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 #include "urcu-pointer.h"
  43 #include "urcu/tls-compat.h"
  44
  45 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  46 #undef _LGPL_SOURCE
  47 #include "urcu-bp.h"
  48 #define _LGPL_SOURCE
  49
  50 #ifndef MAP_ANONYMOUS
  51 #define MAP_ANONYMOUS MAP_ANON
  52 #endif
  53
  54 #ifndef __linux__
  55
  56 #define MREMAP_MAYMOVE  1
  57 #define MREMAP_FIXED    2
  58
  59 /*
  60  * mremap wrapper for non-Linux systems. Maps a RW, anonymous private mapping.
  61  * This is not generic.
  62 */
  63 void *mremap(void *old_address, size_t old_size, size_t new_size, int flags)
  64 {
  65         void *new_address;
  66
  67         assert(flags & MREMAP_MAYMOVE);
  68         assert(!(flags & MREMAP_FIXED));
  69         new_address = mmap(old_address, new_size,
  70                            PROT_READ | PROT_WRITE,
  71                            MAP_ANONYMOUS | MAP_PRIVATE,
  72                            -1, 0);
  73         if (new_address == MAP_FAILED)
  74                 return MAP_FAILED;
  75         if (old_address) {
  76                 memcpy(new_address, old_address, old_size);
  77                 munmap(old_address, old_size);
  78         }
  79         return new_address;
  80 }
  81 #endif
  82
  83 /* Sleep delay in us */
  84 #define RCU_SLEEP_DELAY         1000
  85 #define ARENA_INIT_ALLOC        16
  86
  87 /*
  88  * Active attempts to check for reader Q.S. before calling sleep().
  89  */
  90 #define RCU_QS_ACTIVE_ATTEMPTS 100
  91
  92 void __attribute__((destructor)) rcu_bp_exit(void);
  93
  94 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  95
  96 #ifdef DEBUG_YIELD
  97 unsigned int yield_active;
  98 DEFINE_URCU_TLS(unsigned int, rand_yield);
  99 #endif
 100
 101 /*
 102  * Global grace period counter.
 103  * Contains the current RCU_GP_CTR_PHASE.
 104  * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
 105  * Written to only by writer with mutex taken. Read by both writer and readers.
 106  */
 107 long rcu_gp_ctr = RCU_GP_COUNT;
 108
 109 /*
 110  * Pointer to registry elements. Written to only by each individual reader. Read
 111  * by both the reader and the writers.
 112  */
 113 DEFINE_URCU_TLS(struct rcu_reader *, rcu_reader);
 114
 115 static CDS_LIST_HEAD(registry);
 116
 117 struct registry_arena {
 118         void *p;
 119         size_t len;
 120         size_t used;
 121 };
 122
 123 static struct registry_arena registry_arena;
 124
 125 /* Saved fork signal mask, protected by rcu_gp_lock */
 126 static sigset_t saved_fork_signal_mask;
 127
 128 static void rcu_gc_registry(void);
 129
 130 static void mutex_lock(pthread_mutex_t *mutex)
 131 {
 132         int ret;
 133
 134 #ifndef DISTRUST_SIGNALS_EXTREME
 135         ret = pthread_mutex_lock(mutex);
 136         if (ret) {
 137                 perror("Error in pthread mutex lock");
 138                 exit(-1);
 139         }
 140 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 141         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 142                 if (ret != EBUSY && ret != EINTR) {
 143                         printf("ret = %d, errno = %d\n", ret, errno);
 144                         perror("Error in pthread mutex lock");
 145                         exit(-1);
 146                 }
 147                 poll(NULL,0,10);
 148         }
 149 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 150 }
 151
 152 static void mutex_unlock(pthread_mutex_t *mutex)
 153 {
 154         int ret;
 155
 156         ret = pthread_mutex_unlock(mutex);
 157         if (ret) {
 158                 perror("Error in pthread mutex unlock");
 159                 exit(-1);
 160         }
 161 }
 162
 163 void update_counter_and_wait(void)
 164 {
 165         CDS_LIST_HEAD(qsreaders);
 166         int wait_loops = 0;
 167         struct rcu_reader *index, *tmp;
 168
 169         /* Switch parity: 0 -> 1, 1 -> 0 */
 170         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
 171
 172         /*
 173          * Must commit qparity update to memory before waiting for other parity
 174          * quiescent state. Failure to do so could result in the writer waiting
 175          * forever while new readers are always accessing data (no progress).
 176          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 177          */
 178
 179         /*
 180          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 181          * model easier to understand. It does not have a big performance impact
 182          * anyway, given this is the write-side.
 183          */
 184         cmm_smp_mb();
 185
 186         /*
 187          * Wait for each thread rcu_reader.ctr count to become 0.
 188          */
 189         for (;;) {
 190                 wait_loops++;
 191                 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
 192                         if (!rcu_old_gp_ongoing(&index->ctr))
 193                                 cds_list_move(&index->node, &qsreaders);
 194                 }
 195
 196                 if (cds_list_empty(&registry)) {
 197                         break;
 198                 } else {
 199                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 200                                 usleep(RCU_SLEEP_DELAY);
 201                         else
 202                                 caa_cpu_relax();
 203                 }
 204         }
 205         /* put back the reader list in the registry */
 206         cds_list_splice(&qsreaders, &registry);
 207 }
 208
 209 void synchronize_rcu(void)
 210 {
 211         sigset_t newmask, oldmask;
 212         int ret;
 213
 214         ret = sigemptyset(&newmask);
 215         assert(!ret);
 216         ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
 217         assert(!ret);
 218
 219         mutex_lock(&rcu_gp_lock);
 220
 221         if (cds_list_empty(&registry))
 222                 goto out;
 223
 224         /* All threads should read qparity before accessing data structure
 225          * where new ptr points to. */
 226         /* Write new ptr before changing the qparity */
 227         cmm_smp_mb();
 228
 229         /* Remove old registry elements */
 230         rcu_gc_registry();
 231
 232         /*
 233          * Wait for previous parity to be empty of readers.
 234          */
 235         update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 236
 237         /*
 238          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 239          * model easier to understand. It does not have a big performance impact
 240          * anyway, given this is the write-side.
 241          */
 242         cmm_smp_mb();
 243
 244         /*
 245          * Wait for previous parity to be empty of readers.
 246          */
 247         update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
 248
 249         /*
 250          * Finish waiting for reader threads before letting the old ptr being
 251          * freed.
 252          */
 253         cmm_smp_mb();
 254 out:
 255         mutex_unlock(&rcu_gp_lock);
 256         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 257         assert(!ret);
 258 }
 259
 260 /*
 261  * library wrappers to be used by non-LGPL compatible source code.
 262  */
 263
 264 void rcu_read_lock(void)
 265 {
 266         _rcu_read_lock();
 267 }
 268
 269 void rcu_read_unlock(void)
 270 {
 271         _rcu_read_unlock();
 272 }
 273
 274 /*
 275  * only grow for now.
 276  */
 277 static void resize_arena(struct registry_arena *arena, size_t len)
 278 {
 279         void *new_arena;
 280
 281         if (!arena->p)
 282                 new_arena = mmap(arena->p, len,
 283                                  PROT_READ | PROT_WRITE,
 284                                  MAP_ANONYMOUS | MAP_PRIVATE,
 285                                  -1, 0);
 286         else
 287                 new_arena = mremap(arena->p, arena->len,
 288                                    len, MREMAP_MAYMOVE);
 289         assert(new_arena != MAP_FAILED);
 290
 291         /*
 292          * re-used the same region ?
 293          */
 294         if (new_arena == arena->p)
 295                 return;
 296
 297         bzero(new_arena + arena->len, len - arena->len);
 298         arena->p = new_arena;
 299 }
 300
 301 /* Called with signals off and mutex locked */
 302 static void add_thread(void)
 303 {
 304         struct rcu_reader *rcu_reader_reg;
 305
 306         if (registry_arena.len
 307             < registry_arena.used + sizeof(struct rcu_reader))
 308                 resize_arena(&registry_arena,
 309                 caa_max(registry_arena.len << 1, ARENA_INIT_ALLOC));
 310         /*
 311          * Find a free spot.
 312          */
 313         for (rcu_reader_reg = registry_arena.p;
 314              (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
 315              rcu_reader_reg++) {
 316                 if (!rcu_reader_reg->alloc)
 317                         break;
 318         }
 319         rcu_reader_reg->alloc = 1;
 320         registry_arena.used += sizeof(struct rcu_reader);
 321
 322         /* Add to registry */
 323         rcu_reader_reg->tid = pthread_self();
 324         assert(rcu_reader_reg->ctr == 0);
 325         cds_list_add(&rcu_reader_reg->node, &registry);
 326         URCU_TLS(rcu_reader) = rcu_reader_reg;
 327 }
 328
 329 /* Called with signals off and mutex locked */
 330 static void rcu_gc_registry(void)
 331 {
 332         struct rcu_reader *rcu_reader_reg;
 333         pthread_t tid;
 334         int ret;
 335
 336         for (rcu_reader_reg = registry_arena.p;
 337              (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
 338              rcu_reader_reg++) {
 339                 if (!rcu_reader_reg->alloc)
 340                         continue;
 341                 tid = rcu_reader_reg->tid;
 342                 ret = pthread_kill(tid, 0);
 343                 assert(ret != EINVAL);
 344                 if (ret == ESRCH) {
 345                         cds_list_del(&rcu_reader_reg->node);
 346                         rcu_reader_reg->ctr = 0;
 347                         rcu_reader_reg->alloc = 0;
 348                         registry_arena.used -= sizeof(struct rcu_reader);
 349                 }
 350         }
 351 }
 352
 353 /* Disable signals, take mutex, add to registry */
 354 void rcu_bp_register(void)
 355 {
 356         sigset_t newmask, oldmask;
 357         int ret;
 358
 359         ret = sigemptyset(&newmask);
 360         assert(!ret);
 361         ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
 362         assert(!ret);
 363
 364         /*
 365          * Check if a signal concurrently registered our thread since
 366          * the check in rcu_read_lock(). */
 367         if (URCU_TLS(rcu_reader))
 368                 goto end;
 369
 370         mutex_lock(&rcu_gp_lock);
 371         add_thread();
 372         mutex_unlock(&rcu_gp_lock);
 373 end:
 374         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 375         assert(!ret);
 376 }
 377
 378 void rcu_bp_exit(void)
 379 {
 380         if (registry_arena.p)
 381                 munmap(registry_arena.p, registry_arena.len);
 382 }
 383
 384 /*
 385  * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
 386  * a concurrent thread executing with this same lock held. This ensures that the
 387  * registry is in a coherent state in the child.
 388  */
 389 void rcu_bp_before_fork(void)
 390 {
 391         sigset_t newmask, oldmask;
 392         int ret;
 393
 394         ret = sigemptyset(&newmask);
 395         assert(!ret);
 396         ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask);
 397         assert(!ret);
 398         mutex_lock(&rcu_gp_lock);
 399         saved_fork_signal_mask = oldmask;
 400 }
 401
 402 void rcu_bp_after_fork_parent(void)
 403 {
 404         sigset_t oldmask;
 405         int ret;
 406
 407         oldmask = saved_fork_signal_mask;
 408         mutex_unlock(&rcu_gp_lock);
 409         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 410         assert(!ret);
 411 }
 412
 413 void rcu_bp_after_fork_child(void)
 414 {
 415         sigset_t oldmask;
 416         int ret;
 417
 418         rcu_gc_registry();
 419         oldmask = saved_fork_signal_mask;
 420         mutex_unlock(&rcu_gp_lock);
 421         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 422         assert(!ret);
 423 }
 424
 425 void *rcu_dereference_sym_bp(void *p)
 426 {
 427         return _rcu_dereference(p);
 428 }
 429
 430 void *rcu_set_pointer_sym_bp(void **p, void *v)
 431 {
 432         cmm_wmb();
 433         return uatomic_set(p, v);
 434 }
 435
 436 void *rcu_xchg_pointer_sym_bp(void **p, void *v)
 437 {
 438         cmm_wmb();
 439         return uatomic_xchg(p, v);
 440 }
 441
 442 void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new)
 443 {
 444         cmm_wmb();
 445         return uatomic_cmpxchg(p, old, _new);
 446 }
 447
 448 DEFINE_RCU_FLAVOR(rcu_flavor);
 449
 450 #include "urcu-call-rcu-impl.h"
 451 #include "urcu-defer-impl.h"