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 #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
  37 #include "urcu/map/urcu.h"
  38
  39 #include "urcu/static/urcu.h"
  40 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  41 #include "urcu.h"
  42
  43 #ifdef RCU_MEMBARRIER
  44 static int init_done;
  45 int has_sys_membarrier;
  46
  47 void __attribute__((constructor)) rcu_init(void);
  48 #endif
  49
  50 #ifdef RCU_MB
  51 void rcu_init(void)
  52 {
  53 }
  54 #endif
  55
  56 #ifdef RCU_SIGNAL
  57 static int init_done;
  58
  59 void __attribute__((constructor)) rcu_init(void);
  60 void __attribute__((destructor)) rcu_exit(void);
  61 #endif
  62
  63 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  64
  65 int gp_futex;
  66
  67 /*
  68  * Global grace period counter.
  69  * Contains the current RCU_GP_CTR_PHASE.
  70  * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
  71  * Written to only by writer with mutex taken. Read by both writer and readers.
  72  */
  73 unsigned long rcu_gp_ctr = RCU_GP_COUNT;
  74
  75 /*
  76  * Written to only by each individual reader. Read by both the reader and the
  77  * writers.
  78  */
  79 struct rcu_reader __thread rcu_reader;
  80
  81 #ifdef DEBUG_YIELD
  82 unsigned int yield_active;
  83 unsigned int __thread rand_yield;
  84 #endif
  85
  86 static CDS_LIST_HEAD(registry);
  87
  88 static void mutex_lock(pthread_mutex_t *mutex)
  89 {
  90         int ret;
  91
  92 #ifndef DISTRUST_SIGNALS_EXTREME
  93         ret = pthread_mutex_lock(mutex);
  94         if (ret) {
  95                 perror("Error in pthread mutex lock");
  96                 exit(-1);
  97         }
  98 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  99         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 100                 if (ret != EBUSY && ret != EINTR) {
 101                         printf("ret = %d, errno = %d\n", ret, errno);
 102                         perror("Error in pthread mutex lock");
 103                         exit(-1);
 104                 }
 105                 if (CMM_LOAD_SHARED(rcu_reader.need_mb)) {
 106                         cmm_smp_mb();
 107                         _CMM_STORE_SHARED(rcu_reader.need_mb, 0);
 108                         cmm_smp_mb();
 109                 }
 110                 poll(NULL,0,10);
 111         }
 112 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 113 }
 114
 115 static void mutex_unlock(pthread_mutex_t *mutex)
 116 {
 117         int ret;
 118
 119         ret = pthread_mutex_unlock(mutex);
 120         if (ret) {
 121                 perror("Error in pthread mutex unlock");
 122                 exit(-1);
 123         }
 124 }
 125
 126 #ifdef RCU_MEMBARRIER
 127 static void smp_mb_master(int group)
 128 {
 129         if (likely(has_sys_membarrier))
 130                 membarrier(MEMBARRIER_EXPEDITED);
 131         else
 132                 cmm_smp_mb();
 133 }
 134 #endif
 135
 136 #ifdef RCU_MB
 137 static void smp_mb_master(int group)
 138 {
 139         cmm_smp_mb();
 140 }
 141 #endif
 142
 143 #ifdef RCU_SIGNAL
 144 static void force_mb_all_readers(void)
 145 {
 146         struct rcu_reader *index;
 147
 148         /*
 149          * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 150          * compiler barriers around rcu read lock as real memory barriers.
 151          */
 152         if (cds_list_empty(&registry))
 153                 return;
 154         /*
 155          * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 156          * a cache flush on architectures with non-coherent cache. Let's play
 157          * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 158          * cache flush is enforced.
 159          */
 160         cds_list_for_each_entry(index, &registry, node) {
 161                 CMM_STORE_SHARED(index->need_mb, 1);
 162                 pthread_kill(index->tid, SIGRCU);
 163         }
 164         /*
 165          * Wait for sighandler (and thus mb()) to execute on every thread.
 166          *
 167          * Note that the pthread_kill() will never be executed on systems
 168          * that correctly deliver signals in a timely manner.  However, it
 169          * is not uncommon for kernels to have bugs that can result in
 170          * lost or unduly delayed signals.
 171          *
 172          * If you are seeing the below pthread_kill() executing much at
 173          * all, we suggest testing the underlying kernel and filing the
 174          * relevant bug report.  For Linux kernels, we recommend getting
 175          * the Linux Test Project (LTP).
 176          */
 177         cds_list_for_each_entry(index, &registry, node) {
 178                 while (CMM_LOAD_SHARED(index->need_mb)) {
 179                         pthread_kill(index->tid, SIGRCU);
 180                         poll(NULL, 0, 1);
 181                 }
 182         }
 183         cmm_smp_mb();   /* read ->need_mb before ending the barrier */
 184 }
 185
 186 static void smp_mb_master(int group)
 187 {
 188         force_mb_all_readers();
 189 }
 190 #endif /* #ifdef RCU_SIGNAL */
 191
 192 /*
 193  * synchronize_rcu() waiting. Single thread.
 194  */
 195 static void wait_gp(void)
 196 {
 197         /* Read reader_gp before read futex */
 198         smp_mb_master(RCU_MB_GROUP);
 199         if (uatomic_read(&gp_futex) == -1)
 200                 futex_async(&gp_futex, FUTEX_WAIT, -1,
 201                       NULL, NULL, 0);
 202 }
 203
 204 void update_counter_and_wait(void)
 205 {
 206         CDS_LIST_HEAD(qsreaders);
 207         int wait_loops = 0;
 208         struct rcu_reader *index, *tmp;
 209
 210         /* Switch parity: 0 -> 1, 1 -> 0 */
 211         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
 212
 213         /*
 214          * Must commit rcu_gp_ctr update to memory before waiting for quiescent
 215          * state. Failure to do so could result in the writer waiting forever
 216          * while new readers are always accessing data (no progress). Enforce
 217          * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr.
 218          */
 219         cmm_barrier();
 220
 221         /*
 222          *
 223          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 224          * model easier to understand. It does not have a big performance impact
 225          * anyway, given this is the write-side.
 226          */
 227         cmm_smp_mb();
 228
 229         /*
 230          * Wait for each thread rcu_reader.ctr count to become 0.
 231          */
 232         for (;;) {
 233                 wait_loops++;
 234                 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 235                         uatomic_dec(&gp_futex);
 236                         /* Write futex before read reader_gp */
 237                         smp_mb_master(RCU_MB_GROUP);
 238                 }
 239
 240                 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
 241                         if (!rcu_gp_ongoing(&index->ctr))
 242                                 cds_list_move(&index->node, &qsreaders);
 243                 }
 244
 245 #ifndef HAS_INCOHERENT_CACHES
 246                 if (cds_list_empty(&registry)) {
 247                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 248                                 /* Read reader_gp before write futex */
 249                                 smp_mb_master(RCU_MB_GROUP);
 250                                 uatomic_set(&gp_futex, 0);
 251                         }
 252                         break;
 253                 } else {
 254                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 255                                 wait_gp();
 256                         else
 257                                 caa_cpu_relax();
 258                 }
 259 #else /* #ifndef HAS_INCOHERENT_CACHES */
 260                 /*
 261                  * BUSY-LOOP. Force the reader thread to commit its
 262                  * rcu_reader.ctr update to memory if we wait for too long.
 263                  */
 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                         switch (wait_loops) {
 273                         case RCU_QS_ACTIVE_ATTEMPTS:
 274                                 wait_gp();
 275                                 break; /* only escape switch */
 276                         case KICK_READER_LOOPS:
 277                                 smp_mb_master(RCU_MB_GROUP);
 278                                 wait_loops = 0;
 279                                 break; /* only escape switch */
 280                         default:
 281                                 caa_cpu_relax();
 282                         }
 283                 }
 284 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 285         }
 286         /* put back the reader list in the registry */
 287         cds_list_splice(&qsreaders, &registry);
 288 }
 289
 290 void synchronize_rcu(void)
 291 {
 292         mutex_lock(&rcu_gp_lock);
 293
 294         if (cds_list_empty(&registry))
 295                 goto out;
 296
 297         /* All threads should read qparity before accessing data structure
 298          * where new ptr points to. Must be done within rcu_gp_lock because it
 299          * iterates on reader threads.*/
 300         /* Write new ptr before changing the qparity */
 301         smp_mb_master(RCU_MB_GROUP);
 302
 303         /*
 304          * Wait for previous parity to be empty of readers.
 305          */
 306         update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 307
 308         /*
 309          * Must finish waiting for quiescent state for parity 0 before
 310          * committing next rcu_gp_ctr update to memory. Failure to do so could
 311          * result in the writer waiting forever while new readers are always
 312          * accessing data (no progress).  Enforce compiler-order of load
 313          * rcu_reader ctr before store to rcu_gp_ctr.
 314          */
 315         cmm_barrier();
 316
 317         /*
 318          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 319          * model easier to understand. It does not have a big performance impact
 320          * anyway, given this is the write-side.
 321          */
 322         cmm_smp_mb();
 323
 324         /*
 325          * Wait for previous parity to be empty of readers.
 326          */
 327         update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
 328
 329         /* Finish waiting for reader threads before letting the old ptr being
 330          * freed. Must be done within rcu_gp_lock because it iterates on reader
 331          * threads. */
 332         smp_mb_master(RCU_MB_GROUP);
 333 out:
 334         mutex_unlock(&rcu_gp_lock);
 335 }
 336
 337 /*
 338  * library wrappers to be used by non-LGPL compatible source code.
 339  */
 340
 341 void rcu_read_lock(void)
 342 {
 343         _rcu_read_lock();
 344 }
 345
 346 void rcu_read_unlock(void)
 347 {
 348         _rcu_read_unlock();
 349 }
 350
 351 void rcu_register_thread(void)
 352 {
 353         rcu_reader.tid = pthread_self();
 354         assert(rcu_reader.need_mb == 0);
 355         assert(!(rcu_reader.ctr & RCU_GP_CTR_NEST_MASK));
 356
 357         mutex_lock(&rcu_gp_lock);
 358         rcu_init();     /* In case gcc does not support constructor attribute */
 359         cds_list_add(&rcu_reader.node, &registry);
 360         mutex_unlock(&rcu_gp_lock);
 361 }
 362
 363 void rcu_unregister_thread(void)
 364 {
 365         mutex_lock(&rcu_gp_lock);
 366         cds_list_del(&rcu_reader.node);
 367         mutex_unlock(&rcu_gp_lock);
 368 }
 369
 370 #ifdef RCU_MEMBARRIER
 371 void rcu_init(void)
 372 {
 373         if (init_done)
 374                 return;
 375         init_done = 1;
 376         if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
 377                 has_sys_membarrier = 1;
 378 }
 379 #endif
 380
 381 #ifdef RCU_SIGNAL
 382 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
 383 {
 384         /*
 385          * Executing this cmm_smp_mb() is the only purpose of this signal handler.
 386          * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 387          * executed on.
 388          */
 389         cmm_smp_mb();
 390         _CMM_STORE_SHARED(rcu_reader.need_mb, 0);
 391         cmm_smp_mb();
 392 }
 393
 394 /*
 395  * rcu_init constructor. Called when the library is linked, but also when
 396  * reader threads are calling rcu_register_thread().
 397  * Should only be called by a single thread at a given time. This is ensured by
 398  * holing the rcu_gp_lock from rcu_register_thread() or by running at library
 399  * load time, which should not be executed by multiple threads nor concurrently
 400  * with rcu_register_thread() anyway.
 401  */
 402 void rcu_init(void)
 403 {
 404         struct sigaction act;
 405         int ret;
 406
 407         if (init_done)
 408                 return;
 409         init_done = 1;
 410
 411         act.sa_sigaction = sigrcu_handler;
 412         act.sa_flags = SA_SIGINFO | SA_RESTART;
 413         sigemptyset(&act.sa_mask);
 414         ret = sigaction(SIGRCU, &act, NULL);
 415         if (ret) {
 416                 perror("Error in sigaction");
 417                 exit(-1);
 418         }
 419 }
 420
 421 void rcu_exit(void)
 422 {
 423         struct sigaction act;
 424         int ret;
 425
 426         ret = sigaction(SIGRCU, NULL, &act);
 427         if (ret) {
 428                 perror("Error in sigaction");
 429                 exit(-1);
 430         }
 431         assert(act.sa_sigaction == sigrcu_handler);
 432         assert(cds_list_empty(&registry));
 433 }
 434
 435 #endif /* #ifdef RCU_SIGNAL */
 436
 437 #include "urcu-call-rcu-impl.h"
 438 #include "urcu-defer-impl.h"