urcu (signal/mb): use linked list instead of array for registry.
[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
MD
67
68/* Thread IDs of registered readers */
69#define INIT_NUM_THREADS 4
70
cf380c2f 71#ifdef DEBUG_YIELD
9d335088
MD
72unsigned int yield_active;
73unsigned int __thread rand_yield;
cf380c2f
MD
74#endif
75
e3b0cef0 76static LIST_HEAD(registry);
27b012e2 77
81d1f1f5 78static void internal_urcu_lock(void)
41718ff9
MD
79{
80 int ret;
09a9f986
PM
81
82#ifndef DISTRUST_SIGNALS_EXTREME
41718ff9
MD
83 ret = pthread_mutex_lock(&urcu_mutex);
84 if (ret) {
85 perror("Error in pthread mutex lock");
86 exit(-1);
87 }
09a9f986
PM
88#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
89 while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
90 if (ret != EBUSY && ret != EINTR) {
91 printf("ret = %d, errno = %d\n", ret, errno);
92 perror("Error in pthread mutex lock");
93 exit(-1);
94 }
e3b0cef0 95 if (urcu_reader.need_mb) {
09a9f986 96 smp_mb();
e3b0cef0 97 urcu_reader.need_mb = 0;
09a9f986
PM
98 smp_mb();
99 }
100 poll(NULL,0,10);
101 }
102#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
41718ff9
MD
103}
104
81d1f1f5 105static void internal_urcu_unlock(void)
41718ff9
MD
106{
107 int ret;
108
109 ret = pthread_mutex_unlock(&urcu_mutex);
110 if (ret) {
111 perror("Error in pthread mutex unlock");
112 exit(-1);
113 }
114}
115
27b012e2
MD
116/*
117 * called with urcu_mutex held.
118 */
1430ee0b 119static void switch_next_urcu_qparity(void)
27b012e2 120{
b0d5e790 121 STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
27b012e2
MD
122}
123
0a1d290b 124#ifdef URCU_MB
e3b0cef0 125static void force_mb_single_thread(struct urcu_reader *index)
40e140c9
MD
126{
127 smp_mb();
128}
129
bb488185
MD
130static void force_mb_all_threads(void)
131{
b715b99e 132 smp_mb();
bb488185 133}
0a1d290b 134#else /* #ifdef URCU_MB */
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
MD
156}
157
27b012e2
MD
158static void force_mb_all_threads(void)
159{
e3b0cef0
MD
160 struct urcu_reader *index;
161
27b012e2 162 /*
b715b99e 163 * Ask for each threads to execute a smp_mb() so we can consider the
27b012e2
MD
164 * compiler barriers around rcu read lock as real memory barriers.
165 */
e3b0cef0 166 if (list_empty(&registry))
27b012e2 167 return;
3a86deba
MD
168 /*
169 * pthread_kill has a smp_mb(). But beware, we assume it performs
157dca95
MD
170 * a cache flush on architectures with non-coherent cache. Let's play
171 * safe and don't assume anything : we use smp_mc() to make sure the
172 * cache flush is enforced.
3a86deba 173 */
e3b0cef0
MD
174 list_for_each_entry(index, &registry, head) {
175 index->need_mb = 1;
09a9f986 176 smp_mc(); /* write need_mb before sending the signal */
f69f195a 177 pthread_kill(index->tid, SIGURCU);
09a9f986 178 }
27b012e2
MD
179 /*
180 * Wait for sighandler (and thus mb()) to execute on every thread.
09a9f986
PM
181 *
182 * Note that the pthread_kill() will never be executed on systems
183 * that correctly deliver signals in a timely manner. However, it
184 * is not uncommon for kernels to have bugs that can result in
185 * lost or unduly delayed signals.
186 *
187 * If you are seeing the below pthread_kill() executing much at
188 * all, we suggest testing the underlying kernel and filing the
189 * relevant bug report. For Linux kernels, we recommend getting
190 * the Linux Test Project (LTP).
27b012e2 191 */
e3b0cef0
MD
192 list_for_each_entry(index, &registry, head) {
193 while (index->need_mb) {
09a9f986
PM
194 pthread_kill(index->tid, SIGURCU);
195 poll(NULL, 0, 1);
196 }
197 }
198 smp_mb(); /* read ->need_mb before ending the barrier */
27b012e2 199}
0a1d290b 200#endif /* #else #ifdef URCU_MB */
27b012e2 201
bc6c15bb
MD
202/*
203 * synchronize_rcu() waiting. Single thread.
204 */
e3b0cef0 205static void wait_gp(struct urcu_reader *index)
bc6c15bb 206{
ec4e58a3 207 uatomic_dec(&gp_futex);
bc6c15bb 208 force_mb_single_thread(index); /* Write futex before read reader_gp */
e3b0cef0 209 if (!rcu_old_gp_ongoing(&index->ctr)) {
bc6c15bb
MD
210 /* Read reader_gp before write futex */
211 force_mb_single_thread(index);
212 /* Callbacks are queued, don't wait. */
ec4e58a3 213 uatomic_set(&gp_futex, 0);
bc6c15bb
MD
214 } else {
215 /* Read reader_gp before read futex */
216 force_mb_single_thread(index);
ec4e58a3 217 if (uatomic_read(&gp_futex) == -1)
bc6c15bb
MD
218 futex(&gp_futex, FUTEX_WAIT, -1,
219 NULL, NULL, 0);
220 }
221}
222
1430ee0b 223void wait_for_quiescent_state(void)
27b012e2 224{
e3b0cef0 225 struct urcu_reader *index;
27b012e2 226
e3b0cef0 227 if (list_empty(&registry))
27b012e2 228 return;
40e140c9 229 /*
e3b0cef0 230 * Wait for each thread urcu_reader.ctr count to become 0.
27b012e2 231 */
e3b0cef0 232 list_for_each_entry(index, &registry, head) {
bc6c15bb 233 int wait_loops = 0;
e8043c1b 234#ifndef HAS_INCOHERENT_CACHES
e3b0cef0 235 while (rcu_old_gp_ongoing(&index->ctr)) {
bc6c15bb
MD
236 if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) {
237 wait_gp(index);
238 } else {
239 cpu_relax();
240 }
241 }
e8043c1b 242#else /* #ifndef HAS_INCOHERENT_CACHES */
27b012e2 243 /*
40e140c9 244 * BUSY-LOOP. Force the reader thread to commit its
e3b0cef0 245 * urcu_reader.ctr update to memory if we wait for too long.
27b012e2 246 */
e3b0cef0 247 while (rcu_old_gp_ongoing(&index->ctr)) {
bc6c15bb
MD
248 switch (wait_loops++) {
249 case RCU_QS_ACTIVE_ATTEMPTS:
250 wait_gp(index);
251 break;
252 case KICK_READER_LOOPS:
09a9f986 253 force_mb_single_thread(index);
40e140c9 254 wait_loops = 0;
bc6c15bb
MD
255 break;
256 default:
3b55dbf4 257 cpu_relax();
40e140c9
MD
258 }
259 }
e8043c1b 260#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
27b012e2 261 }
27b012e2
MD
262}
263
9598a481 264void synchronize_rcu(void)
2bc59bd7 265{
135530fd
MD
266 internal_urcu_lock();
267
9598a481 268 /* All threads should read qparity before accessing data structure
135530fd
MD
269 * where new ptr points to. Must be done within internal_urcu_lock
270 * because it iterates on reader threads.*/
9598a481 271 /* Write new ptr before changing the qparity */
2bc59bd7 272 force_mb_all_threads();
9598a481 273
9598a481 274 switch_next_urcu_qparity(); /* 0 -> 1 */
2bc59bd7
PM
275
276 /*
9598a481
MD
277 * Must commit qparity update to memory before waiting for parity
278 * 0 quiescent state. Failure to do so could result in the writer
279 * waiting forever while new readers are always accessing data (no
280 * progress).
b0d5e790 281 * Ensured by STORE_SHARED and LOAD_SHARED.
2bc59bd7 282 */
2bc59bd7 283
67c2d80b 284 /*
5dba80f9
MD
285 * Adding a smp_mb() which is _not_ formally required, but makes the
286 * model easier to understand. It does not have a big performance impact
287 * anyway, given this is the write-side.
67c2d80b 288 */
5dba80f9 289 smp_mb();
67c2d80b 290
9598a481
MD
291 /*
292 * Wait for previous parity to be empty of readers.
293 */
294 wait_for_quiescent_state(); /* Wait readers in parity 0 */
9598a481
MD
295
296 /*
297 * Must finish waiting for quiescent state for parity 0 before
298 * committing qparity update to memory. Failure to do so could result in
299 * the writer waiting forever while new readers are always accessing
300 * data (no progress).
b0d5e790 301 * Ensured by STORE_SHARED and LOAD_SHARED.
9598a481 302 */
9598a481 303
5dba80f9
MD
304 /*
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.
308 */
309 smp_mb();
67c2d80b 310
9598a481 311 switch_next_urcu_qparity(); /* 1 -> 0 */
9598a481
MD
312
313 /*
314 * Must commit qparity update to memory before waiting for parity
315 * 1 quiescent state. Failure to do so could result in the writer
316 * waiting forever while new readers are always accessing data (no
317 * progress).
b0d5e790 318 * Ensured by STORE_SHARED and LOAD_SHARED.
9598a481 319 */
9598a481 320
5dba80f9
MD
321 /*
322 * Adding a smp_mb() which is _not_ formally required, but makes the
323 * model easier to understand. It does not have a big performance impact
324 * anyway, given this is the write-side.
325 */
326 smp_mb();
67c2d80b 327
9598a481
MD
328 /*
329 * Wait for previous parity to be empty of readers.
330 */
331 wait_for_quiescent_state(); /* Wait readers in parity 1 */
9598a481 332
9598a481 333 /* Finish waiting for reader threads before letting the old ptr being
135530fd
MD
334 * freed. Must be done within internal_urcu_lock because it iterates on
335 * reader threads. */
9598a481 336 force_mb_all_threads();
135530fd
MD
337
338 internal_urcu_unlock();
2bc59bd7
PM
339}
340
121a5d44
MD
341/*
342 * library wrappers to be used by non-LGPL compatible source code.
343 */
344
345void rcu_read_lock(void)
346{
347 _rcu_read_lock();
348}
349
350void rcu_read_unlock(void)
351{
352 _rcu_read_unlock();
353}
354
355void *rcu_dereference(void *p)
356{
357 return _rcu_dereference(p);
358}
359
360void *rcu_assign_pointer_sym(void **p, void *v)
361{
362 wmb();
363 return STORE_SHARED(p, v);
364}
365
366void *rcu_xchg_pointer_sym(void **p, void *v)
367{
368 wmb();
ec4e58a3 369 return uatomic_xchg(p, v);
121a5d44
MD
370}
371
4d1ce26f
MD
372void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
373{
374 wmb();
ec4e58a3 375 return uatomic_cmpxchg(p, old, _new);
4d1ce26f
MD
376}
377
121a5d44
MD
378void *rcu_publish_content_sym(void **p, void *v)
379{
380 void *oldptr;
381
382 oldptr = _rcu_xchg_pointer(p, v);
383 synchronize_rcu();
384 return oldptr;
385}
386
121a5d44 387void rcu_register_thread(void)
27b012e2 388{
c265818b 389 internal_urcu_lock();
8a5fb4c9 390 urcu_init(); /* In case gcc does not support constructor attribute */
e3b0cef0
MD
391 urcu_reader.tid = pthread_self();
392 assert(urcu_reader.need_mb == 0);
393 assert(urcu_reader.ctr == 0);
394 list_add(&urcu_reader.head, &registry);
c265818b 395 internal_urcu_unlock();
27b012e2
MD
396}
397
121a5d44 398void rcu_unregister_thread(void)
27b012e2 399{
c265818b 400 internal_urcu_lock();
e3b0cef0 401 list_del(&urcu_reader.head);
c265818b 402 internal_urcu_unlock();
27b012e2
MD
403}
404
0a1d290b 405#ifndef URCU_MB
121a5d44 406static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
27b012e2 407{
40e140c9
MD
408 /*
409 * Executing this smp_mb() is the only purpose of this signal handler.
410 * It punctually promotes barrier() into smp_mb() on every thread it is
411 * executed on.
412 */
b715b99e 413 smp_mb();
e3b0cef0 414 urcu_reader.need_mb = 0;
09a9f986 415 smp_mb();
27b012e2
MD
416}
417
8a5fb4c9
MD
418/*
419 * urcu_init constructor. Called when the library is linked, but also when
420 * reader threads are calling rcu_register_thread().
421 * Should only be called by a single thread at a given time. This is ensured by
422 * holing the internal_urcu_lock() from rcu_register_thread() or by running at
423 * library load time, which should not be executed by multiple threads nor
424 * concurrently with rcu_register_thread() anyway.
425 */
426void urcu_init(void)
27b012e2
MD
427{
428 struct sigaction act;
429 int ret;
430
8a5fb4c9
MD
431 if (init_done)
432 return;
433 init_done = 1;
434
27b012e2 435 act.sa_sigaction = sigurcu_handler;
dd052bd3 436 act.sa_flags = SA_SIGINFO | SA_RESTART;
c297c21c 437 sigemptyset(&act.sa_mask);
27b012e2 438 ret = sigaction(SIGURCU, &act, NULL);
f69f195a
MD
439 if (ret) {
440 perror("Error in sigaction");
27b012e2
MD
441 exit(-1);
442 }
443}
444
8a5fb4c9 445void urcu_exit(void)
27b012e2
MD
446{
447 struct sigaction act;
448 int ret;
449
450 ret = sigaction(SIGURCU, NULL, &act);
f69f195a
MD
451 if (ret) {
452 perror("Error in sigaction");
27b012e2
MD
453 exit(-1);
454 }
455 assert(act.sa_sigaction == sigurcu_handler);
e3b0cef0 456 assert(list_empty(&registry));
27b012e2 457}
0a1d290b 458#endif /* #ifndef URCU_MB */
This page took 0.047184 seconds and 4 git commands to generate.