Fix: deadlock when thread join is issued in read-side C.S.
[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 #include <stdio.h>
30 #include <pthread.h>
31 #include <signal.h>
32 #include <assert.h>
33 #include <stdlib.h>
34 #include <stdint.h>
35 #include <string.h>
36 #include <errno.h>
37 #include <poll.h>
38
39 #include "urcu/wfcqueue.h"
40 #include "urcu/map/urcu.h"
41 #include "urcu/static/urcu.h"
42 #include "urcu-pointer.h"
43 #include "urcu/tls-compat.h"
44
45 #include "urcu-die.h"
46 #include "urcu-wait.h"
47
48 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
49 #undef _LGPL_SOURCE
50 #include "urcu.h"
51 #define _LGPL_SOURCE
52
53 /*
54 * If a reader is really non-cooperative and refuses to commit its
55 * rcu_active_readers count to memory (there is no barrier in the reader
56 * per-se), kick it after 10 loops waiting for it.
57 */
58 #define KICK_READER_LOOPS 10
59
60 /*
61 * Active attempts to check for reader Q.S. before calling futex().
62 */
63 #define RCU_QS_ACTIVE_ATTEMPTS 100
64
65 /*
66 * RCU_MEMBARRIER is only possibly available on Linux.
67 */
68 #if defined(RCU_MEMBARRIER) && defined(__linux__)
69 #include <urcu/syscall-compat.h>
70 #endif
71
72 /* If the headers do not support SYS_membarrier, fall back on RCU_MB */
73 #ifdef SYS_membarrier
74 # define membarrier(...) syscall(SYS_membarrier, __VA_ARGS__)
75 #else
76 # define membarrier(...) -ENOSYS
77 #endif
78
79 #define MEMBARRIER_EXPEDITED (1 << 0)
80 #define MEMBARRIER_DELAYED (1 << 1)
81 #define MEMBARRIER_QUERY (1 << 16)
82
83 #ifdef RCU_MEMBARRIER
84 static int init_done;
85 int rcu_has_sys_membarrier;
86
87 void __attribute__((constructor)) rcu_init(void);
88 #endif
89
90 #ifdef RCU_MB
91 void rcu_init(void)
92 {
93 }
94 #endif
95
96 #ifdef RCU_SIGNAL
97 static int init_done;
98
99 void __attribute__((constructor)) rcu_init(void);
100 void __attribute__((destructor)) rcu_exit(void);
101 #endif
102
103 /*
104 * rcu_gp_lock ensures mutual exclusion between threads calling
105 * synchronize_rcu().
106 */
107 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
108 /*
109 * rcu_registry_lock ensures mutual exclusion between threads
110 * registering and unregistering themselves to/from the registry, and
111 * with threads reading that registry from synchronize_rcu(). However,
112 * this lock is not held all the way through the completion of awaiting
113 * for the grace period. It is sporadically released between iterations
114 * on the registry.
115 * rcu_registry_lock may nest inside rcu_gp_lock.
116 */
117 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
118 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
119
120 /*
121 * Written to only by each individual reader. Read by both the reader and the
122 * writers.
123 */
124 DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
125
126 static CDS_LIST_HEAD(registry);
127
128 /*
129 * Queue keeping threads awaiting to wait for a grace period. Contains
130 * struct gp_waiters_thread objects.
131 */
132 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
133
134 static void mutex_lock(pthread_mutex_t *mutex)
135 {
136 int ret;
137
138 #ifndef DISTRUST_SIGNALS_EXTREME
139 ret = pthread_mutex_lock(mutex);
140 if (ret)
141 urcu_die(ret);
142 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
143 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
144 if (ret != EBUSY && ret != EINTR)
145 urcu_die(ret);
146 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
147 cmm_smp_mb();
148 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
149 cmm_smp_mb();
150 }
151 poll(NULL,0,10);
152 }
153 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
154 }
155
156 static void mutex_unlock(pthread_mutex_t *mutex)
157 {
158 int ret;
159
160 ret = pthread_mutex_unlock(mutex);
161 if (ret)
162 urcu_die(ret);
163 }
164
165 #ifdef RCU_MEMBARRIER
166 static void smp_mb_master(int group)
167 {
168 if (caa_likely(rcu_has_sys_membarrier))
169 (void) membarrier(MEMBARRIER_EXPEDITED);
170 else
171 cmm_smp_mb();
172 }
173 #endif
174
175 #ifdef RCU_MB
176 static void smp_mb_master(int group)
177 {
178 cmm_smp_mb();
179 }
180 #endif
181
182 #ifdef RCU_SIGNAL
183 static void force_mb_all_readers(void)
184 {
185 struct rcu_reader *index;
186
187 /*
188 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
189 * compiler barriers around rcu read lock as real memory barriers.
190 */
191 if (cds_list_empty(&registry))
192 return;
193 /*
194 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
195 * a cache flush on architectures with non-coherent cache. Let's play
196 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
197 * cache flush is enforced.
198 */
199 cds_list_for_each_entry(index, &registry, node) {
200 CMM_STORE_SHARED(index->need_mb, 1);
201 pthread_kill(index->tid, SIGRCU);
202 }
203 /*
204 * Wait for sighandler (and thus mb()) to execute on every thread.
205 *
206 * Note that the pthread_kill() will never be executed on systems
207 * that correctly deliver signals in a timely manner. However, it
208 * is not uncommon for kernels to have bugs that can result in
209 * lost or unduly delayed signals.
210 *
211 * If you are seeing the below pthread_kill() executing much at
212 * all, we suggest testing the underlying kernel and filing the
213 * relevant bug report. For Linux kernels, we recommend getting
214 * the Linux Test Project (LTP).
215 */
216 cds_list_for_each_entry(index, &registry, node) {
217 while (CMM_LOAD_SHARED(index->need_mb)) {
218 pthread_kill(index->tid, SIGRCU);
219 poll(NULL, 0, 1);
220 }
221 }
222 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
223 }
224
225 static void smp_mb_master(int group)
226 {
227 force_mb_all_readers();
228 }
229 #endif /* #ifdef RCU_SIGNAL */
230
231 /*
232 * synchronize_rcu() waiting. Single thread.
233 */
234 static void wait_gp(void)
235 {
236 /* Read reader_gp before read futex */
237 smp_mb_master(RCU_MB_GROUP);
238 if (uatomic_read(&rcu_gp.futex) == -1)
239 futex_async(&rcu_gp.futex, FUTEX_WAIT, -1,
240 NULL, NULL, 0);
241 }
242
243 /*
244 * Always called with rcu_registry lock held. Releases this lock between
245 * iterations and grabs it again. Holds the lock when it returns.
246 */
247 static void wait_for_readers(struct cds_list_head *input_readers,
248 struct cds_list_head *cur_snap_readers,
249 struct cds_list_head *qsreaders)
250 {
251 unsigned int wait_loops = 0;
252 struct rcu_reader *index, *tmp;
253 #ifdef HAS_INCOHERENT_CACHES
254 unsigned int wait_gp_loops = 0;
255 #endif /* HAS_INCOHERENT_CACHES */
256
257 /*
258 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
259 * indicate quiescence (not nested), or observe the current
260 * rcu_gp.ctr value.
261 */
262 for (;;) {
263 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
264 wait_loops++;
265 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
266 uatomic_dec(&rcu_gp.futex);
267 /* Write futex before read reader_gp */
268 smp_mb_master(RCU_MB_GROUP);
269 }
270
271 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
272 switch (rcu_reader_state(&index->ctr)) {
273 case RCU_READER_ACTIVE_CURRENT:
274 if (cur_snap_readers) {
275 cds_list_move(&index->node,
276 cur_snap_readers);
277 break;
278 }
279 /* Fall-through */
280 case RCU_READER_INACTIVE:
281 cds_list_move(&index->node, qsreaders);
282 break;
283 case RCU_READER_ACTIVE_OLD:
284 /*
285 * Old snapshot. Leaving node in
286 * input_readers will make us busy-loop
287 * until the snapshot becomes current or
288 * the reader becomes inactive.
289 */
290 break;
291 }
292 }
293
294 #ifndef HAS_INCOHERENT_CACHES
295 if (cds_list_empty(input_readers)) {
296 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
297 /* Read reader_gp before write futex */
298 smp_mb_master(RCU_MB_GROUP);
299 uatomic_set(&rcu_gp.futex, 0);
300 }
301 break;
302 } else {
303 /* Temporarily unlock the registry lock. */
304 mutex_unlock(&rcu_registry_lock);
305 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
306 wait_gp();
307 else
308 caa_cpu_relax();
309 /* Re-lock the registry lock before the next loop. */
310 mutex_lock(&rcu_registry_lock);
311 }
312 #else /* #ifndef HAS_INCOHERENT_CACHES */
313 /*
314 * BUSY-LOOP. Force the reader thread to commit its
315 * URCU_TLS(rcu_reader).ctr update to memory if we wait
316 * for too long.
317 */
318 if (cds_list_empty(input_readers)) {
319 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
320 /* Read reader_gp before write futex */
321 smp_mb_master(RCU_MB_GROUP);
322 uatomic_set(&rcu_gp.futex, 0);
323 }
324 break;
325 } else {
326 if (wait_gp_loops == KICK_READER_LOOPS) {
327 smp_mb_master(RCU_MB_GROUP);
328 wait_gp_loops = 0;
329 }
330 /* Temporarily unlock the registry lock. */
331 mutex_unlock(&rcu_registry_lock);
332 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
333 wait_gp();
334 wait_gp_loops++;
335 } else {
336 caa_cpu_relax();
337 }
338 /* Re-lock the registry lock before the next loop. */
339 mutex_lock(&rcu_registry_lock);
340 }
341 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
342 }
343 }
344
345 void synchronize_rcu(void)
346 {
347 CDS_LIST_HEAD(cur_snap_readers);
348 CDS_LIST_HEAD(qsreaders);
349 DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
350 struct urcu_waiters waiters;
351
352 /*
353 * Add ourself to gp_waiters queue of threads awaiting to wait
354 * for a grace period. Proceed to perform the grace period only
355 * if we are the first thread added into the queue.
356 * The implicit memory barrier before urcu_wait_add()
357 * orders prior memory accesses of threads put into the wait
358 * queue before their insertion into the wait queue.
359 */
360 if (urcu_wait_add(&gp_waiters, &wait) != 0) {
361 /* Not first in queue: will be awakened by another thread. */
362 urcu_adaptative_busy_wait(&wait);
363 /* Order following memory accesses after grace period. */
364 cmm_smp_mb();
365 return;
366 }
367 /* We won't need to wake ourself up */
368 urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
369
370 mutex_lock(&rcu_gp_lock);
371
372 /*
373 * Move all waiters into our local queue.
374 */
375 urcu_move_waiters(&waiters, &gp_waiters);
376
377 mutex_lock(&rcu_registry_lock);
378
379 if (cds_list_empty(&registry))
380 goto out;
381
382 /*
383 * All threads should read qparity before accessing data structure
384 * where new ptr points to. Must be done within rcu_registry_lock
385 * because it iterates on reader threads.
386 */
387 /* Write new ptr before changing the qparity */
388 smp_mb_master(RCU_MB_GROUP);
389
390 /*
391 * Wait for readers to observe original parity or be quiescent.
392 * wait_for_readers() can release and grab again rcu_registry_lock
393 * interally.
394 */
395 wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
396
397 /*
398 * Must finish waiting for quiescent state for original parity before
399 * committing next rcu_gp.ctr update to memory. Failure to do so could
400 * result in the writer waiting forever while new readers are always
401 * accessing data (no progress). Enforce compiler-order of load
402 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
403 */
404 cmm_barrier();
405
406 /*
407 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
408 * model easier to understand. It does not have a big performance impact
409 * anyway, given this is the write-side.
410 */
411 cmm_smp_mb();
412
413 /* Switch parity: 0 -> 1, 1 -> 0 */
414 CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
415
416 /*
417 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
418 * state. Failure to do so could result in the writer waiting forever
419 * while new readers are always accessing data (no progress). Enforce
420 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
421 */
422 cmm_barrier();
423
424 /*
425 *
426 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
427 * model easier to understand. It does not have a big performance impact
428 * anyway, given this is the write-side.
429 */
430 cmm_smp_mb();
431
432 /*
433 * Wait for readers to observe new parity or be quiescent.
434 * wait_for_readers() can release and grab again rcu_registry_lock
435 * interally.
436 */
437 wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
438
439 /*
440 * Put quiescent reader list back into registry.
441 */
442 cds_list_splice(&qsreaders, &registry);
443
444 /*
445 * Finish waiting for reader threads before letting the old ptr
446 * being freed. Must be done within rcu_registry_lock because it
447 * iterates on reader threads.
448 */
449 smp_mb_master(RCU_MB_GROUP);
450 out:
451 mutex_unlock(&rcu_registry_lock);
452 mutex_unlock(&rcu_gp_lock);
453
454 /*
455 * Wakeup waiters only after we have completed the grace period
456 * and have ensured the memory barriers at the end of the grace
457 * period have been issued.
458 */
459 urcu_wake_all_waiters(&waiters);
460 }
461
462 /*
463 * library wrappers to be used by non-LGPL compatible source code.
464 */
465
466 void rcu_read_lock(void)
467 {
468 _rcu_read_lock();
469 }
470
471 void rcu_read_unlock(void)
472 {
473 _rcu_read_unlock();
474 }
475
476 int rcu_read_ongoing(void)
477 {
478 return _rcu_read_ongoing();
479 }
480
481 void rcu_register_thread(void)
482 {
483 URCU_TLS(rcu_reader).tid = pthread_self();
484 assert(URCU_TLS(rcu_reader).need_mb == 0);
485 assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
486
487 mutex_lock(&rcu_registry_lock);
488 rcu_init(); /* In case gcc does not support constructor attribute */
489 cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
490 mutex_unlock(&rcu_registry_lock);
491 }
492
493 void rcu_unregister_thread(void)
494 {
495 mutex_lock(&rcu_registry_lock);
496 cds_list_del(&URCU_TLS(rcu_reader).node);
497 mutex_unlock(&rcu_registry_lock);
498 }
499
500 #ifdef RCU_MEMBARRIER
501 void rcu_init(void)
502 {
503 if (init_done)
504 return;
505 init_done = 1;
506 if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
507 rcu_has_sys_membarrier = 1;
508 }
509 #endif
510
511 #ifdef RCU_SIGNAL
512 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
513 {
514 /*
515 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
516 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
517 * executed on.
518 */
519 cmm_smp_mb();
520 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
521 cmm_smp_mb();
522 }
523
524 /*
525 * rcu_init constructor. Called when the library is linked, but also when
526 * reader threads are calling rcu_register_thread().
527 * Should only be called by a single thread at a given time. This is ensured by
528 * holing the rcu_registry_lock from rcu_register_thread() or by running
529 * at library load time, which should not be executed by multiple
530 * threads nor concurrently with rcu_register_thread() anyway.
531 */
532 void rcu_init(void)
533 {
534 struct sigaction act;
535 int ret;
536
537 if (init_done)
538 return;
539 init_done = 1;
540
541 act.sa_sigaction = sigrcu_handler;
542 act.sa_flags = SA_SIGINFO | SA_RESTART;
543 sigemptyset(&act.sa_mask);
544 ret = sigaction(SIGRCU, &act, NULL);
545 if (ret)
546 urcu_die(errno);
547 }
548
549 void rcu_exit(void)
550 {
551 /*
552 * Don't unregister the SIGRCU signal handler anymore, because
553 * call_rcu threads could still be using it shortly before the
554 * application exits.
555 * Assertion disabled because call_rcu threads are now rcu
556 * readers, and left running at exit.
557 * assert(cds_list_empty(&registry));
558 */
559 }
560
561 #endif /* #ifdef RCU_SIGNAL */
562
563 DEFINE_RCU_FLAVOR(rcu_flavor);
564
565 #include "urcu-call-rcu-impl.h"
566 #include "urcu-defer-impl.h"
This page took 0.060433 seconds and 5 git commands to generate.