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