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