Commit | Line | Data |
---|---|---|
b257a10b MD |
1 | /* |
2 | * urcu.c | |
3 | * | |
4 | * Userspace RCU library | |
5 | * | |
6982d6d7 | 6 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
af02d47e | 7 | * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. |
b257a10b | 8 | * |
af02d47e MD |
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 | |
54843abc PM |
22 | * |
23 | * IBM's contributions to this file may be relicensed under LGPLv2 or later. | |
b257a10b MD |
24 | */ |
25 | ||
fdf01eed | 26 | #define _BSD_SOURCE |
c1d2c60b | 27 | #define _GNU_SOURCE |
27b012e2 MD |
28 | #include <stdio.h> |
29 | #include <pthread.h> | |
30 | #include <signal.h> | |
31 | #include <assert.h> | |
f69f195a | 32 | #include <stdlib.h> |
6d841bc2 | 33 | #include <stdint.h> |
f69f195a | 34 | #include <string.h> |
09a9f986 | 35 | #include <errno.h> |
e8043c1b | 36 | #include <poll.h> |
27b012e2 | 37 | |
57760d44 | 38 | #include "urcu/map/urcu.h" |
5e77fc1f | 39 | |
af7c2dbe | 40 | #include "urcu/static/urcu.h" |
121a5d44 | 41 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
27b012e2 MD |
42 | #include "urcu.h" |
43 | ||
3a71751e PB |
44 | /* |
45 | * If a reader is really non-cooperative and refuses to commit its | |
46 | * rcu_active_readers count to memory (there is no barrier in the reader | |
47 | * per-se), kick it after a few loops waiting for it. | |
48 | */ | |
49 | #define KICK_READER_LOOPS 10000 | |
50 | ||
51 | /* | |
52 | * Active attempts to check for reader Q.S. before calling futex(). | |
53 | */ | |
54 | #define RCU_QS_ACTIVE_ATTEMPTS 100 | |
55 | ||
fdf01eed | 56 | #ifdef RCU_MEMBARRIER |
834a45ba | 57 | static int init_done; |
fdf01eed | 58 | int has_sys_membarrier; |
834a45ba | 59 | |
02be5561 | 60 | void __attribute__((constructor)) rcu_init(void); |
fdf01eed MD |
61 | #endif |
62 | ||
63 | #ifdef RCU_MB | |
02be5561 | 64 | void rcu_init(void) |
e90a6e9c MD |
65 | { |
66 | } | |
67 | #endif | |
8a5fb4c9 | 68 | |
fdf01eed MD |
69 | #ifdef RCU_SIGNAL |
70 | static int init_done; | |
71 | ||
72 | void __attribute__((constructor)) rcu_init(void); | |
73 | void __attribute__((destructor)) rcu_exit(void); | |
74 | #endif | |
75 | ||
6abb4bd5 | 76 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
27b012e2 | 77 | |
6d841bc2 | 78 | int32_t gp_futex; |
bc6c15bb | 79 | |
128166c9 MD |
80 | /* |
81 | * Global grace period counter. | |
02be5561 | 82 | * Contains the current RCU_GP_CTR_PHASE. |
afb8f2c9 | 83 | * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. |
b0d5e790 | 84 | * Written to only by writer with mutex taken. Read by both writer and readers. |
128166c9 | 85 | */ |
27d65bc5 | 86 | unsigned long rcu_gp_ctr = RCU_GP_COUNT; |
27b012e2 | 87 | |
b0d5e790 MD |
88 | /* |
89 | * Written to only by each individual reader. Read by both the reader and the | |
90 | * writers. | |
91 | */ | |
02be5561 | 92 | struct rcu_reader __thread rcu_reader; |
27b012e2 | 93 | |
cf380c2f | 94 | #ifdef DEBUG_YIELD |
9d335088 MD |
95 | unsigned int yield_active; |
96 | unsigned int __thread rand_yield; | |
cf380c2f MD |
97 | #endif |
98 | ||
16aa9ee8 | 99 | static CDS_LIST_HEAD(registry); |
27b012e2 | 100 | |
6abb4bd5 | 101 | static void mutex_lock(pthread_mutex_t *mutex) |
41718ff9 MD |
102 | { |
103 | int ret; | |
09a9f986 PM |
104 | |
105 | #ifndef DISTRUST_SIGNALS_EXTREME | |
6abb4bd5 | 106 | ret = pthread_mutex_lock(mutex); |
41718ff9 MD |
107 | if (ret) { |
108 | perror("Error in pthread mutex lock"); | |
109 | exit(-1); | |
110 | } | |
09a9f986 | 111 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ |
6abb4bd5 | 112 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
09a9f986 PM |
113 | if (ret != EBUSY && ret != EINTR) { |
114 | printf("ret = %d, errno = %d\n", ret, errno); | |
115 | perror("Error in pthread mutex lock"); | |
116 | exit(-1); | |
117 | } | |
6cf3827c | 118 | if (CMM_LOAD_SHARED(rcu_reader.need_mb)) { |
5481ddb3 | 119 | cmm_smp_mb(); |
6cf3827c | 120 | _CMM_STORE_SHARED(rcu_reader.need_mb, 0); |
5481ddb3 | 121 | cmm_smp_mb(); |
09a9f986 PM |
122 | } |
123 | poll(NULL,0,10); | |
124 | } | |
125 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ | |
41718ff9 MD |
126 | } |
127 | ||
6abb4bd5 | 128 | static void mutex_unlock(pthread_mutex_t *mutex) |
41718ff9 MD |
129 | { |
130 | int ret; | |
131 | ||
6abb4bd5 | 132 | ret = pthread_mutex_unlock(mutex); |
41718ff9 MD |
133 | if (ret) { |
134 | perror("Error in pthread mutex unlock"); | |
135 | exit(-1); | |
136 | } | |
137 | } | |
138 | ||
fdf01eed | 139 | #ifdef RCU_MEMBARRIER |
25cc6d18 | 140 | static void smp_mb_master(int group) |
fdf01eed MD |
141 | { |
142 | if (likely(has_sys_membarrier)) | |
f0708810 | 143 | membarrier(MEMBARRIER_EXPEDITED); |
fdf01eed | 144 | else |
5481ddb3 | 145 | cmm_smp_mb(); |
fdf01eed MD |
146 | } |
147 | #endif | |
148 | ||
02be5561 | 149 | #ifdef RCU_MB |
25cc6d18 | 150 | static void smp_mb_master(int group) |
40e140c9 | 151 | { |
5481ddb3 | 152 | cmm_smp_mb(); |
40e140c9 | 153 | } |
fdf01eed MD |
154 | #endif |
155 | ||
156 | #ifdef RCU_SIGNAL | |
78ff9419 | 157 | static void force_mb_all_readers(void) |
27b012e2 | 158 | { |
02be5561 | 159 | struct rcu_reader *index; |
e3b0cef0 | 160 | |
27b012e2 | 161 | /* |
5481ddb3 | 162 | * Ask for each threads to execute a cmm_smp_mb() so we can consider the |
27b012e2 MD |
163 | * compiler barriers around rcu read lock as real memory barriers. |
164 | */ | |
16aa9ee8 | 165 | if (cds_list_empty(®istry)) |
27b012e2 | 166 | return; |
3a86deba | 167 | /* |
5481ddb3 | 168 | * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs |
157dca95 | 169 | * a cache flush on architectures with non-coherent cache. Let's play |
5481ddb3 | 170 | * safe and don't assume anything : we use cmm_smp_mc() to make sure the |
157dca95 | 171 | * cache flush is enforced. |
3a86deba | 172 | */ |
16aa9ee8 | 173 | cds_list_for_each_entry(index, ®istry, node) { |
6cf3827c | 174 | CMM_STORE_SHARED(index->need_mb, 1); |
02be5561 | 175 | pthread_kill(index->tid, SIGRCU); |
09a9f986 | 176 | } |
27b012e2 MD |
177 | /* |
178 | * Wait for sighandler (and thus mb()) to execute on every thread. | |
09a9f986 PM |
179 | * |
180 | * Note that the pthread_kill() will never be executed on systems | |
181 | * that correctly deliver signals in a timely manner. However, it | |
182 | * is not uncommon for kernels to have bugs that can result in | |
183 | * lost or unduly delayed signals. | |
184 | * | |
185 | * If you are seeing the below pthread_kill() executing much at | |
186 | * all, we suggest testing the underlying kernel and filing the | |
187 | * relevant bug report. For Linux kernels, we recommend getting | |
188 | * the Linux Test Project (LTP). | |
27b012e2 | 189 | */ |
16aa9ee8 | 190 | cds_list_for_each_entry(index, ®istry, node) { |
6cf3827c | 191 | while (CMM_LOAD_SHARED(index->need_mb)) { |
02be5561 | 192 | pthread_kill(index->tid, SIGRCU); |
09a9f986 PM |
193 | poll(NULL, 0, 1); |
194 | } | |
195 | } | |
5481ddb3 | 196 | cmm_smp_mb(); /* read ->need_mb before ending the barrier */ |
27b012e2 | 197 | } |
9d7e3f89 | 198 | |
25cc6d18 | 199 | static void smp_mb_master(int group) |
9d7e3f89 MD |
200 | { |
201 | force_mb_all_readers(); | |
202 | } | |
fdf01eed | 203 | #endif /* #ifdef RCU_SIGNAL */ |
27b012e2 | 204 | |
bc6c15bb MD |
205 | /* |
206 | * synchronize_rcu() waiting. Single thread. | |
207 | */ | |
cfe78e25 | 208 | static void wait_gp(void) |
bc6c15bb | 209 | { |
cfe78e25 | 210 | /* Read reader_gp before read futex */ |
25cc6d18 | 211 | smp_mb_master(RCU_MB_GROUP); |
cfe78e25 | 212 | if (uatomic_read(&gp_futex) == -1) |
0854ccff | 213 | futex_async(&gp_futex, FUTEX_WAIT, -1, |
cfe78e25 | 214 | NULL, NULL, 0); |
bc6c15bb MD |
215 | } |
216 | ||
2dfb8b5e | 217 | void update_counter_and_wait(void) |
27b012e2 | 218 | { |
16aa9ee8 | 219 | CDS_LIST_HEAD(qsreaders); |
cfe78e25 | 220 | int wait_loops = 0; |
02be5561 | 221 | struct rcu_reader *index, *tmp; |
27b012e2 | 222 | |
32c15e4e | 223 | /* Switch parity: 0 -> 1, 1 -> 0 */ |
6cf3827c | 224 | CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); |
2dfb8b5e MD |
225 | |
226 | /* | |
d40fde2c MD |
227 | * Must commit rcu_gp_ctr update to memory before waiting for quiescent |
228 | * state. Failure to do so could result in the writer waiting forever | |
229 | * while new readers are always accessing data (no progress). Enforce | |
230 | * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr. | |
2dfb8b5e | 231 | */ |
5481ddb3 | 232 | cmm_barrier(); |
2dfb8b5e MD |
233 | |
234 | /* | |
935b11ff | 235 | * |
5481ddb3 | 236 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
2dfb8b5e MD |
237 | * model easier to understand. It does not have a big performance impact |
238 | * anyway, given this is the write-side. | |
239 | */ | |
5481ddb3 | 240 | cmm_smp_mb(); |
2dfb8b5e | 241 | |
40e140c9 | 242 | /* |
02be5561 | 243 | * Wait for each thread rcu_reader.ctr count to become 0. |
27b012e2 | 244 | */ |
cfe78e25 MD |
245 | for (;;) { |
246 | wait_loops++; | |
247 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { | |
248 | uatomic_dec(&gp_futex); | |
249 | /* Write futex before read reader_gp */ | |
25cc6d18 | 250 | smp_mb_master(RCU_MB_GROUP); |
cfe78e25 MD |
251 | } |
252 | ||
16aa9ee8 | 253 | cds_list_for_each_entry_safe(index, tmp, ®istry, node) { |
b95a001f | 254 | if (!rcu_gp_ongoing(&index->ctr)) |
16aa9ee8 | 255 | cds_list_move(&index->node, &qsreaders); |
cfe78e25 MD |
256 | } |
257 | ||
e8043c1b | 258 | #ifndef HAS_INCOHERENT_CACHES |
16aa9ee8 | 259 | if (cds_list_empty(®istry)) { |
cfe78e25 MD |
260 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { |
261 | /* Read reader_gp before write futex */ | |
25cc6d18 | 262 | smp_mb_master(RCU_MB_GROUP); |
cfe78e25 | 263 | uatomic_set(&gp_futex, 0); |
bc6c15bb | 264 | } |
cfe78e25 MD |
265 | break; |
266 | } else { | |
267 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) | |
268 | wait_gp(); | |
269 | else | |
06f22bdb | 270 | caa_cpu_relax(); |
bc6c15bb | 271 | } |
e8043c1b | 272 | #else /* #ifndef HAS_INCOHERENT_CACHES */ |
27b012e2 | 273 | /* |
40e140c9 | 274 | * BUSY-LOOP. Force the reader thread to commit its |
02be5561 | 275 | * rcu_reader.ctr update to memory if we wait for too long. |
27b012e2 | 276 | */ |
16aa9ee8 | 277 | if (cds_list_empty(®istry)) { |
cfe78e25 MD |
278 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { |
279 | /* Read reader_gp before write futex */ | |
25cc6d18 | 280 | smp_mb_master(RCU_MB_GROUP); |
cfe78e25 MD |
281 | uatomic_set(&gp_futex, 0); |
282 | } | |
283 | break; | |
284 | } else { | |
285 | switch (wait_loops) { | |
bc6c15bb | 286 | case RCU_QS_ACTIVE_ATTEMPTS: |
cfe78e25 MD |
287 | wait_gp(); |
288 | break; /* only escape switch */ | |
bc6c15bb | 289 | case KICK_READER_LOOPS: |
25cc6d18 | 290 | smp_mb_master(RCU_MB_GROUP); |
40e140c9 | 291 | wait_loops = 0; |
cfe78e25 | 292 | break; /* only escape switch */ |
bc6c15bb | 293 | default: |
06f22bdb | 294 | caa_cpu_relax(); |
40e140c9 MD |
295 | } |
296 | } | |
e8043c1b | 297 | #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ |
27b012e2 | 298 | } |
cfe78e25 | 299 | /* put back the reader list in the registry */ |
16aa9ee8 | 300 | cds_list_splice(&qsreaders, ®istry); |
27b012e2 MD |
301 | } |
302 | ||
9598a481 | 303 | void synchronize_rcu(void) |
2bc59bd7 | 304 | { |
6abb4bd5 | 305 | mutex_lock(&rcu_gp_lock); |
135530fd | 306 | |
16aa9ee8 | 307 | if (cds_list_empty(®istry)) |
2dfb8b5e MD |
308 | goto out; |
309 | ||
9598a481 | 310 | /* All threads should read qparity before accessing data structure |
6abb4bd5 MD |
311 | * where new ptr points to. Must be done within rcu_gp_lock because it |
312 | * iterates on reader threads.*/ | |
9598a481 | 313 | /* Write new ptr before changing the qparity */ |
25cc6d18 | 314 | smp_mb_master(RCU_MB_GROUP); |
9598a481 | 315 | |
9598a481 MD |
316 | /* |
317 | * Wait for previous parity to be empty of readers. | |
318 | */ | |
2dfb8b5e | 319 | update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */ |
9598a481 MD |
320 | |
321 | /* | |
322 | * Must finish waiting for quiescent state for parity 0 before | |
d40fde2c MD |
323 | * committing next rcu_gp_ctr update to memory. Failure to do so could |
324 | * result in the writer waiting forever while new readers are always | |
325 | * accessing data (no progress). Enforce compiler-order of load | |
326 | * rcu_reader ctr before store to rcu_gp_ctr. | |
9598a481 | 327 | */ |
5481ddb3 | 328 | cmm_barrier(); |
9598a481 | 329 | |
5dba80f9 | 330 | /* |
5481ddb3 | 331 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
5dba80f9 MD |
332 | * model easier to understand. It does not have a big performance impact |
333 | * anyway, given this is the write-side. | |
334 | */ | |
5481ddb3 | 335 | cmm_smp_mb(); |
67c2d80b | 336 | |
9598a481 MD |
337 | /* |
338 | * Wait for previous parity to be empty of readers. | |
339 | */ | |
2dfb8b5e | 340 | update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */ |
9598a481 | 341 | |
9598a481 | 342 | /* Finish waiting for reader threads before letting the old ptr being |
6abb4bd5 MD |
343 | * freed. Must be done within rcu_gp_lock because it iterates on reader |
344 | * threads. */ | |
25cc6d18 | 345 | smp_mb_master(RCU_MB_GROUP); |
2dfb8b5e | 346 | out: |
6abb4bd5 | 347 | mutex_unlock(&rcu_gp_lock); |
2bc59bd7 PM |
348 | } |
349 | ||
121a5d44 MD |
350 | /* |
351 | * library wrappers to be used by non-LGPL compatible source code. | |
352 | */ | |
353 | ||
354 | void rcu_read_lock(void) | |
355 | { | |
356 | _rcu_read_lock(); | |
357 | } | |
358 | ||
359 | void rcu_read_unlock(void) | |
360 | { | |
361 | _rcu_read_unlock(); | |
362 | } | |
363 | ||
121a5d44 | 364 | void rcu_register_thread(void) |
27b012e2 | 365 | { |
02be5561 MD |
366 | rcu_reader.tid = pthread_self(); |
367 | assert(rcu_reader.need_mb == 0); | |
4b5be3be | 368 | assert(!(rcu_reader.ctr & RCU_GP_CTR_NEST_MASK)); |
02be5561 | 369 | |
6abb4bd5 | 370 | mutex_lock(&rcu_gp_lock); |
02be5561 | 371 | rcu_init(); /* In case gcc does not support constructor attribute */ |
16aa9ee8 | 372 | cds_list_add(&rcu_reader.node, ®istry); |
6abb4bd5 | 373 | mutex_unlock(&rcu_gp_lock); |
27b012e2 MD |
374 | } |
375 | ||
121a5d44 | 376 | void rcu_unregister_thread(void) |
27b012e2 | 377 | { |
6abb4bd5 | 378 | mutex_lock(&rcu_gp_lock); |
16aa9ee8 | 379 | cds_list_del(&rcu_reader.node); |
6abb4bd5 | 380 | mutex_unlock(&rcu_gp_lock); |
27b012e2 MD |
381 | } |
382 | ||
fdf01eed MD |
383 | #ifdef RCU_MEMBARRIER |
384 | void rcu_init(void) | |
385 | { | |
386 | if (init_done) | |
387 | return; | |
388 | init_done = 1; | |
cf5271ee | 389 | if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY)) |
fdf01eed MD |
390 | has_sys_membarrier = 1; |
391 | } | |
392 | #endif | |
393 | ||
394 | #ifdef RCU_SIGNAL | |
02be5561 | 395 | static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context) |
27b012e2 | 396 | { |
40e140c9 | 397 | /* |
5481ddb3 DG |
398 | * Executing this cmm_smp_mb() is the only purpose of this signal handler. |
399 | * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is | |
40e140c9 MD |
400 | * executed on. |
401 | */ | |
5481ddb3 | 402 | cmm_smp_mb(); |
6cf3827c | 403 | _CMM_STORE_SHARED(rcu_reader.need_mb, 0); |
5481ddb3 | 404 | cmm_smp_mb(); |
27b012e2 MD |
405 | } |
406 | ||
8a5fb4c9 | 407 | /* |
02be5561 | 408 | * rcu_init constructor. Called when the library is linked, but also when |
8a5fb4c9 MD |
409 | * reader threads are calling rcu_register_thread(). |
410 | * Should only be called by a single thread at a given time. This is ensured by | |
6abb4bd5 MD |
411 | * holing the rcu_gp_lock from rcu_register_thread() or by running at library |
412 | * load time, which should not be executed by multiple threads nor concurrently | |
413 | * with rcu_register_thread() anyway. | |
8a5fb4c9 | 414 | */ |
02be5561 | 415 | void rcu_init(void) |
27b012e2 MD |
416 | { |
417 | struct sigaction act; | |
418 | int ret; | |
419 | ||
8a5fb4c9 MD |
420 | if (init_done) |
421 | return; | |
422 | init_done = 1; | |
423 | ||
02be5561 | 424 | act.sa_sigaction = sigrcu_handler; |
dd052bd3 | 425 | act.sa_flags = SA_SIGINFO | SA_RESTART; |
c297c21c | 426 | sigemptyset(&act.sa_mask); |
02be5561 | 427 | ret = sigaction(SIGRCU, &act, NULL); |
f69f195a MD |
428 | if (ret) { |
429 | perror("Error in sigaction"); | |
27b012e2 MD |
430 | exit(-1); |
431 | } | |
432 | } | |
433 | ||
02be5561 | 434 | void rcu_exit(void) |
27b012e2 MD |
435 | { |
436 | struct sigaction act; | |
437 | int ret; | |
438 | ||
02be5561 | 439 | ret = sigaction(SIGRCU, NULL, &act); |
f69f195a MD |
440 | if (ret) { |
441 | perror("Error in sigaction"); | |
27b012e2 MD |
442 | exit(-1); |
443 | } | |
02be5561 | 444 | assert(act.sa_sigaction == sigrcu_handler); |
16aa9ee8 | 445 | assert(cds_list_empty(®istry)); |
27b012e2 | 446 | } |
5e77fc1f | 447 | |
fdf01eed | 448 | #endif /* #ifdef RCU_SIGNAL */ |
5e77fc1f PM |
449 | |
450 | #include "urcu-call-rcu-impl.h" | |
0376e7b2 | 451 | #include "urcu-defer-impl.h" |