Fix: rename RCU_DEBUG to DEBUG_RCU in urcu-qsbr.h
[urcu.git] / urcu.c
CommitLineData
b257a10b
MD
1/*
2 * urcu.c
3 *
4 * Userspace RCU library
5 *
6982d6d7 6 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
af02d47e 7 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
b257a10b 8 *
af02d47e
MD
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
54843abc
PM
22 *
23 * IBM's contributions to this file may be relicensed under LGPLv2 or later.
b257a10b
MD
24 */
25
fdf01eed 26#define _BSD_SOURCE
c1d2c60b 27#define _GNU_SOURCE
71c811bf 28#define _LGPL_SOURCE
27b012e2
MD
29#include <stdio.h>
30#include <pthread.h>
31#include <signal.h>
32#include <assert.h>
f69f195a 33#include <stdlib.h>
6d841bc2 34#include <stdint.h>
f69f195a 35#include <string.h>
09a9f986 36#include <errno.h>
e8043c1b 37#include <poll.h>
27b012e2 38
d73fb81f 39#include "urcu/wfcqueue.h"
57760d44 40#include "urcu/map/urcu.h"
af7c2dbe 41#include "urcu/static/urcu.h"
618b2595 42#include "urcu-pointer.h"
bd252a04 43#include "urcu/tls-compat.h"
71c811bf 44
4a6d7378 45#include "urcu-die.h"
5bffdd5d 46#include "urcu-wait.h"
4a6d7378 47
121a5d44 48/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
71c811bf 49#undef _LGPL_SOURCE
27b012e2 50#include "urcu.h"
71c811bf 51#define _LGPL_SOURCE
27b012e2 52
3a71751e
PB
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
9340c38d 56 * per-se), kick it after 10 loops waiting for it.
3a71751e 57 */
9340c38d 58#define KICK_READER_LOOPS 10
3a71751e
PB
59
60/*
61 * Active attempts to check for reader Q.S. before calling futex().
62 */
63#define RCU_QS_ACTIVE_ATTEMPTS 100
64
553b7eb9
MD
65/*
66 * RCU_MEMBARRIER is only possibly available on Linux.
67 */
68#if defined(RCU_MEMBARRIER) && defined(__linux__)
9ba261bd 69#include <urcu/syscall-compat.h>
553b7eb9
MD
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
fdf01eed 83#ifdef RCU_MEMBARRIER
834a45ba 84static int init_done;
1de4df4b 85int rcu_has_sys_membarrier;
834a45ba 86
02be5561 87void __attribute__((constructor)) rcu_init(void);
fdf01eed
MD
88#endif
89
90#ifdef RCU_MB
02be5561 91void rcu_init(void)
e90a6e9c
MD
92{
93}
94#endif
8a5fb4c9 95
fdf01eed
MD
96#ifdef RCU_SIGNAL
97static int init_done;
98
99void __attribute__((constructor)) rcu_init(void);
100void __attribute__((destructor)) rcu_exit(void);
101#endif
102
6abb4bd5 103static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
4de0cd31 104struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
27b012e2 105
b0d5e790
MD
106/*
107 * Written to only by each individual reader. Read by both the reader and the
108 * writers.
109 */
bd252a04 110DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
27b012e2 111
16aa9ee8 112static CDS_LIST_HEAD(registry);
27b012e2 113
5bffdd5d
MD
114/*
115 * Queue keeping threads awaiting to wait for a grace period. Contains
116 * struct gp_waiters_thread objects.
117 */
118static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
119
6abb4bd5 120static void mutex_lock(pthread_mutex_t *mutex)
41718ff9
MD
121{
122 int ret;
09a9f986
PM
123
124#ifndef DISTRUST_SIGNALS_EXTREME
6abb4bd5 125 ret = pthread_mutex_lock(mutex);
4a6d7378
MD
126 if (ret)
127 urcu_die(ret);
09a9f986 128#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
6abb4bd5 129 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
4a6d7378
MD
130 if (ret != EBUSY && ret != EINTR)
131 urcu_die(ret);
bd252a04 132 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
5481ddb3 133 cmm_smp_mb();
bd252a04 134 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
5481ddb3 135 cmm_smp_mb();
09a9f986
PM
136 }
137 poll(NULL,0,10);
138 }
139#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
41718ff9
MD
140}
141
6abb4bd5 142static void mutex_unlock(pthread_mutex_t *mutex)
41718ff9
MD
143{
144 int ret;
145
6abb4bd5 146 ret = pthread_mutex_unlock(mutex);
4a6d7378
MD
147 if (ret)
148 urcu_die(ret);
41718ff9
MD
149}
150
fdf01eed 151#ifdef RCU_MEMBARRIER
25cc6d18 152static void smp_mb_master(int group)
fdf01eed 153{
1de4df4b 154 if (caa_likely(rcu_has_sys_membarrier))
553b7eb9 155 (void) membarrier(MEMBARRIER_EXPEDITED);
fdf01eed 156 else
5481ddb3 157 cmm_smp_mb();
fdf01eed
MD
158}
159#endif
160
02be5561 161#ifdef RCU_MB
25cc6d18 162static void smp_mb_master(int group)
40e140c9 163{
5481ddb3 164 cmm_smp_mb();
40e140c9 165}
fdf01eed
MD
166#endif
167
168#ifdef RCU_SIGNAL
78ff9419 169static void force_mb_all_readers(void)
27b012e2 170{
02be5561 171 struct rcu_reader *index;
e3b0cef0 172
27b012e2 173 /*
5481ddb3 174 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
27b012e2
MD
175 * compiler barriers around rcu read lock as real memory barriers.
176 */
16aa9ee8 177 if (cds_list_empty(&registry))
27b012e2 178 return;
3a86deba 179 /*
5481ddb3 180 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
157dca95 181 * a cache flush on architectures with non-coherent cache. Let's play
5481ddb3 182 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
157dca95 183 * cache flush is enforced.
3a86deba 184 */
16aa9ee8 185 cds_list_for_each_entry(index, &registry, node) {
6cf3827c 186 CMM_STORE_SHARED(index->need_mb, 1);
02be5561 187 pthread_kill(index->tid, SIGRCU);
09a9f986 188 }
27b012e2
MD
189 /*
190 * Wait for sighandler (and thus mb()) to execute on every thread.
09a9f986
PM
191 *
192 * Note that the pthread_kill() will never be executed on systems
193 * that correctly deliver signals in a timely manner. However, it
194 * is not uncommon for kernels to have bugs that can result in
195 * lost or unduly delayed signals.
196 *
197 * If you are seeing the below pthread_kill() executing much at
198 * all, we suggest testing the underlying kernel and filing the
199 * relevant bug report. For Linux kernels, we recommend getting
200 * the Linux Test Project (LTP).
27b012e2 201 */
16aa9ee8 202 cds_list_for_each_entry(index, &registry, node) {
6cf3827c 203 while (CMM_LOAD_SHARED(index->need_mb)) {
02be5561 204 pthread_kill(index->tid, SIGRCU);
09a9f986
PM
205 poll(NULL, 0, 1);
206 }
207 }
5481ddb3 208 cmm_smp_mb(); /* read ->need_mb before ending the barrier */
27b012e2 209}
9d7e3f89 210
25cc6d18 211static void smp_mb_master(int group)
9d7e3f89
MD
212{
213 force_mb_all_readers();
214}
fdf01eed 215#endif /* #ifdef RCU_SIGNAL */
27b012e2 216
bc6c15bb
MD
217/*
218 * synchronize_rcu() waiting. Single thread.
219 */
cfe78e25 220static void wait_gp(void)
bc6c15bb 221{
cfe78e25 222 /* Read reader_gp before read futex */
25cc6d18 223 smp_mb_master(RCU_MB_GROUP);
ed1b099e
LJ
224 if (uatomic_read(&rcu_gp.futex) == -1)
225 futex_async(&rcu_gp.futex, FUTEX_WAIT, -1,
cfe78e25 226 NULL, NULL, 0);
bc6c15bb
MD
227}
228
fd189fa5
MD
229static void wait_for_readers(struct cds_list_head *input_readers,
230 struct cds_list_head *cur_snap_readers,
231 struct cds_list_head *qsreaders)
27b012e2 232{
9340c38d 233 unsigned int wait_loops = 0;
02be5561 234 struct rcu_reader *index, *tmp;
9340c38d
MD
235#ifdef HAS_INCOHERENT_CACHES
236 unsigned int wait_gp_loops = 0;
237#endif /* HAS_INCOHERENT_CACHES */
27b012e2 238
40e140c9 239 /*
c9488684
MD
240 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
241 * indicate quiescence (not nested), or observe the current
ed1b099e 242 * rcu_gp.ctr value.
27b012e2 243 */
cfe78e25 244 for (;;) {
5e81fed7
MD
245 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
246 wait_loops++;
9340c38d 247 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
ed1b099e 248 uatomic_dec(&rcu_gp.futex);
cfe78e25 249 /* Write futex before read reader_gp */
25cc6d18 250 smp_mb_master(RCU_MB_GROUP);
cfe78e25
MD
251 }
252
fd189fa5
MD
253 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
254 switch (rcu_reader_state(&index->ctr)) {
255 case RCU_READER_ACTIVE_CURRENT:
256 if (cur_snap_readers) {
257 cds_list_move(&index->node,
258 cur_snap_readers);
259 break;
260 }
261 /* Fall-through */
262 case RCU_READER_INACTIVE:
263 cds_list_move(&index->node, qsreaders);
264 break;
265 case RCU_READER_ACTIVE_OLD:
266 /*
267 * Old snapshot. Leaving node in
268 * input_readers will make us busy-loop
269 * until the snapshot becomes current or
270 * the reader becomes inactive.
271 */
272 break;
273 }
cfe78e25
MD
274 }
275
e8043c1b 276#ifndef HAS_INCOHERENT_CACHES
fd189fa5 277 if (cds_list_empty(input_readers)) {
9340c38d 278 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
cfe78e25 279 /* Read reader_gp before write futex */
25cc6d18 280 smp_mb_master(RCU_MB_GROUP);
ed1b099e 281 uatomic_set(&rcu_gp.futex, 0);
bc6c15bb 282 }
cfe78e25
MD
283 break;
284 } else {
9340c38d 285 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
cfe78e25
MD
286 wait_gp();
287 else
06f22bdb 288 caa_cpu_relax();
bc6c15bb 289 }
e8043c1b 290#else /* #ifndef HAS_INCOHERENT_CACHES */
27b012e2 291 /*
40e140c9 292 * BUSY-LOOP. Force the reader thread to commit its
bd252a04
MD
293 * URCU_TLS(rcu_reader).ctr update to memory if we wait
294 * for too long.
27b012e2 295 */
fd189fa5 296 if (cds_list_empty(input_readers)) {
9340c38d 297 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
cfe78e25 298 /* Read reader_gp before write futex */
25cc6d18 299 smp_mb_master(RCU_MB_GROUP);
ed1b099e 300 uatomic_set(&rcu_gp.futex, 0);
cfe78e25
MD
301 }
302 break;
303 } else {
9340c38d 304 if (wait_gp_loops == KICK_READER_LOOPS) {
25cc6d18 305 smp_mb_master(RCU_MB_GROUP);
9340c38d
MD
306 wait_gp_loops = 0;
307 }
308 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
309 wait_gp();
310 wait_gp_loops++;
311 } else {
06f22bdb 312 caa_cpu_relax();
40e140c9
MD
313 }
314 }
e8043c1b 315#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
27b012e2 316 }
27b012e2
MD
317}
318
9598a481 319void synchronize_rcu(void)
2bc59bd7 320{
fd189fa5
MD
321 CDS_LIST_HEAD(cur_snap_readers);
322 CDS_LIST_HEAD(qsreaders);
5bffdd5d
MD
323 DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
324 struct urcu_waiters waiters;
325
326 /*
327 * Add ourself to gp_waiters queue of threads awaiting to wait
328 * for a grace period. Proceed to perform the grace period only
329 * if we are the first thread added into the queue.
330 * The implicit memory barrier before urcu_wait_add()
331 * orders prior memory accesses of threads put into the wait
332 * queue before their insertion into the wait queue.
333 */
334 if (urcu_wait_add(&gp_waiters, &wait) != 0) {
335 /* Not first in queue: will be awakened by another thread. */
336 urcu_adaptative_busy_wait(&wait);
337 /* Order following memory accesses after grace period. */
338 cmm_smp_mb();
339 return;
340 }
341 /* We won't need to wake ourself up */
342 urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
fd189fa5 343
6abb4bd5 344 mutex_lock(&rcu_gp_lock);
135530fd 345
5bffdd5d
MD
346 /*
347 * Move all waiters into our local queue.
348 */
349 urcu_move_waiters(&waiters, &gp_waiters);
350
16aa9ee8 351 if (cds_list_empty(&registry))
2dfb8b5e
MD
352 goto out;
353
9598a481 354 /* All threads should read qparity before accessing data structure
6abb4bd5
MD
355 * where new ptr points to. Must be done within rcu_gp_lock because it
356 * iterates on reader threads.*/
9598a481 357 /* Write new ptr before changing the qparity */
25cc6d18 358 smp_mb_master(RCU_MB_GROUP);
9598a481 359
9598a481 360 /*
c9488684 361 * Wait for readers to observe original parity or be quiescent.
9598a481 362 */
fd189fa5 363 wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
9598a481
MD
364
365 /*
c9488684 366 * Must finish waiting for quiescent state for original parity before
ed1b099e 367 * committing next rcu_gp.ctr update to memory. Failure to do so could
d40fde2c
MD
368 * result in the writer waiting forever while new readers are always
369 * accessing data (no progress). Enforce compiler-order of load
ed1b099e 370 * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
9598a481 371 */
5481ddb3 372 cmm_barrier();
9598a481 373
5dba80f9 374 /*
5481ddb3 375 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
5dba80f9
MD
376 * model easier to understand. It does not have a big performance impact
377 * anyway, given this is the write-side.
378 */
5481ddb3 379 cmm_smp_mb();
67c2d80b 380
c9488684 381 /* Switch parity: 0 -> 1, 1 -> 0 */
ed1b099e 382 CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
c9488684
MD
383
384 /*
ed1b099e 385 * Must commit rcu_gp.ctr update to memory before waiting for quiescent
c9488684
MD
386 * state. Failure to do so could result in the writer waiting forever
387 * while new readers are always accessing data (no progress). Enforce
ed1b099e 388 * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
c9488684
MD
389 */
390 cmm_barrier();
391
392 /*
393 *
394 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
395 * model easier to understand. It does not have a big performance impact
396 * anyway, given this is the write-side.
397 */
398 cmm_smp_mb();
399
9598a481 400 /*
c9488684 401 * Wait for readers to observe new parity or be quiescent.
9598a481 402 */
fd189fa5
MD
403 wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
404
405 /*
406 * Put quiescent reader list back into registry.
407 */
408 cds_list_splice(&qsreaders, &registry);
9598a481 409
9598a481 410 /* Finish waiting for reader threads before letting the old ptr being
6abb4bd5
MD
411 * freed. Must be done within rcu_gp_lock because it iterates on reader
412 * threads. */
25cc6d18 413 smp_mb_master(RCU_MB_GROUP);
2dfb8b5e 414out:
6abb4bd5 415 mutex_unlock(&rcu_gp_lock);
5bffdd5d
MD
416
417 /*
418 * Wakeup waiters only after we have completed the grace period
419 * and have ensured the memory barriers at the end of the grace
420 * period have been issued.
421 */
422 urcu_wake_all_waiters(&waiters);
2bc59bd7
PM
423}
424
121a5d44
MD
425/*
426 * library wrappers to be used by non-LGPL compatible source code.
427 */
428
429void rcu_read_lock(void)
430{
431 _rcu_read_lock();
432}
433
434void rcu_read_unlock(void)
435{
436 _rcu_read_unlock();
437}
438
882f3357
MD
439int rcu_read_ongoing(void)
440{
441 return _rcu_read_ongoing();
442}
443
121a5d44 444void rcu_register_thread(void)
27b012e2 445{
bd252a04
MD
446 URCU_TLS(rcu_reader).tid = pthread_self();
447 assert(URCU_TLS(rcu_reader).need_mb == 0);
448 assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
02be5561 449
6abb4bd5 450 mutex_lock(&rcu_gp_lock);
02be5561 451 rcu_init(); /* In case gcc does not support constructor attribute */
bd252a04 452 cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
6abb4bd5 453 mutex_unlock(&rcu_gp_lock);
27b012e2
MD
454}
455
121a5d44 456void rcu_unregister_thread(void)
27b012e2 457{
6abb4bd5 458 mutex_lock(&rcu_gp_lock);
bd252a04 459 cds_list_del(&URCU_TLS(rcu_reader).node);
6abb4bd5 460 mutex_unlock(&rcu_gp_lock);
27b012e2
MD
461}
462
fdf01eed
MD
463#ifdef RCU_MEMBARRIER
464void rcu_init(void)
465{
466 if (init_done)
467 return;
468 init_done = 1;
cf5271ee 469 if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
1de4df4b 470 rcu_has_sys_membarrier = 1;
fdf01eed
MD
471}
472#endif
473
474#ifdef RCU_SIGNAL
02be5561 475static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
27b012e2 476{
40e140c9 477 /*
5481ddb3
DG
478 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
479 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
40e140c9
MD
480 * executed on.
481 */
5481ddb3 482 cmm_smp_mb();
bd252a04 483 _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
5481ddb3 484 cmm_smp_mb();
27b012e2
MD
485}
486
8a5fb4c9 487/*
02be5561 488 * rcu_init constructor. Called when the library is linked, but also when
8a5fb4c9
MD
489 * reader threads are calling rcu_register_thread().
490 * Should only be called by a single thread at a given time. This is ensured by
6abb4bd5
MD
491 * holing the rcu_gp_lock from rcu_register_thread() or by running at library
492 * load time, which should not be executed by multiple threads nor concurrently
493 * with rcu_register_thread() anyway.
8a5fb4c9 494 */
02be5561 495void rcu_init(void)
27b012e2
MD
496{
497 struct sigaction act;
498 int ret;
499
8a5fb4c9
MD
500 if (init_done)
501 return;
502 init_done = 1;
503
02be5561 504 act.sa_sigaction = sigrcu_handler;
dd052bd3 505 act.sa_flags = SA_SIGINFO | SA_RESTART;
c297c21c 506 sigemptyset(&act.sa_mask);
02be5561 507 ret = sigaction(SIGRCU, &act, NULL);
4a6d7378
MD
508 if (ret)
509 urcu_die(errno);
27b012e2
MD
510}
511
02be5561 512void rcu_exit(void)
27b012e2 513{
71210954
MD
514 /*
515 * Don't unregister the SIGRCU signal handler anymore, because
516 * call_rcu threads could still be using it shortly before the
517 * application exits.
518 * Assertion disabled because call_rcu threads are now rcu
519 * readers, and left running at exit.
520 * assert(cds_list_empty(&registry));
521 */
27b012e2 522}
5e77fc1f 523
fdf01eed 524#endif /* #ifdef RCU_SIGNAL */
5e77fc1f 525
5e6b23a6 526DEFINE_RCU_FLAVOR(rcu_flavor);
541d828d 527
5e77fc1f 528#include "urcu-call-rcu-impl.h"
0376e7b2 529#include "urcu-defer-impl.h"
This page took 0.095931 seconds and 4 git commands to generate.