| 1 | /* |
| 2 | * urcu-bp.c |
| 3 | * |
| 4 | * Userspace RCU library, "bulletproof" version. |
| 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 <string.h> |
| 34 | #include <errno.h> |
| 35 | #include <poll.h> |
| 36 | #include <unistd.h> |
| 37 | #include <sys/mman.h> |
| 38 | |
| 39 | #include "urcu/wfcqueue.h" |
| 40 | #include "urcu/map/urcu-bp.h" |
| 41 | #include "urcu/static/urcu-bp.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-bp.h" |
| 50 | #define _LGPL_SOURCE |
| 51 | |
| 52 | #ifndef MAP_ANONYMOUS |
| 53 | #define MAP_ANONYMOUS MAP_ANON |
| 54 | #endif |
| 55 | |
| 56 | #ifdef __linux__ |
| 57 | static |
| 58 | void *mremap_wrapper(void *old_address, size_t old_size, |
| 59 | size_t new_size, int flags) |
| 60 | { |
| 61 | return mremap(old_address, old_size, new_size, flags); |
| 62 | } |
| 63 | #else |
| 64 | |
| 65 | #define MREMAP_MAYMOVE 1 |
| 66 | #define MREMAP_FIXED 2 |
| 67 | |
| 68 | /* |
| 69 | * mremap wrapper for non-Linux systems. Maps a RW, anonymous private mapping. |
| 70 | * This is not generic. |
| 71 | */ |
| 72 | static |
| 73 | void *mremap_wrapper(void *old_address, size_t old_size, |
| 74 | size_t new_size, int flags) |
| 75 | { |
| 76 | void *new_address; |
| 77 | |
| 78 | assert(flags & MREMAP_MAYMOVE); |
| 79 | assert(!(flags & MREMAP_FIXED)); |
| 80 | new_address = mmap(old_address, new_size, |
| 81 | PROT_READ | PROT_WRITE, |
| 82 | MAP_ANONYMOUS | MAP_PRIVATE, |
| 83 | -1, 0); |
| 84 | if (new_address == MAP_FAILED) |
| 85 | return MAP_FAILED; |
| 86 | if (old_address) { |
| 87 | memcpy(new_address, old_address, old_size); |
| 88 | munmap(old_address, old_size); |
| 89 | } |
| 90 | return new_address; |
| 91 | } |
| 92 | #endif |
| 93 | |
| 94 | /* Sleep delay in us */ |
| 95 | #define RCU_SLEEP_DELAY 1000 |
| 96 | #define ARENA_INIT_ALLOC 16 |
| 97 | |
| 98 | /* |
| 99 | * Active attempts to check for reader Q.S. before calling sleep(). |
| 100 | */ |
| 101 | #define RCU_QS_ACTIVE_ATTEMPTS 100 |
| 102 | |
| 103 | void __attribute__((destructor)) rcu_bp_exit(void); |
| 104 | |
| 105 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
| 106 | |
| 107 | #ifdef DEBUG_YIELD |
| 108 | unsigned int rcu_yield_active; |
| 109 | DEFINE_URCU_TLS(unsigned int, rcu_rand_yield); |
| 110 | #endif |
| 111 | |
| 112 | /* |
| 113 | * Global grace period counter. |
| 114 | * Contains the current RCU_GP_CTR_PHASE. |
| 115 | * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. |
| 116 | * Written to only by writer with mutex taken. Read by both writer and readers. |
| 117 | */ |
| 118 | long rcu_gp_ctr = RCU_GP_COUNT; |
| 119 | |
| 120 | /* |
| 121 | * Pointer to registry elements. Written to only by each individual reader. Read |
| 122 | * by both the reader and the writers. |
| 123 | */ |
| 124 | DEFINE_URCU_TLS(struct rcu_reader *, rcu_reader); |
| 125 | |
| 126 | static CDS_LIST_HEAD(registry); |
| 127 | |
| 128 | struct registry_arena { |
| 129 | void *p; |
| 130 | size_t len; |
| 131 | size_t used; |
| 132 | }; |
| 133 | |
| 134 | static struct registry_arena registry_arena; |
| 135 | |
| 136 | /* Saved fork signal mask, protected by rcu_gp_lock */ |
| 137 | static sigset_t saved_fork_signal_mask; |
| 138 | |
| 139 | static void rcu_gc_registry(void); |
| 140 | |
| 141 | static void mutex_lock(pthread_mutex_t *mutex) |
| 142 | { |
| 143 | int ret; |
| 144 | |
| 145 | #ifndef DISTRUST_SIGNALS_EXTREME |
| 146 | ret = pthread_mutex_lock(mutex); |
| 147 | if (ret) |
| 148 | urcu_die(ret); |
| 149 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ |
| 150 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
| 151 | if (ret != EBUSY && ret != EINTR) |
| 152 | urcu_die(ret); |
| 153 | poll(NULL,0,10); |
| 154 | } |
| 155 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ |
| 156 | } |
| 157 | |
| 158 | static void mutex_unlock(pthread_mutex_t *mutex) |
| 159 | { |
| 160 | int ret; |
| 161 | |
| 162 | ret = pthread_mutex_unlock(mutex); |
| 163 | if (ret) |
| 164 | urcu_die(ret); |
| 165 | } |
| 166 | |
| 167 | static void wait_for_readers(void) |
| 168 | { |
| 169 | CDS_LIST_HEAD(qsreaders); |
| 170 | int wait_loops = 0; |
| 171 | struct rcu_reader *index, *tmp; |
| 172 | |
| 173 | /* |
| 174 | * Wait for each thread URCU_TLS(rcu_reader).ctr to either |
| 175 | * indicate quiescence (not nested), or observe the current |
| 176 | * rcu_gp_ctr value. |
| 177 | */ |
| 178 | for (;;) { |
| 179 | wait_loops++; |
| 180 | cds_list_for_each_entry_safe(index, tmp, ®istry, node) { |
| 181 | if (!rcu_old_gp_ongoing(&index->ctr)) |
| 182 | cds_list_move(&index->node, &qsreaders); |
| 183 | } |
| 184 | |
| 185 | if (cds_list_empty(®istry)) { |
| 186 | break; |
| 187 | } else { |
| 188 | if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) |
| 189 | usleep(RCU_SLEEP_DELAY); |
| 190 | else |
| 191 | caa_cpu_relax(); |
| 192 | } |
| 193 | } |
| 194 | /* put back the reader list in the registry */ |
| 195 | cds_list_splice(&qsreaders, ®istry); |
| 196 | } |
| 197 | |
| 198 | void synchronize_rcu(void) |
| 199 | { |
| 200 | sigset_t newmask, oldmask; |
| 201 | int ret; |
| 202 | |
| 203 | ret = sigemptyset(&newmask); |
| 204 | assert(!ret); |
| 205 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); |
| 206 | assert(!ret); |
| 207 | |
| 208 | mutex_lock(&rcu_gp_lock); |
| 209 | |
| 210 | if (cds_list_empty(®istry)) |
| 211 | goto out; |
| 212 | |
| 213 | /* All threads should read qparity before accessing data structure |
| 214 | * where new ptr points to. */ |
| 215 | /* Write new ptr before changing the qparity */ |
| 216 | cmm_smp_mb(); |
| 217 | |
| 218 | /* Remove old registry elements */ |
| 219 | rcu_gc_registry(); |
| 220 | |
| 221 | /* |
| 222 | * Wait for readers to observe original parity or be quiescent. |
| 223 | */ |
| 224 | wait_for_readers(); |
| 225 | |
| 226 | /* |
| 227 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
| 228 | * model easier to understand. It does not have a big performance impact |
| 229 | * anyway, given this is the write-side. |
| 230 | */ |
| 231 | cmm_smp_mb(); |
| 232 | |
| 233 | /* Switch parity: 0 -> 1, 1 -> 0 */ |
| 234 | CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); |
| 235 | |
| 236 | /* |
| 237 | * Must commit qparity update to memory before waiting for other parity |
| 238 | * quiescent state. Failure to do so could result in the writer waiting |
| 239 | * forever while new readers are always accessing data (no progress). |
| 240 | * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED. |
| 241 | */ |
| 242 | |
| 243 | /* |
| 244 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
| 245 | * model easier to understand. It does not have a big performance impact |
| 246 | * anyway, given this is the write-side. |
| 247 | */ |
| 248 | cmm_smp_mb(); |
| 249 | |
| 250 | /* |
| 251 | * Wait for readers to observe new parity or be quiescent. |
| 252 | */ |
| 253 | wait_for_readers(); |
| 254 | |
| 255 | /* |
| 256 | * Finish waiting for reader threads before letting the old ptr being |
| 257 | * freed. |
| 258 | */ |
| 259 | cmm_smp_mb(); |
| 260 | out: |
| 261 | mutex_unlock(&rcu_gp_lock); |
| 262 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
| 263 | assert(!ret); |
| 264 | } |
| 265 | |
| 266 | /* |
| 267 | * library wrappers to be used by non-LGPL compatible source code. |
| 268 | */ |
| 269 | |
| 270 | void rcu_read_lock(void) |
| 271 | { |
| 272 | _rcu_read_lock(); |
| 273 | } |
| 274 | |
| 275 | void rcu_read_unlock(void) |
| 276 | { |
| 277 | _rcu_read_unlock(); |
| 278 | } |
| 279 | |
| 280 | /* |
| 281 | * only grow for now. |
| 282 | */ |
| 283 | static void resize_arena(struct registry_arena *arena, size_t len) |
| 284 | { |
| 285 | void *new_arena; |
| 286 | |
| 287 | if (!arena->p) |
| 288 | new_arena = mmap(arena->p, len, |
| 289 | PROT_READ | PROT_WRITE, |
| 290 | MAP_ANONYMOUS | MAP_PRIVATE, |
| 291 | -1, 0); |
| 292 | else |
| 293 | new_arena = mremap_wrapper(arena->p, arena->len, |
| 294 | len, MREMAP_MAYMOVE); |
| 295 | assert(new_arena != MAP_FAILED); |
| 296 | |
| 297 | /* |
| 298 | * re-used the same region ? |
| 299 | */ |
| 300 | if (new_arena == arena->p) |
| 301 | return; |
| 302 | |
| 303 | bzero(new_arena + arena->len, len - arena->len); |
| 304 | arena->p = new_arena; |
| 305 | } |
| 306 | |
| 307 | /* Called with signals off and mutex locked */ |
| 308 | static void add_thread(void) |
| 309 | { |
| 310 | struct rcu_reader *rcu_reader_reg; |
| 311 | |
| 312 | if (registry_arena.len |
| 313 | < registry_arena.used + sizeof(struct rcu_reader)) |
| 314 | resize_arena(®istry_arena, |
| 315 | caa_max(registry_arena.len << 1, ARENA_INIT_ALLOC)); |
| 316 | /* |
| 317 | * Find a free spot. |
| 318 | */ |
| 319 | for (rcu_reader_reg = registry_arena.p; |
| 320 | (void *)rcu_reader_reg < registry_arena.p + registry_arena.len; |
| 321 | rcu_reader_reg++) { |
| 322 | if (!rcu_reader_reg->alloc) |
| 323 | break; |
| 324 | } |
| 325 | rcu_reader_reg->alloc = 1; |
| 326 | registry_arena.used += sizeof(struct rcu_reader); |
| 327 | |
| 328 | /* Add to registry */ |
| 329 | rcu_reader_reg->tid = pthread_self(); |
| 330 | assert(rcu_reader_reg->ctr == 0); |
| 331 | cds_list_add(&rcu_reader_reg->node, ®istry); |
| 332 | URCU_TLS(rcu_reader) = rcu_reader_reg; |
| 333 | } |
| 334 | |
| 335 | /* Called with signals off and mutex locked */ |
| 336 | static void rcu_gc_registry(void) |
| 337 | { |
| 338 | struct rcu_reader *rcu_reader_reg; |
| 339 | pthread_t tid; |
| 340 | int ret; |
| 341 | |
| 342 | for (rcu_reader_reg = registry_arena.p; |
| 343 | (void *)rcu_reader_reg < registry_arena.p + registry_arena.len; |
| 344 | rcu_reader_reg++) { |
| 345 | if (!rcu_reader_reg->alloc) |
| 346 | continue; |
| 347 | tid = rcu_reader_reg->tid; |
| 348 | ret = pthread_kill(tid, 0); |
| 349 | assert(ret != EINVAL); |
| 350 | if (ret == ESRCH) { |
| 351 | cds_list_del(&rcu_reader_reg->node); |
| 352 | rcu_reader_reg->ctr = 0; |
| 353 | rcu_reader_reg->alloc = 0; |
| 354 | registry_arena.used -= sizeof(struct rcu_reader); |
| 355 | } |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | /* Disable signals, take mutex, add to registry */ |
| 360 | void rcu_bp_register(void) |
| 361 | { |
| 362 | sigset_t newmask, oldmask; |
| 363 | int ret; |
| 364 | |
| 365 | ret = sigemptyset(&newmask); |
| 366 | assert(!ret); |
| 367 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); |
| 368 | assert(!ret); |
| 369 | |
| 370 | /* |
| 371 | * Check if a signal concurrently registered our thread since |
| 372 | * the check in rcu_read_lock(). */ |
| 373 | if (URCU_TLS(rcu_reader)) |
| 374 | goto end; |
| 375 | |
| 376 | mutex_lock(&rcu_gp_lock); |
| 377 | add_thread(); |
| 378 | mutex_unlock(&rcu_gp_lock); |
| 379 | end: |
| 380 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
| 381 | assert(!ret); |
| 382 | } |
| 383 | |
| 384 | void rcu_bp_exit(void) |
| 385 | { |
| 386 | if (registry_arena.p) |
| 387 | munmap(registry_arena.p, registry_arena.len); |
| 388 | } |
| 389 | |
| 390 | /* |
| 391 | * Holding the rcu_gp_lock across fork will make sure we fork() don't race with |
| 392 | * a concurrent thread executing with this same lock held. This ensures that the |
| 393 | * registry is in a coherent state in the child. |
| 394 | */ |
| 395 | void rcu_bp_before_fork(void) |
| 396 | { |
| 397 | sigset_t newmask, oldmask; |
| 398 | int ret; |
| 399 | |
| 400 | ret = sigemptyset(&newmask); |
| 401 | assert(!ret); |
| 402 | ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); |
| 403 | assert(!ret); |
| 404 | mutex_lock(&rcu_gp_lock); |
| 405 | saved_fork_signal_mask = oldmask; |
| 406 | } |
| 407 | |
| 408 | void rcu_bp_after_fork_parent(void) |
| 409 | { |
| 410 | sigset_t oldmask; |
| 411 | int ret; |
| 412 | |
| 413 | oldmask = saved_fork_signal_mask; |
| 414 | mutex_unlock(&rcu_gp_lock); |
| 415 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
| 416 | assert(!ret); |
| 417 | } |
| 418 | |
| 419 | void rcu_bp_after_fork_child(void) |
| 420 | { |
| 421 | sigset_t oldmask; |
| 422 | int ret; |
| 423 | |
| 424 | rcu_gc_registry(); |
| 425 | oldmask = saved_fork_signal_mask; |
| 426 | mutex_unlock(&rcu_gp_lock); |
| 427 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
| 428 | assert(!ret); |
| 429 | } |
| 430 | |
| 431 | void *rcu_dereference_sym_bp(void *p) |
| 432 | { |
| 433 | return _rcu_dereference(p); |
| 434 | } |
| 435 | |
| 436 | void *rcu_set_pointer_sym_bp(void **p, void *v) |
| 437 | { |
| 438 | cmm_wmb(); |
| 439 | uatomic_set(p, v); |
| 440 | return v; |
| 441 | } |
| 442 | |
| 443 | void *rcu_xchg_pointer_sym_bp(void **p, void *v) |
| 444 | { |
| 445 | cmm_wmb(); |
| 446 | return uatomic_xchg(p, v); |
| 447 | } |
| 448 | |
| 449 | void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new) |
| 450 | { |
| 451 | cmm_wmb(); |
| 452 | return uatomic_cmpxchg(p, old, _new); |
| 453 | } |
| 454 | |
| 455 | DEFINE_RCU_FLAVOR(rcu_flavor); |
| 456 | |
| 457 | #include "urcu-call-rcu-impl.h" |
| 458 | #include "urcu-defer-impl.h" |