src/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 _LGPL_SOURCE
  28 #define _DEFAULT_SOURCE
  29 #include <stdio.h>
  30 #include <pthread.h>
  31 #include <signal.h>
  32 #include <assert.h>
  33 #include <stdlib.h>
  34 #include <stdint.h>
  35 #include <string.h>
  36 #include <errno.h>
  37 #include <poll.h>
  38
  39 #include "urcu/arch.h"
  40 #include "urcu/wfcqueue.h"
  41 #include "urcu/map/urcu.h"
  42 #include "urcu/static/urcu.h"
  43 #include "urcu-pointer.h"
  44 #include "urcu/tls-compat.h"
  45
  46 #include "urcu-die.h"
  47 #include "urcu-wait.h"
  48
  49 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  50 #undef _LGPL_SOURCE
  51 #include "urcu.h"
  52 #define _LGPL_SOURCE
  53
  54 /*
  55  * If a reader is really non-cooperative and refuses to commit its
  56  * rcu_active_readers count to memory (there is no barrier in the reader
  57  * per-se), kick it after 10 loops waiting for it.
  58  */
  59 #define KICK_READER_LOOPS       10
  60
  61 /*
  62  * Active attempts to check for reader Q.S. before calling futex().
  63  */
  64 #define RCU_QS_ACTIVE_ATTEMPTS 100
  65
  66 /* If the headers do not support membarrier system call, fall back on RCU_MB */
  67 #ifdef __NR_membarrier
  68 # define membarrier(...)                syscall(__NR_membarrier, __VA_ARGS__)
  69 #else
  70 # define membarrier(...)                -ENOSYS
  71 #endif
  72
  73 enum membarrier_cmd {
  74         MEMBARRIER_CMD_QUERY = 0,
  75         MEMBARRIER_CMD_SHARED = (1 << 0),
  76 };
  77
  78 #ifdef RCU_MEMBARRIER
  79 static int init_done;
  80 int rcu_has_sys_membarrier;
  81
  82 void __attribute__((constructor)) rcu_init(void);
  83 #endif
  84
  85 #ifdef RCU_MB
  86 void rcu_init(void)
  87 {
  88 }
  89 #endif
  90
  91 #ifdef RCU_SIGNAL
  92 static int init_done;
  93
  94 void __attribute__((constructor)) rcu_init(void);
  95 void __attribute__((destructor)) rcu_exit(void);
  96 #endif
  97
  98 /*
  99  * rcu_gp_lock ensures mutual exclusion between threads calling
 100  * synchronize_rcu().
 101  */
 102 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 103 /*
 104  * rcu_registry_lock ensures mutual exclusion between threads
 105  * registering and unregistering themselves to/from the registry, and
 106  * with threads reading that registry from synchronize_rcu(). However,
 107  * this lock is not held all the way through the completion of awaiting
 108  * for the grace period. It is sporadically released between iterations
 109  * on the registry.
 110  * rcu_registry_lock may nest inside rcu_gp_lock.
 111  */
 112 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 113 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
 114
 115 /*
 116  * Written to only by each individual reader. Read by both the reader and the
 117  * writers.
 118  */
 119 DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
 120
 121 static CDS_LIST_HEAD(registry);
 122
 123 /*
 124  * Queue keeping threads awaiting to wait for a grace period. Contains
 125  * struct gp_waiters_thread objects.
 126  */
 127 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
 128
 129 static void mutex_lock(pthread_mutex_t *mutex)
 130 {
 131         int ret;
 132
 133 #ifndef DISTRUST_SIGNALS_EXTREME
 134         ret = pthread_mutex_lock(mutex);
 135         if (ret)
 136                 urcu_die(ret);
 137 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 138         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 139                 if (ret != EBUSY && ret != EINTR)
 140                         urcu_die(ret);
 141                 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
 142                         cmm_smp_mb();
 143                         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 144                         cmm_smp_mb();
 145                 }
 146                 (void) poll(NULL, 0, 10);
 147         }
 148 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 149 }
 150
 151 static void mutex_unlock(pthread_mutex_t *mutex)
 152 {
 153         int ret;
 154
 155         ret = pthread_mutex_unlock(mutex);
 156         if (ret)
 157                 urcu_die(ret);
 158 }
 159
 160 #ifdef RCU_MEMBARRIER
 161 static void smp_mb_master(void)
 162 {
 163         if (caa_likely(rcu_has_sys_membarrier))
 164                 (void) membarrier(MEMBARRIER_CMD_SHARED, 0);
 165         else
 166                 cmm_smp_mb();
 167 }
 168 #endif
 169
 170 #ifdef RCU_MB
 171 static void smp_mb_master(void)
 172 {
 173         cmm_smp_mb();
 174 }
 175 #endif
 176
 177 #ifdef RCU_SIGNAL
 178 static void force_mb_all_readers(void)
 179 {
 180         struct rcu_reader *index;
 181
 182         /*
 183          * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 184          * compiler barriers around rcu read lock as real memory barriers.
 185          */
 186         if (cds_list_empty(&registry))
 187                 return;
 188         /*
 189          * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 190          * a cache flush on architectures with non-coherent cache. Let's play
 191          * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 192          * cache flush is enforced.
 193          */
 194         cds_list_for_each_entry(index, &registry, node) {
 195                 CMM_STORE_SHARED(index->need_mb, 1);
 196                 pthread_kill(index->tid, SIGRCU);
 197         }
 198         /*
 199          * Wait for sighandler (and thus mb()) to execute on every thread.
 200          *
 201          * Note that the pthread_kill() will never be executed on systems
 202          * that correctly deliver signals in a timely manner.  However, it
 203          * is not uncommon for kernels to have bugs that can result in
 204          * lost or unduly delayed signals.
 205          *
 206          * If you are seeing the below pthread_kill() executing much at
 207          * all, we suggest testing the underlying kernel and filing the
 208          * relevant bug report.  For Linux kernels, we recommend getting
 209          * the Linux Test Project (LTP).
 210          */
 211         cds_list_for_each_entry(index, &registry, node) {
 212                 while (CMM_LOAD_SHARED(index->need_mb)) {
 213                         pthread_kill(index->tid, SIGRCU);
 214                         (void) poll(NULL, 0, 1);
 215                 }
 216         }
 217         cmm_smp_mb();   /* read ->need_mb before ending the barrier */
 218 }
 219
 220 static void smp_mb_master(void)
 221 {
 222         force_mb_all_readers();
 223 }
 224 #endif /* #ifdef RCU_SIGNAL */
 225
 226 /*
 227  * synchronize_rcu() waiting. Single thread.
 228  * Always called with rcu_registry lock held. Releases this lock and
 229  * grabs it again. Holds the lock when it returns.
 230  */
 231 static void wait_gp(void)
 232 {
 233         /*
 234          * Read reader_gp before read futex. smp_mb_master() needs to
 235          * be called with the rcu registry lock held in RCU_SIGNAL
 236          * flavor.
 237          */
 238         smp_mb_master();
 239         /* Temporarily unlock the registry lock. */
 240         mutex_unlock(&rcu_registry_lock);
 241         if (uatomic_read(&rcu_gp.futex) != -1)
 242                 goto end;
 243         while (futex_async(&rcu_gp.futex, FUTEX_WAIT, -1,
 244                         NULL, NULL, 0)) {
 245                 switch (errno) {
 246                 case EWOULDBLOCK:
 247                         /* Value already changed. */
 248                         goto end;
 249                 case EINTR:
 250                         /* Retry if interrupted by signal. */
 251                         break;  /* Get out of switch. */
 252                 default:
 253                         /* Unexpected error. */
 254                         urcu_die(errno);
 255                 }
 256         }
 257 end:
 258         /*
 259          * Re-lock the registry lock before the next loop.
 260          */
 261         mutex_lock(&rcu_registry_lock);
 262 }
 263
 264 /*
 265  * Always called with rcu_registry lock held. Releases this lock between
 266  * iterations and grabs it again. Holds the lock when it returns.
 267  */
 268 static void wait_for_readers(struct cds_list_head *input_readers,
 269                         struct cds_list_head *cur_snap_readers,
 270                         struct cds_list_head *qsreaders)
 271 {
 272         unsigned int wait_loops = 0;
 273         struct rcu_reader *index, *tmp;
 274 #ifdef HAS_INCOHERENT_CACHES
 275         unsigned int wait_gp_loops = 0;
 276 #endif /* HAS_INCOHERENT_CACHES */
 277
 278         /*
 279          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 280          * indicate quiescence (not nested), or observe the current
 281          * rcu_gp.ctr value.
 282          */
 283         for (;;) {
 284                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 285                         wait_loops++;
 286                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 287                         uatomic_dec(&rcu_gp.futex);
 288                         /* Write futex before read reader_gp */
 289                         smp_mb_master();
 290                 }
 291
 292                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 293                         switch (rcu_reader_state(&index->ctr)) {
 294                         case RCU_READER_ACTIVE_CURRENT:
 295                                 if (cur_snap_readers) {
 296                                         cds_list_move(&index->node,
 297                                                 cur_snap_readers);
 298                                         break;
 299                                 }
 300                                 /* Fall-through */
 301                         case RCU_READER_INACTIVE:
 302                                 cds_list_move(&index->node, qsreaders);
 303                                 break;
 304                         case RCU_READER_ACTIVE_OLD:
 305                                 /*
 306                                  * Old snapshot. Leaving node in
 307                                  * input_readers will make us busy-loop
 308                                  * until the snapshot becomes current or
 309                                  * the reader becomes inactive.
 310                                  */
 311                                 break;
 312                         }
 313                 }
 314
 315 #ifndef HAS_INCOHERENT_CACHES
 316                 if (cds_list_empty(input_readers)) {
 317                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 318                                 /* Read reader_gp before write futex */
 319                                 smp_mb_master();
 320                                 uatomic_set(&rcu_gp.futex, 0);
 321                         }
 322                         break;
 323                 } else {
 324                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 325                                 /* wait_gp unlocks/locks registry lock. */
 326                                 wait_gp();
 327                         } else {
 328                                 /* Temporarily unlock the registry lock. */
 329                                 mutex_unlock(&rcu_registry_lock);
 330                                 caa_cpu_relax();
 331                                 /*
 332                                  * Re-lock the registry lock before the
 333                                  * next loop.
 334                                  */
 335                                 mutex_lock(&rcu_registry_lock);
 336                         }
 337                 }
 338 #else /* #ifndef HAS_INCOHERENT_CACHES */
 339                 /*
 340                  * BUSY-LOOP. Force the reader thread to commit its
 341                  * URCU_TLS(rcu_reader).ctr update to memory if we wait
 342                  * for too long.
 343                  */
 344                 if (cds_list_empty(input_readers)) {
 345                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 346                                 /* Read reader_gp before write futex */
 347                                 smp_mb_master();
 348                                 uatomic_set(&rcu_gp.futex, 0);
 349                         }
 350                         break;
 351                 } else {
 352                         if (wait_gp_loops == KICK_READER_LOOPS) {
 353                                 smp_mb_master();
 354                                 wait_gp_loops = 0;
 355                         }
 356                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 357                                 /* wait_gp unlocks/locks registry lock. */
 358                                 wait_gp();
 359                                 wait_gp_loops++;
 360                         } else {
 361                                 /* Temporarily unlock the registry lock. */
 362                                 mutex_unlock(&rcu_registry_lock);
 363                                 caa_cpu_relax();
 364                                 /*
 365                                  * Re-lock the registry lock before the
 366                                  * next loop.
 367                                  */
 368                                 mutex_lock(&rcu_registry_lock);
 369                         }
 370                 }
 371 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 372         }
 373 }
 374
 375 void synchronize_rcu(void)
 376 {
 377         CDS_LIST_HEAD(cur_snap_readers);
 378         CDS_LIST_HEAD(qsreaders);
 379         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 380         struct urcu_waiters waiters;
 381
 382         /*
 383          * Add ourself to gp_waiters queue of threads awaiting to wait
 384          * for a grace period. Proceed to perform the grace period only
 385          * if we are the first thread added into the queue.
 386          * The implicit memory barrier before urcu_wait_add()
 387          * orders prior memory accesses of threads put into the wait
 388          * queue before their insertion into the wait queue.
 389          */
 390         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 391                 /* Not first in queue: will be awakened by another thread. */
 392                 urcu_adaptative_busy_wait(&wait);
 393                 /* Order following memory accesses after grace period. */
 394                 cmm_smp_mb();
 395                 return;
 396         }
 397         /* We won't need to wake ourself up */
 398         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 399
 400         mutex_lock(&rcu_gp_lock);
 401
 402         /*
 403          * Move all waiters into our local queue.
 404          */
 405         urcu_move_waiters(&waiters, &gp_waiters);
 406
 407         mutex_lock(&rcu_registry_lock);
 408
 409         if (cds_list_empty(&registry))
 410                 goto out;
 411
 412         /*
 413          * All threads should read qparity before accessing data structure
 414          * where new ptr points to. Must be done within rcu_registry_lock
 415          * because it iterates on reader threads.
 416          */
 417         /* Write new ptr before changing the qparity */
 418         smp_mb_master();
 419
 420         /*
 421          * Wait for readers to observe original parity or be quiescent.
 422          * wait_for_readers() can release and grab again rcu_registry_lock
 423          * interally.
 424          */
 425         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 426
 427         /*
 428          * Must finish waiting for quiescent state for original parity before
 429          * committing next rcu_gp.ctr update to memory. Failure to do so could
 430          * result in the writer waiting forever while new readers are always
 431          * accessing data (no progress).  Enforce compiler-order of load
 432          * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
 433          */
 434         cmm_barrier();
 435
 436         /*
 437          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 438          * model easier to understand. It does not have a big performance impact
 439          * anyway, given this is the write-side.
 440          */
 441         cmm_smp_mb();
 442
 443         /* Switch parity: 0 -> 1, 1 -> 0 */
 444         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
 445
 446         /*
 447          * Must commit rcu_gp.ctr update to memory before waiting for quiescent
 448          * state. Failure to do so could result in the writer waiting forever
 449          * while new readers are always accessing data (no progress). Enforce
 450          * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
 451          */
 452         cmm_barrier();
 453
 454         /*
 455          *
 456          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 457          * model easier to understand. It does not have a big performance impact
 458          * anyway, given this is the write-side.
 459          */
 460         cmm_smp_mb();
 461
 462         /*
 463          * Wait for readers to observe new parity or be quiescent.
 464          * wait_for_readers() can release and grab again rcu_registry_lock
 465          * interally.
 466          */
 467         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 468
 469         /*
 470          * Put quiescent reader list back into registry.
 471          */
 472         cds_list_splice(&qsreaders, &registry);
 473
 474         /*
 475          * Finish waiting for reader threads before letting the old ptr
 476          * being freed. Must be done within rcu_registry_lock because it
 477          * iterates on reader threads.
 478          */
 479         smp_mb_master();
 480 out:
 481         mutex_unlock(&rcu_registry_lock);
 482         mutex_unlock(&rcu_gp_lock);
 483
 484         /*
 485          * Wakeup waiters only after we have completed the grace period
 486          * and have ensured the memory barriers at the end of the grace
 487          * period have been issued.
 488          */
 489         urcu_wake_all_waiters(&waiters);
 490 }
 491
 492 /*
 493  * library wrappers to be used by non-LGPL compatible source code.
 494  */
 495
 496 void rcu_read_lock(void)
 497 {
 498         _rcu_read_lock();
 499 }
 500
 501 void rcu_read_unlock(void)
 502 {
 503         _rcu_read_unlock();
 504 }
 505
 506 int rcu_read_ongoing(void)
 507 {
 508         return _rcu_read_ongoing();
 509 }
 510
 511 void rcu_register_thread(void)
 512 {
 513         URCU_TLS(rcu_reader).tid = pthread_self();
 514         assert(URCU_TLS(rcu_reader).need_mb == 0);
 515         assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
 516
 517         mutex_lock(&rcu_registry_lock);
 518         assert(!URCU_TLS(rcu_reader).registered);
 519         URCU_TLS(rcu_reader).registered = 1;
 520         rcu_init();     /* In case gcc does not support constructor attribute */
 521         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 522         mutex_unlock(&rcu_registry_lock);
 523 }
 524
 525 void rcu_unregister_thread(void)
 526 {
 527         mutex_lock(&rcu_registry_lock);
 528         assert(URCU_TLS(rcu_reader).registered);
 529         URCU_TLS(rcu_reader).registered = 0;
 530         cds_list_del(&URCU_TLS(rcu_reader).node);
 531         mutex_unlock(&rcu_registry_lock);
 532 }
 533
 534 #ifdef RCU_MEMBARRIER
 535 void rcu_init(void)
 536 {
 537         int ret;
 538
 539         if (init_done)
 540                 return;
 541         init_done = 1;
 542         ret = membarrier(MEMBARRIER_CMD_QUERY, 0);
 543         if (ret >= 0 && (ret & MEMBARRIER_CMD_SHARED)) {
 544                 rcu_has_sys_membarrier = 1;
 545         }
 546 }
 547 #endif
 548
 549 #ifdef RCU_SIGNAL
 550 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
 551 {
 552         /*
 553          * Executing this cmm_smp_mb() is the only purpose of this signal handler.
 554          * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 555          * executed on.
 556          */
 557         cmm_smp_mb();
 558         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 559         cmm_smp_mb();
 560 }
 561
 562 /*
 563  * rcu_init constructor. Called when the library is linked, but also when
 564  * reader threads are calling rcu_register_thread().
 565  * Should only be called by a single thread at a given time. This is ensured by
 566  * holing the rcu_registry_lock from rcu_register_thread() or by running
 567  * at library load time, which should not be executed by multiple
 568  * threads nor concurrently with rcu_register_thread() anyway.
 569  */
 570 void rcu_init(void)
 571 {
 572         struct sigaction act;
 573         int ret;
 574
 575         if (init_done)
 576                 return;
 577         init_done = 1;
 578
 579         act.sa_sigaction = sigrcu_handler;
 580         act.sa_flags = SA_SIGINFO | SA_RESTART;
 581         sigemptyset(&act.sa_mask);
 582         ret = sigaction(SIGRCU, &act, NULL);
 583         if (ret)
 584                 urcu_die(errno);
 585 }
 586
 587 void rcu_exit(void)
 588 {
 589         /*
 590          * Don't unregister the SIGRCU signal handler anymore, because
 591          * call_rcu threads could still be using it shortly before the
 592          * application exits.
 593          * Assertion disabled because call_rcu threads are now rcu
 594          * readers, and left running at exit.
 595          * assert(cds_list_empty(&registry));
 596          */
 597 }
 598
 599 #endif /* #ifdef RCU_SIGNAL */
 600
 601 DEFINE_RCU_FLAVOR(rcu_flavor);
 602
 603 #include "urcu-call-rcu-impl.h"
 604 #include "urcu-defer-impl.h"