| 1 | /* |
| 2 | * urcu-qsbr.c |
| 3 | * |
| 4 | * Userspace RCU QSBR 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 | #include <stdio.h> |
| 27 | #include <pthread.h> |
| 28 | #include <signal.h> |
| 29 | #include <assert.h> |
| 30 | #include <stdlib.h> |
| 31 | #include <string.h> |
| 32 | #include <errno.h> |
| 33 | #include <poll.h> |
| 34 | |
| 35 | #define BUILD_QSBR_LIB |
| 36 | #include "urcu-qsbr-static.h" |
| 37 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
| 38 | #include "urcu-qsbr.h" |
| 39 | |
| 40 | static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; |
| 41 | |
| 42 | /* |
| 43 | * Global grace period counter. |
| 44 | */ |
| 45 | unsigned long urcu_gp_ctr = RCU_GP_ONLINE; |
| 46 | |
| 47 | /* |
| 48 | * Written to only by each individual reader. Read by both the reader and the |
| 49 | * writers. |
| 50 | */ |
| 51 | unsigned long __thread rcu_reader_qs_gp; |
| 52 | |
| 53 | /* Thread IDs of registered readers */ |
| 54 | #define INIT_NUM_THREADS 4 |
| 55 | |
| 56 | struct reader_registry { |
| 57 | pthread_t tid; |
| 58 | unsigned long *rcu_reader_qs_gp; |
| 59 | }; |
| 60 | |
| 61 | #ifdef DEBUG_YIELD |
| 62 | unsigned int yield_active; |
| 63 | unsigned int __thread rand_yield; |
| 64 | #endif |
| 65 | |
| 66 | static struct reader_registry *registry; |
| 67 | static int num_readers, alloc_readers; |
| 68 | |
| 69 | static void internal_urcu_lock(void) |
| 70 | { |
| 71 | int ret; |
| 72 | |
| 73 | #ifndef DISTRUST_SIGNALS_EXTREME |
| 74 | ret = pthread_mutex_lock(&urcu_mutex); |
| 75 | if (ret) { |
| 76 | perror("Error in pthread mutex lock"); |
| 77 | exit(-1); |
| 78 | } |
| 79 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ |
| 80 | while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) { |
| 81 | if (ret != EBUSY && ret != EINTR) { |
| 82 | printf("ret = %d, errno = %d\n", ret, errno); |
| 83 | perror("Error in pthread mutex lock"); |
| 84 | exit(-1); |
| 85 | } |
| 86 | poll(NULL,0,10); |
| 87 | } |
| 88 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ |
| 89 | } |
| 90 | |
| 91 | static void internal_urcu_unlock(void) |
| 92 | { |
| 93 | int ret; |
| 94 | |
| 95 | ret = pthread_mutex_unlock(&urcu_mutex); |
| 96 | if (ret) { |
| 97 | perror("Error in pthread mutex unlock"); |
| 98 | exit(-1); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | static void wait_for_quiescent_state(void) |
| 103 | { |
| 104 | struct reader_registry *index; |
| 105 | |
| 106 | if (!registry) |
| 107 | return; |
| 108 | /* |
| 109 | * Wait for each thread rcu_reader_qs_gp count to become 0. |
| 110 | */ |
| 111 | for (index = registry; index < registry + num_readers; index++) { |
| 112 | #ifndef HAS_INCOHERENT_CACHES |
| 113 | while (rcu_gp_ongoing(index->rcu_reader_qs_gp)) |
| 114 | cpu_relax(); |
| 115 | #else /* #ifndef HAS_INCOHERENT_CACHES */ |
| 116 | while (rcu_gp_ongoing(index->rcu_reader_qs_gp)) |
| 117 | smp_mb(); |
| 118 | #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | /* |
| 123 | * Using a two-subphases algorithm for architectures with smaller than 64-bit |
| 124 | * long-size to ensure we do not encounter an overflow bug. |
| 125 | */ |
| 126 | |
| 127 | #if (BITS_PER_LONG < 64) |
| 128 | /* |
| 129 | * called with urcu_mutex held. |
| 130 | */ |
| 131 | static void switch_next_urcu_qparity(void) |
| 132 | { |
| 133 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR); |
| 134 | } |
| 135 | |
| 136 | void synchronize_rcu(void) |
| 137 | { |
| 138 | unsigned long was_online; |
| 139 | |
| 140 | was_online = rcu_reader_qs_gp; |
| 141 | |
| 142 | /* All threads should read qparity before accessing data structure |
| 143 | * where new ptr points to. |
| 144 | */ |
| 145 | /* Write new ptr before changing the qparity */ |
| 146 | smp_mb(); |
| 147 | |
| 148 | /* |
| 149 | * Mark the writer thread offline to make sure we don't wait for |
| 150 | * our own quiescent state. This allows using synchronize_rcu() in |
| 151 | * threads registered as readers. |
| 152 | */ |
| 153 | if (was_online) |
| 154 | STORE_SHARED(rcu_reader_qs_gp, 0); |
| 155 | |
| 156 | internal_urcu_lock(); |
| 157 | |
| 158 | switch_next_urcu_qparity(); /* 0 -> 1 */ |
| 159 | |
| 160 | /* |
| 161 | * Must commit qparity update to memory before waiting for parity |
| 162 | * 0 quiescent state. Failure to do so could result in the writer |
| 163 | * waiting forever while new readers are always accessing data (no |
| 164 | * progress). |
| 165 | * Ensured by STORE_SHARED and LOAD_SHARED. |
| 166 | */ |
| 167 | |
| 168 | /* |
| 169 | * Wait for previous parity to be empty of readers. |
| 170 | */ |
| 171 | wait_for_quiescent_state(); /* Wait readers in parity 0 */ |
| 172 | |
| 173 | /* |
| 174 | * Must finish waiting for quiescent state for parity 0 before |
| 175 | * committing qparity update to memory. Failure to do so could result in |
| 176 | * the writer waiting forever while new readers are always accessing |
| 177 | * data (no progress). |
| 178 | * Ensured by STORE_SHARED and LOAD_SHARED. |
| 179 | */ |
| 180 | |
| 181 | switch_next_urcu_qparity(); /* 1 -> 0 */ |
| 182 | |
| 183 | /* |
| 184 | * Must commit qparity update to memory before waiting for parity |
| 185 | * 1 quiescent state. Failure to do so could result in the writer |
| 186 | * waiting forever while new readers are always accessing data (no |
| 187 | * progress). |
| 188 | * Ensured by STORE_SHARED and LOAD_SHARED. |
| 189 | */ |
| 190 | |
| 191 | /* |
| 192 | * Wait for previous parity to be empty of readers. |
| 193 | */ |
| 194 | wait_for_quiescent_state(); /* Wait readers in parity 1 */ |
| 195 | |
| 196 | internal_urcu_unlock(); |
| 197 | |
| 198 | /* |
| 199 | * Finish waiting for reader threads before letting the old ptr being |
| 200 | * freed. |
| 201 | */ |
| 202 | if (was_online) |
| 203 | _STORE_SHARED(rcu_reader_qs_gp, LOAD_SHARED(urcu_gp_ctr)); |
| 204 | smp_mb(); |
| 205 | } |
| 206 | #else /* !(BITS_PER_LONG < 64) */ |
| 207 | void synchronize_rcu(void) |
| 208 | { |
| 209 | unsigned long was_online; |
| 210 | |
| 211 | was_online = rcu_reader_qs_gp; |
| 212 | |
| 213 | /* |
| 214 | * Mark the writer thread offline to make sure we don't wait for |
| 215 | * our own quiescent state. This allows using synchronize_rcu() in |
| 216 | * threads registered as readers. |
| 217 | */ |
| 218 | smp_mb(); |
| 219 | if (was_online) |
| 220 | STORE_SHARED(rcu_reader_qs_gp, 0); |
| 221 | |
| 222 | internal_urcu_lock(); |
| 223 | STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_CTR); |
| 224 | wait_for_quiescent_state(); |
| 225 | internal_urcu_unlock(); |
| 226 | |
| 227 | if (was_online) |
| 228 | _STORE_SHARED(rcu_reader_qs_gp, LOAD_SHARED(urcu_gp_ctr)); |
| 229 | smp_mb(); |
| 230 | } |
| 231 | #endif /* !(BITS_PER_LONG < 64) */ |
| 232 | |
| 233 | /* |
| 234 | * library wrappers to be used by non-LGPL compatible source code. |
| 235 | */ |
| 236 | |
| 237 | void rcu_read_lock(void) |
| 238 | { |
| 239 | _rcu_read_lock(); |
| 240 | } |
| 241 | |
| 242 | void rcu_read_unlock(void) |
| 243 | { |
| 244 | _rcu_read_unlock(); |
| 245 | } |
| 246 | |
| 247 | void *rcu_dereference(void *p) |
| 248 | { |
| 249 | return _rcu_dereference(p); |
| 250 | } |
| 251 | |
| 252 | void *rcu_assign_pointer_sym(void **p, void *v) |
| 253 | { |
| 254 | wmb(); |
| 255 | return STORE_SHARED(p, v); |
| 256 | } |
| 257 | |
| 258 | void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new) |
| 259 | { |
| 260 | wmb(); |
| 261 | return cmpxchg(p, old, _new); |
| 262 | } |
| 263 | |
| 264 | void *rcu_xchg_pointer_sym(void **p, void *v) |
| 265 | { |
| 266 | wmb(); |
| 267 | return xchg(p, v); |
| 268 | } |
| 269 | |
| 270 | void *rcu_publish_content_sym(void **p, void *v) |
| 271 | { |
| 272 | void *oldptr; |
| 273 | |
| 274 | oldptr = _rcu_xchg_pointer(p, v); |
| 275 | synchronize_rcu(); |
| 276 | return oldptr; |
| 277 | } |
| 278 | |
| 279 | void rcu_quiescent_state(void) |
| 280 | { |
| 281 | _rcu_quiescent_state(); |
| 282 | } |
| 283 | |
| 284 | void rcu_thread_offline(void) |
| 285 | { |
| 286 | _rcu_thread_offline(); |
| 287 | } |
| 288 | |
| 289 | void rcu_thread_online(void) |
| 290 | { |
| 291 | _rcu_thread_online(); |
| 292 | } |
| 293 | |
| 294 | static void rcu_add_reader(pthread_t id) |
| 295 | { |
| 296 | struct reader_registry *oldarray; |
| 297 | |
| 298 | if (!registry) { |
| 299 | alloc_readers = INIT_NUM_THREADS; |
| 300 | num_readers = 0; |
| 301 | registry = |
| 302 | malloc(sizeof(struct reader_registry) * alloc_readers); |
| 303 | } |
| 304 | if (alloc_readers < num_readers + 1) { |
| 305 | oldarray = registry; |
| 306 | registry = malloc(sizeof(struct reader_registry) |
| 307 | * (alloc_readers << 1)); |
| 308 | memcpy(registry, oldarray, |
| 309 | sizeof(struct reader_registry) * alloc_readers); |
| 310 | alloc_readers <<= 1; |
| 311 | free(oldarray); |
| 312 | } |
| 313 | registry[num_readers].tid = id; |
| 314 | /* reference to the TLS of _this_ reader thread. */ |
| 315 | registry[num_readers].rcu_reader_qs_gp = &rcu_reader_qs_gp; |
| 316 | num_readers++; |
| 317 | } |
| 318 | |
| 319 | /* |
| 320 | * Never shrink (implementation limitation). |
| 321 | * This is O(nb threads). Eventually use a hash table. |
| 322 | */ |
| 323 | static void rcu_remove_reader(pthread_t id) |
| 324 | { |
| 325 | struct reader_registry *index; |
| 326 | |
| 327 | assert(registry != NULL); |
| 328 | for (index = registry; index < registry + num_readers; index++) { |
| 329 | if (pthread_equal(index->tid, id)) { |
| 330 | memcpy(index, ®istry[num_readers - 1], |
| 331 | sizeof(struct reader_registry)); |
| 332 | registry[num_readers - 1].tid = 0; |
| 333 | registry[num_readers - 1].rcu_reader_qs_gp = NULL; |
| 334 | num_readers--; |
| 335 | return; |
| 336 | } |
| 337 | } |
| 338 | /* Hrm not found, forgot to register ? */ |
| 339 | assert(0); |
| 340 | } |
| 341 | |
| 342 | void rcu_register_thread(void) |
| 343 | { |
| 344 | internal_urcu_lock(); |
| 345 | rcu_add_reader(pthread_self()); |
| 346 | internal_urcu_unlock(); |
| 347 | _rcu_thread_online(); |
| 348 | } |
| 349 | |
| 350 | void rcu_unregister_thread(void) |
| 351 | { |
| 352 | /* |
| 353 | * We have to make the thread offline otherwise we end up dealocking |
| 354 | * with a waiting writer. |
| 355 | */ |
| 356 | _rcu_thread_offline(); |
| 357 | internal_urcu_lock(); |
| 358 | rcu_remove_reader(pthread_self()); |
| 359 | internal_urcu_unlock(); |
| 360 | } |