93d781fcf44e5f2c9ee89928631a9d0d6cbb3e21
[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 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
113
114 /*
115 * Pointer to registry elements. Written to only by each individual reader. Read
116 * by both the reader and the writers.
117 */
118 DEFINE_URCU_TLS(struct rcu_reader *, rcu_reader);
119
120 static CDS_LIST_HEAD(registry);
121
122 struct registry_arena {
123 void *p;
124 size_t len;
125 size_t used;
126 };
127
128 static struct registry_arena registry_arena;
129
130 /* Saved fork signal mask, protected by rcu_gp_lock */
131 static sigset_t saved_fork_signal_mask;
132
133 static void rcu_gc_registry(void);
134
135 static void mutex_lock(pthread_mutex_t *mutex)
136 {
137 int ret;
138
139 #ifndef DISTRUST_SIGNALS_EXTREME
140 ret = pthread_mutex_lock(mutex);
141 if (ret)
142 urcu_die(ret);
143 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
144 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
145 if (ret != EBUSY && ret != EINTR)
146 urcu_die(ret);
147 poll(NULL,0,10);
148 }
149 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
150 }
151
152 static void mutex_unlock(pthread_mutex_t *mutex)
153 {
154 int ret;
155
156 ret = pthread_mutex_unlock(mutex);
157 if (ret)
158 urcu_die(ret);
159 }
160
161 static void wait_for_readers(struct cds_list_head *input_readers,
162 struct cds_list_head *cur_snap_readers,
163 struct cds_list_head *qsreaders)
164 {
165 int wait_loops = 0;
166 struct rcu_reader *index, *tmp;
167
168 /*
169 * Wait for each thread URCU_TLS(rcu_reader).ctr to either
170 * indicate quiescence (not nested), or observe the current
171 * rcu_gp.ctr value.
172 */
173 for (;;) {
174 wait_loops++;
175 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
176 switch (rcu_reader_state(&index->ctr)) {
177 case RCU_READER_ACTIVE_CURRENT:
178 if (cur_snap_readers) {
179 cds_list_move(&index->node,
180 cur_snap_readers);
181 break;
182 }
183 /* Fall-through */
184 case RCU_READER_INACTIVE:
185 cds_list_move(&index->node, qsreaders);
186 break;
187 case RCU_READER_ACTIVE_OLD:
188 /*
189 * Old snapshot. Leaving node in
190 * input_readers will make us busy-loop
191 * until the snapshot becomes current or
192 * the reader becomes inactive.
193 */
194 break;
195 }
196 }
197
198 if (cds_list_empty(input_readers)) {
199 break;
200 } else {
201 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
202 usleep(RCU_SLEEP_DELAY);
203 else
204 caa_cpu_relax();
205 }
206 }
207 }
208
209 void synchronize_rcu(void)
210 {
211 CDS_LIST_HEAD(cur_snap_readers);
212 CDS_LIST_HEAD(qsreaders);
213 sigset_t newmask, oldmask;
214 int ret;
215
216 ret = sigfillset(&newmask);
217 assert(!ret);
218 ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
219 assert(!ret);
220
221 mutex_lock(&rcu_gp_lock);
222
223 if (cds_list_empty(&registry))
224 goto out;
225
226 /* All threads should read qparity before accessing data structure
227 * where new ptr points to. */
228 /* Write new ptr before changing the qparity */
229 cmm_smp_mb();
230
231 /* Remove old registry elements */
232 rcu_gc_registry();
233
234 /*
235 * Wait for readers to observe original parity or be quiescent.
236 */
237 wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
238
239 /*
240 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
241 * model easier to understand. It does not have a big performance impact
242 * anyway, given this is the write-side.
243 */
244 cmm_smp_mb();
245
246 /* Switch parity: 0 -> 1, 1 -> 0 */
247 CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
248
249 /*
250 * Must commit qparity update to memory before waiting for other parity
251 * quiescent state. Failure to do so could result in the writer waiting
252 * forever while new readers are always accessing data (no progress).
253 * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
254 */
255
256 /*
257 * Adding a cmm_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 cmm_smp_mb();
262
263 /*
264 * Wait for readers to observe new parity or be quiescent.
265 */
266 wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
267
268 /*
269 * Put quiescent reader list back into registry.
270 */
271 cds_list_splice(&qsreaders, &registry);
272
273 /*
274 * Finish waiting for reader threads before letting the old ptr being
275 * freed.
276 */
277 cmm_smp_mb();
278 out:
279 mutex_unlock(&rcu_gp_lock);
280 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
281 assert(!ret);
282 }
283
284 /*
285 * library wrappers to be used by non-LGPL compatible source code.
286 */
287
288 void rcu_read_lock(void)
289 {
290 _rcu_read_lock();
291 }
292
293 void rcu_read_unlock(void)
294 {
295 _rcu_read_unlock();
296 }
297
298 int rcu_read_ongoing(void)
299 {
300 return _rcu_read_ongoing();
301 }
302
303 /*
304 * only grow for now.
305 */
306 static void resize_arena(struct registry_arena *arena, size_t len)
307 {
308 void *new_p;
309 size_t old_len;
310
311 old_len = arena->len;
312
313 if (!arena->p)
314 new_p = mmap(arena->p, len,
315 PROT_READ | PROT_WRITE,
316 MAP_ANONYMOUS | MAP_PRIVATE,
317 -1, 0);
318 else
319 new_p = mremap_wrapper(arena->p, old_len,
320 len, MREMAP_MAYMOVE);
321 assert(new_p != MAP_FAILED);
322
323 /*
324 * Zero the newly allocated memory. Since mmap() does not
325 * clearly specify if memory is zeroed or not (although it is
326 * very likely that it is), be extra careful by not expecting
327 * the new range to be zeroed by mremap.
328 */
329 bzero(new_p + old_len, len - old_len);
330
331 /*
332 * If we did not re-use the same region, we need to update the
333 * arena pointer.
334 */
335 if (new_p != arena->p)
336 arena->p = new_p;
337
338 arena->len = len;
339 }
340
341 /* Called with signals off and mutex locked */
342 static void add_thread(void)
343 {
344 struct rcu_reader *rcu_reader_reg;
345
346 if (registry_arena.len
347 < registry_arena.used + sizeof(struct rcu_reader))
348 resize_arena(&registry_arena,
349 caa_max(registry_arena.len << 1, ARENA_INIT_ALLOC));
350 /*
351 * Find a free spot.
352 */
353 for (rcu_reader_reg = registry_arena.p;
354 (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
355 rcu_reader_reg++) {
356 if (!rcu_reader_reg->alloc)
357 break;
358 }
359 rcu_reader_reg->alloc = 1;
360 registry_arena.used += sizeof(struct rcu_reader);
361
362 /* Add to registry */
363 rcu_reader_reg->tid = pthread_self();
364 assert(rcu_reader_reg->ctr == 0);
365 cds_list_add(&rcu_reader_reg->node, &registry);
366 URCU_TLS(rcu_reader) = rcu_reader_reg;
367 }
368
369 /* Called with signals off and mutex locked */
370 static void rcu_gc_registry(void)
371 {
372 struct rcu_reader *rcu_reader_reg;
373 pthread_t tid;
374 int ret;
375
376 for (rcu_reader_reg = registry_arena.p;
377 (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
378 rcu_reader_reg++) {
379 if (!rcu_reader_reg->alloc)
380 continue;
381 tid = rcu_reader_reg->tid;
382 ret = pthread_kill(tid, 0);
383 assert(ret != EINVAL);
384 if (ret == ESRCH) {
385 cds_list_del(&rcu_reader_reg->node);
386 rcu_reader_reg->ctr = 0;
387 rcu_reader_reg->alloc = 0;
388 registry_arena.used -= sizeof(struct rcu_reader);
389 }
390 }
391 }
392
393 /* Disable signals, take mutex, add to registry */
394 void rcu_bp_register(void)
395 {
396 sigset_t newmask, oldmask;
397 int ret;
398
399 ret = sigfillset(&newmask);
400 assert(!ret);
401 ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
402 assert(!ret);
403
404 /*
405 * Check if a signal concurrently registered our thread since
406 * the check in rcu_read_lock(). */
407 if (URCU_TLS(rcu_reader))
408 goto end;
409
410 mutex_lock(&rcu_gp_lock);
411 add_thread();
412 mutex_unlock(&rcu_gp_lock);
413 end:
414 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
415 assert(!ret);
416 }
417
418 void rcu_bp_exit(void)
419 {
420 if (registry_arena.p)
421 munmap(registry_arena.p, registry_arena.len);
422 }
423
424 /*
425 * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
426 * a concurrent thread executing with this same lock held. This ensures that the
427 * registry is in a coherent state in the child.
428 */
429 void rcu_bp_before_fork(void)
430 {
431 sigset_t newmask, oldmask;
432 int ret;
433
434 ret = sigfillset(&newmask);
435 assert(!ret);
436 ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
437 assert(!ret);
438 mutex_lock(&rcu_gp_lock);
439 saved_fork_signal_mask = oldmask;
440 }
441
442 void rcu_bp_after_fork_parent(void)
443 {
444 sigset_t oldmask;
445 int ret;
446
447 oldmask = saved_fork_signal_mask;
448 mutex_unlock(&rcu_gp_lock);
449 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
450 assert(!ret);
451 }
452
453 void rcu_bp_after_fork_child(void)
454 {
455 sigset_t oldmask;
456 int ret;
457
458 rcu_gc_registry();
459 oldmask = saved_fork_signal_mask;
460 mutex_unlock(&rcu_gp_lock);
461 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
462 assert(!ret);
463 }
464
465 void *rcu_dereference_sym_bp(void *p)
466 {
467 return _rcu_dereference(p);
468 }
469
470 void *rcu_set_pointer_sym_bp(void **p, void *v)
471 {
472 cmm_wmb();
473 uatomic_set(p, v);
474 return v;
475 }
476
477 void *rcu_xchg_pointer_sym_bp(void **p, void *v)
478 {
479 cmm_wmb();
480 return uatomic_xchg(p, v);
481 }
482
483 void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new)
484 {
485 cmm_wmb();
486 return uatomic_cmpxchg(p, old, _new);
487 }
488
489 DEFINE_RCU_FLAVOR(rcu_flavor);
490
491 #include "urcu-call-rcu-impl.h"
492 #include "urcu-defer-impl.h"
This page took 0.037707 seconds and 3 git commands to generate.