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 | ||
999991c6 | 39 | #include "urcu/arch.h" |
d73fb81f | 40 | #include "urcu/wfcqueue.h" |
57760d44 | 41 | #include "urcu/map/urcu-bp.h" |
af7c2dbe | 42 | #include "urcu/static/urcu-bp.h" |
618b2595 | 43 | #include "urcu-pointer.h" |
bd252a04 | 44 | #include "urcu/tls-compat.h" |
71c811bf | 45 | |
4a6d7378 MD |
46 | #include "urcu-die.h" |
47 | ||
fdee2e6d | 48 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
71c811bf | 49 | #undef _LGPL_SOURCE |
fdee2e6d | 50 | #include "urcu-bp.h" |
71c811bf | 51 | #define _LGPL_SOURCE |
fdee2e6d | 52 | |
4c1ae2ea MD |
53 | #ifndef MAP_ANONYMOUS |
54 | #define MAP_ANONYMOUS MAP_ANON | |
55 | #endif | |
56 | ||
c7eaf61c MD |
57 | #ifdef __linux__ |
58 | static | |
59 | void *mremap_wrapper(void *old_address, size_t old_size, | |
60 | size_t new_size, int flags) | |
61 | { | |
62 | return mremap(old_address, old_size, new_size, flags); | |
63 | } | |
64 | #else | |
45a4872f MD |
65 | |
66 | #define MREMAP_MAYMOVE 1 | |
67 | #define MREMAP_FIXED 2 | |
68 | ||
69 | /* | |
95b94246 | 70 | * mremap wrapper for non-Linux systems not allowing MAYMOVE. |
45a4872f MD |
71 | * This is not generic. |
72 | */ | |
c7eaf61c MD |
73 | static |
74 | void *mremap_wrapper(void *old_address, size_t old_size, | |
75 | size_t new_size, int flags) | |
45a4872f | 76 | { |
95b94246 MD |
77 | assert(!(flags & MREMAP_MAYMOVE)); |
78 | ||
79 | return MAP_FAILED; | |
45a4872f MD |
80 | } |
81 | #endif | |
82 | ||
9340c38d MD |
83 | /* Sleep delay in ms */ |
84 | #define RCU_SLEEP_DELAY_MS 10 | |
95b94246 MD |
85 | #define INIT_NR_THREADS 8 |
86 | #define ARENA_INIT_ALLOC \ | |
87 | sizeof(struct registry_chunk) \ | |
88 | + INIT_NR_THREADS * sizeof(struct rcu_reader) | |
fdee2e6d | 89 | |
b7b6a8f5 PB |
90 | /* |
91 | * Active attempts to check for reader Q.S. before calling sleep(). | |
92 | */ | |
93 | #define RCU_QS_ACTIVE_ATTEMPTS 100 | |
94 | ||
76d6a951 MD |
95 | static |
96 | int rcu_bp_refcount; | |
97 | ||
999991c6 MD |
98 | /* If the headers do not support membarrier system call, fall back smp_mb. */ |
99 | #ifdef __NR_membarrier | |
100 | # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__) | |
f541831e MD |
101 | #else |
102 | # define membarrier(...) -ENOSYS | |
103 | #endif | |
104 | ||
105 | enum membarrier_cmd { | |
3745305b MD |
106 | MEMBARRIER_CMD_QUERY = 0, |
107 | MEMBARRIER_CMD_SHARED = (1 << 0), | |
108 | /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */ | |
109 | /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */ | |
110 | MEMBARRIER_CMD_PRIVATE_EXPEDITED = (1 << 3), | |
111 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED = (1 << 4), | |
f541831e MD |
112 | }; |
113 | ||
c1be8fb9 MD |
114 | static |
115 | void __attribute__((constructor)) rcu_bp_init(void); | |
116 | static | |
02be5561 | 117 | void __attribute__((destructor)) rcu_bp_exit(void); |
fdee2e6d | 118 | |
d8d9a340 | 119 | #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER |
f541831e | 120 | int urcu_bp_has_sys_membarrier; |
d8d9a340 | 121 | #endif |
f541831e | 122 | |
731ccb96 MD |
123 | /* |
124 | * rcu_gp_lock ensures mutual exclusion between threads calling | |
125 | * synchronize_rcu(). | |
126 | */ | |
6abb4bd5 | 127 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
731ccb96 MD |
128 | /* |
129 | * rcu_registry_lock ensures mutual exclusion between threads | |
130 | * registering and unregistering themselves to/from the registry, and | |
131 | * with threads reading that registry from synchronize_rcu(). However, | |
132 | * this lock is not held all the way through the completion of awaiting | |
133 | * for the grace period. It is sporadically released between iterations | |
134 | * on the registry. | |
135 | * rcu_registry_lock may nest inside rcu_gp_lock. | |
136 | */ | |
137 | static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER; | |
fdee2e6d | 138 | |
c1be8fb9 MD |
139 | static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER; |
140 | static int initialized; | |
141 | ||
142 | static pthread_key_t urcu_bp_key; | |
143 | ||
c13c2e55 | 144 | struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT }; |
fdee2e6d MD |
145 | |
146 | /* | |
147 | * Pointer to registry elements. Written to only by each individual reader. Read | |
148 | * by both the reader and the writers. | |
149 | */ | |
2f661865 | 150 | DEFINE_URCU_TLS(struct rcu_reader *, rcu_reader); |
fdee2e6d | 151 | |
16aa9ee8 | 152 | static CDS_LIST_HEAD(registry); |
fdee2e6d | 153 | |
95b94246 MD |
154 | struct registry_chunk { |
155 | size_t data_len; /* data length */ | |
c1be8fb9 | 156 | size_t used; /* amount of data used */ |
95b94246 MD |
157 | struct cds_list_head node; /* chunk_list node */ |
158 | char data[]; | |
159 | }; | |
160 | ||
fdee2e6d | 161 | struct registry_arena { |
95b94246 | 162 | struct cds_list_head chunk_list; |
fdee2e6d MD |
163 | }; |
164 | ||
95b94246 MD |
165 | static struct registry_arena registry_arena = { |
166 | .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list), | |
167 | }; | |
fdee2e6d | 168 | |
4cf1675f MD |
169 | /* Saved fork signal mask, protected by rcu_gp_lock */ |
170 | static sigset_t saved_fork_signal_mask; | |
171 | ||
6abb4bd5 | 172 | static void mutex_lock(pthread_mutex_t *mutex) |
fdee2e6d MD |
173 | { |
174 | int ret; | |
175 | ||
176 | #ifndef DISTRUST_SIGNALS_EXTREME | |
6abb4bd5 | 177 | ret = pthread_mutex_lock(mutex); |
4a6d7378 MD |
178 | if (ret) |
179 | urcu_die(ret); | |
fdee2e6d | 180 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ |
6abb4bd5 | 181 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
4a6d7378 MD |
182 | if (ret != EBUSY && ret != EINTR) |
183 | urcu_die(ret); | |
fdee2e6d MD |
184 | poll(NULL,0,10); |
185 | } | |
186 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
187 | } | |
188 | ||
6abb4bd5 | 189 | static void mutex_unlock(pthread_mutex_t *mutex) |
fdee2e6d MD |
190 | { |
191 | int ret; | |
192 | ||
6abb4bd5 | 193 | ret = pthread_mutex_unlock(mutex); |
4a6d7378 MD |
194 | if (ret) |
195 | urcu_die(ret); | |
fdee2e6d MD |
196 | } |
197 | ||
f541831e MD |
198 | static void smp_mb_master(void) |
199 | { | |
3745305b MD |
200 | if (caa_likely(urcu_bp_has_sys_membarrier)) { |
201 | if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0)) | |
202 | urcu_die(errno); | |
203 | } else { | |
f541831e | 204 | cmm_smp_mb(); |
3745305b | 205 | } |
f541831e MD |
206 | } |
207 | ||
731ccb96 MD |
208 | /* |
209 | * Always called with rcu_registry lock held. Releases this lock between | |
210 | * iterations and grabs it again. Holds the lock when it returns. | |
211 | */ | |
52c75091 MD |
212 | static void wait_for_readers(struct cds_list_head *input_readers, |
213 | struct cds_list_head *cur_snap_readers, | |
214 | struct cds_list_head *qsreaders) | |
fdee2e6d | 215 | { |
9340c38d | 216 | unsigned int wait_loops = 0; |
02be5561 | 217 | struct rcu_reader *index, *tmp; |
fdee2e6d | 218 | |
fdee2e6d | 219 | /* |
dd61d077 MD |
220 | * Wait for each thread URCU_TLS(rcu_reader).ctr to either |
221 | * indicate quiescence (not nested), or observe the current | |
c13c2e55 | 222 | * rcu_gp.ctr value. |
fdee2e6d MD |
223 | */ |
224 | for (;;) { | |
9340c38d MD |
225 | if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS) |
226 | wait_loops++; | |
227 | ||
52c75091 MD |
228 | cds_list_for_each_entry_safe(index, tmp, input_readers, node) { |
229 | switch (rcu_reader_state(&index->ctr)) { | |
230 | case RCU_READER_ACTIVE_CURRENT: | |
231 | if (cur_snap_readers) { | |
232 | cds_list_move(&index->node, | |
233 | cur_snap_readers); | |
234 | break; | |
235 | } | |
236 | /* Fall-through */ | |
237 | case RCU_READER_INACTIVE: | |
238 | cds_list_move(&index->node, qsreaders); | |
239 | break; | |
240 | case RCU_READER_ACTIVE_OLD: | |
241 | /* | |
242 | * Old snapshot. Leaving node in | |
243 | * input_readers will make us busy-loop | |
244 | * until the snapshot becomes current or | |
245 | * the reader becomes inactive. | |
246 | */ | |
247 | break; | |
248 | } | |
fdee2e6d MD |
249 | } |
250 | ||
52c75091 | 251 | if (cds_list_empty(input_readers)) { |
fdee2e6d MD |
252 | break; |
253 | } else { | |
731ccb96 MD |
254 | /* Temporarily unlock the registry lock. */ |
255 | mutex_unlock(&rcu_registry_lock); | |
9340c38d MD |
256 | if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) |
257 | (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS); | |
fdee2e6d | 258 | else |
06f22bdb | 259 | caa_cpu_relax(); |
731ccb96 MD |
260 | /* Re-lock the registry lock before the next loop. */ |
261 | mutex_lock(&rcu_registry_lock); | |
fdee2e6d MD |
262 | } |
263 | } | |
fdee2e6d MD |
264 | } |
265 | ||
266 | void synchronize_rcu(void) | |
267 | { | |
52c75091 MD |
268 | CDS_LIST_HEAD(cur_snap_readers); |
269 | CDS_LIST_HEAD(qsreaders); | |
fdee2e6d MD |
270 | sigset_t newmask, oldmask; |
271 | int ret; | |
272 | ||
6ed4b2e6 | 273 | ret = sigfillset(&newmask); |
fdee2e6d | 274 | assert(!ret); |
6ed4b2e6 | 275 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
fdee2e6d MD |
276 | assert(!ret); |
277 | ||
6abb4bd5 | 278 | mutex_lock(&rcu_gp_lock); |
fdee2e6d | 279 | |
731ccb96 MD |
280 | mutex_lock(&rcu_registry_lock); |
281 | ||
16aa9ee8 | 282 | if (cds_list_empty(®istry)) |
2dfb8b5e | 283 | goto out; |
fdee2e6d MD |
284 | |
285 | /* All threads should read qparity before accessing data structure | |
2dfb8b5e | 286 | * where new ptr points to. */ |
fdee2e6d | 287 | /* Write new ptr before changing the qparity */ |
f541831e | 288 | smp_mb_master(); |
fdee2e6d | 289 | |
fdee2e6d | 290 | /* |
dd61d077 | 291 | * Wait for readers to observe original parity or be quiescent. |
731ccb96 MD |
292 | * wait_for_readers() can release and grab again rcu_registry_lock |
293 | * interally. | |
dd61d077 | 294 | */ |
52c75091 | 295 | wait_for_readers(®istry, &cur_snap_readers, &qsreaders); |
dd61d077 MD |
296 | |
297 | /* | |
298 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the | |
299 | * model easier to understand. It does not have a big performance impact | |
300 | * anyway, given this is the write-side. | |
301 | */ | |
302 | cmm_smp_mb(); | |
303 | ||
304 | /* Switch parity: 0 -> 1, 1 -> 0 */ | |
c13c2e55 | 305 | CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE); |
dd61d077 MD |
306 | |
307 | /* | |
308 | * Must commit qparity update to memory before waiting for other parity | |
309 | * quiescent state. Failure to do so could result in the writer waiting | |
310 | * forever while new readers are always accessing data (no progress). | |
311 | * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED. | |
fdee2e6d | 312 | */ |
fdee2e6d MD |
313 | |
314 | /* | |
5481ddb3 | 315 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
fdee2e6d MD |
316 | * model easier to understand. It does not have a big performance impact |
317 | * anyway, given this is the write-side. | |
318 | */ | |
5481ddb3 | 319 | cmm_smp_mb(); |
fdee2e6d | 320 | |
fdee2e6d | 321 | /* |
dd61d077 | 322 | * Wait for readers to observe new parity or be quiescent. |
731ccb96 MD |
323 | * wait_for_readers() can release and grab again rcu_registry_lock |
324 | * interally. | |
fdee2e6d | 325 | */ |
52c75091 MD |
326 | wait_for_readers(&cur_snap_readers, NULL, &qsreaders); |
327 | ||
328 | /* | |
329 | * Put quiescent reader list back into registry. | |
330 | */ | |
331 | cds_list_splice(&qsreaders, ®istry); | |
fdee2e6d MD |
332 | |
333 | /* | |
2dfb8b5e MD |
334 | * Finish waiting for reader threads before letting the old ptr being |
335 | * freed. | |
fdee2e6d | 336 | */ |
f541831e | 337 | smp_mb_master(); |
2dfb8b5e | 338 | out: |
731ccb96 | 339 | mutex_unlock(&rcu_registry_lock); |
6abb4bd5 | 340 | mutex_unlock(&rcu_gp_lock); |
fdee2e6d MD |
341 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
342 | assert(!ret); | |
343 | } | |
344 | ||
345 | /* | |
346 | * library wrappers to be used by non-LGPL compatible source code. | |
347 | */ | |
348 | ||
349 | void rcu_read_lock(void) | |
350 | { | |
351 | _rcu_read_lock(); | |
352 | } | |
353 | ||
354 | void rcu_read_unlock(void) | |
355 | { | |
356 | _rcu_read_unlock(); | |
357 | } | |
358 | ||
882f3357 MD |
359 | int rcu_read_ongoing(void) |
360 | { | |
361 | return _rcu_read_ongoing(); | |
362 | } | |
363 | ||
fdee2e6d | 364 | /* |
95b94246 MD |
365 | * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk. |
366 | * Else, try expanding the last chunk. If this fails, allocate a new | |
367 | * chunk twice as big as the last chunk. | |
368 | * Memory used by chunks _never_ moves. A chunk could theoretically be | |
369 | * freed when all "used" slots are released, but we don't do it at this | |
370 | * point. | |
fdee2e6d | 371 | */ |
95b94246 MD |
372 | static |
373 | void expand_arena(struct registry_arena *arena) | |
fdee2e6d | 374 | { |
95b94246 MD |
375 | struct registry_chunk *new_chunk, *last_chunk; |
376 | size_t old_chunk_len, new_chunk_len; | |
377 | ||
378 | /* No chunk. */ | |
379 | if (cds_list_empty(&arena->chunk_list)) { | |
380 | assert(ARENA_INIT_ALLOC >= | |
381 | sizeof(struct registry_chunk) | |
382 | + sizeof(struct rcu_reader)); | |
383 | new_chunk_len = ARENA_INIT_ALLOC; | |
5592d049 MJ |
384 | new_chunk = (struct registry_chunk *) mmap(NULL, |
385 | new_chunk_len, | |
9d8612b7 MD |
386 | PROT_READ | PROT_WRITE, |
387 | MAP_ANONYMOUS | MAP_PRIVATE, | |
388 | -1, 0); | |
95b94246 MD |
389 | if (new_chunk == MAP_FAILED) |
390 | abort(); | |
d3ac5bb7 | 391 | memset(new_chunk, 0, new_chunk_len); |
95b94246 MD |
392 | new_chunk->data_len = |
393 | new_chunk_len - sizeof(struct registry_chunk); | |
394 | cds_list_add_tail(&new_chunk->node, &arena->chunk_list); | |
395 | return; /* We're done. */ | |
396 | } | |
9d8612b7 | 397 | |
95b94246 MD |
398 | /* Try expanding last chunk. */ |
399 | last_chunk = cds_list_entry(arena->chunk_list.prev, | |
400 | struct registry_chunk, node); | |
401 | old_chunk_len = | |
402 | last_chunk->data_len + sizeof(struct registry_chunk); | |
403 | new_chunk_len = old_chunk_len << 1; | |
404 | ||
405 | /* Don't allow memory mapping to move, just expand. */ | |
406 | new_chunk = mremap_wrapper(last_chunk, old_chunk_len, | |
407 | new_chunk_len, 0); | |
408 | if (new_chunk != MAP_FAILED) { | |
409 | /* Should not have moved. */ | |
410 | assert(new_chunk == last_chunk); | |
d3ac5bb7 | 411 | memset((char *) last_chunk + old_chunk_len, 0, |
95b94246 MD |
412 | new_chunk_len - old_chunk_len); |
413 | last_chunk->data_len = | |
414 | new_chunk_len - sizeof(struct registry_chunk); | |
415 | return; /* We're done. */ | |
416 | } | |
0617bf4c | 417 | |
95b94246 | 418 | /* Remap did not succeed, we need to add a new chunk. */ |
5592d049 MJ |
419 | new_chunk = (struct registry_chunk *) mmap(NULL, |
420 | new_chunk_len, | |
95b94246 MD |
421 | PROT_READ | PROT_WRITE, |
422 | MAP_ANONYMOUS | MAP_PRIVATE, | |
423 | -1, 0); | |
424 | if (new_chunk == MAP_FAILED) | |
425 | abort(); | |
d3ac5bb7 | 426 | memset(new_chunk, 0, new_chunk_len); |
95b94246 MD |
427 | new_chunk->data_len = |
428 | new_chunk_len - sizeof(struct registry_chunk); | |
429 | cds_list_add_tail(&new_chunk->node, &arena->chunk_list); | |
430 | } | |
fdee2e6d | 431 | |
95b94246 MD |
432 | static |
433 | struct rcu_reader *arena_alloc(struct registry_arena *arena) | |
434 | { | |
435 | struct registry_chunk *chunk; | |
436 | struct rcu_reader *rcu_reader_reg; | |
437 | int expand_done = 0; /* Only allow to expand once per alloc */ | |
438 | size_t len = sizeof(struct rcu_reader); | |
439 | ||
440 | retry: | |
441 | cds_list_for_each_entry(chunk, &arena->chunk_list, node) { | |
442 | if (chunk->data_len - chunk->used < len) | |
443 | continue; | |
444 | /* Find spot */ | |
445 | for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0]; | |
446 | rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len]; | |
447 | rcu_reader_reg++) { | |
448 | if (!rcu_reader_reg->alloc) { | |
449 | rcu_reader_reg->alloc = 1; | |
450 | chunk->used += len; | |
451 | return rcu_reader_reg; | |
452 | } | |
453 | } | |
454 | } | |
455 | ||
456 | if (!expand_done) { | |
457 | expand_arena(arena); | |
458 | expand_done = 1; | |
459 | goto retry; | |
460 | } | |
461 | ||
462 | return NULL; | |
fdee2e6d MD |
463 | } |
464 | ||
465 | /* Called with signals off and mutex locked */ | |
95b94246 MD |
466 | static |
467 | void add_thread(void) | |
fdee2e6d | 468 | { |
02be5561 | 469 | struct rcu_reader *rcu_reader_reg; |
c1be8fb9 | 470 | int ret; |
fdee2e6d | 471 | |
95b94246 MD |
472 | rcu_reader_reg = arena_alloc(®istry_arena); |
473 | if (!rcu_reader_reg) | |
474 | abort(); | |
c1be8fb9 MD |
475 | ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg); |
476 | if (ret) | |
477 | abort(); | |
fdee2e6d MD |
478 | |
479 | /* Add to registry */ | |
02be5561 MD |
480 | rcu_reader_reg->tid = pthread_self(); |
481 | assert(rcu_reader_reg->ctr == 0); | |
16aa9ee8 | 482 | cds_list_add(&rcu_reader_reg->node, ®istry); |
95b94246 MD |
483 | /* |
484 | * Reader threads are pointing to the reader registry. This is | |
485 | * why its memory should never be relocated. | |
486 | */ | |
bd252a04 | 487 | URCU_TLS(rcu_reader) = rcu_reader_reg; |
fdee2e6d MD |
488 | } |
489 | ||
c1be8fb9 MD |
490 | /* Called with mutex locked */ |
491 | static | |
492 | void cleanup_thread(struct registry_chunk *chunk, | |
493 | struct rcu_reader *rcu_reader_reg) | |
494 | { | |
495 | rcu_reader_reg->ctr = 0; | |
496 | cds_list_del(&rcu_reader_reg->node); | |
497 | rcu_reader_reg->tid = 0; | |
498 | rcu_reader_reg->alloc = 0; | |
499 | chunk->used -= sizeof(struct rcu_reader); | |
500 | } | |
501 | ||
502 | static | |
503 | struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg) | |
fdee2e6d | 504 | { |
95b94246 | 505 | struct registry_chunk *chunk; |
fdee2e6d | 506 | |
95b94246 | 507 | cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) { |
c1be8fb9 MD |
508 | if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0]) |
509 | continue; | |
510 | if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len]) | |
511 | continue; | |
512 | return chunk; | |
513 | } | |
514 | return NULL; | |
515 | } | |
95b94246 | 516 | |
c1be8fb9 MD |
517 | /* Called with signals off and mutex locked */ |
518 | static | |
76d6a951 | 519 | void remove_thread(struct rcu_reader *rcu_reader_reg) |
c1be8fb9 | 520 | { |
c1be8fb9 MD |
521 | cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg); |
522 | URCU_TLS(rcu_reader) = NULL; | |
fdee2e6d MD |
523 | } |
524 | ||
525 | /* Disable signals, take mutex, add to registry */ | |
526 | void rcu_bp_register(void) | |
527 | { | |
528 | sigset_t newmask, oldmask; | |
529 | int ret; | |
530 | ||
6ed4b2e6 | 531 | ret = sigfillset(&newmask); |
c1be8fb9 MD |
532 | if (ret) |
533 | abort(); | |
6ed4b2e6 | 534 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
c1be8fb9 MD |
535 | if (ret) |
536 | abort(); | |
fdee2e6d MD |
537 | |
538 | /* | |
539 | * Check if a signal concurrently registered our thread since | |
c1be8fb9 MD |
540 | * the check in rcu_read_lock(). |
541 | */ | |
bd252a04 | 542 | if (URCU_TLS(rcu_reader)) |
fdee2e6d MD |
543 | goto end; |
544 | ||
c1be8fb9 MD |
545 | /* |
546 | * Take care of early registration before urcu_bp constructor. | |
547 | */ | |
548 | rcu_bp_init(); | |
549 | ||
731ccb96 | 550 | mutex_lock(&rcu_registry_lock); |
fdee2e6d | 551 | add_thread(); |
731ccb96 | 552 | mutex_unlock(&rcu_registry_lock); |
fdee2e6d MD |
553 | end: |
554 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
c1be8fb9 MD |
555 | if (ret) |
556 | abort(); | |
557 | } | |
558 | ||
559 | /* Disable signals, take mutex, remove from registry */ | |
560 | static | |
76d6a951 | 561 | void rcu_bp_unregister(struct rcu_reader *rcu_reader_reg) |
c1be8fb9 MD |
562 | { |
563 | sigset_t newmask, oldmask; | |
564 | int ret; | |
565 | ||
566 | ret = sigfillset(&newmask); | |
567 | if (ret) | |
568 | abort(); | |
569 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); | |
570 | if (ret) | |
571 | abort(); | |
572 | ||
731ccb96 | 573 | mutex_lock(&rcu_registry_lock); |
76d6a951 | 574 | remove_thread(rcu_reader_reg); |
731ccb96 | 575 | mutex_unlock(&rcu_registry_lock); |
c1be8fb9 MD |
576 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
577 | if (ret) | |
578 | abort(); | |
76d6a951 | 579 | rcu_bp_exit(); |
c1be8fb9 MD |
580 | } |
581 | ||
582 | /* | |
583 | * Remove thread from the registry when it exits, and flag it as | |
584 | * destroyed so garbage collection can take care of it. | |
585 | */ | |
586 | static | |
587 | void urcu_bp_thread_exit_notifier(void *rcu_key) | |
588 | { | |
76d6a951 | 589 | rcu_bp_unregister(rcu_key); |
c1be8fb9 MD |
590 | } |
591 | ||
d8d9a340 MD |
592 | #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER |
593 | static | |
3745305b | 594 | void rcu_sys_membarrier_status(bool available) |
d8d9a340 MD |
595 | { |
596 | if (!available) | |
597 | abort(); | |
598 | } | |
599 | #else | |
600 | static | |
3745305b | 601 | void rcu_sys_membarrier_status(bool available) |
d8d9a340 | 602 | { |
3745305b MD |
603 | if (!available) |
604 | return; | |
605 | urcu_bp_has_sys_membarrier = 1; | |
d8d9a340 MD |
606 | } |
607 | #endif | |
608 | ||
3745305b MD |
609 | static |
610 | void rcu_sys_membarrier_init(void) | |
611 | { | |
612 | bool available = false; | |
613 | int mask; | |
614 | ||
615 | mask = membarrier(MEMBARRIER_CMD_QUERY, 0); | |
616 | if (mask >= 0) { | |
617 | if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) { | |
618 | if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0)) | |
619 | urcu_die(errno); | |
620 | available = true; | |
621 | } | |
622 | } | |
623 | rcu_sys_membarrier_status(available); | |
624 | } | |
625 | ||
c1be8fb9 MD |
626 | static |
627 | void rcu_bp_init(void) | |
628 | { | |
629 | mutex_lock(&init_lock); | |
76d6a951 | 630 | if (!rcu_bp_refcount++) { |
c1be8fb9 MD |
631 | int ret; |
632 | ||
633 | ret = pthread_key_create(&urcu_bp_key, | |
634 | urcu_bp_thread_exit_notifier); | |
635 | if (ret) | |
636 | abort(); | |
3745305b | 637 | rcu_sys_membarrier_init(); |
c1be8fb9 MD |
638 | initialized = 1; |
639 | } | |
640 | mutex_unlock(&init_lock); | |
fdee2e6d MD |
641 | } |
642 | ||
c1be8fb9 | 643 | static |
9380711a | 644 | void rcu_bp_exit(void) |
fdee2e6d | 645 | { |
76d6a951 MD |
646 | mutex_lock(&init_lock); |
647 | if (!--rcu_bp_refcount) { | |
648 | struct registry_chunk *chunk, *tmp; | |
649 | int ret; | |
95b94246 | 650 | |
76d6a951 MD |
651 | cds_list_for_each_entry_safe(chunk, tmp, |
652 | ®istry_arena.chunk_list, node) { | |
5592d049 | 653 | munmap((void *) chunk, chunk->data_len |
76d6a951 MD |
654 | + sizeof(struct registry_chunk)); |
655 | } | |
7937ae1c | 656 | CDS_INIT_LIST_HEAD(®istry_arena.chunk_list); |
76d6a951 MD |
657 | ret = pthread_key_delete(urcu_bp_key); |
658 | if (ret) | |
659 | abort(); | |
95b94246 | 660 | } |
76d6a951 | 661 | mutex_unlock(&init_lock); |
fdee2e6d | 662 | } |
4cf1675f MD |
663 | |
664 | /* | |
731ccb96 MD |
665 | * Holding the rcu_gp_lock and rcu_registry_lock across fork will make |
666 | * sure we fork() don't race with a concurrent thread executing with | |
667 | * any of those locks held. This ensures that the registry and data | |
668 | * protected by rcu_gp_lock are in a coherent state in the child. | |
4cf1675f MD |
669 | */ |
670 | void rcu_bp_before_fork(void) | |
671 | { | |
672 | sigset_t newmask, oldmask; | |
673 | int ret; | |
674 | ||
6ed4b2e6 | 675 | ret = sigfillset(&newmask); |
4cf1675f | 676 | assert(!ret); |
6ed4b2e6 | 677 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
4cf1675f MD |
678 | assert(!ret); |
679 | mutex_lock(&rcu_gp_lock); | |
731ccb96 | 680 | mutex_lock(&rcu_registry_lock); |
4cf1675f MD |
681 | saved_fork_signal_mask = oldmask; |
682 | } | |
683 | ||
684 | void rcu_bp_after_fork_parent(void) | |
685 | { | |
686 | sigset_t oldmask; | |
687 | int ret; | |
688 | ||
689 | oldmask = saved_fork_signal_mask; | |
731ccb96 | 690 | mutex_unlock(&rcu_registry_lock); |
4cf1675f MD |
691 | mutex_unlock(&rcu_gp_lock); |
692 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
693 | assert(!ret); | |
694 | } | |
695 | ||
c1be8fb9 MD |
696 | /* |
697 | * Prune all entries from registry except our own thread. Fits the Linux | |
731ccb96 | 698 | * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held. |
c1be8fb9 MD |
699 | */ |
700 | static | |
701 | void urcu_bp_prune_registry(void) | |
702 | { | |
703 | struct registry_chunk *chunk; | |
704 | struct rcu_reader *rcu_reader_reg; | |
705 | ||
706 | cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) { | |
707 | for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0]; | |
708 | rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len]; | |
709 | rcu_reader_reg++) { | |
710 | if (!rcu_reader_reg->alloc) | |
711 | continue; | |
712 | if (rcu_reader_reg->tid == pthread_self()) | |
713 | continue; | |
714 | cleanup_thread(chunk, rcu_reader_reg); | |
715 | } | |
716 | } | |
717 | } | |
718 | ||
4cf1675f MD |
719 | void rcu_bp_after_fork_child(void) |
720 | { | |
721 | sigset_t oldmask; | |
722 | int ret; | |
723 | ||
c1be8fb9 | 724 | urcu_bp_prune_registry(); |
4cf1675f | 725 | oldmask = saved_fork_signal_mask; |
731ccb96 | 726 | mutex_unlock(&rcu_registry_lock); |
4cf1675f MD |
727 | mutex_unlock(&rcu_gp_lock); |
728 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); | |
729 | assert(!ret); | |
730 | } | |
5e77fc1f | 731 | |
9b7981bb MD |
732 | void *rcu_dereference_sym_bp(void *p) |
733 | { | |
734 | return _rcu_dereference(p); | |
735 | } | |
736 | ||
5efd3cd2 MD |
737 | void *rcu_set_pointer_sym_bp(void **p, void *v) |
738 | { | |
739 | cmm_wmb(); | |
424d4ed5 MD |
740 | uatomic_set(p, v); |
741 | return v; | |
5efd3cd2 MD |
742 | } |
743 | ||
744 | void *rcu_xchg_pointer_sym_bp(void **p, void *v) | |
745 | { | |
746 | cmm_wmb(); | |
747 | return uatomic_xchg(p, v); | |
748 | } | |
749 | ||
750 | void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new) | |
751 | { | |
752 | cmm_wmb(); | |
753 | return uatomic_cmpxchg(p, old, _new); | |
754 | } | |
755 | ||
5e6b23a6 | 756 | DEFINE_RCU_FLAVOR(rcu_flavor); |
541d828d | 757 | |
5e77fc1f | 758 | #include "urcu-call-rcu-impl.h" |
0376e7b2 | 759 | #include "urcu-defer-impl.h" |