Fix: format string signedness
[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 * The ABI of sys_membarrier changed after its original implementation.
68 * Disable it for now. Use RCU_MB flavor instead.
69 */
70 #define membarrier(...) -ENOSYS
71
72 #define MEMBARRIER_EXPEDITED (1 << 0)
73 #define MEMBARRIER_DELAYED (1 << 1)
74 #define MEMBARRIER_QUERY (1 << 16)
75
76 #ifdef RCU_MEMBARRIER
77 static int init_done;
78 int rcu_has_sys_membarrier;
79
80 void __attribute__((constructor)) rcu_init(void);
81 #endif
82
83 #ifdef RCU_MB
84 void rcu_init(void)
85 {
86 }
87 #endif
88
89 #ifdef RCU_SIGNAL
90 static int init_done;
91
92 void __attribute__((constructor)) rcu_init(void);
93 void __attribute__((destructor)) rcu_exit(void);
94 #endif
95
96 /*
97 * rcu_gp_lock ensures mutual exclusion between threads calling
98 * synchronize_rcu().
99 */
100 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
101 /*
102 * rcu_registry_lock ensures mutual exclusion between threads
103 * registering and unregistering themselves to/from the registry, and
104 * with threads reading that registry from synchronize_rcu(). However,
105 * this lock is not held all the way through the completion of awaiting
106 * for the grace period. It is sporadically released between iterations
107 * on the registry.
108 * rcu_registry_lock may nest inside rcu_gp_lock.
109 */
110 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
111 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
112
113 /*
114 * Written to only by each individual reader. Read by both the reader and the
115 * writers.
116 */
117 __DEFINE_URCU_TLS_GLOBAL(struct rcu_reader, rcu_reader);
118
119 static CDS_LIST_HEAD(registry);
120
121 /*
122 * Queue keeping threads awaiting to wait for a grace period. Contains
123 * struct gp_waiters_thread objects.
124 */
125 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
126
127 static void mutex_lock(pthread_mutex_t *mutex)
128 {
129 int ret;
130
131 #ifndef DISTRUST_SIGNALS_EXTREME
132 ret = pthread_mutex_lock(mutex);
133 if (ret)
134 urcu_die(ret);
135 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
136 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
137 if (ret != EBUSY && ret != EINTR)
138 urcu_die(ret);
139 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
140 cmm_smp_mb();
141 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
142 cmm_smp_mb();
143 }
144 poll(NULL,0,10);
145 }
146 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
147 }
148
149 static void mutex_unlock(pthread_mutex_t *mutex)
150 {
151 int ret;
152
153 ret = pthread_mutex_unlock(mutex);
154 if (ret)
155 urcu_die(ret);
156 }
157
158 #ifdef RCU_MEMBARRIER
159 static void smp_mb_master(int group)
160 {
161 if (caa_likely(rcu_has_sys_membarrier))
162 (void) membarrier(MEMBARRIER_EXPEDITED);
163 else
164 cmm_smp_mb();
165 }
166 #endif
167
168 #ifdef RCU_MB
169 static void smp_mb_master(int group)
170 {
171 cmm_smp_mb();
172 }
173 #endif
174
175 #ifdef RCU_SIGNAL
176 static void force_mb_all_readers(void)
177 {
178 struct rcu_reader *index;
179
180 /*
181 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
182 * compiler barriers around rcu read lock as real memory barriers.
183 */
184 if (cds_list_empty(&registry))
185 return;
186 /*
187 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
188 * a cache flush on architectures with non-coherent cache. Let's play
189 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
190 * cache flush is enforced.
191 */
192 cds_list_for_each_entry(index, &registry, node) {
193 CMM_STORE_SHARED(index->need_mb, 1);
194 pthread_kill(index->tid, SIGRCU);
195 }
196 /*
197 * Wait for sighandler (and thus mb()) to execute on every thread.
198 *
199 * Note that the pthread_kill() will never be executed on systems
200 * that correctly deliver signals in a timely manner. However, it
201 * is not uncommon for kernels to have bugs that can result in
202 * lost or unduly delayed signals.
203 *
204 * If you are seeing the below pthread_kill() executing much at
205 * all, we suggest testing the underlying kernel and filing the
206 * relevant bug report. For Linux kernels, we recommend getting
207 * the Linux Test Project (LTP).
208 */
209 cds_list_for_each_entry(index, &registry, node) {
210 while (CMM_LOAD_SHARED(index->need_mb)) {
211 pthread_kill(index->tid, SIGRCU);
212 poll(NULL, 0, 1);
213 }
214 }
215 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
216 }
217
218 static void smp_mb_master(int group)
219 {
220 force_mb_all_readers();
221 }
222 #endif /* #ifdef RCU_SIGNAL */
223
224 /*
225 * synchronize_rcu() waiting. Single thread.
226 */
227 static void wait_gp(void)
228 {
229 /* Read reader_gp before read futex */
230 smp_mb_master(RCU_MB_GROUP);
231 if (uatomic_read(&rcu_gp.futex) != -1)
232 return;
233 while (futex_async(&rcu_gp.futex, FUTEX_WAIT, -1,
234 NULL, NULL, 0)) {
235 switch (errno) {
236 case EWOULDBLOCK:
237 /* Value already changed. */
238 return;
239 case EINTR:
240 /* Retry if interrupted by signal. */
241 break; /* Get out of switch. */
242 default:
243 /* Unexpected error. */
244 urcu_die(errno);
245 }
246 }
247 }
248
249 /*
250 * Always called with rcu_registry lock held. Releases this lock between
251 * iterations and grabs it again. Holds the lock when it returns.
252 */
253 static void wait_for_readers(struct cds_list_head *input_readers,
254 struct cds_list_head *cur_snap_readers,
255 struct cds_list_head *qsreaders)
256 {
257 unsigned int wait_loops = 0;
258 struct rcu_reader *index, *tmp;
259 #ifdef HAS_INCOHERENT_CACHES
260 unsigned int wait_gp_loops = 0;
261 #endif /* HAS_INCOHERENT_CACHES */
262
263 /*
264 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
265 * indicate quiescence (not nested), or observe the current
266 * rcu_gp.ctr value.
267 */
268 for (;;) {
269 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
270 wait_loops++;
271 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
272 uatomic_dec(&rcu_gp.futex);
273 /* Write futex before read reader_gp */
274 smp_mb_master(RCU_MB_GROUP);
275 }
276
277 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
278 switch (rcu_reader_state(&index->ctr)) {
279 case RCU_READER_ACTIVE_CURRENT:
280 if (cur_snap_readers) {
281 cds_list_move(&index->node,
282 cur_snap_readers);
283 break;
284 }
285 /* Fall-through */
286 case RCU_READER_INACTIVE:
287 cds_list_move(&index->node, qsreaders);
288 break;
289 case RCU_READER_ACTIVE_OLD:
290 /*
291 * Old snapshot. Leaving node in
292 * input_readers will make us busy-loop
293 * until the snapshot becomes current or
294 * the reader becomes inactive.
295 */
296 break;
297 }
298 }
299
300 #ifndef HAS_INCOHERENT_CACHES
301 if (cds_list_empty(input_readers)) {
302 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
303 /* Read reader_gp before write futex */
304 smp_mb_master(RCU_MB_GROUP);
305 uatomic_set(&rcu_gp.futex, 0);
306 }
307 break;
308 } else {
309 /* Temporarily unlock the registry lock. */
310 mutex_unlock(&rcu_registry_lock);
311 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
312 wait_gp();
313 else
314 caa_cpu_relax();
315 /* Re-lock the registry lock before the next loop. */
316 mutex_lock(&rcu_registry_lock);
317 }
318 #else /* #ifndef HAS_INCOHERENT_CACHES */
319 /*
320 * BUSY-LOOP. Force the reader thread to commit its
321 * URCU_TLS(rcu_reader).ctr update to memory if we wait
322 * for too long.
323 */
324 if (cds_list_empty(input_readers)) {
325 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
326 /* Read reader_gp before write futex */
327 smp_mb_master(RCU_MB_GROUP);
328 uatomic_set(&rcu_gp.futex, 0);
329 }
330 break;
331 } else {
332 if (wait_gp_loops == KICK_READER_LOOPS) {
333 smp_mb_master(RCU_MB_GROUP);
334 wait_gp_loops = 0;
335 }
336 /* Temporarily unlock the registry lock. */
337 mutex_unlock(&rcu_registry_lock);
338 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
339 wait_gp();
340 wait_gp_loops++;
341 } else {
342 caa_cpu_relax();
343 }
344 /* Re-lock the registry lock before the next loop. */
345 mutex_lock(&rcu_registry_lock);
346 }
347 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
348 }
349 }
350
351 void synchronize_rcu(void)
352 {
353 CDS_LIST_HEAD(cur_snap_readers);
354 CDS_LIST_HEAD(qsreaders);
355 DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
356 struct urcu_waiters waiters;
357
358 /*
359 * Add ourself to gp_waiters queue of threads awaiting to wait
360 * for a grace period. Proceed to perform the grace period only
361 * if we are the first thread added into the queue.
362 * The implicit memory barrier before urcu_wait_add()
363 * orders prior memory accesses of threads put into the wait
364 * queue before their insertion into the wait queue.
365 */
366 if (urcu_wait_add(&gp_waiters, &wait) != 0) {
367 /* Not first in queue: will be awakened by another thread. */
368 urcu_adaptative_busy_wait(&wait);
369 /* Order following memory accesses after grace period. */
370 cmm_smp_mb();
371 return;
372 }
373 /* We won't need to wake ourself up */
374 urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
375
376 mutex_lock(&rcu_gp_lock);
377
378 /*
379 * Move all waiters into our local queue.
380 */
381 urcu_move_waiters(&waiters, &gp_waiters);
382
383 mutex_lock(&rcu_registry_lock);
384
385 if (cds_list_empty(&registry))
386 goto out;
387
388 /*
389 * All threads should read qparity before accessing data structure
390 * where new ptr points to. Must be done within rcu_registry_lock
391 * because it iterates on reader threads.
392 */
393 /* Write new ptr before changing the qparity */
394 smp_mb_master(RCU_MB_GROUP);
395
396 /*
397 * Wait for readers to observe original parity or be quiescent.
398 * wait_for_readers() can release and grab again rcu_registry_lock
399 * interally.
400 */
401 wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
402
403 /*
404 * Must finish waiting for quiescent state for original parity before
405 * committing next rcu_gp.ctr update to memory. Failure to do so could
406 * result in the writer waiting forever while new readers are always
407 * accessing data (no progress). Enforce compiler-order of load
408 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
409 */
410 cmm_barrier();
411
412 /*
413 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
414 * model easier to understand. It does not have a big performance impact
415 * anyway, given this is the write-side.
416 */
417 cmm_smp_mb();
418
419 /* Switch parity: 0 -> 1, 1 -> 0 */
420 CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
421
422 /*
423 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
424 * state. Failure to do so could result in the writer waiting forever
425 * while new readers are always accessing data (no progress). Enforce
426 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
427 */
428 cmm_barrier();
429
430 /*
431 *
432 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
433 * model easier to understand. It does not have a big performance impact
434 * anyway, given this is the write-side.
435 */
436 cmm_smp_mb();
437
438 /*
439 * Wait for readers to observe new parity or be quiescent.
440 * wait_for_readers() can release and grab again rcu_registry_lock
441 * interally.
442 */
443 wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
444
445 /*
446 * Put quiescent reader list back into registry.
447 */
448 cds_list_splice(&qsreaders, &registry);
449
450 /*
451 * Finish waiting for reader threads before letting the old ptr
452 * being freed. Must be done within rcu_registry_lock because it
453 * iterates on reader threads.
454 */
455 smp_mb_master(RCU_MB_GROUP);
456 out:
457 mutex_unlock(&rcu_registry_lock);
458 mutex_unlock(&rcu_gp_lock);
459
460 /*
461 * Wakeup waiters only after we have completed the grace period
462 * and have ensured the memory barriers at the end of the grace
463 * period have been issued.
464 */
465 urcu_wake_all_waiters(&waiters);
466 }
467
468 /*
469 * library wrappers to be used by non-LGPL compatible source code.
470 */
471
472 void rcu_read_lock(void)
473 {
474 _rcu_read_lock();
475 }
476
477 void rcu_read_unlock(void)
478 {
479 _rcu_read_unlock();
480 }
481
482 int rcu_read_ongoing(void)
483 {
484 return _rcu_read_ongoing();
485 }
486
487 void rcu_register_thread(void)
488 {
489 URCU_TLS(rcu_reader).tid = pthread_self();
490 assert(URCU_TLS(rcu_reader).need_mb == 0);
491 assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
492
493 mutex_lock(&rcu_registry_lock);
494 rcu_init(); /* In case gcc does not support constructor attribute */
495 cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
496 mutex_unlock(&rcu_registry_lock);
497 }
498
499 void rcu_unregister_thread(void)
500 {
501 mutex_lock(&rcu_registry_lock);
502 cds_list_del(&URCU_TLS(rcu_reader).node);
503 mutex_unlock(&rcu_registry_lock);
504 }
505
506 #ifdef RCU_MEMBARRIER
507 void rcu_init(void)
508 {
509 if (init_done)
510 return;
511 init_done = 1;
512 if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
513 rcu_has_sys_membarrier = 1;
514 }
515 #endif
516
517 #ifdef RCU_SIGNAL
518 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
519 {
520 /*
521 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
522 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
523 * executed on.
524 */
525 cmm_smp_mb();
526 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
527 cmm_smp_mb();
528 }
529
530 /*
531 * rcu_init constructor. Called when the library is linked, but also when
532 * reader threads are calling rcu_register_thread().
533 * Should only be called by a single thread at a given time. This is ensured by
534 * holing the rcu_registry_lock from rcu_register_thread() or by running
535 * at library load time, which should not be executed by multiple
536 * threads nor concurrently with rcu_register_thread() anyway.
537 */
538 void rcu_init(void)
539 {
540 struct sigaction act;
541 int ret;
542
543 if (init_done)
544 return;
545 init_done = 1;
546
547 act.sa_sigaction = sigrcu_handler;
548 act.sa_flags = SA_SIGINFO | SA_RESTART;
549 sigemptyset(&act.sa_mask);
550 ret = sigaction(SIGRCU, &act, NULL);
551 if (ret)
552 urcu_die(errno);
553 }
554
555 void rcu_exit(void)
556 {
557 /*
558 * Don't unregister the SIGRCU signal handler anymore, because
559 * call_rcu threads could still be using it shortly before the
560 * application exits.
561 * Assertion disabled because call_rcu threads are now rcu
562 * readers, and left running at exit.
563 * assert(cds_list_empty(&registry));
564 */
565 }
566
567 #endif /* #ifdef RCU_SIGNAL */
568
569 DEFINE_RCU_FLAVOR(rcu_flavor);
570
571 #include "urcu-call-rcu-impl.h"
572 #include "urcu-defer-impl.h"
This page took 0.040959 seconds and 4 git commands to generate.