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 enum crdf_flags {
  85         CRDF_FLAG_JOIN_THREAD = (1 << 0),
  86 };
  87
  88 /*
  89  * List of all call_rcu_data structures to keep valgrind happy.
  90  * Protected by call_rcu_mutex.
  91  */
  92
  93 static CDS_LIST_HEAD(call_rcu_data_list);
  94
  95 /* Link a thread using call_rcu() to its call_rcu thread. */
  96
  97 static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data);
  98
  99 /*
 100  * Guard call_rcu thread creation and atfork handlers.
 101  */
 102 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
 103
 104 /* If a given thread does not have its own call_rcu thread, this is default. */
 105
 106 static struct call_rcu_data *default_call_rcu_data;
 107
 108 static struct urcu_atfork *registered_rculfhash_atfork;
 109
 110 /*
 111  * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
 112  * available, then we can have call_rcu threads assigned to individual
 113  * CPUs rather than only to specific threads.
 114  */
 115
 116 #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
 117
 118 /*
 119  * Pointer to array of pointers to per-CPU call_rcu_data structures
 120  * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
 121  * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
 122  * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
 123  * without mutex. The call_rcu_mutex protects updates.
 124  */
 125
 126 static struct call_rcu_data **per_cpu_call_rcu_data;
 127 static long cpus_array_len;
 128
 129 static void cpus_array_len_reset(void)
 130 {
 131         cpus_array_len = 0;
 132 }
 133
 134 /* Allocate the array if it has not already been allocated. */
 135
 136 static void alloc_cpu_call_rcu_data(void)
 137 {
 138         struct call_rcu_data **p;
 139         static int warned = 0;
 140
 141         if (cpus_array_len != 0)
 142                 return;
 143         cpus_array_len = get_possible_cpus_array_len();
 144         if (cpus_array_len <= 0) {
 145                 return;
 146         }
 147         p = malloc(cpus_array_len * sizeof(*per_cpu_call_rcu_data));
 148         if (p != NULL) {
 149                 memset(p, '\0', cpus_array_len * sizeof(*per_cpu_call_rcu_data));
 150                 rcu_set_pointer(&per_cpu_call_rcu_data, p);
 151         } else {
 152                 if (!warned) {
 153                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 154                 }
 155                 warned = 1;
 156         }
 157 }
 158
 159 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 160
 161 /*
 162  * per_cpu_call_rcu_data should be constant, but some functions below, used both
 163  * for cases where cpu number is available and not available, assume it it not
 164  * constant.
 165  */
 166 static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
 167 static const long cpus_array_len = -1;
 168
 169 static void cpus_array_len_reset(void)
 170 {
 171 }
 172
 173 static void alloc_cpu_call_rcu_data(void)
 174 {
 175 }
 176
 177 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 178
 179 /* Acquire the specified pthread mutex. */
 180
 181 static void call_rcu_lock(pthread_mutex_t *pmp)
 182 {
 183         int ret;
 184
 185         ret = pthread_mutex_lock(pmp);
 186         if (ret)
 187                 urcu_die(ret);
 188 }
 189
 190 /* Release the specified pthread mutex. */
 191
 192 static void call_rcu_unlock(pthread_mutex_t *pmp)
 193 {
 194         int ret;
 195
 196         ret = pthread_mutex_unlock(pmp);
 197         if (ret)
 198                 urcu_die(ret);
 199 }
 200
 201 /*
 202  * Periodically retry setting CPU affinity if we migrate.
 203  * Losing affinity can be caused by CPU hotunplug/hotplug, or by
 204  * cpuset(7).
 205  */
 206 #ifdef HAVE_SCHED_SETAFFINITY
 207 static
 208 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
 209 {
 210         cpu_set_t mask;
 211         int ret;
 212
 213         if (crdp->cpu_affinity < 0)
 214                 return 0;
 215         if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK)
 216                 return 0;
 217         if (urcu_sched_getcpu() == crdp->cpu_affinity)
 218                 return 0;
 219
 220         CPU_ZERO(&mask);
 221         CPU_SET(crdp->cpu_affinity, &mask);
 222         ret = sched_setaffinity(0, sizeof(mask), &mask);
 223
 224         /*
 225          * EINVAL is fine: can be caused by hotunplugged CPUs, or by
 226          * cpuset(7). This is why we should always retry if we detect
 227          * migration.
 228          */
 229         if (ret && errno == EINVAL) {
 230                 ret = 0;
 231                 errno = 0;
 232         }
 233         return ret;
 234 }
 235 #else
 236 static
 237 int set_thread_cpu_affinity(struct call_rcu_data *crdp __attribute__((unused)))
 238 {
 239         return 0;
 240 }
 241 #endif
 242
 243 static void call_rcu_wait(struct call_rcu_data *crdp)
 244 {
 245         /* Read call_rcu list before read futex */
 246         cmm_smp_mb();
 247         while (uatomic_read(&crdp->futex) == -1) {
 248                 if (!futex_async(&crdp->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 249                         /*
 250                          * Prior queued wakeups queued by unrelated code
 251                          * using the same address can cause futex wait to
 252                          * return 0 even through the futex value is still
 253                          * -1 (spurious wakeups). Check the value again
 254                          * in user-space to validate whether it really
 255                          * differs from -1.
 256                          */
 257                         continue;
 258                 }
 259                 switch (errno) {
 260                 case EAGAIN:
 261                         /* Value already changed. */
 262                         return;
 263                 case EINTR:
 264                         /* Retry if interrupted by signal. */
 265                         break;  /* Get out of switch. Check again. */
 266                 default:
 267                         /* Unexpected error. */
 268                         urcu_die(errno);
 269                 }
 270         }
 271 }
 272
 273 static void call_rcu_wake_up(struct call_rcu_data *crdp)
 274 {
 275         /* Write to call_rcu list before reading/writing futex */
 276         cmm_smp_mb();
 277         if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
 278                 uatomic_set(&crdp->futex, 0);
 279                 if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
 280                                 NULL, NULL, 0) < 0)
 281                         urcu_die(errno);
 282         }
 283 }
 284
 285 static void call_rcu_completion_wait(struct call_rcu_completion *completion)
 286 {
 287         /* Read completion barrier count before read futex */
 288         cmm_smp_mb();
 289         while (uatomic_read(&completion->futex) == -1) {
 290                 if (!futex_async(&completion->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 291                         /*
 292                          * Prior queued wakeups queued by unrelated code
 293                          * using the same address can cause futex wait to
 294                          * return 0 even through the futex value is still
 295                          * -1 (spurious wakeups). Check the value again
 296                          * in user-space to validate whether it really
 297                          * differs from -1.
 298                          */
 299                         continue;
 300                 }
 301                 switch (errno) {
 302                 case EAGAIN:
 303                         /* Value already changed. */
 304                         return;
 305                 case EINTR:
 306                         /* Retry if interrupted by signal. */
 307                         break;  /* Get out of switch. Check again. */
 308                 default:
 309                         /* Unexpected error. */
 310                         urcu_die(errno);
 311                 }
 312         }
 313 }
 314
 315 static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
 316 {
 317         /* Write to completion barrier count before reading/writing futex */
 318         cmm_smp_mb();
 319         if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
 320                 uatomic_set(&completion->futex, 0);
 321                 if (futex_async(&completion->futex, FUTEX_WAKE, 1,
 322                                 NULL, NULL, 0) < 0)
 323                         urcu_die(errno);
 324         }
 325 }
 326
 327 /* This is the code run by each call_rcu thread. */
 328
 329 static void *call_rcu_thread(void *arg)
 330 {
 331         unsigned long cbcount;
 332         struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
 333         int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
 334
 335         if (set_thread_cpu_affinity(crdp))
 336                 urcu_die(errno);
 337
 338         /*
 339          * If callbacks take a read-side lock, we need to be registered.
 340          */
 341         rcu_register_thread();
 342
 343         URCU_TLS(thread_call_rcu_data) = crdp;
 344         if (!rt) {
 345                 uatomic_dec(&crdp->futex);
 346                 /* Decrement futex before reading call_rcu list */
 347                 cmm_smp_mb();
 348         }
 349         for (;;) {
 350                 struct cds_wfcq_head cbs_tmp_head;
 351                 struct cds_wfcq_tail cbs_tmp_tail;
 352                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 353                 enum cds_wfcq_ret splice_ret;
 354
 355                 if (set_thread_cpu_affinity(crdp))
 356                         urcu_die(errno);
 357
 358                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
 359                         /*
 360                          * Pause requested. Become quiescent: remove
 361                          * ourself from all global lists, and don't
 362                          * process any callback. The callback lists may
 363                          * still be non-empty though.
 364                          */
 365                         rcu_unregister_thread();
 366                         cmm_smp_mb__before_uatomic_or();
 367                         uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
 368                         while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
 369                                 (void) poll(NULL, 0, 1);
 370                         uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
 371                         cmm_smp_mb__after_uatomic_and();
 372                         rcu_register_thread();
 373                 }
 374
 375                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 376                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 377                         &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail);
 378                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 379                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 380                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 381                         synchronize_rcu();
 382                         cbcount = 0;
 383                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 384                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 385                                 struct rcu_head *rhp;
 386
 387                                 rhp = caa_container_of(cbs,
 388                                         struct rcu_head, next);
 389                                 rhp->func(rhp);
 390                                 cbcount++;
 391                         }
 392                         uatomic_sub(&crdp->qlen, cbcount);
 393                 }
 394                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
 395                         break;
 396                 rcu_thread_offline();
 397                 if (!rt) {
 398                         if (cds_wfcq_empty(&crdp->cbs_head,
 399                                         &crdp->cbs_tail)) {
 400                                 call_rcu_wait(crdp);
 401                                 (void) poll(NULL, 0, 10);
 402                                 uatomic_dec(&crdp->futex);
 403                                 /*
 404                                  * Decrement futex before reading
 405                                  * call_rcu list.
 406                                  */
 407                                 cmm_smp_mb();
 408                         } else {
 409                                 (void) poll(NULL, 0, 10);
 410                         }
 411                 } else {
 412                         (void) poll(NULL, 0, 10);
 413                 }
 414                 rcu_thread_online();
 415         }
 416         if (!rt) {
 417                 /*
 418                  * Read call_rcu list before write futex.
 419                  */
 420                 cmm_smp_mb();
 421                 uatomic_set(&crdp->futex, 0);
 422         }
 423         uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
 424         rcu_unregister_thread();
 425         return NULL;
 426 }
 427
 428 /*
 429  * Create both a call_rcu thread and the corresponding call_rcu_data
 430  * structure, linking the structure in as specified.  Caller must hold
 431  * call_rcu_mutex.
 432  */
 433
 434 static void call_rcu_data_init(struct call_rcu_data **crdpp,
 435                                unsigned long flags,
 436                                int cpu_affinity)
 437 {
 438         struct call_rcu_data *crdp;
 439         int ret;
 440         sigset_t newmask, oldmask;
 441
 442         crdp = malloc(sizeof(*crdp));
 443         if (crdp == NULL)
 444                 urcu_die(errno);
 445         memset(crdp, '\0', sizeof(*crdp));
 446         cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail);
 447         crdp->qlen = 0;
 448         crdp->futex = 0;
 449         crdp->flags = flags;
 450         cds_list_add(&crdp->list, &call_rcu_data_list);
 451         crdp->cpu_affinity = cpu_affinity;
 452         crdp->gp_count = 0;
 453         rcu_set_pointer(crdpp, crdp);
 454
 455         ret = sigfillset(&newmask);
 456         urcu_posix_assert(!ret);
 457         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 458         urcu_posix_assert(!ret);
 459
 460         ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
 461         if (ret)
 462                 urcu_die(ret);
 463
 464         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 465         urcu_posix_assert(!ret);
 466 }
 467
 468 /*
 469  * Return a pointer to the call_rcu_data structure for the specified
 470  * CPU, returning NULL if there is none.  We cannot automatically
 471  * created it because the platform we are running on might not define
 472  * urcu_sched_getcpu().
 473  *
 474  * The call to this function and use of the returned call_rcu_data
 475  * should be protected by RCU read-side lock.
 476  */
 477
 478 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
 479 {
 480         static int warned = 0;
 481         struct call_rcu_data **pcpu_crdp;
 482
 483         pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
 484         if (pcpu_crdp == NULL)
 485                 return NULL;
 486         if (!warned && cpus_array_len > 0 && (cpu < 0 || cpus_array_len <= cpu)) {
 487                 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
 488                 warned = 1;
 489         }
 490         if (cpu < 0 || cpus_array_len <= cpu)
 491                 return NULL;
 492         return rcu_dereference(pcpu_crdp[cpu]);
 493 }
 494
 495 /*
 496  * Return the tid corresponding to the call_rcu thread whose
 497  * call_rcu_data structure is specified.
 498  */
 499
 500 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
 501 {
 502         return crdp->tid;
 503 }
 504
 505 /*
 506  * Create a call_rcu_data structure (with thread) and return a pointer.
 507  */
 508
 509 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
 510                                                     int cpu_affinity)
 511 {
 512         struct call_rcu_data *crdp;
 513
 514         call_rcu_data_init(&crdp, flags, cpu_affinity);
 515         return crdp;
 516 }
 517
 518 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
 519                                            int cpu_affinity)
 520 {
 521         struct call_rcu_data *crdp;
 522
 523         call_rcu_lock(&call_rcu_mutex);
 524         crdp = __create_call_rcu_data(flags, cpu_affinity);
 525         call_rcu_unlock(&call_rcu_mutex);
 526         return crdp;
 527 }
 528
 529 /*
 530  * Set the specified CPU to use the specified call_rcu_data structure.
 531  *
 532  * Use NULL to remove a CPU's call_rcu_data structure, but it is
 533  * the caller's responsibility to dispose of the removed structure.
 534  * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
 535  * (prior to NULLing it out, of course).
 536  *
 537  * The caller must wait for a grace-period to pass between return from
 538  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 539  * previous call rcu data as argument.
 540  */
 541
 542 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
 543 {
 544         static int warned = 0;
 545
 546         call_rcu_lock(&call_rcu_mutex);
 547         alloc_cpu_call_rcu_data();
 548         if (cpu < 0 || cpus_array_len <= cpu) {
 549                 if (!warned) {
 550                         fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
 551                         warned = 1;
 552                 }
 553                 call_rcu_unlock(&call_rcu_mutex);
 554                 errno = EINVAL;
 555                 return -EINVAL;
 556         }
 557
 558         if (per_cpu_call_rcu_data == NULL) {
 559                 call_rcu_unlock(&call_rcu_mutex);
 560                 errno = ENOMEM;
 561                 return -ENOMEM;
 562         }
 563
 564         if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
 565                 call_rcu_unlock(&call_rcu_mutex);
 566                 errno = EEXIST;
 567                 return -EEXIST;
 568         }
 569
 570         rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
 571         call_rcu_unlock(&call_rcu_mutex);
 572         return 0;
 573 }
 574
 575 /*
 576  * Return a pointer to the default call_rcu_data structure, creating
 577  * one if need be.
 578  *
 579  * The call to this function with intent to use the returned
 580  * call_rcu_data should be protected by RCU read-side lock.
 581  */
 582
 583 struct call_rcu_data *get_default_call_rcu_data(void)
 584 {
 585         struct call_rcu_data *crdp;
 586
 587         crdp = rcu_dereference(default_call_rcu_data);
 588         if (crdp != NULL)
 589                 return crdp;
 590
 591         call_rcu_lock(&call_rcu_mutex);
 592         if (default_call_rcu_data == NULL)
 593                 call_rcu_data_init(&default_call_rcu_data, 0, -1);
 594         crdp = default_call_rcu_data;
 595         call_rcu_unlock(&call_rcu_mutex);
 596
 597         return crdp;
 598 }
 599
 600 /*
 601  * Return the call_rcu_data structure that applies to the currently
 602  * running thread.  Any call_rcu_data structure assigned specifically
 603  * to this thread has first priority, followed by any call_rcu_data
 604  * structure assigned to the CPU on which the thread is running,
 605  * followed by the default call_rcu_data structure.  If there is not
 606  * yet a default call_rcu_data structure, one will be created.
 607  *
 608  * Calls to this function and use of the returned call_rcu_data should
 609  * be protected by RCU read-side lock.
 610  */
 611 struct call_rcu_data *get_call_rcu_data(void)
 612 {
 613         struct call_rcu_data *crd;
 614
 615         if (URCU_TLS(thread_call_rcu_data) != NULL)
 616                 return URCU_TLS(thread_call_rcu_data);
 617
 618         if (cpus_array_len > 0) {
 619                 crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
 620                 if (crd)
 621                         return crd;
 622         }
 623
 624         return get_default_call_rcu_data();
 625 }
 626
 627 /*
 628  * Return a pointer to this task's call_rcu_data if there is one.
 629  */
 630
 631 struct call_rcu_data *get_thread_call_rcu_data(void)
 632 {
 633         return URCU_TLS(thread_call_rcu_data);
 634 }
 635
 636 /*
 637  * Set this task's call_rcu_data structure as specified, regardless
 638  * of whether or not this task already had one.  (This allows switching
 639  * to and from real-time call_rcu threads, for example.)
 640  *
 641  * Use NULL to remove a thread's call_rcu_data structure, but it is
 642  * the caller's responsibility to dispose of the removed structure.
 643  * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
 644  * (prior to NULLing it out, of course).
 645  */
 646
 647 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
 648 {
 649         URCU_TLS(thread_call_rcu_data) = crdp;
 650 }
 651
 652 /*
 653  * Create a separate call_rcu thread for each CPU.  This does not
 654  * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
 655  * function if you want that behavior. Should be paired with
 656  * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
 657  * threads.
 658  */
 659
 660 int create_all_cpu_call_rcu_data(unsigned long flags)
 661 {
 662         int i;
 663         struct call_rcu_data *crdp;
 664         int ret;
 665
 666         call_rcu_lock(&call_rcu_mutex);
 667         alloc_cpu_call_rcu_data();
 668         call_rcu_unlock(&call_rcu_mutex);
 669         if (cpus_array_len <= 0) {
 670                 errno = EINVAL;
 671                 return -EINVAL;
 672         }
 673         if (per_cpu_call_rcu_data == NULL) {
 674                 errno = ENOMEM;
 675                 return -ENOMEM;
 676         }
 677         for (i = 0; i < cpus_array_len; i++) {
 678                 call_rcu_lock(&call_rcu_mutex);
 679                 if (get_cpu_call_rcu_data(i)) {
 680                         call_rcu_unlock(&call_rcu_mutex);
 681                         continue;
 682                 }
 683                 crdp = __create_call_rcu_data(flags, i);
 684                 if (crdp == NULL) {
 685                         call_rcu_unlock(&call_rcu_mutex);
 686                         errno = ENOMEM;
 687                         return -ENOMEM;
 688                 }
 689                 call_rcu_unlock(&call_rcu_mutex);
 690                 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
 691                         call_rcu_data_free(crdp);
 692
 693                         /* it has been created by other thread */
 694                         if (ret == -EEXIST)
 695                                 continue;
 696
 697                         return ret;
 698                 }
 699         }
 700         return 0;
 701 }
 702
 703 /*
 704  * Wake up the call_rcu thread corresponding to the specified
 705  * call_rcu_data structure.
 706  */
 707 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
 708 {
 709         if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
 710                 call_rcu_wake_up(crdp);
 711 }
 712
 713 static void _call_rcu(struct rcu_head *head,
 714                       void (*func)(struct rcu_head *head),
 715                       struct call_rcu_data *crdp)
 716 {
 717         cds_wfcq_node_init(&head->next);
 718         head->func = func;
 719         cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
 720         uatomic_inc(&crdp->qlen);
 721         wake_call_rcu_thread(crdp);
 722 }
 723
 724 /*
 725  * Schedule a function to be invoked after a following grace period.
 726  * This is the only function that must be called -- the others are
 727  * only present to allow applications to tune their use of RCU for
 728  * maximum performance.
 729  *
 730  * Note that unless a call_rcu thread has not already been created,
 731  * the first invocation of call_rcu() will create one.  So, if you
 732  * need the first invocation of call_rcu() to be fast, make sure
 733  * to create a call_rcu thread first.  One way to accomplish this is
 734  * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
 735  *
 736  * call_rcu must be called by registered RCU read-side threads.
 737  */
 738 void call_rcu(struct rcu_head *head,
 739               void (*func)(struct rcu_head *head))
 740 {
 741         struct call_rcu_data *crdp;
 742
 743         /* Holding rcu read-side lock across use of per-cpu crdp */
 744         _rcu_read_lock();
 745         crdp = get_call_rcu_data();
 746         _call_rcu(head, func, crdp);
 747         _rcu_read_unlock();
 748 }
 749
 750 /*
 751  * Free up the specified call_rcu_data structure, terminating the
 752  * associated call_rcu thread.  The caller must have previously
 753  * removed the call_rcu_data structure from per-thread or per-CPU
 754  * usage.  For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
 755  * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
 756  * per-thread call_rcu_data structures.
 757  *
 758  * We silently refuse to free up the default call_rcu_data structure
 759  * because that is where we put any leftover callbacks.  Note that
 760  * the possibility of self-spawning callbacks makes it impossible
 761  * to execute all the callbacks in finite time without putting any
 762  * newly spawned callbacks somewhere else.  The "somewhere else" of
 763  * last resort is the default call_rcu_data structure.
 764  *
 765  * We also silently refuse to free NULL pointers.  This simplifies
 766  * the calling code.
 767  *
 768  * The caller must wait for a grace-period to pass between return from
 769  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 770  * previous call rcu data as argument.
 771  *
 772  * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
 773  * a list corruption bug in the 0.7.x series. The equivalent fix
 774  * appeared in 0.6.8 for the stable-0.6 branch.
 775  */
 776 static
 777 void _call_rcu_data_free(struct call_rcu_data *crdp, unsigned int flags)
 778 {
 779         if (crdp == NULL || crdp == default_call_rcu_data) {
 780                 return;
 781         }
 782         if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
 783                 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
 784                 wake_call_rcu_thread(crdp);
 785                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
 786                         (void) poll(NULL, 0, 1);
 787         }
 788         call_rcu_lock(&call_rcu_mutex);
 789         if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
 790                 call_rcu_unlock(&call_rcu_mutex);
 791                 /* Create default call rcu data if need be. */
 792                 /* CBs queued here will be handed to the default list. */
 793                 (void) get_default_call_rcu_data();
 794                 call_rcu_lock(&call_rcu_mutex);
 795                 __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
 796                         &default_call_rcu_data->cbs_tail,
 797                         &crdp->cbs_head, &crdp->cbs_tail);
 798                 uatomic_add(&default_call_rcu_data->qlen,
 799                             uatomic_read(&crdp->qlen));
 800                 wake_call_rcu_thread(default_call_rcu_data);
 801         }
 802
 803         cds_list_del(&crdp->list);
 804         call_rcu_unlock(&call_rcu_mutex);
 805
 806         if (flags & CRDF_FLAG_JOIN_THREAD) {
 807                 int ret;
 808
 809                 ret = pthread_join(get_call_rcu_thread(crdp), NULL);
 810                 if (ret)
 811                         urcu_die(ret);
 812         }
 813         free(crdp);
 814 }
 815
 816 void call_rcu_data_free(struct call_rcu_data *crdp)
 817 {
 818         _call_rcu_data_free(crdp, CRDF_FLAG_JOIN_THREAD);
 819 }
 820
 821 /*
 822  * Clean up all the per-CPU call_rcu threads.
 823  */
 824 void free_all_cpu_call_rcu_data(void)
 825 {
 826         int cpu;
 827         struct call_rcu_data **crdp;
 828         static int warned = 0;
 829
 830         if (cpus_array_len <= 0)
 831                 return;
 832
 833         crdp = malloc(sizeof(*crdp) * cpus_array_len);
 834         if (!crdp) {
 835                 if (!warned) {
 836                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 837                 }
 838                 warned = 1;
 839                 return;
 840         }
 841
 842         for (cpu = 0; cpu < cpus_array_len; cpu++) {
 843                 crdp[cpu] = get_cpu_call_rcu_data(cpu);
 844                 if (crdp[cpu] == NULL)
 845                         continue;
 846                 set_cpu_call_rcu_data(cpu, NULL);
 847         }
 848         /*
 849          * Wait for call_rcu sites acting as RCU readers of the
 850          * call_rcu_data to become quiescent.
 851          */
 852         synchronize_rcu();
 853         for (cpu = 0; cpu < cpus_array_len; cpu++) {
 854                 if (crdp[cpu] == NULL)
 855                         continue;
 856                 call_rcu_data_free(crdp[cpu]);
 857         }
 858         free(crdp);
 859 }
 860
 861 static
 862 void free_completion(struct urcu_ref *ref)
 863 {
 864         struct call_rcu_completion *completion;
 865
 866         completion = caa_container_of(ref, struct call_rcu_completion, ref);
 867         free(completion);
 868 }
 869
 870 static
 871 void _rcu_barrier_complete(struct rcu_head *head)
 872 {
 873         struct call_rcu_completion_work *work;
 874         struct call_rcu_completion *completion;
 875
 876         work = caa_container_of(head, struct call_rcu_completion_work, head);
 877         completion = work->completion;
 878         if (!uatomic_sub_return(&completion->barrier_count, 1))
 879                 call_rcu_completion_wake_up(completion);
 880         urcu_ref_put(&completion->ref, free_completion);
 881         free(work);
 882 }
 883
 884 /*
 885  * Wait for all in-flight call_rcu callbacks to complete execution.
 886  */
 887 void rcu_barrier(void)
 888 {
 889         struct call_rcu_data *crdp;
 890         struct call_rcu_completion *completion;
 891         int count = 0;
 892         int was_online;
 893
 894         /* Put in offline state in QSBR. */
 895         was_online = _rcu_read_ongoing();
 896         if (was_online)
 897                 rcu_thread_offline();
 898         /*
 899          * Calling a rcu_barrier() within a RCU read-side critical
 900          * section is an error.
 901          */
 902         if (_rcu_read_ongoing()) {
 903                 static int warned = 0;
 904
 905                 if (!warned) {
 906                         fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
 907                 }
 908                 warned = 1;
 909                 goto online;
 910         }
 911
 912         completion = calloc(sizeof(*completion), 1);
 913         if (!completion)
 914                 urcu_die(errno);
 915
 916         call_rcu_lock(&call_rcu_mutex);
 917         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 918                 count++;
 919
 920         /* Referenced by rcu_barrier() and each call_rcu thread. */
 921         urcu_ref_set(&completion->ref, count + 1);
 922         completion->barrier_count = count;
 923
 924         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 925                 struct call_rcu_completion_work *work;
 926
 927                 work = calloc(sizeof(*work), 1);
 928                 if (!work)
 929                         urcu_die(errno);
 930                 work->completion = completion;
 931                 _call_rcu(&work->head, _rcu_barrier_complete, crdp);
 932         }
 933         call_rcu_unlock(&call_rcu_mutex);
 934
 935         /* Wait for them */
 936         for (;;) {
 937                 uatomic_dec(&completion->futex);
 938                 /* Decrement futex before reading barrier_count */
 939                 cmm_smp_mb();
 940                 if (!uatomic_read(&completion->barrier_count))
 941                         break;
 942                 call_rcu_completion_wait(completion);
 943         }
 944
 945         urcu_ref_put(&completion->ref, free_completion);
 946
 947 online:
 948         if (was_online)
 949                 rcu_thread_online();
 950 }
 951
 952 /*
 953  * Acquire the call_rcu_mutex in order to ensure that the child sees
 954  * all of the call_rcu() data structures in a consistent state. Ensure
 955  * that all call_rcu threads are in a quiescent state across fork.
 956  * Suitable for pthread_atfork() and friends.
 957  */
 958 void call_rcu_before_fork(void)
 959 {
 960         struct call_rcu_data *crdp;
 961         struct urcu_atfork *atfork;
 962
 963         call_rcu_lock(&call_rcu_mutex);
 964
 965         atfork = registered_rculfhash_atfork;
 966         if (atfork)
 967                 atfork->before_fork(atfork->priv);
 968
 969         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 970                 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
 971                 cmm_smp_mb__after_uatomic_or();
 972                 wake_call_rcu_thread(crdp);
 973         }
 974         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 975                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0)
 976                         (void) poll(NULL, 0, 1);
 977         }
 978 }
 979
 980 /*
 981  * Clean up call_rcu data structures in the parent of a successful fork()
 982  * that is not followed by exec() in the child.  Suitable for
 983  * pthread_atfork() and friends.
 984  */
 985 void call_rcu_after_fork_parent(void)
 986 {
 987         struct call_rcu_data *crdp;
 988         struct urcu_atfork *atfork;
 989
 990         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 991                 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
 992         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 993                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
 994                         (void) poll(NULL, 0, 1);
 995         }
 996         atfork = registered_rculfhash_atfork;
 997         if (atfork)
 998                 atfork->after_fork_parent(atfork->priv);
 999         call_rcu_unlock(&call_rcu_mutex);
1000 }
1001
1002 /*
1003  * Clean up call_rcu data structures in the child of a successful fork()
1004  * that is not followed by exec().  Suitable for pthread_atfork() and
1005  * friends.
1006  */
1007 void call_rcu_after_fork_child(void)
1008 {
1009         struct call_rcu_data *crdp, *next;
1010         struct urcu_atfork *atfork;
1011
1012         /* Release the mutex. */
1013         call_rcu_unlock(&call_rcu_mutex);
1014
1015         atfork = registered_rculfhash_atfork;
1016         if (atfork)
1017                 atfork->after_fork_child(atfork->priv);
1018
1019         /* Do nothing when call_rcu() has not been used */
1020         if (cds_list_empty(&call_rcu_data_list))
1021                 return;
1022
1023         /*
1024          * Allocate a new default call_rcu_data structure in order
1025          * to get a working call_rcu thread to go with it.
1026          */
1027         default_call_rcu_data = NULL;
1028         (void)get_default_call_rcu_data();
1029
1030         /* Cleanup call_rcu_data pointers before use */
1031         cpus_array_len_reset();
1032         free(per_cpu_call_rcu_data);
1033         rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
1034         URCU_TLS(thread_call_rcu_data) = NULL;
1035
1036         /*
1037          * Dispose of all of the rest of the call_rcu_data structures.
1038          * Leftover call_rcu callbacks will be merged into the new
1039          * default call_rcu thread queue.
1040          */
1041         cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
1042                 if (crdp == default_call_rcu_data)
1043                         continue;
1044                 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
1045                 /*
1046                  * Do not join the thread because it does not exist in
1047                  * the child.
1048                  */
1049                 _call_rcu_data_free(crdp, 0);
1050         }
1051 }
1052
1053 void urcu_register_rculfhash_atfork(struct urcu_atfork *atfork)
1054 {
1055         if (CMM_LOAD_SHARED(registered_rculfhash_atfork))
1056                 return;
1057         call_rcu_lock(&call_rcu_mutex);
1058         if (!registered_rculfhash_atfork)
1059                 registered_rculfhash_atfork = atfork;
1060         call_rcu_unlock(&call_rcu_mutex);
1061 }
1062
1063 /*
1064  * This unregistration function is deprecated, meant only for internal
1065  * use by rculfhash.
1066  */
1067 __attribute__((__noreturn__))
1068 void urcu_unregister_rculfhash_atfork(struct urcu_atfork *atfork __attribute__((unused)))
1069 {
1070         urcu_die(EPERM);
1071 }
1072
1073 /*
1074  * Teardown the default call_rcu worker thread if there are no queued
1075  * callbacks on process exit. This prevents leaking memory.
1076  *
1077  * Here is how an application can ensure graceful teardown of this
1078  * worker thread:
1079  *
1080  * - An application queuing call_rcu callbacks should invoke
1081  *   rcu_barrier() before it exits.
1082  * - When chaining call_rcu callbacks, the number of calls to
1083  *   rcu_barrier() on application exit must match at least the maximum
1084  *   number of chained callbacks.
1085  * - If an application chains callbacks endlessly, it would have to be
1086  *   modified to stop chaining callbacks when it detects an application
1087  *   exit (e.g. with a flag), and wait for quiescence with rcu_barrier()
1088  *   after setting that flag.
1089  * - The statements above apply to a library which queues call_rcu
1090  *   callbacks, only it needs to invoke rcu_barrier in its library
1091  *   destructor.
1092  *
1093  * Note that this function does not presume it is being called when the
1094  * application is single-threaded even though this is invoked from a
1095  * destructor: this function synchronizes against concurrent calls to
1096  * get_default_call_rcu_data().
1097  */
1098 static void urcu_call_rcu_exit(void)
1099 {
1100         struct call_rcu_data *crdp;
1101         bool teardown = true;
1102
1103         if (default_call_rcu_data == NULL)
1104                 return;
1105         call_rcu_lock(&call_rcu_mutex);
1106         /*
1107          * If the application leaves callbacks in the default call_rcu
1108          * worker queue, keep the default worker in place.
1109          */
1110         crdp = default_call_rcu_data;
1111         if (!crdp) {
1112                 teardown = false;
1113                 goto unlock;
1114         }
1115         if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
1116                 teardown = false;
1117                 goto unlock;
1118         }
1119         rcu_set_pointer(&default_call_rcu_data, NULL);
1120 unlock:
1121         call_rcu_unlock(&call_rcu_mutex);
1122         if (teardown) {
1123                 synchronize_rcu();
1124                 call_rcu_data_free(crdp);
1125         }
1126 }