version 0.3.4
[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
27b012e2
MD
26#include <stdio.h>
27#include <pthread.h>
28#include <signal.h>
29#include <assert.h>
f69f195a
MD
30#include <stdlib.h>
31#include <string.h>
09a9f986 32#include <errno.h>
e8043c1b 33#include <poll.h>
27b012e2 34
121a5d44
MD
35#include "urcu-static.h"
36/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
27b012e2
MD
37#include "urcu.h"
38
0a1d290b 39#ifndef URCU_MB
834a45ba
MD
40static int init_done;
41
8a5fb4c9
MD
42void __attribute__((constructor)) urcu_init(void);
43void __attribute__((destructor)) urcu_exit(void);
e90a6e9c 44#else
538d7df5 45void urcu_init(void)
e90a6e9c
MD
46{
47}
48#endif
8a5fb4c9 49
81d1f1f5 50static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
27b012e2 51
bc6c15bb
MD
52int gp_futex;
53
128166c9
MD
54/*
55 * Global grace period counter.
56 * Contains the current RCU_GP_CTR_BIT.
afb8f2c9 57 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
b0d5e790 58 * Written to only by writer with mutex taken. Read by both writer and readers.
128166c9
MD
59 */
60long urcu_gp_ctr = RCU_GP_COUNT;
27b012e2 61
b0d5e790
MD
62/*
63 * Written to only by each individual reader. Read by both the reader and the
64 * writers.
65 */
e3b0cef0 66struct urcu_reader __thread urcu_reader;
27b012e2 67
cf380c2f 68#ifdef DEBUG_YIELD
9d335088
MD
69unsigned int yield_active;
70unsigned int __thread rand_yield;
cf380c2f
MD
71#endif
72
e3b0cef0 73static LIST_HEAD(registry);
27b012e2 74
81d1f1f5 75static void internal_urcu_lock(void)
41718ff9
MD
76{
77 int ret;
09a9f986
PM
78
79#ifndef DISTRUST_SIGNALS_EXTREME
41718ff9
MD
80 ret = pthread_mutex_lock(&urcu_mutex);
81 if (ret) {
82 perror("Error in pthread mutex lock");
83 exit(-1);
84 }
09a9f986
PM
85#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
86 while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
87 if (ret != EBUSY && ret != EINTR) {
88 printf("ret = %d, errno = %d\n", ret, errno);
89 perror("Error in pthread mutex lock");
90 exit(-1);
91 }
e3b0cef0 92 if (urcu_reader.need_mb) {
09a9f986 93 smp_mb();
e3b0cef0 94 urcu_reader.need_mb = 0;
09a9f986
PM
95 smp_mb();
96 }
97 poll(NULL,0,10);
98 }
99#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
41718ff9
MD
100}
101
81d1f1f5 102static void internal_urcu_unlock(void)
41718ff9
MD
103{
104 int ret;
105
106 ret = pthread_mutex_unlock(&urcu_mutex);
107 if (ret) {
108 perror("Error in pthread mutex unlock");
109 exit(-1);
110 }
111}
112
27b012e2
MD
113/*
114 * called with urcu_mutex held.
115 */
1430ee0b 116static void switch_next_urcu_qparity(void)
27b012e2 117{
b0d5e790 118 STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
27b012e2
MD
119}
120
0a1d290b 121#ifdef URCU_MB
cfe78e25 122#if 0 /* unused */
e3b0cef0 123static void force_mb_single_thread(struct urcu_reader *index)
40e140c9
MD
124{
125 smp_mb();
126}
cfe78e25 127#endif //0
40e140c9 128
bb488185
MD
129static void force_mb_all_threads(void)
130{
b715b99e 131 smp_mb();
bb488185 132}
0a1d290b 133#else /* #ifdef URCU_MB */
cfe78e25 134#if 0 /* unused */
e3b0cef0 135static void force_mb_single_thread(struct urcu_reader *index)
40e140c9 136{
e3b0cef0 137 assert(!list_empty(&registry));
157dca95
MD
138 /*
139 * pthread_kill has a smp_mb(). But beware, we assume it performs
140 * a cache flush on architectures with non-coherent cache. Let's play
141 * safe and don't assume anything : we use smp_mc() to make sure the
142 * cache flush is enforced.
157dca95 143 */
e3b0cef0 144 index->need_mb = 1;
09a9f986
PM
145 smp_mc(); /* write ->need_mb before sending the signals */
146 pthread_kill(index->tid, SIGURCU);
147 smp_mb();
40e140c9
MD
148 /*
149 * Wait for sighandler (and thus mb()) to execute on every thread.
150 * BUSY-LOOP.
151 */
e3b0cef0 152 while (index->need_mb) {
09a9f986
PM
153 poll(NULL, 0, 1);
154 }
155 smp_mb(); /* read ->need_mb before ending the barrier */
40e140c9 156}
cfe78e25 157#endif //0
40e140c9 158
27b012e2
MD
159static void force_mb_all_threads(void)
160{
e3b0cef0
MD
161 struct urcu_reader *index;
162
27b012e2 163 /*
b715b99e 164 * Ask for each threads to execute a smp_mb() so we can consider the
27b012e2
MD
165 * compiler barriers around rcu read lock as real memory barriers.
166 */
e3b0cef0 167 if (list_empty(&registry))
27b012e2 168 return;
3a86deba
MD
169 /*
170 * pthread_kill has a smp_mb(). But beware, we assume it performs
157dca95
MD
171 * a cache flush on architectures with non-coherent cache. Let's play
172 * safe and don't assume anything : we use smp_mc() to make sure the
173 * cache flush is enforced.
3a86deba 174 */
e3b0cef0
MD
175 list_for_each_entry(index, &registry, head) {
176 index->need_mb = 1;
09a9f986 177 smp_mc(); /* write need_mb before sending the signal */
f69f195a 178 pthread_kill(index->tid, SIGURCU);
09a9f986 179 }
27b012e2
MD
180 /*
181 * Wait for sighandler (and thus mb()) to execute on every thread.
09a9f986
PM
182 *
183 * Note that the pthread_kill() will never be executed on systems
184 * that correctly deliver signals in a timely manner. However, it
185 * is not uncommon for kernels to have bugs that can result in
186 * lost or unduly delayed signals.
187 *
188 * If you are seeing the below pthread_kill() executing much at
189 * all, we suggest testing the underlying kernel and filing the
190 * relevant bug report. For Linux kernels, we recommend getting
191 * the Linux Test Project (LTP).
27b012e2 192 */
e3b0cef0
MD
193 list_for_each_entry(index, &registry, head) {
194 while (index->need_mb) {
09a9f986
PM
195 pthread_kill(index->tid, SIGURCU);
196 poll(NULL, 0, 1);
197 }
198 }
199 smp_mb(); /* read ->need_mb before ending the barrier */
27b012e2 200}
0a1d290b 201#endif /* #else #ifdef URCU_MB */
27b012e2 202
bc6c15bb
MD
203/*
204 * synchronize_rcu() waiting. Single thread.
205 */
cfe78e25 206static void wait_gp(void)
bc6c15bb 207{
cfe78e25
MD
208 /* Read reader_gp before read futex */
209 force_mb_all_threads();
210 if (uatomic_read(&gp_futex) == -1)
0854ccff 211 futex_async(&gp_futex, FUTEX_WAIT, -1,
cfe78e25 212 NULL, NULL, 0);
bc6c15bb
MD
213}
214
1430ee0b 215void wait_for_quiescent_state(void)
27b012e2 216{
cfe78e25
MD
217 LIST_HEAD(qsreaders);
218 int wait_loops = 0;
23758cc9 219 struct urcu_reader *index, *tmp;
27b012e2 220
e3b0cef0 221 if (list_empty(&registry))
27b012e2 222 return;
40e140c9 223 /*
e3b0cef0 224 * Wait for each thread urcu_reader.ctr count to become 0.
27b012e2 225 */
cfe78e25
MD
226 for (;;) {
227 wait_loops++;
228 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
229 uatomic_dec(&gp_futex);
230 /* Write futex before read reader_gp */
231 force_mb_all_threads();
232 }
233
23758cc9 234 list_for_each_entry_safe(index, tmp, &registry, head) {
cfe78e25
MD
235 if (!rcu_old_gp_ongoing(&index->ctr))
236 list_move(&index->head, &qsreaders);
237 }
238
e8043c1b 239#ifndef HAS_INCOHERENT_CACHES
cfe78e25
MD
240 if (list_empty(&registry)) {
241 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
242 /* Read reader_gp before write futex */
243 force_mb_all_threads();
244 uatomic_set(&gp_futex, 0);
bc6c15bb 245 }
cfe78e25
MD
246 break;
247 } else {
248 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
249 wait_gp();
250 else
251 cpu_relax();
bc6c15bb 252 }
e8043c1b 253#else /* #ifndef HAS_INCOHERENT_CACHES */
27b012e2 254 /*
40e140c9 255 * BUSY-LOOP. Force the reader thread to commit its
e3b0cef0 256 * urcu_reader.ctr update to memory if we wait for too long.
27b012e2 257 */
cfe78e25
MD
258 if (list_empty(&registry)) {
259 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
260 /* Read reader_gp before write futex */
261 force_mb_all_threads();
262 uatomic_set(&gp_futex, 0);
263 }
264 break;
265 } else {
266 switch (wait_loops) {
bc6c15bb 267 case RCU_QS_ACTIVE_ATTEMPTS:
cfe78e25
MD
268 wait_gp();
269 break; /* only escape switch */
bc6c15bb 270 case KICK_READER_LOOPS:
cfe78e25 271 force_mb_all_threads();
40e140c9 272 wait_loops = 0;
cfe78e25 273 break; /* only escape switch */
bc6c15bb 274 default:
3b55dbf4 275 cpu_relax();
40e140c9
MD
276 }
277 }
e8043c1b 278#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
27b012e2 279 }
cfe78e25 280 /* put back the reader list in the registry */
23758cc9 281 list_splice(&qsreaders, &registry);
27b012e2
MD
282}
283
9598a481 284void synchronize_rcu(void)
2bc59bd7 285{
135530fd
MD
286 internal_urcu_lock();
287
9598a481 288 /* All threads should read qparity before accessing data structure
135530fd
MD
289 * where new ptr points to. Must be done within internal_urcu_lock
290 * because it iterates on reader threads.*/
9598a481 291 /* Write new ptr before changing the qparity */
2bc59bd7 292 force_mb_all_threads();
9598a481 293
9598a481 294 switch_next_urcu_qparity(); /* 0 -> 1 */
2bc59bd7
PM
295
296 /*
9598a481
MD
297 * Must commit qparity update to memory before waiting for parity
298 * 0 quiescent state. Failure to do so could result in the writer
299 * waiting forever while new readers are always accessing data (no
300 * progress).
b0d5e790 301 * Ensured by STORE_SHARED and LOAD_SHARED.
2bc59bd7 302 */
2bc59bd7 303
67c2d80b 304 /*
5dba80f9
MD
305 * Adding a smp_mb() which is _not_ formally required, but makes the
306 * model easier to understand. It does not have a big performance impact
307 * anyway, given this is the write-side.
67c2d80b 308 */
5dba80f9 309 smp_mb();
67c2d80b 310
9598a481
MD
311 /*
312 * Wait for previous parity to be empty of readers.
313 */
314 wait_for_quiescent_state(); /* Wait readers in parity 0 */
9598a481
MD
315
316 /*
317 * Must finish waiting for quiescent state for parity 0 before
318 * committing qparity update to memory. Failure to do so could result in
319 * the writer waiting forever while new readers are always accessing
320 * data (no progress).
b0d5e790 321 * Ensured by STORE_SHARED and LOAD_SHARED.
9598a481 322 */
9598a481 323
5dba80f9
MD
324 /*
325 * Adding a smp_mb() which is _not_ formally required, but makes the
326 * model easier to understand. It does not have a big performance impact
327 * anyway, given this is the write-side.
328 */
329 smp_mb();
67c2d80b 330
9598a481 331 switch_next_urcu_qparity(); /* 1 -> 0 */
9598a481
MD
332
333 /*
334 * Must commit qparity update to memory before waiting for parity
335 * 1 quiescent state. Failure to do so could result in the writer
336 * waiting forever while new readers are always accessing data (no
337 * progress).
b0d5e790 338 * Ensured by STORE_SHARED and LOAD_SHARED.
9598a481 339 */
9598a481 340
5dba80f9
MD
341 /*
342 * Adding a smp_mb() which is _not_ formally required, but makes the
343 * model easier to understand. It does not have a big performance impact
344 * anyway, given this is the write-side.
345 */
346 smp_mb();
67c2d80b 347
9598a481
MD
348 /*
349 * Wait for previous parity to be empty of readers.
350 */
351 wait_for_quiescent_state(); /* Wait readers in parity 1 */
9598a481 352
9598a481 353 /* Finish waiting for reader threads before letting the old ptr being
135530fd
MD
354 * freed. Must be done within internal_urcu_lock because it iterates on
355 * reader threads. */
9598a481 356 force_mb_all_threads();
135530fd
MD
357
358 internal_urcu_unlock();
2bc59bd7
PM
359}
360
121a5d44
MD
361/*
362 * library wrappers to be used by non-LGPL compatible source code.
363 */
364
365void rcu_read_lock(void)
366{
367 _rcu_read_lock();
368}
369
370void rcu_read_unlock(void)
371{
372 _rcu_read_unlock();
373}
374
121a5d44 375void rcu_register_thread(void)
27b012e2 376{
e3b0cef0
MD
377 urcu_reader.tid = pthread_self();
378 assert(urcu_reader.need_mb == 0);
379 assert(urcu_reader.ctr == 0);
812cadd7
MD
380
381 internal_urcu_lock();
382 urcu_init(); /* In case gcc does not support constructor attribute */
e3b0cef0 383 list_add(&urcu_reader.head, &registry);
c265818b 384 internal_urcu_unlock();
27b012e2
MD
385}
386
121a5d44 387void rcu_unregister_thread(void)
27b012e2 388{
c265818b 389 internal_urcu_lock();
e3b0cef0 390 list_del(&urcu_reader.head);
c265818b 391 internal_urcu_unlock();
27b012e2
MD
392}
393
0a1d290b 394#ifndef URCU_MB
121a5d44 395static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
27b012e2 396{
40e140c9
MD
397 /*
398 * Executing this smp_mb() is the only purpose of this signal handler.
399 * It punctually promotes barrier() into smp_mb() on every thread it is
400 * executed on.
401 */
b715b99e 402 smp_mb();
e3b0cef0 403 urcu_reader.need_mb = 0;
09a9f986 404 smp_mb();
27b012e2
MD
405}
406
8a5fb4c9
MD
407/*
408 * urcu_init constructor. Called when the library is linked, but also when
409 * reader threads are calling rcu_register_thread().
410 * Should only be called by a single thread at a given time. This is ensured by
411 * holing the internal_urcu_lock() from rcu_register_thread() or by running at
412 * library load time, which should not be executed by multiple threads nor
413 * concurrently with rcu_register_thread() anyway.
414 */
415void urcu_init(void)
27b012e2
MD
416{
417 struct sigaction act;
418 int ret;
419
8a5fb4c9
MD
420 if (init_done)
421 return;
422 init_done = 1;
423
27b012e2 424 act.sa_sigaction = sigurcu_handler;
dd052bd3 425 act.sa_flags = SA_SIGINFO | SA_RESTART;
c297c21c 426 sigemptyset(&act.sa_mask);
27b012e2 427 ret = sigaction(SIGURCU, &act, NULL);
f69f195a
MD
428 if (ret) {
429 perror("Error in sigaction");
27b012e2
MD
430 exit(-1);
431 }
432}
433
8a5fb4c9 434void urcu_exit(void)
27b012e2
MD
435{
436 struct sigaction act;
437 int ret;
438
439 ret = sigaction(SIGURCU, NULL, &act);
f69f195a
MD
440 if (ret) {
441 perror("Error in sigaction");
27b012e2
MD
442 exit(-1);
443 }
444 assert(act.sa_sigaction == sigurcu_handler);
e3b0cef0 445 assert(list_empty(&registry));
27b012e2 446}
0a1d290b 447#endif /* #ifndef URCU_MB */
This page took 0.080995 seconds and 4 git commands to generate.