> 1, 1 -> 0 */
428 CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
429
430 /*
431 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
432 * state. Failure to do so could result in the writer waiting forever
433 * while new readers are always accessing data (no progress). Enforce
434 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
435 */
436 cmm_barrier();
437
438 /*
439 *
440 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
441 * model easier to understand. It does not have a big performance impact
442 * anyway, given this is the write-side.
443 */
444 cmm_smp_mb();
445
446 /*
447 * Wait for readers to observe new parity or be quiescent.
448 * wait_for_readers() can release and grab again rcu_registry_lock
449 * interally.
450 */
451 wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
452
453 /*
454 * Put quiescent reader list back into registry.
455 */
456 cds_list_splice(&qsreaders, &registry);
457
458 /*
459 * Finish waiting for reader threads before letting the old ptr
460 * being freed. Must be done within rcu_registry_lock because it
461 * iterates on reader threads.
462 */
463 smp_mb_master(RCU_MB_GROUP);
464 out:
465 mutex_unlock(&rcu_registry_lock);
466 mutex_unlock(&rcu_gp_lock);
467
468 /*
469 * Wakeup waiters only after we have completed the grace period
470 * and have ensured the memory barriers at the end of the grace
471 * period have been issued.
472 */
473 urcu_wake_all_waiters(&waiters);
474 }
475
476 /*
477 * library wrappers to be used by non-LGPL compatible source code.
478 */
479
480 void rcu_read_lock(void)
481 {
482 _rcu_read_lock();
483 }
484
485 void rcu_read_unlock(void)
486 {
487 _rcu_read_unlock();
488 }
489
490 int rcu_read_ongoing(void)
491 {
492 return _rcu_read_ongoing();
493 }
494
495 void rcu_register_thread(void)
496 {
497 URCU_TLS(rcu_reader).tid = pthread_self();
498 assert(URCU_TLS(rcu_reader).need_mb == 0);
499 assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
500
501 mutex_lock(&rcu_registry_lock);
502 rcu_init(); /* In case gcc does not support constructor attribute */
503 cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
504 mutex_unlock(&rcu_registry_lock);
505 }
506
507 void rcu_unregister_thread(void)
508 {
509 mutex_lock(&rcu_registry_lock);
510 cds_list_del(&URCU_TLS(rcu_reader).node);
511 mutex_unlock(&rcu_registry_lock);
512 }
513
514 #ifdef RCU_MEMBARRIER
515 void rcu_init(void)
516 {
517 if (init_done)
518 return;
519 init_done = 1;
520 if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
521 rcu_has_sys_membarrier = 1;
522 }
523 #endif
524
525 #ifdef RCU_SIGNAL
526 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
527 {
528 /*
529 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
530 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
531 * executed on.
532 */
533 cmm_smp_mb();
534 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
535 cmm_smp_mb();
536 }
537
538 /*
539 * rcu_init constructor. Called when the library is linked, but also when
540 * reader threads are calling rcu_register_thread().
541 * Should only be called by a single thread at a given time. This is ensured by
542 * holing the rcu_registry_lock from rcu_register_thread() or by running
543 * at library load time, which should not be executed by multiple
544 * threads nor concurrently with rcu_register_thread() anyway.
545 */
546 void rcu_init(void)
547 {
548 struct sigaction act;
549 int ret;
550
551 if (init_done)
552 return;
553 init_done = 1;
554
555 act.sa_sigaction = sigrcu_handler;
556 act.sa_flags = SA_SIGINFO | SA_RESTART;
557 sigemptyset(&act.sa_mask);
558 ret = sigaction(SIGRCU, &act, NULL);
559 if (ret)
560 urcu_die(errno);
561 }
562
563 void rcu_exit(void)
564 {
565 /*
566 * Don't unregister the SIGRCU signal handler anymore, because
567 * call_rcu threads could still be using it shortly before the
568 * application exits.
569 * Assertion disabled because call_rcu threads are now rcu
570 * readers, and left running at exit.
571 * assert(cds_list_empty(&registry));
572 */
573 }
574
575 #endif /* #ifdef RCU_SIGNAL */
576
577 DEFINE_RCU_FLAVOR(rcu_flavor);
578
579 #include "urcu-call-rcu-impl.h"
580 #include "urcu-defer-impl.h"
This page took 0.060591 seconds and 3 git commands to generate.