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