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