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 <stdlib.h>
  28 #include <stdint.h>
  29 #include <string.h>
  30 #include <errno.h>
  31 #include <poll.h>
  32 #include <sys/time.h>
  33 #include <unistd.h>
  34 #include <sched.h>
  35
  36 #include "compat-getcpu.h"
  37 #include <urcu/assert.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 #include "urcu-utils.h"
  47
  48 #define SET_AFFINITY_CHECK_PERIOD               (1U << 8)       /* 256 */
  49 #define SET_AFFINITY_CHECK_PERIOD_MASK          (SET_AFFINITY_CHECK_PERIOD - 1)
  50
  51 /* Data structure that identifies a call_rcu thread. */
  52
  53 struct call_rcu_data {
  54         /*
  55          * We do not align head on a different cache-line than tail
  56          * mainly because call_rcu callback-invocation threads use
  57          * batching ("splice") to get an entire list of callbacks, which
  58          * effectively empties the queue, and requires to touch the tail
  59          * anyway.
  60          */
  61         struct cds_wfcq_tail cbs_tail;
  62         struct cds_wfcq_head cbs_head;
  63         unsigned long flags;
  64         int32_t futex;
  65         unsigned long qlen; /* maintained for debugging. */
  66         pthread_t tid;
  67         int cpu_affinity;
  68         unsigned long gp_count;
  69         struct cds_list_head list;
  70 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
  71
  72 struct call_rcu_completion {
  73         int barrier_count;
  74         int32_t futex;
  75         struct urcu_ref ref;
  76 };
  77
  78 struct call_rcu_completion_work {
  79         struct rcu_head head;
  80         struct call_rcu_completion *completion;
  81 };
  82
  83 /*
  84  * List of all call_rcu_data structures to keep valgrind happy.
  85  * Protected by call_rcu_mutex.
  86  */
  87
  88 static CDS_LIST_HEAD(call_rcu_data_list);
  89
  90 /* Link a thread using call_rcu() to its call_rcu thread. */
  91
  92 static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data);
  93
  94 /*
  95  * Guard call_rcu thread creation and atfork handlers.
  96  */
  97 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  98
  99 /* If a given thread does not have its own call_rcu thread, this is default. */
 100
 101 static struct call_rcu_data *default_call_rcu_data;
 102
 103 static struct urcu_atfork *registered_rculfhash_atfork;
 104 static unsigned long registered_rculfhash_atfork_refcount;
 105
 106 /*
 107  * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
 108  * available, then we can have call_rcu threads assigned to individual
 109  * CPUs rather than only to specific threads.
 110  */
 111
 112 #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
 113
 114 /*
 115  * Pointer to array of pointers to per-CPU call_rcu_data structures
 116  * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
 117  * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
 118  * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
 119  * without mutex. The call_rcu_mutex protects updates.
 120  */
 121
 122 static struct call_rcu_data **per_cpu_call_rcu_data;
 123 static long maxcpus;
 124
 125 static void maxcpus_reset(void)
 126 {
 127         maxcpus = 0;
 128 }
 129
 130 /* Allocate the array if it has not already been allocated. */
 131
 132 static void alloc_cpu_call_rcu_data(void)
 133 {
 134         struct call_rcu_data **p;
 135         static int warned = 0;
 136
 137         if (maxcpus != 0)
 138                 return;
 139         maxcpus = sysconf(_SC_NPROCESSORS_CONF);
 140         if (maxcpus <= 0) {
 141                 return;
 142         }
 143         p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
 144         if (p != NULL) {
 145                 memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
 146                 rcu_set_pointer(&per_cpu_call_rcu_data, p);
 147         } else {
 148                 if (!warned) {
 149                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 150                 }
 151                 warned = 1;
 152         }
 153 }
 154
 155 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 156
 157 /*
 158  * per_cpu_call_rcu_data should be constant, but some functions below, used both
 159  * for cases where cpu number is available and not available, assume it it not
 160  * constant.
 161  */
 162 static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
 163 static const long maxcpus = -1;
 164
 165 static void maxcpus_reset(void)
 166 {
 167 }
 168
 169 static void alloc_cpu_call_rcu_data(void)
 170 {
 171 }
 172
 173 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 174
 175 /* Acquire the specified pthread mutex. */
 176
 177 static void call_rcu_lock(pthread_mutex_t *pmp)
 178 {
 179         int ret;
 180
 181         ret = pthread_mutex_lock(pmp);
 182         if (ret)
 183                 urcu_die(ret);
 184 }
 185
 186 /* Release the specified pthread mutex. */
 187
 188 static void call_rcu_unlock(pthread_mutex_t *pmp)
 189 {
 190         int ret;
 191
 192         ret = pthread_mutex_unlock(pmp);
 193         if (ret)
 194                 urcu_die(ret);
 195 }
 196
 197 /*
 198  * Periodically retry setting CPU affinity if we migrate.
 199  * Losing affinity can be caused by CPU hotunplug/hotplug, or by
 200  * cpuset(7).
 201  */
 202 #ifdef HAVE_SCHED_SETAFFINITY
 203 static
 204 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
 205 {
 206         cpu_set_t mask;
 207         int ret;
 208
 209         if (crdp->cpu_affinity < 0)
 210                 return 0;
 211         if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK)
 212                 return 0;
 213         if (urcu_sched_getcpu() == crdp->cpu_affinity)
 214                 return 0;
 215
 216         CPU_ZERO(&mask);
 217         CPU_SET(crdp->cpu_affinity, &mask);
 218         ret = sched_setaffinity(0, sizeof(mask), &mask);
 219
 220         /*
 221          * EINVAL is fine: can be caused by hotunplugged CPUs, or by
 222          * cpuset(7). This is why we should always retry if we detect
 223          * migration.
 224          */
 225         if (ret && errno == EINVAL) {
 226                 ret = 0;
 227                 errno = 0;
 228         }
 229         return ret;
 230 }
 231 #else
 232 static
 233 int set_thread_cpu_affinity(struct call_rcu_data *crdp __attribute__((unused)))
 234 {
 235         return 0;
 236 }
 237 #endif
 238
 239 static void call_rcu_wait(struct call_rcu_data *crdp)
 240 {
 241         /* Read call_rcu list before read futex */
 242         cmm_smp_mb();
 243         if (uatomic_read(&crdp->futex) != -1)
 244                 return;
 245         while (futex_async(&crdp->futex, FUTEX_WAIT, -1,
 246                         NULL, NULL, 0)) {
 247                 switch (errno) {
 248                 case EWOULDBLOCK:
 249                         /* Value already changed. */
 250                         return;
 251                 case EINTR:
 252                         /* Retry if interrupted by signal. */
 253                         break;  /* Get out of switch. */
 254                 default:
 255                         /* Unexpected error. */
 256                         urcu_die(errno);
 257                 }
 258         }
 259 }
 260
 261 static void call_rcu_wake_up(struct call_rcu_data *crdp)
 262 {
 263         /* Write to call_rcu list before reading/writing futex */
 264         cmm_smp_mb();
 265         if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
 266                 uatomic_set(&crdp->futex, 0);
 267                 if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
 268                                 NULL, NULL, 0) < 0)
 269                         urcu_die(errno);
 270         }
 271 }
 272
 273 static void call_rcu_completion_wait(struct call_rcu_completion *completion)
 274 {
 275         /* Read completion barrier count before read futex */
 276         cmm_smp_mb();
 277         if (uatomic_read(&completion->futex) != -1)
 278                 return;
 279         while (futex_async(&completion->futex, FUTEX_WAIT, -1,
 280                         NULL, NULL, 0)) {
 281                 switch (errno) {
 282                 case EWOULDBLOCK:
 283                         /* Value already changed. */
 284                         return;
 285                 case EINTR:
 286                         /* Retry if interrupted by signal. */
 287                         break;  /* Get out of switch. */
 288                 default:
 289                         /* Unexpected error. */
 290                         urcu_die(errno);
 291                 }
 292         }
 293 }
 294
 295 static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
 296 {
 297         /* Write to completion barrier count before reading/writing futex */
 298         cmm_smp_mb();
 299         if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
 300                 uatomic_set(&completion->futex, 0);
 301                 if (futex_async(&completion->futex, FUTEX_WAKE, 1,
 302                                 NULL, NULL, 0) < 0)
 303                         urcu_die(errno);
 304         }
 305 }
 306
 307 /* This is the code run by each call_rcu thread. */
 308
 309 static void *call_rcu_thread(void *arg)
 310 {
 311         unsigned long cbcount;
 312         struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
 313         int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
 314
 315         if (set_thread_cpu_affinity(crdp))
 316                 urcu_die(errno);
 317
 318         /*
 319          * If callbacks take a read-side lock, we need to be registered.
 320          */
 321         rcu_register_thread();
 322
 323         URCU_TLS(thread_call_rcu_data) = crdp;
 324         if (!rt) {
 325                 uatomic_dec(&crdp->futex);
 326                 /* Decrement futex before reading call_rcu list */
 327                 cmm_smp_mb();
 328         }
 329         for (;;) {
 330                 struct cds_wfcq_head cbs_tmp_head;
 331                 struct cds_wfcq_tail cbs_tmp_tail;
 332                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 333                 enum cds_wfcq_ret splice_ret;
 334
 335                 if (set_thread_cpu_affinity(crdp))
 336                         urcu_die(errno);
 337
 338                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
 339                         /*
 340                          * Pause requested. Become quiescent: remove
 341                          * ourself from all global lists, and don't
 342                          * process any callback. The callback lists may
 343                          * still be non-empty though.
 344                          */
 345                         rcu_unregister_thread();
 346                         cmm_smp_mb__before_uatomic_or();
 347                         uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
 348                         while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
 349                                 (void) poll(NULL, 0, 1);
 350                         uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
 351                         cmm_smp_mb__after_uatomic_and();
 352                         rcu_register_thread();
 353                 }
 354
 355                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 356                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 357                         &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail);
 358                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 359                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 360                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 361                         synchronize_rcu();
 362                         cbcount = 0;
 363                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 364                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 365                                 struct rcu_head *rhp;
 366
 367                                 rhp = caa_container_of(cbs,
 368                                         struct rcu_head, next);
 369                                 rhp->func(rhp);
 370                                 cbcount++;
 371                         }
 372                         uatomic_sub(&crdp->qlen, cbcount);
 373                 }
 374                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
 375                         break;
 376                 rcu_thread_offline();
 377                 if (!rt) {
 378                         if (cds_wfcq_empty(&crdp->cbs_head,
 379                                         &crdp->cbs_tail)) {
 380                                 call_rcu_wait(crdp);
 381                                 (void) poll(NULL, 0, 10);
 382                                 uatomic_dec(&crdp->futex);
 383                                 /*
 384                                  * Decrement futex before reading
 385                                  * call_rcu list.
 386                                  */
 387                                 cmm_smp_mb();
 388                         } else {
 389                                 (void) poll(NULL, 0, 10);
 390                         }
 391                 } else {
 392                         (void) poll(NULL, 0, 10);
 393                 }
 394                 rcu_thread_online();
 395         }
 396         if (!rt) {
 397                 /*
 398                  * Read call_rcu list before write futex.
 399                  */
 400                 cmm_smp_mb();
 401                 uatomic_set(&crdp->futex, 0);
 402         }
 403         uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
 404         rcu_unregister_thread();
 405         return NULL;
 406 }
 407
 408 /*
 409  * Create both a call_rcu thread and the corresponding call_rcu_data
 410  * structure, linking the structure in as specified.  Caller must hold
 411  * call_rcu_mutex.
 412  */
 413
 414 static void call_rcu_data_init(struct call_rcu_data **crdpp,
 415                                unsigned long flags,
 416                                int cpu_affinity)
 417 {
 418         struct call_rcu_data *crdp;
 419         int ret;
 420
 421         crdp = malloc(sizeof(*crdp));
 422         if (crdp == NULL)
 423                 urcu_die(errno);
 424         memset(crdp, '\0', sizeof(*crdp));
 425         cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail);
 426         crdp->qlen = 0;
 427         crdp->futex = 0;
 428         crdp->flags = flags;
 429         cds_list_add(&crdp->list, &call_rcu_data_list);
 430         crdp->cpu_affinity = cpu_affinity;
 431         crdp->gp_count = 0;
 432         cmm_smp_mb();  /* Structure initialized before pointer is planted. */
 433         *crdpp = crdp;
 434         ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
 435         if (ret)
 436                 urcu_die(ret);
 437 }
 438
 439 /*
 440  * Return a pointer to the call_rcu_data structure for the specified
 441  * CPU, returning NULL if there is none.  We cannot automatically
 442  * created it because the platform we are running on might not define
 443  * urcu_sched_getcpu().
 444  *
 445  * The call to this function and use of the returned call_rcu_data
 446  * should be protected by RCU read-side lock.
 447  */
 448
 449 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
 450 {
 451         static int warned = 0;
 452         struct call_rcu_data **pcpu_crdp;
 453
 454         pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
 455         if (pcpu_crdp == NULL)
 456                 return NULL;
 457         if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
 458                 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
 459                 warned = 1;
 460         }
 461         if (cpu < 0 || maxcpus <= cpu)
 462                 return NULL;
 463         return rcu_dereference(pcpu_crdp[cpu]);
 464 }
 465
 466 /*
 467  * Return the tid corresponding to the call_rcu thread whose
 468  * call_rcu_data structure is specified.
 469  */
 470
 471 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
 472 {
 473         return crdp->tid;
 474 }
 475
 476 /*
 477  * Create a call_rcu_data structure (with thread) and return a pointer.
 478  */
 479
 480 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
 481                                                     int cpu_affinity)
 482 {
 483         struct call_rcu_data *crdp;
 484
 485         call_rcu_data_init(&crdp, flags, cpu_affinity);
 486         return crdp;
 487 }
 488
 489 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
 490                                            int cpu_affinity)
 491 {
 492         struct call_rcu_data *crdp;
 493
 494         call_rcu_lock(&call_rcu_mutex);
 495         crdp = __create_call_rcu_data(flags, cpu_affinity);
 496         call_rcu_unlock(&call_rcu_mutex);
 497         return crdp;
 498 }
 499
 500 /*
 501  * Set the specified CPU to use the specified call_rcu_data structure.
 502  *
 503  * Use NULL to remove a CPU's call_rcu_data structure, but it is
 504  * the caller's responsibility to dispose of the removed structure.
 505  * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
 506  * (prior to NULLing it out, of course).
 507  *
 508  * The caller must wait for a grace-period to pass between return from
 509  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 510  * previous call rcu data as argument.
 511  */
 512
 513 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
 514 {
 515         static int warned = 0;
 516
 517         call_rcu_lock(&call_rcu_mutex);
 518         alloc_cpu_call_rcu_data();
 519         if (cpu < 0 || maxcpus <= cpu) {
 520                 if (!warned) {
 521                         fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
 522                         warned = 1;
 523                 }
 524                 call_rcu_unlock(&call_rcu_mutex);
 525                 errno = EINVAL;
 526                 return -EINVAL;
 527         }
 528
 529         if (per_cpu_call_rcu_data == NULL) {
 530                 call_rcu_unlock(&call_rcu_mutex);
 531                 errno = ENOMEM;
 532                 return -ENOMEM;
 533         }
 534
 535         if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
 536                 call_rcu_unlock(&call_rcu_mutex);
 537                 errno = EEXIST;
 538                 return -EEXIST;
 539         }
 540
 541         rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
 542         call_rcu_unlock(&call_rcu_mutex);
 543         return 0;
 544 }
 545
 546 /*
 547  * Return a pointer to the default call_rcu_data structure, creating
 548  * one if need be.  Because we never free call_rcu_data structures,
 549  * we don't need to be in an RCU read-side critical section.
 550  */
 551
 552 struct call_rcu_data *get_default_call_rcu_data(void)
 553 {
 554         if (default_call_rcu_data != NULL)
 555                 return rcu_dereference(default_call_rcu_data);
 556         call_rcu_lock(&call_rcu_mutex);
 557         if (default_call_rcu_data != NULL) {
 558                 call_rcu_unlock(&call_rcu_mutex);
 559                 return default_call_rcu_data;
 560         }
 561         call_rcu_data_init(&default_call_rcu_data, 0, -1);
 562         call_rcu_unlock(&call_rcu_mutex);
 563         return default_call_rcu_data;
 564 }
 565
 566 /*
 567  * Return the call_rcu_data structure that applies to the currently
 568  * running thread.  Any call_rcu_data structure assigned specifically
 569  * to this thread has first priority, followed by any call_rcu_data
 570  * structure assigned to the CPU on which the thread is running,
 571  * followed by the default call_rcu_data structure.  If there is not
 572  * yet a default call_rcu_data structure, one will be created.
 573  *
 574  * Calls to this function and use of the returned call_rcu_data should
 575  * be protected by RCU read-side lock.
 576  */
 577 struct call_rcu_data *get_call_rcu_data(void)
 578 {
 579         struct call_rcu_data *crd;
 580
 581         if (URCU_TLS(thread_call_rcu_data) != NULL)
 582                 return URCU_TLS(thread_call_rcu_data);
 583
 584         if (maxcpus > 0) {
 585                 crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
 586                 if (crd)
 587                         return crd;
 588         }
 589
 590         return get_default_call_rcu_data();
 591 }
 592
 593 /*
 594  * Return a pointer to this task's call_rcu_data if there is one.
 595  */
 596
 597 struct call_rcu_data *get_thread_call_rcu_data(void)
 598 {
 599         return URCU_TLS(thread_call_rcu_data);
 600 }
 601
 602 /*
 603  * Set this task's call_rcu_data structure as specified, regardless
 604  * of whether or not this task already had one.  (This allows switching
 605  * to and from real-time call_rcu threads, for example.)
 606  *
 607  * Use NULL to remove a thread's call_rcu_data structure, but it is
 608  * the caller's responsibility to dispose of the removed structure.
 609  * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
 610  * (prior to NULLing it out, of course).
 611  */
 612
 613 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
 614 {
 615         URCU_TLS(thread_call_rcu_data) = crdp;
 616 }
 617
 618 /*
 619  * Create a separate call_rcu thread for each CPU.  This does not
 620  * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
 621  * function if you want that behavior. Should be paired with
 622  * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
 623  * threads.
 624  */
 625
 626 int create_all_cpu_call_rcu_data(unsigned long flags)
 627 {
 628         int i;
 629         struct call_rcu_data *crdp;
 630         int ret;
 631
 632         call_rcu_lock(&call_rcu_mutex);
 633         alloc_cpu_call_rcu_data();
 634         call_rcu_unlock(&call_rcu_mutex);
 635         if (maxcpus <= 0) {
 636                 errno = EINVAL;
 637                 return -EINVAL;
 638         }
 639         if (per_cpu_call_rcu_data == NULL) {
 640                 errno = ENOMEM;
 641                 return -ENOMEM;
 642         }
 643         for (i = 0; i < maxcpus; i++) {
 644                 call_rcu_lock(&call_rcu_mutex);
 645                 if (get_cpu_call_rcu_data(i)) {
 646                         call_rcu_unlock(&call_rcu_mutex);
 647                         continue;
 648                 }
 649                 crdp = __create_call_rcu_data(flags, i);
 650                 if (crdp == NULL) {
 651                         call_rcu_unlock(&call_rcu_mutex);
 652                         errno = ENOMEM;
 653                         return -ENOMEM;
 654                 }
 655                 call_rcu_unlock(&call_rcu_mutex);
 656                 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
 657                         call_rcu_data_free(crdp);
 658
 659                         /* it has been created by other thread */
 660                         if (ret == -EEXIST)
 661                                 continue;
 662
 663                         return ret;
 664                 }
 665         }
 666         return 0;
 667 }
 668
 669 /*
 670  * Wake up the call_rcu thread corresponding to the specified
 671  * call_rcu_data structure.
 672  */
 673 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
 674 {
 675         if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
 676                 call_rcu_wake_up(crdp);
 677 }
 678
 679 static void _call_rcu(struct rcu_head *head,
 680                       void (*func)(struct rcu_head *head),
 681                       struct call_rcu_data *crdp)
 682 {
 683         cds_wfcq_node_init(&head->next);
 684         head->func = func;
 685         cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
 686         uatomic_inc(&crdp->qlen);
 687         wake_call_rcu_thread(crdp);
 688 }
 689
 690 /*
 691  * Schedule a function to be invoked after a following grace period.
 692  * This is the only function that must be called -- the others are
 693  * only present to allow applications to tune their use of RCU for
 694  * maximum performance.
 695  *
 696  * Note that unless a call_rcu thread has not already been created,
 697  * the first invocation of call_rcu() will create one.  So, if you
 698  * need the first invocation of call_rcu() to be fast, make sure
 699  * to create a call_rcu thread first.  One way to accomplish this is
 700  * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
 701  *
 702  * call_rcu must be called by registered RCU read-side threads.
 703  */
 704 void call_rcu(struct rcu_head *head,
 705               void (*func)(struct rcu_head *head))
 706 {
 707         struct call_rcu_data *crdp;
 708
 709         /* Holding rcu read-side lock across use of per-cpu crdp */
 710         _rcu_read_lock();
 711         crdp = get_call_rcu_data();
 712         _call_rcu(head, func, crdp);
 713         _rcu_read_unlock();
 714 }
 715
 716 /*
 717  * Free up the specified call_rcu_data structure, terminating the
 718  * associated call_rcu thread.  The caller must have previously
 719  * removed the call_rcu_data structure from per-thread or per-CPU
 720  * usage.  For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
 721  * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
 722  * per-thread call_rcu_data structures.
 723  *
 724  * We silently refuse to free up the default call_rcu_data structure
 725  * because that is where we put any leftover callbacks.  Note that
 726  * the possibility of self-spawning callbacks makes it impossible
 727  * to execute all the callbacks in finite time without putting any
 728  * newly spawned callbacks somewhere else.  The "somewhere else" of
 729  * last resort is the default call_rcu_data structure.
 730  *
 731  * We also silently refuse to free NULL pointers.  This simplifies
 732  * the calling code.
 733  *
 734  * The caller must wait for a grace-period to pass between return from
 735  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 736  * previous call rcu data as argument.
 737  *
 738  * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
 739  * a list corruption bug in the 0.7.x series. The equivalent fix
 740  * appeared in 0.6.8 for the stable-0.6 branch.
 741  */
 742 void call_rcu_data_free(struct call_rcu_data *crdp)
 743 {
 744         if (crdp == NULL || crdp == default_call_rcu_data) {
 745                 return;
 746         }
 747         if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
 748                 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
 749                 wake_call_rcu_thread(crdp);
 750                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
 751                         (void) poll(NULL, 0, 1);
 752         }
 753         call_rcu_lock(&call_rcu_mutex);
 754         if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
 755                 call_rcu_unlock(&call_rcu_mutex);
 756                 /* Create default call rcu data if need be. */
 757                 /* CBs queued here will be handed to the default list. */
 758                 (void) get_default_call_rcu_data();
 759                 call_rcu_lock(&call_rcu_mutex);
 760                 __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
 761                         &default_call_rcu_data->cbs_tail,
 762                         &crdp->cbs_head, &crdp->cbs_tail);
 763                 uatomic_add(&default_call_rcu_data->qlen,
 764                             uatomic_read(&crdp->qlen));
 765                 wake_call_rcu_thread(default_call_rcu_data);
 766         }
 767
 768         cds_list_del(&crdp->list);
 769         call_rcu_unlock(&call_rcu_mutex);
 770
 771         free(crdp);
 772 }
 773
 774 /*
 775  * Clean up all the per-CPU call_rcu threads.
 776  */
 777 void free_all_cpu_call_rcu_data(void)
 778 {
 779         int cpu;
 780         struct call_rcu_data **crdp;
 781         static int warned = 0;
 782
 783         if (maxcpus <= 0)
 784                 return;
 785
 786         crdp = malloc(sizeof(*crdp) * maxcpus);
 787         if (!crdp) {
 788                 if (!warned) {
 789                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 790                 }
 791                 warned = 1;
 792                 return;
 793         }
 794
 795         for (cpu = 0; cpu < maxcpus; cpu++) {
 796                 crdp[cpu] = get_cpu_call_rcu_data(cpu);
 797                 if (crdp[cpu] == NULL)
 798                         continue;
 799                 set_cpu_call_rcu_data(cpu, NULL);
 800         }
 801         /*
 802          * Wait for call_rcu sites acting as RCU readers of the
 803          * call_rcu_data to become quiescent.
 804          */
 805         synchronize_rcu();
 806         for (cpu = 0; cpu < maxcpus; cpu++) {
 807                 if (crdp[cpu] == NULL)
 808                         continue;
 809                 call_rcu_data_free(crdp[cpu]);
 810         }
 811         free(crdp);
 812 }
 813
 814 static
 815 void free_completion(struct urcu_ref *ref)
 816 {
 817         struct call_rcu_completion *completion;
 818
 819         completion = caa_container_of(ref, struct call_rcu_completion, ref);
 820         free(completion);
 821 }
 822
 823 static
 824 void _rcu_barrier_complete(struct rcu_head *head)
 825 {
 826         struct call_rcu_completion_work *work;
 827         struct call_rcu_completion *completion;
 828
 829         work = caa_container_of(head, struct call_rcu_completion_work, head);
 830         completion = work->completion;
 831         if (!uatomic_sub_return(&completion->barrier_count, 1))
 832                 call_rcu_completion_wake_up(completion);
 833         urcu_ref_put(&completion->ref, free_completion);
 834         free(work);
 835 }
 836
 837 /*
 838  * Wait for all in-flight call_rcu callbacks to complete execution.
 839  */
 840 void rcu_barrier(void)
 841 {
 842         struct call_rcu_data *crdp;
 843         struct call_rcu_completion *completion;
 844         int count = 0;
 845         int was_online;
 846
 847         /* Put in offline state in QSBR. */
 848         was_online = _rcu_read_ongoing();
 849         if (was_online)
 850                 rcu_thread_offline();
 851         /*
 852          * Calling a rcu_barrier() within a RCU read-side critical
 853          * section is an error.
 854          */
 855         if (_rcu_read_ongoing()) {
 856                 static int warned = 0;
 857
 858                 if (!warned) {
 859                         fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
 860                 }
 861                 warned = 1;
 862                 goto online;
 863         }
 864
 865         completion = calloc(sizeof(*completion), 1);
 866         if (!completion)
 867                 urcu_die(errno);
 868
 869         call_rcu_lock(&call_rcu_mutex);
 870         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 871                 count++;
 872
 873         /* Referenced by rcu_barrier() and each call_rcu thread. */
 874         urcu_ref_set(&completion->ref, count + 1);
 875         completion->barrier_count = count;
 876
 877         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 878                 struct call_rcu_completion_work *work;
 879
 880                 work = calloc(sizeof(*work), 1);
 881                 if (!work)
 882                         urcu_die(errno);
 883                 work->completion = completion;
 884                 _call_rcu(&work->head, _rcu_barrier_complete, crdp);
 885         }
 886         call_rcu_unlock(&call_rcu_mutex);
 887
 888         /* Wait for them */
 889         for (;;) {
 890                 uatomic_dec(&completion->futex);
 891                 /* Decrement futex before reading barrier_count */
 892                 cmm_smp_mb();
 893                 if (!uatomic_read(&completion->barrier_count))
 894                         break;
 895                 call_rcu_completion_wait(completion);
 896         }
 897
 898         urcu_ref_put(&completion->ref, free_completion);
 899
 900 online:
 901         if (was_online)
 902                 rcu_thread_online();
 903 }
 904
 905 /*
 906  * Acquire the call_rcu_mutex in order to ensure that the child sees
 907  * all of the call_rcu() data structures in a consistent state. Ensure
 908  * that all call_rcu threads are in a quiescent state across fork.
 909  * Suitable for pthread_atfork() and friends.
 910  */
 911 void call_rcu_before_fork(void)
 912 {
 913         struct call_rcu_data *crdp;
 914         struct urcu_atfork *atfork;
 915
 916         call_rcu_lock(&call_rcu_mutex);
 917
 918         atfork = registered_rculfhash_atfork;
 919         if (atfork)
 920                 atfork->before_fork(atfork->priv);
 921
 922         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 923                 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
 924                 cmm_smp_mb__after_uatomic_or();
 925                 wake_call_rcu_thread(crdp);
 926         }
 927         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 928                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0)
 929                         (void) poll(NULL, 0, 1);
 930         }
 931 }
 932
 933 /*
 934  * Clean up call_rcu data structures in the parent of a successful fork()
 935  * that is not followed by exec() in the child.  Suitable for
 936  * pthread_atfork() and friends.
 937  */
 938 void call_rcu_after_fork_parent(void)
 939 {
 940         struct call_rcu_data *crdp;
 941         struct urcu_atfork *atfork;
 942
 943         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 944                 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
 945         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 946                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
 947                         (void) poll(NULL, 0, 1);
 948         }
 949         atfork = registered_rculfhash_atfork;
 950         if (atfork)
 951                 atfork->after_fork_parent(atfork->priv);
 952         call_rcu_unlock(&call_rcu_mutex);
 953 }
 954
 955 /*
 956  * Clean up call_rcu data structures in the child of a successful fork()
 957  * that is not followed by exec().  Suitable for pthread_atfork() and
 958  * friends.
 959  */
 960 void call_rcu_after_fork_child(void)
 961 {
 962         struct call_rcu_data *crdp, *next;
 963         struct urcu_atfork *atfork;
 964
 965         /* Release the mutex. */
 966         call_rcu_unlock(&call_rcu_mutex);
 967
 968         atfork = registered_rculfhash_atfork;
 969         if (atfork)
 970                 atfork->after_fork_child(atfork->priv);
 971
 972         /* Do nothing when call_rcu() has not been used */
 973         if (cds_list_empty(&call_rcu_data_list))
 974                 return;
 975
 976         /*
 977          * Allocate a new default call_rcu_data structure in order
 978          * to get a working call_rcu thread to go with it.
 979          */
 980         default_call_rcu_data = NULL;
 981         (void)get_default_call_rcu_data();
 982
 983         /* Cleanup call_rcu_data pointers before use */
 984         maxcpus_reset();
 985         free(per_cpu_call_rcu_data);
 986         rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
 987         URCU_TLS(thread_call_rcu_data) = NULL;
 988
 989         /*
 990          * Dispose of all of the rest of the call_rcu_data structures.
 991          * Leftover call_rcu callbacks will be merged into the new
 992          * default call_rcu thread queue.
 993          */
 994         cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
 995                 if (crdp == default_call_rcu_data)
 996                         continue;
 997                 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
 998                 call_rcu_data_free(crdp);
 999         }
1000 }
1001
1002 void urcu_register_rculfhash_atfork(struct urcu_atfork *atfork)
1003 {
1004         call_rcu_lock(&call_rcu_mutex);
1005         if (registered_rculfhash_atfork_refcount++)
1006                 goto end;
1007         registered_rculfhash_atfork = atfork;
1008 end:
1009         call_rcu_unlock(&call_rcu_mutex);
1010 }
1011
1012 void urcu_unregister_rculfhash_atfork(struct urcu_atfork *atfork __attribute__((unused)))
1013 {
1014         call_rcu_lock(&call_rcu_mutex);
1015         if (--registered_rculfhash_atfork_refcount)
1016                 goto end;
1017         registered_rculfhash_atfork = NULL;
1018 end:
1019         call_rcu_unlock(&call_rcu_mutex);
1020 }