urcu.c

   1 /*
   2  * urcu.c
   3  *
   4  * Userspace RCU library
   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 _BSD_SOURCE
  27 #define _GNU_SOURCE
  28 #define _LGPL_SOURCE
  29 #define _DEFAULT_SOURCE
  30 #include <stdio.h>
  31 #include <pthread.h>
  32 #include <signal.h>
  33 #include <assert.h>
  34 #include <stdlib.h>
  35 #include <stdint.h>
  36 #include <string.h>
  37 #include <errno.h>
  38 #include <poll.h>
  39
  40 #include "urcu/arch.h"
  41 #include "urcu/wfcqueue.h"
  42 #include "urcu/map/urcu.h"
  43 #include "urcu/static/urcu.h"
  44 #include "urcu-pointer.h"
  45 #include "urcu/tls-compat.h"
  46
  47 #include "urcu-die.h"
  48 #include "urcu-wait.h"
  49
  50 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  51 #undef _LGPL_SOURCE
  52 #include "urcu.h"
  53 #define _LGPL_SOURCE
  54
  55 /*
  56  * If a reader is really non-cooperative and refuses to commit its
  57  * rcu_active_readers count to memory (there is no barrier in the reader
  58  * per-se), kick it after 10 loops waiting for it.
  59  */
  60 #define KICK_READER_LOOPS       10
  61
  62 /*
  63  * Active attempts to check for reader Q.S. before calling futex().
  64  */
  65 #define RCU_QS_ACTIVE_ATTEMPTS 100
  66
  67 /* If the headers do not support membarrier system call, fall back on RCU_MB */
  68 #ifdef __NR_membarrier
  69 # define membarrier(...)                syscall(__NR_membarrier, __VA_ARGS__)
  70 #else
  71 # define membarrier(...)                -ENOSYS
  72 #endif
  73
  74 enum membarrier_cmd {
  75         MEMBARRIER_CMD_QUERY = 0,
  76         MEMBARRIER_CMD_SHARED = (1 << 0),
  77 };
  78
  79 #ifdef RCU_MEMBARRIER
  80 static int init_done;
  81 int rcu_has_sys_membarrier;
  82
  83 void __attribute__((constructor)) rcu_init(void);
  84 #endif
  85
  86 #ifdef RCU_MB
  87 void rcu_init(void)
  88 {
  89 }
  90 #endif
  91
  92 #ifdef RCU_SIGNAL
  93 static int init_done;
  94
  95 void __attribute__((constructor)) rcu_init(void);
  96 void __attribute__((destructor)) rcu_exit(void);
  97 #endif
  98
  99 /*
 100  * rcu_gp_lock ensures mutual exclusion between threads calling
 101  * synchronize_rcu().
 102  */
 103 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 104 /*
 105  * rcu_registry_lock ensures mutual exclusion between threads
 106  * registering and unregistering themselves to/from the registry, and
 107  * with threads reading that registry from synchronize_rcu(). However,
 108  * this lock is not held all the way through the completion of awaiting
 109  * for the grace period. It is sporadically released between iterations
 110  * on the registry.
 111  * rcu_registry_lock may nest inside rcu_gp_lock.
 112  */
 113 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 114 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
 115
 116 /*
 117  * Written to only by each individual reader. Read by both the reader and the
 118  * writers.
 119  */
 120 DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
 121
 122 static CDS_LIST_HEAD(registry);
 123
 124 /*
 125  * Queue keeping threads awaiting to wait for a grace period. Contains
 126  * struct gp_waiters_thread objects.
 127  */
 128 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
 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                 urcu_die(ret);
 138 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 139         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 140                 if (ret != EBUSY && ret != EINTR)
 141                         urcu_die(ret);
 142                 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
 143                         cmm_smp_mb();
 144                         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 145                         cmm_smp_mb();
 146                 }
 147                 (void) 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                 urcu_die(ret);
 159 }
 160
 161 #ifdef RCU_MEMBARRIER
 162 static void smp_mb_master(void)
 163 {
 164         if (caa_likely(rcu_has_sys_membarrier))
 165                 (void) membarrier(MEMBARRIER_CMD_SHARED, 0);
 166         else
 167                 cmm_smp_mb();
 168 }
 169 #endif
 170
 171 #ifdef RCU_MB
 172 static void smp_mb_master(void)
 173 {
 174         cmm_smp_mb();
 175 }
 176 #endif
 177
 178 #ifdef RCU_SIGNAL
 179 static void force_mb_all_readers(void)
 180 {
 181         struct rcu_reader *index;
 182
 183         /*
 184          * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 185          * compiler barriers around rcu read lock as real memory barriers.
 186          */
 187         if (cds_list_empty(&registry))
 188                 return;
 189         /*
 190          * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 191          * a cache flush on architectures with non-coherent cache. Let's play
 192          * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 193          * cache flush is enforced.
 194          */
 195         cds_list_for_each_entry(index, &registry, node) {
 196                 CMM_STORE_SHARED(index->need_mb, 1);
 197                 pthread_kill(index->tid, SIGRCU);
 198         }
 199         /*
 200          * Wait for sighandler (and thus mb()) to execute on every thread.
 201          *
 202          * Note that the pthread_kill() will never be executed on systems
 203          * that correctly deliver signals in a timely manner.  However, it
 204          * is not uncommon for kernels to have bugs that can result in
 205          * lost or unduly delayed signals.
 206          *
 207          * If you are seeing the below pthread_kill() executing much at
 208          * all, we suggest testing the underlying kernel and filing the
 209          * relevant bug report.  For Linux kernels, we recommend getting
 210          * the Linux Test Project (LTP).
 211          */
 212         cds_list_for_each_entry(index, &registry, node) {
 213                 while (CMM_LOAD_SHARED(index->need_mb)) {
 214                         pthread_kill(index->tid, SIGRCU);
 215                         (void) poll(NULL, 0, 1);
 216                 }
 217         }
 218         cmm_smp_mb();   /* read ->need_mb before ending the barrier */
 219 }
 220
 221 static void smp_mb_master(void)
 222 {
 223         force_mb_all_readers();
 224 }
 225 #endif /* #ifdef RCU_SIGNAL */
 226
 227 /*
 228  * synchronize_rcu() waiting. Single thread.
 229  * Always called with rcu_registry lock held. Releases this lock and
 230  * grabs it again. Holds the lock when it returns.
 231  */
 232 static void wait_gp(void)
 233 {
 234         /*
 235          * Read reader_gp before read futex. smp_mb_master() needs to
 236          * be called with the rcu registry lock held in RCU_SIGNAL
 237          * flavor.
 238          */
 239         smp_mb_master();
 240         /* Temporarily unlock the registry lock. */
 241         mutex_unlock(&rcu_registry_lock);
 242         if (uatomic_read(&rcu_gp.futex) != -1)
 243                 goto end;
 244         while (futex_async(&rcu_gp.futex, FUTEX_WAIT, -1,
 245                         NULL, NULL, 0)) {
 246                 switch (errno) {
 247                 case EWOULDBLOCK:
 248                         /* Value already changed. */
 249                         goto end;
 250                 case EINTR:
 251                         /* Retry if interrupted by signal. */
 252                         break;  /* Get out of switch. */
 253                 default:
 254                         /* Unexpected error. */
 255                         urcu_die(errno);
 256                 }
 257         }
 258 end:
 259         /*
 260          * Re-lock the registry lock before the next loop.
 261          */
 262         mutex_lock(&rcu_registry_lock);
 263 }
 264
 265 /*
 266  * Always called with rcu_registry lock held. Releases this lock between
 267  * iterations and grabs it again. Holds the lock when it returns.
 268  */
 269 static void wait_for_readers(struct cds_list_head *input_readers,
 270                         struct cds_list_head *cur_snap_readers,
 271                         struct cds_list_head *qsreaders)
 272 {
 273         unsigned int wait_loops = 0;
 274         struct rcu_reader *index, *tmp;
 275 #ifdef HAS_INCOHERENT_CACHES
 276         unsigned int wait_gp_loops = 0;
 277 #endif /* HAS_INCOHERENT_CACHES */
 278
 279         /*
 280          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 281          * indicate quiescence (not nested), or observe the current
 282          * rcu_gp.ctr value.
 283          */
 284         for (;;) {
 285                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 286                         wait_loops++;
 287                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 288                         uatomic_dec(&rcu_gp.futex);
 289                         /* Write futex before read reader_gp */
 290                         smp_mb_master();
 291                 }
 292
 293                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 294                         switch (rcu_reader_state(&index->ctr)) {
 295                         case RCU_READER_ACTIVE_CURRENT:
 296                                 if (cur_snap_readers) {
 297                                         cds_list_move(&index->node,
 298                                                 cur_snap_readers);
 299                                         break;
 300                                 }
 301                                 /* Fall-through */
 302                         case RCU_READER_INACTIVE:
 303                                 cds_list_move(&index->node, qsreaders);
 304                                 break;
 305                         case RCU_READER_ACTIVE_OLD:
 306                                 /*
 307                                  * Old snapshot. Leaving node in
 308                                  * input_readers will make us busy-loop
 309                                  * until the snapshot becomes current or
 310                                  * the reader becomes inactive.
 311                                  */
 312                                 break;
 313                         }
 314                 }
 315
 316 #ifndef HAS_INCOHERENT_CACHES
 317                 if (cds_list_empty(input_readers)) {
 318                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 319                                 /* Read reader_gp before write futex */
 320                                 smp_mb_master();
 321                                 uatomic_set(&rcu_gp.futex, 0);
 322                         }
 323                         break;
 324                 } else {
 325                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 326                                 /* wait_gp unlocks/locks registry lock. */
 327                                 wait_gp();
 328                         } else {
 329                                 /* Temporarily unlock the registry lock. */
 330                                 mutex_unlock(&rcu_registry_lock);
 331                                 caa_cpu_relax();
 332                                 /*
 333                                  * Re-lock the registry lock before the
 334                                  * next loop.
 335                                  */
 336                                 mutex_lock(&rcu_registry_lock);
 337                         }
 338                 }
 339 #else /* #ifndef HAS_INCOHERENT_CACHES */
 340                 /*
 341                  * BUSY-LOOP. Force the reader thread to commit its
 342                  * URCU_TLS(rcu_reader).ctr update to memory if we wait
 343                  * for too long.
 344                  */
 345                 if (cds_list_empty(input_readers)) {
 346                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 347                                 /* Read reader_gp before write futex */
 348                                 smp_mb_master();
 349                                 uatomic_set(&rcu_gp.futex, 0);
 350                         }
 351                         break;
 352                 } else {
 353                         if (wait_gp_loops == KICK_READER_LOOPS) {
 354                                 smp_mb_master();
 355                                 wait_gp_loops = 0;
 356                         }
 357                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 358                                 /* wait_gp unlocks/locks registry lock. */
 359                                 wait_gp();
 360                                 wait_gp_loops++;
 361                         } else {
 362                                 /* Temporarily unlock the registry lock. */
 363                                 mutex_unlock(&rcu_registry_lock);
 364                                 caa_cpu_relax();
 365                                 /*
 366                                  * Re-lock the registry lock before the
 367                                  * next loop.
 368                                  */
 369                                 mutex_lock(&rcu_registry_lock);
 370                         }
 371                 }
 372 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 373         }
 374 }
 375
 376 void synchronize_rcu(void)
 377 {
 378         CDS_LIST_HEAD(cur_snap_readers);
 379         CDS_LIST_HEAD(qsreaders);
 380         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 381         struct urcu_waiters waiters;
 382
 383         /*
 384          * Add ourself to gp_waiters queue of threads awaiting to wait
 385          * for a grace period. Proceed to perform the grace period only
 386          * if we are the first thread added into the queue.
 387          * The implicit memory barrier before urcu_wait_add()
 388          * orders prior memory accesses of threads put into the wait
 389          * queue before their insertion into the wait queue.
 390          */
 391         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 392                 /* Not first in queue: will be awakened by another thread. */
 393                 urcu_adaptative_busy_wait(&wait);
 394                 /* Order following memory accesses after grace period. */
 395                 cmm_smp_mb();
 396                 return;
 397         }
 398         /* We won't need to wake ourself up */
 399         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 400
 401         mutex_lock(&rcu_gp_lock);
 402
 403         /*
 404          * Move all waiters into our local queue.
 405          */
 406         urcu_move_waiters(&waiters, &gp_waiters);
 407
 408         mutex_lock(&rcu_registry_lock);
 409
 410         if (cds_list_empty(&registry))
 411                 goto out;
 412
 413         /*
 414          * All threads should read qparity before accessing data structure
 415          * where new ptr points to. Must be done within rcu_registry_lock
 416          * because it iterates on reader threads.
 417          */
 418         /* Write new ptr before changing the qparity */
 419         smp_mb_master();
 420
 421         /*
 422          * Wait for readers to observe original parity or be quiescent.
 423          * wait_for_readers() can release and grab again rcu_registry_lock
 424          * interally.
 425          */
 426         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 427
 428         /*
 429          * Must finish waiting for quiescent state for original parity before
 430          * committing next rcu_gp.ctr update to memory. Failure to do so could
 431          * result in the writer waiting forever while new readers are always
 432          * accessing data (no progress).  Enforce compiler-order of load
 433          * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
 434          */
 435         cmm_barrier();
 436
 437         /*
 438          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 439          * model easier to understand. It does not have a big performance impact
 440          * anyway, given this is the write-side.
 441          */
 442         cmm_smp_mb();
 443
 444         /* Switch parity: 0 -> 1, 1 -> 0 */
 445         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
 446
 447         /*
 448          * Must commit rcu_gp.ctr update to memory before waiting for quiescent
 449          * state. Failure to do so could result in the writer waiting forever
 450          * while new readers are always accessing data (no progress). Enforce
 451          * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
 452          */
 453         cmm_barrier();
 454
 455         /*
 456          *
 457          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 458          * model easier to understand. It does not have a big performance impact
 459          * anyway, given this is the write-side.
 460          */
 461         cmm_smp_mb();
 462
 463         /*
 464          * Wait for readers to observe new parity or be quiescent.
 465          * wait_for_readers() can release and grab again rcu_registry_lock
 466          * interally.
 467          */
 468         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 469
 470         /*
 471          * Put quiescent reader list back into registry.
 472          */
 473         cds_list_splice(&qsreaders, &registry);
 474
 475         /*
 476          * Finish waiting for reader threads before letting the old ptr
 477          * being freed. Must be done within rcu_registry_lock because it
 478          * iterates on reader threads.
 479          */
 480         smp_mb_master();
 481 out:
 482         mutex_unlock(&rcu_registry_lock);
 483         mutex_unlock(&rcu_gp_lock);
 484
 485         /*
 486          * Wakeup waiters only after we have completed the grace period
 487          * and have ensured the memory barriers at the end of the grace
 488          * period have been issued.
 489          */
 490         urcu_wake_all_waiters(&waiters);
 491 }
 492
 493 /*
 494  * library wrappers to be used by non-LGPL compatible source code.
 495  */
 496
 497 void rcu_read_lock(void)
 498 {
 499         _rcu_read_lock();
 500 }
 501
 502 void rcu_read_unlock(void)
 503 {
 504         _rcu_read_unlock();
 505 }
 506
 507 int rcu_read_ongoing(void)
 508 {
 509         return _rcu_read_ongoing();
 510 }
 511
 512 void rcu_register_thread(void)
 513 {
 514         URCU_TLS(rcu_reader).tid = pthread_self();
 515         assert(URCU_TLS(rcu_reader).need_mb == 0);
 516         assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
 517
 518         mutex_lock(&rcu_registry_lock);
 519         assert(!URCU_TLS(rcu_reader).registered);
 520         URCU_TLS(rcu_reader).registered = 1;
 521         rcu_init();     /* In case gcc does not support constructor attribute */
 522         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 523         mutex_unlock(&rcu_registry_lock);
 524 }
 525
 526 void rcu_unregister_thread(void)
 527 {
 528         mutex_lock(&rcu_registry_lock);
 529         assert(URCU_TLS(rcu_reader).registered);
 530         URCU_TLS(rcu_reader).registered = 0;
 531         cds_list_del(&URCU_TLS(rcu_reader).node);
 532         mutex_unlock(&rcu_registry_lock);
 533 }
 534
 535 #ifdef RCU_MEMBARRIER
 536 void rcu_init(void)
 537 {
 538         int ret;
 539
 540         if (init_done)
 541                 return;
 542         init_done = 1;
 543         ret = membarrier(MEMBARRIER_CMD_QUERY, 0);
 544         if (ret >= 0 && (ret & MEMBARRIER_CMD_SHARED)) {
 545                 rcu_has_sys_membarrier = 1;
 546         }
 547 }
 548 #endif
 549
 550 #ifdef RCU_SIGNAL
 551 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
 552 {
 553         /*
 554          * Executing this cmm_smp_mb() is the only purpose of this signal handler.
 555          * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 556          * executed on.
 557          */
 558         cmm_smp_mb();
 559         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 560         cmm_smp_mb();
 561 }
 562
 563 /*
 564  * rcu_init constructor. Called when the library is linked, but also when
 565  * reader threads are calling rcu_register_thread().
 566  * Should only be called by a single thread at a given time. This is ensured by
 567  * holing the rcu_registry_lock from rcu_register_thread() or by running
 568  * at library load time, which should not be executed by multiple
 569  * threads nor concurrently with rcu_register_thread() anyway.
 570  */
 571 void rcu_init(void)
 572 {
 573         struct sigaction act;
 574         int ret;
 575
 576         if (init_done)
 577                 return;
 578         init_done = 1;
 579
 580         act.sa_sigaction = sigrcu_handler;
 581         act.sa_flags = SA_SIGINFO | SA_RESTART;
 582         sigemptyset(&act.sa_mask);
 583         ret = sigaction(SIGRCU, &act, NULL);
 584         if (ret)
 585                 urcu_die(errno);
 586 }
 587
 588 void rcu_exit(void)
 589 {
 590         /*
 591          * Don't unregister the SIGRCU signal handler anymore, because
 592          * call_rcu threads could still be using it shortly before the
 593          * application exits.
 594          * Assertion disabled because call_rcu threads are now rcu
 595          * readers, and left running at exit.
 596          * assert(cds_list_empty(&registry));
 597          */
 598 }
 599
 600 #endif /* #ifdef RCU_SIGNAL */
 601
 602 DEFINE_RCU_FLAVOR(rcu_flavor);
 603
 604 #include "urcu-call-rcu-impl.h"
 605 #include "urcu-defer-impl.h"