urcu-qsbr.c

   1 /*
   2  * urcu-qsbr.c
   3  *
   4  * Userspace RCU QSBR 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 #define BUILD_QSBR_LIB
  36 #include "urcu-qsbr-static.h"
  37 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  38 #include "urcu-qsbr.h"
  39
  40 static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  41
  42 /*
  43  * Global grace period counter.
  44  */
  45 unsigned long urcu_gp_ctr = RCU_GP_ONLINE;
  46
  47 /*
  48  * Written to only by each individual reader. Read by both the reader and the
  49  * writers.
  50  */
  51 unsigned long __thread rcu_reader_qs_gp;
  52
  53 /* Thread IDs of registered readers */
  54 #define INIT_NUM_THREADS 4
  55
  56 struct reader_registry {
  57         pthread_t tid;
  58         unsigned long *rcu_reader_qs_gp;
  59 };
  60
  61 #ifdef DEBUG_YIELD
  62 unsigned int yield_active;
  63 unsigned int __thread rand_yield;
  64 #endif
  65
  66 static struct reader_registry *registry;
  67 static int num_readers, alloc_readers;
  68
  69 static void internal_urcu_lock(void)
  70 {
  71         int ret;
  72
  73 #ifndef DISTRUST_SIGNALS_EXTREME
  74         ret = pthread_mutex_lock(&urcu_mutex);
  75         if (ret) {
  76                 perror("Error in pthread mutex lock");
  77                 exit(-1);
  78         }
  79 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  80         while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
  81                 if (ret != EBUSY && ret != EINTR) {
  82                         printf("ret = %d, errno = %d\n", ret, errno);
  83                         perror("Error in pthread mutex lock");
  84                         exit(-1);
  85                 }
  86                 poll(NULL,0,10);
  87         }
  88 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
  89 }
  90
  91 static void internal_urcu_unlock(void)
  92 {
  93         int ret;
  94
  95         ret = pthread_mutex_unlock(&urcu_mutex);
  96         if (ret) {
  97                 perror("Error in pthread mutex unlock");
  98                 exit(-1);
  99         }
 100 }
 101
 102 /*
 103  * synchronize_rcu() waiting. Single thread.
 104  */
 105 static void wait_for_quiescent_state(void)
 106 {
 107         struct reader_registry *index;
 108
 109         if (!registry)
 110                 return;
 111         /*
 112          * Wait for each thread rcu_reader_qs_gp count to become 0.
 113          */
 114         for (index = registry; index < registry + num_readers; index++) {
 115                 int wait_loops = 0;
 116
 117                 while (rcu_gp_ongoing(index->rcu_reader_qs_gp)) {
 118                         if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) {
 119                                 sched_yield();  /* ideally sched_yield_to() */
 120                         } else {
 121 #ifndef HAS_INCOHERENT_CACHES
 122                                 cpu_relax();
 123 #else /* #ifndef HAS_INCOHERENT_CACHES */
 124                                 smp_mb();
 125 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 126                         }
 127                 }
 128         }
 129 }
 130
 131 /*
 132  * Using a two-subphases algorithm for architectures with smaller than 64-bit
 133  * long-size to ensure we do not encounter an overflow bug.
 134  */
 135
 136 #if (BITS_PER_LONG < 64)
 137 /*
 138  * called with urcu_mutex held.
 139  */
 140 static void switch_next_urcu_qparity(void)
 141 {
 142         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR);
 143 }
 144
 145 void synchronize_rcu(void)
 146 {
 147         unsigned long was_online;
 148
 149         was_online = rcu_reader_qs_gp;
 150
 151         /* All threads should read qparity before accessing data structure
 152          * where new ptr points to.
 153          */
 154         /* Write new ptr before changing the qparity */
 155         smp_mb();
 156
 157         /*
 158          * Mark the writer thread offline to make sure we don't wait for
 159          * our own quiescent state. This allows using synchronize_rcu() in
 160          * threads registered as readers.
 161          */
 162         if (was_online)
 163                 STORE_SHARED(rcu_reader_qs_gp, 0);
 164
 165         internal_urcu_lock();
 166
 167         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING);
 168
 169         switch_next_urcu_qparity();     /* 0 -> 1 */
 170
 171         /*
 172          * Must commit qparity update to memory before waiting for parity
 173          * 0 quiescent state. Failure to do so could result in the writer
 174          * waiting forever while new readers are always accessing data (no
 175          * progress).
 176          * Ensured by STORE_SHARED and LOAD_SHARED.
 177          */
 178
 179         /*
 180          * Wait for previous parity to be empty of readers.
 181          */
 182         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 183
 184         /*
 185          * Must finish waiting for quiescent state for parity 0 before
 186          * committing qparity update to memory. Failure to do so could result in
 187          * the writer waiting forever while new readers are always accessing
 188          * data (no progress).
 189          * Ensured by STORE_SHARED and LOAD_SHARED.
 190          */
 191
 192         switch_next_urcu_qparity();     /* 1 -> 0 */
 193
 194         /*
 195          * Must commit qparity update to memory before waiting for parity
 196          * 1 quiescent state. Failure to do so could result in the writer
 197          * waiting forever while new readers are always accessing data (no
 198          * progress).
 199          * Ensured by STORE_SHARED and LOAD_SHARED.
 200          */
 201
 202         /*
 203          * Wait for previous parity to be empty of readers.
 204          */
 205         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 206
 207         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING);
 208
 209         internal_urcu_unlock();
 210
 211         /*
 212          * Finish waiting for reader threads before letting the old ptr being
 213          * freed.
 214          */
 215         if (was_online)
 216                 _STORE_SHARED(rcu_reader_qs_gp, LOAD_SHARED(urcu_gp_ctr));
 217         smp_mb();
 218 }
 219 #else /* !(BITS_PER_LONG < 64) */
 220 void synchronize_rcu(void)
 221 {
 222         unsigned long was_online;
 223
 224         was_online = rcu_reader_qs_gp;
 225
 226         /*
 227          * Mark the writer thread offline to make sure we don't wait for
 228          * our own quiescent state. This allows using synchronize_rcu() in
 229          * threads registered as readers.
 230          */
 231         smp_mb();
 232         if (was_online)
 233                 STORE_SHARED(rcu_reader_qs_gp, 0);
 234
 235         internal_urcu_lock();
 236         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING);
 237         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_CTR);
 238         wait_for_quiescent_state();
 239         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING);
 240         internal_urcu_unlock();
 241
 242         if (was_online)
 243                 _STORE_SHARED(rcu_reader_qs_gp, LOAD_SHARED(urcu_gp_ctr));
 244         smp_mb();
 245 }
 246 #endif  /* !(BITS_PER_LONG < 64) */
 247
 248 /*
 249  * library wrappers to be used by non-LGPL compatible source code.
 250  */
 251
 252 void rcu_read_lock(void)
 253 {
 254         _rcu_read_lock();
 255 }
 256
 257 void rcu_read_unlock(void)
 258 {
 259         _rcu_read_unlock();
 260 }
 261
 262 void *rcu_dereference(void *p)
 263 {
 264         return _rcu_dereference(p);
 265 }
 266
 267 void *rcu_assign_pointer_sym(void **p, void *v)
 268 {
 269         wmb();
 270         return STORE_SHARED(p, v);
 271 }
 272
 273 void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 274 {
 275         wmb();
 276         return cmpxchg(p, old, _new);
 277 }
 278
 279 void *rcu_xchg_pointer_sym(void **p, void *v)
 280 {
 281         wmb();
 282         return xchg(p, v);
 283 }
 284
 285 void *rcu_publish_content_sym(void **p, void *v)
 286 {
 287         void *oldptr;
 288
 289         oldptr = _rcu_xchg_pointer(p, v);
 290         synchronize_rcu();
 291         return oldptr;
 292 }
 293
 294 void rcu_quiescent_state(void)
 295 {
 296         _rcu_quiescent_state();
 297 }
 298
 299 void rcu_thread_offline(void)
 300 {
 301         _rcu_thread_offline();
 302 }
 303
 304 void rcu_thread_online(void)
 305 {
 306         _rcu_thread_online();
 307 }
 308
 309 static void rcu_add_reader(pthread_t id)
 310 {
 311         struct reader_registry *oldarray;
 312
 313         if (!registry) {
 314                 alloc_readers = INIT_NUM_THREADS;
 315                 num_readers = 0;
 316                 registry =
 317                         malloc(sizeof(struct reader_registry) * alloc_readers);
 318         }
 319         if (alloc_readers < num_readers + 1) {
 320                 oldarray = registry;
 321                 registry = malloc(sizeof(struct reader_registry)
 322                                 * (alloc_readers << 1));
 323                 memcpy(registry, oldarray,
 324                         sizeof(struct reader_registry) * alloc_readers);
 325                 alloc_readers <<= 1;
 326                 free(oldarray);
 327         }
 328         registry[num_readers].tid = id;
 329         /* reference to the TLS of _this_ reader thread. */
 330         registry[num_readers].rcu_reader_qs_gp = &rcu_reader_qs_gp;
 331         num_readers++;
 332 }
 333
 334 /*
 335  * Never shrink (implementation limitation).
 336  * This is O(nb threads). Eventually use a hash table.
 337  */
 338 static void rcu_remove_reader(pthread_t id)
 339 {
 340         struct reader_registry *index;
 341
 342         assert(registry != NULL);
 343         for (index = registry; index < registry + num_readers; index++) {
 344                 if (pthread_equal(index->tid, id)) {
 345                         memcpy(index, &registry[num_readers - 1],
 346                                 sizeof(struct reader_registry));
 347                         registry[num_readers - 1].tid = 0;
 348                         registry[num_readers - 1].rcu_reader_qs_gp = NULL;
 349                         num_readers--;
 350                         return;
 351                 }
 352         }
 353         /* Hrm not found, forgot to register ? */
 354         assert(0);
 355 }
 356
 357 void rcu_register_thread(void)
 358 {
 359         internal_urcu_lock();
 360         rcu_add_reader(pthread_self());
 361         internal_urcu_unlock();
 362         _rcu_thread_online();
 363 }
 364
 365 void rcu_unregister_thread(void)
 366 {
 367         /*
 368          * We have to make the thread offline otherwise we end up dealocking
 369          * with a waiting writer.
 370          */
 371         _rcu_thread_offline();
 372         internal_urcu_lock();
 373         rcu_remove_reader(pthread_self());
 374         internal_urcu_unlock();
 375 }