urcu-bp.c

   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/wfqueue.h"
  40 #include "urcu/map/urcu-bp.h"
  41 #include "urcu/static/urcu-bp.h"
  42
  43 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  44 #undef _LGPL_SOURCE
  45 #include "urcu-bp.h"
  46 #define _LGPL_SOURCE
  47
  48 #ifndef MAP_ANONYMOUS
  49 #define MAP_ANONYMOUS MAP_ANON
  50 #endif
  51
  52 #ifndef __linux__
  53
  54 #define MREMAP_MAYMOVE  1
  55 #define MREMAP_FIXED    2
  56
  57 /*
  58  * mremap wrapper for non-Linux systems. Maps a RW, anonymous private mapping.
  59  * This is not generic.
  60 */
  61 void *mremap(void *old_address, size_t old_size, size_t new_size, int flags)
  62 {
  63         void *new_address;
  64
  65         assert(flags & MREMAP_MAYMOVE);
  66         assert(!(flags & MREMAP_FIXED));
  67         new_address = mmap(old_address, new_size,
  68                            PROT_READ | PROT_WRITE,
  69                            MAP_ANONYMOUS | MAP_PRIVATE,
  70                            -1, 0);
  71         if (new_address == MAP_FAILED)
  72                 return MAP_FAILED;
  73         if (old_address) {
  74                 memcpy(new_address, old_address, old_size);
  75                 munmap(old_address, old_size);
  76         }
  77         return new_address;
  78 }
  79 #endif
  80
  81 /* Sleep delay in us */
  82 #define RCU_SLEEP_DELAY         1000
  83 #define ARENA_INIT_ALLOC        16
  84
  85 /*
  86  * Active attempts to check for reader Q.S. before calling sleep().
  87  */
  88 #define RCU_QS_ACTIVE_ATTEMPTS 100
  89
  90 void __attribute__((destructor)) rcu_bp_exit(void);
  91
  92 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  93
  94 #ifdef DEBUG_YIELD
  95 unsigned int yield_active;
  96 unsigned int __thread rand_yield;
  97 #endif
  98
  99 /*
 100  * Global grace period counter.
 101  * Contains the current RCU_GP_CTR_PHASE.
 102  * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
 103  * Written to only by writer with mutex taken. Read by both writer and readers.
 104  */
 105 long rcu_gp_ctr = RCU_GP_COUNT;
 106
 107 /*
 108  * Pointer to registry elements. Written to only by each individual reader. Read
 109  * by both the reader and the writers.
 110  */
 111 struct rcu_reader __thread *rcu_reader;
 112
 113 static CDS_LIST_HEAD(registry);
 114
 115 struct registry_arena {
 116         void *p;
 117         size_t len;
 118         size_t used;
 119 };
 120
 121 static struct registry_arena registry_arena;
 122
 123 /* Saved fork signal mask, protected by rcu_gp_lock */
 124 static sigset_t saved_fork_signal_mask;
 125
 126 static void rcu_gc_registry(void);
 127
 128 static void mutex_lock(pthread_mutex_t *mutex)
 129 {
 130         int ret;
 131
 132 #ifndef DISTRUST_SIGNALS_EXTREME
 133         ret = pthread_mutex_lock(mutex);
 134         if (ret) {
 135                 perror("Error in pthread mutex lock");
 136                 exit(-1);
 137         }
 138 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 139         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 140                 if (ret != EBUSY && ret != EINTR) {
 141                         printf("ret = %d, errno = %d\n", ret, errno);
 142                         perror("Error in pthread mutex lock");
 143                         exit(-1);
 144                 }
 145                 poll(NULL,0,10);
 146         }
 147 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 148 }
 149
 150 static void mutex_unlock(pthread_mutex_t *mutex)
 151 {
 152         int ret;
 153
 154         ret = pthread_mutex_unlock(mutex);
 155         if (ret) {
 156                 perror("Error in pthread mutex unlock");
 157                 exit(-1);
 158         }
 159 }
 160
 161 void update_counter_and_wait(void)
 162 {
 163         CDS_LIST_HEAD(qsreaders);
 164         int wait_loops = 0;
 165         struct rcu_reader *index, *tmp;
 166
 167         /* Switch parity: 0 -> 1, 1 -> 0 */
 168         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
 169
 170         /*
 171          * Must commit qparity update to memory before waiting for other parity
 172          * quiescent state. Failure to do so could result in the writer waiting
 173          * forever while new readers are always accessing data (no progress).
 174          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 175          */
 176
 177         /*
 178          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 179          * model easier to understand. It does not have a big performance impact
 180          * anyway, given this is the write-side.
 181          */
 182         cmm_smp_mb();
 183
 184         /*
 185          * Wait for each thread rcu_reader.ctr count to become 0.
 186          */
 187         for (;;) {
 188                 wait_loops++;
 189                 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
 190                         if (!rcu_old_gp_ongoing(&index->ctr))
 191                                 cds_list_move(&index->node, &qsreaders);
 192                 }
 193
 194                 if (cds_list_empty(&registry)) {
 195                         break;
 196                 } else {
 197                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 198                                 usleep(RCU_SLEEP_DELAY);
 199                         else
 200                                 caa_cpu_relax();
 201                 }
 202         }
 203         /* put back the reader list in the registry */
 204         cds_list_splice(&qsreaders, &registry);
 205 }
 206
 207 void synchronize_rcu(void)
 208 {
 209         sigset_t newmask, oldmask;
 210         int ret;
 211
 212         ret = sigemptyset(&newmask);
 213         assert(!ret);
 214         ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
 215         assert(!ret);
 216
 217         mutex_lock(&rcu_gp_lock);
 218
 219         if (cds_list_empty(&registry))
 220                 goto out;
 221
 222         /* All threads should read qparity before accessing data structure
 223          * where new ptr points to. */
 224         /* Write new ptr before changing the qparity */
 225         cmm_smp_mb();
 226
 227         /* Remove old registry elements */
 228         rcu_gc_registry();
 229
 230         /*
 231          * Wait for previous parity to be empty of readers.
 232          */
 233         update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 234
 235         /*
 236          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 237          * model easier to understand. It does not have a big performance impact
 238          * anyway, given this is the write-side.
 239          */
 240         cmm_smp_mb();
 241
 242         /*
 243          * Wait for previous parity to be empty of readers.
 244          */
 245         update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
 246
 247         /*
 248          * Finish waiting for reader threads before letting the old ptr being
 249          * freed.
 250          */
 251         cmm_smp_mb();
 252 out:
 253         mutex_unlock(&rcu_gp_lock);
 254         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 255         assert(!ret);
 256 }
 257
 258 /*
 259  * library wrappers to be used by non-LGPL compatible source code.
 260  */
 261
 262 void rcu_read_lock(void)
 263 {
 264         _rcu_read_lock();
 265 }
 266
 267 void rcu_read_unlock(void)
 268 {
 269         _rcu_read_unlock();
 270 }
 271
 272 /*
 273  * only grow for now.
 274  */
 275 static void resize_arena(struct registry_arena *arena, size_t len)
 276 {
 277         void *new_arena;
 278
 279         if (!arena->p)
 280                 new_arena = mmap(arena->p, len,
 281                                  PROT_READ | PROT_WRITE,
 282                                  MAP_ANONYMOUS | MAP_PRIVATE,
 283                                  -1, 0);
 284         else
 285                 new_arena = mremap(arena->p, arena->len,
 286                                    len, MREMAP_MAYMOVE);
 287         assert(new_arena != MAP_FAILED);
 288
 289         /*
 290          * re-used the same region ?
 291          */
 292         if (new_arena == arena->p)
 293                 return;
 294
 295         bzero(new_arena + arena->len, len - arena->len);
 296         arena->p = new_arena;
 297 }
 298
 299 /* Called with signals off and mutex locked */
 300 static void add_thread(void)
 301 {
 302         struct rcu_reader *rcu_reader_reg;
 303
 304         if (registry_arena.len
 305             < registry_arena.used + sizeof(struct rcu_reader))
 306                 resize_arena(&registry_arena,
 307                 caa_max(registry_arena.len << 1, ARENA_INIT_ALLOC));
 308         /*
 309          * Find a free spot.
 310          */
 311         for (rcu_reader_reg = registry_arena.p;
 312              (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
 313              rcu_reader_reg++) {
 314                 if (!rcu_reader_reg->alloc)
 315                         break;
 316         }
 317         rcu_reader_reg->alloc = 1;
 318         registry_arena.used += sizeof(struct rcu_reader);
 319
 320         /* Add to registry */
 321         rcu_reader_reg->tid = pthread_self();
 322         assert(rcu_reader_reg->ctr == 0);
 323         cds_list_add(&rcu_reader_reg->node, &registry);
 324         rcu_reader = rcu_reader_reg;
 325 }
 326
 327 /* Called with signals off and mutex locked */
 328 static void rcu_gc_registry(void)
 329 {
 330         struct rcu_reader *rcu_reader_reg;
 331         pthread_t tid;
 332         int ret;
 333
 334         for (rcu_reader_reg = registry_arena.p;
 335              (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
 336              rcu_reader_reg++) {
 337                 if (!rcu_reader_reg->alloc)
 338                         continue;
 339                 tid = rcu_reader_reg->tid;
 340                 ret = pthread_kill(tid, 0);
 341                 assert(ret != EINVAL);
 342                 if (ret == ESRCH) {
 343                         cds_list_del(&rcu_reader_reg->node);
 344                         rcu_reader_reg->ctr = 0;
 345                         rcu_reader_reg->alloc = 0;
 346                         registry_arena.used -= sizeof(struct rcu_reader);
 347                 }
 348         }
 349 }
 350
 351 /* Disable signals, take mutex, add to registry */
 352 void rcu_bp_register(void)
 353 {
 354         sigset_t newmask, oldmask;
 355         int ret;
 356
 357         ret = sigemptyset(&newmask);
 358         assert(!ret);
 359         ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
 360         assert(!ret);
 361
 362         /*
 363          * Check if a signal concurrently registered our thread since
 364          * the check in rcu_read_lock(). */
 365         if (rcu_reader)
 366                 goto end;
 367
 368         mutex_lock(&rcu_gp_lock);
 369         add_thread();
 370         mutex_unlock(&rcu_gp_lock);
 371 end:
 372         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 373         assert(!ret);
 374 }
 375
 376 void rcu_bp_exit(void)
 377 {
 378         if (registry_arena.p)
 379                 munmap(registry_arena.p, registry_arena.len);
 380 }
 381
 382 /*
 383  * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
 384  * a concurrent thread executing with this same lock held. This ensures that the
 385  * registry is in a coherent state in the child.
 386  */
 387 void rcu_bp_before_fork(void)
 388 {
 389         sigset_t newmask, oldmask;
 390         int ret;
 391
 392         ret = sigemptyset(&newmask);
 393         assert(!ret);
 394         ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
 395         assert(!ret);
 396         mutex_lock(&rcu_gp_lock);
 397         saved_fork_signal_mask = oldmask;
 398 }
 399
 400 void rcu_bp_after_fork_parent(void)
 401 {
 402         sigset_t oldmask;
 403         int ret;
 404
 405         oldmask = saved_fork_signal_mask;
 406         mutex_unlock(&rcu_gp_lock);
 407         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 408         assert(!ret);
 409 }
 410
 411 void rcu_bp_after_fork_child(void)
 412 {
 413         sigset_t oldmask;
 414         int ret;
 415
 416         rcu_gc_registry();
 417         oldmask = saved_fork_signal_mask;
 418         mutex_unlock(&rcu_gp_lock);
 419         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 420         assert(!ret);
 421 }
 422
 423 #include "urcu-call-rcu-impl.h"
 424 #include "urcu-defer-impl.h"