| 1 | /* |
| 2 | * urcu-qsbr.c |
| 3 | * |
| 4 | * Userspace RCU QSBR library |
| 5 | * |
| 6 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 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 _GNU_SOURCE |
| 27 | #define _LGPL_SOURCE |
| 28 | #include <stdio.h> |
| 29 | #include <pthread.h> |
| 30 | #include <signal.h> |
| 31 | #include <assert.h> |
| 32 | #include <stdlib.h> |
| 33 | #include <stdint.h> |
| 34 | #include <string.h> |
| 35 | #include <errno.h> |
| 36 | #include <poll.h> |
| 37 | |
| 38 | #include "urcu/wfqueue.h" |
| 39 | #include "urcu/map/urcu-qsbr.h" |
| 40 | #define BUILD_QSBR_LIB |
| 41 | #include "urcu/static/urcu-qsbr.h" |
| 42 | #include "urcu-pointer.h" |
| 43 | #include "urcu/tls-compat.h" |
| 44 | |
| 45 | #include "urcu-die.h" |
| 46 | |
| 47 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
| 48 | #undef _LGPL_SOURCE |
| 49 | #include "urcu-qsbr.h" |
| 50 | #define _LGPL_SOURCE |
| 51 | |
| 52 | void __attribute__((destructor)) rcu_exit(void); |
| 53 | |
| 54 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
| 55 | |
| 56 | int32_t gp_futex; |
| 57 | |
| 58 | /* |
| 59 | * Global grace period counter. |
| 60 | */ |
| 61 | unsigned long rcu_gp_ctr = RCU_GP_ONLINE; |
| 62 | |
| 63 | /* |
| 64 | * Active attempts to check for reader Q.S. before calling futex(). |
| 65 | */ |
| 66 | #define RCU_QS_ACTIVE_ATTEMPTS 100 |
| 67 | |
| 68 | /* |
| 69 | * Written to only by each individual reader. Read by both the reader and the |
| 70 | * writers. |
| 71 | */ |
| 72 | DEFINE_URCU_TLS(struct rcu_reader, rcu_reader); |
| 73 | |
| 74 | #ifdef DEBUG_YIELD |
| 75 | unsigned int yield_active; |
| 76 | DEFINE_URCU_TLS(unsigned int, rand_yield); |
| 77 | #endif |
| 78 | |
| 79 | static CDS_LIST_HEAD(registry); |
| 80 | |
| 81 | static void mutex_lock(pthread_mutex_t *mutex) |
| 82 | { |
| 83 | int ret; |
| 84 | |
| 85 | #ifndef DISTRUST_SIGNALS_EXTREME |
| 86 | ret = pthread_mutex_lock(mutex); |
| 87 | if (ret) |
| 88 | urcu_die(ret); |
| 89 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ |
| 90 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
| 91 | if (ret != EBUSY && ret != EINTR) |
| 92 | urcu_die(ret); |
| 93 | poll(NULL,0,10); |
| 94 | } |
| 95 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ |
| 96 | } |
| 97 | |
| 98 | static void mutex_unlock(pthread_mutex_t *mutex) |
| 99 | { |
| 100 | int ret; |
| 101 | |
| 102 | ret = pthread_mutex_unlock(mutex); |
| 103 | if (ret) |
| 104 | urcu_die(ret); |
| 105 | } |
| 106 | |
| 107 | /* |
| 108 | * synchronize_rcu() waiting. Single thread. |
| 109 | */ |
| 110 | static void wait_gp(void) |
| 111 | { |
| 112 | /* Read reader_gp before read futex */ |
| 113 | cmm_smp_rmb(); |
| 114 | if (uatomic_read(&gp_futex) == -1) |
| 115 | futex_noasync(&gp_futex, FUTEX_WAIT, -1, |
| 116 | NULL, NULL, 0); |
| 117 | } |
| 118 | |
| 119 | static void update_counter_and_wait(void) |
| 120 | { |
| 121 | CDS_LIST_HEAD(qsreaders); |
| 122 | int wait_loops = 0; |
| 123 | struct rcu_reader *index, *tmp; |
| 124 | |
| 125 | #if (CAA_BITS_PER_LONG < 64) |
| 126 | /* Switch parity: 0 -> 1, 1 -> 0 */ |
| 127 | CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR); |
| 128 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
| 129 | /* Increment current G.P. */ |
| 130 | CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR); |
| 131 | #endif /* !(CAA_BITS_PER_LONG < 64) */ |
| 132 | |
| 133 | /* |
| 134 | * Must commit rcu_gp_ctr update to memory before waiting for |
| 135 | * quiescent state. Failure to do so could result in the writer |
| 136 | * waiting forever while new readers are always accessing data |
| 137 | * (no progress). Enforce compiler-order of store to rcu_gp_ctr |
| 138 | * before load URCU_TLS(rcu_reader).ctr. |
| 139 | */ |
| 140 | cmm_barrier(); |
| 141 | |
| 142 | /* |
| 143 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
| 144 | * model easier to understand. It does not have a big performance impact |
| 145 | * anyway, given this is the write-side. |
| 146 | */ |
| 147 | cmm_smp_mb(); |
| 148 | |
| 149 | /* |
| 150 | * Wait for each thread rcu_reader_qs_gp count to become 0. |
| 151 | */ |
| 152 | for (;;) { |
| 153 | wait_loops++; |
| 154 | if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { |
| 155 | uatomic_set(&gp_futex, -1); |
| 156 | /* |
| 157 | * Write futex before write waiting (the other side |
| 158 | * reads them in the opposite order). |
| 159 | */ |
| 160 | cmm_smp_wmb(); |
| 161 | cds_list_for_each_entry(index, ®istry, node) { |
| 162 | _CMM_STORE_SHARED(index->waiting, 1); |
| 163 | } |
| 164 | /* Write futex before read reader_gp */ |
| 165 | cmm_smp_mb(); |
| 166 | } |
| 167 | cds_list_for_each_entry_safe(index, tmp, ®istry, node) { |
| 168 | if (!rcu_gp_ongoing(&index->ctr)) |
| 169 | cds_list_move(&index->node, &qsreaders); |
| 170 | } |
| 171 | |
| 172 | if (cds_list_empty(®istry)) { |
| 173 | if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { |
| 174 | /* Read reader_gp before write futex */ |
| 175 | cmm_smp_mb(); |
| 176 | uatomic_set(&gp_futex, 0); |
| 177 | } |
| 178 | break; |
| 179 | } else { |
| 180 | if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { |
| 181 | wait_gp(); |
| 182 | } else { |
| 183 | #ifndef HAS_INCOHERENT_CACHES |
| 184 | caa_cpu_relax(); |
| 185 | #else /* #ifndef HAS_INCOHERENT_CACHES */ |
| 186 | cmm_smp_mb(); |
| 187 | #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ |
| 188 | } |
| 189 | } |
| 190 | } |
| 191 | /* put back the reader list in the registry */ |
| 192 | cds_list_splice(&qsreaders, ®istry); |
| 193 | } |
| 194 | |
| 195 | /* |
| 196 | * Using a two-subphases algorithm for architectures with smaller than 64-bit |
| 197 | * long-size to ensure we do not encounter an overflow bug. |
| 198 | */ |
| 199 | |
| 200 | #if (CAA_BITS_PER_LONG < 64) |
| 201 | void synchronize_rcu(void) |
| 202 | { |
| 203 | unsigned long was_online; |
| 204 | |
| 205 | was_online = URCU_TLS(rcu_reader).ctr; |
| 206 | |
| 207 | /* All threads should read qparity before accessing data structure |
| 208 | * where new ptr points to. In the "then" case, rcu_thread_offline |
| 209 | * includes a memory barrier. |
| 210 | * |
| 211 | * Mark the writer thread offline to make sure we don't wait for |
| 212 | * our own quiescent state. This allows using synchronize_rcu() |
| 213 | * in threads registered as readers. |
| 214 | */ |
| 215 | if (was_online) |
| 216 | rcu_thread_offline(); |
| 217 | else |
| 218 | cmm_smp_mb(); |
| 219 | |
| 220 | mutex_lock(&rcu_gp_lock); |
| 221 | |
| 222 | if (cds_list_empty(®istry)) |
| 223 | goto out; |
| 224 | |
| 225 | /* |
| 226 | * Wait for previous parity to be empty of readers. |
| 227 | */ |
| 228 | update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */ |
| 229 | |
| 230 | /* |
| 231 | * Must finish waiting for quiescent state for parity 0 before |
| 232 | * committing next rcu_gp_ctr update to memory. Failure to |
| 233 | * do so could result in the writer waiting forever while new |
| 234 | * readers are always accessing data (no progress). Enforce |
| 235 | * compiler-order of load URCU_TLS(rcu_reader).ctr before store to |
| 236 | * rcu_gp_ctr. |
| 237 | */ |
| 238 | cmm_barrier(); |
| 239 | |
| 240 | /* |
| 241 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
| 242 | * model easier to understand. It does not have a big performance impact |
| 243 | * anyway, given this is the write-side. |
| 244 | */ |
| 245 | cmm_smp_mb(); |
| 246 | |
| 247 | /* |
| 248 | * Wait for previous parity to be empty of readers. |
| 249 | */ |
| 250 | update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */ |
| 251 | out: |
| 252 | mutex_unlock(&rcu_gp_lock); |
| 253 | |
| 254 | /* |
| 255 | * Finish waiting for reader threads before letting the old ptr being |
| 256 | * freed. |
| 257 | */ |
| 258 | if (was_online) |
| 259 | rcu_thread_online(); |
| 260 | else |
| 261 | cmm_smp_mb(); |
| 262 | } |
| 263 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
| 264 | void synchronize_rcu(void) |
| 265 | { |
| 266 | unsigned long was_online; |
| 267 | |
| 268 | was_online = URCU_TLS(rcu_reader).ctr; |
| 269 | |
| 270 | /* |
| 271 | * Mark the writer thread offline to make sure we don't wait for |
| 272 | * our own quiescent state. This allows using synchronize_rcu() |
| 273 | * in threads registered as readers. |
| 274 | */ |
| 275 | if (was_online) |
| 276 | rcu_thread_offline(); |
| 277 | else |
| 278 | cmm_smp_mb(); |
| 279 | |
| 280 | mutex_lock(&rcu_gp_lock); |
| 281 | if (cds_list_empty(®istry)) |
| 282 | goto out; |
| 283 | update_counter_and_wait(); |
| 284 | out: |
| 285 | mutex_unlock(&rcu_gp_lock); |
| 286 | |
| 287 | if (was_online) |
| 288 | rcu_thread_online(); |
| 289 | else |
| 290 | cmm_smp_mb(); |
| 291 | } |
| 292 | #endif /* !(CAA_BITS_PER_LONG < 64) */ |
| 293 | |
| 294 | /* |
| 295 | * library wrappers to be used by non-LGPL compatible source code. |
| 296 | */ |
| 297 | |
| 298 | void rcu_read_lock(void) |
| 299 | { |
| 300 | _rcu_read_lock(); |
| 301 | } |
| 302 | |
| 303 | void rcu_read_unlock(void) |
| 304 | { |
| 305 | _rcu_read_unlock(); |
| 306 | } |
| 307 | |
| 308 | void rcu_quiescent_state(void) |
| 309 | { |
| 310 | _rcu_quiescent_state(); |
| 311 | } |
| 312 | |
| 313 | void rcu_thread_offline(void) |
| 314 | { |
| 315 | _rcu_thread_offline(); |
| 316 | } |
| 317 | |
| 318 | void rcu_thread_online(void) |
| 319 | { |
| 320 | _rcu_thread_online(); |
| 321 | } |
| 322 | |
| 323 | void rcu_register_thread(void) |
| 324 | { |
| 325 | URCU_TLS(rcu_reader).tid = pthread_self(); |
| 326 | assert(URCU_TLS(rcu_reader).ctr == 0); |
| 327 | |
| 328 | mutex_lock(&rcu_gp_lock); |
| 329 | cds_list_add(&URCU_TLS(rcu_reader).node, ®istry); |
| 330 | mutex_unlock(&rcu_gp_lock); |
| 331 | _rcu_thread_online(); |
| 332 | } |
| 333 | |
| 334 | void rcu_unregister_thread(void) |
| 335 | { |
| 336 | /* |
| 337 | * We have to make the thread offline otherwise we end up dealocking |
| 338 | * with a waiting writer. |
| 339 | */ |
| 340 | _rcu_thread_offline(); |
| 341 | mutex_lock(&rcu_gp_lock); |
| 342 | cds_list_del(&URCU_TLS(rcu_reader).node); |
| 343 | mutex_unlock(&rcu_gp_lock); |
| 344 | } |
| 345 | |
| 346 | void rcu_exit(void) |
| 347 | { |
| 348 | /* |
| 349 | * Assertion disabled because call_rcu threads are now rcu |
| 350 | * readers, and left running at exit. |
| 351 | * assert(cds_list_empty(®istry)); |
| 352 | */ |
| 353 | } |
| 354 | |
| 355 | DEFINE_RCU_FLAVOR(rcu_flavor); |
| 356 | |
| 357 | #include "urcu-call-rcu-impl.h" |
| 358 | #include "urcu-defer-impl.h" |