urcu-mb/signal/membarrier: improve 2-phase wait scheme
[urcu.git] / urcu-bp.c
1 /*
2 * urcu-bp.c
3 *
4 * Userspace RCU library, "bulletproof" version.
5 *
6 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
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 #define _GNU_SOURCE
27 #define _LGPL_SOURCE
28 #include <stdio.h>
29 #include <pthread.h>
30 #include <signal.h>
31 #include <assert.h>
32 #include <stdlib.h>
33 #include <string.h>
34 #include <errno.h>
35 #include <poll.h>
36 #include <unistd.h>
37 #include <sys/mman.h>
38
39 #include "urcu/wfcqueue.h"
40 #include "urcu/map/urcu-bp.h"
41 #include "urcu/static/urcu-bp.h"
42 #include "urcu-pointer.h"
43 #include "urcu/tls-compat.h"
44
45 #include "urcu-die.h"
46
47 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
48 #undef _LGPL_SOURCE
49 #include "urcu-bp.h"
50 #define _LGPL_SOURCE
51
52 #ifndef MAP_ANONYMOUS
53 #define MAP_ANONYMOUS MAP_ANON
54 #endif
55
56 #ifdef __linux__
57 static
58 void *mremap_wrapper(void *old_address, size_t old_size,
59 size_t new_size, int flags)
60 {
61 return mremap(old_address, old_size, new_size, flags);
62 }
63 #else
64
65 #define MREMAP_MAYMOVE 1
66 #define MREMAP_FIXED 2
67
68 /*
69 * mremap wrapper for non-Linux systems. Maps a RW, anonymous private mapping.
70 * This is not generic.
71 */
72 static
73 void *mremap_wrapper(void *old_address, size_t old_size,
74 size_t new_size, int flags)
75 {
76 void *new_address;
77
78 assert(flags & MREMAP_MAYMOVE);
79 assert(!(flags & MREMAP_FIXED));
80 new_address = mmap(old_address, new_size,
81 PROT_READ | PROT_WRITE,
82 MAP_ANONYMOUS | MAP_PRIVATE,
83 -1, 0);
84 if (new_address == MAP_FAILED)
85 return MAP_FAILED;
86 if (old_address) {
87 memcpy(new_address, old_address, old_size);
88 munmap(old_address, old_size);
89 }
90 return new_address;
91 }
92 #endif
93
94 /* Sleep delay in us */
95 #define RCU_SLEEP_DELAY 1000
96 #define ARENA_INIT_ALLOC 16
97
98 /*
99 * Active attempts to check for reader Q.S. before calling sleep().
100 */
101 #define RCU_QS_ACTIVE_ATTEMPTS 100
102
103 void __attribute__((destructor)) rcu_bp_exit(void);
104
105 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
106
107 #ifdef DEBUG_YIELD
108 unsigned int rcu_yield_active;
109 DEFINE_URCU_TLS(unsigned int, rcu_rand_yield);
110 #endif
111
112 /*
113 * Global grace period counter.
114 * Contains the current RCU_GP_CTR_PHASE.
115 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
116 * Written to only by writer with mutex taken. Read by both writer and readers.
117 */
118 long rcu_gp_ctr = RCU_GP_COUNT;
119
120 /*
121 * Pointer to registry elements. Written to only by each individual reader. Read
122 * by both the reader and the writers.
123 */
124 DEFINE_URCU_TLS(struct rcu_reader *, rcu_reader);
125
126 static CDS_LIST_HEAD(registry);
127
128 struct registry_arena {
129 void *p;
130 size_t len;
131 size_t used;
132 };
133
134 static struct registry_arena registry_arena;
135
136 /* Saved fork signal mask, protected by rcu_gp_lock */
137 static sigset_t saved_fork_signal_mask;
138
139 static void rcu_gc_registry(void);
140
141 static void mutex_lock(pthread_mutex_t *mutex)
142 {
143 int ret;
144
145 #ifndef DISTRUST_SIGNALS_EXTREME
146 ret = pthread_mutex_lock(mutex);
147 if (ret)
148 urcu_die(ret);
149 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
150 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
151 if (ret != EBUSY && ret != EINTR)
152 urcu_die(ret);
153 poll(NULL,0,10);
154 }
155 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
156 }
157
158 static void mutex_unlock(pthread_mutex_t *mutex)
159 {
160 int ret;
161
162 ret = pthread_mutex_unlock(mutex);
163 if (ret)
164 urcu_die(ret);
165 }
166
167 static void update_counter_and_wait(void)
168 {
169 CDS_LIST_HEAD(qsreaders);
170 int wait_loops = 0;
171 struct rcu_reader *index, *tmp;
172
173 /* Switch parity: 0 -> 1, 1 -> 0 */
174 CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
175
176 /*
177 * Must commit qparity update to memory before waiting for other parity
178 * quiescent state. Failure to do so could result in the writer waiting
179 * forever while new readers are always accessing data (no progress).
180 * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
181 */
182
183 /*
184 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
185 * model easier to understand. It does not have a big performance impact
186 * anyway, given this is the write-side.
187 */
188 cmm_smp_mb();
189
190 /*
191 * Wait for each thread rcu_reader.ctr count to become 0.
192 */
193 for (;;) {
194 wait_loops++;
195 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
196 if (!rcu_old_gp_ongoing(&index->ctr))
197 cds_list_move(&index->node, &qsreaders);
198 }
199
200 if (cds_list_empty(&registry)) {
201 break;
202 } else {
203 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
204 usleep(RCU_SLEEP_DELAY);
205 else
206 caa_cpu_relax();
207 }
208 }
209 /* put back the reader list in the registry */
210 cds_list_splice(&qsreaders, &registry);
211 }
212
213 void synchronize_rcu(void)
214 {
215 sigset_t newmask, oldmask;
216 int ret;
217
218 ret = sigemptyset(&newmask);
219 assert(!ret);
220 ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
221 assert(!ret);
222
223 mutex_lock(&rcu_gp_lock);
224
225 if (cds_list_empty(&registry))
226 goto out;
227
228 /* All threads should read qparity before accessing data structure
229 * where new ptr points to. */
230 /* Write new ptr before changing the qparity */
231 cmm_smp_mb();
232
233 /* Remove old registry elements */
234 rcu_gc_registry();
235
236 /*
237 * Wait for previous parity to be empty of readers.
238 */
239 update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
240
241 /*
242 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
243 * model easier to understand. It does not have a big performance impact
244 * anyway, given this is the write-side.
245 */
246 cmm_smp_mb();
247
248 /*
249 * Wait for previous parity to be empty of readers.
250 */
251 update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
252
253 /*
254 * Finish waiting for reader threads before letting the old ptr being
255 * freed.
256 */
257 cmm_smp_mb();
258 out:
259 mutex_unlock(&rcu_gp_lock);
260 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
261 assert(!ret);
262 }
263
264 /*
265 * library wrappers to be used by non-LGPL compatible source code.
266 */
267
268 void rcu_read_lock(void)
269 {
270 _rcu_read_lock();
271 }
272
273 void rcu_read_unlock(void)
274 {
275 _rcu_read_unlock();
276 }
277
278 /*
279 * only grow for now.
280 */
281 static void resize_arena(struct registry_arena *arena, size_t len)
282 {
283 void *new_arena;
284
285 if (!arena->p)
286 new_arena = mmap(arena->p, len,
287 PROT_READ | PROT_WRITE,
288 MAP_ANONYMOUS | MAP_PRIVATE,
289 -1, 0);
290 else
291 new_arena = mremap_wrapper(arena->p, arena->len,
292 len, MREMAP_MAYMOVE);
293 assert(new_arena != MAP_FAILED);
294
295 /*
296 * re-used the same region ?
297 */
298 if (new_arena == arena->p)
299 return;
300
301 bzero(new_arena + arena->len, len - arena->len);
302 arena->p = new_arena;
303 }
304
305 /* Called with signals off and mutex locked */
306 static void add_thread(void)
307 {
308 struct rcu_reader *rcu_reader_reg;
309
310 if (registry_arena.len
311 < registry_arena.used + sizeof(struct rcu_reader))
312 resize_arena(&registry_arena,
313 caa_max(registry_arena.len << 1, ARENA_INIT_ALLOC));
314 /*
315 * Find a free spot.
316 */
317 for (rcu_reader_reg = registry_arena.p;
318 (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
319 rcu_reader_reg++) {
320 if (!rcu_reader_reg->alloc)
321 break;
322 }
323 rcu_reader_reg->alloc = 1;
324 registry_arena.used += sizeof(struct rcu_reader);
325
326 /* Add to registry */
327 rcu_reader_reg->tid = pthread_self();
328 assert(rcu_reader_reg->ctr == 0);
329 cds_list_add(&rcu_reader_reg->node, &registry);
330 URCU_TLS(rcu_reader) = rcu_reader_reg;
331 }
332
333 /* Called with signals off and mutex locked */
334 static void rcu_gc_registry(void)
335 {
336 struct rcu_reader *rcu_reader_reg;
337 pthread_t tid;
338 int ret;
339
340 for (rcu_reader_reg = registry_arena.p;
341 (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
342 rcu_reader_reg++) {
343 if (!rcu_reader_reg->alloc)
344 continue;
345 tid = rcu_reader_reg->tid;
346 ret = pthread_kill(tid, 0);
347 assert(ret != EINVAL);
348 if (ret == ESRCH) {
349 cds_list_del(&rcu_reader_reg->node);
350 rcu_reader_reg->ctr = 0;
351 rcu_reader_reg->alloc = 0;
352 registry_arena.used -= sizeof(struct rcu_reader);
353 }
354 }
355 }
356
357 /* Disable signals, take mutex, add to registry */
358 void rcu_bp_register(void)
359 {
360 sigset_t newmask, oldmask;
361 int ret;
362
363 ret = sigemptyset(&newmask);
364 assert(!ret);
365 ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
366 assert(!ret);
367
368 /*
369 * Check if a signal concurrently registered our thread since
370 * the check in rcu_read_lock(). */
371 if (URCU_TLS(rcu_reader))
372 goto end;
373
374 mutex_lock(&rcu_gp_lock);
375 add_thread();
376 mutex_unlock(&rcu_gp_lock);
377 end:
378 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
379 assert(!ret);
380 }
381
382 void rcu_bp_exit(void)
383 {
384 if (registry_arena.p)
385 munmap(registry_arena.p, registry_arena.len);
386 }
387
388 /*
389 * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
390 * a concurrent thread executing with this same lock held. This ensures that the
391 * registry is in a coherent state in the child.
392 */
393 void rcu_bp_before_fork(void)
394 {
395 sigset_t newmask, oldmask;
396 int ret;
397
398 ret = sigemptyset(&newmask);
399 assert(!ret);
400 ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
401 assert(!ret);
402 mutex_lock(&rcu_gp_lock);
403 saved_fork_signal_mask = oldmask;
404 }
405
406 void rcu_bp_after_fork_parent(void)
407 {
408 sigset_t oldmask;
409 int ret;
410
411 oldmask = saved_fork_signal_mask;
412 mutex_unlock(&rcu_gp_lock);
413 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
414 assert(!ret);
415 }
416
417 void rcu_bp_after_fork_child(void)
418 {
419 sigset_t oldmask;
420 int ret;
421
422 rcu_gc_registry();
423 oldmask = saved_fork_signal_mask;
424 mutex_unlock(&rcu_gp_lock);
425 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
426 assert(!ret);
427 }
428
429 void *rcu_dereference_sym_bp(void *p)
430 {
431 return _rcu_dereference(p);
432 }
433
434 void *rcu_set_pointer_sym_bp(void **p, void *v)
435 {
436 cmm_wmb();
437 uatomic_set(p, v);
438 return v;
439 }
440
441 void *rcu_xchg_pointer_sym_bp(void **p, void *v)
442 {
443 cmm_wmb();
444 return uatomic_xchg(p, v);
445 }
446
447 void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new)
448 {
449 cmm_wmb();
450 return uatomic_cmpxchg(p, old, _new);
451 }
452
453 DEFINE_RCU_FLAVOR(rcu_flavor);
454
455 #include "urcu-call-rcu-impl.h"
456 #include "urcu-defer-impl.h"
This page took 0.037374 seconds and 5 git commands to generate.