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 | ||
71c811bf | 26 | #define _LGPL_SOURCE |
fdee2e6d MD |
27 | #include <stdio.h> |
28 | #include <pthread.h> | |
29 | #include <signal.h> | |
30 | #include <assert.h> | |
31 | #include <stdlib.h> | |
32 | #include <string.h> | |
33 | #include <errno.h> | |
34 | #include <poll.h> | |
35 | #include <unistd.h> | |
3745305b | 36 | #include <stdbool.h> |
fdee2e6d MD |
37 | #include <sys/mman.h> |
38 | ||
4477a870 MD |
39 | #include <urcu/arch.h> |
40 | #include <urcu/wfcqueue.h> | |
41 | #include <urcu/map/urcu-bp.h> | |
42 | #include <urcu/static/urcu-bp.h> | |
43 | #include <urcu/pointer.h> | |
44 | #include <urcu/tls-compat.h> | |
71c811bf | 45 | |
4a6d7378 | 46 | #include "urcu-die.h" |
ce28e67a | 47 | #include "urcu-utils.h" |
4a6d7378 | 48 | |
4477a870 | 49 | #define URCU_API_MAP |
fdee2e6d | 50 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
71c811bf | 51 | #undef _LGPL_SOURCE |
4477a870 | 52 | #include <urcu/urcu-bp.h> |
71c811bf | 53 | #define _LGPL_SOURCE |
fdee2e6d | 54 | |
4c1ae2ea MD |
55 | #ifndef MAP_ANONYMOUS |
56 | #define MAP_ANONYMOUS MAP_ANON | |
57 | #endif | |
58 | ||
c7eaf61c MD |
59 | #ifdef __linux__ |
60 | static | |
61 | void *mremap_wrapper(void *old_address, size_t old_size, | |
62 | size_t new_size, int flags) | |
63 | { | |
64 | return mremap(old_address, old_size, new_size, flags); | |
65 | } | |
66 | #else | |
45a4872f MD |
67 | |
68 | #define MREMAP_MAYMOVE 1 | |
69 | #define MREMAP_FIXED 2 | |
70 | ||
71 | /* | |
95b94246 | 72 | * mremap wrapper for non-Linux systems not allowing MAYMOVE. |
45a4872f MD |
73 | * This is not generic. |
74 | */ | |
c7eaf61c MD |
75 | static |
76 | void *mremap_wrapper(void *old_address, size_t old_size, | |
77 | size_t new_size, int flags) | |
45a4872f | 78 | { |
95b94246 MD |
79 | assert(!(flags & MREMAP_MAYMOVE)); |
80 | ||
81 | return MAP_FAILED; | |
45a4872f MD |
82 | } |
83 | #endif | |
84 | ||
9340c38d MD |
85 | /* Sleep delay in ms */ |
86 | #define RCU_SLEEP_DELAY_MS 10 | |
95b94246 MD |
87 | #define INIT_NR_THREADS 8 |
88 | #define ARENA_INIT_ALLOC \ | |
89 | sizeof(struct registry_chunk) \ | |
4477a870 | 90 | + INIT_NR_THREADS * sizeof(struct urcu_bp_reader) |
fdee2e6d | 91 | |
b7b6a8f5 PB |
92 | /* |
93 | * Active attempts to check for reader Q.S. before calling sleep(). | |
94 | */ | |
95 | #define RCU_QS_ACTIVE_ATTEMPTS 100 | |
96 | ||
76d6a951 | 97 | static |
4477a870 | 98 | int urcu_bp_refcount; |
76d6a951 | 99 | |
999991c6 MD |
100 | /* If the headers do not support membarrier system call, fall back smp_mb. */ |
101 | #ifdef __NR_membarrier | |
102 | # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__) | |
f541831e MD |
103 | #else |
104 | # define membarrier(...) -ENOSYS | |
105 | #endif | |
106 | ||
107 | enum membarrier_cmd { | |
3745305b MD |
108 | MEMBARRIER_CMD_QUERY = 0, |
109 | MEMBARRIER_CMD_SHARED = (1 << 0), | |
110 | /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */ | |
111 | /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */ | |
112 | MEMBARRIER_CMD_PRIVATE_EXPEDITED = (1 << 3), | |
113 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED = (1 << 4), | |
f541831e MD |
114 | }; |
115 | ||
c1be8fb9 | 116 | static |
4477a870 | 117 | void __attribute__((constructor)) _urcu_bp_init(void); |
c1be8fb9 | 118 | static |
4477a870 | 119 | void __attribute__((destructor)) urcu_bp_exit(void); |
fdee2e6d | 120 | |
d8d9a340 | 121 | #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER |
f541831e | 122 | int urcu_bp_has_sys_membarrier; |
d8d9a340 | 123 | #endif |
f541831e | 124 | |
731ccb96 MD |
125 | /* |
126 | * rcu_gp_lock ensures mutual exclusion between threads calling | |
127 | * synchronize_rcu(). | |
128 | */ | |
6abb4bd5 | 129 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
731ccb96 MD |
130 | /* |
131 | * rcu_registry_lock ensures mutual exclusion between threads | |
132 | * registering and unregistering themselves to/from the registry, and | |
133 | * with threads reading that registry from synchronize_rcu(). However, | |
134 | * this lock is not held all the way through the completion of awaiting | |
135 | * for the grace period. It is sporadically released between iterations | |
136 | * on the registry. | |
137 | * rcu_registry_lock may nest inside rcu_gp_lock. | |
138 | */ | |
139 | static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER; | |
fdee2e6d | 140 | |
c1be8fb9 MD |
141 | static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER; |
142 | static int initialized; | |
143 | ||
144 | static pthread_key_t urcu_bp_key; | |
145 | ||
4477a870 | 146 | struct urcu_bp_gp urcu_bp_gp = { .ctr = URCU_BP_GP_COUNT }; |
ce28e67a | 147 | URCU_ATTR_ALIAS("urcu_bp_gp") extern struct urcu_bp_gp rcu_gp_bp; |
fdee2e6d MD |
148 | |
149 | /* | |
150 | * Pointer to registry elements. Written to only by each individual reader. Read | |
151 | * by both the reader and the writers. | |
152 | */ | |
4477a870 | 153 | DEFINE_URCU_TLS(struct urcu_bp_reader *, urcu_bp_reader); |
99bfa9e9 | 154 | DEFINE_URCU_TLS_ALIAS(struct urcu_bp_reader *, urcu_bp_reader, rcu_reader_bp); |
fdee2e6d | 155 | |
16aa9ee8 | 156 | static CDS_LIST_HEAD(registry); |
fdee2e6d | 157 | |
95b94246 MD |
158 | struct registry_chunk { |
159 | size_t data_len; /* data length */ | |
c1be8fb9 | 160 | size_t used; /* amount of data used */ |
95b94246 MD |
161 | struct cds_list_head node; /* chunk_list node */ |
162 | char data[]; | |
163 | }; | |
164 | ||
fdee2e6d | 165 | struct registry_arena { |
95b94246 | 166 | struct cds_list_head chunk_list; |
fdee2e6d MD |
167 | }; |
168 | ||
95b94246 MD |
169 | static struct registry_arena registry_arena = { |
170 | .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list), | |
171 | }; | |
fdee2e6d | 172 | |
4cf1675f MD |
173 | /* Saved fork signal mask, protected by rcu_gp_lock */ |
174 | static sigset_t saved_fork_signal_mask; | |
175 | ||
6abb4bd5 | 176 | static void mutex_lock(pthread_mutex_t *mutex) |
fdee2e6d MD |
177 | { |
178 | int ret; | |
179 | ||
180 | #ifndef DISTRUST_SIGNALS_EXTREME | |
6abb4bd5 | 181 | ret = pthread_mutex_lock(mutex); |
4a6d7378 MD |
182 | if (ret) |
183 | urcu_die(ret); | |
fdee2e6d | 184 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ |
6abb4bd5 | 185 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
4a6d7378 MD |
186 | if (ret != EBUSY && ret != EINTR) |
187 | urcu_die(ret); | |
fdee2e6d MD |
188 | poll(NULL,0,10); |
189 | } | |
190 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
191 | } | |
192 | ||
6abb4bd5 | 193 | static void mutex_unlock(pthread_mutex_t *mutex) |
fdee2e6d MD |
194 | { |
195 | int ret; | |
196 | ||
6abb4bd5 | 197 | ret = pthread_mutex_unlock(mutex); |
4a6d7378 MD |
198 | if (ret) |
199 | urcu_die(ret); | |
fdee2e6d MD |
200 | } |
201 | ||
f541831e MD |
202 | static void smp_mb_master(void) |
203 | { | |
3745305b MD |
204 | if (caa_likely(urcu_bp_has_sys_membarrier)) { |
205 | if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0)) | |
206 | urcu_die(errno); | |
207 | } else { | |
f541831e | 208 | cmm_smp_mb(); |
3745305b | 209 | } |
f541831e MD |
210 | } |
211 | ||
731ccb96 MD |
212 | /* |
213 | * Always called with rcu_registry lock held. Releases this lock between | |
214 | * iterations and grabs it again. Holds the lock when it returns. | |
215 | */ | |
52c75091 MD |
216 | static void wait_for_readers(struct cds_list_head *input_readers, |
217 | struct cds_list_head *cur_snap_readers, | |
218 | struct cds_list_head *qsreaders) | |
fdee2e6d | 219 | { |
9340c38d | 220 | unsigned int wait_loops = 0; |
4477a870 | 221 | struct urcu_bp_reader *index, *tmp; |
fdee2e6d | 222 | |
fdee2e6d | 223 | /* |
4477a870 | 224 | * Wait for each thread URCU_TLS(urcu_bp_reader).ctr to either |
dd61d077 | 225 | * indicate quiescence (not nested), or observe the current |
c13c2e55 | 226 | * rcu_gp.ctr value. |
fdee2e6d MD |
227 | */ |
228 | for (;;) { | |
9340c38d MD |
229 | if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS) |
230 | wait_loops++; | |
231 | ||
52c75091 | 232 | cds_list_for_each_entry_safe(index, tmp, input_readers, node) { |
4477a870 MD |
233 | switch (urcu_bp_reader_state(&index->ctr)) { |
234 | case URCU_BP_READER_ACTIVE_CURRENT: | |
52c75091 MD |
235 | if (cur_snap_readers) { |
236 | cds_list_move(&index->node, | |
237 | cur_snap_readers); | |
238 | break; | |
239 | } | |
240 | /* Fall-through */ | |
4477a870 | 241 | case URCU_BP_READER_INACTIVE: |
52c75091 MD |
242 | cds_list_move(&index->node, qsreaders); |
243 | break; | |
4477a870 | 244 | case URCU_BP_READER_ACTIVE_OLD: |
52c75091 MD |
245 | /* |
246 | * Old snapshot. Leaving node in | |
247 | * input_readers will make us busy-loop | |
248 | * until the snapshot becomes current or | |
249 | * the reader becomes inactive. | |
250 | */ | |
251 | break; | |
252 | } | |
fdee2e6d MD |
253 | } |
254 | ||
52c75091 | 255 | if (cds_list_empty(input_readers)) { |
fdee2e6d MD |
256 | break; |
257 | } else { | |
731ccb96 MD |
258 | /* Temporarily unlock the registry lock. */ |
259 | mutex_unlock(&rcu_registry_lock); | |
9340c38d MD |
260 | if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) |
261 | (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS); | |
fdee2e6d | 262 | else |
06f22bdb | 263 | caa_cpu_relax(); |
731ccb96 MD |
264 | /* Re-lock the registry lock before the next loop. */ |
265 | mutex_lock(&rcu_registry_lock); | |
fdee2e6d MD |
266 | } |
267 | } | |
fdee2e6d MD |
268 | } |
269 | ||
4477a870 | 270 | void urcu_bp_synchronize_rcu(void) |
fdee2e6d | 271 | { |
52c75091 MD |
272 | CDS_LIST_HEAD(cur_snap_readers); |
273 | CDS_LIST_HEAD(qsreaders); | |
fdee2e6d MD |
274 | sigset_t newmask, oldmask; |
275 | int ret; | |
276 | ||
6ed4b2e6 | 277 | ret = sigfillset(&newmask); |
fdee2e6d | 278 | assert(!ret); |
6ed4b2e6 | 279 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
fdee2e6d MD |
280 | assert(!ret); |
281 | ||
6abb4bd5 | 282 | mutex_lock(&rcu_gp_lock); |
fdee2e6d | 283 | |
731ccb96 MD |
284 | mutex_lock(&rcu_registry_lock); |
285 | ||
16aa9ee8 | 286 | if (cds_list_empty(®istry)) |
2dfb8b5e | 287 | goto out; |
fdee2e6d MD |
288 | |
289 | /* All threads should read qparity before accessing data structure | |
2dfb8b5e | 290 | * where new ptr points to. */ |
fdee2e6d | 291 | /* Write new ptr before changing the qparity */ |
f541831e | 292 | smp_mb_master(); |
fdee2e6d | 293 | |
fdee2e6d | 294 | /* |
dd61d077 | 295 | * Wait for readers to observe original parity or be quiescent. |
731ccb96 MD |
296 | * wait_for_readers() can release and grab again rcu_registry_lock |
297 | * interally. | |
dd61d077 | 298 | */ |
52c75091 | 299 | wait_for_readers(®istry, &cur_snap_readers, &qsreaders); |
dd61d077 MD |
300 | |
301 | /* | |
302 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the | |
303 | * model easier to understand. It does not have a big performance impact | |
304 | * anyway, given this is the write-side. | |
305 | */ | |
306 | cmm_smp_mb(); | |
307 | ||
308 | /* Switch parity: 0 -> 1, 1 -> 0 */ | |
4477a870 | 309 | CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_BP_GP_CTR_PHASE); |
dd61d077 MD |
310 | |
311 | /* | |
312 | * Must commit qparity update to memory before waiting for other parity | |
313 | * quiescent state. Failure to do so could result in the writer waiting | |
314 | * forever while new readers are always accessing data (no progress). | |
315 | * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED. | |
fdee2e6d | 316 | */ |
fdee2e6d MD |
317 | |
318 | /* | |
5481ddb3 | 319 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
fdee2e6d MD |
320 | * model easier to understand. It does not have a big performance impact |
321 | * anyway, given this is the write-side. | |
322 | */ | |
5481ddb3 | 323 | cmm_smp_mb(); |
fdee2e6d | 324 | |
fdee2e6d | 325 | /* |
dd61d077 | 326 | * Wait for readers to observe new parity or be quiescent. |
731ccb96 MD |
327 | * wait_for_readers() can release and grab again rcu_registry_lock |
328 | * interally. | |
fdee2e6d | 329 | */ |
52c75091 MD |
330 | wait_for_readers(&cur_snap_readers, NULL, &qsreaders); |
331 | ||
332 | /* | |
333 | * Put quiescent reader list back into registry. | |
334 | */ | |
335 | cds_list_splice(&qsreaders, ®istry); | |
fdee2e6d MD |
336 | |
337 | /* | |
2dfb8b5e MD |
338 | * Finish waiting for reader threads before letting the old ptr being |
339 | * freed. | |
fdee2e6d | 340 | */ |
f541831e | 341 | smp_mb_master(); |
2dfb8b5e | 342 | out: |
731ccb96 | 343 | mutex_unlock(&rcu_registry_lock); |
6abb4bd5 | 344 | mutex_unlock(&rcu_gp_lock); |
fdee2e6d MD |
345 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
346 | assert(!ret); | |
347 | } | |
ce28e67a | 348 | URCU_ATTR_ALIAS("urcu_bp_synchronize_rcu") void synchronize_rcu_bp(); |
fdee2e6d MD |
349 | |
350 | /* | |
351 | * library wrappers to be used by non-LGPL compatible source code. | |
352 | */ | |
353 | ||
4477a870 | 354 | void urcu_bp_read_lock(void) |
fdee2e6d | 355 | { |
4477a870 | 356 | _urcu_bp_read_lock(); |
fdee2e6d | 357 | } |
ce28e67a | 358 | URCU_ATTR_ALIAS("urcu_bp_read_lock") void rcu_read_lock_bp(); |
fdee2e6d | 359 | |
4477a870 | 360 | void urcu_bp_read_unlock(void) |
fdee2e6d | 361 | { |
4477a870 | 362 | _urcu_bp_read_unlock(); |
fdee2e6d | 363 | } |
ce28e67a | 364 | URCU_ATTR_ALIAS("urcu_bp_read_unlock") void rcu_read_unlock_bp(); |
fdee2e6d | 365 | |
4477a870 | 366 | int urcu_bp_read_ongoing(void) |
882f3357 | 367 | { |
4477a870 | 368 | return _urcu_bp_read_ongoing(); |
882f3357 | 369 | } |
ce28e67a | 370 | URCU_ATTR_ALIAS("urcu_bp_read_ongoing") int rcu_read_ongoing_bp(); |
882f3357 | 371 | |
fdee2e6d | 372 | /* |
95b94246 MD |
373 | * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk. |
374 | * Else, try expanding the last chunk. If this fails, allocate a new | |
375 | * chunk twice as big as the last chunk. | |
376 | * Memory used by chunks _never_ moves. A chunk could theoretically be | |
377 | * freed when all "used" slots are released, but we don't do it at this | |
378 | * point. | |
fdee2e6d | 379 | */ |
95b94246 MD |
380 | static |
381 | void expand_arena(struct registry_arena *arena) | |
fdee2e6d | 382 | { |
95b94246 MD |
383 | struct registry_chunk *new_chunk, *last_chunk; |
384 | size_t old_chunk_len, new_chunk_len; | |
385 | ||
386 | /* No chunk. */ | |
387 | if (cds_list_empty(&arena->chunk_list)) { | |
388 | assert(ARENA_INIT_ALLOC >= | |
389 | sizeof(struct registry_chunk) | |
390 | + sizeof(struct rcu_reader)); | |
391 | new_chunk_len = ARENA_INIT_ALLOC; | |
5592d049 MJ |
392 | new_chunk = (struct registry_chunk *) mmap(NULL, |
393 | new_chunk_len, | |
9d8612b7 MD |
394 | PROT_READ | PROT_WRITE, |
395 | MAP_ANONYMOUS | MAP_PRIVATE, | |
396 | -1, 0); | |
95b94246 MD |
397 | if (new_chunk == MAP_FAILED) |
398 | abort(); | |
d3ac5bb7 | 399 | memset(new_chunk, 0, new_chunk_len); |
95b94246 MD |
400 | new_chunk->data_len = |
401 | new_chunk_len - sizeof(struct registry_chunk); | |
402 | cds_list_add_tail(&new_chunk->node, &arena->chunk_list); | |
403 | return; /* We're done. */ | |
404 | } | |
9d8612b7 | 405 | |
95b94246 MD |
406 | /* Try expanding last chunk. */ |
407 | last_chunk = cds_list_entry(arena->chunk_list.prev, | |
408 | struct registry_chunk, node); | |
409 | old_chunk_len = | |
410 | last_chunk->data_len + sizeof(struct registry_chunk); | |
411 | new_chunk_len = old_chunk_len << 1; | |
412 | ||
413 | /* Don't allow memory mapping to move, just expand. */ | |
414 | new_chunk = mremap_wrapper(last_chunk, old_chunk_len, | |
415 | new_chunk_len, 0); | |
416 | if (new_chunk != MAP_FAILED) { | |
417 | /* Should not have moved. */ | |
418 | assert(new_chunk == last_chunk); | |
d3ac5bb7 | 419 | memset((char *) last_chunk + old_chunk_len, 0, |
95b94246 MD |
420 | new_chunk_len - old_chunk_len); |
421 | last_chunk->data_len = | |
422 | new_chunk_len - sizeof(struct registry_chunk); | |
423 | return; /* We're done. */ | |
424 | } | |
0617bf4c | 425 | |
95b94246 | 426 | /* Remap did not succeed, we need to add a new chunk. */ |
5592d049 MJ |
427 | new_chunk = (struct registry_chunk *) mmap(NULL, |
428 | new_chunk_len, | |
95b94246 MD |
429 | PROT_READ | PROT_WRITE, |
430 | MAP_ANONYMOUS | MAP_PRIVATE, | |
431 | -1, 0); | |
432 | if (new_chunk == MAP_FAILED) | |
433 | abort(); | |
d3ac5bb7 | 434 | memset(new_chunk, 0, new_chunk_len); |
95b94246 MD |
435 | new_chunk->data_len = |
436 | new_chunk_len - sizeof(struct registry_chunk); | |
437 | cds_list_add_tail(&new_chunk->node, &arena->chunk_list); | |
438 | } | |
fdee2e6d | 439 | |
95b94246 MD |
440 | static |
441 | struct rcu_reader *arena_alloc(struct registry_arena *arena) | |
442 | { | |
443 | struct registry_chunk *chunk; | |
444 | struct rcu_reader *rcu_reader_reg; | |
445 | int expand_done = 0; /* Only allow to expand once per alloc */ | |
446 | size_t len = sizeof(struct rcu_reader); | |
447 | ||
448 | retry: | |
449 | cds_list_for_each_entry(chunk, &arena->chunk_list, node) { | |
450 | if (chunk->data_len - chunk->used < len) | |
451 | continue; | |
452 | /* Find spot */ | |
453 | for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0]; | |
454 | rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len]; | |
455 | rcu_reader_reg++) { | |
456 | if (!rcu_reader_reg->alloc) { | |
457 | rcu_reader_reg->alloc = 1; | |
458 | chunk->used += len; | |
459 | return rcu_reader_reg; | |
460 | } | |
461 | } | |
462 | } | |
463 | ||
464 | if (!expand_done) { | |
465 | expand_arena(arena); | |
466 | expand_done = 1; | |
467 | goto retry; | |
468 | } | |
469 | ||
470 | return NULL; | |
fdee2e6d MD |
471 | } |
472 | ||
473 | /* Called with signals off and mutex locked */ | |
95b94246 MD |
474 | static |
475 | void add_thread(void) | |
fdee2e6d | 476 | { |
02be5561 | 477 | struct rcu_reader *rcu_reader_reg; |
c1be8fb9 | 478 | int ret; |
fdee2e6d | 479 | |
95b94246 MD |
480 | rcu_reader_reg = arena_alloc(®istry_arena); |
481 | if (!rcu_reader_reg) | |
482 | abort(); | |
c1be8fb9 MD |
483 | ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg); |
484 | if (ret) | |
485 | abort(); | |
fdee2e6d MD |
486 | |
487 | /* Add to registry */ | |
02be5561 MD |
488 | rcu_reader_reg->tid = pthread_self(); |
489 | assert(rcu_reader_reg->ctr == 0); | |
16aa9ee8 | 490 | cds_list_add(&rcu_reader_reg->node, ®istry); |
95b94246 MD |
491 | /* |
492 | * Reader threads are pointing to the reader registry. This is | |
493 | * why its memory should never be relocated. | |
494 | */ | |
4477a870 | 495 | URCU_TLS(urcu_bp_reader) = rcu_reader_reg; |
fdee2e6d MD |
496 | } |
497 | ||
c1be8fb9 MD |
498 | /* Called with mutex locked */ |
499 | static | |
500 | void cleanup_thread(struct registry_chunk *chunk, | |
501 | struct rcu_reader *rcu_reader_reg) | |
502 | { | |
503 | rcu_reader_reg->ctr = 0; | |
504 | cds_list_del(&rcu_reader_reg->node); | |
505 | rcu_reader_reg->tid = 0; | |
506 | rcu_reader_reg->alloc = 0; | |
507 | chunk->used -= sizeof(struct rcu_reader); | |
508 | } | |
509 | ||
510 | static | |
511 | struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg) | |
fdee2e6d | 512 | { |
95b94246 | 513 | struct registry_chunk *chunk; |
fdee2e6d | 514 | |
95b94246 | 515 | cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) { |
c1be8fb9 MD |
516 | if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0]) |
517 | continue; | |
518 | if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len]) | |
519 | continue; | |
520 | return chunk; | |
521 | } | |
522 | return NULL; | |
523 | } | |
95b94246 | 524 | |
c1be8fb9 MD |
525 | /* Called with signals off and mutex locked */ |
526 | static | |
76d6a951 | 527 | void remove_thread(struct rcu_reader *rcu_reader_reg) |
c1be8fb9 | 528 | { |
c1be8fb9 | 529 | cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg); |
4477a870 | 530 | URCU_TLS(urcu_bp_reader) = NULL; |
fdee2e6d MD |
531 | } |
532 | ||
533 | /* Disable signals, take mutex, add to registry */ | |
4477a870 | 534 | void urcu_bp_register(void) |
fdee2e6d MD |
535 | { |
536 | sigset_t newmask, oldmask; | |
537 | int ret; | |
538 | ||
6ed4b2e6 | 539 | ret = sigfillset(&newmask); |
c1be8fb9 MD |
540 | if (ret) |
541 | abort(); | |
6ed4b2e6 | 542 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
c1be8fb9 MD |
543 | if (ret) |
544 | abort(); | |
fdee2e6d MD |
545 | |
546 | /* | |
547 | * Check if a signal concurrently registered our thread since | |
c1be8fb9 MD |
548 | * the check in rcu_read_lock(). |
549 | */ | |
4477a870 | 550 | if (URCU_TLS(urcu_bp_reader)) |
fdee2e6d MD |
551 | goto end; |
552 | ||
c1be8fb9 MD |
553 | /* |
554 | * Take care of early registration before urcu_bp constructor. | |
555 | */ | |
4477a870 | 556 | _urcu_bp_init(); |
c1be8fb9 | 557 | |
731ccb96 | 558 | mutex_lock(&rcu_registry_lock); |
fdee2e6d | 559 | add_thread(); |
731ccb96 | 560 | mutex_unlock(&rcu_registry_lock); |
fdee2e6d MD |
561 | end: |
562 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
c1be8fb9 MD |
563 | if (ret) |
564 | abort(); | |
565 | } | |
ce28e67a | 566 | URCU_ATTR_ALIAS("urcu_bp_register") void rcu_bp_register(); |
c1be8fb9 MD |
567 | |
568 | /* Disable signals, take mutex, remove from registry */ | |
569 | static | |
4477a870 | 570 | void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg) |
c1be8fb9 MD |
571 | { |
572 | sigset_t newmask, oldmask; | |
573 | int ret; | |
574 | ||
575 | ret = sigfillset(&newmask); | |
576 | if (ret) | |
577 | abort(); | |
578 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
579 | if (ret) | |
580 | abort(); | |
581 | ||
731ccb96 | 582 | mutex_lock(&rcu_registry_lock); |
76d6a951 | 583 | remove_thread(rcu_reader_reg); |
731ccb96 | 584 | mutex_unlock(&rcu_registry_lock); |
c1be8fb9 MD |
585 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
586 | if (ret) | |
587 | abort(); | |
4477a870 | 588 | urcu_bp_exit(); |
c1be8fb9 MD |
589 | } |
590 | ||
591 | /* | |
592 | * Remove thread from the registry when it exits, and flag it as | |
593 | * destroyed so garbage collection can take care of it. | |
594 | */ | |
595 | static | |
596 | void urcu_bp_thread_exit_notifier(void *rcu_key) | |
597 | { | |
4477a870 | 598 | urcu_bp_unregister(rcu_key); |
c1be8fb9 MD |
599 | } |
600 | ||
d8d9a340 MD |
601 | #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER |
602 | static | |
4477a870 | 603 | void urcu_bp_sys_membarrier_status(bool available) |
d8d9a340 MD |
604 | { |
605 | if (!available) | |
606 | abort(); | |
607 | } | |
608 | #else | |
609 | static | |
4477a870 | 610 | void urcu_bp_sys_membarrier_status(bool available) |
d8d9a340 | 611 | { |
3745305b MD |
612 | if (!available) |
613 | return; | |
614 | urcu_bp_has_sys_membarrier = 1; | |
d8d9a340 MD |
615 | } |
616 | #endif | |
617 | ||
3745305b | 618 | static |
4477a870 | 619 | void urcu_bp_sys_membarrier_init(void) |
3745305b MD |
620 | { |
621 | bool available = false; | |
622 | int mask; | |
623 | ||
624 | mask = membarrier(MEMBARRIER_CMD_QUERY, 0); | |
625 | if (mask >= 0) { | |
626 | if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) { | |
627 | if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0)) | |
628 | urcu_die(errno); | |
629 | available = true; | |
630 | } | |
631 | } | |
4477a870 | 632 | urcu_bp_sys_membarrier_status(available); |
3745305b MD |
633 | } |
634 | ||
c1be8fb9 | 635 | static |
4477a870 | 636 | void _urcu_bp_init(void) |
c1be8fb9 MD |
637 | { |
638 | mutex_lock(&init_lock); | |
4477a870 | 639 | if (!urcu_bp_refcount++) { |
c1be8fb9 MD |
640 | int ret; |
641 | ||
642 | ret = pthread_key_create(&urcu_bp_key, | |
643 | urcu_bp_thread_exit_notifier); | |
644 | if (ret) | |
645 | abort(); | |
4477a870 | 646 | urcu_bp_sys_membarrier_init(); |
c1be8fb9 MD |
647 | initialized = 1; |
648 | } | |
649 | mutex_unlock(&init_lock); | |
fdee2e6d MD |
650 | } |
651 | ||
c1be8fb9 | 652 | static |
4477a870 | 653 | void urcu_bp_exit(void) |
fdee2e6d | 654 | { |
76d6a951 | 655 | mutex_lock(&init_lock); |
4477a870 | 656 | if (!--urcu_bp_refcount) { |
76d6a951 MD |
657 | struct registry_chunk *chunk, *tmp; |
658 | int ret; | |
95b94246 | 659 | |
76d6a951 MD |
660 | cds_list_for_each_entry_safe(chunk, tmp, |
661 | ®istry_arena.chunk_list, node) { | |
5592d049 | 662 | munmap((void *) chunk, chunk->data_len |
76d6a951 MD |
663 | + sizeof(struct registry_chunk)); |
664 | } | |
7937ae1c | 665 | CDS_INIT_LIST_HEAD(®istry_arena.chunk_list); |
76d6a951 MD |
666 | ret = pthread_key_delete(urcu_bp_key); |
667 | if (ret) | |
668 | abort(); | |
95b94246 | 669 | } |
76d6a951 | 670 | mutex_unlock(&init_lock); |
fdee2e6d | 671 | } |
4cf1675f MD |
672 | |
673 | /* | |
731ccb96 MD |
674 | * Holding the rcu_gp_lock and rcu_registry_lock across fork will make |
675 | * sure we fork() don't race with a concurrent thread executing with | |
676 | * any of those locks held. This ensures that the registry and data | |
677 | * protected by rcu_gp_lock are in a coherent state in the child. | |
4cf1675f | 678 | */ |
4477a870 | 679 | void urcu_bp_before_fork(void) |
4cf1675f MD |
680 | { |
681 | sigset_t newmask, oldmask; | |
682 | int ret; | |
683 | ||
6ed4b2e6 | 684 | ret = sigfillset(&newmask); |
4cf1675f | 685 | assert(!ret); |
6ed4b2e6 | 686 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
4cf1675f MD |
687 | assert(!ret); |
688 | mutex_lock(&rcu_gp_lock); | |
731ccb96 | 689 | mutex_lock(&rcu_registry_lock); |
4cf1675f MD |
690 | saved_fork_signal_mask = oldmask; |
691 | } | |
ce28e67a | 692 | URCU_ATTR_ALIAS("urcu_bp_before_fork") void rcu_bp_before_fork(); |
4cf1675f | 693 | |
4477a870 | 694 | void urcu_bp_after_fork_parent(void) |
4cf1675f MD |
695 | { |
696 | sigset_t oldmask; | |
697 | int ret; | |
698 | ||
699 | oldmask = saved_fork_signal_mask; | |
731ccb96 | 700 | mutex_unlock(&rcu_registry_lock); |
4cf1675f MD |
701 | mutex_unlock(&rcu_gp_lock); |
702 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
703 | assert(!ret); | |
704 | } | |
ce28e67a | 705 | URCU_ATTR_ALIAS("urcu_bp_after_fork_parent") |
4477a870 | 706 | void rcu_bp_after_fork_parent(void); |
4cf1675f | 707 | |
c1be8fb9 MD |
708 | /* |
709 | * Prune all entries from registry except our own thread. Fits the Linux | |
731ccb96 | 710 | * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held. |
c1be8fb9 MD |
711 | */ |
712 | static | |
713 | void urcu_bp_prune_registry(void) | |
714 | { | |
715 | struct registry_chunk *chunk; | |
4477a870 | 716 | struct urcu_bp_reader *rcu_reader_reg; |
c1be8fb9 MD |
717 | |
718 | cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) { | |
4477a870 MD |
719 | for (rcu_reader_reg = (struct urcu_bp_reader *) &chunk->data[0]; |
720 | rcu_reader_reg < (struct urcu_bp_reader *) &chunk->data[chunk->data_len]; | |
c1be8fb9 MD |
721 | rcu_reader_reg++) { |
722 | if (!rcu_reader_reg->alloc) | |
723 | continue; | |
724 | if (rcu_reader_reg->tid == pthread_self()) | |
725 | continue; | |
726 | cleanup_thread(chunk, rcu_reader_reg); | |
727 | } | |
728 | } | |
729 | } | |
730 | ||
4477a870 | 731 | void urcu_bp_after_fork_child(void) |
4cf1675f MD |
732 | { |
733 | sigset_t oldmask; | |
734 | int ret; | |
735 | ||
c1be8fb9 | 736 | urcu_bp_prune_registry(); |
4cf1675f | 737 | oldmask = saved_fork_signal_mask; |
731ccb96 | 738 | mutex_unlock(&rcu_registry_lock); |
4cf1675f MD |
739 | mutex_unlock(&rcu_gp_lock); |
740 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
741 | assert(!ret); | |
742 | } | |
ce28e67a | 743 | URCU_ATTR_ALIAS("urcu_bp_after_fork_child") |
4477a870 | 744 | void rcu_bp_after_fork_child(void); |
5e77fc1f | 745 | |
4477a870 | 746 | void *urcu_bp_dereference_sym(void *p) |
9b7981bb MD |
747 | { |
748 | return _rcu_dereference(p); | |
749 | } | |
ce28e67a | 750 | URCU_ATTR_ALIAS("urcu_bp_dereference_sym") |
4477a870 | 751 | void *rcu_dereference_sym_bp(); |
9b7981bb | 752 | |
4477a870 | 753 | void *urcu_bp_set_pointer_sym(void **p, void *v) |
5efd3cd2 MD |
754 | { |
755 | cmm_wmb(); | |
424d4ed5 MD |
756 | uatomic_set(p, v); |
757 | return v; | |
5efd3cd2 | 758 | } |
ce28e67a | 759 | URCU_ATTR_ALIAS("urcu_bp_set_pointer_sym") |
4477a870 | 760 | void *rcu_set_pointer_sym_bp(); |
5efd3cd2 | 761 | |
4477a870 | 762 | void *urcu_bp_xchg_pointer_sym(void **p, void *v) |
5efd3cd2 MD |
763 | { |
764 | cmm_wmb(); | |
765 | return uatomic_xchg(p, v); | |
766 | } | |
ce28e67a | 767 | URCU_ATTR_ALIAS("urcu_bp_xchg_pointer_sym") |
4477a870 | 768 | void *rcu_xchg_pointer_sym_bp(); |
5efd3cd2 | 769 | |
4477a870 | 770 | void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new) |
5efd3cd2 MD |
771 | { |
772 | cmm_wmb(); | |
773 | return uatomic_cmpxchg(p, old, _new); | |
774 | } | |
ce28e67a | 775 | URCU_ATTR_ALIAS("urcu_bp_cmpxchg_pointer_sym") |
4477a870 | 776 | void *rcu_cmpxchg_pointer_sym_bp(); |
5efd3cd2 | 777 | |
5e6b23a6 | 778 | DEFINE_RCU_FLAVOR(rcu_flavor); |
4477a870 | 779 | DEFINE_RCU_FLAVOR_ALIAS(rcu_flavor, alias_rcu_flavor); |
541d828d | 780 | |
5e77fc1f | 781 | #include "urcu-call-rcu-impl.h" |
0376e7b2 | 782 | #include "urcu-defer-impl.h" |