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 #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 yield_active;
 109 DEFINE_URCU_TLS(unsigned int, 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 void update_counter_and_wait(void)
 168 {
 169         CDS_LIST_HEAD(qsreaders);
 170         int wait_loops = 0;
 171         struct rcu_reader *index, *tmp;
 172
 173         /* Switch parity: 0 -> 1, 1 -> 0 */
 174         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
 175
 176         /*
 177          * Must commit qparity update to memory before waiting for other parity
 178          * quiescent state. Failure to do so could result in the writer waiting
 179          * forever while new readers are always accessing data (no progress).
 180          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 181          */
 182
 183         /*
 184          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 185          * model easier to understand. It does not have a big performance impact
 186          * anyway, given this is the write-side.
 187          */
 188         cmm_smp_mb();
 189
 190         /*
 191          * Wait for each thread rcu_reader.ctr count to become 0.
 192          */
 193         for (;;) {
 194                 wait_loops++;
 195                 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
 196                         if (!rcu_old_gp_ongoing(&index->ctr))
 197                                 cds_list_move(&index->node, &qsreaders);
 198                 }
 199
 200                 if (cds_list_empty(&registry)) {
 201                         break;
 202                 } else {
 203                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 204                                 usleep(RCU_SLEEP_DELAY);
 205                         else
 206                                 caa_cpu_relax();
 207                 }
 208         }
 209         /* put back the reader list in the registry */
 210         cds_list_splice(&qsreaders, &registry);
 211 }
 212
 213 void synchronize_rcu(void)
 214 {
 215         sigset_t newmask, oldmask;
 216         int ret;
 217
 218         ret = sigfillset(&newmask);
 219         assert(!ret);
 220         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 221         assert(!ret);
 222
 223         mutex_lock(&rcu_gp_lock);
 224
 225         if (cds_list_empty(&registry))
 226                 goto out;
 227
 228         /* All threads should read qparity before accessing data structure
 229          * where new ptr points to. */
 230         /* Write new ptr before changing the qparity */
 231         cmm_smp_mb();
 232
 233         /* Remove old registry elements */
 234         rcu_gc_registry();
 235
 236         /*
 237          * Wait for previous parity to be empty of readers.
 238          */
 239         update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 240
 241         /*
 242          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 243          * model easier to understand. It does not have a big performance impact
 244          * anyway, given this is the write-side.
 245          */
 246         cmm_smp_mb();
 247
 248         /*
 249          * Wait for previous parity to be empty of readers.
 250          */
 251         update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
 252
 253         /*
 254          * Finish waiting for reader threads before letting the old ptr being
 255          * freed.
 256          */
 257         cmm_smp_mb();
 258 out:
 259         mutex_unlock(&rcu_gp_lock);
 260         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 261         assert(!ret);
 262 }
 263
 264 /*
 265  * library wrappers to be used by non-LGPL compatible source code.
 266  */
 267
 268 void rcu_read_lock(void)
 269 {
 270         _rcu_read_lock();
 271 }
 272
 273 void rcu_read_unlock(void)
 274 {
 275         _rcu_read_unlock();
 276 }
 277
 278 /*
 279  * only grow for now.
 280  */
 281 static void resize_arena(struct registry_arena *arena, size_t len)
 282 {
 283         void *new_p;
 284         size_t old_len;
 285
 286         old_len = arena->len;
 287
 288         if (!arena->p)
 289                 new_p = mmap(arena->p, len,
 290                         PROT_READ | PROT_WRITE,
 291                         MAP_ANONYMOUS | MAP_PRIVATE,
 292                         -1, 0);
 293         else
 294                 new_p = mremap_wrapper(arena->p, old_len,
 295                         len, MREMAP_MAYMOVE);
 296         assert(new_p != MAP_FAILED);
 297
 298         /*
 299          * Zero the newly allocated memory. Since mmap() does not
 300          * clearly specify if memory is zeroed or not (although it is
 301          * very likely that it is), be extra careful by not expecting
 302          * the new range to be zeroed by mremap.
 303          */
 304         bzero(new_p + old_len, len - old_len);
 305
 306         /*
 307          * If we did not re-use the same region, we need to update the
 308          * arena pointer.
 309          */
 310         if (new_p != arena->p)
 311                 arena->p = new_p;
 312
 313         arena->len = len;
 314 }
 315
 316 /* Called with signals off and mutex locked */
 317 static void add_thread(void)
 318 {
 319         struct rcu_reader *rcu_reader_reg;
 320
 321         if (registry_arena.len
 322             < registry_arena.used + sizeof(struct rcu_reader))
 323                 resize_arena(&registry_arena,
 324                 caa_max(registry_arena.len << 1, ARENA_INIT_ALLOC));
 325         /*
 326          * Find a free spot.
 327          */
 328         for (rcu_reader_reg = registry_arena.p;
 329              (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
 330              rcu_reader_reg++) {
 331                 if (!rcu_reader_reg->alloc)
 332                         break;
 333         }
 334         rcu_reader_reg->alloc = 1;
 335         registry_arena.used += sizeof(struct rcu_reader);
 336
 337         /* Add to registry */
 338         rcu_reader_reg->tid = pthread_self();
 339         assert(rcu_reader_reg->ctr == 0);
 340         cds_list_add(&rcu_reader_reg->node, &registry);
 341         URCU_TLS(rcu_reader) = rcu_reader_reg;
 342 }
 343
 344 /* Called with signals off and mutex locked */
 345 static void rcu_gc_registry(void)
 346 {
 347         struct rcu_reader *rcu_reader_reg;
 348         pthread_t tid;
 349         int ret;
 350
 351         for (rcu_reader_reg = registry_arena.p;
 352              (void *)rcu_reader_reg < registry_arena.p + registry_arena.len;
 353              rcu_reader_reg++) {
 354                 if (!rcu_reader_reg->alloc)
 355                         continue;
 356                 tid = rcu_reader_reg->tid;
 357                 ret = pthread_kill(tid, 0);
 358                 assert(ret != EINVAL);
 359                 if (ret == ESRCH) {
 360                         cds_list_del(&rcu_reader_reg->node);
 361                         rcu_reader_reg->ctr = 0;
 362                         rcu_reader_reg->alloc = 0;
 363                         registry_arena.used -= sizeof(struct rcu_reader);
 364                 }
 365         }
 366 }
 367
 368 /* Disable signals, take mutex, add to registry */
 369 void rcu_bp_register(void)
 370 {
 371         sigset_t newmask, oldmask;
 372         int ret;
 373
 374         ret = sigfillset(&newmask);
 375         assert(!ret);
 376         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 377         assert(!ret);
 378
 379         /*
 380          * Check if a signal concurrently registered our thread since
 381          * the check in rcu_read_lock(). */
 382         if (URCU_TLS(rcu_reader))
 383                 goto end;
 384
 385         mutex_lock(&rcu_gp_lock);
 386         add_thread();
 387         mutex_unlock(&rcu_gp_lock);
 388 end:
 389         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 390         assert(!ret);
 391 }
 392
 393 void rcu_bp_exit(void)
 394 {
 395         if (registry_arena.p)
 396                 munmap(registry_arena.p, registry_arena.len);
 397 }
 398
 399 /*
 400  * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
 401  * a concurrent thread executing with this same lock held. This ensures that the
 402  * registry is in a coherent state in the child.
 403  */
 404 void rcu_bp_before_fork(void)
 405 {
 406         sigset_t newmask, oldmask;
 407         int ret;
 408
 409         ret = sigfillset(&newmask);
 410         assert(!ret);
 411         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 412         assert(!ret);
 413         mutex_lock(&rcu_gp_lock);
 414         saved_fork_signal_mask = oldmask;
 415 }
 416
 417 void rcu_bp_after_fork_parent(void)
 418 {
 419         sigset_t oldmask;
 420         int ret;
 421
 422         oldmask = saved_fork_signal_mask;
 423         mutex_unlock(&rcu_gp_lock);
 424         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 425         assert(!ret);
 426 }
 427
 428 void rcu_bp_after_fork_child(void)
 429 {
 430         sigset_t oldmask;
 431         int ret;
 432
 433         rcu_gc_registry();
 434         oldmask = saved_fork_signal_mask;
 435         mutex_unlock(&rcu_gp_lock);
 436         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 437         assert(!ret);
 438 }
 439
 440 void *rcu_dereference_sym_bp(void *p)
 441 {
 442         return _rcu_dereference(p);
 443 }
 444
 445 void *rcu_set_pointer_sym_bp(void **p, void *v)
 446 {
 447         cmm_wmb();
 448         uatomic_set(p, v);
 449         return v;
 450 }
 451
 452 void *rcu_xchg_pointer_sym_bp(void **p, void *v)
 453 {
 454         cmm_wmb();
 455         return uatomic_xchg(p, v);
 456 }
 457
 458 void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new)
 459 {
 460         cmm_wmb();
 461         return uatomic_cmpxchg(p, old, _new);
 462 }
 463
 464 DEFINE_RCU_FLAVOR(rcu_flavor);
 465
 466 #include "urcu-call-rcu-impl.h"
 467 #include "urcu-defer-impl.h"