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 /*
  66  * RCU_MEMBARRIER is only possibly available on Linux.
  67  */
  68 #if defined(RCU_MEMBARRIER) && defined(__linux__)
  69 #include <syscall.h>
  70 #endif
  71
  72 /* If the headers do not support SYS_membarrier, fall back on RCU_MB */
  73 #ifdef SYS_membarrier
  74 # define membarrier(...)                syscall(SYS_membarrier, __VA_ARGS__)
  75 #else
  76 # define membarrier(...)                -ENOSYS
  77 #endif
  78
  79 #define MEMBARRIER_EXPEDITED            (1 << 0)
  80 #define MEMBARRIER_DELAYED              (1 << 1)
  81 #define MEMBARRIER_QUERY                (1 << 16)
  82
  83 #ifdef RCU_MEMBARRIER
  84 static int init_done;
  85 int rcu_has_sys_membarrier;
  86
  87 void __attribute__((constructor)) rcu_init(void);
  88 #endif
  89
  90 #ifdef RCU_MB
  91 void rcu_init(void)
  92 {
  93 }
  94 #endif
  95
  96 #ifdef RCU_SIGNAL
  97 static int init_done;
  98
  99 void __attribute__((constructor)) rcu_init(void);
 100 void __attribute__((destructor)) rcu_exit(void);
 101 #endif
 102
 103 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 104 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
 105
 106 /*
 107  * Written to only by each individual reader. Read by both the reader and the
 108  * writers.
 109  */
 110 __DEFINE_URCU_TLS_GLOBAL(struct rcu_reader, rcu_reader);
 111
 112 #ifdef DEBUG_YIELD
 113 unsigned int rcu_yield_active;
 114 __DEFINE_URCU_TLS_GLOBAL(unsigned int, rcu_rand_yield);
 115 #endif
 116
 117 static CDS_LIST_HEAD(registry);
 118
 119 /*
 120  * Queue keeping threads awaiting to wait for a grace period. Contains
 121  * struct gp_waiters_thread objects.
 122  */
 123 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
 124
 125 static void mutex_lock(pthread_mutex_t *mutex)
 126 {
 127         int ret;
 128
 129 #ifndef DISTRUST_SIGNALS_EXTREME
 130         ret = pthread_mutex_lock(mutex);
 131         if (ret)
 132                 urcu_die(ret);
 133 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 134         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 135                 if (ret != EBUSY && ret != EINTR)
 136                         urcu_die(ret);
 137                 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
 138                         cmm_smp_mb();
 139                         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 140                         cmm_smp_mb();
 141                 }
 142                 poll(NULL,0,10);
 143         }
 144 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 145 }
 146
 147 static void mutex_unlock(pthread_mutex_t *mutex)
 148 {
 149         int ret;
 150
 151         ret = pthread_mutex_unlock(mutex);
 152         if (ret)
 153                 urcu_die(ret);
 154 }
 155
 156 #ifdef RCU_MEMBARRIER
 157 static void smp_mb_master(int group)
 158 {
 159         if (caa_likely(rcu_has_sys_membarrier))
 160                 (void) membarrier(MEMBARRIER_EXPEDITED);
 161         else
 162                 cmm_smp_mb();
 163 }
 164 #endif
 165
 166 #ifdef RCU_MB
 167 static void smp_mb_master(int group)
 168 {
 169         cmm_smp_mb();
 170 }
 171 #endif
 172
 173 #ifdef RCU_SIGNAL
 174 static void force_mb_all_readers(void)
 175 {
 176         struct rcu_reader *index;
 177
 178         /*
 179          * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 180          * compiler barriers around rcu read lock as real memory barriers.
 181          */
 182         if (cds_list_empty(&registry))
 183                 return;
 184         /*
 185          * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 186          * a cache flush on architectures with non-coherent cache. Let's play
 187          * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 188          * cache flush is enforced.
 189          */
 190         cds_list_for_each_entry(index, &registry, node) {
 191                 CMM_STORE_SHARED(index->need_mb, 1);
 192                 pthread_kill(index->tid, SIGRCU);
 193         }
 194         /*
 195          * Wait for sighandler (and thus mb()) to execute on every thread.
 196          *
 197          * Note that the pthread_kill() will never be executed on systems
 198          * that correctly deliver signals in a timely manner.  However, it
 199          * is not uncommon for kernels to have bugs that can result in
 200          * lost or unduly delayed signals.
 201          *
 202          * If you are seeing the below pthread_kill() executing much at
 203          * all, we suggest testing the underlying kernel and filing the
 204          * relevant bug report.  For Linux kernels, we recommend getting
 205          * the Linux Test Project (LTP).
 206          */
 207         cds_list_for_each_entry(index, &registry, node) {
 208                 while (CMM_LOAD_SHARED(index->need_mb)) {
 209                         pthread_kill(index->tid, SIGRCU);
 210                         poll(NULL, 0, 1);
 211                 }
 212         }
 213         cmm_smp_mb();   /* read ->need_mb before ending the barrier */
 214 }
 215
 216 static void smp_mb_master(int group)
 217 {
 218         force_mb_all_readers();
 219 }
 220 #endif /* #ifdef RCU_SIGNAL */
 221
 222 /*
 223  * synchronize_rcu() waiting. Single thread.
 224  */
 225 static void wait_gp(void)
 226 {
 227         /* Read reader_gp before read futex */
 228         smp_mb_master(RCU_MB_GROUP);
 229         if (uatomic_read(&rcu_gp.futex) == -1)
 230                 futex_async(&rcu_gp.futex, FUTEX_WAIT, -1,
 231                       NULL, NULL, 0);
 232 }
 233
 234 static void wait_for_readers(struct cds_list_head *input_readers,
 235                         struct cds_list_head *cur_snap_readers,
 236                         struct cds_list_head *qsreaders)
 237 {
 238         int wait_loops = 0;
 239         struct rcu_reader *index, *tmp;
 240
 241         /*
 242          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 243          * indicate quiescence (not nested), or observe the current
 244          * rcu_gp.ctr value.
 245          */
 246         for (;;) {
 247                 wait_loops++;
 248                 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 249                         uatomic_dec(&rcu_gp.futex);
 250                         /* Write futex before read reader_gp */
 251                         smp_mb_master(RCU_MB_GROUP);
 252                 }
 253
 254                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 255                         switch (rcu_reader_state(&index->ctr)) {
 256                         case RCU_READER_ACTIVE_CURRENT:
 257                                 if (cur_snap_readers) {
 258                                         cds_list_move(&index->node,
 259                                                 cur_snap_readers);
 260                                         break;
 261                                 }
 262                                 /* Fall-through */
 263                         case RCU_READER_INACTIVE:
 264                                 cds_list_move(&index->node, qsreaders);
 265                                 break;
 266                         case RCU_READER_ACTIVE_OLD:
 267                                 /*
 268                                  * Old snapshot. Leaving node in
 269                                  * input_readers will make us busy-loop
 270                                  * until the snapshot becomes current or
 271                                  * the reader becomes inactive.
 272                                  */
 273                                 break;
 274                         }
 275                 }
 276
 277 #ifndef HAS_INCOHERENT_CACHES
 278                 if (cds_list_empty(input_readers)) {
 279                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 280                                 /* Read reader_gp before write futex */
 281                                 smp_mb_master(RCU_MB_GROUP);
 282                                 uatomic_set(&rcu_gp.futex, 0);
 283                         }
 284                         break;
 285                 } else {
 286                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 287                                 wait_gp();
 288                         else
 289                                 caa_cpu_relax();
 290                 }
 291 #else /* #ifndef HAS_INCOHERENT_CACHES */
 292                 /*
 293                  * BUSY-LOOP. Force the reader thread to commit its
 294                  * URCU_TLS(rcu_reader).ctr update to memory if we wait
 295                  * for too long.
 296                  */
 297                 if (cds_list_empty(input_readers)) {
 298                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 299                                 /* Read reader_gp before write futex */
 300                                 smp_mb_master(RCU_MB_GROUP);
 301                                 uatomic_set(&rcu_gp.futex, 0);
 302                         }
 303                         break;
 304                 } else {
 305                         switch (wait_loops) {
 306                         case RCU_QS_ACTIVE_ATTEMPTS:
 307                                 wait_gp();
 308                                 break; /* only escape switch */
 309                         case KICK_READER_LOOPS:
 310                                 smp_mb_master(RCU_MB_GROUP);
 311                                 wait_loops = 0;
 312                                 break; /* only escape switch */
 313                         default:
 314                                 caa_cpu_relax();
 315                         }
 316                 }
 317 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 318         }
 319 }
 320
 321 void synchronize_rcu(void)
 322 {
 323         CDS_LIST_HEAD(cur_snap_readers);
 324         CDS_LIST_HEAD(qsreaders);
 325         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 326         struct urcu_waiters waiters;
 327
 328         /*
 329          * Add ourself to gp_waiters queue of threads awaiting to wait
 330          * for a grace period. Proceed to perform the grace period only
 331          * if we are the first thread added into the queue.
 332          * The implicit memory barrier before urcu_wait_add()
 333          * orders prior memory accesses of threads put into the wait
 334          * queue before their insertion into the wait queue.
 335          */
 336         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 337                 /* Not first in queue: will be awakened by another thread. */
 338                 urcu_adaptative_busy_wait(&wait);
 339                 /* Order following memory accesses after grace period. */
 340                 cmm_smp_mb();
 341                 return;
 342         }
 343         /* We won't need to wake ourself up */
 344         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 345
 346         mutex_lock(&rcu_gp_lock);
 347
 348         /*
 349          * Move all waiters into our local queue.
 350          */
 351         urcu_move_waiters(&waiters, &gp_waiters);
 352
 353         if (cds_list_empty(&registry))
 354                 goto out;
 355
 356         /* All threads should read qparity before accessing data structure
 357          * where new ptr points to. Must be done within rcu_gp_lock because it
 358          * iterates on reader threads.*/
 359         /* Write new ptr before changing the qparity */
 360         smp_mb_master(RCU_MB_GROUP);
 361
 362         /*
 363          * Wait for readers to observe original parity or be quiescent.
 364          */
 365         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 366
 367         /*
 368          * Must finish waiting for quiescent state for original parity before
 369          * committing next rcu_gp.ctr update to memory. Failure to do so could
 370          * result in the writer waiting forever while new readers are always
 371          * accessing data (no progress).  Enforce compiler-order of load
 372          * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
 373          */
 374         cmm_barrier();
 375
 376         /*
 377          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 378          * model easier to understand. It does not have a big performance impact
 379          * anyway, given this is the write-side.
 380          */
 381         cmm_smp_mb();
 382
 383         /* Switch parity: 0 -> 1, 1 -> 0 */
 384         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
 385
 386         /*
 387          * Must commit rcu_gp.ctr update to memory before waiting for quiescent
 388          * state. Failure to do so could result in the writer waiting forever
 389          * while new readers are always accessing data (no progress). Enforce
 390          * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
 391          */
 392         cmm_barrier();
 393
 394         /*
 395          *
 396          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 397          * model easier to understand. It does not have a big performance impact
 398          * anyway, given this is the write-side.
 399          */
 400         cmm_smp_mb();
 401
 402         /*
 403          * Wait for readers to observe new parity or be quiescent.
 404          */
 405         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 406
 407         /*
 408          * Put quiescent reader list back into registry.
 409          */
 410         cds_list_splice(&qsreaders, &registry);
 411
 412         /* Finish waiting for reader threads before letting the old ptr being
 413          * freed. Must be done within rcu_gp_lock because it iterates on reader
 414          * threads. */
 415         smp_mb_master(RCU_MB_GROUP);
 416 out:
 417         mutex_unlock(&rcu_gp_lock);
 418
 419         /*
 420          * Wakeup waiters only after we have completed the grace period
 421          * and have ensured the memory barriers at the end of the grace
 422          * period have been issued.
 423          */
 424         urcu_wake_all_waiters(&waiters);
 425 }
 426
 427 /*
 428  * library wrappers to be used by non-LGPL compatible source code.
 429  */
 430
 431 void rcu_read_lock(void)
 432 {
 433         _rcu_read_lock();
 434 }
 435
 436 void rcu_read_unlock(void)
 437 {
 438         _rcu_read_unlock();
 439 }
 440
 441 int rcu_read_ongoing(void)
 442 {
 443         return _rcu_read_ongoing();
 444 }
 445
 446 void rcu_register_thread(void)
 447 {
 448         URCU_TLS(rcu_reader).tid = pthread_self();
 449         assert(URCU_TLS(rcu_reader).need_mb == 0);
 450         assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
 451
 452         mutex_lock(&rcu_gp_lock);
 453         rcu_init();     /* In case gcc does not support constructor attribute */
 454         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 455         mutex_unlock(&rcu_gp_lock);
 456 }
 457
 458 void rcu_unregister_thread(void)
 459 {
 460         mutex_lock(&rcu_gp_lock);
 461         cds_list_del(&URCU_TLS(rcu_reader).node);
 462         mutex_unlock(&rcu_gp_lock);
 463 }
 464
 465 #ifdef RCU_MEMBARRIER
 466 void rcu_init(void)
 467 {
 468         if (init_done)
 469                 return;
 470         init_done = 1;
 471         if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
 472                 rcu_has_sys_membarrier = 1;
 473 }
 474 #endif
 475
 476 #ifdef RCU_SIGNAL
 477 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
 478 {
 479         /*
 480          * Executing this cmm_smp_mb() is the only purpose of this signal handler.
 481          * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 482          * executed on.
 483          */
 484         cmm_smp_mb();
 485         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 486         cmm_smp_mb();
 487 }
 488
 489 /*
 490  * rcu_init constructor. Called when the library is linked, but also when
 491  * reader threads are calling rcu_register_thread().
 492  * Should only be called by a single thread at a given time. This is ensured by
 493  * holing the rcu_gp_lock from rcu_register_thread() or by running at library
 494  * load time, which should not be executed by multiple threads nor concurrently
 495  * with rcu_register_thread() anyway.
 496  */
 497 void rcu_init(void)
 498 {
 499         struct sigaction act;
 500         int ret;
 501
 502         if (init_done)
 503                 return;
 504         init_done = 1;
 505
 506         act.sa_sigaction = sigrcu_handler;
 507         act.sa_flags = SA_SIGINFO | SA_RESTART;
 508         sigemptyset(&act.sa_mask);
 509         ret = sigaction(SIGRCU, &act, NULL);
 510         if (ret)
 511                 urcu_die(errno);
 512 }
 513
 514 void rcu_exit(void)
 515 {
 516         /*
 517          * Don't unregister the SIGRCU signal handler anymore, because
 518          * call_rcu threads could still be using it shortly before the
 519          * application exits.
 520          * Assertion disabled because call_rcu threads are now rcu
 521          * readers, and left running at exit.
 522          * assert(cds_list_empty(&registry));
 523          */
 524 }
 525
 526 #endif /* #ifdef RCU_SIGNAL */
 527
 528 DEFINE_RCU_FLAVOR(rcu_flavor);
 529
 530 #include "urcu-call-rcu-impl.h"
 531 #include "urcu-defer-impl.h"