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