src/urcu-call-rcu-impl.h

   1 // SPDX-FileCopyrightText: 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
   2 //
   3 // SPDX-License-Identifier: LGPL-2.1-or-later
   4
   5 /*
   6  * Userspace RCU library - batch memory reclamation with kernel API
   7  */
   8
   9 #define _LGPL_SOURCE
  10 #include <stdio.h>
  11 #include <pthread.h>
  12 #include <signal.h>
  13 #include <stdlib.h>
  14 #include <stdint.h>
  15 #include <string.h>
  16 #include <errno.h>
  17 #include <poll.h>
  18 #include <sys/time.h>
  19 #include <unistd.h>
  20 #include <sched.h>
  21
  22 #include "compat-getcpu.h"
  23 #include <urcu/assert.h>
  24 #include <urcu/wfcqueue.h>
  25 #include <urcu/call-rcu.h>
  26 #include <urcu/pointer.h>
  27 #include <urcu/list.h>
  28 #include <urcu/futex.h>
  29 #include <urcu/tls-compat.h>
  30 #include <urcu/ref.h>
  31 #include "urcu-die.h"
  32 #include "urcu-utils.h"
  33 #include "compat-smp.h"
  34
  35 #define SET_AFFINITY_CHECK_PERIOD               (1U << 8)       /* 256 */
  36 #define SET_AFFINITY_CHECK_PERIOD_MASK          (SET_AFFINITY_CHECK_PERIOD - 1)
  37
  38 /* Data structure that identifies a call_rcu thread. */
  39
  40 struct call_rcu_data {
  41         /*
  42          * We do not align head on a different cache-line than tail
  43          * mainly because call_rcu callback-invocation threads use
  44          * batching ("splice") to get an entire list of callbacks, which
  45          * effectively empties the queue, and requires to touch the tail
  46          * anyway.
  47          */
  48         struct cds_wfcq_tail cbs_tail;
  49         struct cds_wfcq_head cbs_head;
  50         unsigned long flags;
  51         int32_t futex;
  52         unsigned long qlen; /* maintained for debugging. */
  53         pthread_t tid;
  54         int cpu_affinity;
  55         unsigned long gp_count;
  56         struct cds_list_head list;
  57 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
  58
  59 struct call_rcu_completion {
  60         int barrier_count;
  61         int32_t futex;
  62         struct urcu_ref ref;
  63 };
  64
  65 struct call_rcu_completion_work {
  66         struct rcu_head head;
  67         struct call_rcu_completion *completion;
  68 };
  69
  70 enum crdf_flags {
  71         CRDF_FLAG_JOIN_THREAD = (1 << 0),
  72 };
  73
  74 /*
  75  * List of all call_rcu_data structures to keep valgrind happy.
  76  * Protected by call_rcu_mutex.
  77  */
  78
  79 static CDS_LIST_HEAD(call_rcu_data_list);
  80
  81 /* Link a thread using call_rcu() to its call_rcu thread. */
  82
  83 static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data);
  84
  85 /*
  86  * Guard call_rcu thread creation and atfork handlers.
  87  */
  88 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  89
  90 /* If a given thread does not have its own call_rcu thread, this is default. */
  91
  92 static struct call_rcu_data *default_call_rcu_data;
  93
  94 static struct urcu_atfork *registered_rculfhash_atfork;
  95
  96 /*
  97  * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
  98  * available, then we can have call_rcu threads assigned to individual
  99  * CPUs rather than only to specific threads.
 100  */
 101
 102 #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
 103
 104 /*
 105  * Pointer to array of pointers to per-CPU call_rcu_data structures
 106  * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
 107  * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
 108  * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
 109  * without mutex. The call_rcu_mutex protects updates.
 110  */
 111
 112 static struct call_rcu_data **per_cpu_call_rcu_data;
 113 static long cpus_array_len;
 114
 115 static void cpus_array_len_reset(void)
 116 {
 117         cpus_array_len = 0;
 118 }
 119
 120 /* Allocate the array if it has not already been allocated. */
 121
 122 static void alloc_cpu_call_rcu_data(void)
 123 {
 124         struct call_rcu_data **p;
 125         static int warned = 0;
 126
 127         if (cpus_array_len != 0)
 128                 return;
 129         cpus_array_len = get_possible_cpus_array_len();
 130         if (cpus_array_len <= 0) {
 131                 return;
 132         }
 133         p = malloc(cpus_array_len * sizeof(*per_cpu_call_rcu_data));
 134         if (p != NULL) {
 135                 memset(p, '\0', cpus_array_len * sizeof(*per_cpu_call_rcu_data));
 136                 rcu_set_pointer(&per_cpu_call_rcu_data, p);
 137         } else {
 138                 if (!warned) {
 139                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 140                 }
 141                 warned = 1;
 142         }
 143 }
 144
 145 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 146
 147 /*
 148  * per_cpu_call_rcu_data should be constant, but some functions below, used both
 149  * for cases where cpu number is available and not available, assume it it not
 150  * constant.
 151  */
 152 static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
 153 static const long cpus_array_len = -1;
 154
 155 static void cpus_array_len_reset(void)
 156 {
 157 }
 158
 159 static void alloc_cpu_call_rcu_data(void)
 160 {
 161 }
 162
 163 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 164
 165 /* Acquire the specified pthread mutex. */
 166
 167 static void call_rcu_lock(pthread_mutex_t *pmp)
 168 {
 169         int ret;
 170
 171         ret = pthread_mutex_lock(pmp);
 172         if (ret)
 173                 urcu_die(ret);
 174 }
 175
 176 /* Release the specified pthread mutex. */
 177
 178 static void call_rcu_unlock(pthread_mutex_t *pmp)
 179 {
 180         int ret;
 181
 182         ret = pthread_mutex_unlock(pmp);
 183         if (ret)
 184                 urcu_die(ret);
 185 }
 186
 187 /*
 188  * Periodically retry setting CPU affinity if we migrate.
 189  * Losing affinity can be caused by CPU hotunplug/hotplug, or by
 190  * cpuset(7).
 191  */
 192 #ifdef HAVE_SCHED_SETAFFINITY
 193 static
 194 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
 195 {
 196         cpu_set_t mask;
 197         int ret;
 198
 199         if (crdp->cpu_affinity < 0)
 200                 return 0;
 201         if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK)
 202                 return 0;
 203         if (urcu_sched_getcpu() == crdp->cpu_affinity)
 204                 return 0;
 205
 206         CPU_ZERO(&mask);
 207         CPU_SET(crdp->cpu_affinity, &mask);
 208         ret = sched_setaffinity(0, sizeof(mask), &mask);
 209
 210         /*
 211          * EINVAL is fine: can be caused by hotunplugged CPUs, or by
 212          * cpuset(7). This is why we should always retry if we detect
 213          * migration.
 214          */
 215         if (ret && errno == EINVAL) {
 216                 ret = 0;
 217                 errno = 0;
 218         }
 219         return ret;
 220 }
 221 #else
 222 static
 223 int set_thread_cpu_affinity(struct call_rcu_data *crdp __attribute__((unused)))
 224 {
 225         return 0;
 226 }
 227 #endif
 228
 229 static void call_rcu_wait(struct call_rcu_data *crdp)
 230 {
 231         /* Read call_rcu list before read futex */
 232         cmm_smp_mb();
 233         while (uatomic_read(&crdp->futex) == -1) {
 234                 if (!futex_async(&crdp->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 235                         /*
 236                          * Prior queued wakeups queued by unrelated code
 237                          * using the same address can cause futex wait to
 238                          * return 0 even through the futex value is still
 239                          * -1 (spurious wakeups). Check the value again
 240                          * in user-space to validate whether it really
 241                          * differs from -1.
 242                          */
 243                         continue;
 244                 }
 245                 switch (errno) {
 246                 case EAGAIN:
 247                         /* Value already changed. */
 248                         return;
 249                 case EINTR:
 250                         /* Retry if interrupted by signal. */
 251                         break;  /* Get out of switch. Check again. */
 252                 default:
 253                         /* Unexpected error. */
 254                         urcu_die(errno);
 255                 }
 256         }
 257 }
 258
 259 static void call_rcu_wake_up(struct call_rcu_data *crdp)
 260 {
 261         /* Write to call_rcu list before reading/writing futex */
 262         cmm_smp_mb();
 263         if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
 264                 uatomic_set(&crdp->futex, 0);
 265                 if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
 266                                 NULL, NULL, 0) < 0)
 267                         urcu_die(errno);
 268         }
 269 }
 270
 271 static void call_rcu_completion_wait(struct call_rcu_completion *completion)
 272 {
 273         /* Read completion barrier count before read futex */
 274         cmm_smp_mb();
 275         while (uatomic_read(&completion->futex) == -1) {
 276                 if (!futex_async(&completion->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 277                         /*
 278                          * Prior queued wakeups queued by unrelated code
 279                          * using the same address can cause futex wait to
 280                          * return 0 even through the futex value is still
 281                          * -1 (spurious wakeups). Check the value again
 282                          * in user-space to validate whether it really
 283                          * differs from -1.
 284                          */
 285                         continue;
 286                 }
 287                 switch (errno) {
 288                 case EAGAIN:
 289                         /* Value already changed. */
 290                         return;
 291                 case EINTR:
 292                         /* Retry if interrupted by signal. */
 293                         break;  /* Get out of switch. Check again. */
 294                 default:
 295                         /* Unexpected error. */
 296                         urcu_die(errno);
 297                 }
 298         }
 299 }
 300
 301 static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
 302 {
 303         /* Write to completion barrier count before reading/writing futex */
 304         cmm_smp_mb();
 305         if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
 306                 uatomic_set(&completion->futex, 0);
 307                 if (futex_async(&completion->futex, FUTEX_WAKE, 1,
 308                                 NULL, NULL, 0) < 0)
 309                         urcu_die(errno);
 310         }
 311 }
 312
 313 /* This is the code run by each call_rcu thread. */
 314
 315 static void *call_rcu_thread(void *arg)
 316 {
 317         unsigned long cbcount;
 318         struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
 319         int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
 320
 321         if (set_thread_cpu_affinity(crdp))
 322                 urcu_die(errno);
 323
 324         /*
 325          * If callbacks take a read-side lock, we need to be registered.
 326          */
 327         rcu_register_thread();
 328
 329         URCU_TLS(thread_call_rcu_data) = crdp;
 330         if (!rt) {
 331                 uatomic_dec(&crdp->futex);
 332                 /* Decrement futex before reading call_rcu list */
 333                 cmm_smp_mb();
 334         }
 335         for (;;) {
 336                 struct cds_wfcq_head cbs_tmp_head;
 337                 struct cds_wfcq_tail cbs_tmp_tail;
 338                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 339                 enum cds_wfcq_ret splice_ret;
 340
 341                 if (set_thread_cpu_affinity(crdp))
 342                         urcu_die(errno);
 343
 344                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
 345                         /*
 346                          * Pause requested. Become quiescent: remove
 347                          * ourself from all global lists, and don't
 348                          * process any callback. The callback lists may
 349                          * still be non-empty though.
 350                          */
 351                         rcu_unregister_thread();
 352                         cmm_smp_mb__before_uatomic_or();
 353                         uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
 354                         while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
 355                                 (void) poll(NULL, 0, 1);
 356                         uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
 357                         cmm_smp_mb__after_uatomic_and();
 358                         rcu_register_thread();
 359                 }
 360
 361                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 362                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 363                         &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail);
 364                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 365                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 366                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 367                         synchronize_rcu();
 368                         cbcount = 0;
 369                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 370                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 371                                 struct rcu_head *rhp;
 372
 373                                 rhp = caa_container_of(cbs,
 374                                         struct rcu_head, next);
 375                                 rhp->func(rhp);
 376                                 cbcount++;
 377                         }
 378                         uatomic_sub(&crdp->qlen, cbcount);
 379                 }
 380                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
 381                         break;
 382                 rcu_thread_offline();
 383                 if (!rt) {
 384                         if (cds_wfcq_empty(&crdp->cbs_head,
 385                                         &crdp->cbs_tail)) {
 386                                 call_rcu_wait(crdp);
 387                                 (void) poll(NULL, 0, 10);
 388                                 uatomic_dec(&crdp->futex);
 389                                 /*
 390                                  * Decrement futex before reading
 391                                  * call_rcu list.
 392                                  */
 393                                 cmm_smp_mb();
 394                         } else {
 395                                 (void) poll(NULL, 0, 10);
 396                         }
 397                 } else {
 398                         (void) poll(NULL, 0, 10);
 399                 }
 400                 rcu_thread_online();
 401         }
 402         if (!rt) {
 403                 /*
 404                  * Read call_rcu list before write futex.
 405                  */
 406                 cmm_smp_mb();
 407                 uatomic_set(&crdp->futex, 0);
 408         }
 409         uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
 410         rcu_unregister_thread();
 411         return NULL;
 412 }
 413
 414 /*
 415  * Create both a call_rcu thread and the corresponding call_rcu_data
 416  * structure, linking the structure in as specified.  Caller must hold
 417  * call_rcu_mutex.
 418  */
 419
 420 static void call_rcu_data_init(struct call_rcu_data **crdpp,
 421                                unsigned long flags,
 422                                int cpu_affinity)
 423 {
 424         struct call_rcu_data *crdp;
 425         int ret;
 426         sigset_t newmask, oldmask;
 427
 428         crdp = malloc(sizeof(*crdp));
 429         if (crdp == NULL)
 430                 urcu_die(errno);
 431         memset(crdp, '\0', sizeof(*crdp));
 432         cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail);
 433         crdp->qlen = 0;
 434         crdp->futex = 0;
 435         crdp->flags = flags;
 436         cds_list_add(&crdp->list, &call_rcu_data_list);
 437         crdp->cpu_affinity = cpu_affinity;
 438         crdp->gp_count = 0;
 439         rcu_set_pointer(crdpp, crdp);
 440
 441         ret = sigfillset(&newmask);
 442         urcu_posix_assert(!ret);
 443         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 444         urcu_posix_assert(!ret);
 445
 446         ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
 447         if (ret)
 448                 urcu_die(ret);
 449
 450         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 451         urcu_posix_assert(!ret);
 452 }
 453
 454 /*
 455  * Return a pointer to the call_rcu_data structure for the specified
 456  * CPU, returning NULL if there is none.  We cannot automatically
 457  * created it because the platform we are running on might not define
 458  * urcu_sched_getcpu().
 459  *
 460  * The call to this function and use of the returned call_rcu_data
 461  * should be protected by RCU read-side lock.
 462  */
 463
 464 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
 465 {
 466         static int warned = 0;
 467         struct call_rcu_data **pcpu_crdp;
 468
 469         pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
 470         if (pcpu_crdp == NULL)
 471                 return NULL;
 472         if (!warned && cpus_array_len > 0 && (cpu < 0 || cpus_array_len <= cpu)) {
 473                 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
 474                 warned = 1;
 475         }
 476         if (cpu < 0 || cpus_array_len <= cpu)
 477                 return NULL;
 478         return rcu_dereference(pcpu_crdp[cpu]);
 479 }
 480
 481 /*
 482  * Return the tid corresponding to the call_rcu thread whose
 483  * call_rcu_data structure is specified.
 484  */
 485
 486 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
 487 {
 488         return crdp->tid;
 489 }
 490
 491 /*
 492  * Create a call_rcu_data structure (with thread) and return a pointer.
 493  */
 494
 495 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
 496                                                     int cpu_affinity)
 497 {
 498         struct call_rcu_data *crdp;
 499
 500         call_rcu_data_init(&crdp, flags, cpu_affinity);
 501         return crdp;
 502 }
 503
 504 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
 505                                            int cpu_affinity)
 506 {
 507         struct call_rcu_data *crdp;
 508
 509         call_rcu_lock(&call_rcu_mutex);
 510         crdp = __create_call_rcu_data(flags, cpu_affinity);
 511         call_rcu_unlock(&call_rcu_mutex);
 512         return crdp;
 513 }
 514
 515 /*
 516  * Set the specified CPU to use the specified call_rcu_data structure.
 517  *
 518  * Use NULL to remove a CPU's call_rcu_data structure, but it is
 519  * the caller's responsibility to dispose of the removed structure.
 520  * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
 521  * (prior to NULLing it out, of course).
 522  *
 523  * The caller must wait for a grace-period to pass between return from
 524  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 525  * previous call rcu data as argument.
 526  */
 527
 528 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
 529 {
 530         static int warned = 0;
 531
 532         call_rcu_lock(&call_rcu_mutex);
 533         alloc_cpu_call_rcu_data();
 534         if (cpu < 0 || cpus_array_len <= cpu) {
 535                 if (!warned) {
 536                         fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
 537                         warned = 1;
 538                 }
 539                 call_rcu_unlock(&call_rcu_mutex);
 540                 errno = EINVAL;
 541                 return -EINVAL;
 542         }
 543
 544         if (per_cpu_call_rcu_data == NULL) {
 545                 call_rcu_unlock(&call_rcu_mutex);
 546                 errno = ENOMEM;
 547                 return -ENOMEM;
 548         }
 549
 550         if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
 551                 call_rcu_unlock(&call_rcu_mutex);
 552                 errno = EEXIST;
 553                 return -EEXIST;
 554         }
 555
 556         rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
 557         call_rcu_unlock(&call_rcu_mutex);
 558         return 0;
 559 }
 560
 561 /*
 562  * Return a pointer to the default call_rcu_data structure, creating
 563  * one if need be.
 564  *
 565  * The call to this function with intent to use the returned
 566  * call_rcu_data should be protected by RCU read-side lock.
 567  */
 568
 569 struct call_rcu_data *get_default_call_rcu_data(void)
 570 {
 571         struct call_rcu_data *crdp;
 572
 573         crdp = rcu_dereference(default_call_rcu_data);
 574         if (crdp != NULL)
 575                 return crdp;
 576
 577         call_rcu_lock(&call_rcu_mutex);
 578         if (default_call_rcu_data == NULL)
 579                 call_rcu_data_init(&default_call_rcu_data, 0, -1);
 580         crdp = default_call_rcu_data;
 581         call_rcu_unlock(&call_rcu_mutex);
 582
 583         return crdp;
 584 }
 585
 586 /*
 587  * Return the call_rcu_data structure that applies to the currently
 588  * running thread.  Any call_rcu_data structure assigned specifically
 589  * to this thread has first priority, followed by any call_rcu_data
 590  * structure assigned to the CPU on which the thread is running,
 591  * followed by the default call_rcu_data structure.  If there is not
 592  * yet a default call_rcu_data structure, one will be created.
 593  *
 594  * Calls to this function and use of the returned call_rcu_data should
 595  * be protected by RCU read-side lock.
 596  */
 597 struct call_rcu_data *get_call_rcu_data(void)
 598 {
 599         struct call_rcu_data *crd;
 600
 601         if (URCU_TLS(thread_call_rcu_data) != NULL)
 602                 return URCU_TLS(thread_call_rcu_data);
 603
 604         if (cpus_array_len > 0) {
 605                 crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
 606                 if (crd)
 607                         return crd;
 608         }
 609
 610         return get_default_call_rcu_data();
 611 }
 612
 613 /*
 614  * Return a pointer to this task's call_rcu_data if there is one.
 615  */
 616
 617 struct call_rcu_data *get_thread_call_rcu_data(void)
 618 {
 619         return URCU_TLS(thread_call_rcu_data);
 620 }
 621
 622 /*
 623  * Set this task's call_rcu_data structure as specified, regardless
 624  * of whether or not this task already had one.  (This allows switching
 625  * to and from real-time call_rcu threads, for example.)
 626  *
 627  * Use NULL to remove a thread's call_rcu_data structure, but it is
 628  * the caller's responsibility to dispose of the removed structure.
 629  * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
 630  * (prior to NULLing it out, of course).
 631  */
 632
 633 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
 634 {
 635         URCU_TLS(thread_call_rcu_data) = crdp;
 636 }
 637
 638 /*
 639  * Create a separate call_rcu thread for each CPU.  This does not
 640  * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
 641  * function if you want that behavior. Should be paired with
 642  * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
 643  * threads.
 644  */
 645
 646 int create_all_cpu_call_rcu_data(unsigned long flags)
 647 {
 648         int i;
 649         struct call_rcu_data *crdp;
 650         int ret;
 651
 652         call_rcu_lock(&call_rcu_mutex);
 653         alloc_cpu_call_rcu_data();
 654         call_rcu_unlock(&call_rcu_mutex);
 655         if (cpus_array_len <= 0) {
 656                 errno = EINVAL;
 657                 return -EINVAL;
 658         }
 659         if (per_cpu_call_rcu_data == NULL) {
 660                 errno = ENOMEM;
 661                 return -ENOMEM;
 662         }
 663         for (i = 0; i < cpus_array_len; i++) {
 664                 call_rcu_lock(&call_rcu_mutex);
 665                 if (get_cpu_call_rcu_data(i)) {
 666                         call_rcu_unlock(&call_rcu_mutex);
 667                         continue;
 668                 }
 669                 crdp = __create_call_rcu_data(flags, i);
 670                 if (crdp == NULL) {
 671                         call_rcu_unlock(&call_rcu_mutex);
 672                         errno = ENOMEM;
 673                         return -ENOMEM;
 674                 }
 675                 call_rcu_unlock(&call_rcu_mutex);
 676                 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
 677                         call_rcu_data_free(crdp);
 678
 679                         /* it has been created by other thread */
 680                         if (ret == -EEXIST)
 681                                 continue;
 682
 683                         return ret;
 684                 }
 685         }
 686         return 0;
 687 }
 688
 689 /*
 690  * Wake up the call_rcu thread corresponding to the specified
 691  * call_rcu_data structure.
 692  */
 693 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
 694 {
 695         if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
 696                 call_rcu_wake_up(crdp);
 697 }
 698
 699 static void _call_rcu(struct rcu_head *head,
 700                       void (*func)(struct rcu_head *head),
 701                       struct call_rcu_data *crdp)
 702 {
 703         cds_wfcq_node_init(&head->next);
 704         head->func = func;
 705         cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
 706         uatomic_inc(&crdp->qlen);
 707         wake_call_rcu_thread(crdp);
 708 }
 709
 710 /*
 711  * Schedule a function to be invoked after a following grace period.
 712  * This is the only function that must be called -- the others are
 713  * only present to allow applications to tune their use of RCU for
 714  * maximum performance.
 715  *
 716  * Note that unless a call_rcu thread has not already been created,
 717  * the first invocation of call_rcu() will create one.  So, if you
 718  * need the first invocation of call_rcu() to be fast, make sure
 719  * to create a call_rcu thread first.  One way to accomplish this is
 720  * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
 721  *
 722  * call_rcu must be called by registered RCU read-side threads.
 723  */
 724 void call_rcu(struct rcu_head *head,
 725               void (*func)(struct rcu_head *head))
 726 {
 727         struct call_rcu_data *crdp;
 728
 729         /* Holding rcu read-side lock across use of per-cpu crdp */
 730         _rcu_read_lock();
 731         crdp = get_call_rcu_data();
 732         _call_rcu(head, func, crdp);
 733         _rcu_read_unlock();
 734 }
 735
 736 /*
 737  * Free up the specified call_rcu_data structure, terminating the
 738  * associated call_rcu thread.  The caller must have previously
 739  * removed the call_rcu_data structure from per-thread or per-CPU
 740  * usage.  For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
 741  * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
 742  * per-thread call_rcu_data structures.
 743  *
 744  * We silently refuse to free up the default call_rcu_data structure
 745  * because that is where we put any leftover callbacks.  Note that
 746  * the possibility of self-spawning callbacks makes it impossible
 747  * to execute all the callbacks in finite time without putting any
 748  * newly spawned callbacks somewhere else.  The "somewhere else" of
 749  * last resort is the default call_rcu_data structure.
 750  *
 751  * We also silently refuse to free NULL pointers.  This simplifies
 752  * the calling code.
 753  *
 754  * The caller must wait for a grace-period to pass between return from
 755  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 756  * previous call rcu data as argument.
 757  *
 758  * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
 759  * a list corruption bug in the 0.7.x series. The equivalent fix
 760  * appeared in 0.6.8 for the stable-0.6 branch.
 761  */
 762 static
 763 void _call_rcu_data_free(struct call_rcu_data *crdp, unsigned int flags)
 764 {
 765         if (crdp == NULL || crdp == default_call_rcu_data) {
 766                 return;
 767         }
 768         if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
 769                 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
 770                 wake_call_rcu_thread(crdp);
 771                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
 772                         (void) poll(NULL, 0, 1);
 773         }
 774         call_rcu_lock(&call_rcu_mutex);
 775         if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
 776                 call_rcu_unlock(&call_rcu_mutex);
 777                 /* Create default call rcu data if need be. */
 778                 /* CBs queued here will be handed to the default list. */
 779                 (void) get_default_call_rcu_data();
 780                 call_rcu_lock(&call_rcu_mutex);
 781                 __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
 782                         &default_call_rcu_data->cbs_tail,
 783                         &crdp->cbs_head, &crdp->cbs_tail);
 784                 uatomic_add(&default_call_rcu_data->qlen,
 785                             uatomic_read(&crdp->qlen));
 786                 wake_call_rcu_thread(default_call_rcu_data);
 787         }
 788
 789         cds_list_del(&crdp->list);
 790         call_rcu_unlock(&call_rcu_mutex);
 791
 792         if (flags & CRDF_FLAG_JOIN_THREAD) {
 793                 int ret;
 794
 795                 ret = pthread_join(get_call_rcu_thread(crdp), NULL);
 796                 if (ret)
 797                         urcu_die(ret);
 798         }
 799         free(crdp);
 800 }
 801
 802 void call_rcu_data_free(struct call_rcu_data *crdp)
 803 {
 804         _call_rcu_data_free(crdp, CRDF_FLAG_JOIN_THREAD);
 805 }
 806
 807 /*
 808  * Clean up all the per-CPU call_rcu threads.
 809  */
 810 void free_all_cpu_call_rcu_data(void)
 811 {
 812         int cpu;
 813         struct call_rcu_data **crdp;
 814         static int warned = 0;
 815
 816         if (cpus_array_len <= 0)
 817                 return;
 818
 819         crdp = malloc(sizeof(*crdp) * cpus_array_len);
 820         if (!crdp) {
 821                 if (!warned) {
 822                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 823                 }
 824                 warned = 1;
 825                 return;
 826         }
 827
 828         for (cpu = 0; cpu < cpus_array_len; cpu++) {
 829                 crdp[cpu] = get_cpu_call_rcu_data(cpu);
 830                 if (crdp[cpu] == NULL)
 831                         continue;
 832                 set_cpu_call_rcu_data(cpu, NULL);
 833         }
 834         /*
 835          * Wait for call_rcu sites acting as RCU readers of the
 836          * call_rcu_data to become quiescent.
 837          */
 838         synchronize_rcu();
 839         for (cpu = 0; cpu < cpus_array_len; cpu++) {
 840                 if (crdp[cpu] == NULL)
 841                         continue;
 842                 call_rcu_data_free(crdp[cpu]);
 843         }
 844         free(crdp);
 845 }
 846
 847 static
 848 void free_completion(struct urcu_ref *ref)
 849 {
 850         struct call_rcu_completion *completion;
 851
 852         completion = caa_container_of(ref, struct call_rcu_completion, ref);
 853         free(completion);
 854 }
 855
 856 static
 857 void _rcu_barrier_complete(struct rcu_head *head)
 858 {
 859         struct call_rcu_completion_work *work;
 860         struct call_rcu_completion *completion;
 861
 862         work = caa_container_of(head, struct call_rcu_completion_work, head);
 863         completion = work->completion;
 864         if (!uatomic_sub_return(&completion->barrier_count, 1))
 865                 call_rcu_completion_wake_up(completion);
 866         urcu_ref_put(&completion->ref, free_completion);
 867         free(work);
 868 }
 869
 870 /*
 871  * Wait for all in-flight call_rcu callbacks to complete execution.
 872  */
 873 void rcu_barrier(void)
 874 {
 875         struct call_rcu_data *crdp;
 876         struct call_rcu_completion *completion;
 877         int count = 0;
 878         int was_online;
 879
 880         /* Put in offline state in QSBR. */
 881         was_online = _rcu_read_ongoing();
 882         if (was_online)
 883                 rcu_thread_offline();
 884         /*
 885          * Calling a rcu_barrier() within a RCU read-side critical
 886          * section is an error.
 887          */
 888         if (_rcu_read_ongoing()) {
 889                 static int warned = 0;
 890
 891                 if (!warned) {
 892                         fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
 893                 }
 894                 warned = 1;
 895                 goto online;
 896         }
 897
 898         completion = calloc(1, sizeof(*completion));
 899         if (!completion)
 900                 urcu_die(errno);
 901
 902         call_rcu_lock(&call_rcu_mutex);
 903         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 904                 count++;
 905
 906         /* Referenced by rcu_barrier() and each call_rcu thread. */
 907         urcu_ref_set(&completion->ref, count + 1);
 908         completion->barrier_count = count;
 909
 910         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 911                 struct call_rcu_completion_work *work;
 912
 913                 work = calloc(1, sizeof(*work));
 914                 if (!work)
 915                         urcu_die(errno);
 916                 work->completion = completion;
 917                 _call_rcu(&work->head, _rcu_barrier_complete, crdp);
 918         }
 919         call_rcu_unlock(&call_rcu_mutex);
 920
 921         /* Wait for them */
 922         for (;;) {
 923                 uatomic_dec(&completion->futex);
 924                 /* Decrement futex before reading barrier_count */
 925                 cmm_smp_mb();
 926                 if (!uatomic_read(&completion->barrier_count))
 927                         break;
 928                 call_rcu_completion_wait(completion);
 929         }
 930
 931         urcu_ref_put(&completion->ref, free_completion);
 932
 933 online:
 934         if (was_online)
 935                 rcu_thread_online();
 936 }
 937
 938 /*
 939  * Acquire the call_rcu_mutex in order to ensure that the child sees
 940  * all of the call_rcu() data structures in a consistent state. Ensure
 941  * that all call_rcu threads are in a quiescent state across fork.
 942  * Suitable for pthread_atfork() and friends.
 943  */
 944 void call_rcu_before_fork(void)
 945 {
 946         struct call_rcu_data *crdp;
 947         struct urcu_atfork *atfork;
 948
 949         call_rcu_lock(&call_rcu_mutex);
 950
 951         atfork = registered_rculfhash_atfork;
 952         if (atfork)
 953                 atfork->before_fork(atfork->priv);
 954
 955         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 956                 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
 957                 cmm_smp_mb__after_uatomic_or();
 958                 wake_call_rcu_thread(crdp);
 959         }
 960         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 961                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0)
 962                         (void) poll(NULL, 0, 1);
 963         }
 964 }
 965
 966 /*
 967  * Clean up call_rcu data structures in the parent of a successful fork()
 968  * that is not followed by exec() in the child.  Suitable for
 969  * pthread_atfork() and friends.
 970  */
 971 void call_rcu_after_fork_parent(void)
 972 {
 973         struct call_rcu_data *crdp;
 974         struct urcu_atfork *atfork;
 975
 976         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 977                 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
 978         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 979                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
 980                         (void) poll(NULL, 0, 1);
 981         }
 982         atfork = registered_rculfhash_atfork;
 983         if (atfork)
 984                 atfork->after_fork_parent(atfork->priv);
 985         call_rcu_unlock(&call_rcu_mutex);
 986 }
 987
 988 /*
 989  * Clean up call_rcu data structures in the child of a successful fork()
 990  * that is not followed by exec().  Suitable for pthread_atfork() and
 991  * friends.
 992  */
 993 void call_rcu_after_fork_child(void)
 994 {
 995         struct call_rcu_data *crdp, *next;
 996         struct urcu_atfork *atfork;
 997
 998         /* Release the mutex. */
 999         call_rcu_unlock(&call_rcu_mutex);
1000
1001         atfork = registered_rculfhash_atfork;
1002         if (atfork)
1003                 atfork->after_fork_child(atfork->priv);
1004
1005         /* Do nothing when call_rcu() has not been used */
1006         if (cds_list_empty(&call_rcu_data_list))
1007                 return;
1008
1009         /*
1010          * Allocate a new default call_rcu_data structure in order
1011          * to get a working call_rcu thread to go with it.
1012          */
1013         default_call_rcu_data = NULL;
1014         (void)get_default_call_rcu_data();
1015
1016         /* Cleanup call_rcu_data pointers before use */
1017         cpus_array_len_reset();
1018         free(per_cpu_call_rcu_data);
1019         rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
1020         URCU_TLS(thread_call_rcu_data) = NULL;
1021
1022         /*
1023          * Dispose of all of the rest of the call_rcu_data structures.
1024          * Leftover call_rcu callbacks will be merged into the new
1025          * default call_rcu thread queue.
1026          */
1027         cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
1028                 if (crdp == default_call_rcu_data)
1029                         continue;
1030                 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
1031                 /*
1032                  * Do not join the thread because it does not exist in
1033                  * the child.
1034                  */
1035                 _call_rcu_data_free(crdp, 0);
1036         }
1037 }
1038
1039 void urcu_register_rculfhash_atfork(struct urcu_atfork *atfork)
1040 {
1041         if (CMM_LOAD_SHARED(registered_rculfhash_atfork))
1042                 return;
1043         call_rcu_lock(&call_rcu_mutex);
1044         if (!registered_rculfhash_atfork)
1045                 registered_rculfhash_atfork = atfork;
1046         call_rcu_unlock(&call_rcu_mutex);
1047 }
1048
1049 /*
1050  * This unregistration function is deprecated, meant only for internal
1051  * use by rculfhash.
1052  */
1053 __attribute__((__noreturn__))
1054 void urcu_unregister_rculfhash_atfork(struct urcu_atfork *atfork __attribute__((unused)))
1055 {
1056         urcu_die(EPERM);
1057 }
1058
1059 /*
1060  * Teardown the default call_rcu worker thread if there are no queued
1061  * callbacks on process exit. This prevents leaking memory.
1062  *
1063  * Here is how an application can ensure graceful teardown of this
1064  * worker thread:
1065  *
1066  * - An application queuing call_rcu callbacks should invoke
1067  *   rcu_barrier() before it exits.
1068  * - When chaining call_rcu callbacks, the number of calls to
1069  *   rcu_barrier() on application exit must match at least the maximum
1070  *   number of chained callbacks.
1071  * - If an application chains callbacks endlessly, it would have to be
1072  *   modified to stop chaining callbacks when it detects an application
1073  *   exit (e.g. with a flag), and wait for quiescence with rcu_barrier()
1074  *   after setting that flag.
1075  * - The statements above apply to a library which queues call_rcu
1076  *   callbacks, only it needs to invoke rcu_barrier in its library
1077  *   destructor.
1078  *
1079  * Note that this function does not presume it is being called when the
1080  * application is single-threaded even though this is invoked from a
1081  * destructor: this function synchronizes against concurrent calls to
1082  * get_default_call_rcu_data().
1083  */
1084 static void urcu_call_rcu_exit(void)
1085 {
1086         struct call_rcu_data *crdp;
1087         bool teardown = true;
1088
1089         if (default_call_rcu_data == NULL)
1090                 return;
1091         call_rcu_lock(&call_rcu_mutex);
1092         /*
1093          * If the application leaves callbacks in the default call_rcu
1094          * worker queue, keep the default worker in place.
1095          */
1096         crdp = default_call_rcu_data;
1097         if (!crdp) {
1098                 teardown = false;
1099                 goto unlock;
1100         }
1101         if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
1102                 teardown = false;
1103                 goto unlock;
1104         }
1105         rcu_set_pointer(&default_call_rcu_data, NULL);
1106 unlock:
1107         call_rcu_unlock(&call_rcu_mutex);
1108         if (teardown) {
1109                 synchronize_rcu();
1110                 call_rcu_data_free(crdp);
1111         }
1112 }