| 1 | /* |
| 2 | * urcu.c |
| 3 | * |
| 4 | * Userspace RCU library |
| 5 | * |
| 6 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> |
| 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 | |
| 26 | #define _BSD_SOURCE |
| 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 | |
| 36 | #include "urcu-static.h" |
| 37 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
| 38 | #include "urcu.h" |
| 39 | |
| 40 | #ifdef RCU_MEMBARRIER |
| 41 | static int init_done; |
| 42 | int has_sys_membarrier; |
| 43 | |
| 44 | void __attribute__((constructor)) rcu_init(void); |
| 45 | #endif |
| 46 | |
| 47 | #ifdef RCU_MB |
| 48 | void rcu_init(void) |
| 49 | { |
| 50 | } |
| 51 | #endif |
| 52 | |
| 53 | #ifdef RCU_SIGNAL |
| 54 | static int init_done; |
| 55 | |
| 56 | void __attribute__((constructor)) rcu_init(void); |
| 57 | void __attribute__((destructor)) rcu_exit(void); |
| 58 | #endif |
| 59 | |
| 60 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
| 61 | |
| 62 | int gp_futex; |
| 63 | |
| 64 | /* |
| 65 | * Global grace period counter. |
| 66 | * Contains the current RCU_GP_CTR_PHASE. |
| 67 | * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. |
| 68 | * Written to only by writer with mutex taken. Read by both writer and readers. |
| 69 | */ |
| 70 | long rcu_gp_ctr = RCU_GP_COUNT; |
| 71 | |
| 72 | /* |
| 73 | * Written to only by each individual reader. Read by both the reader and the |
| 74 | * writers. |
| 75 | */ |
| 76 | struct rcu_reader __thread rcu_reader; |
| 77 | |
| 78 | #ifdef DEBUG_YIELD |
| 79 | unsigned int yield_active; |
| 80 | unsigned int __thread rand_yield; |
| 81 | #endif |
| 82 | |
| 83 | static LIST_HEAD(registry); |
| 84 | |
| 85 | static void mutex_lock(pthread_mutex_t *mutex) |
| 86 | { |
| 87 | int ret; |
| 88 | |
| 89 | #ifndef DISTRUST_SIGNALS_EXTREME |
| 90 | ret = pthread_mutex_lock(mutex); |
| 91 | if (ret) { |
| 92 | perror("Error in pthread mutex lock"); |
| 93 | exit(-1); |
| 94 | } |
| 95 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ |
| 96 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
| 97 | if (ret != EBUSY && ret != EINTR) { |
| 98 | printf("ret = %d, errno = %d\n", ret, errno); |
| 99 | perror("Error in pthread mutex lock"); |
| 100 | exit(-1); |
| 101 | } |
| 102 | if (rcu_reader.need_mb) { |
| 103 | smp_mb(); |
| 104 | rcu_reader.need_mb = 0; |
| 105 | smp_mb(); |
| 106 | } |
| 107 | poll(NULL,0,10); |
| 108 | } |
| 109 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ |
| 110 | } |
| 111 | |
| 112 | static void mutex_unlock(pthread_mutex_t *mutex) |
| 113 | { |
| 114 | int ret; |
| 115 | |
| 116 | ret = pthread_mutex_unlock(mutex); |
| 117 | if (ret) { |
| 118 | perror("Error in pthread mutex unlock"); |
| 119 | exit(-1); |
| 120 | } |
| 121 | } |
| 122 | |
| 123 | /* |
| 124 | * called with rcu_gp_lock held. |
| 125 | */ |
| 126 | static void switch_next_rcu_qparity(void) |
| 127 | { |
| 128 | STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); |
| 129 | } |
| 130 | |
| 131 | #ifdef RCU_MEMBARRIER |
| 132 | static void smp_mb_heavy(void) |
| 133 | { |
| 134 | if (likely(has_sys_membarrier)) |
| 135 | membarrier(MEMBARRIER_EXPEDITED); |
| 136 | else |
| 137 | smp_mb(); |
| 138 | } |
| 139 | #endif |
| 140 | |
| 141 | #ifdef RCU_MB |
| 142 | static void smp_mb_heavy(void) |
| 143 | { |
| 144 | smp_mb(); |
| 145 | } |
| 146 | #endif |
| 147 | |
| 148 | #ifdef RCU_SIGNAL |
| 149 | static void force_mb_all_readers(void) |
| 150 | { |
| 151 | struct rcu_reader *index; |
| 152 | |
| 153 | /* |
| 154 | * Ask for each threads to execute a smp_mb() so we can consider the |
| 155 | * compiler barriers around rcu read lock as real memory barriers. |
| 156 | */ |
| 157 | if (list_empty(®istry)) |
| 158 | return; |
| 159 | /* |
| 160 | * pthread_kill has a smp_mb(). But beware, we assume it performs |
| 161 | * a cache flush on architectures with non-coherent cache. Let's play |
| 162 | * safe and don't assume anything : we use smp_mc() to make sure the |
| 163 | * cache flush is enforced. |
| 164 | */ |
| 165 | list_for_each_entry(index, ®istry, head) { |
| 166 | index->need_mb = 1; |
| 167 | smp_mc(); /* write need_mb before sending the signal */ |
| 168 | pthread_kill(index->tid, SIGRCU); |
| 169 | } |
| 170 | /* |
| 171 | * Wait for sighandler (and thus mb()) to execute on every thread. |
| 172 | * |
| 173 | * Note that the pthread_kill() will never be executed on systems |
| 174 | * that correctly deliver signals in a timely manner. However, it |
| 175 | * is not uncommon for kernels to have bugs that can result in |
| 176 | * lost or unduly delayed signals. |
| 177 | * |
| 178 | * If you are seeing the below pthread_kill() executing much at |
| 179 | * all, we suggest testing the underlying kernel and filing the |
| 180 | * relevant bug report. For Linux kernels, we recommend getting |
| 181 | * the Linux Test Project (LTP). |
| 182 | */ |
| 183 | list_for_each_entry(index, ®istry, head) { |
| 184 | while (index->need_mb) { |
| 185 | pthread_kill(index->tid, SIGRCU); |
| 186 | poll(NULL, 0, 1); |
| 187 | } |
| 188 | } |
| 189 | smp_mb(); /* read ->need_mb before ending the barrier */ |
| 190 | } |
| 191 | |
| 192 | static void smp_mb_heavy(void) |
| 193 | { |
| 194 | force_mb_all_readers(); |
| 195 | } |
| 196 | #endif /* #ifdef RCU_SIGNAL */ |
| 197 | |
| 198 | /* |
| 199 | * synchronize_rcu() waiting. Single thread. |
| 200 | */ |
| 201 | static void wait_gp(void) |
| 202 | { |
| 203 | /* Read reader_gp before read futex */ |
| 204 | smp_mb_heavy(); |
| 205 | if (uatomic_read(&gp_futex) == -1) |
| 206 | futex_async(&gp_futex, FUTEX_WAIT, -1, |
| 207 | NULL, NULL, 0); |
| 208 | } |
| 209 | |
| 210 | void wait_for_quiescent_state(void) |
| 211 | { |
| 212 | LIST_HEAD(qsreaders); |
| 213 | int wait_loops = 0; |
| 214 | struct rcu_reader *index, *tmp; |
| 215 | |
| 216 | if (list_empty(®istry)) |
| 217 | return; |
| 218 | /* |
| 219 | * Wait for each thread rcu_reader.ctr count to become 0. |
| 220 | */ |
| 221 | for (;;) { |
| 222 | wait_loops++; |
| 223 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { |
| 224 | uatomic_dec(&gp_futex); |
| 225 | /* Write futex before read reader_gp */ |
| 226 | smp_mb_heavy(); |
| 227 | } |
| 228 | |
| 229 | list_for_each_entry_safe(index, tmp, ®istry, head) { |
| 230 | if (!rcu_old_gp_ongoing(&index->ctr)) |
| 231 | list_move(&index->head, &qsreaders); |
| 232 | } |
| 233 | |
| 234 | #ifndef HAS_INCOHERENT_CACHES |
| 235 | if (list_empty(®istry)) { |
| 236 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { |
| 237 | /* Read reader_gp before write futex */ |
| 238 | smp_mb_heavy(); |
| 239 | uatomic_set(&gp_futex, 0); |
| 240 | } |
| 241 | break; |
| 242 | } else { |
| 243 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) |
| 244 | wait_gp(); |
| 245 | else |
| 246 | cpu_relax(); |
| 247 | } |
| 248 | #else /* #ifndef HAS_INCOHERENT_CACHES */ |
| 249 | /* |
| 250 | * BUSY-LOOP. Force the reader thread to commit its |
| 251 | * rcu_reader.ctr update to memory if we wait for too long. |
| 252 | */ |
| 253 | if (list_empty(®istry)) { |
| 254 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { |
| 255 | /* Read reader_gp before write futex */ |
| 256 | smp_mb_heavy(); |
| 257 | uatomic_set(&gp_futex, 0); |
| 258 | } |
| 259 | break; |
| 260 | } else { |
| 261 | switch (wait_loops) { |
| 262 | case RCU_QS_ACTIVE_ATTEMPTS: |
| 263 | wait_gp(); |
| 264 | break; /* only escape switch */ |
| 265 | case KICK_READER_LOOPS: |
| 266 | smp_mb_heavy(); |
| 267 | wait_loops = 0; |
| 268 | break; /* only escape switch */ |
| 269 | default: |
| 270 | cpu_relax(); |
| 271 | } |
| 272 | } |
| 273 | #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ |
| 274 | } |
| 275 | /* put back the reader list in the registry */ |
| 276 | list_splice(&qsreaders, ®istry); |
| 277 | } |
| 278 | |
| 279 | void synchronize_rcu(void) |
| 280 | { |
| 281 | mutex_lock(&rcu_gp_lock); |
| 282 | |
| 283 | /* All threads should read qparity before accessing data structure |
| 284 | * where new ptr points to. Must be done within rcu_gp_lock because it |
| 285 | * iterates on reader threads.*/ |
| 286 | /* Write new ptr before changing the qparity */ |
| 287 | smp_mb_heavy(); |
| 288 | |
| 289 | switch_next_rcu_qparity(); /* 0 -> 1 */ |
| 290 | |
| 291 | /* |
| 292 | * Must commit qparity update to memory before waiting for parity |
| 293 | * 0 quiescent state. Failure to do so could result in the writer |
| 294 | * waiting forever while new readers are always accessing data (no |
| 295 | * progress). |
| 296 | * Ensured by STORE_SHARED and LOAD_SHARED. |
| 297 | */ |
| 298 | |
| 299 | /* |
| 300 | * Adding a smp_mb() which is _not_ formally required, but makes the |
| 301 | * model easier to understand. It does not have a big performance impact |
| 302 | * anyway, given this is the write-side. |
| 303 | */ |
| 304 | smp_mb(); |
| 305 | |
| 306 | /* |
| 307 | * Wait for previous parity to be empty of readers. |
| 308 | */ |
| 309 | wait_for_quiescent_state(); /* Wait readers in parity 0 */ |
| 310 | |
| 311 | /* |
| 312 | * Must finish waiting for quiescent state for parity 0 before |
| 313 | * committing qparity update to memory. Failure to do so could result in |
| 314 | * the writer waiting forever while new readers are always accessing |
| 315 | * data (no progress). |
| 316 | * Ensured by STORE_SHARED and LOAD_SHARED. |
| 317 | */ |
| 318 | |
| 319 | /* |
| 320 | * Adding a smp_mb() which is _not_ formally required, but makes the |
| 321 | * model easier to understand. It does not have a big performance impact |
| 322 | * anyway, given this is the write-side. |
| 323 | */ |
| 324 | smp_mb(); |
| 325 | |
| 326 | switch_next_rcu_qparity(); /* 1 -> 0 */ |
| 327 | |
| 328 | /* |
| 329 | * Must commit qparity update to memory before waiting for parity |
| 330 | * 1 quiescent state. Failure to do so could result in the writer |
| 331 | * waiting forever while new readers are always accessing data (no |
| 332 | * progress). |
| 333 | * Ensured by STORE_SHARED and LOAD_SHARED. |
| 334 | */ |
| 335 | |
| 336 | /* |
| 337 | * Adding a smp_mb() which is _not_ formally required, but makes the |
| 338 | * model easier to understand. It does not have a big performance impact |
| 339 | * anyway, given this is the write-side. |
| 340 | */ |
| 341 | smp_mb(); |
| 342 | |
| 343 | /* |
| 344 | * Wait for previous parity to be empty of readers. |
| 345 | */ |
| 346 | wait_for_quiescent_state(); /* Wait readers in parity 1 */ |
| 347 | |
| 348 | /* Finish waiting for reader threads before letting the old ptr being |
| 349 | * freed. Must be done within rcu_gp_lock because it iterates on reader |
| 350 | * threads. */ |
| 351 | smp_mb_heavy(); |
| 352 | |
| 353 | mutex_unlock(&rcu_gp_lock); |
| 354 | } |
| 355 | |
| 356 | /* |
| 357 | * library wrappers to be used by non-LGPL compatible source code. |
| 358 | */ |
| 359 | |
| 360 | void rcu_read_lock(void) |
| 361 | { |
| 362 | _rcu_read_lock(); |
| 363 | } |
| 364 | |
| 365 | void rcu_read_unlock(void) |
| 366 | { |
| 367 | _rcu_read_unlock(); |
| 368 | } |
| 369 | |
| 370 | void rcu_register_thread(void) |
| 371 | { |
| 372 | rcu_reader.tid = pthread_self(); |
| 373 | assert(rcu_reader.need_mb == 0); |
| 374 | assert(rcu_reader.ctr == 0); |
| 375 | |
| 376 | mutex_lock(&rcu_gp_lock); |
| 377 | rcu_init(); /* In case gcc does not support constructor attribute */ |
| 378 | list_add(&rcu_reader.head, ®istry); |
| 379 | mutex_unlock(&rcu_gp_lock); |
| 380 | } |
| 381 | |
| 382 | void rcu_unregister_thread(void) |
| 383 | { |
| 384 | mutex_lock(&rcu_gp_lock); |
| 385 | list_del(&rcu_reader.head); |
| 386 | mutex_unlock(&rcu_gp_lock); |
| 387 | } |
| 388 | |
| 389 | #ifdef RCU_MEMBARRIER |
| 390 | void rcu_init(void) |
| 391 | { |
| 392 | if (init_done) |
| 393 | return; |
| 394 | init_done = 1; |
| 395 | if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY)) |
| 396 | has_sys_membarrier = 1; |
| 397 | } |
| 398 | #endif |
| 399 | |
| 400 | #ifdef RCU_SIGNAL |
| 401 | static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context) |
| 402 | { |
| 403 | /* |
| 404 | * Executing this smp_mb() is the only purpose of this signal handler. |
| 405 | * It punctually promotes barrier() into smp_mb() on every thread it is |
| 406 | * executed on. |
| 407 | */ |
| 408 | smp_mb(); |
| 409 | rcu_reader.need_mb = 0; |
| 410 | smp_mb(); |
| 411 | } |
| 412 | |
| 413 | /* |
| 414 | * rcu_init constructor. Called when the library is linked, but also when |
| 415 | * reader threads are calling rcu_register_thread(). |
| 416 | * Should only be called by a single thread at a given time. This is ensured by |
| 417 | * holing the rcu_gp_lock from rcu_register_thread() or by running at library |
| 418 | * load time, which should not be executed by multiple threads nor concurrently |
| 419 | * with rcu_register_thread() anyway. |
| 420 | */ |
| 421 | void rcu_init(void) |
| 422 | { |
| 423 | struct sigaction act; |
| 424 | int ret; |
| 425 | |
| 426 | if (init_done) |
| 427 | return; |
| 428 | init_done = 1; |
| 429 | |
| 430 | act.sa_sigaction = sigrcu_handler; |
| 431 | act.sa_flags = SA_SIGINFO | SA_RESTART; |
| 432 | sigemptyset(&act.sa_mask); |
| 433 | ret = sigaction(SIGRCU, &act, NULL); |
| 434 | if (ret) { |
| 435 | perror("Error in sigaction"); |
| 436 | exit(-1); |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | void rcu_exit(void) |
| 441 | { |
| 442 | struct sigaction act; |
| 443 | int ret; |
| 444 | |
| 445 | ret = sigaction(SIGRCU, NULL, &act); |
| 446 | if (ret) { |
| 447 | perror("Error in sigaction"); |
| 448 | exit(-1); |
| 449 | } |
| 450 | assert(act.sa_sigaction == sigrcu_handler); |
| 451 | assert(list_empty(®istry)); |
| 452 | } |
| 453 | #endif /* #ifdef RCU_SIGNAL */ |