Version 0.7.11
[userspace-rcu.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/wfqueue.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 not allowing MAYMOVE.
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 assert(!(flags & MREMAP_MAYMOVE));
77
78 return MAP_FAILED;
79 }
80 #endif
81
82 /* Sleep delay in ms */
83 #define RCU_SLEEP_DELAY_MS 10
84 #define INIT_NR_THREADS 8
85 #define ARENA_INIT_ALLOC \
86 sizeof(struct registry_chunk) \
87 + INIT_NR_THREADS * sizeof(struct rcu_reader)
88
89 /*
90 * Active attempts to check for reader Q.S. before calling sleep().
91 */
92 #define RCU_QS_ACTIVE_ATTEMPTS 100
93
94 static
95 int rcu_bp_refcount;
96
97 static
98 void __attribute__((constructor)) rcu_bp_init(void);
99 static
100 void __attribute__((destructor)) rcu_bp_exit(void);
101
102 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
103
104 static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
105 static int initialized;
106
107 static pthread_key_t urcu_bp_key;
108
109 #ifdef DEBUG_YIELD
110 unsigned int yield_active;
111 __DEFINE_URCU_TLS_GLOBAL(unsigned int, rand_yield);
112 #endif
113
114 /*
115 * Global grace period counter.
116 * Contains the current RCU_GP_CTR_PHASE.
117 * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
118 * Written to only by writer with mutex taken. Read by both writer and readers.
119 */
120 long rcu_gp_ctr = RCU_GP_COUNT;
121
122 /*
123 * Pointer to registry elements. Written to only by each individual reader. Read
124 * by both the reader and the writers.
125 */
126 __DEFINE_URCU_TLS_GLOBAL(struct rcu_reader *, rcu_reader);
127
128 static CDS_LIST_HEAD(registry);
129
130 struct registry_chunk {
131 size_t data_len; /* data length */
132 size_t used; /* amount of data used */
133 struct cds_list_head node; /* chunk_list node */
134 char data[];
135 };
136
137 struct registry_arena {
138 struct cds_list_head chunk_list;
139 };
140
141 static struct registry_arena registry_arena = {
142 .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list),
143 };
144
145 /* Saved fork signal mask, protected by rcu_gp_lock */
146 static sigset_t saved_fork_signal_mask;
147
148 static void mutex_lock(pthread_mutex_t *mutex)
149 {
150 int ret;
151
152 #ifndef DISTRUST_SIGNALS_EXTREME
153 ret = pthread_mutex_lock(mutex);
154 if (ret)
155 urcu_die(ret);
156 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
157 while ((ret = pthread_mutex_trylock(mutex)) != 0) {
158 if (ret != EBUSY && ret != EINTR)
159 urcu_die(ret);
160 poll(NULL,0,10);
161 }
162 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
163 }
164
165 static void mutex_unlock(pthread_mutex_t *mutex)
166 {
167 int ret;
168
169 ret = pthread_mutex_unlock(mutex);
170 if (ret)
171 urcu_die(ret);
172 }
173
174 void update_counter_and_wait(void)
175 {
176 CDS_LIST_HEAD(qsreaders);
177 unsigned int wait_loops = 0;
178 struct rcu_reader *index, *tmp;
179
180 /* Switch parity: 0 -> 1, 1 -> 0 */
181 CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
182
183 /*
184 * Must commit qparity update to memory before waiting for other parity
185 * quiescent state. Failure to do so could result in the writer waiting
186 * forever while new readers are always accessing data (no progress).
187 * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
188 */
189
190 /*
191 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
192 * model easier to understand. It does not have a big performance impact
193 * anyway, given this is the write-side.
194 */
195 cmm_smp_mb();
196
197 /*
198 * Wait for each thread rcu_reader.ctr count to become 0.
199 */
200 for (;;) {
201 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
202 wait_loops++;
203
204 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
205 if (!rcu_old_gp_ongoing(&index->ctr))
206 cds_list_move(&index->node, &qsreaders);
207 }
208
209 if (cds_list_empty(&registry)) {
210 break;
211 } else {
212 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
213 (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS);
214 else
215 caa_cpu_relax();
216 }
217 }
218 /* put back the reader list in the registry */
219 cds_list_splice(&qsreaders, &registry);
220 }
221
222 void synchronize_rcu(void)
223 {
224 sigset_t newmask, oldmask;
225 int ret;
226
227 ret = sigfillset(&newmask);
228 assert(!ret);
229 ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
230 assert(!ret);
231
232 mutex_lock(&rcu_gp_lock);
233
234 if (cds_list_empty(&registry))
235 goto out;
236
237 /* All threads should read qparity before accessing data structure
238 * where new ptr points to. */
239 /* Write new ptr before changing the qparity */
240 cmm_smp_mb();
241
242 /*
243 * Wait for previous parity to be empty of readers.
244 */
245 update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
246
247 /*
248 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
249 * model easier to understand. It does not have a big performance impact
250 * anyway, given this is the write-side.
251 */
252 cmm_smp_mb();
253
254 /*
255 * Wait for previous parity to be empty of readers.
256 */
257 update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
258
259 /*
260 * Finish waiting for reader threads before letting the old ptr being
261 * freed.
262 */
263 cmm_smp_mb();
264 out:
265 mutex_unlock(&rcu_gp_lock);
266 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
267 assert(!ret);
268 }
269
270 /*
271 * library wrappers to be used by non-LGPL compatible source code.
272 */
273
274 void rcu_read_lock(void)
275 {
276 _rcu_read_lock();
277 }
278
279 void rcu_read_unlock(void)
280 {
281 _rcu_read_unlock();
282 }
283
284 /*
285 * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
286 * Else, try expanding the last chunk. If this fails, allocate a new
287 * chunk twice as big as the last chunk.
288 * Memory used by chunks _never_ moves. A chunk could theoretically be
289 * freed when all "used" slots are released, but we don't do it at this
290 * point.
291 */
292 static
293 void expand_arena(struct registry_arena *arena)
294 {
295 struct registry_chunk *new_chunk, *last_chunk;
296 size_t old_chunk_len, new_chunk_len;
297
298 /* No chunk. */
299 if (cds_list_empty(&arena->chunk_list)) {
300 assert(ARENA_INIT_ALLOC >=
301 sizeof(struct registry_chunk)
302 + sizeof(struct rcu_reader));
303 new_chunk_len = ARENA_INIT_ALLOC;
304 new_chunk = mmap(NULL, new_chunk_len,
305 PROT_READ | PROT_WRITE,
306 MAP_ANONYMOUS | MAP_PRIVATE,
307 -1, 0);
308 if (new_chunk == MAP_FAILED)
309 abort();
310 bzero(new_chunk, new_chunk_len);
311 new_chunk->data_len =
312 new_chunk_len - sizeof(struct registry_chunk);
313 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
314 return; /* We're done. */
315 }
316
317 /* Try expanding last chunk. */
318 last_chunk = cds_list_entry(arena->chunk_list.prev,
319 struct registry_chunk, node);
320 old_chunk_len =
321 last_chunk->data_len + sizeof(struct registry_chunk);
322 new_chunk_len = old_chunk_len << 1;
323
324 /* Don't allow memory mapping to move, just expand. */
325 new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
326 new_chunk_len, 0);
327 if (new_chunk != MAP_FAILED) {
328 /* Should not have moved. */
329 assert(new_chunk == last_chunk);
330 bzero((char *) last_chunk + old_chunk_len,
331 new_chunk_len - old_chunk_len);
332 last_chunk->data_len =
333 new_chunk_len - sizeof(struct registry_chunk);
334 return; /* We're done. */
335 }
336
337 /* Remap did not succeed, we need to add a new chunk. */
338 new_chunk = mmap(NULL, new_chunk_len,
339 PROT_READ | PROT_WRITE,
340 MAP_ANONYMOUS | MAP_PRIVATE,
341 -1, 0);
342 if (new_chunk == MAP_FAILED)
343 abort();
344 bzero(new_chunk, new_chunk_len);
345 new_chunk->data_len =
346 new_chunk_len - sizeof(struct registry_chunk);
347 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
348 }
349
350 static
351 struct rcu_reader *arena_alloc(struct registry_arena *arena)
352 {
353 struct registry_chunk *chunk;
354 struct rcu_reader *rcu_reader_reg;
355 int expand_done = 0; /* Only allow to expand once per alloc */
356 size_t len = sizeof(struct rcu_reader);
357
358 retry:
359 cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
360 if (chunk->data_len - chunk->used < len)
361 continue;
362 /* Find spot */
363 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
364 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
365 rcu_reader_reg++) {
366 if (!rcu_reader_reg->alloc) {
367 rcu_reader_reg->alloc = 1;
368 chunk->used += len;
369 return rcu_reader_reg;
370 }
371 }
372 }
373
374 if (!expand_done) {
375 expand_arena(arena);
376 expand_done = 1;
377 goto retry;
378 }
379
380 return NULL;
381 }
382
383 /* Called with signals off and mutex locked */
384 static
385 void add_thread(void)
386 {
387 struct rcu_reader *rcu_reader_reg;
388 int ret;
389
390 rcu_reader_reg = arena_alloc(&registry_arena);
391 if (!rcu_reader_reg)
392 abort();
393 ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
394 if (ret)
395 abort();
396
397 /* Add to registry */
398 rcu_reader_reg->tid = pthread_self();
399 assert(rcu_reader_reg->ctr == 0);
400 cds_list_add(&rcu_reader_reg->node, &registry);
401 /*
402 * Reader threads are pointing to the reader registry. This is
403 * why its memory should never be relocated.
404 */
405 URCU_TLS(rcu_reader) = rcu_reader_reg;
406 }
407
408 /* Called with mutex locked */
409 static
410 void cleanup_thread(struct registry_chunk *chunk,
411 struct rcu_reader *rcu_reader_reg)
412 {
413 rcu_reader_reg->ctr = 0;
414 cds_list_del(&rcu_reader_reg->node);
415 rcu_reader_reg->tid = 0;
416 rcu_reader_reg->alloc = 0;
417 chunk->used -= sizeof(struct rcu_reader);
418 }
419
420 static
421 struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
422 {
423 struct registry_chunk *chunk;
424
425 cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
426 if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
427 continue;
428 if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
429 continue;
430 return chunk;
431 }
432 return NULL;
433 }
434
435 /* Called with signals off and mutex locked */
436 static
437 void remove_thread(struct rcu_reader *rcu_reader_reg)
438 {
439 cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
440 URCU_TLS(rcu_reader) = NULL;
441 }
442
443 /* Disable signals, take mutex, add to registry */
444 void rcu_bp_register(void)
445 {
446 sigset_t newmask, oldmask;
447 int ret;
448
449 ret = sigfillset(&newmask);
450 if (ret)
451 abort();
452 ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
453 if (ret)
454 abort();
455
456 /*
457 * Check if a signal concurrently registered our thread since
458 * the check in rcu_read_lock().
459 */
460 if (URCU_TLS(rcu_reader))
461 goto end;
462
463 /*
464 * Take care of early registration before urcu_bp constructor.
465 */
466 rcu_bp_init();
467
468 mutex_lock(&rcu_gp_lock);
469 add_thread();
470 mutex_unlock(&rcu_gp_lock);
471 end:
472 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
473 if (ret)
474 abort();
475 }
476
477 /* Disable signals, take mutex, remove from registry */
478 static
479 void rcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
480 {
481 sigset_t newmask, oldmask;
482 int ret;
483
484 ret = sigfillset(&newmask);
485 if (ret)
486 abort();
487 ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
488 if (ret)
489 abort();
490
491 mutex_lock(&rcu_gp_lock);
492 remove_thread(rcu_reader_reg);
493 mutex_unlock(&rcu_gp_lock);
494 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
495 if (ret)
496 abort();
497 rcu_bp_exit();
498 }
499
500 /*
501 * Remove thread from the registry when it exits, and flag it as
502 * destroyed so garbage collection can take care of it.
503 */
504 static
505 void urcu_bp_thread_exit_notifier(void *rcu_key)
506 {
507 rcu_bp_unregister(rcu_key);
508 }
509
510 static
511 void rcu_bp_init(void)
512 {
513 mutex_lock(&init_lock);
514 if (!rcu_bp_refcount++) {
515 int ret;
516
517 ret = pthread_key_create(&urcu_bp_key,
518 urcu_bp_thread_exit_notifier);
519 if (ret)
520 abort();
521 initialized = 1;
522 }
523 mutex_unlock(&init_lock);
524 }
525
526 static
527 void rcu_bp_exit(void)
528 {
529 mutex_lock(&init_lock);
530 if (!--rcu_bp_refcount) {
531 struct registry_chunk *chunk, *tmp;
532 int ret;
533
534 cds_list_for_each_entry_safe(chunk, tmp,
535 &registry_arena.chunk_list, node) {
536 munmap(chunk, chunk->data_len
537 + sizeof(struct registry_chunk));
538 }
539 ret = pthread_key_delete(urcu_bp_key);
540 if (ret)
541 abort();
542 }
543 mutex_unlock(&init_lock);
544 }
545
546 /*
547 * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
548 * a concurrent thread executing with this same lock held. This ensures that the
549 * registry is in a coherent state in the child.
550 */
551 void rcu_bp_before_fork(void)
552 {
553 sigset_t newmask, oldmask;
554 int ret;
555
556 ret = sigfillset(&newmask);
557 assert(!ret);
558 ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
559 assert(!ret);
560 mutex_lock(&rcu_gp_lock);
561 saved_fork_signal_mask = oldmask;
562 }
563
564 void rcu_bp_after_fork_parent(void)
565 {
566 sigset_t oldmask;
567 int ret;
568
569 oldmask = saved_fork_signal_mask;
570 mutex_unlock(&rcu_gp_lock);
571 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
572 assert(!ret);
573 }
574
575 /*
576 * Prune all entries from registry except our own thread. Fits the Linux
577 * fork behavior. Called with rcu_gp_lock held.
578 */
579 static
580 void urcu_bp_prune_registry(void)
581 {
582 struct registry_chunk *chunk;
583 struct rcu_reader *rcu_reader_reg;
584
585 cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
586 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
587 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
588 rcu_reader_reg++) {
589 if (!rcu_reader_reg->alloc)
590 continue;
591 if (rcu_reader_reg->tid == pthread_self())
592 continue;
593 cleanup_thread(chunk, rcu_reader_reg);
594 }
595 }
596 }
597
598 void rcu_bp_after_fork_child(void)
599 {
600 sigset_t oldmask;
601 int ret;
602
603 urcu_bp_prune_registry();
604 oldmask = saved_fork_signal_mask;
605 mutex_unlock(&rcu_gp_lock);
606 ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
607 assert(!ret);
608 }
609
610 void *rcu_dereference_sym_bp(void *p)
611 {
612 return _rcu_dereference(p);
613 }
614
615 void *rcu_set_pointer_sym_bp(void **p, void *v)
616 {
617 cmm_wmb();
618 uatomic_set(p, v);
619 return v;
620 }
621
622 void *rcu_xchg_pointer_sym_bp(void **p, void *v)
623 {
624 cmm_wmb();
625 return uatomic_xchg(p, v);
626 }
627
628 void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new)
629 {
630 cmm_wmb();
631 return uatomic_cmpxchg(p, old, _new);
632 }
633
634 DEFINE_RCU_FLAVOR(rcu_flavor);
635
636 #include "urcu-call-rcu-impl.h"
637 #include "urcu-defer-impl.h"
This page took 0.04154 seconds and 4 git commands to generate.