src/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 _LGPL_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 "compat-getcpu.h"
  38 #include "urcu/wfcqueue.h"
  39 #include "urcu-call-rcu.h"
  40 #include "urcu-pointer.h"
  41 #include "urcu/list.h"
  42 #include "urcu/futex.h"
  43 #include "urcu/tls-compat.h"
  44 #include "urcu/ref.h"
  45 #include "urcu-die.h"
  46
  47 #define SET_AFFINITY_CHECK_PERIOD               (1U << 8)       /* 256 */
  48 #define SET_AFFINITY_CHECK_PERIOD_MASK          (SET_AFFINITY_CHECK_PERIOD - 1)
  49
  50 /* Data structure that identifies a call_rcu thread. */
  51
  52 struct call_rcu_data {
  53         /*
  54          * We do not align head on a different cache-line than tail
  55          * mainly because call_rcu callback-invocation threads use
  56          * batching ("splice") to get an entire list of callbacks, which
  57          * effectively empties the queue, and requires to touch the tail
  58          * anyway.
  59          */
  60         struct cds_wfcq_tail cbs_tail;
  61         struct cds_wfcq_head cbs_head;
  62         unsigned long flags;
  63         int32_t futex;
  64         unsigned long qlen; /* maintained for debugging. */
  65         pthread_t tid;
  66         int cpu_affinity;
  67         unsigned long gp_count;
  68         struct cds_list_head list;
  69 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
  70
  71 struct call_rcu_completion {
  72         int barrier_count;
  73         int32_t futex;
  74         struct urcu_ref ref;
  75 };
  76
  77 struct call_rcu_completion_work {
  78         struct rcu_head head;
  79         struct call_rcu_completion *completion;
  80 };
  81
  82 /*
  83  * List of all call_rcu_data structures to keep valgrind happy.
  84  * Protected by call_rcu_mutex.
  85  */
  86
  87 static CDS_LIST_HEAD(call_rcu_data_list);
  88
  89 /* Link a thread using call_rcu() to its call_rcu thread. */
  90
  91 static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data);
  92
  93 /*
  94  * Guard call_rcu thread creation and atfork handlers.
  95  */
  96 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  97
  98 /* If a given thread does not have its own call_rcu thread, this is default. */
  99
 100 static struct call_rcu_data *default_call_rcu_data;
 101
 102 static struct urcu_atfork *registered_rculfhash_atfork;
 103 static unsigned long registered_rculfhash_atfork_refcount;
 104
 105 /*
 106  * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
 107  * available, then we can have call_rcu threads assigned to individual
 108  * CPUs rather than only to specific threads.
 109  */
 110
 111 #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
 112
 113 /*
 114  * Pointer to array of pointers to per-CPU call_rcu_data structures
 115  * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
 116  * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
 117  * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
 118  * without mutex. The call_rcu_mutex protects updates.
 119  */
 120
 121 static struct call_rcu_data **per_cpu_call_rcu_data;
 122 static long maxcpus;
 123
 124 static void maxcpus_reset(void)
 125 {
 126         maxcpus = 0;
 127 }
 128
 129 /* Allocate the array if it has not already been allocated. */
 130
 131 static void alloc_cpu_call_rcu_data(void)
 132 {
 133         struct call_rcu_data **p;
 134         static int warned = 0;
 135
 136         if (maxcpus != 0)
 137                 return;
 138         maxcpus = sysconf(_SC_NPROCESSORS_CONF);
 139         if (maxcpus <= 0) {
 140                 return;
 141         }
 142         p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
 143         if (p != NULL) {
 144                 memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
 145                 rcu_set_pointer(&per_cpu_call_rcu_data, p);
 146         } else {
 147                 if (!warned) {
 148                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 149                 }
 150                 warned = 1;
 151         }
 152 }
 153
 154 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 155
 156 /*
 157  * per_cpu_call_rcu_data should be constant, but some functions below, used both
 158  * for cases where cpu number is available and not available, assume it it not
 159  * constant.
 160  */
 161 static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
 162 static const long maxcpus = -1;
 163
 164 static void maxcpus_reset(void)
 165 {
 166 }
 167
 168 static void alloc_cpu_call_rcu_data(void)
 169 {
 170 }
 171
 172 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 173
 174 /* Acquire the specified pthread mutex. */
 175
 176 static void call_rcu_lock(pthread_mutex_t *pmp)
 177 {
 178         int ret;
 179
 180         ret = pthread_mutex_lock(pmp);
 181         if (ret)
 182                 urcu_die(ret);
 183 }
 184
 185 /* Release the specified pthread mutex. */
 186
 187 static void call_rcu_unlock(pthread_mutex_t *pmp)
 188 {
 189         int ret;
 190
 191         ret = pthread_mutex_unlock(pmp);
 192         if (ret)
 193                 urcu_die(ret);
 194 }
 195
 196 /*
 197  * Periodically retry setting CPU affinity if we migrate.
 198  * Losing affinity can be caused by CPU hotunplug/hotplug, or by
 199  * cpuset(7).
 200  */
 201 #if HAVE_SCHED_SETAFFINITY
 202 static
 203 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
 204 {
 205         cpu_set_t mask;
 206         int ret;
 207
 208         if (crdp->cpu_affinity < 0)
 209                 return 0;
 210         if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK)
 211                 return 0;
 212         if (urcu_sched_getcpu() == crdp->cpu_affinity)
 213                 return 0;
 214
 215         CPU_ZERO(&mask);
 216         CPU_SET(crdp->cpu_affinity, &mask);
 217 #if SCHED_SETAFFINITY_ARGS == 2
 218         ret = sched_setaffinity(0, &mask);
 219 #else
 220         ret = sched_setaffinity(0, sizeof(mask), &mask);
 221 #endif
 222         /*
 223          * EINVAL is fine: can be caused by hotunplugged CPUs, or by
 224          * cpuset(7). This is why we should always retry if we detect
 225          * migration.
 226          */
 227         if (ret && errno == EINVAL) {
 228                 ret = 0;
 229                 errno = 0;
 230         }
 231         return ret;
 232 }
 233 #else
 234 static
 235 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
 236 {
 237         return 0;
 238 }
 239 #endif
 240
 241 static void call_rcu_wait(struct call_rcu_data *crdp)
 242 {
 243         /* Read call_rcu list before read futex */
 244         cmm_smp_mb();
 245         if (uatomic_read(&crdp->futex) != -1)
 246                 return;
 247         while (futex_async(&crdp->futex, FUTEX_WAIT, -1,
 248                         NULL, NULL, 0)) {
 249                 switch (errno) {
 250                 case EWOULDBLOCK:
 251                         /* Value already changed. */
 252                         return;
 253                 case EINTR:
 254                         /* Retry if interrupted by signal. */
 255                         break;  /* Get out of switch. */
 256                 default:
 257                         /* Unexpected error. */
 258                         urcu_die(errno);
 259                 }
 260         }
 261 }
 262
 263 static void call_rcu_wake_up(struct call_rcu_data *crdp)
 264 {
 265         /* Write to call_rcu list before reading/writing futex */
 266         cmm_smp_mb();
 267         if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
 268                 uatomic_set(&crdp->futex, 0);
 269                 if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
 270                                 NULL, NULL, 0) < 0)
 271                         urcu_die(errno);
 272         }
 273 }
 274
 275 static void call_rcu_completion_wait(struct call_rcu_completion *completion)
 276 {
 277         /* Read completion barrier count before read futex */
 278         cmm_smp_mb();
 279         if (uatomic_read(&completion->futex) != -1)
 280                 return;
 281         while (futex_async(&completion->futex, FUTEX_WAIT, -1,
 282                         NULL, NULL, 0)) {
 283                 switch (errno) {
 284                 case EWOULDBLOCK:
 285                         /* Value already changed. */
 286                         return;
 287                 case EINTR:
 288                         /* Retry if interrupted by signal. */
 289                         break;  /* Get out of switch. */
 290                 default:
 291                         /* Unexpected error. */
 292                         urcu_die(errno);
 293                 }
 294         }
 295 }
 296
 297 static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
 298 {
 299         /* Write to completion barrier count before reading/writing futex */
 300         cmm_smp_mb();
 301         if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
 302                 uatomic_set(&completion->futex, 0);
 303                 if (futex_async(&completion->futex, FUTEX_WAKE, 1,
 304                                 NULL, NULL, 0) < 0)
 305                         urcu_die(errno);
 306         }
 307 }
 308
 309 /* This is the code run by each call_rcu thread. */
 310
 311 static void *call_rcu_thread(void *arg)
 312 {
 313         unsigned long cbcount;
 314         struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
 315         int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
 316
 317         if (set_thread_cpu_affinity(crdp))
 318                 urcu_die(errno);
 319
 320         /*
 321          * If callbacks take a read-side lock, we need to be registered.
 322          */
 323         rcu_register_thread();
 324
 325         URCU_TLS(thread_call_rcu_data) = crdp;
 326         if (!rt) {
 327                 uatomic_dec(&crdp->futex);
 328                 /* Decrement futex before reading call_rcu list */
 329                 cmm_smp_mb();
 330         }
 331         for (;;) {
 332                 struct cds_wfcq_head cbs_tmp_head;
 333                 struct cds_wfcq_tail cbs_tmp_tail;
 334                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 335                 enum cds_wfcq_ret splice_ret;
 336
 337                 if (set_thread_cpu_affinity(crdp))
 338                         urcu_die(errno);
 339
 340                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
 341                         /*
 342                          * Pause requested. Become quiescent: remove
 343                          * ourself from all global lists, and don't
 344                          * process any callback. The callback lists may
 345                          * still be non-empty though.
 346                          */
 347                         rcu_unregister_thread();
 348                         cmm_smp_mb__before_uatomic_or();
 349                         uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
 350                         while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
 351                                 (void) poll(NULL, 0, 1);
 352                         uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
 353                         cmm_smp_mb__after_uatomic_and();
 354                         rcu_register_thread();
 355                 }
 356
 357                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 358                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 359                         &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail);
 360                 assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 361                 assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 362                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 363                         synchronize_rcu();
 364                         cbcount = 0;
 365                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 366                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 367                                 struct rcu_head *rhp;
 368
 369                                 rhp = caa_container_of(cbs,
 370                                         struct rcu_head, next);
 371                                 rhp->func(rhp);
 372                                 cbcount++;
 373                         }
 374                         uatomic_sub(&crdp->qlen, cbcount);
 375                 }
 376                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
 377                         break;
 378                 rcu_thread_offline();
 379                 if (!rt) {
 380                         if (cds_wfcq_empty(&crdp->cbs_head,
 381                                         &crdp->cbs_tail)) {
 382                                 call_rcu_wait(crdp);
 383                                 (void) poll(NULL, 0, 10);
 384                                 uatomic_dec(&crdp->futex);
 385                                 /*
 386                                  * Decrement futex before reading
 387                                  * call_rcu list.
 388                                  */
 389                                 cmm_smp_mb();
 390                         } else {
 391                                 (void) poll(NULL, 0, 10);
 392                         }
 393                 } else {
 394                         (void) poll(NULL, 0, 10);
 395                 }
 396                 rcu_thread_online();
 397         }
 398         if (!rt) {
 399                 /*
 400                  * Read call_rcu list before write futex.
 401                  */
 402                 cmm_smp_mb();
 403                 uatomic_set(&crdp->futex, 0);
 404         }
 405         uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
 406         rcu_unregister_thread();
 407         return NULL;
 408 }
 409
 410 /*
 411  * Create both a call_rcu thread and the corresponding call_rcu_data
 412  * structure, linking the structure in as specified.  Caller must hold
 413  * call_rcu_mutex.
 414  */
 415
 416 static void call_rcu_data_init(struct call_rcu_data **crdpp,
 417                                unsigned long flags,
 418                                int cpu_affinity)
 419 {
 420         struct call_rcu_data *crdp;
 421         int ret;
 422
 423         crdp = malloc(sizeof(*crdp));
 424         if (crdp == NULL)
 425                 urcu_die(errno);
 426         memset(crdp, '\0', sizeof(*crdp));
 427         cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail);
 428         crdp->qlen = 0;
 429         crdp->futex = 0;
 430         crdp->flags = flags;
 431         cds_list_add(&crdp->list, &call_rcu_data_list);
 432         crdp->cpu_affinity = cpu_affinity;
 433         crdp->gp_count = 0;
 434         cmm_smp_mb();  /* Structure initialized before pointer is planted. */
 435         *crdpp = crdp;
 436         ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
 437         if (ret)
 438                 urcu_die(ret);
 439 }
 440
 441 /*
 442  * Return a pointer to the call_rcu_data structure for the specified
 443  * CPU, returning NULL if there is none.  We cannot automatically
 444  * created it because the platform we are running on might not define
 445  * urcu_sched_getcpu().
 446  *
 447  * The call to this function and use of the returned call_rcu_data
 448  * should be protected by RCU read-side lock.
 449  */
 450
 451 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
 452 {
 453         static int warned = 0;
 454         struct call_rcu_data **pcpu_crdp;
 455
 456         pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
 457         if (pcpu_crdp == NULL)
 458                 return NULL;
 459         if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
 460                 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
 461                 warned = 1;
 462         }
 463         if (cpu < 0 || maxcpus <= cpu)
 464                 return NULL;
 465         return rcu_dereference(pcpu_crdp[cpu]);
 466 }
 467
 468 /*
 469  * Return the tid corresponding to the call_rcu thread whose
 470  * call_rcu_data structure is specified.
 471  */
 472
 473 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
 474 {
 475         return crdp->tid;
 476 }
 477
 478 /*
 479  * Create a call_rcu_data structure (with thread) and return a pointer.
 480  */
 481
 482 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
 483                                                     int cpu_affinity)
 484 {
 485         struct call_rcu_data *crdp;
 486
 487         call_rcu_data_init(&crdp, flags, cpu_affinity);
 488         return crdp;
 489 }
 490
 491 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
 492                                            int cpu_affinity)
 493 {
 494         struct call_rcu_data *crdp;
 495
 496         call_rcu_lock(&call_rcu_mutex);
 497         crdp = __create_call_rcu_data(flags, cpu_affinity);
 498         call_rcu_unlock(&call_rcu_mutex);
 499         return crdp;
 500 }
 501
 502 /*
 503  * Set the specified CPU to use the specified call_rcu_data structure.
 504  *
 505  * Use NULL to remove a CPU's call_rcu_data structure, but it is
 506  * the caller's responsibility to dispose of the removed structure.
 507  * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
 508  * (prior to NULLing it out, of course).
 509  *
 510  * The caller must wait for a grace-period to pass between return from
 511  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 512  * previous call rcu data as argument.
 513  */
 514
 515 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
 516 {
 517         static int warned = 0;
 518
 519         call_rcu_lock(&call_rcu_mutex);
 520         alloc_cpu_call_rcu_data();
 521         if (cpu < 0 || maxcpus <= cpu) {
 522                 if (!warned) {
 523                         fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
 524                         warned = 1;
 525                 }
 526                 call_rcu_unlock(&call_rcu_mutex);
 527                 errno = EINVAL;
 528                 return -EINVAL;
 529         }
 530
 531         if (per_cpu_call_rcu_data == NULL) {
 532                 call_rcu_unlock(&call_rcu_mutex);
 533                 errno = ENOMEM;
 534                 return -ENOMEM;
 535         }
 536
 537         if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
 538                 call_rcu_unlock(&call_rcu_mutex);
 539                 errno = EEXIST;
 540                 return -EEXIST;
 541         }
 542
 543         rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
 544         call_rcu_unlock(&call_rcu_mutex);
 545         return 0;
 546 }
 547
 548 /*
 549  * Return a pointer to the default call_rcu_data structure, creating
 550  * one if need be.  Because we never free call_rcu_data structures,
 551  * we don't need to be in an RCU read-side critical section.
 552  */
 553
 554 struct call_rcu_data *get_default_call_rcu_data(void)
 555 {
 556         if (default_call_rcu_data != NULL)
 557                 return rcu_dereference(default_call_rcu_data);
 558         call_rcu_lock(&call_rcu_mutex);
 559         if (default_call_rcu_data != NULL) {
 560                 call_rcu_unlock(&call_rcu_mutex);
 561                 return default_call_rcu_data;
 562         }
 563         call_rcu_data_init(&default_call_rcu_data, 0, -1);
 564         call_rcu_unlock(&call_rcu_mutex);
 565         return default_call_rcu_data;
 566 }
 567
 568 /*
 569  * Return the call_rcu_data structure that applies to the currently
 570  * running thread.  Any call_rcu_data structure assigned specifically
 571  * to this thread has first priority, followed by any call_rcu_data
 572  * structure assigned to the CPU on which the thread is running,
 573  * followed by the default call_rcu_data structure.  If there is not
 574  * yet a default call_rcu_data structure, one will be created.
 575  *
 576  * Calls to this function and use of the returned call_rcu_data should
 577  * be protected by RCU read-side lock.
 578  */
 579 struct call_rcu_data *get_call_rcu_data(void)
 580 {
 581         struct call_rcu_data *crd;
 582
 583         if (URCU_TLS(thread_call_rcu_data) != NULL)
 584                 return URCU_TLS(thread_call_rcu_data);
 585
 586         if (maxcpus > 0) {
 587                 crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
 588                 if (crd)
 589                         return crd;
 590         }
 591
 592         return get_default_call_rcu_data();
 593 }
 594
 595 /*
 596  * Return a pointer to this task's call_rcu_data if there is one.
 597  */
 598
 599 struct call_rcu_data *get_thread_call_rcu_data(void)
 600 {
 601         return URCU_TLS(thread_call_rcu_data);
 602 }
 603
 604 /*
 605  * Set this task's call_rcu_data structure as specified, regardless
 606  * of whether or not this task already had one.  (This allows switching
 607  * to and from real-time call_rcu threads, for example.)
 608  *
 609  * Use NULL to remove a thread's call_rcu_data structure, but it is
 610  * the caller's responsibility to dispose of the removed structure.
 611  * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
 612  * (prior to NULLing it out, of course).
 613  */
 614
 615 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
 616 {
 617         URCU_TLS(thread_call_rcu_data) = crdp;
 618 }
 619
 620 /*
 621  * Create a separate call_rcu thread for each CPU.  This does not
 622  * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
 623  * function if you want that behavior. Should be paired with
 624  * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
 625  * threads.
 626  */
 627
 628 int create_all_cpu_call_rcu_data(unsigned long flags)
 629 {
 630         int i;
 631         struct call_rcu_data *crdp;
 632         int ret;
 633
 634         call_rcu_lock(&call_rcu_mutex);
 635         alloc_cpu_call_rcu_data();
 636         call_rcu_unlock(&call_rcu_mutex);
 637         if (maxcpus <= 0) {
 638                 errno = EINVAL;
 639                 return -EINVAL;
 640         }
 641         if (per_cpu_call_rcu_data == NULL) {
 642                 errno = ENOMEM;
 643                 return -ENOMEM;
 644         }
 645         for (i = 0; i < maxcpus; i++) {
 646                 call_rcu_lock(&call_rcu_mutex);
 647                 if (get_cpu_call_rcu_data(i)) {
 648                         call_rcu_unlock(&call_rcu_mutex);
 649                         continue;
 650                 }
 651                 crdp = __create_call_rcu_data(flags, i);
 652                 if (crdp == NULL) {
 653                         call_rcu_unlock(&call_rcu_mutex);
 654                         errno = ENOMEM;
 655                         return -ENOMEM;
 656                 }
 657                 call_rcu_unlock(&call_rcu_mutex);
 658                 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
 659                         call_rcu_data_free(crdp);
 660
 661                         /* it has been created by other thread */
 662                         if (ret == -EEXIST)
 663                                 continue;
 664
 665                         return ret;
 666                 }
 667         }
 668         return 0;
 669 }
 670
 671 /*
 672  * Wake up the call_rcu thread corresponding to the specified
 673  * call_rcu_data structure.
 674  */
 675 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
 676 {
 677         if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
 678                 call_rcu_wake_up(crdp);
 679 }
 680
 681 static void _call_rcu(struct rcu_head *head,
 682                       void (*func)(struct rcu_head *head),
 683                       struct call_rcu_data *crdp)
 684 {
 685         cds_wfcq_node_init(&head->next);
 686         head->func = func;
 687         cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
 688         uatomic_inc(&crdp->qlen);
 689         wake_call_rcu_thread(crdp);
 690 }
 691
 692 /*
 693  * Schedule a function to be invoked after a following grace period.
 694  * This is the only function that must be called -- the others are
 695  * only present to allow applications to tune their use of RCU for
 696  * maximum performance.
 697  *
 698  * Note that unless a call_rcu thread has not already been created,
 699  * the first invocation of call_rcu() will create one.  So, if you
 700  * need the first invocation of call_rcu() to be fast, make sure
 701  * to create a call_rcu thread first.  One way to accomplish this is
 702  * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
 703  *
 704  * call_rcu must be called by registered RCU read-side threads.
 705  */
 706 void call_rcu(struct rcu_head *head,
 707               void (*func)(struct rcu_head *head))
 708 {
 709         struct call_rcu_data *crdp;
 710
 711         /* Holding rcu read-side lock across use of per-cpu crdp */
 712         _rcu_read_lock();
 713         crdp = get_call_rcu_data();
 714         _call_rcu(head, func, crdp);
 715         _rcu_read_unlock();
 716 }
 717
 718 /*
 719  * Free up the specified call_rcu_data structure, terminating the
 720  * associated call_rcu thread.  The caller must have previously
 721  * removed the call_rcu_data structure from per-thread or per-CPU
 722  * usage.  For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
 723  * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
 724  * per-thread call_rcu_data structures.
 725  *
 726  * We silently refuse to free up the default call_rcu_data structure
 727  * because that is where we put any leftover callbacks.  Note that
 728  * the possibility of self-spawning callbacks makes it impossible
 729  * to execute all the callbacks in finite time without putting any
 730  * newly spawned callbacks somewhere else.  The "somewhere else" of
 731  * last resort is the default call_rcu_data structure.
 732  *
 733  * We also silently refuse to free NULL pointers.  This simplifies
 734  * the calling code.
 735  *
 736  * The caller must wait for a grace-period to pass between return from
 737  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 738  * previous call rcu data as argument.
 739  *
 740  * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
 741  * a list corruption bug in the 0.7.x series. The equivalent fix
 742  * appeared in 0.6.8 for the stable-0.6 branch.
 743  */
 744 void call_rcu_data_free(struct call_rcu_data *crdp)
 745 {
 746         if (crdp == NULL || crdp == default_call_rcu_data) {
 747                 return;
 748         }
 749         if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
 750                 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
 751                 wake_call_rcu_thread(crdp);
 752                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
 753                         (void) poll(NULL, 0, 1);
 754         }
 755         if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
 756                 /* Create default call rcu data if need be */
 757                 (void) get_default_call_rcu_data();
 758                 __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
 759                         &default_call_rcu_data->cbs_tail,
 760                         &crdp->cbs_head, &crdp->cbs_tail);
 761                 uatomic_add(&default_call_rcu_data->qlen,
 762                             uatomic_read(&crdp->qlen));
 763                 wake_call_rcu_thread(default_call_rcu_data);
 764         }
 765
 766         call_rcu_lock(&call_rcu_mutex);
 767         cds_list_del(&crdp->list);
 768         call_rcu_unlock(&call_rcu_mutex);
 769
 770         free(crdp);
 771 }
 772
 773 /*
 774  * Clean up all the per-CPU call_rcu threads.
 775  */
 776 void free_all_cpu_call_rcu_data(void)
 777 {
 778         int cpu;
 779         struct call_rcu_data **crdp;
 780         static int warned = 0;
 781
 782         if (maxcpus <= 0)
 783                 return;
 784
 785         crdp = malloc(sizeof(*crdp) * maxcpus);
 786         if (!crdp) {
 787                 if (!warned) {
 788                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 789                 }
 790                 warned = 1;
 791                 return;
 792         }
 793
 794         for (cpu = 0; cpu < maxcpus; cpu++) {
 795                 crdp[cpu] = get_cpu_call_rcu_data(cpu);
 796                 if (crdp[cpu] == NULL)
 797                         continue;
 798                 set_cpu_call_rcu_data(cpu, NULL);
 799         }
 800         /*
 801          * Wait for call_rcu sites acting as RCU readers of the
 802          * call_rcu_data to become quiescent.
 803          */
 804         synchronize_rcu();
 805         for (cpu = 0; cpu < maxcpus; cpu++) {
 806                 if (crdp[cpu] == NULL)
 807                         continue;
 808                 call_rcu_data_free(crdp[cpu]);
 809         }
 810         free(crdp);
 811 }
 812
 813 static
 814 void free_completion(struct urcu_ref *ref)
 815 {
 816         struct call_rcu_completion *completion;
 817
 818         completion = caa_container_of(ref, struct call_rcu_completion, ref);
 819         free(completion);
 820 }
 821
 822 static
 823 void _rcu_barrier_complete(struct rcu_head *head)
 824 {
 825         struct call_rcu_completion_work *work;
 826         struct call_rcu_completion *completion;
 827
 828         work = caa_container_of(head, struct call_rcu_completion_work, head);
 829         completion = work->completion;
 830         if (!uatomic_sub_return(&completion->barrier_count, 1))
 831                 call_rcu_completion_wake_up(completion);
 832         urcu_ref_put(&completion->ref, free_completion);
 833         free(work);
 834 }
 835
 836 /*
 837  * Wait for all in-flight call_rcu callbacks to complete execution.
 838  */
 839 void rcu_barrier(void)
 840 {
 841         struct call_rcu_data *crdp;
 842         struct call_rcu_completion *completion;
 843         int count = 0;
 844         int was_online;
 845
 846         /* Put in offline state in QSBR. */
 847         was_online = _rcu_read_ongoing();
 848         if (was_online)
 849                 rcu_thread_offline();
 850         /*
 851          * Calling a rcu_barrier() within a RCU read-side critical
 852          * section is an error.
 853          */
 854         if (_rcu_read_ongoing()) {
 855                 static int warned = 0;
 856
 857                 if (!warned) {
 858                         fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
 859                 }
 860                 warned = 1;
 861                 goto online;
 862         }
 863
 864         completion = calloc(sizeof(*completion), 1);
 865         if (!completion)
 866                 urcu_die(errno);
 867
 868         call_rcu_lock(&call_rcu_mutex);
 869         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 870                 count++;
 871
 872         /* Referenced by rcu_barrier() and each call_rcu thread. */
 873         urcu_ref_set(&completion->ref, count + 1);
 874         completion->barrier_count = count;
 875
 876         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 877                 struct call_rcu_completion_work *work;
 878
 879                 work = calloc(sizeof(*work), 1);
 880                 if (!work)
 881                         urcu_die(errno);
 882                 work->completion = completion;
 883                 _call_rcu(&work->head, _rcu_barrier_complete, crdp);
 884         }
 885         call_rcu_unlock(&call_rcu_mutex);
 886
 887         /* Wait for them */
 888         for (;;) {
 889                 uatomic_dec(&completion->futex);
 890                 /* Decrement futex before reading barrier_count */
 891                 cmm_smp_mb();
 892                 if (!uatomic_read(&completion->barrier_count))
 893                         break;
 894                 call_rcu_completion_wait(completion);
 895         }
 896
 897         urcu_ref_put(&completion->ref, free_completion);
 898
 899 online:
 900         if (was_online)
 901                 rcu_thread_online();
 902 }
 903
 904 /*
 905  * Acquire the call_rcu_mutex in order to ensure that the child sees
 906  * all of the call_rcu() data structures in a consistent state. Ensure
 907  * that all call_rcu threads are in a quiescent state across fork.
 908  * Suitable for pthread_atfork() and friends.
 909  */
 910 void call_rcu_before_fork(void)
 911 {
 912         struct call_rcu_data *crdp;
 913         struct urcu_atfork *atfork;
 914
 915         call_rcu_lock(&call_rcu_mutex);
 916
 917         atfork = registered_rculfhash_atfork;
 918         if (atfork)
 919                 atfork->before_fork(atfork->priv);
 920
 921         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 922                 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
 923                 cmm_smp_mb__after_uatomic_or();
 924                 wake_call_rcu_thread(crdp);
 925         }
 926         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 927                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0)
 928                         (void) poll(NULL, 0, 1);
 929         }
 930 }
 931
 932 /*
 933  * Clean up call_rcu data structures in the parent of a successful fork()
 934  * that is not followed by exec() in the child.  Suitable for
 935  * pthread_atfork() and friends.
 936  */
 937 void call_rcu_after_fork_parent(void)
 938 {
 939         struct call_rcu_data *crdp;
 940         struct urcu_atfork *atfork;
 941
 942         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 943                 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
 944         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 945                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
 946                         (void) poll(NULL, 0, 1);
 947         }
 948         atfork = registered_rculfhash_atfork;
 949         if (atfork)
 950                 atfork->after_fork_parent(atfork->priv);
 951         call_rcu_unlock(&call_rcu_mutex);
 952 }
 953
 954 /*
 955  * Clean up call_rcu data structures in the child of a successful fork()
 956  * that is not followed by exec().  Suitable for pthread_atfork() and
 957  * friends.
 958  */
 959 void call_rcu_after_fork_child(void)
 960 {
 961         struct call_rcu_data *crdp, *next;
 962         struct urcu_atfork *atfork;
 963
 964         /* Release the mutex. */
 965         call_rcu_unlock(&call_rcu_mutex);
 966
 967         atfork = registered_rculfhash_atfork;
 968         if (atfork)
 969                 atfork->after_fork_child(atfork->priv);
 970
 971         /* Do nothing when call_rcu() has not been used */
 972         if (cds_list_empty(&call_rcu_data_list))
 973                 return;
 974
 975         /*
 976          * Allocate a new default call_rcu_data structure in order
 977          * to get a working call_rcu thread to go with it.
 978          */
 979         default_call_rcu_data = NULL;
 980         (void)get_default_call_rcu_data();
 981
 982         /* Cleanup call_rcu_data pointers before use */
 983         maxcpus_reset();
 984         free(per_cpu_call_rcu_data);
 985         rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
 986         URCU_TLS(thread_call_rcu_data) = NULL;
 987
 988         /*
 989          * Dispose of all of the rest of the call_rcu_data structures.
 990          * Leftover call_rcu callbacks will be merged into the new
 991          * default call_rcu thread queue.
 992          */
 993         cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
 994                 if (crdp == default_call_rcu_data)
 995                         continue;
 996                 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
 997                 call_rcu_data_free(crdp);
 998         }
 999 }
1000
1001 void urcu_register_rculfhash_atfork(struct urcu_atfork *atfork)
1002 {
1003         call_rcu_lock(&call_rcu_mutex);
1004         if (registered_rculfhash_atfork_refcount++)
1005                 goto end;
1006         registered_rculfhash_atfork = atfork;
1007 end:
1008         call_rcu_unlock(&call_rcu_mutex);
1009 }
1010
1011 void urcu_unregister_rculfhash_atfork(struct urcu_atfork *atfork)
1012 {
1013         call_rcu_lock(&call_rcu_mutex);
1014         if (--registered_rculfhash_atfork_refcount)
1015                 goto end;
1016         registered_rculfhash_atfork = NULL;
1017 end:
1018         call_rcu_unlock(&call_rcu_mutex);
1019 }