urcu-call-rcu-impl.h

   1 /*
   2  * urcu-call-rcu.c
   3  *
   4  * Userspace RCU library - batch memory reclamation with kernel API
   5  *
   6  * Copyright (c) 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
   7  *
   8  * This library is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * This library is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with this library; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #define _GNU_SOURCE
  24 #include <stdio.h>
  25 #include <pthread.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <stdlib.h>
  29 #include <stdint.h>
  30 #include <string.h>
  31 #include <errno.h>
  32 #include <poll.h>
  33 #include <sys/time.h>
  34 #include <unistd.h>
  35 #include <sched.h>
  36
  37 #include "config.h"
  38 #include "urcu/wfqueue.h"
  39 #include "urcu-call-rcu.h"
  40 #include "urcu-pointer.h"
  41 #include "urcu/list.h"
  42 #include "urcu/futex.h"
  43
  44 /* Data structure that identifies a call_rcu thread. */
  45
  46 struct call_rcu_data {
  47         struct cds_wfq_queue cbs;
  48         unsigned long flags;
  49         int32_t futex;
  50         unsigned long qlen; /* maintained for debugging. */
  51         pthread_t tid;
  52         int cpu_affinity;
  53         struct cds_list_head list;
  54 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
  55
  56 /*
  57  * List of all call_rcu_data structures to keep valgrind happy.
  58  * Protected by call_rcu_mutex.
  59  */
  60
  61 CDS_LIST_HEAD(call_rcu_data_list);
  62
  63 /* Link a thread using call_rcu() to its call_rcu thread. */
  64
  65 static __thread struct call_rcu_data *thread_call_rcu_data;
  66
  67 /*
  68  * Guard call_rcu thread creation and atfork handlers.
  69  */
  70 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  71
  72 /* If a given thread does not have its own call_rcu thread, this is default. */
  73
  74 static struct call_rcu_data *default_call_rcu_data;
  75
  76 /*
  77  * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
  78  * available, then we can have call_rcu threads assigned to individual
  79  * CPUs rather than only to specific threads.
  80  */
  81
  82 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
  83
  84 /*
  85  * Pointer to array of pointers to per-CPU call_rcu_data structures
  86  * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
  87  * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
  88  * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
  89  * without mutex. The call_rcu_mutex protects updates.
  90  */
  91
  92 static struct call_rcu_data **per_cpu_call_rcu_data;
  93 static long maxcpus;
  94
  95 static void maxcpus_reset(void)
  96 {
  97         maxcpus = 0;
  98 }
  99
 100 /* Allocate the array if it has not already been allocated. */
 101
 102 static void alloc_cpu_call_rcu_data(void)
 103 {
 104         struct call_rcu_data **p;
 105         static int warned = 0;
 106
 107         if (maxcpus != 0)
 108                 return;
 109         maxcpus = sysconf(_SC_NPROCESSORS_CONF);
 110         if (maxcpus <= 0) {
 111                 return;
 112         }
 113         p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
 114         if (p != NULL) {
 115                 memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
 116                 rcu_set_pointer(&per_cpu_call_rcu_data, p);
 117         } else {
 118                 if (!warned) {
 119                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 120                 }
 121                 warned = 1;
 122         }
 123 }
 124
 125 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
 126
 127 /*
 128  * per_cpu_call_rcu_data should be constant, but some functions below, used both
 129  * for cases where cpu number is available and not available, assume it it not
 130  * constant.
 131  */
 132 static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
 133 static const long maxcpus = -1;
 134
 135 static void maxcpus_reset(void)
 136 {
 137 }
 138
 139 static void alloc_cpu_call_rcu_data(void)
 140 {
 141 }
 142
 143 static int sched_getcpu(void)
 144 {
 145         return -1;
 146 }
 147
 148 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
 149
 150 /* Acquire the specified pthread mutex. */
 151
 152 static void call_rcu_lock(pthread_mutex_t *pmp)
 153 {
 154         if (pthread_mutex_lock(pmp) != 0) {
 155                 perror("pthread_mutex_lock");
 156                 exit(-1);
 157         }
 158 }
 159
 160 /* Release the specified pthread mutex. */
 161
 162 static void call_rcu_unlock(pthread_mutex_t *pmp)
 163 {
 164         if (pthread_mutex_unlock(pmp) != 0) {
 165                 perror("pthread_mutex_unlock");
 166                 exit(-1);
 167         }
 168 }
 169
 170 #if HAVE_SCHED_SETAFFINITY
 171 static
 172 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
 173 {
 174         cpu_set_t mask;
 175
 176         if (crdp->cpu_affinity < 0)
 177                 return 0;
 178
 179         CPU_ZERO(&mask);
 180         CPU_SET(crdp->cpu_affinity, &mask);
 181 #if SCHED_SETAFFINITY_ARGS == 2
 182         return sched_setaffinity(0, &mask);
 183 #else
 184         return sched_setaffinity(0, sizeof(mask), &mask);
 185 #endif
 186 }
 187 #else
 188 static
 189 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
 190 {
 191         return 0;
 192 }
 193 #endif
 194
 195 static void call_rcu_wait(struct call_rcu_data *crdp)
 196 {
 197         /* Read call_rcu list before read futex */
 198         cmm_smp_mb();
 199         if (uatomic_read(&crdp->futex) == -1)
 200                 futex_async(&crdp->futex, FUTEX_WAIT, -1,
 201                       NULL, NULL, 0);
 202 }
 203
 204 static void call_rcu_wake_up(struct call_rcu_data *crdp)
 205 {
 206         /* Write to call_rcu list before reading/writing futex */
 207         cmm_smp_mb();
 208         if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
 209                 uatomic_set(&crdp->futex, 0);
 210                 futex_async(&crdp->futex, FUTEX_WAKE, 1,
 211                       NULL, NULL, 0);
 212         }
 213 }
 214
 215 /* This is the code run by each call_rcu thread. */
 216
 217 static void *call_rcu_thread(void *arg)
 218 {
 219         unsigned long cbcount;
 220         struct cds_wfq_node *cbs;
 221         struct cds_wfq_node **cbs_tail;
 222         struct call_rcu_data *crdp = (struct call_rcu_data *)arg;
 223         struct rcu_head *rhp;
 224         int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
 225
 226         if (set_thread_cpu_affinity(crdp) != 0) {
 227                 perror("pthread_setaffinity_np");
 228                 exit(-1);
 229         }
 230
 231         /*
 232          * If callbacks take a read-side lock, we need to be registered.
 233          */
 234         rcu_register_thread();
 235
 236         thread_call_rcu_data = crdp;
 237         if (!rt) {
 238                 uatomic_dec(&crdp->futex);
 239                 /* Decrement futex before reading call_rcu list */
 240                 cmm_smp_mb();
 241         }
 242         for (;;) {
 243                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
 244                         /*
 245                          * Pause requested. Become quiescent: remove
 246                          * ourself from all global lists, and don't
 247                          * process any callback. The callback lists may
 248                          * still be non-empty though.
 249                          */
 250                         rcu_unregister_thread();
 251                         cmm_smp_mb__before_uatomic_or();
 252                         uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
 253                         while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
 254                                 poll(NULL, 0, 1);
 255                         rcu_register_thread();
 256                 }
 257
 258                 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
 259                         while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
 260                                 poll(NULL, 0, 1);
 261                         _CMM_STORE_SHARED(crdp->cbs.head, NULL);
 262                         cbs_tail = (struct cds_wfq_node **)
 263                                 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
 264                         synchronize_rcu();
 265                         cbcount = 0;
 266                         do {
 267                                 while (cbs->next == NULL &&
 268                                        &cbs->next != cbs_tail)
 269                                         poll(NULL, 0, 1);
 270                                 if (cbs == &crdp->cbs.dummy) {
 271                                         cbs = cbs->next;
 272                                         continue;
 273                                 }
 274                                 rhp = (struct rcu_head *)cbs;
 275                                 cbs = cbs->next;
 276                                 rhp->func(rhp);
 277                                 cbcount++;
 278                         } while (cbs != NULL);
 279                         uatomic_sub(&crdp->qlen, cbcount);
 280                 }
 281                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
 282                         break;
 283                 rcu_thread_offline();
 284                 if (!rt) {
 285                         if (&crdp->cbs.head
 286                             == _CMM_LOAD_SHARED(crdp->cbs.tail)) {
 287                                 call_rcu_wait(crdp);
 288                                 poll(NULL, 0, 10);
 289                                 uatomic_dec(&crdp->futex);
 290                                 /*
 291                                  * Decrement futex before reading
 292                                  * call_rcu list.
 293                                  */
 294                                 cmm_smp_mb();
 295                         } else {
 296                                 poll(NULL, 0, 10);
 297                         }
 298                 } else {
 299                         poll(NULL, 0, 10);
 300                 }
 301                 rcu_thread_online();
 302         }
 303         if (!rt) {
 304                 /*
 305                  * Read call_rcu list before write futex.
 306                  */
 307                 cmm_smp_mb();
 308                 uatomic_set(&crdp->futex, 0);
 309         }
 310         uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
 311         rcu_unregister_thread();
 312         return NULL;
 313 }
 314
 315 /*
 316  * Create both a call_rcu thread and the corresponding call_rcu_data
 317  * structure, linking the structure in as specified.  Caller must hold
 318  * call_rcu_mutex.
 319  */
 320
 321 static void call_rcu_data_init(struct call_rcu_data **crdpp,
 322                                unsigned long flags,
 323                                int cpu_affinity)
 324 {
 325         struct call_rcu_data *crdp;
 326
 327         crdp = malloc(sizeof(*crdp));
 328         if (crdp == NULL) {
 329                 fprintf(stderr, "Out of memory.\n");
 330                 exit(-1);
 331         }
 332         memset(crdp, '\0', sizeof(*crdp));
 333         cds_wfq_init(&crdp->cbs);
 334         crdp->qlen = 0;
 335         crdp->futex = 0;
 336         crdp->flags = flags;
 337         cds_list_add(&crdp->list, &call_rcu_data_list);
 338         crdp->cpu_affinity = cpu_affinity;
 339         cmm_smp_mb();  /* Structure initialized before pointer is planted. */
 340         *crdpp = crdp;
 341         if (pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp) != 0) {
 342                 perror("pthread_create");
 343                 exit(-1);
 344         }
 345 }
 346
 347 /*
 348  * Return a pointer to the call_rcu_data structure for the specified
 349  * CPU, returning NULL if there is none.  We cannot automatically
 350  * created it because the platform we are running on might not define
 351  * sched_getcpu().
 352  *
 353  * The call to this function and use of the returned call_rcu_data
 354  * should be protected by RCU read-side lock.
 355  */
 356
 357 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
 358 {
 359         static int warned = 0;
 360         struct call_rcu_data **pcpu_crdp;
 361
 362         pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
 363         if (pcpu_crdp == NULL)
 364                 return NULL;
 365         if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
 366                 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
 367                 warned = 1;
 368         }
 369         if (cpu < 0 || maxcpus <= cpu)
 370                 return NULL;
 371         return rcu_dereference(pcpu_crdp[cpu]);
 372 }
 373
 374 /*
 375  * Return the tid corresponding to the call_rcu thread whose
 376  * call_rcu_data structure is specified.
 377  */
 378
 379 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
 380 {
 381         return crdp->tid;
 382 }
 383
 384 /*
 385  * Create a call_rcu_data structure (with thread) and return a pointer.
 386  */
 387
 388 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
 389                                                     int cpu_affinity)
 390 {
 391         struct call_rcu_data *crdp;
 392
 393         call_rcu_data_init(&crdp, flags, cpu_affinity);
 394         return crdp;
 395 }
 396
 397 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
 398                                            int cpu_affinity)
 399 {
 400         struct call_rcu_data *crdp;
 401
 402         call_rcu_lock(&call_rcu_mutex);
 403         crdp = __create_call_rcu_data(flags, cpu_affinity);
 404         call_rcu_unlock(&call_rcu_mutex);
 405         return crdp;
 406 }
 407
 408 /*
 409  * Set the specified CPU to use the specified call_rcu_data structure.
 410  *
 411  * Use NULL to remove a CPU's call_rcu_data structure, but it is
 412  * the caller's responsibility to dispose of the removed structure.
 413  * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
 414  * (prior to NULLing it out, of course).
 415  *
 416  * The caller must wait for a grace-period to pass between return from
 417  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 418  * previous call rcu data as argument.
 419  */
 420
 421 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
 422 {
 423         static int warned = 0;
 424
 425         call_rcu_lock(&call_rcu_mutex);
 426         alloc_cpu_call_rcu_data();
 427         if (cpu < 0 || maxcpus <= cpu) {
 428                 if (!warned) {
 429                         fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
 430                         warned = 1;
 431                 }
 432                 call_rcu_unlock(&call_rcu_mutex);
 433                 errno = EINVAL;
 434                 return -EINVAL;
 435         }
 436
 437         if (per_cpu_call_rcu_data == NULL) {
 438                 call_rcu_unlock(&call_rcu_mutex);
 439                 errno = ENOMEM;
 440                 return -ENOMEM;
 441         }
 442
 443         if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
 444                 call_rcu_unlock(&call_rcu_mutex);
 445                 errno = EEXIST;
 446                 return -EEXIST;
 447         }
 448
 449         rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
 450         call_rcu_unlock(&call_rcu_mutex);
 451         return 0;
 452 }
 453
 454 /*
 455  * Return a pointer to the default call_rcu_data structure, creating
 456  * one if need be.  Because we never free call_rcu_data structures,
 457  * we don't need to be in an RCU read-side critical section.
 458  */
 459
 460 struct call_rcu_data *get_default_call_rcu_data(void)
 461 {
 462         if (default_call_rcu_data != NULL)
 463                 return rcu_dereference(default_call_rcu_data);
 464         call_rcu_lock(&call_rcu_mutex);
 465         if (default_call_rcu_data != NULL) {
 466                 call_rcu_unlock(&call_rcu_mutex);
 467                 return default_call_rcu_data;
 468         }
 469         call_rcu_data_init(&default_call_rcu_data, 0, -1);
 470         call_rcu_unlock(&call_rcu_mutex);
 471         return default_call_rcu_data;
 472 }
 473
 474 /*
 475  * Return the call_rcu_data structure that applies to the currently
 476  * running thread.  Any call_rcu_data structure assigned specifically
 477  * to this thread has first priority, followed by any call_rcu_data
 478  * structure assigned to the CPU on which the thread is running,
 479  * followed by the default call_rcu_data structure.  If there is not
 480  * yet a default call_rcu_data structure, one will be created.
 481  *
 482  * Calls to this function and use of the returned call_rcu_data should
 483  * be protected by RCU read-side lock.
 484  */
 485 struct call_rcu_data *get_call_rcu_data(void)
 486 {
 487         struct call_rcu_data *crd;
 488
 489         if (thread_call_rcu_data != NULL)
 490                 return thread_call_rcu_data;
 491
 492         if (maxcpus > 0) {
 493                 crd = get_cpu_call_rcu_data(sched_getcpu());
 494                 if (crd)
 495                         return crd;
 496         }
 497
 498         return get_default_call_rcu_data();
 499 }
 500
 501 /*
 502  * Return a pointer to this task's call_rcu_data if there is one.
 503  */
 504
 505 struct call_rcu_data *get_thread_call_rcu_data(void)
 506 {
 507         return thread_call_rcu_data;
 508 }
 509
 510 /*
 511  * Set this task's call_rcu_data structure as specified, regardless
 512  * of whether or not this task already had one.  (This allows switching
 513  * to and from real-time call_rcu threads, for example.)
 514  *
 515  * Use NULL to remove a thread's call_rcu_data structure, but it is
 516  * the caller's responsibility to dispose of the removed structure.
 517  * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
 518  * (prior to NULLing it out, of course).
 519  */
 520
 521 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
 522 {
 523         thread_call_rcu_data = crdp;
 524 }
 525
 526 /*
 527  * Create a separate call_rcu thread for each CPU.  This does not
 528  * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
 529  * function if you want that behavior. Should be paired with
 530  * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
 531  * threads.
 532  */
 533
 534 int create_all_cpu_call_rcu_data(unsigned long flags)
 535 {
 536         int i;
 537         struct call_rcu_data *crdp;
 538         int ret;
 539
 540         call_rcu_lock(&call_rcu_mutex);
 541         alloc_cpu_call_rcu_data();
 542         call_rcu_unlock(&call_rcu_mutex);
 543         if (maxcpus <= 0) {
 544                 errno = EINVAL;
 545                 return -EINVAL;
 546         }
 547         if (per_cpu_call_rcu_data == NULL) {
 548                 errno = ENOMEM;
 549                 return -ENOMEM;
 550         }
 551         for (i = 0; i < maxcpus; i++) {
 552                 call_rcu_lock(&call_rcu_mutex);
 553                 if (get_cpu_call_rcu_data(i)) {
 554                         call_rcu_unlock(&call_rcu_mutex);
 555                         continue;
 556                 }
 557                 crdp = __create_call_rcu_data(flags, i);
 558                 if (crdp == NULL) {
 559                         call_rcu_unlock(&call_rcu_mutex);
 560                         errno = ENOMEM;
 561                         return -ENOMEM;
 562                 }
 563                 call_rcu_unlock(&call_rcu_mutex);
 564                 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
 565                         call_rcu_data_free(crdp);
 566
 567                         /* it has been created by other thread */
 568                         if (ret == -EEXIST)
 569                                 continue;
 570
 571                         return ret;
 572                 }
 573         }
 574         return 0;
 575 }
 576
 577 /*
 578  * Wake up the call_rcu thread corresponding to the specified
 579  * call_rcu_data structure.
 580  */
 581 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
 582 {
 583         if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
 584                 call_rcu_wake_up(crdp);
 585 }
 586
 587 /*
 588  * Schedule a function to be invoked after a following grace period.
 589  * This is the only function that must be called -- the others are
 590  * only present to allow applications to tune their use of RCU for
 591  * maximum performance.
 592  *
 593  * Note that unless a call_rcu thread has not already been created,
 594  * the first invocation of call_rcu() will create one.  So, if you
 595  * need the first invocation of call_rcu() to be fast, make sure
 596  * to create a call_rcu thread first.  One way to accomplish this is
 597  * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
 598  *
 599  * call_rcu must be called by registered RCU read-side threads.
 600  */
 601
 602 void call_rcu(struct rcu_head *head,
 603               void (*func)(struct rcu_head *head))
 604 {
 605         struct call_rcu_data *crdp;
 606
 607         cds_wfq_node_init(&head->next);
 608         head->func = func;
 609         /* Holding rcu read-side lock across use of per-cpu crdp */
 610         rcu_read_lock();
 611         crdp = get_call_rcu_data();
 612         cds_wfq_enqueue(&crdp->cbs, &head->next);
 613         uatomic_inc(&crdp->qlen);
 614         wake_call_rcu_thread(crdp);
 615         rcu_read_unlock();
 616 }
 617
 618 /*
 619  * Free up the specified call_rcu_data structure, terminating the
 620  * associated call_rcu thread.  The caller must have previously
 621  * removed the call_rcu_data structure from per-thread or per-CPU
 622  * usage.  For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
 623  * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
 624  * per-thread call_rcu_data structures.
 625  *
 626  * We silently refuse to free up the default call_rcu_data structure
 627  * because that is where we put any leftover callbacks.  Note that
 628  * the possibility of self-spawning callbacks makes it impossible
 629  * to execute all the callbacks in finite time without putting any
 630  * newly spawned callbacks somewhere else.  The "somewhere else" of
 631  * last resort is the default call_rcu_data structure.
 632  *
 633  * We also silently refuse to free NULL pointers.  This simplifies
 634  * the calling code.
 635  *
 636  * The caller must wait for a grace-period to pass between return from
 637  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 638  * previous call rcu data as argument.
 639  */
 640 void call_rcu_data_free(struct call_rcu_data *crdp)
 641 {
 642         struct cds_wfq_node *cbs;
 643         struct cds_wfq_node **cbs_tail;
 644         struct cds_wfq_node **cbs_endprev;
 645
 646         if (crdp == NULL || crdp == default_call_rcu_data) {
 647                 return;
 648         }
 649         if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
 650                 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
 651                 wake_call_rcu_thread(crdp);
 652                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
 653                         poll(NULL, 0, 1);
 654         }
 655         if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
 656                 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
 657                         poll(NULL, 0, 1);
 658                 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
 659                 cbs_tail = (struct cds_wfq_node **)
 660                         uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
 661                 /* Create default call rcu data if need be */
 662                 (void) get_default_call_rcu_data();
 663                 cbs_endprev = (struct cds_wfq_node **)
 664                         uatomic_xchg(&default_call_rcu_data->cbs.tail,
 665                                         cbs_tail);
 666                 _CMM_STORE_SHARED(*cbs_endprev, cbs);
 667                 uatomic_add(&default_call_rcu_data->qlen,
 668                             uatomic_read(&crdp->qlen));
 669                 wake_call_rcu_thread(default_call_rcu_data);
 670         }
 671
 672         call_rcu_lock(&call_rcu_mutex);
 673         cds_list_del(&crdp->list);
 674         call_rcu_unlock(&call_rcu_mutex);
 675
 676         free(crdp);
 677 }
 678
 679 /*
 680  * Clean up all the per-CPU call_rcu threads.
 681  */
 682 void free_all_cpu_call_rcu_data(void)
 683 {
 684         int cpu;
 685         struct call_rcu_data **crdp;
 686         static int warned = 0;
 687
 688         if (maxcpus <= 0)
 689                 return;
 690
 691         crdp = malloc(sizeof(*crdp) * maxcpus);
 692         if (!crdp) {
 693                 if (!warned) {
 694                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 695                 }
 696                 warned = 1;
 697                 return;
 698         }
 699
 700         for (cpu = 0; cpu < maxcpus; cpu++) {
 701                 crdp[cpu] = get_cpu_call_rcu_data(cpu);
 702                 if (crdp[cpu] == NULL)
 703                         continue;
 704                 set_cpu_call_rcu_data(cpu, NULL);
 705         }
 706         /*
 707          * Wait for call_rcu sites acting as RCU readers of the
 708          * call_rcu_data to become quiescent.
 709          */
 710         synchronize_rcu();
 711         for (cpu = 0; cpu < maxcpus; cpu++) {
 712                 if (crdp[cpu] == NULL)
 713                         continue;
 714                 call_rcu_data_free(crdp[cpu]);
 715         }
 716         free(crdp);
 717 }
 718
 719 /*
 720  * Acquire the call_rcu_mutex in order to ensure that the child sees
 721  * all of the call_rcu() data structures in a consistent state. Ensure
 722  * that all call_rcu threads are in a quiescent state across fork.
 723  * Suitable for pthread_atfork() and friends.
 724  */
 725 void call_rcu_before_fork(void)
 726 {
 727         struct call_rcu_data *crdp;
 728
 729         call_rcu_lock(&call_rcu_mutex);
 730
 731         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 732                 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
 733                 cmm_smp_mb__after_uatomic_or();
 734                 wake_call_rcu_thread(crdp);
 735         }
 736         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 737                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0)
 738                         poll(NULL, 0, 1);
 739         }
 740 }
 741
 742 /*
 743  * Clean up call_rcu data structures in the parent of a successful fork()
 744  * that is not followed by exec() in the child.  Suitable for
 745  * pthread_atfork() and friends.
 746  */
 747 void call_rcu_after_fork_parent(void)
 748 {
 749         struct call_rcu_data *crdp;
 750
 751         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 752                 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
 753         call_rcu_unlock(&call_rcu_mutex);
 754 }
 755
 756 /*
 757  * Clean up call_rcu data structures in the child of a successful fork()
 758  * that is not followed by exec().  Suitable for pthread_atfork() and
 759  * friends.
 760  */
 761 void call_rcu_after_fork_child(void)
 762 {
 763         struct call_rcu_data *crdp, *next;
 764
 765         /* Release the mutex. */
 766         call_rcu_unlock(&call_rcu_mutex);
 767
 768         /* Do nothing when call_rcu() has not been used */
 769         if (cds_list_empty(&call_rcu_data_list))
 770                 return;
 771
 772         /*
 773          * Allocate a new default call_rcu_data structure in order
 774          * to get a working call_rcu thread to go with it.
 775          */
 776         default_call_rcu_data = NULL;
 777         (void)get_default_call_rcu_data();
 778
 779         /* Cleanup call_rcu_data pointers before use */
 780         maxcpus_reset();
 781         free(per_cpu_call_rcu_data);
 782         rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
 783         thread_call_rcu_data = NULL;
 784
 785         /*
 786          * Dispose of all of the rest of the call_rcu_data structures.
 787          * Leftover call_rcu callbacks will be merged into the new
 788          * default call_rcu thread queue.
 789          */
 790         cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
 791                 if (crdp == default_call_rcu_data)
 792                         continue;
 793                 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
 794                 call_rcu_data_free(crdp);
 795         }
 796 }