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/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 not allowing MAYMOVE.
  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         assert(!(flags & MREMAP_MAYMOVE));
  77
  78         return MAP_FAILED;
  79 }
  80 #endif
  81
  82 /* Sleep delay in ms */
  83 #define RCU_SLEEP_DELAY_MS      10
  84 #define INIT_NR_THREADS         8
  85 #define ARENA_INIT_ALLOC                \
  86         sizeof(struct registry_chunk)   \
  87         + INIT_NR_THREADS * sizeof(struct rcu_reader)
  88
  89 /*
  90  * Active attempts to check for reader Q.S. before calling sleep().
  91  */
  92 #define RCU_QS_ACTIVE_ATTEMPTS 100
  93
  94 static
  95 int rcu_bp_refcount;
  96
  97 static
  98 void __attribute__((constructor)) rcu_bp_init(void);
  99 static
 100 void __attribute__((destructor)) _rcu_bp_exit(void);
 101
 102 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 103
 104 static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
 105 static int initialized;
 106
 107 static pthread_key_t urcu_bp_key;
 108
 109 #ifdef DEBUG_YIELD
 110 unsigned int rcu_yield_active;
 111 __DEFINE_URCU_TLS_GLOBAL(unsigned int, rcu_rand_yield);
 112 #endif
 113
 114 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
 115
 116 /*
 117  * Pointer to registry elements. Written to only by each individual reader. Read
 118  * by both the reader and the writers.
 119  */
 120 __DEFINE_URCU_TLS_GLOBAL(struct rcu_reader *, rcu_reader);
 121
 122 static CDS_LIST_HEAD(registry);
 123
 124 struct registry_chunk {
 125         size_t data_len;                /* data length */
 126         size_t used;                    /* amount of data used */
 127         struct cds_list_head node;      /* chunk_list node */
 128         char data[];
 129 };
 130
 131 struct registry_arena {
 132         struct cds_list_head chunk_list;
 133 };
 134
 135 static struct registry_arena registry_arena = {
 136         .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list),
 137 };
 138
 139 /* Saved fork signal mask, protected by rcu_gp_lock */
 140 static sigset_t saved_fork_signal_mask;
 141
 142 static void mutex_lock(pthread_mutex_t *mutex)
 143 {
 144         int ret;
 145
 146 #ifndef DISTRUST_SIGNALS_EXTREME
 147         ret = pthread_mutex_lock(mutex);
 148         if (ret)
 149                 urcu_die(ret);
 150 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 151         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 152                 if (ret != EBUSY && ret != EINTR)
 153                         urcu_die(ret);
 154                 poll(NULL,0,10);
 155         }
 156 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 157 }
 158
 159 static void mutex_unlock(pthread_mutex_t *mutex)
 160 {
 161         int ret;
 162
 163         ret = pthread_mutex_unlock(mutex);
 164         if (ret)
 165                 urcu_die(ret);
 166 }
 167
 168 static void wait_for_readers(struct cds_list_head *input_readers,
 169                         struct cds_list_head *cur_snap_readers,
 170                         struct cds_list_head *qsreaders)
 171 {
 172         unsigned int wait_loops = 0;
 173         struct rcu_reader *index, *tmp;
 174
 175         /*
 176          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 177          * indicate quiescence (not nested), or observe the current
 178          * rcu_gp.ctr value.
 179          */
 180         for (;;) {
 181                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 182                         wait_loops++;
 183
 184                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 185                         switch (rcu_reader_state(&index->ctr)) {
 186                         case RCU_READER_ACTIVE_CURRENT:
 187                                 if (cur_snap_readers) {
 188                                         cds_list_move(&index->node,
 189                                                 cur_snap_readers);
 190                                         break;
 191                                 }
 192                                 /* Fall-through */
 193                         case RCU_READER_INACTIVE:
 194                                 cds_list_move(&index->node, qsreaders);
 195                                 break;
 196                         case RCU_READER_ACTIVE_OLD:
 197                                 /*
 198                                  * Old snapshot. Leaving node in
 199                                  * input_readers will make us busy-loop
 200                                  * until the snapshot becomes current or
 201                                  * the reader becomes inactive.
 202                                  */
 203                                 break;
 204                         }
 205                 }
 206
 207                 if (cds_list_empty(input_readers)) {
 208                         break;
 209                 } else {
 210                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
 211                                 (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS);
 212                         else
 213                                 caa_cpu_relax();
 214                 }
 215         }
 216 }
 217
 218 void synchronize_rcu(void)
 219 {
 220         CDS_LIST_HEAD(cur_snap_readers);
 221         CDS_LIST_HEAD(qsreaders);
 222         sigset_t newmask, oldmask;
 223         int ret;
 224
 225         ret = sigfillset(&newmask);
 226         assert(!ret);
 227         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 228         assert(!ret);
 229
 230         mutex_lock(&rcu_gp_lock);
 231
 232         if (cds_list_empty(&registry))
 233                 goto out;
 234
 235         /* All threads should read qparity before accessing data structure
 236          * where new ptr points to. */
 237         /* Write new ptr before changing the qparity */
 238         cmm_smp_mb();
 239
 240         /*
 241          * Wait for readers to observe original parity or be quiescent.
 242          */
 243         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 244
 245         /*
 246          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 247          * model easier to understand. It does not have a big performance impact
 248          * anyway, given this is the write-side.
 249          */
 250         cmm_smp_mb();
 251
 252         /* Switch parity: 0 -> 1, 1 -> 0 */
 253         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
 254
 255         /*
 256          * Must commit qparity update to memory before waiting for other parity
 257          * quiescent state. Failure to do so could result in the writer waiting
 258          * forever while new readers are always accessing data (no progress).
 259          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 260          */
 261
 262         /*
 263          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 264          * model easier to understand. It does not have a big performance impact
 265          * anyway, given this is the write-side.
 266          */
 267         cmm_smp_mb();
 268
 269         /*
 270          * Wait for readers to observe new parity or be quiescent.
 271          */
 272         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 273
 274         /*
 275          * Put quiescent reader list back into registry.
 276          */
 277         cds_list_splice(&qsreaders, &registry);
 278
 279         /*
 280          * Finish waiting for reader threads before letting the old ptr being
 281          * freed.
 282          */
 283         cmm_smp_mb();
 284 out:
 285         mutex_unlock(&rcu_gp_lock);
 286         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 287         assert(!ret);
 288 }
 289
 290 /*
 291  * library wrappers to be used by non-LGPL compatible source code.
 292  */
 293
 294 void rcu_read_lock(void)
 295 {
 296         _rcu_read_lock();
 297 }
 298
 299 void rcu_read_unlock(void)
 300 {
 301         _rcu_read_unlock();
 302 }
 303
 304 int rcu_read_ongoing(void)
 305 {
 306         return _rcu_read_ongoing();
 307 }
 308
 309 /*
 310  * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
 311  * Else, try expanding the last chunk. If this fails, allocate a new
 312  * chunk twice as big as the last chunk.
 313  * Memory used by chunks _never_ moves. A chunk could theoretically be
 314  * freed when all "used" slots are released, but we don't do it at this
 315  * point.
 316  */
 317 static
 318 void expand_arena(struct registry_arena *arena)
 319 {
 320         struct registry_chunk *new_chunk, *last_chunk;
 321         size_t old_chunk_len, new_chunk_len;
 322
 323         /* No chunk. */
 324         if (cds_list_empty(&arena->chunk_list)) {
 325                 assert(ARENA_INIT_ALLOC >=
 326                         sizeof(struct registry_chunk)
 327                         + sizeof(struct rcu_reader));
 328                 new_chunk_len = ARENA_INIT_ALLOC;
 329                 new_chunk = mmap(NULL, new_chunk_len,
 330                         PROT_READ | PROT_WRITE,
 331                         MAP_ANONYMOUS | MAP_PRIVATE,
 332                         -1, 0);
 333                 if (new_chunk == MAP_FAILED)
 334                         abort();
 335                 bzero(new_chunk, new_chunk_len);
 336                 new_chunk->data_len =
 337                         new_chunk_len - sizeof(struct registry_chunk);
 338                 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 339                 return;         /* We're done. */
 340         }
 341
 342         /* Try expanding last chunk. */
 343         last_chunk = cds_list_entry(arena->chunk_list.prev,
 344                 struct registry_chunk, node);
 345         old_chunk_len =
 346                 last_chunk->data_len + sizeof(struct registry_chunk);
 347         new_chunk_len = old_chunk_len << 1;
 348
 349         /* Don't allow memory mapping to move, just expand. */
 350         new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
 351                 new_chunk_len, 0);
 352         if (new_chunk != MAP_FAILED) {
 353                 /* Should not have moved. */
 354                 assert(new_chunk == last_chunk);
 355                 bzero((char *) last_chunk + old_chunk_len,
 356                         new_chunk_len - old_chunk_len);
 357                 last_chunk->data_len =
 358                         new_chunk_len - sizeof(struct registry_chunk);
 359                 return;         /* We're done. */
 360         }
 361
 362         /* Remap did not succeed, we need to add a new chunk. */
 363         new_chunk = mmap(NULL, new_chunk_len,
 364                 PROT_READ | PROT_WRITE,
 365                 MAP_ANONYMOUS | MAP_PRIVATE,
 366                 -1, 0);
 367         if (new_chunk == MAP_FAILED)
 368                 abort();
 369         bzero(new_chunk, new_chunk_len);
 370         new_chunk->data_len =
 371                 new_chunk_len - sizeof(struct registry_chunk);
 372         cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 373 }
 374
 375 static
 376 struct rcu_reader *arena_alloc(struct registry_arena *arena)
 377 {
 378         struct registry_chunk *chunk;
 379         struct rcu_reader *rcu_reader_reg;
 380         int expand_done = 0;    /* Only allow to expand once per alloc */
 381         size_t len = sizeof(struct rcu_reader);
 382
 383 retry:
 384         cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
 385                 if (chunk->data_len - chunk->used < len)
 386                         continue;
 387                 /* Find spot */
 388                 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
 389                                 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
 390                                 rcu_reader_reg++) {
 391                         if (!rcu_reader_reg->alloc) {
 392                                 rcu_reader_reg->alloc = 1;
 393                                 chunk->used += len;
 394                                 return rcu_reader_reg;
 395                         }
 396                 }
 397         }
 398
 399         if (!expand_done) {
 400                 expand_arena(arena);
 401                 expand_done = 1;
 402                 goto retry;
 403         }
 404
 405         return NULL;
 406 }
 407
 408 /* Called with signals off and mutex locked */
 409 static
 410 void add_thread(void)
 411 {
 412         struct rcu_reader *rcu_reader_reg;
 413         int ret;
 414
 415         rcu_reader_reg = arena_alloc(&registry_arena);
 416         if (!rcu_reader_reg)
 417                 abort();
 418         ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
 419         if (ret)
 420                 abort();
 421
 422         /* Add to registry */
 423         rcu_reader_reg->tid = pthread_self();
 424         assert(rcu_reader_reg->ctr == 0);
 425         cds_list_add(&rcu_reader_reg->node, &registry);
 426         /*
 427          * Reader threads are pointing to the reader registry. This is
 428          * why its memory should never be relocated.
 429          */
 430         URCU_TLS(rcu_reader) = rcu_reader_reg;
 431 }
 432
 433 /* Called with mutex locked */
 434 static
 435 void cleanup_thread(struct registry_chunk *chunk,
 436                 struct rcu_reader *rcu_reader_reg)
 437 {
 438         rcu_reader_reg->ctr = 0;
 439         cds_list_del(&rcu_reader_reg->node);
 440         rcu_reader_reg->tid = 0;
 441         rcu_reader_reg->alloc = 0;
 442         chunk->used -= sizeof(struct rcu_reader);
 443 }
 444
 445 static
 446 struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
 447 {
 448         struct registry_chunk *chunk;
 449
 450         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 451                 if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
 452                         continue;
 453                 if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
 454                         continue;
 455                 return chunk;
 456         }
 457         return NULL;
 458 }
 459
 460 /* Called with signals off and mutex locked */
 461 static
 462 void remove_thread(struct rcu_reader *rcu_reader_reg)
 463 {
 464         cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 465         URCU_TLS(rcu_reader) = NULL;
 466 }
 467
 468 /* Disable signals, take mutex, add to registry */
 469 void rcu_bp_register(void)
 470 {
 471         sigset_t newmask, oldmask;
 472         int ret;
 473
 474         ret = sigfillset(&newmask);
 475         if (ret)
 476                 abort();
 477         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 478         if (ret)
 479                 abort();
 480
 481         /*
 482          * Check if a signal concurrently registered our thread since
 483          * the check in rcu_read_lock().
 484          */
 485         if (URCU_TLS(rcu_reader))
 486                 goto end;
 487
 488         /*
 489          * Take care of early registration before urcu_bp constructor.
 490          */
 491         rcu_bp_init();
 492
 493         mutex_lock(&rcu_gp_lock);
 494         add_thread();
 495         mutex_unlock(&rcu_gp_lock);
 496 end:
 497         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 498         if (ret)
 499                 abort();
 500 }
 501
 502 /* Disable signals, take mutex, remove from registry */
 503 static
 504 void rcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
 505 {
 506         sigset_t newmask, oldmask;
 507         int ret;
 508
 509         ret = sigfillset(&newmask);
 510         if (ret)
 511                 abort();
 512         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 513         if (ret)
 514                 abort();
 515
 516         mutex_lock(&rcu_gp_lock);
 517         remove_thread(rcu_reader_reg);
 518         mutex_unlock(&rcu_gp_lock);
 519         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 520         if (ret)
 521                 abort();
 522         _rcu_bp_exit();
 523 }
 524
 525 /*
 526  * Remove thread from the registry when it exits, and flag it as
 527  * destroyed so garbage collection can take care of it.
 528  */
 529 static
 530 void urcu_bp_thread_exit_notifier(void *rcu_key)
 531 {
 532         rcu_bp_unregister(rcu_key);
 533 }
 534
 535 static
 536 void rcu_bp_init(void)
 537 {
 538         mutex_lock(&init_lock);
 539         if (!rcu_bp_refcount++) {
 540                 int ret;
 541
 542                 ret = pthread_key_create(&urcu_bp_key,
 543                                 urcu_bp_thread_exit_notifier);
 544                 if (ret)
 545                         abort();
 546                 initialized = 1;
 547         }
 548         mutex_unlock(&init_lock);
 549 }
 550
 551 static
 552 void _rcu_bp_exit(void)
 553 {
 554         mutex_lock(&init_lock);
 555         if (!--rcu_bp_refcount) {
 556                 struct registry_chunk *chunk, *tmp;
 557                 int ret;
 558
 559                 cds_list_for_each_entry_safe(chunk, tmp,
 560                                 &registry_arena.chunk_list, node) {
 561                         munmap(chunk, chunk->data_len
 562                                         + sizeof(struct registry_chunk));
 563                 }
 564                 ret = pthread_key_delete(urcu_bp_key);
 565                 if (ret)
 566                         abort();
 567         }
 568         mutex_unlock(&init_lock);
 569 }
 570
 571 /*
 572  * Keep ABI compability within stable versions. This has never been
 573  * exposed through a header, but needs to stay in the .so until the
 574  * soname is bumped.
 575  */
 576 void rcu_bp_exit(void)
 577 {
 578 }
 579
 580 /*
 581  * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
 582  * a concurrent thread executing with this same lock held. This ensures that the
 583  * registry is in a coherent state in the child.
 584  */
 585 void rcu_bp_before_fork(void)
 586 {
 587         sigset_t newmask, oldmask;
 588         int ret;
 589
 590         ret = sigfillset(&newmask);
 591         assert(!ret);
 592         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 593         assert(!ret);
 594         mutex_lock(&rcu_gp_lock);
 595         saved_fork_signal_mask = oldmask;
 596 }
 597
 598 void rcu_bp_after_fork_parent(void)
 599 {
 600         sigset_t oldmask;
 601         int ret;
 602
 603         oldmask = saved_fork_signal_mask;
 604         mutex_unlock(&rcu_gp_lock);
 605         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 606         assert(!ret);
 607 }
 608
 609 /*
 610  * Prune all entries from registry except our own thread. Fits the Linux
 611  * fork behavior. Called with rcu_gp_lock held.
 612  */
 613 static
 614 void urcu_bp_prune_registry(void)
 615 {
 616         struct registry_chunk *chunk;
 617         struct rcu_reader *rcu_reader_reg;
 618
 619         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 620                 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
 621                                 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
 622                                 rcu_reader_reg++) {
 623                         if (!rcu_reader_reg->alloc)
 624                                 continue;
 625                         if (rcu_reader_reg->tid == pthread_self())
 626                                 continue;
 627                         cleanup_thread(chunk, rcu_reader_reg);
 628                 }
 629         }
 630 }
 631
 632 void rcu_bp_after_fork_child(void)
 633 {
 634         sigset_t oldmask;
 635         int ret;
 636
 637         urcu_bp_prune_registry();
 638         oldmask = saved_fork_signal_mask;
 639         mutex_unlock(&rcu_gp_lock);
 640         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 641         assert(!ret);
 642 }
 643
 644 void *rcu_dereference_sym_bp(void *p)
 645 {
 646         return _rcu_dereference(p);
 647 }
 648
 649 void *rcu_set_pointer_sym_bp(void **p, void *v)
 650 {
 651         cmm_wmb();
 652         uatomic_set(p, v);
 653         return v;
 654 }
 655
 656 void *rcu_xchg_pointer_sym_bp(void **p, void *v)
 657 {
 658         cmm_wmb();
 659         return uatomic_xchg(p, v);
 660 }
 661
 662 void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new)
 663 {
 664         cmm_wmb();
 665         return uatomic_cmpxchg(p, old, _new);
 666 }
 667
 668 DEFINE_RCU_FLAVOR(rcu_flavor);
 669
 670 #include "urcu-call-rcu-impl.h"
 671 #include "urcu-defer-impl.h"