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 | ||
e37faee1 | 26 | #define URCU_NO_COMPAT_IDENTIFIERS |
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> | |
3745305b | 37 | #include <stdbool.h> |
fdee2e6d MD |
38 | #include <sys/mman.h> |
39 | ||
375db287 | 40 | #include <urcu/config.h> |
4477a870 MD |
41 | #include <urcu/arch.h> |
42 | #include <urcu/wfcqueue.h> | |
43 | #include <urcu/map/urcu-bp.h> | |
44 | #include <urcu/static/urcu-bp.h> | |
45 | #include <urcu/pointer.h> | |
46 | #include <urcu/tls-compat.h> | |
71c811bf | 47 | |
4a6d7378 | 48 | #include "urcu-die.h" |
ce28e67a | 49 | #include "urcu-utils.h" |
4a6d7378 | 50 | |
4477a870 | 51 | #define URCU_API_MAP |
fdee2e6d | 52 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
71c811bf | 53 | #undef _LGPL_SOURCE |
4477a870 | 54 | #include <urcu/urcu-bp.h> |
71c811bf | 55 | #define _LGPL_SOURCE |
fdee2e6d | 56 | |
4c1ae2ea MD |
57 | #ifndef MAP_ANONYMOUS |
58 | #define MAP_ANONYMOUS MAP_ANON | |
59 | #endif | |
60 | ||
c7eaf61c MD |
61 | #ifdef __linux__ |
62 | static | |
63 | void *mremap_wrapper(void *old_address, size_t old_size, | |
64 | size_t new_size, int flags) | |
65 | { | |
66 | return mremap(old_address, old_size, new_size, flags); | |
67 | } | |
68 | #else | |
45a4872f MD |
69 | |
70 | #define MREMAP_MAYMOVE 1 | |
71 | #define MREMAP_FIXED 2 | |
72 | ||
73 | /* | |
95b94246 | 74 | * mremap wrapper for non-Linux systems not allowing MAYMOVE. |
45a4872f MD |
75 | * This is not generic. |
76 | */ | |
c7eaf61c MD |
77 | static |
78 | void *mremap_wrapper(void *old_address, size_t old_size, | |
79 | size_t new_size, int flags) | |
45a4872f | 80 | { |
95b94246 MD |
81 | assert(!(flags & MREMAP_MAYMOVE)); |
82 | ||
83 | return MAP_FAILED; | |
45a4872f MD |
84 | } |
85 | #endif | |
86 | ||
9340c38d MD |
87 | /* Sleep delay in ms */ |
88 | #define RCU_SLEEP_DELAY_MS 10 | |
95b94246 MD |
89 | #define INIT_NR_THREADS 8 |
90 | #define ARENA_INIT_ALLOC \ | |
91 | sizeof(struct registry_chunk) \ | |
4477a870 | 92 | + INIT_NR_THREADS * sizeof(struct urcu_bp_reader) |
fdee2e6d | 93 | |
b7b6a8f5 PB |
94 | /* |
95 | * Active attempts to check for reader Q.S. before calling sleep(). | |
96 | */ | |
97 | #define RCU_QS_ACTIVE_ATTEMPTS 100 | |
98 | ||
76d6a951 | 99 | static |
4477a870 | 100 | int urcu_bp_refcount; |
76d6a951 | 101 | |
999991c6 MD |
102 | /* If the headers do not support membarrier system call, fall back smp_mb. */ |
103 | #ifdef __NR_membarrier | |
104 | # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__) | |
f541831e MD |
105 | #else |
106 | # define membarrier(...) -ENOSYS | |
107 | #endif | |
108 | ||
109 | enum membarrier_cmd { | |
3745305b MD |
110 | MEMBARRIER_CMD_QUERY = 0, |
111 | MEMBARRIER_CMD_SHARED = (1 << 0), | |
112 | /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */ | |
113 | /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */ | |
114 | MEMBARRIER_CMD_PRIVATE_EXPEDITED = (1 << 3), | |
115 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED = (1 << 4), | |
f541831e MD |
116 | }; |
117 | ||
c1be8fb9 | 118 | static |
4477a870 | 119 | void __attribute__((constructor)) _urcu_bp_init(void); |
c1be8fb9 | 120 | static |
4477a870 | 121 | void __attribute__((destructor)) urcu_bp_exit(void); |
fdee2e6d | 122 | |
d8d9a340 | 123 | #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER |
f541831e | 124 | int urcu_bp_has_sys_membarrier; |
d8d9a340 | 125 | #endif |
f541831e | 126 | |
731ccb96 MD |
127 | /* |
128 | * rcu_gp_lock ensures mutual exclusion between threads calling | |
129 | * synchronize_rcu(). | |
130 | */ | |
6abb4bd5 | 131 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
731ccb96 MD |
132 | /* |
133 | * rcu_registry_lock ensures mutual exclusion between threads | |
134 | * registering and unregistering themselves to/from the registry, and | |
135 | * with threads reading that registry from synchronize_rcu(). However, | |
136 | * this lock is not held all the way through the completion of awaiting | |
137 | * for the grace period. It is sporadically released between iterations | |
138 | * on the registry. | |
139 | * rcu_registry_lock may nest inside rcu_gp_lock. | |
140 | */ | |
141 | static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER; | |
fdee2e6d | 142 | |
c1be8fb9 MD |
143 | static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER; |
144 | static int initialized; | |
145 | ||
146 | static pthread_key_t urcu_bp_key; | |
147 | ||
4477a870 | 148 | struct urcu_bp_gp urcu_bp_gp = { .ctr = URCU_BP_GP_COUNT }; |
ce28e67a | 149 | URCU_ATTR_ALIAS("urcu_bp_gp") extern struct urcu_bp_gp rcu_gp_bp; |
fdee2e6d MD |
150 | |
151 | /* | |
152 | * Pointer to registry elements. Written to only by each individual reader. Read | |
153 | * by both the reader and the writers. | |
154 | */ | |
4477a870 | 155 | DEFINE_URCU_TLS(struct urcu_bp_reader *, urcu_bp_reader); |
99bfa9e9 | 156 | DEFINE_URCU_TLS_ALIAS(struct urcu_bp_reader *, urcu_bp_reader, rcu_reader_bp); |
fdee2e6d | 157 | |
16aa9ee8 | 158 | static CDS_LIST_HEAD(registry); |
fdee2e6d | 159 | |
95b94246 MD |
160 | struct registry_chunk { |
161 | size_t data_len; /* data length */ | |
c1be8fb9 | 162 | size_t used; /* amount of data used */ |
95b94246 MD |
163 | struct cds_list_head node; /* chunk_list node */ |
164 | char data[]; | |
165 | }; | |
166 | ||
fdee2e6d | 167 | struct registry_arena { |
95b94246 | 168 | struct cds_list_head chunk_list; |
fdee2e6d MD |
169 | }; |
170 | ||
95b94246 MD |
171 | static struct registry_arena registry_arena = { |
172 | .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list), | |
173 | }; | |
fdee2e6d | 174 | |
4cf1675f MD |
175 | /* Saved fork signal mask, protected by rcu_gp_lock */ |
176 | static sigset_t saved_fork_signal_mask; | |
177 | ||
6abb4bd5 | 178 | static void mutex_lock(pthread_mutex_t *mutex) |
fdee2e6d MD |
179 | { |
180 | int ret; | |
181 | ||
182 | #ifndef DISTRUST_SIGNALS_EXTREME | |
6abb4bd5 | 183 | ret = pthread_mutex_lock(mutex); |
4a6d7378 MD |
184 | if (ret) |
185 | urcu_die(ret); | |
fdee2e6d | 186 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ |
6abb4bd5 | 187 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
4a6d7378 MD |
188 | if (ret != EBUSY && ret != EINTR) |
189 | urcu_die(ret); | |
fdee2e6d MD |
190 | poll(NULL,0,10); |
191 | } | |
192 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
193 | } | |
194 | ||
6abb4bd5 | 195 | static void mutex_unlock(pthread_mutex_t *mutex) |
fdee2e6d MD |
196 | { |
197 | int ret; | |
198 | ||
6abb4bd5 | 199 | ret = pthread_mutex_unlock(mutex); |
4a6d7378 MD |
200 | if (ret) |
201 | urcu_die(ret); | |
fdee2e6d MD |
202 | } |
203 | ||
f541831e MD |
204 | static void smp_mb_master(void) |
205 | { | |
3745305b MD |
206 | if (caa_likely(urcu_bp_has_sys_membarrier)) { |
207 | if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0)) | |
208 | urcu_die(errno); | |
209 | } else { | |
f541831e | 210 | cmm_smp_mb(); |
3745305b | 211 | } |
f541831e MD |
212 | } |
213 | ||
731ccb96 MD |
214 | /* |
215 | * Always called with rcu_registry lock held. Releases this lock between | |
216 | * iterations and grabs it again. Holds the lock when it returns. | |
217 | */ | |
52c75091 MD |
218 | static void wait_for_readers(struct cds_list_head *input_readers, |
219 | struct cds_list_head *cur_snap_readers, | |
220 | struct cds_list_head *qsreaders) | |
fdee2e6d | 221 | { |
9340c38d | 222 | unsigned int wait_loops = 0; |
4477a870 | 223 | struct urcu_bp_reader *index, *tmp; |
fdee2e6d | 224 | |
fdee2e6d | 225 | /* |
4477a870 | 226 | * Wait for each thread URCU_TLS(urcu_bp_reader).ctr to either |
dd61d077 | 227 | * indicate quiescence (not nested), or observe the current |
c13c2e55 | 228 | * rcu_gp.ctr value. |
fdee2e6d MD |
229 | */ |
230 | for (;;) { | |
9340c38d MD |
231 | if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS) |
232 | wait_loops++; | |
233 | ||
52c75091 | 234 | cds_list_for_each_entry_safe(index, tmp, input_readers, node) { |
4477a870 MD |
235 | switch (urcu_bp_reader_state(&index->ctr)) { |
236 | case URCU_BP_READER_ACTIVE_CURRENT: | |
52c75091 MD |
237 | if (cur_snap_readers) { |
238 | cds_list_move(&index->node, | |
239 | cur_snap_readers); | |
240 | break; | |
241 | } | |
242 | /* Fall-through */ | |
4477a870 | 243 | case URCU_BP_READER_INACTIVE: |
52c75091 MD |
244 | cds_list_move(&index->node, qsreaders); |
245 | break; | |
4477a870 | 246 | case URCU_BP_READER_ACTIVE_OLD: |
52c75091 MD |
247 | /* |
248 | * Old snapshot. Leaving node in | |
249 | * input_readers will make us busy-loop | |
250 | * until the snapshot becomes current or | |
251 | * the reader becomes inactive. | |
252 | */ | |
253 | break; | |
254 | } | |
fdee2e6d MD |
255 | } |
256 | ||
52c75091 | 257 | if (cds_list_empty(input_readers)) { |
fdee2e6d MD |
258 | break; |
259 | } else { | |
731ccb96 MD |
260 | /* Temporarily unlock the registry lock. */ |
261 | mutex_unlock(&rcu_registry_lock); | |
9340c38d MD |
262 | if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) |
263 | (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS); | |
fdee2e6d | 264 | else |
06f22bdb | 265 | caa_cpu_relax(); |
731ccb96 MD |
266 | /* Re-lock the registry lock before the next loop. */ |
267 | mutex_lock(&rcu_registry_lock); | |
fdee2e6d MD |
268 | } |
269 | } | |
fdee2e6d MD |
270 | } |
271 | ||
4477a870 | 272 | void urcu_bp_synchronize_rcu(void) |
fdee2e6d | 273 | { |
52c75091 MD |
274 | CDS_LIST_HEAD(cur_snap_readers); |
275 | CDS_LIST_HEAD(qsreaders); | |
fdee2e6d MD |
276 | sigset_t newmask, oldmask; |
277 | int ret; | |
278 | ||
6ed4b2e6 | 279 | ret = sigfillset(&newmask); |
fdee2e6d | 280 | assert(!ret); |
6ed4b2e6 | 281 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
fdee2e6d MD |
282 | assert(!ret); |
283 | ||
6abb4bd5 | 284 | mutex_lock(&rcu_gp_lock); |
fdee2e6d | 285 | |
731ccb96 MD |
286 | mutex_lock(&rcu_registry_lock); |
287 | ||
16aa9ee8 | 288 | if (cds_list_empty(®istry)) |
2dfb8b5e | 289 | goto out; |
fdee2e6d MD |
290 | |
291 | /* All threads should read qparity before accessing data structure | |
2dfb8b5e | 292 | * where new ptr points to. */ |
fdee2e6d | 293 | /* Write new ptr before changing the qparity */ |
f541831e | 294 | smp_mb_master(); |
fdee2e6d | 295 | |
fdee2e6d | 296 | /* |
dd61d077 | 297 | * Wait for readers to observe original parity or be quiescent. |
731ccb96 MD |
298 | * wait_for_readers() can release and grab again rcu_registry_lock |
299 | * interally. | |
dd61d077 | 300 | */ |
52c75091 | 301 | wait_for_readers(®istry, &cur_snap_readers, &qsreaders); |
dd61d077 MD |
302 | |
303 | /* | |
304 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the | |
305 | * model easier to understand. It does not have a big performance impact | |
306 | * anyway, given this is the write-side. | |
307 | */ | |
308 | cmm_smp_mb(); | |
309 | ||
310 | /* Switch parity: 0 -> 1, 1 -> 0 */ | |
4477a870 | 311 | CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_BP_GP_CTR_PHASE); |
dd61d077 MD |
312 | |
313 | /* | |
314 | * Must commit qparity update to memory before waiting for other parity | |
315 | * quiescent state. Failure to do so could result in the writer waiting | |
316 | * forever while new readers are always accessing data (no progress). | |
317 | * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED. | |
fdee2e6d | 318 | */ |
fdee2e6d MD |
319 | |
320 | /* | |
5481ddb3 | 321 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
fdee2e6d MD |
322 | * model easier to understand. It does not have a big performance impact |
323 | * anyway, given this is the write-side. | |
324 | */ | |
5481ddb3 | 325 | cmm_smp_mb(); |
fdee2e6d | 326 | |
fdee2e6d | 327 | /* |
dd61d077 | 328 | * Wait for readers to observe new parity or be quiescent. |
731ccb96 MD |
329 | * wait_for_readers() can release and grab again rcu_registry_lock |
330 | * interally. | |
fdee2e6d | 331 | */ |
52c75091 MD |
332 | wait_for_readers(&cur_snap_readers, NULL, &qsreaders); |
333 | ||
334 | /* | |
335 | * Put quiescent reader list back into registry. | |
336 | */ | |
337 | cds_list_splice(&qsreaders, ®istry); | |
fdee2e6d MD |
338 | |
339 | /* | |
2dfb8b5e MD |
340 | * Finish waiting for reader threads before letting the old ptr being |
341 | * freed. | |
fdee2e6d | 342 | */ |
f541831e | 343 | smp_mb_master(); |
2dfb8b5e | 344 | out: |
731ccb96 | 345 | mutex_unlock(&rcu_registry_lock); |
6abb4bd5 | 346 | mutex_unlock(&rcu_gp_lock); |
fdee2e6d MD |
347 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
348 | assert(!ret); | |
349 | } | |
ce28e67a | 350 | URCU_ATTR_ALIAS("urcu_bp_synchronize_rcu") void synchronize_rcu_bp(); |
fdee2e6d MD |
351 | |
352 | /* | |
353 | * library wrappers to be used by non-LGPL compatible source code. | |
354 | */ | |
355 | ||
4477a870 | 356 | void urcu_bp_read_lock(void) |
fdee2e6d | 357 | { |
4477a870 | 358 | _urcu_bp_read_lock(); |
fdee2e6d | 359 | } |
ce28e67a | 360 | URCU_ATTR_ALIAS("urcu_bp_read_lock") void rcu_read_lock_bp(); |
fdee2e6d | 361 | |
4477a870 | 362 | void urcu_bp_read_unlock(void) |
fdee2e6d | 363 | { |
4477a870 | 364 | _urcu_bp_read_unlock(); |
fdee2e6d | 365 | } |
ce28e67a | 366 | URCU_ATTR_ALIAS("urcu_bp_read_unlock") void rcu_read_unlock_bp(); |
fdee2e6d | 367 | |
4477a870 | 368 | int urcu_bp_read_ongoing(void) |
882f3357 | 369 | { |
4477a870 | 370 | return _urcu_bp_read_ongoing(); |
882f3357 | 371 | } |
ce28e67a | 372 | URCU_ATTR_ALIAS("urcu_bp_read_ongoing") int rcu_read_ongoing_bp(); |
882f3357 | 373 | |
fdee2e6d | 374 | /* |
95b94246 MD |
375 | * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk. |
376 | * Else, try expanding the last chunk. If this fails, allocate a new | |
377 | * chunk twice as big as the last chunk. | |
378 | * Memory used by chunks _never_ moves. A chunk could theoretically be | |
379 | * freed when all "used" slots are released, but we don't do it at this | |
380 | * point. | |
fdee2e6d | 381 | */ |
95b94246 MD |
382 | static |
383 | void expand_arena(struct registry_arena *arena) | |
fdee2e6d | 384 | { |
95b94246 MD |
385 | struct registry_chunk *new_chunk, *last_chunk; |
386 | size_t old_chunk_len, new_chunk_len; | |
387 | ||
388 | /* No chunk. */ | |
389 | if (cds_list_empty(&arena->chunk_list)) { | |
390 | assert(ARENA_INIT_ALLOC >= | |
391 | sizeof(struct registry_chunk) | |
392 | + sizeof(struct rcu_reader)); | |
393 | new_chunk_len = ARENA_INIT_ALLOC; | |
5592d049 MJ |
394 | new_chunk = (struct registry_chunk *) mmap(NULL, |
395 | new_chunk_len, | |
9d8612b7 MD |
396 | PROT_READ | PROT_WRITE, |
397 | MAP_ANONYMOUS | MAP_PRIVATE, | |
398 | -1, 0); | |
95b94246 MD |
399 | if (new_chunk == MAP_FAILED) |
400 | abort(); | |
d3ac5bb7 | 401 | memset(new_chunk, 0, new_chunk_len); |
95b94246 MD |
402 | new_chunk->data_len = |
403 | new_chunk_len - sizeof(struct registry_chunk); | |
404 | cds_list_add_tail(&new_chunk->node, &arena->chunk_list); | |
405 | return; /* We're done. */ | |
406 | } | |
9d8612b7 | 407 | |
95b94246 MD |
408 | /* Try expanding last chunk. */ |
409 | last_chunk = cds_list_entry(arena->chunk_list.prev, | |
410 | struct registry_chunk, node); | |
411 | old_chunk_len = | |
412 | last_chunk->data_len + sizeof(struct registry_chunk); | |
413 | new_chunk_len = old_chunk_len << 1; | |
414 | ||
415 | /* Don't allow memory mapping to move, just expand. */ | |
416 | new_chunk = mremap_wrapper(last_chunk, old_chunk_len, | |
417 | new_chunk_len, 0); | |
418 | if (new_chunk != MAP_FAILED) { | |
419 | /* Should not have moved. */ | |
420 | assert(new_chunk == last_chunk); | |
d3ac5bb7 | 421 | memset((char *) last_chunk + old_chunk_len, 0, |
95b94246 MD |
422 | new_chunk_len - old_chunk_len); |
423 | last_chunk->data_len = | |
424 | new_chunk_len - sizeof(struct registry_chunk); | |
425 | return; /* We're done. */ | |
426 | } | |
0617bf4c | 427 | |
95b94246 | 428 | /* Remap did not succeed, we need to add a new chunk. */ |
5592d049 MJ |
429 | new_chunk = (struct registry_chunk *) mmap(NULL, |
430 | new_chunk_len, | |
95b94246 MD |
431 | PROT_READ | PROT_WRITE, |
432 | MAP_ANONYMOUS | MAP_PRIVATE, | |
433 | -1, 0); | |
434 | if (new_chunk == MAP_FAILED) | |
435 | abort(); | |
d3ac5bb7 | 436 | memset(new_chunk, 0, new_chunk_len); |
95b94246 MD |
437 | new_chunk->data_len = |
438 | new_chunk_len - sizeof(struct registry_chunk); | |
439 | cds_list_add_tail(&new_chunk->node, &arena->chunk_list); | |
440 | } | |
fdee2e6d | 441 | |
95b94246 MD |
442 | static |
443 | struct rcu_reader *arena_alloc(struct registry_arena *arena) | |
444 | { | |
445 | struct registry_chunk *chunk; | |
446 | struct rcu_reader *rcu_reader_reg; | |
447 | int expand_done = 0; /* Only allow to expand once per alloc */ | |
448 | size_t len = sizeof(struct rcu_reader); | |
449 | ||
450 | retry: | |
451 | cds_list_for_each_entry(chunk, &arena->chunk_list, node) { | |
452 | if (chunk->data_len - chunk->used < len) | |
453 | continue; | |
454 | /* Find spot */ | |
455 | for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0]; | |
456 | rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len]; | |
457 | rcu_reader_reg++) { | |
458 | if (!rcu_reader_reg->alloc) { | |
459 | rcu_reader_reg->alloc = 1; | |
460 | chunk->used += len; | |
461 | return rcu_reader_reg; | |
462 | } | |
463 | } | |
464 | } | |
465 | ||
466 | if (!expand_done) { | |
467 | expand_arena(arena); | |
468 | expand_done = 1; | |
469 | goto retry; | |
470 | } | |
471 | ||
472 | return NULL; | |
fdee2e6d MD |
473 | } |
474 | ||
475 | /* Called with signals off and mutex locked */ | |
95b94246 MD |
476 | static |
477 | void add_thread(void) | |
fdee2e6d | 478 | { |
02be5561 | 479 | struct rcu_reader *rcu_reader_reg; |
c1be8fb9 | 480 | int ret; |
fdee2e6d | 481 | |
95b94246 MD |
482 | rcu_reader_reg = arena_alloc(®istry_arena); |
483 | if (!rcu_reader_reg) | |
484 | abort(); | |
c1be8fb9 MD |
485 | ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg); |
486 | if (ret) | |
487 | abort(); | |
fdee2e6d MD |
488 | |
489 | /* Add to registry */ | |
02be5561 MD |
490 | rcu_reader_reg->tid = pthread_self(); |
491 | assert(rcu_reader_reg->ctr == 0); | |
16aa9ee8 | 492 | cds_list_add(&rcu_reader_reg->node, ®istry); |
95b94246 MD |
493 | /* |
494 | * Reader threads are pointing to the reader registry. This is | |
495 | * why its memory should never be relocated. | |
496 | */ | |
4477a870 | 497 | URCU_TLS(urcu_bp_reader) = rcu_reader_reg; |
fdee2e6d MD |
498 | } |
499 | ||
c1be8fb9 MD |
500 | /* Called with mutex locked */ |
501 | static | |
502 | void cleanup_thread(struct registry_chunk *chunk, | |
503 | struct rcu_reader *rcu_reader_reg) | |
504 | { | |
505 | rcu_reader_reg->ctr = 0; | |
506 | cds_list_del(&rcu_reader_reg->node); | |
507 | rcu_reader_reg->tid = 0; | |
508 | rcu_reader_reg->alloc = 0; | |
509 | chunk->used -= sizeof(struct rcu_reader); | |
510 | } | |
511 | ||
512 | static | |
513 | struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg) | |
fdee2e6d | 514 | { |
95b94246 | 515 | struct registry_chunk *chunk; |
fdee2e6d | 516 | |
95b94246 | 517 | cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) { |
c1be8fb9 MD |
518 | if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0]) |
519 | continue; | |
520 | if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len]) | |
521 | continue; | |
522 | return chunk; | |
523 | } | |
524 | return NULL; | |
525 | } | |
95b94246 | 526 | |
c1be8fb9 MD |
527 | /* Called with signals off and mutex locked */ |
528 | static | |
76d6a951 | 529 | void remove_thread(struct rcu_reader *rcu_reader_reg) |
c1be8fb9 | 530 | { |
c1be8fb9 | 531 | cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg); |
4477a870 | 532 | URCU_TLS(urcu_bp_reader) = NULL; |
fdee2e6d MD |
533 | } |
534 | ||
535 | /* Disable signals, take mutex, add to registry */ | |
4477a870 | 536 | void urcu_bp_register(void) |
fdee2e6d MD |
537 | { |
538 | sigset_t newmask, oldmask; | |
539 | int ret; | |
540 | ||
6ed4b2e6 | 541 | ret = sigfillset(&newmask); |
c1be8fb9 MD |
542 | if (ret) |
543 | abort(); | |
6ed4b2e6 | 544 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
c1be8fb9 MD |
545 | if (ret) |
546 | abort(); | |
fdee2e6d MD |
547 | |
548 | /* | |
549 | * Check if a signal concurrently registered our thread since | |
c1be8fb9 MD |
550 | * the check in rcu_read_lock(). |
551 | */ | |
4477a870 | 552 | if (URCU_TLS(urcu_bp_reader)) |
fdee2e6d MD |
553 | goto end; |
554 | ||
c1be8fb9 MD |
555 | /* |
556 | * Take care of early registration before urcu_bp constructor. | |
557 | */ | |
4477a870 | 558 | _urcu_bp_init(); |
c1be8fb9 | 559 | |
731ccb96 | 560 | mutex_lock(&rcu_registry_lock); |
fdee2e6d | 561 | add_thread(); |
731ccb96 | 562 | mutex_unlock(&rcu_registry_lock); |
fdee2e6d MD |
563 | end: |
564 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
c1be8fb9 MD |
565 | if (ret) |
566 | abort(); | |
567 | } | |
ce28e67a | 568 | URCU_ATTR_ALIAS("urcu_bp_register") void rcu_bp_register(); |
c1be8fb9 | 569 | |
5b46e39d MD |
570 | void urcu_bp_register_thread(void) |
571 | { | |
572 | if (caa_unlikely(!URCU_TLS(urcu_bp_reader))) | |
573 | urcu_bp_register(); /* If not yet registered. */ | |
574 | } | |
575 | ||
c1be8fb9 MD |
576 | /* Disable signals, take mutex, remove from registry */ |
577 | static | |
4477a870 | 578 | void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg) |
c1be8fb9 MD |
579 | { |
580 | sigset_t newmask, oldmask; | |
581 | int ret; | |
582 | ||
583 | ret = sigfillset(&newmask); | |
584 | if (ret) | |
585 | abort(); | |
586 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
587 | if (ret) | |
588 | abort(); | |
589 | ||
731ccb96 | 590 | mutex_lock(&rcu_registry_lock); |
76d6a951 | 591 | remove_thread(rcu_reader_reg); |
731ccb96 | 592 | mutex_unlock(&rcu_registry_lock); |
c1be8fb9 MD |
593 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
594 | if (ret) | |
595 | abort(); | |
4477a870 | 596 | urcu_bp_exit(); |
c1be8fb9 MD |
597 | } |
598 | ||
599 | /* | |
600 | * Remove thread from the registry when it exits, and flag it as | |
601 | * destroyed so garbage collection can take care of it. | |
602 | */ | |
603 | static | |
604 | void urcu_bp_thread_exit_notifier(void *rcu_key) | |
605 | { | |
4477a870 | 606 | urcu_bp_unregister(rcu_key); |
c1be8fb9 MD |
607 | } |
608 | ||
d8d9a340 MD |
609 | #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER |
610 | static | |
4477a870 | 611 | void urcu_bp_sys_membarrier_status(bool available) |
d8d9a340 MD |
612 | { |
613 | if (!available) | |
614 | abort(); | |
615 | } | |
616 | #else | |
617 | static | |
4477a870 | 618 | void urcu_bp_sys_membarrier_status(bool available) |
d8d9a340 | 619 | { |
3745305b MD |
620 | if (!available) |
621 | return; | |
622 | urcu_bp_has_sys_membarrier = 1; | |
d8d9a340 MD |
623 | } |
624 | #endif | |
625 | ||
3745305b | 626 | static |
4477a870 | 627 | void urcu_bp_sys_membarrier_init(void) |
3745305b MD |
628 | { |
629 | bool available = false; | |
630 | int mask; | |
631 | ||
632 | mask = membarrier(MEMBARRIER_CMD_QUERY, 0); | |
633 | if (mask >= 0) { | |
634 | if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) { | |
635 | if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0)) | |
636 | urcu_die(errno); | |
637 | available = true; | |
638 | } | |
639 | } | |
4477a870 | 640 | urcu_bp_sys_membarrier_status(available); |
3745305b MD |
641 | } |
642 | ||
c1be8fb9 | 643 | static |
4477a870 | 644 | void _urcu_bp_init(void) |
c1be8fb9 MD |
645 | { |
646 | mutex_lock(&init_lock); | |
4477a870 | 647 | if (!urcu_bp_refcount++) { |
c1be8fb9 MD |
648 | int ret; |
649 | ||
650 | ret = pthread_key_create(&urcu_bp_key, | |
651 | urcu_bp_thread_exit_notifier); | |
652 | if (ret) | |
653 | abort(); | |
4477a870 | 654 | urcu_bp_sys_membarrier_init(); |
c1be8fb9 MD |
655 | initialized = 1; |
656 | } | |
657 | mutex_unlock(&init_lock); | |
fdee2e6d MD |
658 | } |
659 | ||
c1be8fb9 | 660 | static |
4477a870 | 661 | void urcu_bp_exit(void) |
fdee2e6d | 662 | { |
76d6a951 | 663 | mutex_lock(&init_lock); |
4477a870 | 664 | if (!--urcu_bp_refcount) { |
76d6a951 MD |
665 | struct registry_chunk *chunk, *tmp; |
666 | int ret; | |
95b94246 | 667 | |
76d6a951 MD |
668 | cds_list_for_each_entry_safe(chunk, tmp, |
669 | ®istry_arena.chunk_list, node) { | |
5592d049 | 670 | munmap((void *) chunk, chunk->data_len |
76d6a951 MD |
671 | + sizeof(struct registry_chunk)); |
672 | } | |
7937ae1c | 673 | CDS_INIT_LIST_HEAD(®istry_arena.chunk_list); |
76d6a951 MD |
674 | ret = pthread_key_delete(urcu_bp_key); |
675 | if (ret) | |
676 | abort(); | |
95b94246 | 677 | } |
76d6a951 | 678 | mutex_unlock(&init_lock); |
fdee2e6d | 679 | } |
4cf1675f MD |
680 | |
681 | /* | |
731ccb96 MD |
682 | * Holding the rcu_gp_lock and rcu_registry_lock across fork will make |
683 | * sure we fork() don't race with a concurrent thread executing with | |
684 | * any of those locks held. This ensures that the registry and data | |
685 | * protected by rcu_gp_lock are in a coherent state in the child. | |
4cf1675f | 686 | */ |
4477a870 | 687 | void urcu_bp_before_fork(void) |
4cf1675f MD |
688 | { |
689 | sigset_t newmask, oldmask; | |
690 | int ret; | |
691 | ||
6ed4b2e6 | 692 | ret = sigfillset(&newmask); |
4cf1675f | 693 | assert(!ret); |
6ed4b2e6 | 694 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
4cf1675f MD |
695 | assert(!ret); |
696 | mutex_lock(&rcu_gp_lock); | |
731ccb96 | 697 | mutex_lock(&rcu_registry_lock); |
4cf1675f MD |
698 | saved_fork_signal_mask = oldmask; |
699 | } | |
ce28e67a | 700 | URCU_ATTR_ALIAS("urcu_bp_before_fork") void rcu_bp_before_fork(); |
4cf1675f | 701 | |
4477a870 | 702 | void urcu_bp_after_fork_parent(void) |
4cf1675f MD |
703 | { |
704 | sigset_t oldmask; | |
705 | int ret; | |
706 | ||
707 | oldmask = saved_fork_signal_mask; | |
731ccb96 | 708 | mutex_unlock(&rcu_registry_lock); |
4cf1675f MD |
709 | mutex_unlock(&rcu_gp_lock); |
710 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
711 | assert(!ret); | |
712 | } | |
ce28e67a | 713 | URCU_ATTR_ALIAS("urcu_bp_after_fork_parent") |
4477a870 | 714 | void rcu_bp_after_fork_parent(void); |
4cf1675f | 715 | |
c1be8fb9 MD |
716 | /* |
717 | * Prune all entries from registry except our own thread. Fits the Linux | |
731ccb96 | 718 | * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held. |
c1be8fb9 MD |
719 | */ |
720 | static | |
721 | void urcu_bp_prune_registry(void) | |
722 | { | |
723 | struct registry_chunk *chunk; | |
4477a870 | 724 | struct urcu_bp_reader *rcu_reader_reg; |
c1be8fb9 MD |
725 | |
726 | cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) { | |
4477a870 MD |
727 | for (rcu_reader_reg = (struct urcu_bp_reader *) &chunk->data[0]; |
728 | rcu_reader_reg < (struct urcu_bp_reader *) &chunk->data[chunk->data_len]; | |
c1be8fb9 MD |
729 | rcu_reader_reg++) { |
730 | if (!rcu_reader_reg->alloc) | |
731 | continue; | |
732 | if (rcu_reader_reg->tid == pthread_self()) | |
733 | continue; | |
734 | cleanup_thread(chunk, rcu_reader_reg); | |
735 | } | |
736 | } | |
737 | } | |
738 | ||
4477a870 | 739 | void urcu_bp_after_fork_child(void) |
4cf1675f MD |
740 | { |
741 | sigset_t oldmask; | |
742 | int ret; | |
743 | ||
c1be8fb9 | 744 | urcu_bp_prune_registry(); |
4cf1675f | 745 | oldmask = saved_fork_signal_mask; |
731ccb96 | 746 | mutex_unlock(&rcu_registry_lock); |
4cf1675f MD |
747 | mutex_unlock(&rcu_gp_lock); |
748 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
749 | assert(!ret); | |
750 | } | |
ce28e67a | 751 | URCU_ATTR_ALIAS("urcu_bp_after_fork_child") |
4477a870 | 752 | void rcu_bp_after_fork_child(void); |
5e77fc1f | 753 | |
4477a870 | 754 | void *urcu_bp_dereference_sym(void *p) |
9b7981bb MD |
755 | { |
756 | return _rcu_dereference(p); | |
757 | } | |
ce28e67a | 758 | URCU_ATTR_ALIAS("urcu_bp_dereference_sym") |
4477a870 | 759 | void *rcu_dereference_sym_bp(); |
9b7981bb | 760 | |
4477a870 | 761 | void *urcu_bp_set_pointer_sym(void **p, void *v) |
5efd3cd2 MD |
762 | { |
763 | cmm_wmb(); | |
424d4ed5 MD |
764 | uatomic_set(p, v); |
765 | return v; | |
5efd3cd2 | 766 | } |
ce28e67a | 767 | URCU_ATTR_ALIAS("urcu_bp_set_pointer_sym") |
4477a870 | 768 | void *rcu_set_pointer_sym_bp(); |
5efd3cd2 | 769 | |
4477a870 | 770 | void *urcu_bp_xchg_pointer_sym(void **p, void *v) |
5efd3cd2 MD |
771 | { |
772 | cmm_wmb(); | |
773 | return uatomic_xchg(p, v); | |
774 | } | |
ce28e67a | 775 | URCU_ATTR_ALIAS("urcu_bp_xchg_pointer_sym") |
4477a870 | 776 | void *rcu_xchg_pointer_sym_bp(); |
5efd3cd2 | 777 | |
4477a870 | 778 | void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new) |
5efd3cd2 MD |
779 | { |
780 | cmm_wmb(); | |
781 | return uatomic_cmpxchg(p, old, _new); | |
782 | } | |
ce28e67a | 783 | URCU_ATTR_ALIAS("urcu_bp_cmpxchg_pointer_sym") |
4477a870 | 784 | void *rcu_cmpxchg_pointer_sym_bp(); |
5efd3cd2 | 785 | |
5e6b23a6 | 786 | DEFINE_RCU_FLAVOR(rcu_flavor); |
4477a870 | 787 | DEFINE_RCU_FLAVOR_ALIAS(rcu_flavor, alias_rcu_flavor); |
541d828d | 788 | |
5e77fc1f | 789 | #include "urcu-call-rcu-impl.h" |
0376e7b2 | 790 | #include "urcu-defer-impl.h" |