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