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 #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/wfcqueue.h"
  40 #include "urcu/map/urcu.h"
  41 #include "urcu/static/urcu.h"
  42 #include "urcu-pointer.h"
  43 #include "urcu/tls-compat.h"
  44
  45 #include "urcu-die.h"
  46 #include "urcu-wait.h"
  47
  48 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  49 #undef _LGPL_SOURCE
  50 #include "urcu.h"
  51 #define _LGPL_SOURCE
  52
  53 /*
  54  * If a reader is really non-cooperative and refuses to commit its
  55  * rcu_active_readers count to memory (there is no barrier in the reader
  56  * per-se), kick it after a few loops waiting for it.
  57  */
  58 #define KICK_READER_LOOPS 10000
  59
  60 /*
  61  * Active attempts to check for reader Q.S. before calling futex().
  62  */
  63 #define RCU_QS_ACTIVE_ATTEMPTS 100
  64
  65 #ifdef RCU_MEMBARRIER
  66 static int init_done;
  67 int rcu_has_sys_membarrier;
  68
  69 void __attribute__((constructor)) rcu_init(void);
  70 #endif
  71
  72 #ifdef RCU_MB
  73 void rcu_init(void)
  74 {
  75 }
  76 #endif
  77
  78 #ifdef RCU_SIGNAL
  79 static int init_done;
  80
  81 void __attribute__((constructor)) rcu_init(void);
  82 void __attribute__((destructor)) rcu_exit(void);
  83 #endif
  84
  85 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  86
  87 int32_t rcu_gp_futex;
  88
  89 /*
  90  * Global grace period counter.
  91  * Contains the current RCU_GP_CTR_PHASE.
  92  * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
  93  * Written to only by writer with mutex taken. Read by both writer and readers.
  94  */
  95 unsigned long rcu_gp_ctr = RCU_GP_COUNT;
  96
  97 /*
  98  * Written to only by each individual reader. Read by both the reader and the
  99  * writers.
 100  */
 101 DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
 102
 103 #ifdef DEBUG_YIELD
 104 unsigned int rcu_yield_active;
 105 DEFINE_URCU_TLS(unsigned int, rcu_rand_yield);
 106 #endif
 107
 108 static CDS_LIST_HEAD(registry);
 109
 110 /*
 111  * Queue keeping threads awaiting to wait for a grace period. Contains
 112  * struct gp_waiters_thread objects.
 113  */
 114 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
 115
 116 static void mutex_lock(pthread_mutex_t *mutex)
 117 {
 118         int ret;
 119
 120 #ifndef DISTRUST_SIGNALS_EXTREME
 121         ret = pthread_mutex_lock(mutex);
 122         if (ret)
 123                 urcu_die(ret);
 124 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 125         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 126                 if (ret != EBUSY && ret != EINTR)
 127                         urcu_die(ret);
 128                 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
 129                         cmm_smp_mb();
 130                         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 131                         cmm_smp_mb();
 132                 }
 133                 poll(NULL,0,10);
 134         }
 135 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 136 }
 137
 138 static void mutex_unlock(pthread_mutex_t *mutex)
 139 {
 140         int ret;
 141
 142         ret = pthread_mutex_unlock(mutex);
 143         if (ret)
 144                 urcu_die(ret);
 145 }
 146
 147 #ifdef RCU_MEMBARRIER
 148 static void smp_mb_master(int group)
 149 {
 150         if (caa_likely(rcu_has_sys_membarrier))
 151                 membarrier(MEMBARRIER_EXPEDITED);
 152         else
 153                 cmm_smp_mb();
 154 }
 155 #endif
 156
 157 #ifdef RCU_MB
 158 static void smp_mb_master(int group)
 159 {
 160         cmm_smp_mb();
 161 }
 162 #endif
 163
 164 #ifdef RCU_SIGNAL
 165 static void force_mb_all_readers(void)
 166 {
 167         struct rcu_reader *index;
 168
 169         /*
 170          * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 171          * compiler barriers around rcu read lock as real memory barriers.
 172          */
 173         if (cds_list_empty(&registry))
 174                 return;
 175         /*
 176          * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 177          * a cache flush on architectures with non-coherent cache. Let's play
 178          * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 179          * cache flush is enforced.
 180          */
 181         cds_list_for_each_entry(index, &registry, node) {
 182                 CMM_STORE_SHARED(index->need_mb, 1);
 183                 pthread_kill(index->tid, SIGRCU);
 184         }
 185         /*
 186          * Wait for sighandler (and thus mb()) to execute on every thread.
 187          *
 188          * Note that the pthread_kill() will never be executed on systems
 189          * that correctly deliver signals in a timely manner.  However, it
 190          * is not uncommon for kernels to have bugs that can result in
 191          * lost or unduly delayed signals.
 192          *
 193          * If you are seeing the below pthread_kill() executing much at
 194          * all, we suggest testing the underlying kernel and filing the
 195          * relevant bug report.  For Linux kernels, we recommend getting
 196          * the Linux Test Project (LTP).
 197          */
 198         cds_list_for_each_entry(index, &registry, node) {
 199                 while (CMM_LOAD_SHARED(index->need_mb)) {
 200                         pthread_kill(index->tid, SIGRCU);
 201                         poll(NULL, 0, 1);
 202                 }
 203         }
 204         cmm_smp_mb();   /* read ->need_mb before ending the barrier */
 205 }
 206
 207 static void smp_mb_master(int group)
 208 {
 209         force_mb_all_readers();
 210 }
 211 #endif /* #ifdef RCU_SIGNAL */
 212
 213 /*
 214  * synchronize_rcu() waiting. Single thread.
 215  */
 216 static void wait_gp(void)
 217 {
 218         /* Read reader_gp before read futex */
 219         smp_mb_master(RCU_MB_GROUP);
 220         if (uatomic_read(&rcu_gp_futex) == -1)
 221                 futex_async(&rcu_gp_futex, FUTEX_WAIT, -1,
 222                       NULL, NULL, 0);
 223 }
 224
 225 static void wait_for_readers(struct cds_list_head *input_readers,
 226                         struct cds_list_head *cur_snap_readers,
 227                         struct cds_list_head *qsreaders)
 228 {
 229         int wait_loops = 0;
 230         struct rcu_reader *index, *tmp;
 231
 232         /*
 233          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 234          * indicate quiescence (not nested), or observe the current
 235          * rcu_gp_ctr value.
 236          */
 237         for (;;) {
 238                 wait_loops++;
 239                 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 240                         uatomic_dec(&rcu_gp_futex);
 241                         /* Write futex before read reader_gp */
 242                         smp_mb_master(RCU_MB_GROUP);
 243                 }
 244
 245                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 246                         switch (rcu_reader_state(&index->ctr)) {
 247                         case RCU_READER_ACTIVE_CURRENT:
 248                                 if (cur_snap_readers) {
 249                                         cds_list_move(&index->node,
 250                                                 cur_snap_readers);
 251                                         break;
 252                                 }
 253                                 /* Fall-through */
 254                         case RCU_READER_INACTIVE:
 255                                 cds_list_move(&index->node, qsreaders);
 256                                 break;
 257                         case RCU_READER_ACTIVE_OLD:
 258                                 /*
 259                                  * Old snapshot. Leaving node in
 260                                  * input_readers will make us busy-loop
 261                                  * until the snapshot becomes current or
 262                                  * the reader becomes inactive.
 263                                  */
 264                                 break;
 265                         }
 266                 }
 267
 268 #ifndef HAS_INCOHERENT_CACHES
 269                 if (cds_list_empty(input_readers)) {
 270                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 271                                 /* Read reader_gp before write futex */
 272                                 smp_mb_master(RCU_MB_GROUP);
 273                                 uatomic_set(&rcu_gp_futex, 0);
 274                         }
 275                         break;
 276                 } else {
 277                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 278                                 wait_gp();
 279                         else
 280                                 caa_cpu_relax();
 281                 }
 282 #else /* #ifndef HAS_INCOHERENT_CACHES */
 283                 /*
 284                  * BUSY-LOOP. Force the reader thread to commit its
 285                  * URCU_TLS(rcu_reader).ctr update to memory if we wait
 286                  * for too long.
 287                  */
 288                 if (cds_list_empty(input_readers)) {
 289                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 290                                 /* Read reader_gp before write futex */
 291                                 smp_mb_master(RCU_MB_GROUP);
 292                                 uatomic_set(&rcu_gp_futex, 0);
 293                         }
 294                         break;
 295                 } else {
 296                         switch (wait_loops) {
 297                         case RCU_QS_ACTIVE_ATTEMPTS:
 298                                 wait_gp();
 299                                 break; /* only escape switch */
 300                         case KICK_READER_LOOPS:
 301                                 smp_mb_master(RCU_MB_GROUP);
 302                                 wait_loops = 0;
 303                                 break; /* only escape switch */
 304                         default:
 305                                 caa_cpu_relax();
 306                         }
 307                 }
 308 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 309         }
 310 }
 311
 312 void synchronize_rcu(void)
 313 {
 314         CDS_LIST_HEAD(cur_snap_readers);
 315         CDS_LIST_HEAD(qsreaders);
 316         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 317         struct urcu_waiters waiters;
 318
 319         /*
 320          * Add ourself to gp_waiters queue of threads awaiting to wait
 321          * for a grace period. Proceed to perform the grace period only
 322          * if we are the first thread added into the queue.
 323          * The implicit memory barrier before urcu_wait_add()
 324          * orders prior memory accesses of threads put into the wait
 325          * queue before their insertion into the wait queue.
 326          */
 327         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 328                 /* Not first in queue: will be awakened by another thread. */
 329                 urcu_adaptative_busy_wait(&wait);
 330                 /* Order following memory accesses after grace period. */
 331                 cmm_smp_mb();
 332                 return;
 333         }
 334         /* We won't need to wake ourself up */
 335         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 336
 337         mutex_lock(&rcu_gp_lock);
 338
 339         /*
 340          * Move all waiters into our local queue.
 341          */
 342         urcu_move_waiters(&waiters, &gp_waiters);
 343
 344         if (cds_list_empty(&registry))
 345                 goto out;
 346
 347         /* All threads should read qparity before accessing data structure
 348          * where new ptr points to. Must be done within rcu_gp_lock because it
 349          * iterates on reader threads.*/
 350         /* Write new ptr before changing the qparity */
 351         smp_mb_master(RCU_MB_GROUP);
 352
 353         /*
 354          * Wait for readers to observe original parity or be quiescent.
 355          */
 356         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 357
 358         /*
 359          * Must finish waiting for quiescent state for original parity before
 360          * committing next rcu_gp_ctr update to memory. Failure to do so could
 361          * result in the writer waiting forever while new readers are always
 362          * accessing data (no progress).  Enforce compiler-order of load
 363          * URCU_TLS(rcu_reader).ctr before store to rcu_gp_ctr.
 364          */
 365         cmm_barrier();
 366
 367         /*
 368          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 369          * model easier to understand. It does not have a big performance impact
 370          * anyway, given this is the write-side.
 371          */
 372         cmm_smp_mb();
 373
 374         /* Switch parity: 0 -> 1, 1 -> 0 */
 375         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
 376
 377         /*
 378          * Must commit rcu_gp_ctr update to memory before waiting for quiescent
 379          * state. Failure to do so could result in the writer waiting forever
 380          * while new readers are always accessing data (no progress). Enforce
 381          * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr.
 382          */
 383         cmm_barrier();
 384
 385         /*
 386          *
 387          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 388          * model easier to understand. It does not have a big performance impact
 389          * anyway, given this is the write-side.
 390          */
 391         cmm_smp_mb();
 392
 393         /*
 394          * Wait for readers to observe new parity or be quiescent.
 395          */
 396         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 397
 398         /*
 399          * Put quiescent reader list back into registry.
 400          */
 401         cds_list_splice(&qsreaders, &registry);
 402
 403         /* Finish waiting for reader threads before letting the old ptr being
 404          * freed. Must be done within rcu_gp_lock because it iterates on reader
 405          * threads. */
 406         smp_mb_master(RCU_MB_GROUP);
 407 out:
 408         mutex_unlock(&rcu_gp_lock);
 409
 410         /*
 411          * Wakeup waiters only after we have completed the grace period
 412          * and have ensured the memory barriers at the end of the grace
 413          * period have been issued.
 414          */
 415         urcu_wake_all_waiters(&waiters);
 416 }
 417
 418 /*
 419  * library wrappers to be used by non-LGPL compatible source code.
 420  */
 421
 422 void rcu_read_lock(void)
 423 {
 424         _rcu_read_lock();
 425 }
 426
 427 void rcu_read_unlock(void)
 428 {
 429         _rcu_read_unlock();
 430 }
 431
 432 int rcu_read_ongoing(void)
 433 {
 434         return _rcu_read_ongoing();
 435 }
 436
 437 void rcu_register_thread(void)
 438 {
 439         URCU_TLS(rcu_reader).tid = pthread_self();
 440         assert(URCU_TLS(rcu_reader).need_mb == 0);
 441         assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
 442
 443         mutex_lock(&rcu_gp_lock);
 444         rcu_init();     /* In case gcc does not support constructor attribute */
 445         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 446         mutex_unlock(&rcu_gp_lock);
 447 }
 448
 449 void rcu_unregister_thread(void)
 450 {
 451         mutex_lock(&rcu_gp_lock);
 452         cds_list_del(&URCU_TLS(rcu_reader).node);
 453         mutex_unlock(&rcu_gp_lock);
 454 }
 455
 456 #ifdef RCU_MEMBARRIER
 457 void rcu_init(void)
 458 {
 459         if (init_done)
 460                 return;
 461         init_done = 1;
 462         if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
 463                 rcu_has_sys_membarrier = 1;
 464 }
 465 #endif
 466
 467 #ifdef RCU_SIGNAL
 468 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
 469 {
 470         /*
 471          * Executing this cmm_smp_mb() is the only purpose of this signal handler.
 472          * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 473          * executed on.
 474          */
 475         cmm_smp_mb();
 476         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 477         cmm_smp_mb();
 478 }
 479
 480 /*
 481  * rcu_init constructor. Called when the library is linked, but also when
 482  * reader threads are calling rcu_register_thread().
 483  * Should only be called by a single thread at a given time. This is ensured by
 484  * holing the rcu_gp_lock from rcu_register_thread() or by running at library
 485  * load time, which should not be executed by multiple threads nor concurrently
 486  * with rcu_register_thread() anyway.
 487  */
 488 void rcu_init(void)
 489 {
 490         struct sigaction act;
 491         int ret;
 492
 493         if (init_done)
 494                 return;
 495         init_done = 1;
 496
 497         act.sa_sigaction = sigrcu_handler;
 498         act.sa_flags = SA_SIGINFO | SA_RESTART;
 499         sigemptyset(&act.sa_mask);
 500         ret = sigaction(SIGRCU, &act, NULL);
 501         if (ret)
 502                 urcu_die(errno);
 503 }
 504
 505 void rcu_exit(void)
 506 {
 507         struct sigaction act;
 508         int ret;
 509
 510         ret = sigaction(SIGRCU, NULL, &act);
 511         if (ret)
 512                 urcu_die(errno);
 513         assert(act.sa_sigaction == sigrcu_handler);
 514         assert(cds_list_empty(&registry));
 515 }
 516
 517 #endif /* #ifdef RCU_SIGNAL */
 518
 519 DEFINE_RCU_FLAVOR(rcu_flavor);
 520
 521 #include "urcu-call-rcu-impl.h"
 522 #include "urcu-defer-impl.h"