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 #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
 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 cpus_array_len;
 124
 125 static void cpus_array_len_reset(void)
 126 {
 127         cpus_array_len = 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 (cpus_array_len != 0)
 138                 return;
 139         cpus_array_len = get_possible_cpus_array_len();
 140         if (cpus_array_len <= 0) {
 141                 return;
 142         }
 143         p = malloc(cpus_array_len * sizeof(*per_cpu_call_rcu_data));
 144         if (p != NULL) {
 145                 memset(p, '\0', cpus_array_len * 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 cpus_array_len = -1;
 164
 165 static void cpus_array_len_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         while (uatomic_read(&crdp->futex) == -1) {
 244                 if (!futex_async(&crdp->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 245                         /*
 246                          * Prior queued wakeups queued by unrelated code
 247                          * using the same address can cause futex wait to
 248                          * return 0 even through the futex value is still
 249                          * -1 (spurious wakeups). Check the value again
 250                          * in user-space to validate whether it really
 251                          * differs from -1.
 252                          */
 253                         continue;
 254                 }
 255                 switch (errno) {
 256                 case EAGAIN:
 257                         /* Value already changed. */
 258                         return;
 259                 case EINTR:
 260                         /* Retry if interrupted by signal. */
 261                         break;  /* Get out of switch. Check again. */
 262                 default:
 263                         /* Unexpected error. */
 264                         urcu_die(errno);
 265                 }
 266         }
 267 }
 268
 269 static void call_rcu_wake_up(struct call_rcu_data *crdp)
 270 {
 271         /* Write to call_rcu list before reading/writing futex */
 272         cmm_smp_mb();
 273         if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
 274                 uatomic_set(&crdp->futex, 0);
 275                 if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
 276                                 NULL, NULL, 0) < 0)
 277                         urcu_die(errno);
 278         }
 279 }
 280
 281 static void call_rcu_completion_wait(struct call_rcu_completion *completion)
 282 {
 283         /* Read completion barrier count before read futex */
 284         cmm_smp_mb();
 285         while (uatomic_read(&completion->futex) == -1) {
 286                 if (!futex_async(&completion->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 287                         /*
 288                          * Prior queued wakeups queued by unrelated code
 289                          * using the same address can cause futex wait to
 290                          * return 0 even through the futex value is still
 291                          * -1 (spurious wakeups). Check the value again
 292                          * in user-space to validate whether it really
 293                          * differs from -1.
 294                          */
 295                         continue;
 296                 }
 297                 switch (errno) {
 298                 case EAGAIN:
 299                         /* Value already changed. */
 300                         return;
 301                 case EINTR:
 302                         /* Retry if interrupted by signal. */
 303                         break;  /* Get out of switch. Check again. */
 304                 default:
 305                         /* Unexpected error. */
 306                         urcu_die(errno);
 307                 }
 308         }
 309 }
 310
 311 static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
 312 {
 313         /* Write to completion barrier count before reading/writing futex */
 314         cmm_smp_mb();
 315         if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
 316                 uatomic_set(&completion->futex, 0);
 317                 if (futex_async(&completion->futex, FUTEX_WAKE, 1,
 318                                 NULL, NULL, 0) < 0)
 319                         urcu_die(errno);
 320         }
 321 }
 322
 323 /* This is the code run by each call_rcu thread. */
 324
 325 static void *call_rcu_thread(void *arg)
 326 {
 327         unsigned long cbcount;
 328         struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
 329         int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
 330
 331         if (set_thread_cpu_affinity(crdp))
 332                 urcu_die(errno);
 333
 334         /*
 335          * If callbacks take a read-side lock, we need to be registered.
 336          */
 337         rcu_register_thread();
 338
 339         URCU_TLS(thread_call_rcu_data) = crdp;
 340         if (!rt) {
 341                 uatomic_dec(&crdp->futex);
 342                 /* Decrement futex before reading call_rcu list */
 343                 cmm_smp_mb();
 344         }
 345         for (;;) {
 346                 struct cds_wfcq_head cbs_tmp_head;
 347                 struct cds_wfcq_tail cbs_tmp_tail;
 348                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 349                 enum cds_wfcq_ret splice_ret;
 350
 351                 if (set_thread_cpu_affinity(crdp))
 352                         urcu_die(errno);
 353
 354                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
 355                         /*
 356                          * Pause requested. Become quiescent: remove
 357                          * ourself from all global lists, and don't
 358                          * process any callback. The callback lists may
 359                          * still be non-empty though.
 360                          */
 361                         rcu_unregister_thread();
 362                         cmm_smp_mb__before_uatomic_or();
 363                         uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
 364                         while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
 365                                 (void) poll(NULL, 0, 1);
 366                         uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
 367                         cmm_smp_mb__after_uatomic_and();
 368                         rcu_register_thread();
 369                 }
 370
 371                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 372                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 373                         &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail);
 374                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 375                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 376                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 377                         synchronize_rcu();
 378                         cbcount = 0;
 379                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 380                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 381                                 struct rcu_head *rhp;
 382
 383                                 rhp = caa_container_of(cbs,
 384                                         struct rcu_head, next);
 385                                 rhp->func(rhp);
 386                                 cbcount++;
 387                         }
 388                         uatomic_sub(&crdp->qlen, cbcount);
 389                 }
 390                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
 391                         break;
 392                 rcu_thread_offline();
 393                 if (!rt) {
 394                         if (cds_wfcq_empty(&crdp->cbs_head,
 395                                         &crdp->cbs_tail)) {
 396                                 call_rcu_wait(crdp);
 397                                 (void) poll(NULL, 0, 10);
 398                                 uatomic_dec(&crdp->futex);
 399                                 /*
 400                                  * Decrement futex before reading
 401                                  * call_rcu list.
 402                                  */
 403                                 cmm_smp_mb();
 404                         } else {
 405                                 (void) poll(NULL, 0, 10);
 406                         }
 407                 } else {
 408                         (void) poll(NULL, 0, 10);
 409                 }
 410                 rcu_thread_online();
 411         }
 412         if (!rt) {
 413                 /*
 414                  * Read call_rcu list before write futex.
 415                  */
 416                 cmm_smp_mb();
 417                 uatomic_set(&crdp->futex, 0);
 418         }
 419         uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
 420         rcu_unregister_thread();
 421         return NULL;
 422 }
 423
 424 /*
 425  * Create both a call_rcu thread and the corresponding call_rcu_data
 426  * structure, linking the structure in as specified.  Caller must hold
 427  * call_rcu_mutex.
 428  */
 429
 430 static void call_rcu_data_init(struct call_rcu_data **crdpp,
 431                                unsigned long flags,
 432                                int cpu_affinity)
 433 {
 434         struct call_rcu_data *crdp;
 435         int ret;
 436         sigset_t newmask, oldmask;
 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
 452         ret = sigfillset(&newmask);
 453         urcu_posix_assert(!ret);
 454         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 455         urcu_posix_assert(!ret);
 456
 457         ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
 458         if (ret)
 459                 urcu_die(ret);
 460
 461         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 462         urcu_posix_assert(!ret);
 463 }
 464
 465 /*
 466  * Return a pointer to the call_rcu_data structure for the specified
 467  * CPU, returning NULL if there is none.  We cannot automatically
 468  * created it because the platform we are running on might not define
 469  * urcu_sched_getcpu().
 470  *
 471  * The call to this function and use of the returned call_rcu_data
 472  * should be protected by RCU read-side lock.
 473  */
 474
 475 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
 476 {
 477         static int warned = 0;
 478         struct call_rcu_data **pcpu_crdp;
 479
 480         pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
 481         if (pcpu_crdp == NULL)
 482                 return NULL;
 483         if (!warned && cpus_array_len > 0 && (cpu < 0 || cpus_array_len <= cpu)) {
 484                 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
 485                 warned = 1;
 486         }
 487         if (cpu < 0 || cpus_array_len <= cpu)
 488                 return NULL;
 489         return rcu_dereference(pcpu_crdp[cpu]);
 490 }
 491
 492 /*
 493  * Return the tid corresponding to the call_rcu thread whose
 494  * call_rcu_data structure is specified.
 495  */
 496
 497 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
 498 {
 499         return crdp->tid;
 500 }
 501
 502 /*
 503  * Create a call_rcu_data structure (with thread) and return a pointer.
 504  */
 505
 506 static 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_data_init(&crdp, flags, cpu_affinity);
 512         return crdp;
 513 }
 514
 515 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
 516                                            int cpu_affinity)
 517 {
 518         struct call_rcu_data *crdp;
 519
 520         call_rcu_lock(&call_rcu_mutex);
 521         crdp = __create_call_rcu_data(flags, cpu_affinity);
 522         call_rcu_unlock(&call_rcu_mutex);
 523         return crdp;
 524 }
 525
 526 /*
 527  * Set the specified CPU to use the specified call_rcu_data structure.
 528  *
 529  * Use NULL to remove a CPU's call_rcu_data structure, but it is
 530  * the caller's responsibility to dispose of the removed structure.
 531  * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
 532  * (prior to NULLing it out, of course).
 533  *
 534  * The caller must wait for a grace-period to pass between return from
 535  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 536  * previous call rcu data as argument.
 537  */
 538
 539 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
 540 {
 541         static int warned = 0;
 542
 543         call_rcu_lock(&call_rcu_mutex);
 544         alloc_cpu_call_rcu_data();
 545         if (cpu < 0 || cpus_array_len <= cpu) {
 546                 if (!warned) {
 547                         fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
 548                         warned = 1;
 549                 }
 550                 call_rcu_unlock(&call_rcu_mutex);
 551                 errno = EINVAL;
 552                 return -EINVAL;
 553         }
 554
 555         if (per_cpu_call_rcu_data == NULL) {
 556                 call_rcu_unlock(&call_rcu_mutex);
 557                 errno = ENOMEM;
 558                 return -ENOMEM;
 559         }
 560
 561         if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
 562                 call_rcu_unlock(&call_rcu_mutex);
 563                 errno = EEXIST;
 564                 return -EEXIST;
 565         }
 566
 567         rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
 568         call_rcu_unlock(&call_rcu_mutex);
 569         return 0;
 570 }
 571
 572 /*
 573  * Return a pointer to the default call_rcu_data structure, creating
 574  * one if need be.  Because we never free call_rcu_data structures,
 575  * we don't need to be in an RCU read-side critical section.
 576  */
 577
 578 struct call_rcu_data *get_default_call_rcu_data(void)
 579 {
 580         if (default_call_rcu_data != NULL)
 581                 return rcu_dereference(default_call_rcu_data);
 582         call_rcu_lock(&call_rcu_mutex);
 583         if (default_call_rcu_data != NULL) {
 584                 call_rcu_unlock(&call_rcu_mutex);
 585                 return default_call_rcu_data;
 586         }
 587         call_rcu_data_init(&default_call_rcu_data, 0, -1);
 588         call_rcu_unlock(&call_rcu_mutex);
 589         return default_call_rcu_data;
 590 }
 591
 592 /*
 593  * Return the call_rcu_data structure that applies to the currently
 594  * running thread.  Any call_rcu_data structure assigned specifically
 595  * to this thread has first priority, followed by any call_rcu_data
 596  * structure assigned to the CPU on which the thread is running,
 597  * followed by the default call_rcu_data structure.  If there is not
 598  * yet a default call_rcu_data structure, one will be created.
 599  *
 600  * Calls to this function and use of the returned call_rcu_data should
 601  * be protected by RCU read-side lock.
 602  */
 603 struct call_rcu_data *get_call_rcu_data(void)
 604 {
 605         struct call_rcu_data *crd;
 606
 607         if (URCU_TLS(thread_call_rcu_data) != NULL)
 608                 return URCU_TLS(thread_call_rcu_data);
 609
 610         if (cpus_array_len > 0) {
 611                 crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
 612                 if (crd)
 613                         return crd;
 614         }
 615
 616         return get_default_call_rcu_data();
 617 }
 618
 619 /*
 620  * Return a pointer to this task's call_rcu_data if there is one.
 621  */
 622
 623 struct call_rcu_data *get_thread_call_rcu_data(void)
 624 {
 625         return URCU_TLS(thread_call_rcu_data);
 626 }
 627
 628 /*
 629  * Set this task's call_rcu_data structure as specified, regardless
 630  * of whether or not this task already had one.  (This allows switching
 631  * to and from real-time call_rcu threads, for example.)
 632  *
 633  * Use NULL to remove a thread's call_rcu_data structure, but it is
 634  * the caller's responsibility to dispose of the removed structure.
 635  * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
 636  * (prior to NULLing it out, of course).
 637  */
 638
 639 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
 640 {
 641         URCU_TLS(thread_call_rcu_data) = crdp;
 642 }
 643
 644 /*
 645  * Create a separate call_rcu thread for each CPU.  This does not
 646  * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
 647  * function if you want that behavior. Should be paired with
 648  * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
 649  * threads.
 650  */
 651
 652 int create_all_cpu_call_rcu_data(unsigned long flags)
 653 {
 654         int i;
 655         struct call_rcu_data *crdp;
 656         int ret;
 657
 658         call_rcu_lock(&call_rcu_mutex);
 659         alloc_cpu_call_rcu_data();
 660         call_rcu_unlock(&call_rcu_mutex);
 661         if (cpus_array_len <= 0) {
 662                 errno = EINVAL;
 663                 return -EINVAL;
 664         }
 665         if (per_cpu_call_rcu_data == NULL) {
 666                 errno = ENOMEM;
 667                 return -ENOMEM;
 668         }
 669         for (i = 0; i < cpus_array_len; i++) {
 670                 call_rcu_lock(&call_rcu_mutex);
 671                 if (get_cpu_call_rcu_data(i)) {
 672                         call_rcu_unlock(&call_rcu_mutex);
 673                         continue;
 674                 }
 675                 crdp = __create_call_rcu_data(flags, i);
 676                 if (crdp == NULL) {
 677                         call_rcu_unlock(&call_rcu_mutex);
 678                         errno = ENOMEM;
 679                         return -ENOMEM;
 680                 }
 681                 call_rcu_unlock(&call_rcu_mutex);
 682                 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
 683                         call_rcu_data_free(crdp);
 684
 685                         /* it has been created by other thread */
 686                         if (ret == -EEXIST)
 687                                 continue;
 688
 689                         return ret;
 690                 }
 691         }
 692         return 0;
 693 }
 694
 695 /*
 696  * Wake up the call_rcu thread corresponding to the specified
 697  * call_rcu_data structure.
 698  */
 699 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
 700 {
 701         if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
 702                 call_rcu_wake_up(crdp);
 703 }
 704
 705 static void _call_rcu(struct rcu_head *head,
 706                       void (*func)(struct rcu_head *head),
 707                       struct call_rcu_data *crdp)
 708 {
 709         cds_wfcq_node_init(&head->next);
 710         head->func = func;
 711         cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
 712         uatomic_inc(&crdp->qlen);
 713         wake_call_rcu_thread(crdp);
 714 }
 715
 716 /*
 717  * Schedule a function to be invoked after a following grace period.
 718  * This is the only function that must be called -- the others are
 719  * only present to allow applications to tune their use of RCU for
 720  * maximum performance.
 721  *
 722  * Note that unless a call_rcu thread has not already been created,
 723  * the first invocation of call_rcu() will create one.  So, if you
 724  * need the first invocation of call_rcu() to be fast, make sure
 725  * to create a call_rcu thread first.  One way to accomplish this is
 726  * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
 727  *
 728  * call_rcu must be called by registered RCU read-side threads.
 729  */
 730 void call_rcu(struct rcu_head *head,
 731               void (*func)(struct rcu_head *head))
 732 {
 733         struct call_rcu_data *crdp;
 734
 735         /* Holding rcu read-side lock across use of per-cpu crdp */
 736         _rcu_read_lock();
 737         crdp = get_call_rcu_data();
 738         _call_rcu(head, func, crdp);
 739         _rcu_read_unlock();
 740 }
 741
 742 /*
 743  * Free up the specified call_rcu_data structure, terminating the
 744  * associated call_rcu thread.  The caller must have previously
 745  * removed the call_rcu_data structure from per-thread or per-CPU
 746  * usage.  For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
 747  * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
 748  * per-thread call_rcu_data structures.
 749  *
 750  * We silently refuse to free up the default call_rcu_data structure
 751  * because that is where we put any leftover callbacks.  Note that
 752  * the possibility of self-spawning callbacks makes it impossible
 753  * to execute all the callbacks in finite time without putting any
 754  * newly spawned callbacks somewhere else.  The "somewhere else" of
 755  * last resort is the default call_rcu_data structure.
 756  *
 757  * We also silently refuse to free NULL pointers.  This simplifies
 758  * the calling code.
 759  *
 760  * The caller must wait for a grace-period to pass between return from
 761  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 762  * previous call rcu data as argument.
 763  *
 764  * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
 765  * a list corruption bug in the 0.7.x series. The equivalent fix
 766  * appeared in 0.6.8 for the stable-0.6 branch.
 767  */
 768 void call_rcu_data_free(struct call_rcu_data *crdp)
 769 {
 770         if (crdp == NULL || crdp == default_call_rcu_data) {
 771                 return;
 772         }
 773         if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
 774                 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
 775                 wake_call_rcu_thread(crdp);
 776                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
 777                         (void) poll(NULL, 0, 1);
 778         }
 779         call_rcu_lock(&call_rcu_mutex);
 780         if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
 781                 call_rcu_unlock(&call_rcu_mutex);
 782                 /* Create default call rcu data if need be. */
 783                 /* CBs queued here will be handed to the default list. */
 784                 (void) get_default_call_rcu_data();
 785                 call_rcu_lock(&call_rcu_mutex);
 786                 __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
 787                         &default_call_rcu_data->cbs_tail,
 788                         &crdp->cbs_head, &crdp->cbs_tail);
 789                 uatomic_add(&default_call_rcu_data->qlen,
 790                             uatomic_read(&crdp->qlen));
 791                 wake_call_rcu_thread(default_call_rcu_data);
 792         }
 793
 794         cds_list_del(&crdp->list);
 795         call_rcu_unlock(&call_rcu_mutex);
 796
 797         free(crdp);
 798 }
 799
 800 /*
 801  * Clean up all the per-CPU call_rcu threads.
 802  */
 803 void free_all_cpu_call_rcu_data(void)
 804 {
 805         int cpu;
 806         struct call_rcu_data **crdp;
 807         static int warned = 0;
 808
 809         if (cpus_array_len <= 0)
 810                 return;
 811
 812         crdp = malloc(sizeof(*crdp) * cpus_array_len);
 813         if (!crdp) {
 814                 if (!warned) {
 815                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 816                 }
 817                 warned = 1;
 818                 return;
 819         }
 820
 821         for (cpu = 0; cpu < cpus_array_len; cpu++) {
 822                 crdp[cpu] = get_cpu_call_rcu_data(cpu);
 823                 if (crdp[cpu] == NULL)
 824                         continue;
 825                 set_cpu_call_rcu_data(cpu, NULL);
 826         }
 827         /*
 828          * Wait for call_rcu sites acting as RCU readers of the
 829          * call_rcu_data to become quiescent.
 830          */
 831         synchronize_rcu();
 832         for (cpu = 0; cpu < cpus_array_len; cpu++) {
 833                 if (crdp[cpu] == NULL)
 834                         continue;
 835                 call_rcu_data_free(crdp[cpu]);
 836         }
 837         free(crdp);
 838 }
 839
 840 static
 841 void free_completion(struct urcu_ref *ref)
 842 {
 843         struct call_rcu_completion *completion;
 844
 845         completion = caa_container_of(ref, struct call_rcu_completion, ref);
 846         free(completion);
 847 }
 848
 849 static
 850 void _rcu_barrier_complete(struct rcu_head *head)
 851 {
 852         struct call_rcu_completion_work *work;
 853         struct call_rcu_completion *completion;
 854
 855         work = caa_container_of(head, struct call_rcu_completion_work, head);
 856         completion = work->completion;
 857         if (!uatomic_sub_return(&completion->barrier_count, 1))
 858                 call_rcu_completion_wake_up(completion);
 859         urcu_ref_put(&completion->ref, free_completion);
 860         free(work);
 861 }
 862
 863 /*
 864  * Wait for all in-flight call_rcu callbacks to complete execution.
 865  */
 866 void rcu_barrier(void)
 867 {
 868         struct call_rcu_data *crdp;
 869         struct call_rcu_completion *completion;
 870         int count = 0;
 871         int was_online;
 872
 873         /* Put in offline state in QSBR. */
 874         was_online = _rcu_read_ongoing();
 875         if (was_online)
 876                 rcu_thread_offline();
 877         /*
 878          * Calling a rcu_barrier() within a RCU read-side critical
 879          * section is an error.
 880          */
 881         if (_rcu_read_ongoing()) {
 882                 static int warned = 0;
 883
 884                 if (!warned) {
 885                         fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
 886                 }
 887                 warned = 1;
 888                 goto online;
 889         }
 890
 891         completion = calloc(sizeof(*completion), 1);
 892         if (!completion)
 893                 urcu_die(errno);
 894
 895         call_rcu_lock(&call_rcu_mutex);
 896         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 897                 count++;
 898
 899         /* Referenced by rcu_barrier() and each call_rcu thread. */
 900         urcu_ref_set(&completion->ref, count + 1);
 901         completion->barrier_count = count;
 902
 903         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 904                 struct call_rcu_completion_work *work;
 905
 906                 work = calloc(sizeof(*work), 1);
 907                 if (!work)
 908                         urcu_die(errno);
 909                 work->completion = completion;
 910                 _call_rcu(&work->head, _rcu_barrier_complete, crdp);
 911         }
 912         call_rcu_unlock(&call_rcu_mutex);
 913
 914         /* Wait for them */
 915         for (;;) {
 916                 uatomic_dec(&completion->futex);
 917                 /* Decrement futex before reading barrier_count */
 918                 cmm_smp_mb();
 919                 if (!uatomic_read(&completion->barrier_count))
 920                         break;
 921                 call_rcu_completion_wait(completion);
 922         }
 923
 924         urcu_ref_put(&completion->ref, free_completion);
 925
 926 online:
 927         if (was_online)
 928                 rcu_thread_online();
 929 }
 930
 931 /*
 932  * Acquire the call_rcu_mutex in order to ensure that the child sees
 933  * all of the call_rcu() data structures in a consistent state. Ensure
 934  * that all call_rcu threads are in a quiescent state across fork.
 935  * Suitable for pthread_atfork() and friends.
 936  */
 937 void call_rcu_before_fork(void)
 938 {
 939         struct call_rcu_data *crdp;
 940         struct urcu_atfork *atfork;
 941
 942         call_rcu_lock(&call_rcu_mutex);
 943
 944         atfork = registered_rculfhash_atfork;
 945         if (atfork)
 946                 atfork->before_fork(atfork->priv);
 947
 948         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 949                 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
 950                 cmm_smp_mb__after_uatomic_or();
 951                 wake_call_rcu_thread(crdp);
 952         }
 953         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 954                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0)
 955                         (void) poll(NULL, 0, 1);
 956         }
 957 }
 958
 959 /*
 960  * Clean up call_rcu data structures in the parent of a successful fork()
 961  * that is not followed by exec() in the child.  Suitable for
 962  * pthread_atfork() and friends.
 963  */
 964 void call_rcu_after_fork_parent(void)
 965 {
 966         struct call_rcu_data *crdp;
 967         struct urcu_atfork *atfork;
 968
 969         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 970                 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
 971         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 972                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
 973                         (void) poll(NULL, 0, 1);
 974         }
 975         atfork = registered_rculfhash_atfork;
 976         if (atfork)
 977                 atfork->after_fork_parent(atfork->priv);
 978         call_rcu_unlock(&call_rcu_mutex);
 979 }
 980
 981 /*
 982  * Clean up call_rcu data structures in the child of a successful fork()
 983  * that is not followed by exec().  Suitable for pthread_atfork() and
 984  * friends.
 985  */
 986 void call_rcu_after_fork_child(void)
 987 {
 988         struct call_rcu_data *crdp, *next;
 989         struct urcu_atfork *atfork;
 990
 991         /* Release the mutex. */
 992         call_rcu_unlock(&call_rcu_mutex);
 993
 994         atfork = registered_rculfhash_atfork;
 995         if (atfork)
 996                 atfork->after_fork_child(atfork->priv);
 997
 998         /* Do nothing when call_rcu() has not been used */
 999         if (cds_list_empty(&call_rcu_data_list))
1000                 return;
1001
1002         /*
1003          * Allocate a new default call_rcu_data structure in order
1004          * to get a working call_rcu thread to go with it.
1005          */
1006         default_call_rcu_data = NULL;
1007         (void)get_default_call_rcu_data();
1008
1009         /* Cleanup call_rcu_data pointers before use */
1010         cpus_array_len_reset();
1011         free(per_cpu_call_rcu_data);
1012         rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
1013         URCU_TLS(thread_call_rcu_data) = NULL;
1014
1015         /*
1016          * Dispose of all of the rest of the call_rcu_data structures.
1017          * Leftover call_rcu callbacks will be merged into the new
1018          * default call_rcu thread queue.
1019          */
1020         cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
1021                 if (crdp == default_call_rcu_data)
1022                         continue;
1023                 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
1024                 call_rcu_data_free(crdp);
1025         }
1026 }
1027
1028 void urcu_register_rculfhash_atfork(struct urcu_atfork *atfork)
1029 {
1030         if (CMM_LOAD_SHARED(registered_rculfhash_atfork))
1031                 return;
1032         call_rcu_lock(&call_rcu_mutex);
1033         if (!registered_rculfhash_atfork)
1034                 registered_rculfhash_atfork = atfork;
1035         call_rcu_unlock(&call_rcu_mutex);
1036 }
1037
1038 /*
1039  * This unregistration function is deprecated, meant only for internal
1040  * use by rculfhash.
1041  */
1042 __attribute__((noreturn))
1043 void urcu_unregister_rculfhash_atfork(struct urcu_atfork *atfork __attribute__((unused)))
1044 {
1045         urcu_die(EPERM);
1046 }