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 #include "urcu-utils.h"
  47 #include "compat-smp.h"
  48
  49 #define SET_AFFINITY_CHECK_PERIOD               (1U << 8)       /* 256 */
  50 #define SET_AFFINITY_CHECK_PERIOD_MASK          (SET_AFFINITY_CHECK_PERIOD - 1)
  51
  52 /* Data structure that identifies a call_rcu thread. */
  53
  54 struct call_rcu_data {
  55         /*
  56          * We do not align head on a different cache-line than tail
  57          * mainly because call_rcu callback-invocation threads use
  58          * batching ("splice") to get an entire list of callbacks, which
  59          * effectively empties the queue, and requires to touch the tail
  60          * anyway.
  61          */
  62         struct cds_wfcq_tail cbs_tail;
  63         struct cds_wfcq_head cbs_head;
  64         unsigned long flags;
  65         int32_t futex;
  66         unsigned long qlen; /* maintained for debugging. */
  67         pthread_t tid;
  68         int cpu_affinity;
  69         unsigned long gp_count;
  70         struct cds_list_head list;
  71 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
  72
  73 struct call_rcu_completion {
  74         int barrier_count;
  75         int32_t futex;
  76         struct urcu_ref ref;
  77 };
  78
  79 struct call_rcu_completion_work {
  80         struct rcu_head head;
  81         struct call_rcu_completion *completion;
  82 };
  83
  84 /*
  85  * List of all call_rcu_data structures to keep valgrind happy.
  86  * Protected by call_rcu_mutex.
  87  */
  88
  89 static CDS_LIST_HEAD(call_rcu_data_list);
  90
  91 /* Link a thread using call_rcu() to its call_rcu thread. */
  92
  93 static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data);
  94
  95 /*
  96  * Guard call_rcu thread creation and atfork handlers.
  97  */
  98 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  99
 100 /* If a given thread does not have its own call_rcu thread, this is default. */
 101
 102 static struct call_rcu_data *default_call_rcu_data;
 103
 104 static struct urcu_atfork *registered_rculfhash_atfork;
 105 static unsigned long registered_rculfhash_atfork_refcount;
 106
 107 /*
 108  * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
 109  * available, then we can have call_rcu threads assigned to individual
 110  * CPUs rather than only to specific threads.
 111  */
 112
 113 #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
 114
 115 /*
 116  * Pointer to array of pointers to per-CPU call_rcu_data structures
 117  * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
 118  * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
 119  * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
 120  * without mutex. The call_rcu_mutex protects updates.
 121  */
 122
 123 static struct call_rcu_data **per_cpu_call_rcu_data;
 124 static long cpus_array_len;
 125
 126 static void cpus_array_len_reset(void)
 127 {
 128         cpus_array_len = 0;
 129 }
 130
 131 /* Allocate the array if it has not already been allocated. */
 132
 133 static void alloc_cpu_call_rcu_data(void)
 134 {
 135         struct call_rcu_data **p;
 136         static int warned = 0;
 137
 138         if (cpus_array_len != 0)
 139                 return;
 140         cpus_array_len = get_possible_cpus_array_len();
 141         if (cpus_array_len <= 0) {
 142                 return;
 143         }
 144         p = malloc(cpus_array_len * sizeof(*per_cpu_call_rcu_data));
 145         if (p != NULL) {
 146                 memset(p, '\0', cpus_array_len * sizeof(*per_cpu_call_rcu_data));
 147                 rcu_set_pointer(&per_cpu_call_rcu_data, p);
 148         } else {
 149                 if (!warned) {
 150                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 151                 }
 152                 warned = 1;
 153         }
 154 }
 155
 156 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 157
 158 /*
 159  * per_cpu_call_rcu_data should be constant, but some functions below, used both
 160  * for cases where cpu number is available and not available, assume it it not
 161  * constant.
 162  */
 163 static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
 164 static const long cpus_array_len = -1;
 165
 166 static void cpus_array_len_reset(void)
 167 {
 168 }
 169
 170 static void alloc_cpu_call_rcu_data(void)
 171 {
 172 }
 173
 174 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 175
 176 /* Acquire the specified pthread mutex. */
 177
 178 static void call_rcu_lock(pthread_mutex_t *pmp)
 179 {
 180         int ret;
 181
 182         ret = pthread_mutex_lock(pmp);
 183         if (ret)
 184                 urcu_die(ret);
 185 }
 186
 187 /* Release the specified pthread mutex. */
 188
 189 static void call_rcu_unlock(pthread_mutex_t *pmp)
 190 {
 191         int ret;
 192
 193         ret = pthread_mutex_unlock(pmp);
 194         if (ret)
 195                 urcu_die(ret);
 196 }
 197
 198 /*
 199  * Periodically retry setting CPU affinity if we migrate.
 200  * Losing affinity can be caused by CPU hotunplug/hotplug, or by
 201  * cpuset(7).
 202  */
 203 #ifdef HAVE_SCHED_SETAFFINITY
 204 static
 205 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
 206 {
 207         cpu_set_t mask;
 208         int ret;
 209
 210         if (crdp->cpu_affinity < 0)
 211                 return 0;
 212         if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK)
 213                 return 0;
 214         if (urcu_sched_getcpu() == crdp->cpu_affinity)
 215                 return 0;
 216
 217         CPU_ZERO(&mask);
 218         CPU_SET(crdp->cpu_affinity, &mask);
 219         ret = sched_setaffinity(0, sizeof(mask), &mask);
 220
 221         /*
 222          * EINVAL is fine: can be caused by hotunplugged CPUs, or by
 223          * cpuset(7). This is why we should always retry if we detect
 224          * migration.
 225          */
 226         if (ret && errno == EINVAL) {
 227                 ret = 0;
 228                 errno = 0;
 229         }
 230         return ret;
 231 }
 232 #else
 233 static
 234 int set_thread_cpu_affinity(struct call_rcu_data *crdp __attribute__((unused)))
 235 {
 236         return 0;
 237 }
 238 #endif
 239
 240 static void call_rcu_wait(struct call_rcu_data *crdp)
 241 {
 242         /* Read call_rcu list before read futex */
 243         cmm_smp_mb();
 244         while (uatomic_read(&crdp->futex) == -1) {
 245                 if (!futex_async(&crdp->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 246                         /*
 247                          * Prior queued wakeups queued by unrelated code
 248                          * using the same address can cause futex wait to
 249                          * return 0 even through the futex value is still
 250                          * -1 (spurious wakeups). Check the value again
 251                          * in user-space to validate whether it really
 252                          * differs from -1.
 253                          */
 254                         continue;
 255                 }
 256                 switch (errno) {
 257                 case EAGAIN:
 258                         /* Value already changed. */
 259                         return;
 260                 case EINTR:
 261                         /* Retry if interrupted by signal. */
 262                         break;  /* Get out of switch. Check again. */
 263                 default:
 264                         /* Unexpected error. */
 265                         urcu_die(errno);
 266                 }
 267         }
 268 }
 269
 270 static void call_rcu_wake_up(struct call_rcu_data *crdp)
 271 {
 272         /* Write to call_rcu list before reading/writing futex */
 273         cmm_smp_mb();
 274         if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
 275                 uatomic_set(&crdp->futex, 0);
 276                 if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
 277                                 NULL, NULL, 0) < 0)
 278                         urcu_die(errno);
 279         }
 280 }
 281
 282 static void call_rcu_completion_wait(struct call_rcu_completion *completion)
 283 {
 284         /* Read completion barrier count before read futex */
 285         cmm_smp_mb();
 286         while (uatomic_read(&completion->futex) == -1) {
 287                 if (!futex_async(&completion->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 288                         /*
 289                          * Prior queued wakeups queued by unrelated code
 290                          * using the same address can cause futex wait to
 291                          * return 0 even through the futex value is still
 292                          * -1 (spurious wakeups). Check the value again
 293                          * in user-space to validate whether it really
 294                          * differs from -1.
 295                          */
 296                         continue;
 297                 }
 298                 switch (errno) {
 299                 case EAGAIN:
 300                         /* Value already changed. */
 301                         return;
 302                 case EINTR:
 303                         /* Retry if interrupted by signal. */
 304                         break;  /* Get out of switch. Check again. */
 305                 default:
 306                         /* Unexpected error. */
 307                         urcu_die(errno);
 308                 }
 309         }
 310 }
 311
 312 static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
 313 {
 314         /* Write to completion barrier count before reading/writing futex */
 315         cmm_smp_mb();
 316         if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
 317                 uatomic_set(&completion->futex, 0);
 318                 if (futex_async(&completion->futex, FUTEX_WAKE, 1,
 319                                 NULL, NULL, 0) < 0)
 320                         urcu_die(errno);
 321         }
 322 }
 323
 324 /* This is the code run by each call_rcu thread. */
 325
 326 static void *call_rcu_thread(void *arg)
 327 {
 328         unsigned long cbcount;
 329         struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
 330         int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
 331
 332         if (set_thread_cpu_affinity(crdp))
 333                 urcu_die(errno);
 334
 335         /*
 336          * If callbacks take a read-side lock, we need to be registered.
 337          */
 338         rcu_register_thread();
 339
 340         URCU_TLS(thread_call_rcu_data) = crdp;
 341         if (!rt) {
 342                 uatomic_dec(&crdp->futex);
 343                 /* Decrement futex before reading call_rcu list */
 344                 cmm_smp_mb();
 345         }
 346         for (;;) {
 347                 struct cds_wfcq_head cbs_tmp_head;
 348                 struct cds_wfcq_tail cbs_tmp_tail;
 349                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 350                 enum cds_wfcq_ret splice_ret;
 351
 352                 if (set_thread_cpu_affinity(crdp))
 353                         urcu_die(errno);
 354
 355                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
 356                         /*
 357                          * Pause requested. Become quiescent: remove
 358                          * ourself from all global lists, and don't
 359                          * process any callback. The callback lists may
 360                          * still be non-empty though.
 361                          */
 362                         rcu_unregister_thread();
 363                         cmm_smp_mb__before_uatomic_or();
 364                         uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
 365                         while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
 366                                 (void) poll(NULL, 0, 1);
 367                         uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
 368                         cmm_smp_mb__after_uatomic_and();
 369                         rcu_register_thread();
 370                 }
 371
 372                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 373                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 374                         &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail);
 375                 assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 376                 assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 377                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 378                         synchronize_rcu();
 379                         cbcount = 0;
 380                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 381                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 382                                 struct rcu_head *rhp;
 383
 384                                 rhp = caa_container_of(cbs,
 385                                         struct rcu_head, next);
 386                                 rhp->func(rhp);
 387                                 cbcount++;
 388                         }
 389                         uatomic_sub(&crdp->qlen, cbcount);
 390                 }
 391                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
 392                         break;
 393                 rcu_thread_offline();
 394                 if (!rt) {
 395                         if (cds_wfcq_empty(&crdp->cbs_head,
 396                                         &crdp->cbs_tail)) {
 397                                 call_rcu_wait(crdp);
 398                                 (void) poll(NULL, 0, 10);
 399                                 uatomic_dec(&crdp->futex);
 400                                 /*
 401                                  * Decrement futex before reading
 402                                  * call_rcu list.
 403                                  */
 404                                 cmm_smp_mb();
 405                         } else {
 406                                 (void) poll(NULL, 0, 10);
 407                         }
 408                 } else {
 409                         (void) poll(NULL, 0, 10);
 410                 }
 411                 rcu_thread_online();
 412         }
 413         if (!rt) {
 414                 /*
 415                  * Read call_rcu list before write futex.
 416                  */
 417                 cmm_smp_mb();
 418                 uatomic_set(&crdp->futex, 0);
 419         }
 420         uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
 421         rcu_unregister_thread();
 422         return NULL;
 423 }
 424
 425 /*
 426  * Create both a call_rcu thread and the corresponding call_rcu_data
 427  * structure, linking the structure in as specified.  Caller must hold
 428  * call_rcu_mutex.
 429  */
 430
 431 static void call_rcu_data_init(struct call_rcu_data **crdpp,
 432                                unsigned long flags,
 433                                int cpu_affinity)
 434 {
 435         struct call_rcu_data *crdp;
 436         int ret;
 437
 438         crdp = malloc(sizeof(*crdp));
 439         if (crdp == NULL)
 440                 urcu_die(errno);
 441         memset(crdp, '\0', sizeof(*crdp));
 442         cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail);
 443         crdp->qlen = 0;
 444         crdp->futex = 0;
 445         crdp->flags = flags;
 446         cds_list_add(&crdp->list, &call_rcu_data_list);
 447         crdp->cpu_affinity = cpu_affinity;
 448         crdp->gp_count = 0;
 449         cmm_smp_mb();  /* Structure initialized before pointer is planted. */
 450         *crdpp = crdp;
 451         ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
 452         if (ret)
 453                 urcu_die(ret);
 454 }
 455
 456 /*
 457  * Return a pointer to the call_rcu_data structure for the specified
 458  * CPU, returning NULL if there is none.  We cannot automatically
 459  * created it because the platform we are running on might not define
 460  * urcu_sched_getcpu().
 461  *
 462  * The call to this function and use of the returned call_rcu_data
 463  * should be protected by RCU read-side lock.
 464  */
 465
 466 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
 467 {
 468         static int warned = 0;
 469         struct call_rcu_data **pcpu_crdp;
 470
 471         pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
 472         if (pcpu_crdp == NULL)
 473                 return NULL;
 474         if (!warned && cpus_array_len > 0 && (cpu < 0 || cpus_array_len <= cpu)) {
 475                 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
 476                 warned = 1;
 477         }
 478         if (cpu < 0 || cpus_array_len <= cpu)
 479                 return NULL;
 480         return rcu_dereference(pcpu_crdp[cpu]);
 481 }
 482
 483 /*
 484  * Return the tid corresponding to the call_rcu thread whose
 485  * call_rcu_data structure is specified.
 486  */
 487
 488 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
 489 {
 490         return crdp->tid;
 491 }
 492
 493 /*
 494  * Create a call_rcu_data structure (with thread) and return a pointer.
 495  */
 496
 497 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
 498                                                     int cpu_affinity)
 499 {
 500         struct call_rcu_data *crdp;
 501
 502         call_rcu_data_init(&crdp, flags, cpu_affinity);
 503         return crdp;
 504 }
 505
 506 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
 507                                            int cpu_affinity)
 508 {
 509         struct call_rcu_data *crdp;
 510
 511         call_rcu_lock(&call_rcu_mutex);
 512         crdp = __create_call_rcu_data(flags, cpu_affinity);
 513         call_rcu_unlock(&call_rcu_mutex);
 514         return crdp;
 515 }
 516
 517 /*
 518  * Set the specified CPU to use the specified call_rcu_data structure.
 519  *
 520  * Use NULL to remove a CPU's call_rcu_data structure, but it is
 521  * the caller's responsibility to dispose of the removed structure.
 522  * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
 523  * (prior to NULLing it out, of course).
 524  *
 525  * The caller must wait for a grace-period to pass between return from
 526  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 527  * previous call rcu data as argument.
 528  */
 529
 530 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
 531 {
 532         static int warned = 0;
 533
 534         call_rcu_lock(&call_rcu_mutex);
 535         alloc_cpu_call_rcu_data();
 536         if (cpu < 0 || cpus_array_len <= cpu) {
 537                 if (!warned) {
 538                         fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
 539                         warned = 1;
 540                 }
 541                 call_rcu_unlock(&call_rcu_mutex);
 542                 errno = EINVAL;
 543                 return -EINVAL;
 544         }
 545
 546         if (per_cpu_call_rcu_data == NULL) {
 547                 call_rcu_unlock(&call_rcu_mutex);
 548                 errno = ENOMEM;
 549                 return -ENOMEM;
 550         }
 551
 552         if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
 553                 call_rcu_unlock(&call_rcu_mutex);
 554                 errno = EEXIST;
 555                 return -EEXIST;
 556         }
 557
 558         rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
 559         call_rcu_unlock(&call_rcu_mutex);
 560         return 0;
 561 }
 562
 563 /*
 564  * Return a pointer to the default call_rcu_data structure, creating
 565  * one if need be.  Because we never free call_rcu_data structures,
 566  * we don't need to be in an RCU read-side critical section.
 567  */
 568
 569 struct call_rcu_data *get_default_call_rcu_data(void)
 570 {
 571         if (default_call_rcu_data != NULL)
 572                 return rcu_dereference(default_call_rcu_data);
 573         call_rcu_lock(&call_rcu_mutex);
 574         if (default_call_rcu_data != NULL) {
 575                 call_rcu_unlock(&call_rcu_mutex);
 576                 return default_call_rcu_data;
 577         }
 578         call_rcu_data_init(&default_call_rcu_data, 0, -1);
 579         call_rcu_unlock(&call_rcu_mutex);
 580         return default_call_rcu_data;
 581 }
 582
 583 /*
 584  * Return the call_rcu_data structure that applies to the currently
 585  * running thread.  Any call_rcu_data structure assigned specifically
 586  * to this thread has first priority, followed by any call_rcu_data
 587  * structure assigned to the CPU on which the thread is running,
 588  * followed by the default call_rcu_data structure.  If there is not
 589  * yet a default call_rcu_data structure, one will be created.
 590  *
 591  * Calls to this function and use of the returned call_rcu_data should
 592  * be protected by RCU read-side lock.
 593  */
 594 struct call_rcu_data *get_call_rcu_data(void)
 595 {
 596         struct call_rcu_data *crd;
 597
 598         if (URCU_TLS(thread_call_rcu_data) != NULL)
 599                 return URCU_TLS(thread_call_rcu_data);
 600
 601         if (cpus_array_len > 0) {
 602                 crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
 603                 if (crd)
 604                         return crd;
 605         }
 606
 607         return get_default_call_rcu_data();
 608 }
 609
 610 /*
 611  * Return a pointer to this task's call_rcu_data if there is one.
 612  */
 613
 614 struct call_rcu_data *get_thread_call_rcu_data(void)
 615 {
 616         return URCU_TLS(thread_call_rcu_data);
 617 }
 618
 619 /*
 620  * Set this task's call_rcu_data structure as specified, regardless
 621  * of whether or not this task already had one.  (This allows switching
 622  * to and from real-time call_rcu threads, for example.)
 623  *
 624  * Use NULL to remove a thread's call_rcu_data structure, but it is
 625  * the caller's responsibility to dispose of the removed structure.
 626  * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
 627  * (prior to NULLing it out, of course).
 628  */
 629
 630 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
 631 {
 632         URCU_TLS(thread_call_rcu_data) = crdp;
 633 }
 634
 635 /*
 636  * Create a separate call_rcu thread for each CPU.  This does not
 637  * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
 638  * function if you want that behavior. Should be paired with
 639  * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
 640  * threads.
 641  */
 642
 643 int create_all_cpu_call_rcu_data(unsigned long flags)
 644 {
 645         int i;
 646         struct call_rcu_data *crdp;
 647         int ret;
 648
 649         call_rcu_lock(&call_rcu_mutex);
 650         alloc_cpu_call_rcu_data();
 651         call_rcu_unlock(&call_rcu_mutex);
 652         if (cpus_array_len <= 0) {
 653                 errno = EINVAL;
 654                 return -EINVAL;
 655         }
 656         if (per_cpu_call_rcu_data == NULL) {
 657                 errno = ENOMEM;
 658                 return -ENOMEM;
 659         }
 660         for (i = 0; i < cpus_array_len; i++) {
 661                 call_rcu_lock(&call_rcu_mutex);
 662                 if (get_cpu_call_rcu_data(i)) {
 663                         call_rcu_unlock(&call_rcu_mutex);
 664                         continue;
 665                 }
 666                 crdp = __create_call_rcu_data(flags, i);
 667                 if (crdp == NULL) {
 668                         call_rcu_unlock(&call_rcu_mutex);
 669                         errno = ENOMEM;
 670                         return -ENOMEM;
 671                 }
 672                 call_rcu_unlock(&call_rcu_mutex);
 673                 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
 674                         call_rcu_data_free(crdp);
 675
 676                         /* it has been created by other thread */
 677                         if (ret == -EEXIST)
 678                                 continue;
 679
 680                         return ret;
 681                 }
 682         }
 683         return 0;
 684 }
 685
 686 /*
 687  * Wake up the call_rcu thread corresponding to the specified
 688  * call_rcu_data structure.
 689  */
 690 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
 691 {
 692         if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
 693                 call_rcu_wake_up(crdp);
 694 }
 695
 696 static void _call_rcu(struct rcu_head *head,
 697                       void (*func)(struct rcu_head *head),
 698                       struct call_rcu_data *crdp)
 699 {
 700         cds_wfcq_node_init(&head->next);
 701         head->func = func;
 702         cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
 703         uatomic_inc(&crdp->qlen);
 704         wake_call_rcu_thread(crdp);
 705 }
 706
 707 /*
 708  * Schedule a function to be invoked after a following grace period.
 709  * This is the only function that must be called -- the others are
 710  * only present to allow applications to tune their use of RCU for
 711  * maximum performance.
 712  *
 713  * Note that unless a call_rcu thread has not already been created,
 714  * the first invocation of call_rcu() will create one.  So, if you
 715  * need the first invocation of call_rcu() to be fast, make sure
 716  * to create a call_rcu thread first.  One way to accomplish this is
 717  * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
 718  *
 719  * call_rcu must be called by registered RCU read-side threads.
 720  */
 721 void call_rcu(struct rcu_head *head,
 722               void (*func)(struct rcu_head *head))
 723 {
 724         struct call_rcu_data *crdp;
 725
 726         /* Holding rcu read-side lock across use of per-cpu crdp */
 727         _rcu_read_lock();
 728         crdp = get_call_rcu_data();
 729         _call_rcu(head, func, crdp);
 730         _rcu_read_unlock();
 731 }
 732
 733 /*
 734  * Free up the specified call_rcu_data structure, terminating the
 735  * associated call_rcu thread.  The caller must have previously
 736  * removed the call_rcu_data structure from per-thread or per-CPU
 737  * usage.  For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
 738  * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
 739  * per-thread call_rcu_data structures.
 740  *
 741  * We silently refuse to free up the default call_rcu_data structure
 742  * because that is where we put any leftover callbacks.  Note that
 743  * the possibility of self-spawning callbacks makes it impossible
 744  * to execute all the callbacks in finite time without putting any
 745  * newly spawned callbacks somewhere else.  The "somewhere else" of
 746  * last resort is the default call_rcu_data structure.
 747  *
 748  * We also silently refuse to free NULL pointers.  This simplifies
 749  * the calling code.
 750  *
 751  * The caller must wait for a grace-period to pass between return from
 752  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 753  * previous call rcu data as argument.
 754  *
 755  * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
 756  * a list corruption bug in the 0.7.x series. The equivalent fix
 757  * appeared in 0.6.8 for the stable-0.6 branch.
 758  */
 759 void call_rcu_data_free(struct call_rcu_data *crdp)
 760 {
 761         if (crdp == NULL || crdp == default_call_rcu_data) {
 762                 return;
 763         }
 764         if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
 765                 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
 766                 wake_call_rcu_thread(crdp);
 767                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
 768                         (void) poll(NULL, 0, 1);
 769         }
 770         call_rcu_lock(&call_rcu_mutex);
 771         if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
 772                 call_rcu_unlock(&call_rcu_mutex);
 773                 /* Create default call rcu data if need be. */
 774                 /* CBs queued here will be handed to the default list. */
 775                 (void) get_default_call_rcu_data();
 776                 call_rcu_lock(&call_rcu_mutex);
 777                 __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
 778                         &default_call_rcu_data->cbs_tail,
 779                         &crdp->cbs_head, &crdp->cbs_tail);
 780                 uatomic_add(&default_call_rcu_data->qlen,
 781                             uatomic_read(&crdp->qlen));
 782                 wake_call_rcu_thread(default_call_rcu_data);
 783         }
 784
 785         cds_list_del(&crdp->list);
 786         call_rcu_unlock(&call_rcu_mutex);
 787
 788         free(crdp);
 789 }
 790
 791 /*
 792  * Clean up all the per-CPU call_rcu threads.
 793  */
 794 void free_all_cpu_call_rcu_data(void)
 795 {
 796         int cpu;
 797         struct call_rcu_data **crdp;
 798         static int warned = 0;
 799
 800         if (cpus_array_len <= 0)
 801                 return;
 802
 803         crdp = malloc(sizeof(*crdp) * cpus_array_len);
 804         if (!crdp) {
 805                 if (!warned) {
 806                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 807                 }
 808                 warned = 1;
 809                 return;
 810         }
 811
 812         for (cpu = 0; cpu < cpus_array_len; cpu++) {
 813                 crdp[cpu] = get_cpu_call_rcu_data(cpu);
 814                 if (crdp[cpu] == NULL)
 815                         continue;
 816                 set_cpu_call_rcu_data(cpu, NULL);
 817         }
 818         /*
 819          * Wait for call_rcu sites acting as RCU readers of the
 820          * call_rcu_data to become quiescent.
 821          */
 822         synchronize_rcu();
 823         for (cpu = 0; cpu < cpus_array_len; cpu++) {
 824                 if (crdp[cpu] == NULL)
 825                         continue;
 826                 call_rcu_data_free(crdp[cpu]);
 827         }
 828         free(crdp);
 829 }
 830
 831 static
 832 void free_completion(struct urcu_ref *ref)
 833 {
 834         struct call_rcu_completion *completion;
 835
 836         completion = caa_container_of(ref, struct call_rcu_completion, ref);
 837         free(completion);
 838 }
 839
 840 static
 841 void _rcu_barrier_complete(struct rcu_head *head)
 842 {
 843         struct call_rcu_completion_work *work;
 844         struct call_rcu_completion *completion;
 845
 846         work = caa_container_of(head, struct call_rcu_completion_work, head);
 847         completion = work->completion;
 848         if (!uatomic_sub_return(&completion->barrier_count, 1))
 849                 call_rcu_completion_wake_up(completion);
 850         urcu_ref_put(&completion->ref, free_completion);
 851         free(work);
 852 }
 853
 854 /*
 855  * Wait for all in-flight call_rcu callbacks to complete execution.
 856  */
 857 void rcu_barrier(void)
 858 {
 859         struct call_rcu_data *crdp;
 860         struct call_rcu_completion *completion;
 861         int count = 0;
 862         int was_online;
 863
 864         /* Put in offline state in QSBR. */
 865         was_online = _rcu_read_ongoing();
 866         if (was_online)
 867                 rcu_thread_offline();
 868         /*
 869          * Calling a rcu_barrier() within a RCU read-side critical
 870          * section is an error.
 871          */
 872         if (_rcu_read_ongoing()) {
 873                 static int warned = 0;
 874
 875                 if (!warned) {
 876                         fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
 877                 }
 878                 warned = 1;
 879                 goto online;
 880         }
 881
 882         completion = calloc(sizeof(*completion), 1);
 883         if (!completion)
 884                 urcu_die(errno);
 885
 886         call_rcu_lock(&call_rcu_mutex);
 887         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 888                 count++;
 889
 890         /* Referenced by rcu_barrier() and each call_rcu thread. */
 891         urcu_ref_set(&completion->ref, count + 1);
 892         completion->barrier_count = count;
 893
 894         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 895                 struct call_rcu_completion_work *work;
 896
 897                 work = calloc(sizeof(*work), 1);
 898                 if (!work)
 899                         urcu_die(errno);
 900                 work->completion = completion;
 901                 _call_rcu(&work->head, _rcu_barrier_complete, crdp);
 902         }
 903         call_rcu_unlock(&call_rcu_mutex);
 904
 905         /* Wait for them */
 906         for (;;) {
 907                 uatomic_dec(&completion->futex);
 908                 /* Decrement futex before reading barrier_count */
 909                 cmm_smp_mb();
 910                 if (!uatomic_read(&completion->barrier_count))
 911                         break;
 912                 call_rcu_completion_wait(completion);
 913         }
 914
 915         urcu_ref_put(&completion->ref, free_completion);
 916
 917 online:
 918         if (was_online)
 919                 rcu_thread_online();
 920 }
 921
 922 /*
 923  * Acquire the call_rcu_mutex in order to ensure that the child sees
 924  * all of the call_rcu() data structures in a consistent state. Ensure
 925  * that all call_rcu threads are in a quiescent state across fork.
 926  * Suitable for pthread_atfork() and friends.
 927  */
 928 void call_rcu_before_fork(void)
 929 {
 930         struct call_rcu_data *crdp;
 931         struct urcu_atfork *atfork;
 932
 933         call_rcu_lock(&call_rcu_mutex);
 934
 935         atfork = registered_rculfhash_atfork;
 936         if (atfork)
 937                 atfork->before_fork(atfork->priv);
 938
 939         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 940                 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
 941                 cmm_smp_mb__after_uatomic_or();
 942                 wake_call_rcu_thread(crdp);
 943         }
 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 }
 949
 950 /*
 951  * Clean up call_rcu data structures in the parent of a successful fork()
 952  * that is not followed by exec() in the child.  Suitable for
 953  * pthread_atfork() and friends.
 954  */
 955 void call_rcu_after_fork_parent(void)
 956 {
 957         struct call_rcu_data *crdp;
 958         struct urcu_atfork *atfork;
 959
 960         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 961                 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
 962         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 963                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
 964                         (void) poll(NULL, 0, 1);
 965         }
 966         atfork = registered_rculfhash_atfork;
 967         if (atfork)
 968                 atfork->after_fork_parent(atfork->priv);
 969         call_rcu_unlock(&call_rcu_mutex);
 970 }
 971
 972 /*
 973  * Clean up call_rcu data structures in the child of a successful fork()
 974  * that is not followed by exec().  Suitable for pthread_atfork() and
 975  * friends.
 976  */
 977 void call_rcu_after_fork_child(void)
 978 {
 979         struct call_rcu_data *crdp, *next;
 980         struct urcu_atfork *atfork;
 981
 982         /* Release the mutex. */
 983         call_rcu_unlock(&call_rcu_mutex);
 984
 985         atfork = registered_rculfhash_atfork;
 986         if (atfork)
 987                 atfork->after_fork_child(atfork->priv);
 988
 989         /* Do nothing when call_rcu() has not been used */
 990         if (cds_list_empty(&call_rcu_data_list))
 991                 return;
 992
 993         /*
 994          * Allocate a new default call_rcu_data structure in order
 995          * to get a working call_rcu thread to go with it.
 996          */
 997         default_call_rcu_data = NULL;
 998         (void)get_default_call_rcu_data();
 999
1000         /* Cleanup call_rcu_data pointers before use */
1001         cpus_array_len_reset();
1002         free(per_cpu_call_rcu_data);
1003         rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
1004         URCU_TLS(thread_call_rcu_data) = NULL;
1005
1006         /*
1007          * Dispose of all of the rest of the call_rcu_data structures.
1008          * Leftover call_rcu callbacks will be merged into the new
1009          * default call_rcu thread queue.
1010          */
1011         cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
1012                 if (crdp == default_call_rcu_data)
1013                         continue;
1014                 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
1015                 call_rcu_data_free(crdp);
1016         }
1017 }
1018
1019 void urcu_register_rculfhash_atfork(struct urcu_atfork *atfork)
1020 {
1021         call_rcu_lock(&call_rcu_mutex);
1022         if (registered_rculfhash_atfork_refcount++)
1023                 goto end;
1024         registered_rculfhash_atfork = atfork;
1025 end:
1026         call_rcu_unlock(&call_rcu_mutex);
1027 }
1028
1029 void urcu_unregister_rculfhash_atfork(struct urcu_atfork *atfork __attribute__((unused)))
1030 {
1031         call_rcu_lock(&call_rcu_mutex);
1032         if (--registered_rculfhash_atfork_refcount)
1033                 goto end;
1034         registered_rculfhash_atfork = NULL;
1035 end:
1036         call_rcu_unlock(&call_rcu_mutex);
1037 }