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 | ||
c6bc6503 | 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 | 77 | static |
0d657320 MJ |
78 | void *mremap_wrapper(void *old_address __attribute__((unused)), |
79 | size_t old_size __attribute__((unused)), | |
80 | size_t new_size __attribute__((unused)), | |
81 | int flags) | |
45a4872f | 82 | { |
95b94246 MD |
83 | assert(!(flags & MREMAP_MAYMOVE)); |
84 | ||
85 | return MAP_FAILED; | |
45a4872f MD |
86 | } |
87 | #endif | |
88 | ||
9340c38d MD |
89 | /* Sleep delay in ms */ |
90 | #define RCU_SLEEP_DELAY_MS 10 | |
95b94246 MD |
91 | #define INIT_NR_THREADS 8 |
92 | #define ARENA_INIT_ALLOC \ | |
93 | sizeof(struct registry_chunk) \ | |
4477a870 | 94 | + INIT_NR_THREADS * sizeof(struct urcu_bp_reader) |
fdee2e6d | 95 | |
b7b6a8f5 PB |
96 | /* |
97 | * Active attempts to check for reader Q.S. before calling sleep(). | |
98 | */ | |
99 | #define RCU_QS_ACTIVE_ATTEMPTS 100 | |
100 | ||
76d6a951 | 101 | static |
4477a870 | 102 | int urcu_bp_refcount; |
76d6a951 | 103 | |
999991c6 MD |
104 | /* If the headers do not support membarrier system call, fall back smp_mb. */ |
105 | #ifdef __NR_membarrier | |
106 | # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__) | |
f541831e MD |
107 | #else |
108 | # define membarrier(...) -ENOSYS | |
109 | #endif | |
110 | ||
111 | enum membarrier_cmd { | |
3745305b MD |
112 | MEMBARRIER_CMD_QUERY = 0, |
113 | MEMBARRIER_CMD_SHARED = (1 << 0), | |
114 | /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */ | |
115 | /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */ | |
116 | MEMBARRIER_CMD_PRIVATE_EXPEDITED = (1 << 3), | |
117 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED = (1 << 4), | |
f541831e MD |
118 | }; |
119 | ||
c1be8fb9 | 120 | static |
4477a870 | 121 | void __attribute__((constructor)) _urcu_bp_init(void); |
c1be8fb9 | 122 | static |
4477a870 | 123 | void __attribute__((destructor)) urcu_bp_exit(void); |
457eeeee | 124 | static void urcu_call_rcu_exit(void); |
fdee2e6d | 125 | |
d8d9a340 | 126 | #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER |
f541831e | 127 | int urcu_bp_has_sys_membarrier; |
d8d9a340 | 128 | #endif |
f541831e | 129 | |
731ccb96 MD |
130 | /* |
131 | * rcu_gp_lock ensures mutual exclusion between threads calling | |
132 | * synchronize_rcu(). | |
133 | */ | |
6abb4bd5 | 134 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
731ccb96 MD |
135 | /* |
136 | * rcu_registry_lock ensures mutual exclusion between threads | |
137 | * registering and unregistering themselves to/from the registry, and | |
138 | * with threads reading that registry from synchronize_rcu(). However, | |
139 | * this lock is not held all the way through the completion of awaiting | |
140 | * for the grace period. It is sporadically released between iterations | |
141 | * on the registry. | |
142 | * rcu_registry_lock may nest inside rcu_gp_lock. | |
143 | */ | |
144 | static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER; | |
fdee2e6d | 145 | |
c1be8fb9 MD |
146 | static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER; |
147 | static int initialized; | |
148 | ||
149 | static pthread_key_t urcu_bp_key; | |
150 | ||
4477a870 | 151 | struct urcu_bp_gp urcu_bp_gp = { .ctr = URCU_BP_GP_COUNT }; |
ce28e67a | 152 | URCU_ATTR_ALIAS("urcu_bp_gp") extern struct urcu_bp_gp rcu_gp_bp; |
fdee2e6d MD |
153 | |
154 | /* | |
155 | * Pointer to registry elements. Written to only by each individual reader. Read | |
156 | * by both the reader and the writers. | |
157 | */ | |
4477a870 | 158 | DEFINE_URCU_TLS(struct urcu_bp_reader *, urcu_bp_reader); |
99bfa9e9 | 159 | DEFINE_URCU_TLS_ALIAS(struct urcu_bp_reader *, urcu_bp_reader, rcu_reader_bp); |
fdee2e6d | 160 | |
16aa9ee8 | 161 | static CDS_LIST_HEAD(registry); |
fdee2e6d | 162 | |
95b94246 MD |
163 | struct registry_chunk { |
164 | size_t data_len; /* data length */ | |
c1be8fb9 | 165 | size_t used; /* amount of data used */ |
95b94246 MD |
166 | struct cds_list_head node; /* chunk_list node */ |
167 | char data[]; | |
168 | }; | |
169 | ||
fdee2e6d | 170 | struct registry_arena { |
95b94246 | 171 | struct cds_list_head chunk_list; |
fdee2e6d MD |
172 | }; |
173 | ||
95b94246 MD |
174 | static struct registry_arena registry_arena = { |
175 | .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list), | |
176 | }; | |
fdee2e6d | 177 | |
4cf1675f MD |
178 | /* Saved fork signal mask, protected by rcu_gp_lock */ |
179 | static sigset_t saved_fork_signal_mask; | |
180 | ||
6abb4bd5 | 181 | static void mutex_lock(pthread_mutex_t *mutex) |
fdee2e6d MD |
182 | { |
183 | int ret; | |
184 | ||
185 | #ifndef DISTRUST_SIGNALS_EXTREME | |
6abb4bd5 | 186 | ret = pthread_mutex_lock(mutex); |
4a6d7378 MD |
187 | if (ret) |
188 | urcu_die(ret); | |
fdee2e6d | 189 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ |
6abb4bd5 | 190 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
4a6d7378 MD |
191 | if (ret != EBUSY && ret != EINTR) |
192 | urcu_die(ret); | |
fdee2e6d MD |
193 | poll(NULL,0,10); |
194 | } | |
195 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
196 | } | |
197 | ||
6abb4bd5 | 198 | static void mutex_unlock(pthread_mutex_t *mutex) |
fdee2e6d MD |
199 | { |
200 | int ret; | |
201 | ||
6abb4bd5 | 202 | ret = pthread_mutex_unlock(mutex); |
4a6d7378 MD |
203 | if (ret) |
204 | urcu_die(ret); | |
fdee2e6d MD |
205 | } |
206 | ||
f541831e MD |
207 | static void smp_mb_master(void) |
208 | { | |
3745305b MD |
209 | if (caa_likely(urcu_bp_has_sys_membarrier)) { |
210 | if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0)) | |
211 | urcu_die(errno); | |
212 | } else { | |
f541831e | 213 | cmm_smp_mb(); |
3745305b | 214 | } |
f541831e MD |
215 | } |
216 | ||
731ccb96 MD |
217 | /* |
218 | * Always called with rcu_registry lock held. Releases this lock between | |
219 | * iterations and grabs it again. Holds the lock when it returns. | |
220 | */ | |
52c75091 MD |
221 | static void wait_for_readers(struct cds_list_head *input_readers, |
222 | struct cds_list_head *cur_snap_readers, | |
223 | struct cds_list_head *qsreaders) | |
fdee2e6d | 224 | { |
9340c38d | 225 | unsigned int wait_loops = 0; |
4477a870 | 226 | struct urcu_bp_reader *index, *tmp; |
fdee2e6d | 227 | |
fdee2e6d | 228 | /* |
4477a870 | 229 | * Wait for each thread URCU_TLS(urcu_bp_reader).ctr to either |
dd61d077 | 230 | * indicate quiescence (not nested), or observe the current |
c13c2e55 | 231 | * rcu_gp.ctr value. |
fdee2e6d MD |
232 | */ |
233 | for (;;) { | |
9340c38d MD |
234 | if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS) |
235 | wait_loops++; | |
236 | ||
52c75091 | 237 | cds_list_for_each_entry_safe(index, tmp, input_readers, node) { |
4477a870 MD |
238 | switch (urcu_bp_reader_state(&index->ctr)) { |
239 | case URCU_BP_READER_ACTIVE_CURRENT: | |
52c75091 MD |
240 | if (cur_snap_readers) { |
241 | cds_list_move(&index->node, | |
242 | cur_snap_readers); | |
243 | break; | |
244 | } | |
245 | /* Fall-through */ | |
4477a870 | 246 | case URCU_BP_READER_INACTIVE: |
52c75091 MD |
247 | cds_list_move(&index->node, qsreaders); |
248 | break; | |
4477a870 | 249 | case URCU_BP_READER_ACTIVE_OLD: |
52c75091 MD |
250 | /* |
251 | * Old snapshot. Leaving node in | |
252 | * input_readers will make us busy-loop | |
253 | * until the snapshot becomes current or | |
254 | * the reader becomes inactive. | |
255 | */ | |
256 | break; | |
257 | } | |
fdee2e6d MD |
258 | } |
259 | ||
52c75091 | 260 | if (cds_list_empty(input_readers)) { |
fdee2e6d MD |
261 | break; |
262 | } else { | |
731ccb96 MD |
263 | /* Temporarily unlock the registry lock. */ |
264 | mutex_unlock(&rcu_registry_lock); | |
9340c38d MD |
265 | if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) |
266 | (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS); | |
fdee2e6d | 267 | else |
06f22bdb | 268 | caa_cpu_relax(); |
731ccb96 MD |
269 | /* Re-lock the registry lock before the next loop. */ |
270 | mutex_lock(&rcu_registry_lock); | |
fdee2e6d MD |
271 | } |
272 | } | |
fdee2e6d MD |
273 | } |
274 | ||
4477a870 | 275 | void urcu_bp_synchronize_rcu(void) |
fdee2e6d | 276 | { |
52c75091 MD |
277 | CDS_LIST_HEAD(cur_snap_readers); |
278 | CDS_LIST_HEAD(qsreaders); | |
fdee2e6d MD |
279 | sigset_t newmask, oldmask; |
280 | int ret; | |
281 | ||
6ed4b2e6 | 282 | ret = sigfillset(&newmask); |
fdee2e6d | 283 | assert(!ret); |
6ed4b2e6 | 284 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
fdee2e6d MD |
285 | assert(!ret); |
286 | ||
6abb4bd5 | 287 | mutex_lock(&rcu_gp_lock); |
fdee2e6d | 288 | |
731ccb96 MD |
289 | mutex_lock(&rcu_registry_lock); |
290 | ||
16aa9ee8 | 291 | if (cds_list_empty(®istry)) |
2dfb8b5e | 292 | goto out; |
fdee2e6d MD |
293 | |
294 | /* All threads should read qparity before accessing data structure | |
2dfb8b5e | 295 | * where new ptr points to. */ |
fdee2e6d | 296 | /* Write new ptr before changing the qparity */ |
f541831e | 297 | smp_mb_master(); |
fdee2e6d | 298 | |
fdee2e6d | 299 | /* |
dd61d077 | 300 | * Wait for readers to observe original parity or be quiescent. |
731ccb96 MD |
301 | * wait_for_readers() can release and grab again rcu_registry_lock |
302 | * interally. | |
dd61d077 | 303 | */ |
52c75091 | 304 | wait_for_readers(®istry, &cur_snap_readers, &qsreaders); |
dd61d077 MD |
305 | |
306 | /* | |
307 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the | |
308 | * model easier to understand. It does not have a big performance impact | |
309 | * anyway, given this is the write-side. | |
310 | */ | |
311 | cmm_smp_mb(); | |
312 | ||
313 | /* Switch parity: 0 -> 1, 1 -> 0 */ | |
4477a870 | 314 | CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_BP_GP_CTR_PHASE); |
dd61d077 MD |
315 | |
316 | /* | |
317 | * Must commit qparity update to memory before waiting for other parity | |
318 | * quiescent state. Failure to do so could result in the writer waiting | |
319 | * forever while new readers are always accessing data (no progress). | |
320 | * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED. | |
fdee2e6d | 321 | */ |
fdee2e6d MD |
322 | |
323 | /* | |
5481ddb3 | 324 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
fdee2e6d MD |
325 | * model easier to understand. It does not have a big performance impact |
326 | * anyway, given this is the write-side. | |
327 | */ | |
5481ddb3 | 328 | cmm_smp_mb(); |
fdee2e6d | 329 | |
fdee2e6d | 330 | /* |
dd61d077 | 331 | * Wait for readers to observe new parity or be quiescent. |
731ccb96 MD |
332 | * wait_for_readers() can release and grab again rcu_registry_lock |
333 | * interally. | |
fdee2e6d | 334 | */ |
52c75091 MD |
335 | wait_for_readers(&cur_snap_readers, NULL, &qsreaders); |
336 | ||
337 | /* | |
338 | * Put quiescent reader list back into registry. | |
339 | */ | |
340 | cds_list_splice(&qsreaders, ®istry); | |
fdee2e6d MD |
341 | |
342 | /* | |
2dfb8b5e MD |
343 | * Finish waiting for reader threads before letting the old ptr being |
344 | * freed. | |
fdee2e6d | 345 | */ |
f541831e | 346 | smp_mb_master(); |
2dfb8b5e | 347 | out: |
731ccb96 | 348 | mutex_unlock(&rcu_registry_lock); |
6abb4bd5 | 349 | mutex_unlock(&rcu_gp_lock); |
fdee2e6d MD |
350 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
351 | assert(!ret); | |
352 | } | |
ce28e67a | 353 | URCU_ATTR_ALIAS("urcu_bp_synchronize_rcu") void synchronize_rcu_bp(); |
fdee2e6d MD |
354 | |
355 | /* | |
356 | * library wrappers to be used by non-LGPL compatible source code. | |
357 | */ | |
358 | ||
4477a870 | 359 | void urcu_bp_read_lock(void) |
fdee2e6d | 360 | { |
4477a870 | 361 | _urcu_bp_read_lock(); |
fdee2e6d | 362 | } |
ce28e67a | 363 | URCU_ATTR_ALIAS("urcu_bp_read_lock") void rcu_read_lock_bp(); |
fdee2e6d | 364 | |
4477a870 | 365 | void urcu_bp_read_unlock(void) |
fdee2e6d | 366 | { |
4477a870 | 367 | _urcu_bp_read_unlock(); |
fdee2e6d | 368 | } |
ce28e67a | 369 | URCU_ATTR_ALIAS("urcu_bp_read_unlock") void rcu_read_unlock_bp(); |
fdee2e6d | 370 | |
4477a870 | 371 | int urcu_bp_read_ongoing(void) |
882f3357 | 372 | { |
4477a870 | 373 | return _urcu_bp_read_ongoing(); |
882f3357 | 374 | } |
ce28e67a | 375 | URCU_ATTR_ALIAS("urcu_bp_read_ongoing") int rcu_read_ongoing_bp(); |
882f3357 | 376 | |
fdee2e6d | 377 | /* |
95b94246 MD |
378 | * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk. |
379 | * Else, try expanding the last chunk. If this fails, allocate a new | |
380 | * chunk twice as big as the last chunk. | |
381 | * Memory used by chunks _never_ moves. A chunk could theoretically be | |
382 | * freed when all "used" slots are released, but we don't do it at this | |
383 | * point. | |
fdee2e6d | 384 | */ |
95b94246 MD |
385 | static |
386 | void expand_arena(struct registry_arena *arena) | |
fdee2e6d | 387 | { |
95b94246 MD |
388 | struct registry_chunk *new_chunk, *last_chunk; |
389 | size_t old_chunk_len, new_chunk_len; | |
390 | ||
391 | /* No chunk. */ | |
392 | if (cds_list_empty(&arena->chunk_list)) { | |
393 | assert(ARENA_INIT_ALLOC >= | |
394 | sizeof(struct registry_chunk) | |
395 | + sizeof(struct rcu_reader)); | |
396 | new_chunk_len = ARENA_INIT_ALLOC; | |
5592d049 MJ |
397 | new_chunk = (struct registry_chunk *) mmap(NULL, |
398 | new_chunk_len, | |
9d8612b7 MD |
399 | PROT_READ | PROT_WRITE, |
400 | MAP_ANONYMOUS | MAP_PRIVATE, | |
401 | -1, 0); | |
95b94246 MD |
402 | if (new_chunk == MAP_FAILED) |
403 | abort(); | |
d3ac5bb7 | 404 | memset(new_chunk, 0, new_chunk_len); |
95b94246 MD |
405 | new_chunk->data_len = |
406 | new_chunk_len - sizeof(struct registry_chunk); | |
407 | cds_list_add_tail(&new_chunk->node, &arena->chunk_list); | |
408 | return; /* We're done. */ | |
409 | } | |
9d8612b7 | 410 | |
95b94246 MD |
411 | /* Try expanding last chunk. */ |
412 | last_chunk = cds_list_entry(arena->chunk_list.prev, | |
413 | struct registry_chunk, node); | |
414 | old_chunk_len = | |
415 | last_chunk->data_len + sizeof(struct registry_chunk); | |
416 | new_chunk_len = old_chunk_len << 1; | |
417 | ||
418 | /* Don't allow memory mapping to move, just expand. */ | |
419 | new_chunk = mremap_wrapper(last_chunk, old_chunk_len, | |
420 | new_chunk_len, 0); | |
421 | if (new_chunk != MAP_FAILED) { | |
422 | /* Should not have moved. */ | |
423 | assert(new_chunk == last_chunk); | |
d3ac5bb7 | 424 | memset((char *) last_chunk + old_chunk_len, 0, |
95b94246 MD |
425 | new_chunk_len - old_chunk_len); |
426 | last_chunk->data_len = | |
427 | new_chunk_len - sizeof(struct registry_chunk); | |
428 | return; /* We're done. */ | |
429 | } | |
0617bf4c | 430 | |
95b94246 | 431 | /* Remap did not succeed, we need to add a new chunk. */ |
5592d049 MJ |
432 | new_chunk = (struct registry_chunk *) mmap(NULL, |
433 | new_chunk_len, | |
95b94246 MD |
434 | PROT_READ | PROT_WRITE, |
435 | MAP_ANONYMOUS | MAP_PRIVATE, | |
436 | -1, 0); | |
437 | if (new_chunk == MAP_FAILED) | |
438 | abort(); | |
d3ac5bb7 | 439 | memset(new_chunk, 0, new_chunk_len); |
95b94246 MD |
440 | new_chunk->data_len = |
441 | new_chunk_len - sizeof(struct registry_chunk); | |
442 | cds_list_add_tail(&new_chunk->node, &arena->chunk_list); | |
443 | } | |
fdee2e6d | 444 | |
95b94246 MD |
445 | static |
446 | struct rcu_reader *arena_alloc(struct registry_arena *arena) | |
447 | { | |
448 | struct registry_chunk *chunk; | |
449 | struct rcu_reader *rcu_reader_reg; | |
450 | int expand_done = 0; /* Only allow to expand once per alloc */ | |
451 | size_t len = sizeof(struct rcu_reader); | |
452 | ||
453 | retry: | |
454 | cds_list_for_each_entry(chunk, &arena->chunk_list, node) { | |
455 | if (chunk->data_len - chunk->used < len) | |
456 | continue; | |
457 | /* Find spot */ | |
458 | for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0]; | |
459 | rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len]; | |
460 | rcu_reader_reg++) { | |
461 | if (!rcu_reader_reg->alloc) { | |
462 | rcu_reader_reg->alloc = 1; | |
463 | chunk->used += len; | |
464 | return rcu_reader_reg; | |
465 | } | |
466 | } | |
467 | } | |
468 | ||
469 | if (!expand_done) { | |
470 | expand_arena(arena); | |
471 | expand_done = 1; | |
472 | goto retry; | |
473 | } | |
474 | ||
475 | return NULL; | |
fdee2e6d MD |
476 | } |
477 | ||
478 | /* Called with signals off and mutex locked */ | |
95b94246 MD |
479 | static |
480 | void add_thread(void) | |
fdee2e6d | 481 | { |
02be5561 | 482 | struct rcu_reader *rcu_reader_reg; |
c1be8fb9 | 483 | int ret; |
fdee2e6d | 484 | |
95b94246 MD |
485 | rcu_reader_reg = arena_alloc(®istry_arena); |
486 | if (!rcu_reader_reg) | |
487 | abort(); | |
c1be8fb9 MD |
488 | ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg); |
489 | if (ret) | |
490 | abort(); | |
fdee2e6d MD |
491 | |
492 | /* Add to registry */ | |
02be5561 MD |
493 | rcu_reader_reg->tid = pthread_self(); |
494 | assert(rcu_reader_reg->ctr == 0); | |
16aa9ee8 | 495 | cds_list_add(&rcu_reader_reg->node, ®istry); |
95b94246 MD |
496 | /* |
497 | * Reader threads are pointing to the reader registry. This is | |
498 | * why its memory should never be relocated. | |
499 | */ | |
4477a870 | 500 | URCU_TLS(urcu_bp_reader) = rcu_reader_reg; |
fdee2e6d MD |
501 | } |
502 | ||
c1be8fb9 MD |
503 | /* Called with mutex locked */ |
504 | static | |
505 | void cleanup_thread(struct registry_chunk *chunk, | |
506 | struct rcu_reader *rcu_reader_reg) | |
507 | { | |
508 | rcu_reader_reg->ctr = 0; | |
509 | cds_list_del(&rcu_reader_reg->node); | |
510 | rcu_reader_reg->tid = 0; | |
511 | rcu_reader_reg->alloc = 0; | |
512 | chunk->used -= sizeof(struct rcu_reader); | |
513 | } | |
514 | ||
515 | static | |
516 | struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg) | |
fdee2e6d | 517 | { |
95b94246 | 518 | struct registry_chunk *chunk; |
fdee2e6d | 519 | |
95b94246 | 520 | cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) { |
c1be8fb9 MD |
521 | if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0]) |
522 | continue; | |
523 | if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len]) | |
524 | continue; | |
525 | return chunk; | |
526 | } | |
527 | return NULL; | |
528 | } | |
95b94246 | 529 | |
c1be8fb9 MD |
530 | /* Called with signals off and mutex locked */ |
531 | static | |
76d6a951 | 532 | void remove_thread(struct rcu_reader *rcu_reader_reg) |
c1be8fb9 | 533 | { |
c1be8fb9 | 534 | cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg); |
4477a870 | 535 | URCU_TLS(urcu_bp_reader) = NULL; |
fdee2e6d MD |
536 | } |
537 | ||
538 | /* Disable signals, take mutex, add to registry */ | |
4477a870 | 539 | void urcu_bp_register(void) |
fdee2e6d MD |
540 | { |
541 | sigset_t newmask, oldmask; | |
542 | int ret; | |
543 | ||
6ed4b2e6 | 544 | ret = sigfillset(&newmask); |
c1be8fb9 MD |
545 | if (ret) |
546 | abort(); | |
6ed4b2e6 | 547 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
c1be8fb9 MD |
548 | if (ret) |
549 | abort(); | |
fdee2e6d MD |
550 | |
551 | /* | |
552 | * Check if a signal concurrently registered our thread since | |
c1be8fb9 MD |
553 | * the check in rcu_read_lock(). |
554 | */ | |
4477a870 | 555 | if (URCU_TLS(urcu_bp_reader)) |
fdee2e6d MD |
556 | goto end; |
557 | ||
c1be8fb9 MD |
558 | /* |
559 | * Take care of early registration before urcu_bp constructor. | |
560 | */ | |
4477a870 | 561 | _urcu_bp_init(); |
c1be8fb9 | 562 | |
731ccb96 | 563 | mutex_lock(&rcu_registry_lock); |
fdee2e6d | 564 | add_thread(); |
731ccb96 | 565 | mutex_unlock(&rcu_registry_lock); |
fdee2e6d MD |
566 | end: |
567 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
c1be8fb9 MD |
568 | if (ret) |
569 | abort(); | |
570 | } | |
ce28e67a | 571 | URCU_ATTR_ALIAS("urcu_bp_register") void rcu_bp_register(); |
c1be8fb9 | 572 | |
5b46e39d MD |
573 | void urcu_bp_register_thread(void) |
574 | { | |
575 | if (caa_unlikely(!URCU_TLS(urcu_bp_reader))) | |
576 | urcu_bp_register(); /* If not yet registered. */ | |
577 | } | |
578 | ||
c1be8fb9 MD |
579 | /* Disable signals, take mutex, remove from registry */ |
580 | static | |
4477a870 | 581 | void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg) |
c1be8fb9 MD |
582 | { |
583 | sigset_t newmask, oldmask; | |
584 | int ret; | |
585 | ||
586 | ret = sigfillset(&newmask); | |
587 | if (ret) | |
588 | abort(); | |
589 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
590 | if (ret) | |
591 | abort(); | |
592 | ||
731ccb96 | 593 | mutex_lock(&rcu_registry_lock); |
76d6a951 | 594 | remove_thread(rcu_reader_reg); |
731ccb96 | 595 | mutex_unlock(&rcu_registry_lock); |
c1be8fb9 MD |
596 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
597 | if (ret) | |
598 | abort(); | |
4477a870 | 599 | urcu_bp_exit(); |
c1be8fb9 MD |
600 | } |
601 | ||
602 | /* | |
603 | * Remove thread from the registry when it exits, and flag it as | |
604 | * destroyed so garbage collection can take care of it. | |
605 | */ | |
606 | static | |
607 | void urcu_bp_thread_exit_notifier(void *rcu_key) | |
608 | { | |
4477a870 | 609 | urcu_bp_unregister(rcu_key); |
c1be8fb9 MD |
610 | } |
611 | ||
d8d9a340 MD |
612 | #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER |
613 | static | |
4477a870 | 614 | void urcu_bp_sys_membarrier_status(bool available) |
d8d9a340 MD |
615 | { |
616 | if (!available) | |
617 | abort(); | |
618 | } | |
619 | #else | |
620 | static | |
4477a870 | 621 | void urcu_bp_sys_membarrier_status(bool available) |
d8d9a340 | 622 | { |
3745305b MD |
623 | if (!available) |
624 | return; | |
625 | urcu_bp_has_sys_membarrier = 1; | |
d8d9a340 MD |
626 | } |
627 | #endif | |
628 | ||
3745305b | 629 | static |
4477a870 | 630 | void urcu_bp_sys_membarrier_init(void) |
3745305b MD |
631 | { |
632 | bool available = false; | |
633 | int mask; | |
634 | ||
635 | mask = membarrier(MEMBARRIER_CMD_QUERY, 0); | |
636 | if (mask >= 0) { | |
637 | if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) { | |
638 | if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0)) | |
639 | urcu_die(errno); | |
640 | available = true; | |
641 | } | |
642 | } | |
4477a870 | 643 | urcu_bp_sys_membarrier_status(available); |
3745305b MD |
644 | } |
645 | ||
c1be8fb9 | 646 | static |
4477a870 | 647 | void _urcu_bp_init(void) |
c1be8fb9 MD |
648 | { |
649 | mutex_lock(&init_lock); | |
4477a870 | 650 | if (!urcu_bp_refcount++) { |
c1be8fb9 MD |
651 | int ret; |
652 | ||
653 | ret = pthread_key_create(&urcu_bp_key, | |
654 | urcu_bp_thread_exit_notifier); | |
655 | if (ret) | |
656 | abort(); | |
4477a870 | 657 | urcu_bp_sys_membarrier_init(); |
c1be8fb9 MD |
658 | initialized = 1; |
659 | } | |
660 | mutex_unlock(&init_lock); | |
fdee2e6d MD |
661 | } |
662 | ||
c1be8fb9 | 663 | static |
4477a870 | 664 | void urcu_bp_exit(void) |
fdee2e6d | 665 | { |
457eeeee MD |
666 | urcu_call_rcu_exit(); |
667 | ||
76d6a951 | 668 | mutex_lock(&init_lock); |
4477a870 | 669 | if (!--urcu_bp_refcount) { |
76d6a951 MD |
670 | struct registry_chunk *chunk, *tmp; |
671 | int ret; | |
95b94246 | 672 | |
76d6a951 MD |
673 | cds_list_for_each_entry_safe(chunk, tmp, |
674 | ®istry_arena.chunk_list, node) { | |
5592d049 | 675 | munmap((void *) chunk, chunk->data_len |
76d6a951 MD |
676 | + sizeof(struct registry_chunk)); |
677 | } | |
7937ae1c | 678 | CDS_INIT_LIST_HEAD(®istry_arena.chunk_list); |
76d6a951 MD |
679 | ret = pthread_key_delete(urcu_bp_key); |
680 | if (ret) | |
681 | abort(); | |
95b94246 | 682 | } |
76d6a951 | 683 | mutex_unlock(&init_lock); |
fdee2e6d | 684 | } |
4cf1675f MD |
685 | |
686 | /* | |
731ccb96 MD |
687 | * Holding the rcu_gp_lock and rcu_registry_lock across fork will make |
688 | * sure we fork() don't race with a concurrent thread executing with | |
689 | * any of those locks held. This ensures that the registry and data | |
690 | * protected by rcu_gp_lock are in a coherent state in the child. | |
4cf1675f | 691 | */ |
4477a870 | 692 | void urcu_bp_before_fork(void) |
4cf1675f MD |
693 | { |
694 | sigset_t newmask, oldmask; | |
695 | int ret; | |
696 | ||
6ed4b2e6 | 697 | ret = sigfillset(&newmask); |
4cf1675f | 698 | assert(!ret); |
6ed4b2e6 | 699 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
4cf1675f MD |
700 | assert(!ret); |
701 | mutex_lock(&rcu_gp_lock); | |
731ccb96 | 702 | mutex_lock(&rcu_registry_lock); |
4cf1675f MD |
703 | saved_fork_signal_mask = oldmask; |
704 | } | |
ce28e67a | 705 | URCU_ATTR_ALIAS("urcu_bp_before_fork") void rcu_bp_before_fork(); |
4cf1675f | 706 | |
4477a870 | 707 | void urcu_bp_after_fork_parent(void) |
4cf1675f MD |
708 | { |
709 | sigset_t oldmask; | |
710 | int ret; | |
711 | ||
712 | oldmask = saved_fork_signal_mask; | |
731ccb96 | 713 | mutex_unlock(&rcu_registry_lock); |
4cf1675f MD |
714 | mutex_unlock(&rcu_gp_lock); |
715 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
716 | assert(!ret); | |
717 | } | |
ce28e67a | 718 | URCU_ATTR_ALIAS("urcu_bp_after_fork_parent") |
4477a870 | 719 | void rcu_bp_after_fork_parent(void); |
4cf1675f | 720 | |
c1be8fb9 MD |
721 | /* |
722 | * Prune all entries from registry except our own thread. Fits the Linux | |
731ccb96 | 723 | * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held. |
c1be8fb9 MD |
724 | */ |
725 | static | |
726 | void urcu_bp_prune_registry(void) | |
727 | { | |
728 | struct registry_chunk *chunk; | |
4477a870 | 729 | struct urcu_bp_reader *rcu_reader_reg; |
c1be8fb9 MD |
730 | |
731 | cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) { | |
4477a870 MD |
732 | for (rcu_reader_reg = (struct urcu_bp_reader *) &chunk->data[0]; |
733 | rcu_reader_reg < (struct urcu_bp_reader *) &chunk->data[chunk->data_len]; | |
c1be8fb9 MD |
734 | rcu_reader_reg++) { |
735 | if (!rcu_reader_reg->alloc) | |
736 | continue; | |
737 | if (rcu_reader_reg->tid == pthread_self()) | |
738 | continue; | |
739 | cleanup_thread(chunk, rcu_reader_reg); | |
740 | } | |
741 | } | |
742 | } | |
743 | ||
4477a870 | 744 | void urcu_bp_after_fork_child(void) |
4cf1675f MD |
745 | { |
746 | sigset_t oldmask; | |
747 | int ret; | |
748 | ||
c1be8fb9 | 749 | urcu_bp_prune_registry(); |
4cf1675f | 750 | oldmask = saved_fork_signal_mask; |
731ccb96 | 751 | mutex_unlock(&rcu_registry_lock); |
4cf1675f MD |
752 | mutex_unlock(&rcu_gp_lock); |
753 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
754 | assert(!ret); | |
755 | } | |
ce28e67a | 756 | URCU_ATTR_ALIAS("urcu_bp_after_fork_child") |
4477a870 | 757 | void rcu_bp_after_fork_child(void); |
5e77fc1f | 758 | |
4477a870 | 759 | void *urcu_bp_dereference_sym(void *p) |
9b7981bb MD |
760 | { |
761 | return _rcu_dereference(p); | |
762 | } | |
ce28e67a | 763 | URCU_ATTR_ALIAS("urcu_bp_dereference_sym") |
4477a870 | 764 | void *rcu_dereference_sym_bp(); |
9b7981bb | 765 | |
4477a870 | 766 | void *urcu_bp_set_pointer_sym(void **p, void *v) |
5efd3cd2 MD |
767 | { |
768 | cmm_wmb(); | |
424d4ed5 MD |
769 | uatomic_set(p, v); |
770 | return v; | |
5efd3cd2 | 771 | } |
ce28e67a | 772 | URCU_ATTR_ALIAS("urcu_bp_set_pointer_sym") |
4477a870 | 773 | void *rcu_set_pointer_sym_bp(); |
5efd3cd2 | 774 | |
4477a870 | 775 | void *urcu_bp_xchg_pointer_sym(void **p, void *v) |
5efd3cd2 MD |
776 | { |
777 | cmm_wmb(); | |
778 | return uatomic_xchg(p, v); | |
779 | } | |
ce28e67a | 780 | URCU_ATTR_ALIAS("urcu_bp_xchg_pointer_sym") |
4477a870 | 781 | void *rcu_xchg_pointer_sym_bp(); |
5efd3cd2 | 782 | |
4477a870 | 783 | void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new) |
5efd3cd2 MD |
784 | { |
785 | cmm_wmb(); | |
786 | return uatomic_cmpxchg(p, old, _new); | |
787 | } | |
ce28e67a | 788 | URCU_ATTR_ALIAS("urcu_bp_cmpxchg_pointer_sym") |
4477a870 | 789 | void *rcu_cmpxchg_pointer_sym_bp(); |
5efd3cd2 | 790 | |
5e6b23a6 | 791 | DEFINE_RCU_FLAVOR(rcu_flavor); |
4477a870 | 792 | DEFINE_RCU_FLAVOR_ALIAS(rcu_flavor, alias_rcu_flavor); |
541d828d | 793 | |
5e77fc1f | 794 | #include "urcu-call-rcu-impl.h" |
0376e7b2 | 795 | #include "urcu-defer-impl.h" |