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