Commit | Line | Data |
---|---|---|
fdee2e6d MD |
1 | /* |
2 | * urcu-bp.c | |
3 | * | |
4 | * Userspace RCU library, "bulletproof" version. | |
5 | * | |
6982d6d7 | 6 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
fdee2e6d MD |
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 | ||
0617bf4c | 26 | #define _GNU_SOURCE |
71c811bf | 27 | #define _LGPL_SOURCE |
fdee2e6d MD |
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 | ||
71c811bf | 39 | #include "urcu/wfqueue.h" |
57760d44 | 40 | #include "urcu/map/urcu-bp.h" |
af7c2dbe | 41 | #include "urcu/static/urcu-bp.h" |
618b2595 | 42 | #include "urcu-pointer.h" |
71c811bf | 43 | |
fdee2e6d | 44 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
71c811bf | 45 | #undef _LGPL_SOURCE |
fdee2e6d | 46 | #include "urcu-bp.h" |
71c811bf | 47 | #define _LGPL_SOURCE |
fdee2e6d | 48 | |
4c1ae2ea MD |
49 | #ifndef MAP_ANONYMOUS |
50 | #define MAP_ANONYMOUS MAP_ANON | |
51 | #endif | |
52 | ||
45a4872f MD |
53 | #ifndef __linux__ |
54 | ||
55 | #define MREMAP_MAYMOVE 1 | |
56 | #define MREMAP_FIXED 2 | |
57 | ||
58 | /* | |
59 | * mremap wrapper for non-Linux systems. Maps a RW, anonymous private mapping. | |
60 | * This is not generic. | |
61 | */ | |
62 | void *mremap(void *old_address, size_t old_size, size_t new_size, int flags) | |
63 | { | |
dc745ef6 | 64 | void *new_address; |
45a4872f MD |
65 | |
66 | assert(flags & MREMAP_MAYMOVE); | |
67 | assert(!(flags & MREMAP_FIXED)); | |
68 | new_address = mmap(old_address, new_size, | |
69 | PROT_READ | PROT_WRITE, | |
70 | MAP_ANONYMOUS | MAP_PRIVATE, | |
71 | -1, 0); | |
72 | if (new_address == MAP_FAILED) | |
73 | return MAP_FAILED; | |
74 | if (old_address) { | |
75 | memcpy(new_address, old_address, old_size); | |
76 | munmap(old_address, old_size); | |
77 | } | |
78 | return new_address; | |
79 | } | |
80 | #endif | |
81 | ||
fdee2e6d MD |
82 | /* Sleep delay in us */ |
83 | #define RCU_SLEEP_DELAY 1000 | |
84 | #define ARENA_INIT_ALLOC 16 | |
85 | ||
b7b6a8f5 PB |
86 | /* |
87 | * Active attempts to check for reader Q.S. before calling sleep(). | |
88 | */ | |
89 | #define RCU_QS_ACTIVE_ATTEMPTS 100 | |
90 | ||
02be5561 | 91 | void __attribute__((destructor)) rcu_bp_exit(void); |
fdee2e6d | 92 | |
6abb4bd5 | 93 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
fdee2e6d MD |
94 | |
95 | #ifdef DEBUG_YIELD | |
96 | unsigned int yield_active; | |
97 | unsigned int __thread rand_yield; | |
98 | #endif | |
99 | ||
100 | /* | |
101 | * Global grace period counter. | |
02be5561 | 102 | * Contains the current RCU_GP_CTR_PHASE. |
fdee2e6d MD |
103 | * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. |
104 | * Written to only by writer with mutex taken. Read by both writer and readers. | |
105 | */ | |
02be5561 | 106 | long rcu_gp_ctr = RCU_GP_COUNT; |
fdee2e6d MD |
107 | |
108 | /* | |
109 | * Pointer to registry elements. Written to only by each individual reader. Read | |
110 | * by both the reader and the writers. | |
111 | */ | |
02be5561 | 112 | struct rcu_reader __thread *rcu_reader; |
fdee2e6d | 113 | |
16aa9ee8 | 114 | static CDS_LIST_HEAD(registry); |
fdee2e6d MD |
115 | |
116 | struct registry_arena { | |
117 | void *p; | |
118 | size_t len; | |
119 | size_t used; | |
120 | }; | |
121 | ||
122 | static struct registry_arena registry_arena; | |
123 | ||
4cf1675f MD |
124 | /* Saved fork signal mask, protected by rcu_gp_lock */ |
125 | static sigset_t saved_fork_signal_mask; | |
126 | ||
fdee2e6d MD |
127 | static void rcu_gc_registry(void); |
128 | ||
6abb4bd5 | 129 | static void mutex_lock(pthread_mutex_t *mutex) |
fdee2e6d MD |
130 | { |
131 | int ret; | |
132 | ||
133 | #ifndef DISTRUST_SIGNALS_EXTREME | |
6abb4bd5 | 134 | ret = pthread_mutex_lock(mutex); |
fdee2e6d MD |
135 | if (ret) { |
136 | perror("Error in pthread mutex lock"); | |
137 | exit(-1); | |
138 | } | |
139 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ | |
6abb4bd5 | 140 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
fdee2e6d MD |
141 | if (ret != EBUSY && ret != EINTR) { |
142 | printf("ret = %d, errno = %d\n", ret, errno); | |
143 | perror("Error in pthread mutex lock"); | |
144 | exit(-1); | |
145 | } | |
fdee2e6d MD |
146 | poll(NULL,0,10); |
147 | } | |
148 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
149 | } | |
150 | ||
6abb4bd5 | 151 | static void mutex_unlock(pthread_mutex_t *mutex) |
fdee2e6d MD |
152 | { |
153 | int ret; | |
154 | ||
6abb4bd5 | 155 | ret = pthread_mutex_unlock(mutex); |
fdee2e6d MD |
156 | if (ret) { |
157 | perror("Error in pthread mutex unlock"); | |
158 | exit(-1); | |
159 | } | |
160 | } | |
161 | ||
2dfb8b5e | 162 | void update_counter_and_wait(void) |
fdee2e6d | 163 | { |
16aa9ee8 | 164 | CDS_LIST_HEAD(qsreaders); |
fdee2e6d | 165 | int wait_loops = 0; |
02be5561 | 166 | struct rcu_reader *index, *tmp; |
fdee2e6d | 167 | |
32c15e4e | 168 | /* Switch parity: 0 -> 1, 1 -> 0 */ |
6cf3827c | 169 | CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); |
2dfb8b5e MD |
170 | |
171 | /* | |
172 | * Must commit qparity update to memory before waiting for other parity | |
173 | * quiescent state. Failure to do so could result in the writer waiting | |
174 | * forever while new readers are always accessing data (no progress). | |
6cf3827c | 175 | * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED. |
2dfb8b5e MD |
176 | */ |
177 | ||
178 | /* | |
5481ddb3 | 179 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
2dfb8b5e MD |
180 | * model easier to understand. It does not have a big performance impact |
181 | * anyway, given this is the write-side. | |
182 | */ | |
5481ddb3 | 183 | cmm_smp_mb(); |
2dfb8b5e | 184 | |
fdee2e6d | 185 | /* |
02be5561 | 186 | * Wait for each thread rcu_reader.ctr count to become 0. |
fdee2e6d MD |
187 | */ |
188 | for (;;) { | |
189 | wait_loops++; | |
16aa9ee8 | 190 | cds_list_for_each_entry_safe(index, tmp, ®istry, node) { |
fdee2e6d | 191 | if (!rcu_old_gp_ongoing(&index->ctr)) |
16aa9ee8 | 192 | cds_list_move(&index->node, &qsreaders); |
fdee2e6d MD |
193 | } |
194 | ||
16aa9ee8 | 195 | if (cds_list_empty(®istry)) { |
fdee2e6d MD |
196 | break; |
197 | } else { | |
198 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) | |
199 | usleep(RCU_SLEEP_DELAY); | |
200 | else | |
06f22bdb | 201 | caa_cpu_relax(); |
fdee2e6d MD |
202 | } |
203 | } | |
204 | /* put back the reader list in the registry */ | |
16aa9ee8 | 205 | cds_list_splice(&qsreaders, ®istry); |
fdee2e6d MD |
206 | } |
207 | ||
208 | void synchronize_rcu(void) | |
209 | { | |
210 | sigset_t newmask, oldmask; | |
211 | int ret; | |
212 | ||
213 | ret = sigemptyset(&newmask); | |
214 | assert(!ret); | |
215 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); | |
216 | assert(!ret); | |
217 | ||
6abb4bd5 | 218 | mutex_lock(&rcu_gp_lock); |
fdee2e6d | 219 | |
16aa9ee8 | 220 | if (cds_list_empty(®istry)) |
2dfb8b5e | 221 | goto out; |
fdee2e6d MD |
222 | |
223 | /* All threads should read qparity before accessing data structure | |
2dfb8b5e | 224 | * where new ptr points to. */ |
fdee2e6d | 225 | /* Write new ptr before changing the qparity */ |
5481ddb3 | 226 | cmm_smp_mb(); |
fdee2e6d | 227 | |
2dfb8b5e MD |
228 | /* Remove old registry elements */ |
229 | rcu_gc_registry(); | |
fdee2e6d MD |
230 | |
231 | /* | |
232 | * Wait for previous parity to be empty of readers. | |
233 | */ | |
2dfb8b5e | 234 | update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */ |
fdee2e6d MD |
235 | |
236 | /* | |
5481ddb3 | 237 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
fdee2e6d MD |
238 | * model easier to understand. It does not have a big performance impact |
239 | * anyway, given this is the write-side. | |
240 | */ | |
5481ddb3 | 241 | cmm_smp_mb(); |
fdee2e6d | 242 | |
fdee2e6d | 243 | /* |
2dfb8b5e | 244 | * Wait for previous parity to be empty of readers. |
fdee2e6d | 245 | */ |
2dfb8b5e | 246 | update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */ |
fdee2e6d MD |
247 | |
248 | /* | |
2dfb8b5e MD |
249 | * Finish waiting for reader threads before letting the old ptr being |
250 | * freed. | |
fdee2e6d | 251 | */ |
5481ddb3 | 252 | cmm_smp_mb(); |
2dfb8b5e | 253 | out: |
6abb4bd5 | 254 | mutex_unlock(&rcu_gp_lock); |
fdee2e6d MD |
255 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
256 | assert(!ret); | |
257 | } | |
258 | ||
259 | /* | |
260 | * library wrappers to be used by non-LGPL compatible source code. | |
261 | */ | |
262 | ||
263 | void rcu_read_lock(void) | |
264 | { | |
265 | _rcu_read_lock(); | |
266 | } | |
267 | ||
268 | void rcu_read_unlock(void) | |
269 | { | |
270 | _rcu_read_unlock(); | |
271 | } | |
272 | ||
273 | /* | |
274 | * only grow for now. | |
275 | */ | |
276 | static void resize_arena(struct registry_arena *arena, size_t len) | |
277 | { | |
278 | void *new_arena; | |
279 | ||
0617bf4c MD |
280 | if (!arena->p) |
281 | new_arena = mmap(arena->p, len, | |
282 | PROT_READ | PROT_WRITE, | |
283 | MAP_ANONYMOUS | MAP_PRIVATE, | |
284 | -1, 0); | |
285 | else | |
286 | new_arena = mremap(arena->p, arena->len, | |
287 | len, MREMAP_MAYMOVE); | |
288 | assert(new_arena != MAP_FAILED); | |
289 | ||
fdee2e6d MD |
290 | /* |
291 | * re-used the same region ? | |
292 | */ | |
293 | if (new_arena == arena->p) | |
294 | return; | |
295 | ||
fdee2e6d MD |
296 | bzero(new_arena + arena->len, len - arena->len); |
297 | arena->p = new_arena; | |
298 | } | |
299 | ||
300 | /* Called with signals off and mutex locked */ | |
301 | static void add_thread(void) | |
302 | { | |
02be5561 | 303 | struct rcu_reader *rcu_reader_reg; |
fdee2e6d MD |
304 | |
305 | if (registry_arena.len | |
02be5561 | 306 | < registry_arena.used + sizeof(struct rcu_reader)) |
fdee2e6d | 307 | resize_arena(®istry_arena, |
2f8a5ae7 | 308 | caa_max(registry_arena.len << 1, ARENA_INIT_ALLOC)); |
fdee2e6d MD |
309 | /* |
310 | * Find a free spot. | |
311 | */ | |
02be5561 MD |
312 | for (rcu_reader_reg = registry_arena.p; |
313 | (void *)rcu_reader_reg < registry_arena.p + registry_arena.len; | |
314 | rcu_reader_reg++) { | |
315 | if (!rcu_reader_reg->alloc) | |
fdee2e6d MD |
316 | break; |
317 | } | |
02be5561 MD |
318 | rcu_reader_reg->alloc = 1; |
319 | registry_arena.used += sizeof(struct rcu_reader); | |
fdee2e6d MD |
320 | |
321 | /* Add to registry */ | |
02be5561 MD |
322 | rcu_reader_reg->tid = pthread_self(); |
323 | assert(rcu_reader_reg->ctr == 0); | |
16aa9ee8 | 324 | cds_list_add(&rcu_reader_reg->node, ®istry); |
02be5561 | 325 | rcu_reader = rcu_reader_reg; |
fdee2e6d MD |
326 | } |
327 | ||
328 | /* Called with signals off and mutex locked */ | |
329 | static void rcu_gc_registry(void) | |
330 | { | |
02be5561 | 331 | struct rcu_reader *rcu_reader_reg; |
fdee2e6d MD |
332 | pthread_t tid; |
333 | int ret; | |
334 | ||
02be5561 MD |
335 | for (rcu_reader_reg = registry_arena.p; |
336 | (void *)rcu_reader_reg < registry_arena.p + registry_arena.len; | |
337 | rcu_reader_reg++) { | |
338 | if (!rcu_reader_reg->alloc) | |
fdee2e6d | 339 | continue; |
02be5561 | 340 | tid = rcu_reader_reg->tid; |
fdee2e6d MD |
341 | ret = pthread_kill(tid, 0); |
342 | assert(ret != EINVAL); | |
343 | if (ret == ESRCH) { | |
16aa9ee8 | 344 | cds_list_del(&rcu_reader_reg->node); |
79266051 | 345 | rcu_reader_reg->ctr = 0; |
02be5561 MD |
346 | rcu_reader_reg->alloc = 0; |
347 | registry_arena.used -= sizeof(struct rcu_reader); | |
fdee2e6d MD |
348 | } |
349 | } | |
350 | } | |
351 | ||
352 | /* Disable signals, take mutex, add to registry */ | |
353 | void rcu_bp_register(void) | |
354 | { | |
355 | sigset_t newmask, oldmask; | |
356 | int ret; | |
357 | ||
358 | ret = sigemptyset(&newmask); | |
359 | assert(!ret); | |
360 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); | |
361 | assert(!ret); | |
362 | ||
363 | /* | |
364 | * Check if a signal concurrently registered our thread since | |
365 | * the check in rcu_read_lock(). */ | |
02be5561 | 366 | if (rcu_reader) |
fdee2e6d MD |
367 | goto end; |
368 | ||
6abb4bd5 | 369 | mutex_lock(&rcu_gp_lock); |
fdee2e6d | 370 | add_thread(); |
6abb4bd5 | 371 | mutex_unlock(&rcu_gp_lock); |
fdee2e6d MD |
372 | end: |
373 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
374 | assert(!ret); | |
375 | } | |
376 | ||
9380711a | 377 | void rcu_bp_exit(void) |
fdee2e6d | 378 | { |
9380711a MD |
379 | if (registry_arena.p) |
380 | munmap(registry_arena.p, registry_arena.len); | |
fdee2e6d | 381 | } |
4cf1675f MD |
382 | |
383 | /* | |
384 | * Holding the rcu_gp_lock across fork will make sure we fork() don't race with | |
385 | * a concurrent thread executing with this same lock held. This ensures that the | |
386 | * registry is in a coherent state in the child. | |
387 | */ | |
388 | void rcu_bp_before_fork(void) | |
389 | { | |
390 | sigset_t newmask, oldmask; | |
391 | int ret; | |
392 | ||
393 | ret = sigemptyset(&newmask); | |
394 | assert(!ret); | |
395 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); | |
396 | assert(!ret); | |
397 | mutex_lock(&rcu_gp_lock); | |
398 | saved_fork_signal_mask = oldmask; | |
399 | } | |
400 | ||
401 | void rcu_bp_after_fork_parent(void) | |
402 | { | |
403 | sigset_t oldmask; | |
404 | int ret; | |
405 | ||
406 | oldmask = saved_fork_signal_mask; | |
407 | mutex_unlock(&rcu_gp_lock); | |
408 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
409 | assert(!ret); | |
410 | } | |
411 | ||
412 | void rcu_bp_after_fork_child(void) | |
413 | { | |
414 | sigset_t oldmask; | |
415 | int ret; | |
416 | ||
417 | rcu_gc_registry(); | |
418 | oldmask = saved_fork_signal_mask; | |
419 | mutex_unlock(&rcu_gp_lock); | |
420 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
421 | assert(!ret); | |
422 | } | |
5e77fc1f | 423 | |
9b7981bb MD |
424 | void *rcu_dereference_sym_bp(void *p) |
425 | { | |
426 | return _rcu_dereference(p); | |
427 | } | |
428 | ||
5efd3cd2 MD |
429 | void *rcu_set_pointer_sym_bp(void **p, void *v) |
430 | { | |
431 | cmm_wmb(); | |
432 | return uatomic_set(p, v); | |
433 | } | |
434 | ||
435 | void *rcu_xchg_pointer_sym_bp(void **p, void *v) | |
436 | { | |
437 | cmm_wmb(); | |
438 | return uatomic_xchg(p, v); | |
439 | } | |
440 | ||
441 | void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new) | |
442 | { | |
443 | cmm_wmb(); | |
444 | return uatomic_cmpxchg(p, old, _new); | |
445 | } | |
446 | ||
541d828d LJ |
447 | DEFINE_RCU_FLAVOR() |
448 | ||
5e77fc1f | 449 | #include "urcu-call-rcu-impl.h" |
0376e7b2 | 450 | #include "urcu-defer-impl.h" |