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 #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 #include <unistd.h>
  36 #include <sys/mman.h>
  37
  38 #include "urcu/map/urcu-bp.h"
  39
  40 #include "urcu/static/urcu-bp.h"
  41 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  42 #include "urcu-bp.h"
  43
  44 #ifndef MAP_ANONYMOUS
  45 #define MAP_ANONYMOUS MAP_ANON
  46 #endif
  47
  48 #ifndef __linux__
  49
  50 #define MREMAP_MAYMOVE  1
  51 #define MREMAP_FIXED    2
  52
  53 /*
  54  * mremap wrapper for non-Linux systems. Maps a RW, anonymous private mapping.
  55  * This is not generic.
  56 */
  57 void *mremap(void *old_address, size_t old_size, size_t new_size, int flags)
  58 {
  59         void *new_address;
  60
  61         assert(flags & MREMAP_MAYMOVE);
  62         assert(!(flags & MREMAP_FIXED));
  63         new_address = mmap(old_address, new_size,
  64                            PROT_READ | PROT_WRITE,
  65                            MAP_ANONYMOUS | MAP_PRIVATE,
  66                            -1, 0);
  67         if (new_address == MAP_FAILED)
  68                 return MAP_FAILED;
  69         if (old_address) {
  70                 memcpy(new_address, old_address, old_size);
  71                 munmap(old_address, old_size);
  72         }
  73         return new_address;
  74 }
  75 #endif
  76
  77 /* Sleep delay in us */
  78 #define RCU_SLEEP_DELAY         1000
  79 #define ARENA_INIT_ALLOC        16
  80
  81 /*
  82  * Active attempts to check for reader Q.S. before calling sleep().
  83  */
  84 #define RCU_QS_ACTIVE_ATTEMPTS 100
  85
  86 void __attribute__((destructor)) rcu_bp_exit(void);
  87
  88 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  89
  90 #ifdef DEBUG_YIELD
  91 unsigned int yield_active;
  92 unsigned int __thread rand_yield;
  93 #endif
  94
  95 /*
  96  * Global grace period counter.
  97  * Contains the current RCU_GP_CTR_PHASE.
  98  * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
  99  * Written to only by writer with mutex taken. Read by both writer and readers.
 100  */
 101 long rcu_gp_ctr = RCU_GP_COUNT;
 102
 103 /*
 104  * Pointer to registry elements. Written to only by each individual reader. Read
 105  * by both the reader and the writers.
 106  */
 107 struct rcu_reader __thread *rcu_reader;
 108
 109 static CDS_LIST_HEAD(registry);
 110
 111 struct registry_arena {
 112         void *p;
 113         size_t len;
 114         size_t used;
 115 };
 116
 117 static struct registry_arena registry_arena;
 118
 119 /* Saved fork signal mask, protected by rcu_gp_lock */
 120 static sigset_t saved_fork_signal_mask;
 121
 122 static void rcu_gc_registry(void);
 123
 124 static void mutex_lock(pthread_mutex_t *mutex)
 125 {
 126         int ret;
 127
 128 #ifndef DISTRUST_SIGNALS_EXTREME
 129         ret = pthread_mutex_lock(mutex);
 130         if (ret) {
 131                 perror("Error in pthread mutex lock");
 132                 exit(-1);
 133         }
 134 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 135         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 136                 if (ret != EBUSY && ret != EINTR) {
 137                         printf("ret = %d, errno = %d\n", ret, errno);
 138                         perror("Error in pthread mutex lock");
 139                         exit(-1);
 140                 }
 141                 poll(NULL,0,10);
 142         }
 143 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 144 }
 145
 146 static void mutex_unlock(pthread_mutex_t *mutex)
 147 {
 148         int ret;
 149
 150         ret = pthread_mutex_unlock(mutex);
 151         if (ret) {
 152                 perror("Error in pthread mutex unlock");
 153                 exit(-1);
 154         }
 155 }
 156
 157 void update_counter_and_wait(void)
 158 {
 159         CDS_LIST_HEAD(qsreaders);
 160         int wait_loops = 0;
 161         struct rcu_reader *index, *tmp;
 162
 163         /* Switch parity: 0 -> 1, 1 -> 0 */
 164         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
 165
 166         /*
 167          * Must commit qparity update to memory before waiting for other parity
 168          * quiescent state. Failure to do so could result in the writer waiting
 169          * forever while new readers are always accessing data (no progress).
 170          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 171          */
 172
 173         /*
 174          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 175          * model easier to understand. It does not have a big performance impact
 176          * anyway, given this is the write-side.
 177          */
 178         cmm_smp_mb();
 179
 180         /*
 181          * Wait for each thread rcu_reader.ctr count to become 0.
 182          */
 183         for (;;) {
 184                 wait_loops++;
 185                 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
 186                         if (!rcu_old_gp_ongoing(&index->ctr))
 187                                 cds_list_move(&index->node, &qsreaders);
 188                 }
 189
 190                 if (cds_list_empty(&registry)) {
 191                         break;
 192                 } else {
 193                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 194                                 usleep(RCU_SLEEP_DELAY);
 195                         else
 196                                 caa_cpu_relax();
 197                 }
 198         }
 199         /* put back the reader list in the registry */
 200         cds_list_splice(&qsreaders, &registry);
 201 }
 202
 203 void synchronize_rcu(void)
 204 {
 205         sigset_t newmask, oldmask;
 206         int ret;
 207
 208         ret = sigemptyset(&newmask);
 209         assert(!ret);
 210         ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
 211         assert(!ret);
 212
 213         mutex_lock(&rcu_gp_lock);
 214
 215         if (cds_list_empty(&registry))
 216                 goto out;
 217
 218         /* All threads should read qparity before accessing data structure
 219          * where new ptr points to. */
 220         /* Write new ptr before changing the qparity */
 221         cmm_smp_mb();
 222
 223         /* Remove old registry elements */
 224         rcu_gc_registry();
 225
 226         /*
 227          * Wait for previous parity to be empty of readers.
 228          */
 229         update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 230
 231         /*
 232          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 233          * model easier to understand. It does not have a big performance impact
 234          * anyway, given this is the write-side.
 235          */
 236         cmm_smp_mb();
 237
 238         /*
 239          * Wait for previous parity to be empty of readers.
 240          */
 241         update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
 242
 243         /*
 244          * Finish waiting for reader threads before letting the old ptr being
 245          * freed.
 246          */
 247         cmm_smp_mb();
 248 out:
 249         mutex_unlock(&rcu_gp_lock);
 250         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 251         assert(!ret);
 252 }
 253
 254 /*
 255  * library wrappers to be used by non-LGPL compatible source code.
 256  */
 257
 258 void rcu_read_lock(void)
 259 {
 260         _rcu_read_lock();
 261 }
 262
 263 void rcu_read_unlock(void)
 264 {
 265         _rcu_read_unlock();
 266 }
 267
 268 /*
 269  * only grow for now.
 270  */
 271 static void resize_arena(struct registry_arena *arena, size_t len)
 272 {
 273         void *new_arena;
 274
 275         if (!arena->p)
 276                 new_arena = mmap(arena->p, len,
 277                                  PROT_READ | PROT_WRITE,
 278                                  MAP_ANONYMOUS | MAP_PRIVATE,
 279                                  -1, 0);
 280         else
 281                 new_arena = mremap(arena->p, arena->len,
 282                                    len, MREMAP_MAYMOVE);
 283         assert(new_arena != MAP_FAILED);
 284
 285         /*
 286          * re-used the same region ?
 287          */
 288         if (new_arena == arena->p)
 289                 return;
 290
 291         bzero(new_arena + arena->len, len - arena->len);
 292         arena->p = new_arena;
 293 }
 294
 295 /* Called with signals off and mutex locked */
 296 static void add_thread(void)
 297 {
 298         struct rcu_reader *rcu_reader_reg;
 299
 300         if (registry_arena.len
 301             < registry_arena.used + sizeof(struct rcu_reader))
 302                 resize_arena(&registry_arena,
 303                 caa_max(registry_arena.len << 1, ARENA_INIT_ALLOC));
 304         /*
 305          * Find a free spot.
 306          */
 307         for (rcu_reader_reg = registry_arena.p;
 308              (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
 309              rcu_reader_reg++) {
 310                 if (!rcu_reader_reg->alloc)
 311                         break;
 312         }
 313         rcu_reader_reg->alloc = 1;
 314         registry_arena.used += sizeof(struct rcu_reader);
 315
 316         /* Add to registry */
 317         rcu_reader_reg->tid = pthread_self();
 318         assert(rcu_reader_reg->ctr == 0);
 319         cds_list_add(&rcu_reader_reg->node, &registry);
 320         rcu_reader = rcu_reader_reg;
 321 }
 322
 323 /* Called with signals off and mutex locked */
 324 static void rcu_gc_registry(void)
 325 {
 326         struct rcu_reader *rcu_reader_reg;
 327         pthread_t tid;
 328         int ret;
 329
 330         for (rcu_reader_reg = registry_arena.p;
 331              (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
 332              rcu_reader_reg++) {
 333                 if (!rcu_reader_reg->alloc)
 334                         continue;
 335                 tid = rcu_reader_reg->tid;
 336                 ret = pthread_kill(tid, 0);
 337                 assert(ret != EINVAL);
 338                 if (ret == ESRCH) {
 339                         cds_list_del(&rcu_reader_reg->node);
 340                         rcu_reader_reg->ctr = 0;
 341                         rcu_reader_reg->alloc = 0;
 342                         registry_arena.used -= sizeof(struct rcu_reader);
 343                 }
 344         }
 345 }
 346
 347 /* Disable signals, take mutex, add to registry */
 348 void rcu_bp_register(void)
 349 {
 350         sigset_t newmask, oldmask;
 351         int ret;
 352
 353         ret = sigemptyset(&newmask);
 354         assert(!ret);
 355         ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
 356         assert(!ret);
 357
 358         /*
 359          * Check if a signal concurrently registered our thread since
 360          * the check in rcu_read_lock(). */
 361         if (rcu_reader)
 362                 goto end;
 363
 364         mutex_lock(&rcu_gp_lock);
 365         add_thread();
 366         mutex_unlock(&rcu_gp_lock);
 367 end:
 368         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 369         assert(!ret);
 370 }
 371
 372 void rcu_bp_exit()
 373 {
 374         munmap(registry_arena.p, registry_arena.len);
 375 }
 376
 377 /*
 378  * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
 379  * a concurrent thread executing with this same lock held. This ensures that the
 380  * registry is in a coherent state in the child.
 381  */
 382 void rcu_bp_before_fork(void)
 383 {
 384         sigset_t newmask, oldmask;
 385         int ret;
 386
 387         ret = sigemptyset(&newmask);
 388         assert(!ret);
 389         ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
 390         assert(!ret);
 391         mutex_lock(&rcu_gp_lock);
 392         saved_fork_signal_mask = oldmask;
 393 }
 394
 395 void rcu_bp_after_fork_parent(void)
 396 {
 397         sigset_t oldmask;
 398         int ret;
 399
 400         oldmask = saved_fork_signal_mask;
 401         mutex_unlock(&rcu_gp_lock);
 402         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 403         assert(!ret);
 404 }
 405
 406 void rcu_bp_after_fork_child(void)
 407 {
 408         sigset_t oldmask;
 409         int ret;
 410
 411         rcu_gc_registry();
 412         oldmask = saved_fork_signal_mask;
 413         mutex_unlock(&rcu_gp_lock);
 414         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 415         assert(!ret);
 416 }
 417
 418 #include "urcu-call-rcu-impl.h"
 419 #include "urcu-defer-impl.h"