urcu.c

   1 /*
   2  * urcu.c
   3  *
   4  * Userspace RCU library
   5  *
   6  * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
   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 #include <stdio.h>
  27 #include <pthread.h>
  28 #include <signal.h>
  29 #include <assert.h>
  30 #include <stdlib.h>
  31 #include <string.h>
  32 #include <errno.h>
  33 #include <poll.h>
  34
  35 #include "urcu-static.h"
  36 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  37 #include "urcu.h"
  38
  39 #ifndef DEBUG_FULL_MB
  40 void __attribute__((constructor)) urcu_init(void);
  41 void __attribute__((destructor)) urcu_exit(void);
  42 #else
  43 static inline void urcu_init(void)
  44 {
  45 }
  46 #endif
  47
  48 int init_done;
  49
  50 pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  51
  52 /*
  53  * Global grace period counter.
  54  * Contains the current RCU_GP_CTR_BIT.
  55  * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
  56  * Written to only by writer with mutex taken. Read by both writer and readers.
  57  */
  58 long urcu_gp_ctr = RCU_GP_COUNT;
  59
  60 /*
  61  * Written to only by each individual reader. Read by both the reader and the
  62  * writers.
  63  */
  64 long __thread urcu_active_readers;
  65
  66 /* Thread IDs of registered readers */
  67 #define INIT_NUM_THREADS 4
  68
  69 struct reader_registry {
  70         pthread_t tid;
  71         long *urcu_active_readers;
  72         char *need_mb;
  73 };
  74
  75 #ifdef DEBUG_YIELD
  76 unsigned int yield_active;
  77 unsigned int __thread rand_yield;
  78 #endif
  79
  80 static struct reader_registry *registry;
  81 static char __thread need_mb;
  82 static int num_readers, alloc_readers;
  83
  84 void internal_urcu_lock(void)
  85 {
  86         int ret;
  87
  88 #ifndef DISTRUST_SIGNALS_EXTREME
  89         ret = pthread_mutex_lock(&urcu_mutex);
  90         if (ret) {
  91                 perror("Error in pthread mutex lock");
  92                 exit(-1);
  93         }
  94 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  95         while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
  96                 if (ret != EBUSY && ret != EINTR) {
  97                         printf("ret = %d, errno = %d\n", ret, errno);
  98                         perror("Error in pthread mutex lock");
  99                         exit(-1);
 100                 }
 101                 if (need_mb) {
 102                         smp_mb();
 103                         need_mb = 0;
 104                         smp_mb();
 105                 }
 106                 poll(NULL,0,10);
 107         }
 108 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 109 }
 110
 111 void internal_urcu_unlock(void)
 112 {
 113         int ret;
 114
 115         ret = pthread_mutex_unlock(&urcu_mutex);
 116         if (ret) {
 117                 perror("Error in pthread mutex unlock");
 118                 exit(-1);
 119         }
 120 }
 121
 122 /*
 123  * called with urcu_mutex held.
 124  */
 125 static void switch_next_urcu_qparity(void)
 126 {
 127         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 128 }
 129
 130 #ifdef DEBUG_FULL_MB
 131 #ifdef HAS_INCOHERENT_CACHES
 132 static void force_mb_single_thread(struct reader_registry *index)
 133 {
 134         smp_mb();
 135 }
 136 #endif /* #ifdef HAS_INCOHERENT_CACHES */
 137
 138 static void force_mb_all_threads(void)
 139 {
 140         smp_mb();
 141 }
 142 #else /* #ifdef DEBUG_FULL_MB */
 143 #ifdef HAS_INCOHERENT_CACHES
 144 static void force_mb_single_thread(struct reader_registry *index)
 145 {
 146         assert(registry);
 147         /*
 148          * pthread_kill has a smp_mb(). But beware, we assume it performs
 149          * a cache flush on architectures with non-coherent cache. Let's play
 150          * safe and don't assume anything : we use smp_mc() to make sure the
 151          * cache flush is enforced.
 152          */
 153         *index->need_mb = 1;
 154         smp_mc();       /* write ->need_mb before sending the signals */
 155         pthread_kill(index->tid, SIGURCU);
 156         smp_mb();
 157         /*
 158          * Wait for sighandler (and thus mb()) to execute on every thread.
 159          * BUSY-LOOP.
 160          */
 161         while (*index->need_mb) {
 162                 poll(NULL, 0, 1);
 163         }
 164         smp_mb();       /* read ->need_mb before ending the barrier */
 165 }
 166 #endif /* #ifdef HAS_INCOHERENT_CACHES */
 167
 168 static void force_mb_all_threads(void)
 169 {
 170         struct reader_registry *index;
 171         /*
 172          * Ask for each threads to execute a smp_mb() so we can consider the
 173          * compiler barriers around rcu read lock as real memory barriers.
 174          */
 175         if (!registry)
 176                 return;
 177         /*
 178          * pthread_kill has a smp_mb(). But beware, we assume it performs
 179          * a cache flush on architectures with non-coherent cache. Let's play
 180          * safe and don't assume anything : we use smp_mc() to make sure the
 181          * cache flush is enforced.
 182          */
 183         for (index = registry; index < registry + num_readers; index++) {
 184                 *index->need_mb = 1;
 185                 smp_mc();       /* write need_mb before sending the signal */
 186                 pthread_kill(index->tid, SIGURCU);
 187         }
 188         /*
 189          * Wait for sighandler (and thus mb()) to execute on every thread.
 190          *
 191          * Note that the pthread_kill() will never be executed on systems
 192          * that correctly deliver signals in a timely manner.  However, it
 193          * is not uncommon for kernels to have bugs that can result in
 194          * lost or unduly delayed signals.
 195          *
 196          * If you are seeing the below pthread_kill() executing much at
 197          * all, we suggest testing the underlying kernel and filing the
 198          * relevant bug report.  For Linux kernels, we recommend getting
 199          * the Linux Test Project (LTP).
 200          */
 201         for (index = registry; index < registry + num_readers; index++) {
 202                 while (*index->need_mb) {
 203                         pthread_kill(index->tid, SIGURCU);
 204                         poll(NULL, 0, 1);
 205                 }
 206         }
 207         smp_mb();       /* read ->need_mb before ending the barrier */
 208 }
 209 #endif /* #else #ifdef DEBUG_FULL_MB */
 210
 211 void wait_for_quiescent_state(void)
 212 {
 213         struct reader_registry *index;
 214
 215         if (!registry)
 216                 return;
 217         /*
 218          * Wait for each thread urcu_active_readers count to become 0.
 219          */
 220         for (index = registry; index < registry + num_readers; index++) {
 221 #ifndef HAS_INCOHERENT_CACHES
 222                 while (rcu_old_gp_ongoing(index->urcu_active_readers))
 223                         cpu_relax();
 224 #else /* #ifndef HAS_INCOHERENT_CACHES */
 225                 int wait_loops = 0;
 226                 /*
 227                  * BUSY-LOOP. Force the reader thread to commit its
 228                  * urcu_active_readers update to memory if we wait for too long.
 229                  */
 230                 while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
 231                         if (wait_loops++ == KICK_READER_LOOPS) {
 232                                 force_mb_single_thread(index);
 233                                 wait_loops = 0;
 234                         } else {
 235                                 cpu_relax();
 236                         }
 237                 }
 238 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 239         }
 240 }
 241
 242 void synchronize_rcu(void)
 243 {
 244         internal_urcu_lock();
 245
 246         /* All threads should read qparity before accessing data structure
 247          * where new ptr points to. Must be done within internal_urcu_lock
 248          * because it iterates on reader threads.*/
 249         /* Write new ptr before changing the qparity */
 250         force_mb_all_threads();
 251
 252         switch_next_urcu_qparity();     /* 0 -> 1 */
 253
 254         /*
 255          * Must commit qparity update to memory before waiting for parity
 256          * 0 quiescent state. Failure to do so could result in the writer
 257          * waiting forever while new readers are always accessing data (no
 258          * progress).
 259          * Ensured by STORE_SHARED and LOAD_SHARED.
 260          */
 261
 262         /*
 263          * Adding a smp_mb() which is _not_ formally required, but makes the
 264          * model easier to understand. It does not have a big performance impact
 265          * anyway, given this is the write-side.
 266          */
 267         smp_mb();
 268
 269         /*
 270          * Wait for previous parity to be empty of readers.
 271          */
 272         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 273
 274         /*
 275          * Must finish waiting for quiescent state for parity 0 before
 276          * committing qparity update to memory. Failure to do so could result in
 277          * the writer waiting forever while new readers are always accessing
 278          * data (no progress).
 279          * Ensured by STORE_SHARED and LOAD_SHARED.
 280          */
 281
 282         /*
 283          * Adding a smp_mb() which is _not_ formally required, but makes the
 284          * model easier to understand. It does not have a big performance impact
 285          * anyway, given this is the write-side.
 286          */
 287         smp_mb();
 288
 289         switch_next_urcu_qparity();     /* 1 -> 0 */
 290
 291         /*
 292          * Must commit qparity update to memory before waiting for parity
 293          * 1 quiescent state. Failure to do so could result in the writer
 294          * waiting forever while new readers are always accessing data (no
 295          * progress).
 296          * Ensured by STORE_SHARED and LOAD_SHARED.
 297          */
 298
 299         /*
 300          * Adding a smp_mb() which is _not_ formally required, but makes the
 301          * model easier to understand. It does not have a big performance impact
 302          * anyway, given this is the write-side.
 303          */
 304         smp_mb();
 305
 306         /*
 307          * Wait for previous parity to be empty of readers.
 308          */
 309         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 310
 311         /* Finish waiting for reader threads before letting the old ptr being
 312          * freed. Must be done within internal_urcu_lock because it iterates on
 313          * reader threads. */
 314         force_mb_all_threads();
 315
 316         internal_urcu_unlock();
 317 }
 318
 319 /*
 320  * library wrappers to be used by non-LGPL compatible source code.
 321  */
 322
 323 void rcu_read_lock(void)
 324 {
 325         _rcu_read_lock();
 326 }
 327
 328 void rcu_read_unlock(void)
 329 {
 330         _rcu_read_unlock();
 331 }
 332
 333 void *rcu_dereference(void *p)
 334 {
 335         return _rcu_dereference(p);
 336 }
 337
 338 void *rcu_assign_pointer_sym(void **p, void *v)
 339 {
 340         wmb();
 341         return STORE_SHARED(p, v);
 342 }
 343
 344 void *rcu_xchg_pointer_sym(void **p, void *v)
 345 {
 346         wmb();
 347         return xchg(p, v);
 348 }
 349
 350 void *rcu_publish_content_sym(void **p, void *v)
 351 {
 352         void *oldptr;
 353
 354         oldptr = _rcu_xchg_pointer(p, v);
 355         synchronize_rcu();
 356         return oldptr;
 357 }
 358
 359 static void rcu_add_reader(pthread_t id)
 360 {
 361         struct reader_registry *oldarray;
 362
 363         if (!registry) {
 364                 alloc_readers = INIT_NUM_THREADS;
 365                 num_readers = 0;
 366                 registry =
 367                         malloc(sizeof(struct reader_registry) * alloc_readers);
 368         }
 369         if (alloc_readers < num_readers + 1) {
 370                 oldarray = registry;
 371                 registry = malloc(sizeof(struct reader_registry)
 372                                 * (alloc_readers << 1));
 373                 memcpy(registry, oldarray,
 374                         sizeof(struct reader_registry) * alloc_readers);
 375                 alloc_readers <<= 1;
 376                 free(oldarray);
 377         }
 378         registry[num_readers].tid = id;
 379         /* reference to the TLS of _this_ reader thread. */
 380         registry[num_readers].urcu_active_readers = &urcu_active_readers;
 381         registry[num_readers].need_mb = &need_mb;
 382         num_readers++;
 383 }
 384
 385 /*
 386  * Never shrink (implementation limitation).
 387  * This is O(nb threads). Eventually use a hash table.
 388  */
 389 static void rcu_remove_reader(pthread_t id)
 390 {
 391         struct reader_registry *index;
 392
 393         assert(registry != NULL);
 394         for (index = registry; index < registry + num_readers; index++) {
 395                 if (pthread_equal(index->tid, id)) {
 396                         memcpy(index, &registry[num_readers - 1],
 397                                 sizeof(struct reader_registry));
 398                         registry[num_readers - 1].tid = 0;
 399                         registry[num_readers - 1].urcu_active_readers = NULL;
 400                         num_readers--;
 401                         return;
 402                 }
 403         }
 404         /* Hrm not found, forgot to register ? */
 405         assert(0);
 406 }
 407
 408 void rcu_register_thread(void)
 409 {
 410         internal_urcu_lock();
 411         urcu_init();    /* In case gcc does not support constructor attribute */
 412         rcu_add_reader(pthread_self());
 413         internal_urcu_unlock();
 414 }
 415
 416 void rcu_unregister_thread(void)
 417 {
 418         internal_urcu_lock();
 419         rcu_remove_reader(pthread_self());
 420         internal_urcu_unlock();
 421 }
 422
 423 #ifndef DEBUG_FULL_MB
 424 static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 425 {
 426         /*
 427          * Executing this smp_mb() is the only purpose of this signal handler.
 428          * It punctually promotes barrier() into smp_mb() on every thread it is
 429          * executed on.
 430          */
 431         smp_mb();
 432         need_mb = 0;
 433         smp_mb();
 434 }
 435
 436 /*
 437  * urcu_init constructor. Called when the library is linked, but also when
 438  * reader threads are calling rcu_register_thread().
 439  * Should only be called by a single thread at a given time. This is ensured by
 440  * holing the internal_urcu_lock() from rcu_register_thread() or by running at
 441  * library load time, which should not be executed by multiple threads nor
 442  * concurrently with rcu_register_thread() anyway.
 443  */
 444 void urcu_init(void)
 445 {
 446         struct sigaction act;
 447         int ret;
 448
 449         if (init_done)
 450                 return;
 451         init_done = 1;
 452
 453         act.sa_sigaction = sigurcu_handler;
 454         act.sa_flags = SA_SIGINFO;
 455         sigemptyset(&act.sa_mask);
 456         ret = sigaction(SIGURCU, &act, NULL);
 457         if (ret) {
 458                 perror("Error in sigaction");
 459                 exit(-1);
 460         }
 461 }
 462
 463 void urcu_exit(void)
 464 {
 465         struct sigaction act;
 466         int ret;
 467
 468         ret = sigaction(SIGURCU, NULL, &act);
 469         if (ret) {
 470                 perror("Error in sigaction");
 471                 exit(-1);
 472         }
 473         assert(act.sa_sigaction == sigurcu_handler);
 474         free(registry);
 475 }
 476 #endif /* #ifndef DEBUG_FULL_MB */