src/urcu-qsbr.c

   1 // SPDX-FileCopyrightText: 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   2 // SPDX-FileCopyrightText: 2009 Paul E. McKenney, IBM Corporation.
   3 //
   4 // SPDX-License-Identifier: LGPL-2.1-or-later
   5
   6 /*
   7  * Userspace RCU QSBR library
   8  *
   9  * IBM's contributions to this file may be relicensed under LGPLv2 or later.
  10  */
  11
  12 #define URCU_NO_COMPAT_IDENTIFIERS
  13 #define _LGPL_SOURCE
  14 #include <stdio.h>
  15 #include <pthread.h>
  16 #include <signal.h>
  17 #include <stdlib.h>
  18 #include <stdint.h>
  19 #include <string.h>
  20 #include <errno.h>
  21 #include <poll.h>
  22
  23 #include <urcu/assert.h>
  24 #include <urcu/wfcqueue.h>
  25 #include <urcu/map/urcu-qsbr.h>
  26 #define BUILD_QSBR_LIB
  27 #include <urcu/static/urcu-qsbr.h>
  28 #include <urcu/pointer.h>
  29 #include <urcu/tls-compat.h>
  30
  31 #include "urcu-die.h"
  32 #include "urcu-wait.h"
  33 #include "urcu-utils.h"
  34
  35 #define URCU_API_MAP
  36 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  37 #undef _LGPL_SOURCE
  38 #include <urcu/urcu-qsbr.h>
  39 #define _LGPL_SOURCE
  40
  41 void __attribute__((destructor)) urcu_qsbr_exit(void);
  42 static void urcu_call_rcu_exit(void);
  43
  44 /*
  45  * rcu_gp_lock ensures mutual exclusion between threads calling
  46  * synchronize_rcu().
  47  */
  48 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  49 /*
  50  * rcu_registry_lock ensures mutual exclusion between threads
  51  * registering and unregistering themselves to/from the registry, and
  52  * with threads reading that registry from synchronize_rcu(). However,
  53  * this lock is not held all the way through the completion of awaiting
  54  * for the grace period. It is sporadically released between iterations
  55  * on the registry.
  56  * rcu_registry_lock may nest inside rcu_gp_lock.
  57  */
  58 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
  59 struct urcu_gp urcu_qsbr_gp = { .ctr = URCU_QSBR_GP_ONLINE };
  60
  61 /*
  62  * Active attempts to check for reader Q.S. before calling futex().
  63  */
  64 #define RCU_QS_ACTIVE_ATTEMPTS 100
  65
  66 /*
  67  * Written to only by each individual reader. Read by both the reader and the
  68  * writers.
  69  */
  70 DEFINE_URCU_TLS(struct urcu_qsbr_reader, urcu_qsbr_reader);
  71
  72 static CDS_LIST_HEAD(registry);
  73
  74 /*
  75  * Queue keeping threads awaiting to wait for a grace period. Contains
  76  * struct gp_waiters_thread objects.
  77  */
  78 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
  79
  80 static void mutex_lock(pthread_mutex_t *mutex)
  81 {
  82         int ret;
  83
  84 #ifndef DISTRUST_SIGNALS_EXTREME
  85         ret = pthread_mutex_lock(mutex);
  86         if (ret)
  87                 urcu_die(ret);
  88 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  89         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
  90                 if (ret != EBUSY && ret != EINTR)
  91                         urcu_die(ret);
  92                 poll(NULL,0,10);
  93         }
  94 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
  95 }
  96
  97 static void mutex_unlock(pthread_mutex_t *mutex)
  98 {
  99         int ret;
 100
 101         ret = pthread_mutex_unlock(mutex);
 102         if (ret)
 103                 urcu_die(ret);
 104 }
 105
 106 /*
 107  * synchronize_rcu() waiting. Single thread.
 108  */
 109 static void wait_gp(void)
 110 {
 111         /* Read reader_gp before read futex */
 112         cmm_smp_rmb();
 113         while (uatomic_read(&urcu_qsbr_gp.futex) == -1) {
 114                 if (!futex_noasync(&urcu_qsbr_gp.futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 115                         /*
 116                          * Prior queued wakeups queued by unrelated code
 117                          * using the same address can cause futex wait to
 118                          * return 0 even through the futex value is still
 119                          * -1 (spurious wakeups). Check the value again
 120                          * in user-space to validate whether it really
 121                          * differs from -1.
 122                          */
 123                         continue;
 124                 }
 125                 switch (errno) {
 126                 case EAGAIN:
 127                         /* Value already changed. */
 128                         return;
 129                 case EINTR:
 130                         /* Retry if interrupted by signal. */
 131                         break;  /* Get out of switch. Check again. */
 132                 default:
 133                         /* Unexpected error. */
 134                         urcu_die(errno);
 135                 }
 136         }
 137 }
 138
 139 /*
 140  * Always called with rcu_registry lock held. Releases this lock between
 141  * iterations and grabs it again. Holds the lock when it returns.
 142  */
 143 static void wait_for_readers(struct cds_list_head *input_readers,
 144                         struct cds_list_head *cur_snap_readers,
 145                         struct cds_list_head *qsreaders)
 146 {
 147         unsigned int wait_loops = 0;
 148         struct urcu_qsbr_reader *index, *tmp;
 149
 150         /*
 151          * Wait for each thread URCU_TLS(urcu_qsbr_reader).ctr to either
 152          * indicate quiescence (offline), or for them to observe the
 153          * current urcu_qsbr_gp.ctr value.
 154          */
 155         for (;;) {
 156                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 157                         wait_loops++;
 158                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 159                         uatomic_set(&urcu_qsbr_gp.futex, -1);
 160                         /*
 161                          * Write futex before write waiting (the other side
 162                          * reads them in the opposite order).
 163                          */
 164                         cmm_smp_wmb();
 165                         cds_list_for_each_entry(index, input_readers, node) {
 166                                 _CMM_STORE_SHARED(index->waiting, 1);
 167                         }
 168                         /* Write futex before read reader_gp */
 169                         cmm_smp_mb();
 170                 }
 171                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 172                         switch (urcu_qsbr_reader_state(&index->ctr)) {
 173                         case URCU_READER_ACTIVE_CURRENT:
 174                                 if (cur_snap_readers) {
 175                                         cds_list_move(&index->node,
 176                                                 cur_snap_readers);
 177                                         break;
 178                                 }
 179                                 /* Fall-through */
 180                         case URCU_READER_INACTIVE:
 181                                 cds_list_move(&index->node, qsreaders);
 182                                 break;
 183                         case URCU_READER_ACTIVE_OLD:
 184                                 /*
 185                                  * Old snapshot. Leaving node in
 186                                  * input_readers will make us busy-loop
 187                                  * until the snapshot becomes current or
 188                                  * the reader becomes inactive.
 189                                  */
 190                                 break;
 191                         }
 192                 }
 193
 194                 if (cds_list_empty(input_readers)) {
 195                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 196                                 /* Read reader_gp before write futex */
 197                                 cmm_smp_mb();
 198                                 uatomic_set(&urcu_qsbr_gp.futex, 0);
 199                         }
 200                         break;
 201                 } else {
 202                         /* Temporarily unlock the registry lock. */
 203                         mutex_unlock(&rcu_registry_lock);
 204                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 205                                 wait_gp();
 206                         } else {
 207 #ifndef HAS_INCOHERENT_CACHES
 208                                 caa_cpu_relax();
 209 #else /* #ifndef HAS_INCOHERENT_CACHES */
 210                                 cmm_smp_mb();
 211 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 212                         }
 213                         /* Re-lock the registry lock before the next loop. */
 214                         mutex_lock(&rcu_registry_lock);
 215                 }
 216         }
 217 }
 218
 219 /*
 220  * Using a two-subphases algorithm for architectures with smaller than 64-bit
 221  * long-size to ensure we do not encounter an overflow bug.
 222  */
 223
 224 #if (CAA_BITS_PER_LONG < 64)
 225 void urcu_qsbr_synchronize_rcu(void)
 226 {
 227         CDS_LIST_HEAD(cur_snap_readers);
 228         CDS_LIST_HEAD(qsreaders);
 229         unsigned long was_online;
 230         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 231         struct urcu_waiters waiters;
 232
 233         was_online = urcu_qsbr_read_ongoing();
 234
 235         /* All threads should read qparity before accessing data structure
 236          * where new ptr points to.  In the "then" case, rcu_thread_offline
 237          * includes a memory barrier.
 238          *
 239          * Mark the writer thread offline to make sure we don't wait for
 240          * our own quiescent state. This allows using synchronize_rcu()
 241          * in threads registered as readers.
 242          */
 243         if (was_online)
 244                 urcu_qsbr_thread_offline();
 245         else
 246                 cmm_smp_mb();
 247
 248         /*
 249          * Add ourself to gp_waiters queue of threads awaiting to wait
 250          * for a grace period. Proceed to perform the grace period only
 251          * if we are the first thread added into the queue.
 252          */
 253         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 254                 /* Not first in queue: will be awakened by another thread. */
 255                 urcu_adaptative_busy_wait(&wait);
 256                 goto gp_end;
 257         }
 258         /* We won't need to wake ourself up */
 259         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 260
 261         mutex_lock(&rcu_gp_lock);
 262
 263         /*
 264          * Move all waiters into our local queue.
 265          */
 266         urcu_move_waiters(&waiters, &gp_waiters);
 267
 268         mutex_lock(&rcu_registry_lock);
 269
 270         if (cds_list_empty(&registry))
 271                 goto out;
 272
 273         /*
 274          * Wait for readers to observe original parity or be quiescent.
 275          * wait_for_readers() can release and grab again rcu_registry_lock
 276          * internally.
 277          */
 278         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 279
 280         /*
 281          * Must finish waiting for quiescent state for original parity
 282          * before committing next urcu_qsbr_gp.ctr update to memory. Failure
 283          * to do so could result in the writer waiting forever while new
 284          * readers are always accessing data (no progress).  Enforce
 285          * compiler-order of load URCU_TLS(urcu_qsbr_reader).ctr before store
 286          * to urcu_qsbr_gp.ctr.
 287          */
 288         cmm_barrier();
 289
 290         /*
 291          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 292          * model easier to understand. It does not have a big performance impact
 293          * anyway, given this is the write-side.
 294          */
 295         cmm_smp_mb();
 296
 297         /* Switch parity: 0 -> 1, 1 -> 0 */
 298         CMM_STORE_SHARED(urcu_qsbr_gp.ctr, urcu_qsbr_gp.ctr ^ URCU_QSBR_GP_CTR);
 299
 300         /*
 301          * Must commit urcu_qsbr_gp.ctr update to memory before waiting for
 302          * quiescent state. Failure to do so could result in the writer
 303          * waiting forever while new readers are always accessing data
 304          * (no progress). Enforce compiler-order of store to urcu_qsbr_gp.ctr
 305          * before load URCU_TLS(urcu_qsbr_reader).ctr.
 306          */
 307         cmm_barrier();
 308
 309         /*
 310          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 311          * model easier to understand. It does not have a big performance impact
 312          * anyway, given this is the write-side.
 313          */
 314         cmm_smp_mb();
 315
 316         /*
 317          * Wait for readers to observe new parity or be quiescent.
 318          * wait_for_readers() can release and grab again rcu_registry_lock
 319          * internally.
 320          */
 321         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 322
 323         /*
 324          * Put quiescent reader list back into registry.
 325          */
 326         cds_list_splice(&qsreaders, &registry);
 327 out:
 328         mutex_unlock(&rcu_registry_lock);
 329         mutex_unlock(&rcu_gp_lock);
 330         urcu_wake_all_waiters(&waiters);
 331 gp_end:
 332         /*
 333          * Finish waiting for reader threads before letting the old ptr being
 334          * freed.
 335          */
 336         if (was_online)
 337                 urcu_qsbr_thread_online();
 338         else
 339                 cmm_smp_mb();
 340 }
 341 #else /* !(CAA_BITS_PER_LONG < 64) */
 342 void urcu_qsbr_synchronize_rcu(void)
 343 {
 344         CDS_LIST_HEAD(qsreaders);
 345         unsigned long was_online;
 346         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 347         struct urcu_waiters waiters;
 348
 349         was_online = urcu_qsbr_read_ongoing();
 350
 351         /*
 352          * Mark the writer thread offline to make sure we don't wait for
 353          * our own quiescent state. This allows using synchronize_rcu()
 354          * in threads registered as readers.
 355          */
 356         if (was_online)
 357                 urcu_qsbr_thread_offline();
 358         else
 359                 cmm_smp_mb();
 360
 361         /*
 362          * Add ourself to gp_waiters queue of threads awaiting to wait
 363          * for a grace period. Proceed to perform the grace period only
 364          * if we are the first thread added into the queue.
 365          */
 366         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 367                 /* Not first in queue: will be awakened by another thread. */
 368                 urcu_adaptative_busy_wait(&wait);
 369                 goto gp_end;
 370         }
 371         /* We won't need to wake ourself up */
 372         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 373
 374         mutex_lock(&rcu_gp_lock);
 375
 376         /*
 377          * Move all waiters into our local queue.
 378          */
 379         urcu_move_waiters(&waiters, &gp_waiters);
 380
 381         mutex_lock(&rcu_registry_lock);
 382
 383         if (cds_list_empty(&registry))
 384                 goto out;
 385
 386         /* Increment current G.P. */
 387         CMM_STORE_SHARED(urcu_qsbr_gp.ctr, urcu_qsbr_gp.ctr + URCU_QSBR_GP_CTR);
 388
 389         /*
 390          * Must commit urcu_qsbr_gp.ctr update to memory before waiting for
 391          * quiescent state. Failure to do so could result in the writer
 392          * waiting forever while new readers are always accessing data
 393          * (no progress). Enforce compiler-order of store to urcu_qsbr_gp.ctr
 394          * before load URCU_TLS(urcu_qsbr_reader).ctr.
 395          */
 396         cmm_barrier();
 397
 398         /*
 399          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 400          * model easier to understand. It does not have a big performance impact
 401          * anyway, given this is the write-side.
 402          */
 403         cmm_smp_mb();
 404
 405         /*
 406          * Wait for readers to observe new count of be quiescent.
 407          * wait_for_readers() can release and grab again rcu_registry_lock
 408          * internally.
 409          */
 410         wait_for_readers(&registry, NULL, &qsreaders);
 411
 412         /*
 413          * Put quiescent reader list back into registry.
 414          */
 415         cds_list_splice(&qsreaders, &registry);
 416 out:
 417         mutex_unlock(&rcu_registry_lock);
 418         mutex_unlock(&rcu_gp_lock);
 419         urcu_wake_all_waiters(&waiters);
 420 gp_end:
 421         if (was_online)
 422                 urcu_qsbr_thread_online();
 423         else
 424                 cmm_smp_mb();
 425 }
 426 #endif  /* !(CAA_BITS_PER_LONG < 64) */
 427
 428 /*
 429  * library wrappers to be used by non-LGPL compatible source code.
 430  */
 431
 432 void urcu_qsbr_read_lock(void)
 433 {
 434         _urcu_qsbr_read_lock();
 435 }
 436
 437 void urcu_qsbr_read_unlock(void)
 438 {
 439         _urcu_qsbr_read_unlock();
 440 }
 441
 442 int urcu_qsbr_read_ongoing(void)
 443 {
 444         return _urcu_qsbr_read_ongoing();
 445 }
 446 void rcu_read_ongoing_qsbr();
 447
 448 void urcu_qsbr_quiescent_state(void)
 449 {
 450         _urcu_qsbr_quiescent_state();
 451 }
 452 void rcu_quiescent_state_qsbr();
 453
 454 void urcu_qsbr_thread_offline(void)
 455 {
 456         _urcu_qsbr_thread_offline();
 457 }
 458 void rcu_thread_offline_qsbr();
 459
 460 void urcu_qsbr_thread_online(void)
 461 {
 462         _urcu_qsbr_thread_online();
 463 }
 464
 465 void urcu_qsbr_register_thread(void)
 466 {
 467         URCU_TLS(urcu_qsbr_reader).tid = pthread_self();
 468         urcu_posix_assert(URCU_TLS(urcu_qsbr_reader).ctr == 0);
 469
 470         mutex_lock(&rcu_registry_lock);
 471         urcu_posix_assert(!URCU_TLS(urcu_qsbr_reader).registered);
 472         URCU_TLS(urcu_qsbr_reader).registered = 1;
 473         cds_list_add(&URCU_TLS(urcu_qsbr_reader).node, &registry);
 474         mutex_unlock(&rcu_registry_lock);
 475         _urcu_qsbr_thread_online();
 476 }
 477
 478 void urcu_qsbr_unregister_thread(void)
 479 {
 480         /*
 481          * We have to make the thread offline otherwise we end up dealocking
 482          * with a waiting writer.
 483          */
 484         _urcu_qsbr_thread_offline();
 485         urcu_posix_assert(URCU_TLS(urcu_qsbr_reader).registered);
 486         URCU_TLS(urcu_qsbr_reader).registered = 0;
 487         mutex_lock(&rcu_registry_lock);
 488         cds_list_del(&URCU_TLS(urcu_qsbr_reader).node);
 489         mutex_unlock(&rcu_registry_lock);
 490 }
 491
 492 void urcu_qsbr_exit(void)
 493 {
 494         /*
 495          * Assertion disabled because call_rcu threads are now rcu
 496          * readers, and left running at exit.
 497          * urcu_posix_assert(cds_list_empty(&registry));
 498          */
 499         urcu_call_rcu_exit();
 500 }
 501
 502 DEFINE_RCU_FLAVOR(rcu_flavor);
 503
 504 #include "urcu-call-rcu-impl.h"
 505 #include "urcu-defer-impl.h"
 506 #include "urcu-poll-impl.h"