update test timings for writer
[urcu.git] / urcu.c
1 /*
2 * urcu.c
3 *
4 * Userspace RCU library
5 *
6 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
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 #include <stdio.h>
27 #include <pthread.h>
28 #include <signal.h>
29 #include <assert.h>
30 #include <stdlib.h>
31 #include <string.h>
32 #include <errno.h>
33 #include <poll.h>
34
35 #include "urcu-static.h"
36 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
37 #include "urcu.h"
38
39 void __attribute__((constructor)) urcu_init(void);
40 void __attribute__((destructor)) urcu_exit(void);
41
42 int init_done;
43
44 pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
45
46 /*
47 * Global grace period counter.
48 * Contains the current RCU_GP_CTR_BIT.
49 * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
50 * Written to only by writer with mutex taken. Read by both writer and readers.
51 */
52 long urcu_gp_ctr = RCU_GP_COUNT;
53
54 /*
55 * Written to only by each individual reader. Read by both the reader and the
56 * writers.
57 */
58 long __thread urcu_active_readers;
59
60 /* Thread IDs of registered readers */
61 #define INIT_NUM_THREADS 4
62
63 struct reader_registry {
64 pthread_t tid;
65 long *urcu_active_readers;
66 char *need_mb;
67 };
68
69 #ifdef DEBUG_YIELD
70 unsigned int yield_active;
71 unsigned int __thread rand_yield;
72 #endif
73
74 static struct reader_registry *registry;
75 static char __thread need_mb;
76 static int num_readers, alloc_readers;
77
78 void internal_urcu_lock(void)
79 {
80 int ret;
81
82 #ifndef DISTRUST_SIGNALS_EXTREME
83 ret = pthread_mutex_lock(&urcu_mutex);
84 if (ret) {
85 perror("Error in pthread mutex lock");
86 exit(-1);
87 }
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 }
95 if (need_mb) {
96 smp_mb();
97 need_mb = 0;
98 smp_mb();
99 }
100 poll(NULL,0,10);
101 }
102 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
103 }
104
105 void internal_urcu_unlock(void)
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
116 /*
117 * called with urcu_mutex held.
118 */
119 static void switch_next_urcu_qparity(void)
120 {
121 STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
122 }
123
124 #ifdef DEBUG_FULL_MB
125 #ifdef HAS_INCOHERENT_CACHES
126 static void force_mb_single_thread(struct reader_registry *index)
127 {
128 smp_mb();
129 }
130 #endif /* #ifdef HAS_INCOHERENT_CACHES */
131
132 static void force_mb_all_threads(void)
133 {
134 smp_mb();
135 }
136 #else /* #ifdef DEBUG_FULL_MB */
137 #ifdef HAS_INCOHERENT_CACHES
138 static void force_mb_single_thread(struct reader_registry *index)
139 {
140 assert(registry);
141 /*
142 * pthread_kill has a smp_mb(). But beware, we assume it performs
143 * a cache flush on architectures with non-coherent cache. Let's play
144 * safe and don't assume anything : we use smp_mc() to make sure the
145 * cache flush is enforced.
146 */
147 *index->need_mb = 1;
148 smp_mc(); /* write ->need_mb before sending the signals */
149 pthread_kill(index->tid, SIGURCU);
150 smp_mb();
151 /*
152 * Wait for sighandler (and thus mb()) to execute on every thread.
153 * BUSY-LOOP.
154 */
155 while (*index->need_mb) {
156 poll(NULL, 0, 1);
157 }
158 smp_mb(); /* read ->need_mb before ending the barrier */
159 }
160 #endif /* #ifdef HAS_INCOHERENT_CACHES */
161
162 static void force_mb_all_threads(void)
163 {
164 struct reader_registry *index;
165 /*
166 * Ask for each threads to execute a smp_mb() so we can consider the
167 * compiler barriers around rcu read lock as real memory barriers.
168 */
169 if (!registry)
170 return;
171 /*
172 * pthread_kill has a smp_mb(). But beware, we assume it performs
173 * a cache flush on architectures with non-coherent cache. Let's play
174 * safe and don't assume anything : we use smp_mc() to make sure the
175 * cache flush is enforced.
176 */
177 for (index = registry; index < registry + num_readers; index++) {
178 *index->need_mb = 1;
179 smp_mc(); /* write need_mb before sending the signal */
180 pthread_kill(index->tid, SIGURCU);
181 }
182 /*
183 * Wait for sighandler (and thus mb()) to execute on every thread.
184 *
185 * Note that the pthread_kill() will never be executed on systems
186 * that correctly deliver signals in a timely manner. However, it
187 * is not uncommon for kernels to have bugs that can result in
188 * lost or unduly delayed signals.
189 *
190 * If you are seeing the below pthread_kill() executing much at
191 * all, we suggest testing the underlying kernel and filing the
192 * relevant bug report. For Linux kernels, we recommend getting
193 * the Linux Test Project (LTP).
194 */
195 for (index = registry; index < registry + num_readers; index++) {
196 while (*index->need_mb) {
197 pthread_kill(index->tid, SIGURCU);
198 poll(NULL, 0, 1);
199 }
200 }
201 smp_mb(); /* read ->need_mb before ending the barrier */
202 }
203 #endif /* #else #ifdef DEBUG_FULL_MB */
204
205 void wait_for_quiescent_state(void)
206 {
207 struct reader_registry *index;
208
209 if (!registry)
210 return;
211 /*
212 * Wait for each thread urcu_active_readers count to become 0.
213 */
214 for (index = registry; index < registry + num_readers; index++) {
215 #ifndef HAS_INCOHERENT_CACHES
216 while (rcu_old_gp_ongoing(index->urcu_active_readers))
217 cpu_relax();
218 #else /* #ifndef HAS_INCOHERENT_CACHES */
219 int wait_loops = 0;
220 /*
221 * BUSY-LOOP. Force the reader thread to commit its
222 * urcu_active_readers update to memory if we wait for too long.
223 */
224 while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
225 if (wait_loops++ == KICK_READER_LOOPS) {
226 force_mb_single_thread(index);
227 wait_loops = 0;
228 } else {
229 cpu_relax();
230 }
231 }
232 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
233 }
234 }
235
236 void synchronize_rcu(void)
237 {
238 internal_urcu_lock();
239
240 /* All threads should read qparity before accessing data structure
241 * where new ptr points to. Must be done within internal_urcu_lock
242 * because it iterates on reader threads.*/
243 /* Write new ptr before changing the qparity */
244 force_mb_all_threads();
245
246 switch_next_urcu_qparity(); /* 0 -> 1 */
247
248 /*
249 * Must commit qparity update to memory before waiting for parity
250 * 0 quiescent state. Failure to do so could result in the writer
251 * waiting forever while new readers are always accessing data (no
252 * progress).
253 * Ensured by STORE_SHARED and LOAD_SHARED.
254 */
255
256 /*
257 * Adding a smp_mb() which is _not_ formally required, but makes the
258 * model easier to understand. It does not have a big performance impact
259 * anyway, given this is the write-side.
260 */
261 smp_mb();
262
263 /*
264 * Wait for previous parity to be empty of readers.
265 */
266 wait_for_quiescent_state(); /* Wait readers in parity 0 */
267
268 /*
269 * Must finish waiting for quiescent state for parity 0 before
270 * committing qparity update to memory. Failure to do so could result in
271 * the writer waiting forever while new readers are always accessing
272 * data (no progress).
273 * Ensured by STORE_SHARED and LOAD_SHARED.
274 */
275
276 /*
277 * Adding a smp_mb() which is _not_ formally required, but makes the
278 * model easier to understand. It does not have a big performance impact
279 * anyway, given this is the write-side.
280 */
281 smp_mb();
282
283 switch_next_urcu_qparity(); /* 1 -> 0 */
284
285 /*
286 * Must commit qparity update to memory before waiting for parity
287 * 1 quiescent state. Failure to do so could result in the writer
288 * waiting forever while new readers are always accessing data (no
289 * progress).
290 * Ensured by STORE_SHARED and LOAD_SHARED.
291 */
292
293 /*
294 * Adding a smp_mb() which is _not_ formally required, but makes the
295 * model easier to understand. It does not have a big performance impact
296 * anyway, given this is the write-side.
297 */
298 smp_mb();
299
300 /*
301 * Wait for previous parity to be empty of readers.
302 */
303 wait_for_quiescent_state(); /* Wait readers in parity 1 */
304
305 /* Finish waiting for reader threads before letting the old ptr being
306 * freed. Must be done within internal_urcu_lock because it iterates on
307 * reader threads. */
308 force_mb_all_threads();
309
310 internal_urcu_unlock();
311 }
312
313 /*
314 * library wrappers to be used by non-LGPL compatible source code.
315 */
316
317 void rcu_read_lock(void)
318 {
319 _rcu_read_lock();
320 }
321
322 void rcu_read_unlock(void)
323 {
324 _rcu_read_unlock();
325 }
326
327 void *rcu_dereference(void *p)
328 {
329 return _rcu_dereference(p);
330 }
331
332 void *rcu_assign_pointer_sym(void **p, void *v)
333 {
334 wmb();
335 return STORE_SHARED(p, v);
336 }
337
338 void *rcu_xchg_pointer_sym(void **p, void *v)
339 {
340 wmb();
341 return xchg(p, v);
342 }
343
344 void *rcu_publish_content_sym(void **p, void *v)
345 {
346 void *oldptr;
347
348 oldptr = _rcu_xchg_pointer(p, v);
349 synchronize_rcu();
350 return oldptr;
351 }
352
353 static void rcu_add_reader(pthread_t id)
354 {
355 struct reader_registry *oldarray;
356
357 if (!registry) {
358 alloc_readers = INIT_NUM_THREADS;
359 num_readers = 0;
360 registry =
361 malloc(sizeof(struct reader_registry) * alloc_readers);
362 }
363 if (alloc_readers < num_readers + 1) {
364 oldarray = registry;
365 registry = malloc(sizeof(struct reader_registry)
366 * (alloc_readers << 1));
367 memcpy(registry, oldarray,
368 sizeof(struct reader_registry) * alloc_readers);
369 alloc_readers <<= 1;
370 free(oldarray);
371 }
372 registry[num_readers].tid = id;
373 /* reference to the TLS of _this_ reader thread. */
374 registry[num_readers].urcu_active_readers = &urcu_active_readers;
375 registry[num_readers].need_mb = &need_mb;
376 num_readers++;
377 }
378
379 /*
380 * Never shrink (implementation limitation).
381 * This is O(nb threads). Eventually use a hash table.
382 */
383 static void rcu_remove_reader(pthread_t id)
384 {
385 struct reader_registry *index;
386
387 assert(registry != NULL);
388 for (index = registry; index < registry + num_readers; index++) {
389 if (pthread_equal(index->tid, id)) {
390 memcpy(index, &registry[num_readers - 1],
391 sizeof(struct reader_registry));
392 registry[num_readers - 1].tid = 0;
393 registry[num_readers - 1].urcu_active_readers = NULL;
394 num_readers--;
395 return;
396 }
397 }
398 /* Hrm not found, forgot to register ? */
399 assert(0);
400 }
401
402 void rcu_register_thread(void)
403 {
404 internal_urcu_lock();
405 urcu_init(); /* In case gcc does not support constructor attribute */
406 rcu_add_reader(pthread_self());
407 internal_urcu_unlock();
408 }
409
410 void rcu_unregister_thread(void)
411 {
412 internal_urcu_lock();
413 rcu_remove_reader(pthread_self());
414 internal_urcu_unlock();
415 }
416
417 #ifndef DEBUG_FULL_MB
418 static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
419 {
420 /*
421 * Executing this smp_mb() is the only purpose of this signal handler.
422 * It punctually promotes barrier() into smp_mb() on every thread it is
423 * executed on.
424 */
425 smp_mb();
426 need_mb = 0;
427 smp_mb();
428 }
429
430 /*
431 * urcu_init constructor. Called when the library is linked, but also when
432 * reader threads are calling rcu_register_thread().
433 * Should only be called by a single thread at a given time. This is ensured by
434 * holing the internal_urcu_lock() from rcu_register_thread() or by running at
435 * library load time, which should not be executed by multiple threads nor
436 * concurrently with rcu_register_thread() anyway.
437 */
438 void urcu_init(void)
439 {
440 struct sigaction act;
441 int ret;
442
443 if (init_done)
444 return;
445 init_done = 1;
446
447 act.sa_sigaction = sigurcu_handler;
448 ret = sigaction(SIGURCU, &act, NULL);
449 if (ret) {
450 perror("Error in sigaction");
451 exit(-1);
452 }
453 }
454
455 void urcu_exit(void)
456 {
457 struct sigaction act;
458 int ret;
459
460 ret = sigaction(SIGURCU, NULL, &act);
461 if (ret) {
462 perror("Error in sigaction");
463 exit(-1);
464 }
465 assert(act.sa_sigaction == sigurcu_handler);
466 free(registry);
467 }
468 #endif /* #ifndef DEBUG_FULL_MB */
This page took 0.038244 seconds and 4 git commands to generate.