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 <stdint.h>
  34 #include <string.h>
  35 #include <errno.h>
  36 #include <poll.h>
  37
  38 #include "urcu/map/urcu.h"
  39
  40 #include "urcu/static/urcu.h"
  41 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  42 #include "urcu.h"
  43
  44 /*
  45  * If a reader is really non-cooperative and refuses to commit its
  46  * rcu_active_readers count to memory (there is no barrier in the reader
  47  * per-se), kick it after a few loops waiting for it.
  48  */
  49 #define KICK_READER_LOOPS 10000
  50
  51 /*
  52  * Active attempts to check for reader Q.S. before calling futex().
  53  */
  54 #define RCU_QS_ACTIVE_ATTEMPTS 100
  55
  56 #ifdef RCU_MEMBARRIER
  57 static int init_done;
  58 int has_sys_membarrier;
  59
  60 void __attribute__((constructor)) rcu_init(void);
  61 #endif
  62
  63 #ifdef RCU_MB
  64 void rcu_init(void)
  65 {
  66 }
  67 #endif
  68
  69 #ifdef RCU_SIGNAL
  70 static int init_done;
  71
  72 void __attribute__((constructor)) rcu_init(void);
  73 void __attribute__((destructor)) rcu_exit(void);
  74 #endif
  75
  76 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  77
  78 int32_t gp_futex;
  79
  80 /*
  81  * Global grace period counter.
  82  * Contains the current RCU_GP_CTR_PHASE.
  83  * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
  84  * Written to only by writer with mutex taken. Read by both writer and readers.
  85  */
  86 unsigned long rcu_gp_ctr = RCU_GP_COUNT;
  87
  88 /*
  89  * Written to only by each individual reader. Read by both the reader and the
  90  * writers.
  91  */
  92 struct rcu_reader __thread rcu_reader;
  93
  94 #ifdef DEBUG_YIELD
  95 unsigned int yield_active;
  96 unsigned int __thread rand_yield;
  97 #endif
  98
  99 static CDS_LIST_HEAD(registry);
 100
 101 static void mutex_lock(pthread_mutex_t *mutex)
 102 {
 103         int ret;
 104
 105 #ifndef DISTRUST_SIGNALS_EXTREME
 106         ret = pthread_mutex_lock(mutex);
 107         if (ret) {
 108                 perror("Error in pthread mutex lock");
 109                 exit(-1);
 110         }
 111 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 112         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 113                 if (ret != EBUSY && ret != EINTR) {
 114                         printf("ret = %d, errno = %d\n", ret, errno);
 115                         perror("Error in pthread mutex lock");
 116                         exit(-1);
 117                 }
 118                 if (CMM_LOAD_SHARED(rcu_reader.need_mb)) {
 119                         cmm_smp_mb();
 120                         _CMM_STORE_SHARED(rcu_reader.need_mb, 0);
 121                         cmm_smp_mb();
 122                 }
 123                 poll(NULL,0,10);
 124         }
 125 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 126 }
 127
 128 static void mutex_unlock(pthread_mutex_t *mutex)
 129 {
 130         int ret;
 131
 132         ret = pthread_mutex_unlock(mutex);
 133         if (ret) {
 134                 perror("Error in pthread mutex unlock");
 135                 exit(-1);
 136         }
 137 }
 138
 139 #ifdef RCU_MEMBARRIER
 140 static void smp_mb_master(int group)
 141 {
 142         if (likely(has_sys_membarrier))
 143                 membarrier(MEMBARRIER_EXPEDITED);
 144         else
 145                 cmm_smp_mb();
 146 }
 147 #endif
 148
 149 #ifdef RCU_MB
 150 static void smp_mb_master(int group)
 151 {
 152         cmm_smp_mb();
 153 }
 154 #endif
 155
 156 #ifdef RCU_SIGNAL
 157 static void force_mb_all_readers(void)
 158 {
 159         struct rcu_reader *index;
 160
 161         /*
 162          * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 163          * compiler barriers around rcu read lock as real memory barriers.
 164          */
 165         if (cds_list_empty(&registry))
 166                 return;
 167         /*
 168          * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 169          * a cache flush on architectures with non-coherent cache. Let's play
 170          * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 171          * cache flush is enforced.
 172          */
 173         cds_list_for_each_entry(index, &registry, node) {
 174                 CMM_STORE_SHARED(index->need_mb, 1);
 175                 pthread_kill(index->tid, SIGRCU);
 176         }
 177         /*
 178          * Wait for sighandler (and thus mb()) to execute on every thread.
 179          *
 180          * Note that the pthread_kill() will never be executed on systems
 181          * that correctly deliver signals in a timely manner.  However, it
 182          * is not uncommon for kernels to have bugs that can result in
 183          * lost or unduly delayed signals.
 184          *
 185          * If you are seeing the below pthread_kill() executing much at
 186          * all, we suggest testing the underlying kernel and filing the
 187          * relevant bug report.  For Linux kernels, we recommend getting
 188          * the Linux Test Project (LTP).
 189          */
 190         cds_list_for_each_entry(index, &registry, node) {
 191                 while (CMM_LOAD_SHARED(index->need_mb)) {
 192                         pthread_kill(index->tid, SIGRCU);
 193                         poll(NULL, 0, 1);
 194                 }
 195         }
 196         cmm_smp_mb();   /* read ->need_mb before ending the barrier */
 197 }
 198
 199 static void smp_mb_master(int group)
 200 {
 201         force_mb_all_readers();
 202 }
 203 #endif /* #ifdef RCU_SIGNAL */
 204
 205 /*
 206  * synchronize_rcu() waiting. Single thread.
 207  */
 208 static void wait_gp(void)
 209 {
 210         /* Read reader_gp before read futex */
 211         smp_mb_master(RCU_MB_GROUP);
 212         if (uatomic_read(&gp_futex) == -1)
 213                 futex_async(&gp_futex, FUTEX_WAIT, -1,
 214                       NULL, NULL, 0);
 215 }
 216
 217 void update_counter_and_wait(void)
 218 {
 219         CDS_LIST_HEAD(qsreaders);
 220         int wait_loops = 0;
 221         struct rcu_reader *index, *tmp;
 222
 223         /* Switch parity: 0 -> 1, 1 -> 0 */
 224         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
 225
 226         /*
 227          * Must commit rcu_gp_ctr update to memory before waiting for quiescent
 228          * state. Failure to do so could result in the writer waiting forever
 229          * while new readers are always accessing data (no progress). Enforce
 230          * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr.
 231          */
 232         cmm_barrier();
 233
 234         /*
 235          *
 236          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 237          * model easier to understand. It does not have a big performance impact
 238          * anyway, given this is the write-side.
 239          */
 240         cmm_smp_mb();
 241
 242         /*
 243          * Wait for each thread rcu_reader.ctr count to become 0.
 244          */
 245         for (;;) {
 246                 wait_loops++;
 247                 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 248                         uatomic_dec(&gp_futex);
 249                         /* Write futex before read reader_gp */
 250                         smp_mb_master(RCU_MB_GROUP);
 251                 }
 252
 253                 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
 254                         if (!rcu_gp_ongoing(&index->ctr))
 255                                 cds_list_move(&index->node, &qsreaders);
 256                 }
 257
 258 #ifndef HAS_INCOHERENT_CACHES
 259                 if (cds_list_empty(&registry)) {
 260                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 261                                 /* Read reader_gp before write futex */
 262                                 smp_mb_master(RCU_MB_GROUP);
 263                                 uatomic_set(&gp_futex, 0);
 264                         }
 265                         break;
 266                 } else {
 267                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 268                                 wait_gp();
 269                         else
 270                                 caa_cpu_relax();
 271                 }
 272 #else /* #ifndef HAS_INCOHERENT_CACHES */
 273                 /*
 274                  * BUSY-LOOP. Force the reader thread to commit its
 275                  * rcu_reader.ctr update to memory if we wait for too long.
 276                  */
 277                 if (cds_list_empty(&registry)) {
 278                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 279                                 /* Read reader_gp before write futex */
 280                                 smp_mb_master(RCU_MB_GROUP);
 281                                 uatomic_set(&gp_futex, 0);
 282                         }
 283                         break;
 284                 } else {
 285                         switch (wait_loops) {
 286                         case RCU_QS_ACTIVE_ATTEMPTS:
 287                                 wait_gp();
 288                                 break; /* only escape switch */
 289                         case KICK_READER_LOOPS:
 290                                 smp_mb_master(RCU_MB_GROUP);
 291                                 wait_loops = 0;
 292                                 break; /* only escape switch */
 293                         default:
 294                                 caa_cpu_relax();
 295                         }
 296                 }
 297 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 298         }
 299         /* put back the reader list in the registry */
 300         cds_list_splice(&qsreaders, &registry);
 301 }
 302
 303 void synchronize_rcu(void)
 304 {
 305         mutex_lock(&rcu_gp_lock);
 306
 307         if (cds_list_empty(&registry))
 308                 goto out;
 309
 310         /* All threads should read qparity before accessing data structure
 311          * where new ptr points to. Must be done within rcu_gp_lock because it
 312          * iterates on reader threads.*/
 313         /* Write new ptr before changing the qparity */
 314         smp_mb_master(RCU_MB_GROUP);
 315
 316         /*
 317          * Wait for previous parity to be empty of readers.
 318          */
 319         update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 320
 321         /*
 322          * Must finish waiting for quiescent state for parity 0 before
 323          * committing next rcu_gp_ctr update to memory. Failure to do so could
 324          * result in the writer waiting forever while new readers are always
 325          * accessing data (no progress).  Enforce compiler-order of load
 326          * rcu_reader ctr before store to rcu_gp_ctr.
 327          */
 328         cmm_barrier();
 329
 330         /*
 331          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 332          * model easier to understand. It does not have a big performance impact
 333          * anyway, given this is the write-side.
 334          */
 335         cmm_smp_mb();
 336
 337         /*
 338          * Wait for previous parity to be empty of readers.
 339          */
 340         update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
 341
 342         /* Finish waiting for reader threads before letting the old ptr being
 343          * freed. Must be done within rcu_gp_lock because it iterates on reader
 344          * threads. */
 345         smp_mb_master(RCU_MB_GROUP);
 346 out:
 347         mutex_unlock(&rcu_gp_lock);
 348 }
 349
 350 /*
 351  * library wrappers to be used by non-LGPL compatible source code.
 352  */
 353
 354 void rcu_read_lock(void)
 355 {
 356         _rcu_read_lock();
 357 }
 358
 359 void rcu_read_unlock(void)
 360 {
 361         _rcu_read_unlock();
 362 }
 363
 364 void rcu_register_thread(void)
 365 {
 366         rcu_reader.tid = pthread_self();
 367         assert(rcu_reader.need_mb == 0);
 368         assert(!(rcu_reader.ctr & RCU_GP_CTR_NEST_MASK));
 369
 370         mutex_lock(&rcu_gp_lock);
 371         rcu_init();     /* In case gcc does not support constructor attribute */
 372         cds_list_add(&rcu_reader.node, &registry);
 373         mutex_unlock(&rcu_gp_lock);
 374 }
 375
 376 void rcu_unregister_thread(void)
 377 {
 378         mutex_lock(&rcu_gp_lock);
 379         cds_list_del(&rcu_reader.node);
 380         mutex_unlock(&rcu_gp_lock);
 381 }
 382
 383 #ifdef RCU_MEMBARRIER
 384 void rcu_init(void)
 385 {
 386         if (init_done)
 387                 return;
 388         init_done = 1;
 389         if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
 390                 has_sys_membarrier = 1;
 391 }
 392 #endif
 393
 394 #ifdef RCU_SIGNAL
 395 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
 396 {
 397         /*
 398          * Executing this cmm_smp_mb() is the only purpose of this signal handler.
 399          * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 400          * executed on.
 401          */
 402         cmm_smp_mb();
 403         _CMM_STORE_SHARED(rcu_reader.need_mb, 0);
 404         cmm_smp_mb();
 405 }
 406
 407 /*
 408  * rcu_init constructor. Called when the library is linked, but also when
 409  * reader threads are calling rcu_register_thread().
 410  * Should only be called by a single thread at a given time. This is ensured by
 411  * holing the rcu_gp_lock from rcu_register_thread() or by running at library
 412  * load time, which should not be executed by multiple threads nor concurrently
 413  * with rcu_register_thread() anyway.
 414  */
 415 void rcu_init(void)
 416 {
 417         struct sigaction act;
 418         int ret;
 419
 420         if (init_done)
 421                 return;
 422         init_done = 1;
 423
 424         act.sa_sigaction = sigrcu_handler;
 425         act.sa_flags = SA_SIGINFO | SA_RESTART;
 426         sigemptyset(&act.sa_mask);
 427         ret = sigaction(SIGRCU, &act, NULL);
 428         if (ret) {
 429                 perror("Error in sigaction");
 430                 exit(-1);
 431         }
 432 }
 433
 434 void rcu_exit(void)
 435 {
 436         struct sigaction act;
 437         int ret;
 438
 439         ret = sigaction(SIGRCU, NULL, &act);
 440         if (ret) {
 441                 perror("Error in sigaction");
 442                 exit(-1);
 443         }
 444         assert(act.sa_sigaction == sigrcu_handler);
 445         assert(cds_list_empty(&registry));
 446 }
 447
 448 #endif /* #ifdef RCU_SIGNAL */
 449
 450 #include "urcu-call-rcu-impl.h"
 451 #include "urcu-defer-impl.h"