Support sys_membarrier v6 ABI
[urcu.git] / urcu.c
CommitLineData
b257a10b
MD
1/*
2 * urcu.c
3 *
4 * Userspace RCU library
5 *
af02d47e
MD
6 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
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
27b012e2
MD
27#include <stdio.h>
28#include <pthread.h>
29#include <signal.h>
30#include <assert.h>
f69f195a
MD
31#include <stdlib.h>
32#include <string.h>
09a9f986 33#include <errno.h>
e8043c1b 34#include <poll.h>
27b012e2 35
121a5d44
MD
36#include "urcu-static.h"
37/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
27b012e2
MD
38#include "urcu.h"
39
fdf01eed 40#ifdef RCU_MEMBARRIER
834a45ba 41static int init_done;
fdf01eed 42int has_sys_membarrier;
834a45ba 43
02be5561 44void __attribute__((constructor)) rcu_init(void);
fdf01eed
MD
45#endif
46
47#ifdef RCU_MB
02be5561 48void rcu_init(void)
e90a6e9c
MD
49{
50}
51#endif
8a5fb4c9 52
fdf01eed
MD
53#ifdef RCU_SIGNAL
54static int init_done;
55
56void __attribute__((constructor)) rcu_init(void);
57void __attribute__((destructor)) rcu_exit(void);
58#endif
59
02be5561 60static pthread_mutex_t rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
27b012e2 61
bc6c15bb
MD
62int gp_futex;
63
128166c9
MD
64/*
65 * Global grace period counter.
02be5561 66 * Contains the current RCU_GP_CTR_PHASE.
afb8f2c9 67 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
b0d5e790 68 * Written to only by writer with mutex taken. Read by both writer and readers.
128166c9 69 */
02be5561 70long rcu_gp_ctr = RCU_GP_COUNT;
27b012e2 71
b0d5e790
MD
72/*
73 * Written to only by each individual reader. Read by both the reader and the
74 * writers.
75 */
02be5561 76struct rcu_reader __thread rcu_reader;
27b012e2 77
cf380c2f 78#ifdef DEBUG_YIELD
9d335088
MD
79unsigned int yield_active;
80unsigned int __thread rand_yield;
cf380c2f
MD
81#endif
82
e3b0cef0 83static LIST_HEAD(registry);
27b012e2 84
02be5561 85static void internal_rcu_lock(void)
41718ff9
MD
86{
87 int ret;
09a9f986
PM
88
89#ifndef DISTRUST_SIGNALS_EXTREME
02be5561 90 ret = pthread_mutex_lock(&rcu_mutex);
41718ff9
MD
91 if (ret) {
92 perror("Error in pthread mutex lock");
93 exit(-1);
94 }
09a9f986 95#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
02be5561 96 while ((ret = pthread_mutex_trylock(&rcu_mutex)) != 0) {
09a9f986
PM
97 if (ret != EBUSY && ret != EINTR) {
98 printf("ret = %d, errno = %d\n", ret, errno);
99 perror("Error in pthread mutex lock");
100 exit(-1);
101 }
02be5561 102 if (rcu_reader.need_mb) {
09a9f986 103 smp_mb();
02be5561 104 rcu_reader.need_mb = 0;
09a9f986
PM
105 smp_mb();
106 }
107 poll(NULL,0,10);
108 }
109#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
41718ff9
MD
110}
111
02be5561 112static void internal_rcu_unlock(void)
41718ff9
MD
113{
114 int ret;
115
02be5561 116 ret = pthread_mutex_unlock(&rcu_mutex);
41718ff9
MD
117 if (ret) {
118 perror("Error in pthread mutex unlock");
119 exit(-1);
120 }
121}
122
27b012e2 123/*
02be5561 124 * called with rcu_mutex held.
27b012e2 125 */
02be5561 126static void switch_next_rcu_qparity(void)
27b012e2 127{
02be5561 128 STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
27b012e2
MD
129}
130
fdf01eed
MD
131#ifdef RCU_MEMBARRIER
132static void smp_mb_heavy(void)
133{
134 if (likely(has_sys_membarrier))
f0708810 135 membarrier(MEMBARRIER_EXPEDITED);
fdf01eed
MD
136 else
137 smp_mb();
138}
139#endif
140
02be5561 141#ifdef RCU_MB
fdf01eed 142static void smp_mb_heavy(void)
40e140c9
MD
143{
144 smp_mb();
145}
fdf01eed
MD
146#endif
147
148#ifdef RCU_SIGNAL
78ff9419 149static void force_mb_all_readers(void)
27b012e2 150{
02be5561 151 struct rcu_reader *index;
e3b0cef0 152
27b012e2 153 /*
b715b99e 154 * Ask for each threads to execute a smp_mb() so we can consider the
27b012e2
MD
155 * compiler barriers around rcu read lock as real memory barriers.
156 */
e3b0cef0 157 if (list_empty(&registry))
27b012e2 158 return;
3a86deba
MD
159 /*
160 * pthread_kill has a smp_mb(). But beware, we assume it performs
157dca95
MD
161 * a cache flush on architectures with non-coherent cache. Let's play
162 * safe and don't assume anything : we use smp_mc() to make sure the
163 * cache flush is enforced.
3a86deba 164 */
e3b0cef0
MD
165 list_for_each_entry(index, &registry, head) {
166 index->need_mb = 1;
09a9f986 167 smp_mc(); /* write need_mb before sending the signal */
02be5561 168 pthread_kill(index->tid, SIGRCU);
09a9f986 169 }
27b012e2
MD
170 /*
171 * Wait for sighandler (and thus mb()) to execute on every thread.
09a9f986
PM
172 *
173 * Note that the pthread_kill() will never be executed on systems
174 * that correctly deliver signals in a timely manner. However, it
175 * is not uncommon for kernels to have bugs that can result in
176 * lost or unduly delayed signals.
177 *
178 * If you are seeing the below pthread_kill() executing much at
179 * all, we suggest testing the underlying kernel and filing the
180 * relevant bug report. For Linux kernels, we recommend getting
181 * the Linux Test Project (LTP).
27b012e2 182 */
e3b0cef0
MD
183 list_for_each_entry(index, &registry, head) {
184 while (index->need_mb) {
02be5561 185 pthread_kill(index->tid, SIGRCU);
09a9f986
PM
186 poll(NULL, 0, 1);
187 }
188 }
189 smp_mb(); /* read ->need_mb before ending the barrier */
27b012e2 190}
9d7e3f89 191
fdf01eed 192static void smp_mb_heavy(void)
9d7e3f89
MD
193{
194 force_mb_all_readers();
195}
fdf01eed 196#endif /* #ifdef RCU_SIGNAL */
27b012e2 197
bc6c15bb
MD
198/*
199 * synchronize_rcu() waiting. Single thread.
200 */
cfe78e25 201static void wait_gp(void)
bc6c15bb 202{
cfe78e25 203 /* Read reader_gp before read futex */
9d7e3f89 204 smp_mb_heavy();
cfe78e25 205 if (uatomic_read(&gp_futex) == -1)
0854ccff 206 futex_async(&gp_futex, FUTEX_WAIT, -1,
cfe78e25 207 NULL, NULL, 0);
bc6c15bb
MD
208}
209
1430ee0b 210void wait_for_quiescent_state(void)
27b012e2 211{
cfe78e25
MD
212 LIST_HEAD(qsreaders);
213 int wait_loops = 0;
02be5561 214 struct rcu_reader *index, *tmp;
27b012e2 215
e3b0cef0 216 if (list_empty(&registry))
27b012e2 217 return;
40e140c9 218 /*
02be5561 219 * Wait for each thread rcu_reader.ctr count to become 0.
27b012e2 220 */
cfe78e25
MD
221 for (;;) {
222 wait_loops++;
223 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
224 uatomic_dec(&gp_futex);
225 /* Write futex before read reader_gp */
9d7e3f89 226 smp_mb_heavy();
cfe78e25
MD
227 }
228
23758cc9 229 list_for_each_entry_safe(index, tmp, &registry, head) {
cfe78e25
MD
230 if (!rcu_old_gp_ongoing(&index->ctr))
231 list_move(&index->head, &qsreaders);
232 }
233
e8043c1b 234#ifndef HAS_INCOHERENT_CACHES
cfe78e25
MD
235 if (list_empty(&registry)) {
236 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
237 /* Read reader_gp before write futex */
9d7e3f89 238 smp_mb_heavy();
cfe78e25 239 uatomic_set(&gp_futex, 0);
bc6c15bb 240 }
cfe78e25
MD
241 break;
242 } else {
243 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
244 wait_gp();
245 else
246 cpu_relax();
bc6c15bb 247 }
e8043c1b 248#else /* #ifndef HAS_INCOHERENT_CACHES */
27b012e2 249 /*
40e140c9 250 * BUSY-LOOP. Force the reader thread to commit its
02be5561 251 * rcu_reader.ctr update to memory if we wait for too long.
27b012e2 252 */
cfe78e25
MD
253 if (list_empty(&registry)) {
254 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
255 /* Read reader_gp before write futex */
9d7e3f89 256 smp_mb_heavy();
cfe78e25
MD
257 uatomic_set(&gp_futex, 0);
258 }
259 break;
260 } else {
261 switch (wait_loops) {
bc6c15bb 262 case RCU_QS_ACTIVE_ATTEMPTS:
cfe78e25
MD
263 wait_gp();
264 break; /* only escape switch */
bc6c15bb 265 case KICK_READER_LOOPS:
9d7e3f89 266 smp_mb_heavy();
40e140c9 267 wait_loops = 0;
cfe78e25 268 break; /* only escape switch */
bc6c15bb 269 default:
3b55dbf4 270 cpu_relax();
40e140c9
MD
271 }
272 }
e8043c1b 273#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
27b012e2 274 }
cfe78e25 275 /* put back the reader list in the registry */
23758cc9 276 list_splice(&qsreaders, &registry);
27b012e2
MD
277}
278
9598a481 279void synchronize_rcu(void)
2bc59bd7 280{
02be5561 281 internal_rcu_lock();
135530fd 282
9598a481 283 /* All threads should read qparity before accessing data structure
02be5561 284 * where new ptr points to. Must be done within internal_rcu_lock
135530fd 285 * because it iterates on reader threads.*/
9598a481 286 /* Write new ptr before changing the qparity */
9d7e3f89 287 smp_mb_heavy();
9598a481 288
02be5561 289 switch_next_rcu_qparity(); /* 0 -> 1 */
2bc59bd7
PM
290
291 /*
9598a481
MD
292 * Must commit qparity update to memory before waiting for parity
293 * 0 quiescent state. Failure to do so could result in the writer
294 * waiting forever while new readers are always accessing data (no
295 * progress).
b0d5e790 296 * Ensured by STORE_SHARED and LOAD_SHARED.
2bc59bd7 297 */
2bc59bd7 298
67c2d80b 299 /*
5dba80f9
MD
300 * Adding a smp_mb() which is _not_ formally required, but makes the
301 * model easier to understand. It does not have a big performance impact
302 * anyway, given this is the write-side.
67c2d80b 303 */
5dba80f9 304 smp_mb();
67c2d80b 305
9598a481
MD
306 /*
307 * Wait for previous parity to be empty of readers.
308 */
309 wait_for_quiescent_state(); /* Wait readers in parity 0 */
9598a481
MD
310
311 /*
312 * Must finish waiting for quiescent state for parity 0 before
313 * committing qparity update to memory. Failure to do so could result in
314 * the writer waiting forever while new readers are always accessing
315 * data (no progress).
b0d5e790 316 * Ensured by STORE_SHARED and LOAD_SHARED.
9598a481 317 */
9598a481 318
5dba80f9
MD
319 /*
320 * Adding a smp_mb() which is _not_ formally required, but makes the
321 * model easier to understand. It does not have a big performance impact
322 * anyway, given this is the write-side.
323 */
324 smp_mb();
67c2d80b 325
02be5561 326 switch_next_rcu_qparity(); /* 1 -> 0 */
9598a481
MD
327
328 /*
329 * Must commit qparity update to memory before waiting for parity
330 * 1 quiescent state. Failure to do so could result in the writer
331 * waiting forever while new readers are always accessing data (no
332 * progress).
b0d5e790 333 * Ensured by STORE_SHARED and LOAD_SHARED.
9598a481 334 */
9598a481 335
5dba80f9
MD
336 /*
337 * Adding a smp_mb() which is _not_ formally required, but makes the
338 * model easier to understand. It does not have a big performance impact
339 * anyway, given this is the write-side.
340 */
341 smp_mb();
67c2d80b 342
9598a481
MD
343 /*
344 * Wait for previous parity to be empty of readers.
345 */
346 wait_for_quiescent_state(); /* Wait readers in parity 1 */
9598a481 347
9598a481 348 /* Finish waiting for reader threads before letting the old ptr being
02be5561 349 * freed. Must be done within internal_rcu_lock because it iterates on
135530fd 350 * reader threads. */
9d7e3f89 351 smp_mb_heavy();
135530fd 352
02be5561 353 internal_rcu_unlock();
2bc59bd7
PM
354}
355
121a5d44
MD
356/*
357 * library wrappers to be used by non-LGPL compatible source code.
358 */
359
360void rcu_read_lock(void)
361{
362 _rcu_read_lock();
363}
364
365void rcu_read_unlock(void)
366{
367 _rcu_read_unlock();
368}
369
121a5d44 370void rcu_register_thread(void)
27b012e2 371{
02be5561
MD
372 rcu_reader.tid = pthread_self();
373 assert(rcu_reader.need_mb == 0);
374 assert(rcu_reader.ctr == 0);
375
376 internal_rcu_lock();
377 rcu_init(); /* In case gcc does not support constructor attribute */
378 list_add(&rcu_reader.head, &registry);
379 internal_rcu_unlock();
27b012e2
MD
380}
381
121a5d44 382void rcu_unregister_thread(void)
27b012e2 383{
02be5561
MD
384 internal_rcu_lock();
385 list_del(&rcu_reader.head);
386 internal_rcu_unlock();
27b012e2
MD
387}
388
fdf01eed
MD
389#ifdef RCU_MEMBARRIER
390void rcu_init(void)
391{
392 if (init_done)
393 return;
394 init_done = 1;
f0708810 395 if (!membarrier(MEMBARRIER_EXPEDITED))
fdf01eed
MD
396 has_sys_membarrier = 1;
397}
398#endif
399
400#ifdef RCU_SIGNAL
02be5561 401static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
27b012e2 402{
40e140c9
MD
403 /*
404 * Executing this smp_mb() is the only purpose of this signal handler.
405 * It punctually promotes barrier() into smp_mb() on every thread it is
406 * executed on.
407 */
b715b99e 408 smp_mb();
02be5561 409 rcu_reader.need_mb = 0;
09a9f986 410 smp_mb();
27b012e2
MD
411}
412
8a5fb4c9 413/*
02be5561 414 * rcu_init constructor. Called when the library is linked, but also when
8a5fb4c9
MD
415 * reader threads are calling rcu_register_thread().
416 * Should only be called by a single thread at a given time. This is ensured by
02be5561 417 * holing the internal_rcu_lock() from rcu_register_thread() or by running at
8a5fb4c9
MD
418 * library load time, which should not be executed by multiple threads nor
419 * concurrently with rcu_register_thread() anyway.
420 */
02be5561 421void rcu_init(void)
27b012e2
MD
422{
423 struct sigaction act;
424 int ret;
425
8a5fb4c9
MD
426 if (init_done)
427 return;
428 init_done = 1;
429
02be5561 430 act.sa_sigaction = sigrcu_handler;
dd052bd3 431 act.sa_flags = SA_SIGINFO | SA_RESTART;
c297c21c 432 sigemptyset(&act.sa_mask);
02be5561 433 ret = sigaction(SIGRCU, &act, NULL);
f69f195a
MD
434 if (ret) {
435 perror("Error in sigaction");
27b012e2
MD
436 exit(-1);
437 }
438}
439
02be5561 440void rcu_exit(void)
27b012e2
MD
441{
442 struct sigaction act;
443 int ret;
444
02be5561 445 ret = sigaction(SIGRCU, NULL, &act);
f69f195a
MD
446 if (ret) {
447 perror("Error in sigaction");
27b012e2
MD
448 exit(-1);
449 }
02be5561 450 assert(act.sa_sigaction == sigrcu_handler);
e3b0cef0 451 assert(list_empty(&registry));
27b012e2 452}
fdf01eed 453#endif /* #ifdef RCU_SIGNAL */
This page took 0.07601 seconds and 4 git commands to generate.