src/urcu-call-rcu-impl.h

   1 /*
   2  * urcu-call-rcu.c
   3  *
   4  * Userspace RCU library - batch memory reclamation with kernel API
   5  *
   6  * Copyright (c) 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
   7  *
   8  * This library is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * This library is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with this library; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #define _LGPL_SOURCE
  24 #include <stdio.h>
  25 #include <pthread.h>
  26 #include <signal.h>
  27 #include <assert.h>
  28 #include <stdlib.h>
  29 #include <stdint.h>
  30 #include <string.h>
  31 #include <errno.h>
  32 #include <poll.h>
  33 #include <sys/time.h>
  34 #include <unistd.h>
  35 #include <sched.h>
  36
  37 #include "compat-getcpu.h"
  38 #include <urcu/wfcqueue.h>
  39 #include <urcu/call-rcu.h>
  40 #include <urcu/pointer.h>
  41 #include <urcu/list.h>
  42 #include <urcu/futex.h>
  43 #include <urcu/tls-compat.h>
  44 #include <urcu/ref.h>
  45 #include "urcu-die.h"
  46 #include "urcu-utils.h"
  47
  48 #define SET_AFFINITY_CHECK_PERIOD               (1U << 8)       /* 256 */
  49 #define SET_AFFINITY_CHECK_PERIOD_MASK          (SET_AFFINITY_CHECK_PERIOD - 1)
  50
  51 /* Data structure that identifies a call_rcu thread. */
  52
  53 struct call_rcu_data {
  54         /*
  55          * We do not align head on a different cache-line than tail
  56          * mainly because call_rcu callback-invocation threads use
  57          * batching ("splice") to get an entire list of callbacks, which
  58          * effectively empties the queue, and requires to touch the tail
  59          * anyway.
  60          */
  61         struct cds_wfcq_tail cbs_tail;
  62         struct cds_wfcq_head cbs_head;
  63         unsigned long flags;
  64         int32_t futex;
  65         unsigned long qlen; /* maintained for debugging. */
  66         pthread_t tid;
  67         int cpu_affinity;
  68         unsigned long gp_count;
  69         struct cds_list_head list;
  70 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
  71
  72 struct call_rcu_completion {
  73         int barrier_count;
  74         int32_t futex;
  75         struct urcu_ref ref;
  76 };
  77
  78 struct call_rcu_completion_work {
  79         struct rcu_head head;
  80         struct call_rcu_completion *completion;
  81 };
  82
  83 /*
  84  * List of all call_rcu_data structures to keep valgrind happy.
  85  * Protected by call_rcu_mutex.
  86  */
  87
  88 static CDS_LIST_HEAD(call_rcu_data_list);
  89
  90 /* Link a thread using call_rcu() to its call_rcu thread. */
  91
  92 static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data);
  93
  94 /*
  95  * Guard call_rcu thread creation and atfork handlers.
  96  */
  97 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  98
  99 /* If a given thread does not have its own call_rcu thread, this is default. */
 100
 101 static struct call_rcu_data *default_call_rcu_data;
 102
 103 static struct urcu_atfork *registered_rculfhash_atfork;
 104 static unsigned long registered_rculfhash_atfork_refcount;
 105
 106 /*
 107  * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
 108  * available, then we can have call_rcu threads assigned to individual
 109  * CPUs rather than only to specific threads.
 110  */
 111
 112 #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
 113
 114 /*
 115  * Pointer to array of pointers to per-CPU call_rcu_data structures
 116  * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
 117  * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
 118  * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
 119  * without mutex. The call_rcu_mutex protects updates.
 120  */
 121
 122 static struct call_rcu_data **per_cpu_call_rcu_data;
 123 static long maxcpus;
 124
 125 static void maxcpus_reset(void)
 126 {
 127         maxcpus = 0;
 128 }
 129
 130 /* Allocate the array if it has not already been allocated. */
 131
 132 static void alloc_cpu_call_rcu_data(void)
 133 {
 134         struct call_rcu_data **p;
 135         static int warned = 0;
 136
 137         if (maxcpus != 0)
 138                 return;
 139         maxcpus = sysconf(_SC_NPROCESSORS_CONF);
 140         if (maxcpus <= 0) {
 141                 return;
 142         }
 143         p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
 144         if (p != NULL) {
 145                 memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
 146                 rcu_set_pointer(&per_cpu_call_rcu_data, p);
 147         } else {
 148                 if (!warned) {
 149                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 150                 }
 151                 warned = 1;
 152         }
 153 }
 154
 155 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 156
 157 /*
 158  * per_cpu_call_rcu_data should be constant, but some functions below, used both
 159  * for cases where cpu number is available and not available, assume it it not
 160  * constant.
 161  */
 162 static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
 163 static const long maxcpus = -1;
 164
 165 static void maxcpus_reset(void)
 166 {
 167 }
 168
 169 static void alloc_cpu_call_rcu_data(void)
 170 {
 171 }
 172
 173 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
 174
 175 /* Acquire the specified pthread mutex. */
 176
 177 static void call_rcu_lock(pthread_mutex_t *pmp)
 178 {
 179         int ret;
 180
 181         ret = pthread_mutex_lock(pmp);
 182         if (ret)
 183                 urcu_die(ret);
 184 }
 185
 186 /* Release the specified pthread mutex. */
 187
 188 static void call_rcu_unlock(pthread_mutex_t *pmp)
 189 {
 190         int ret;
 191
 192         ret = pthread_mutex_unlock(pmp);
 193         if (ret)
 194                 urcu_die(ret);
 195 }
 196
 197 /*
 198  * Periodically retry setting CPU affinity if we migrate.
 199  * Losing affinity can be caused by CPU hotunplug/hotplug, or by
 200  * cpuset(7).
 201  */
 202 #if 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 #if SCHED_SETAFFINITY_ARGS == 2
 219         ret = sched_setaffinity(0, &mask);
 220 #else
 221         ret = sched_setaffinity(0, sizeof(mask), &mask);
 222 #endif
 223         /*
 224          * EINVAL is fine: can be caused by hotunplugged CPUs, or by
 225          * cpuset(7). This is why we should always retry if we detect
 226          * migration.
 227          */
 228         if (ret && errno == EINVAL) {
 229                 ret = 0;
 230                 errno = 0;
 231         }
 232         return ret;
 233 }
 234 #else
 235 static
 236 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
 237 {
 238         return 0;
 239 }
 240 #endif
 241
 242 static void call_rcu_wait(struct call_rcu_data *crdp)
 243 {
 244         /* Read call_rcu list before read futex */
 245         cmm_smp_mb();
 246         if (uatomic_read(&crdp->futex) != -1)
 247                 return;
 248         while (futex_async(&crdp->futex, FUTEX_WAIT, -1,
 249                         NULL, NULL, 0)) {
 250                 switch (errno) {
 251                 case EWOULDBLOCK:
 252                         /* Value already changed. */
 253                         return;
 254                 case EINTR:
 255                         /* Retry if interrupted by signal. */
 256                         break;  /* Get out of switch. */
 257                 default:
 258                         /* Unexpected error. */
 259                         urcu_die(errno);
 260                 }
 261         }
 262 }
 263
 264 static void call_rcu_wake_up(struct call_rcu_data *crdp)
 265 {
 266         /* Write to call_rcu list before reading/writing futex */
 267         cmm_smp_mb();
 268         if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
 269                 uatomic_set(&crdp->futex, 0);
 270                 if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
 271                                 NULL, NULL, 0) < 0)
 272                         urcu_die(errno);
 273         }
 274 }
 275
 276 static void call_rcu_completion_wait(struct call_rcu_completion *completion)
 277 {
 278         /* Read completion barrier count before read futex */
 279         cmm_smp_mb();
 280         if (uatomic_read(&completion->futex) != -1)
 281                 return;
 282         while (futex_async(&completion->futex, FUTEX_WAIT, -1,
 283                         NULL, NULL, 0)) {
 284                 switch (errno) {
 285                 case EWOULDBLOCK:
 286                         /* Value already changed. */
 287                         return;
 288                 case EINTR:
 289                         /* Retry if interrupted by signal. */
 290                         break;  /* Get out of switch. */
 291                 default:
 292                         /* Unexpected error. */
 293                         urcu_die(errno);
 294                 }
 295         }
 296 }
 297
 298 static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
 299 {
 300         /* Write to completion barrier count before reading/writing futex */
 301         cmm_smp_mb();
 302         if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
 303                 uatomic_set(&completion->futex, 0);
 304                 if (futex_async(&completion->futex, FUTEX_WAKE, 1,
 305                                 NULL, NULL, 0) < 0)
 306                         urcu_die(errno);
 307         }
 308 }
 309
 310 /* This is the code run by each call_rcu thread. */
 311
 312 static void *call_rcu_thread(void *arg)
 313 {
 314         unsigned long cbcount;
 315         struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
 316         int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
 317
 318         if (set_thread_cpu_affinity(crdp))
 319                 urcu_die(errno);
 320
 321         /*
 322          * If callbacks take a read-side lock, we need to be registered.
 323          */
 324         rcu_register_thread();
 325
 326         URCU_TLS(thread_call_rcu_data) = crdp;
 327         if (!rt) {
 328                 uatomic_dec(&crdp->futex);
 329                 /* Decrement futex before reading call_rcu list */
 330                 cmm_smp_mb();
 331         }
 332         for (;;) {
 333                 struct cds_wfcq_head cbs_tmp_head;
 334                 struct cds_wfcq_tail cbs_tmp_tail;
 335                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 336                 enum cds_wfcq_ret splice_ret;
 337
 338                 if (set_thread_cpu_affinity(crdp))
 339                         urcu_die(errno);
 340
 341                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
 342                         /*
 343                          * Pause requested. Become quiescent: remove
 344                          * ourself from all global lists, and don't
 345                          * process any callback. The callback lists may
 346                          * still be non-empty though.
 347                          */
 348                         rcu_unregister_thread();
 349                         cmm_smp_mb__before_uatomic_or();
 350                         uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
 351                         while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
 352                                 (void) poll(NULL, 0, 1);
 353                         uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
 354                         cmm_smp_mb__after_uatomic_and();
 355                         rcu_register_thread();
 356                 }
 357
 358                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 359                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 360                         &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail);
 361                 assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 362                 assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 363                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 364                         synchronize_rcu();
 365                         cbcount = 0;
 366                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 367                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 368                                 struct rcu_head *rhp;
 369
 370                                 rhp = caa_container_of(cbs,
 371                                         struct rcu_head, next);
 372                                 rhp->func(rhp);
 373                                 cbcount++;
 374                         }
 375                         uatomic_sub(&crdp->qlen, cbcount);
 376                 }
 377                 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
 378                         break;
 379                 rcu_thread_offline();
 380                 if (!rt) {
 381                         if (cds_wfcq_empty(&crdp->cbs_head,
 382                                         &crdp->cbs_tail)) {
 383                                 call_rcu_wait(crdp);
 384                                 (void) poll(NULL, 0, 10);
 385                                 uatomic_dec(&crdp->futex);
 386                                 /*
 387                                  * Decrement futex before reading
 388                                  * call_rcu list.
 389                                  */
 390                                 cmm_smp_mb();
 391                         } else {
 392                                 (void) poll(NULL, 0, 10);
 393                         }
 394                 } else {
 395                         (void) poll(NULL, 0, 10);
 396                 }
 397                 rcu_thread_online();
 398         }
 399         if (!rt) {
 400                 /*
 401                  * Read call_rcu list before write futex.
 402                  */
 403                 cmm_smp_mb();
 404                 uatomic_set(&crdp->futex, 0);
 405         }
 406         uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
 407         rcu_unregister_thread();
 408         return NULL;
 409 }
 410
 411 /*
 412  * Create both a call_rcu thread and the corresponding call_rcu_data
 413  * structure, linking the structure in as specified.  Caller must hold
 414  * call_rcu_mutex.
 415  */
 416
 417 static void call_rcu_data_init(struct call_rcu_data **crdpp,
 418                                unsigned long flags,
 419                                int cpu_affinity)
 420 {
 421         struct call_rcu_data *crdp;
 422         int ret;
 423
 424         crdp = malloc(sizeof(*crdp));
 425         if (crdp == NULL)
 426                 urcu_die(errno);
 427         memset(crdp, '\0', sizeof(*crdp));
 428         cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail);
 429         crdp->qlen = 0;
 430         crdp->futex = 0;
 431         crdp->flags = flags;
 432         cds_list_add(&crdp->list, &call_rcu_data_list);
 433         crdp->cpu_affinity = cpu_affinity;
 434         crdp->gp_count = 0;
 435         cmm_smp_mb();  /* Structure initialized before pointer is planted. */
 436         *crdpp = crdp;
 437         ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
 438         if (ret)
 439                 urcu_die(ret);
 440 }
 441
 442 /*
 443  * Return a pointer to the call_rcu_data structure for the specified
 444  * CPU, returning NULL if there is none.  We cannot automatically
 445  * created it because the platform we are running on might not define
 446  * urcu_sched_getcpu().
 447  *
 448  * The call to this function and use of the returned call_rcu_data
 449  * should be protected by RCU read-side lock.
 450  */
 451
 452 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
 453 {
 454         static int warned = 0;
 455         struct call_rcu_data **pcpu_crdp;
 456
 457         pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
 458         if (pcpu_crdp == NULL)
 459                 return NULL;
 460         if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
 461                 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
 462                 warned = 1;
 463         }
 464         if (cpu < 0 || maxcpus <= cpu)
 465                 return NULL;
 466         return rcu_dereference(pcpu_crdp[cpu]);
 467 }
 468 URCU_ATTR_ALIAS(urcu_stringify(get_cpu_call_rcu_data))
 469 struct call_rcu_data *alias_get_cpu_call_rcu_data();
 470
 471 /*
 472  * Return the tid corresponding to the call_rcu thread whose
 473  * call_rcu_data structure is specified.
 474  */
 475
 476 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
 477 {
 478         return crdp->tid;
 479 }
 480 URCU_ATTR_ALIAS(urcu_stringify(get_call_rcu_thread))
 481 pthread_t alias_get_call_rcu_thread();
 482
 483 /*
 484  * Create a call_rcu_data structure (with thread) and return a pointer.
 485  */
 486
 487 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
 488                                                     int cpu_affinity)
 489 {
 490         struct call_rcu_data *crdp;
 491
 492         call_rcu_data_init(&crdp, flags, cpu_affinity);
 493         return crdp;
 494 }
 495
 496 URCU_ATTR_ALIAS(urcu_stringify(create_call_rcu_data))
 497 struct call_rcu_data *alias_create_call_rcu_data();
 498 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
 499                                            int cpu_affinity)
 500 {
 501         struct call_rcu_data *crdp;
 502
 503         call_rcu_lock(&call_rcu_mutex);
 504         crdp = __create_call_rcu_data(flags, cpu_affinity);
 505         call_rcu_unlock(&call_rcu_mutex);
 506         return crdp;
 507 }
 508
 509 /*
 510  * Set the specified CPU to use the specified call_rcu_data structure.
 511  *
 512  * Use NULL to remove a CPU's call_rcu_data structure, but it is
 513  * the caller's responsibility to dispose of the removed structure.
 514  * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
 515  * (prior to NULLing it out, of course).
 516  *
 517  * The caller must wait for a grace-period to pass between return from
 518  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 519  * previous call rcu data as argument.
 520  */
 521
 522 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
 523 {
 524         static int warned = 0;
 525
 526         call_rcu_lock(&call_rcu_mutex);
 527         alloc_cpu_call_rcu_data();
 528         if (cpu < 0 || maxcpus <= cpu) {
 529                 if (!warned) {
 530                         fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
 531                         warned = 1;
 532                 }
 533                 call_rcu_unlock(&call_rcu_mutex);
 534                 errno = EINVAL;
 535                 return -EINVAL;
 536         }
 537
 538         if (per_cpu_call_rcu_data == NULL) {
 539                 call_rcu_unlock(&call_rcu_mutex);
 540                 errno = ENOMEM;
 541                 return -ENOMEM;
 542         }
 543
 544         if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
 545                 call_rcu_unlock(&call_rcu_mutex);
 546                 errno = EEXIST;
 547                 return -EEXIST;
 548         }
 549
 550         rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
 551         call_rcu_unlock(&call_rcu_mutex);
 552         return 0;
 553 }
 554 URCU_ATTR_ALIAS(urcu_stringify(set_cpu_call_rcu_data))
 555 int alias_set_cpu_call_rcu_data();
 556
 557 /*
 558  * Return a pointer to the default call_rcu_data structure, creating
 559  * one if need be.  Because we never free call_rcu_data structures,
 560  * we don't need to be in an RCU read-side critical section.
 561  */
 562
 563 struct call_rcu_data *get_default_call_rcu_data(void)
 564 {
 565         if (default_call_rcu_data != NULL)
 566                 return rcu_dereference(default_call_rcu_data);
 567         call_rcu_lock(&call_rcu_mutex);
 568         if (default_call_rcu_data != NULL) {
 569                 call_rcu_unlock(&call_rcu_mutex);
 570                 return default_call_rcu_data;
 571         }
 572         call_rcu_data_init(&default_call_rcu_data, 0, -1);
 573         call_rcu_unlock(&call_rcu_mutex);
 574         return default_call_rcu_data;
 575 }
 576 URCU_ATTR_ALIAS(urcu_stringify(get_default_call_rcu_data))
 577 struct call_rcu_data *alias_get_default_call_rcu_data();
 578
 579 /*
 580  * Return the call_rcu_data structure that applies to the currently
 581  * running thread.  Any call_rcu_data structure assigned specifically
 582  * to this thread has first priority, followed by any call_rcu_data
 583  * structure assigned to the CPU on which the thread is running,
 584  * followed by the default call_rcu_data structure.  If there is not
 585  * yet a default call_rcu_data structure, one will be created.
 586  *
 587  * Calls to this function and use of the returned call_rcu_data should
 588  * be protected by RCU read-side lock.
 589  */
 590 struct call_rcu_data *get_call_rcu_data(void)
 591 {
 592         struct call_rcu_data *crd;
 593
 594         if (URCU_TLS(thread_call_rcu_data) != NULL)
 595                 return URCU_TLS(thread_call_rcu_data);
 596
 597         if (maxcpus > 0) {
 598                 crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
 599                 if (crd)
 600                         return crd;
 601         }
 602
 603         return get_default_call_rcu_data();
 604 }
 605 URCU_ATTR_ALIAS(urcu_stringify(get_call_rcu_data))
 606 struct call_rcu_data *alias_get_call_rcu_data();
 607
 608 /*
 609  * Return a pointer to this task's call_rcu_data if there is one.
 610  */
 611
 612 struct call_rcu_data *get_thread_call_rcu_data(void)
 613 {
 614         return URCU_TLS(thread_call_rcu_data);
 615 }
 616 URCU_ATTR_ALIAS(urcu_stringify(get_thread_call_rcu_data))
 617 struct call_rcu_data *alias_get_thread_call_rcu_data();
 618
 619 /*
 620  * Set this task's call_rcu_data structure as specified, regardless
 621  * of whether or not this task already had one.  (This allows switching
 622  * to and from real-time call_rcu threads, for example.)
 623  *
 624  * Use NULL to remove a thread's call_rcu_data structure, but it is
 625  * the caller's responsibility to dispose of the removed structure.
 626  * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
 627  * (prior to NULLing it out, of course).
 628  */
 629
 630 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
 631 {
 632         URCU_TLS(thread_call_rcu_data) = crdp;
 633 }
 634 URCU_ATTR_ALIAS(urcu_stringify(set_thread_call_rcu_data))
 635 void alias_set_thread_call_rcu_data();
 636
 637 /*
 638  * Create a separate call_rcu thread for each CPU.  This does not
 639  * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
 640  * function if you want that behavior. Should be paired with
 641  * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
 642  * threads.
 643  */
 644
 645 int create_all_cpu_call_rcu_data(unsigned long flags)
 646 {
 647         int i;
 648         struct call_rcu_data *crdp;
 649         int ret;
 650
 651         call_rcu_lock(&call_rcu_mutex);
 652         alloc_cpu_call_rcu_data();
 653         call_rcu_unlock(&call_rcu_mutex);
 654         if (maxcpus <= 0) {
 655                 errno = EINVAL;
 656                 return -EINVAL;
 657         }
 658         if (per_cpu_call_rcu_data == NULL) {
 659                 errno = ENOMEM;
 660                 return -ENOMEM;
 661         }
 662         for (i = 0; i < maxcpus; i++) {
 663                 call_rcu_lock(&call_rcu_mutex);
 664                 if (get_cpu_call_rcu_data(i)) {
 665                         call_rcu_unlock(&call_rcu_mutex);
 666                         continue;
 667                 }
 668                 crdp = __create_call_rcu_data(flags, i);
 669                 if (crdp == NULL) {
 670                         call_rcu_unlock(&call_rcu_mutex);
 671                         errno = ENOMEM;
 672                         return -ENOMEM;
 673                 }
 674                 call_rcu_unlock(&call_rcu_mutex);
 675                 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
 676                         call_rcu_data_free(crdp);
 677
 678                         /* it has been created by other thread */
 679                         if (ret == -EEXIST)
 680                                 continue;
 681
 682                         return ret;
 683                 }
 684         }
 685         return 0;
 686 }
 687 URCU_ATTR_ALIAS(urcu_stringify(create_all_cpu_call_rcu_data))
 688 int alias_create_all_cpu_call_rcu_data();
 689
 690 /*
 691  * Wake up the call_rcu thread corresponding to the specified
 692  * call_rcu_data structure.
 693  */
 694 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
 695 {
 696         if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
 697                 call_rcu_wake_up(crdp);
 698 }
 699
 700 static void _call_rcu(struct rcu_head *head,
 701                       void (*func)(struct rcu_head *head),
 702                       struct call_rcu_data *crdp)
 703 {
 704         cds_wfcq_node_init(&head->next);
 705         head->func = func;
 706         cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
 707         uatomic_inc(&crdp->qlen);
 708         wake_call_rcu_thread(crdp);
 709 }
 710
 711 /*
 712  * Schedule a function to be invoked after a following grace period.
 713  * This is the only function that must be called -- the others are
 714  * only present to allow applications to tune their use of RCU for
 715  * maximum performance.
 716  *
 717  * Note that unless a call_rcu thread has not already been created,
 718  * the first invocation of call_rcu() will create one.  So, if you
 719  * need the first invocation of call_rcu() to be fast, make sure
 720  * to create a call_rcu thread first.  One way to accomplish this is
 721  * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
 722  *
 723  * call_rcu must be called by registered RCU read-side threads.
 724  */
 725 void call_rcu(struct rcu_head *head,
 726               void (*func)(struct rcu_head *head))
 727 {
 728         struct call_rcu_data *crdp;
 729
 730         /* Holding rcu read-side lock across use of per-cpu crdp */
 731         _rcu_read_lock();
 732         crdp = get_call_rcu_data();
 733         _call_rcu(head, func, crdp);
 734         _rcu_read_unlock();
 735 }
 736 URCU_ATTR_ALIAS(urcu_stringify(call_rcu)) void alias_call_rcu();
 737
 738 /*
 739  * Free up the specified call_rcu_data structure, terminating the
 740  * associated call_rcu thread.  The caller must have previously
 741  * removed the call_rcu_data structure from per-thread or per-CPU
 742  * usage.  For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
 743  * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
 744  * per-thread call_rcu_data structures.
 745  *
 746  * We silently refuse to free up the default call_rcu_data structure
 747  * because that is where we put any leftover callbacks.  Note that
 748  * the possibility of self-spawning callbacks makes it impossible
 749  * to execute all the callbacks in finite time without putting any
 750  * newly spawned callbacks somewhere else.  The "somewhere else" of
 751  * last resort is the default call_rcu_data structure.
 752  *
 753  * We also silently refuse to free NULL pointers.  This simplifies
 754  * the calling code.
 755  *
 756  * The caller must wait for a grace-period to pass between return from
 757  * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
 758  * previous call rcu data as argument.
 759  *
 760  * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
 761  * a list corruption bug in the 0.7.x series. The equivalent fix
 762  * appeared in 0.6.8 for the stable-0.6 branch.
 763  */
 764 void call_rcu_data_free(struct call_rcu_data *crdp)
 765 {
 766         if (crdp == NULL || crdp == default_call_rcu_data) {
 767                 return;
 768         }
 769         if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
 770                 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
 771                 wake_call_rcu_thread(crdp);
 772                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
 773                         (void) poll(NULL, 0, 1);
 774         }
 775         if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
 776                 /* Create default call rcu data if need be */
 777                 (void) get_default_call_rcu_data();
 778                 __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
 779                         &default_call_rcu_data->cbs_tail,
 780                         &crdp->cbs_head, &crdp->cbs_tail);
 781                 uatomic_add(&default_call_rcu_data->qlen,
 782                             uatomic_read(&crdp->qlen));
 783                 wake_call_rcu_thread(default_call_rcu_data);
 784         }
 785
 786         call_rcu_lock(&call_rcu_mutex);
 787         cds_list_del(&crdp->list);
 788         call_rcu_unlock(&call_rcu_mutex);
 789
 790         free(crdp);
 791 }
 792 URCU_ATTR_ALIAS(urcu_stringify(call_rcu_data_free))
 793 void alias_call_rcu_data_free();
 794
 795 /*
 796  * Clean up all the per-CPU call_rcu threads.
 797  */
 798 void free_all_cpu_call_rcu_data(void)
 799 {
 800         int cpu;
 801         struct call_rcu_data **crdp;
 802         static int warned = 0;
 803
 804         if (maxcpus <= 0)
 805                 return;
 806
 807         crdp = malloc(sizeof(*crdp) * maxcpus);
 808         if (!crdp) {
 809                 if (!warned) {
 810                         fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
 811                 }
 812                 warned = 1;
 813                 return;
 814         }
 815
 816         for (cpu = 0; cpu < maxcpus; cpu++) {
 817                 crdp[cpu] = get_cpu_call_rcu_data(cpu);
 818                 if (crdp[cpu] == NULL)
 819                         continue;
 820                 set_cpu_call_rcu_data(cpu, NULL);
 821         }
 822         /*
 823          * Wait for call_rcu sites acting as RCU readers of the
 824          * call_rcu_data to become quiescent.
 825          */
 826         synchronize_rcu();
 827         for (cpu = 0; cpu < maxcpus; cpu++) {
 828                 if (crdp[cpu] == NULL)
 829                         continue;
 830                 call_rcu_data_free(crdp[cpu]);
 831         }
 832         free(crdp);
 833 }
 834 #ifdef RCU_QSBR
 835 /* ABI6 has a non-namespaced free_all_cpu_call_rcu_data for qsbr */
 836 #undef free_all_cpu_call_rcu_data
 837 URCU_ATTR_ALIAS("urcu_qsbr_free_all_cpu_call_rcu_data")
 838 void free_all_cpu_call_rcu_data();
 839 #define free_all_cpu_call_rcu_data urcu_qsbr_free_all_cpu_call_rcu_data
 840 #else
 841 URCU_ATTR_ALIAS(urcu_stringify(free_all_cpu_call_rcu_data))
 842 void alias_free_all_cpu_call_rcu_data();
 843 #endif
 844
 845 static
 846 void free_completion(struct urcu_ref *ref)
 847 {
 848         struct call_rcu_completion *completion;
 849
 850         completion = caa_container_of(ref, struct call_rcu_completion, ref);
 851         free(completion);
 852 }
 853
 854 static
 855 void _rcu_barrier_complete(struct rcu_head *head)
 856 {
 857         struct call_rcu_completion_work *work;
 858         struct call_rcu_completion *completion;
 859
 860         work = caa_container_of(head, struct call_rcu_completion_work, head);
 861         completion = work->completion;
 862         if (!uatomic_sub_return(&completion->barrier_count, 1))
 863                 call_rcu_completion_wake_up(completion);
 864         urcu_ref_put(&completion->ref, free_completion);
 865         free(work);
 866 }
 867
 868 /*
 869  * Wait for all in-flight call_rcu callbacks to complete execution.
 870  */
 871 void rcu_barrier(void)
 872 {
 873         struct call_rcu_data *crdp;
 874         struct call_rcu_completion *completion;
 875         int count = 0;
 876         int was_online;
 877
 878         /* Put in offline state in QSBR. */
 879         was_online = _rcu_read_ongoing();
 880         if (was_online)
 881                 rcu_thread_offline();
 882         /*
 883          * Calling a rcu_barrier() within a RCU read-side critical
 884          * section is an error.
 885          */
 886         if (_rcu_read_ongoing()) {
 887                 static int warned = 0;
 888
 889                 if (!warned) {
 890                         fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
 891                 }
 892                 warned = 1;
 893                 goto online;
 894         }
 895
 896         completion = calloc(sizeof(*completion), 1);
 897         if (!completion)
 898                 urcu_die(errno);
 899
 900         call_rcu_lock(&call_rcu_mutex);
 901         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 902                 count++;
 903
 904         /* Referenced by rcu_barrier() and each call_rcu thread. */
 905         urcu_ref_set(&completion->ref, count + 1);
 906         completion->barrier_count = count;
 907
 908         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 909                 struct call_rcu_completion_work *work;
 910
 911                 work = calloc(sizeof(*work), 1);
 912                 if (!work)
 913                         urcu_die(errno);
 914                 work->completion = completion;
 915                 _call_rcu(&work->head, _rcu_barrier_complete, crdp);
 916         }
 917         call_rcu_unlock(&call_rcu_mutex);
 918
 919         /* Wait for them */
 920         for (;;) {
 921                 uatomic_dec(&completion->futex);
 922                 /* Decrement futex before reading barrier_count */
 923                 cmm_smp_mb();
 924                 if (!uatomic_read(&completion->barrier_count))
 925                         break;
 926                 call_rcu_completion_wait(completion);
 927         }
 928
 929         urcu_ref_put(&completion->ref, free_completion);
 930
 931 online:
 932         if (was_online)
 933                 rcu_thread_online();
 934 }
 935 URCU_ATTR_ALIAS(urcu_stringify(rcu_barrier))
 936 void alias_rcu_barrier();
 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 URCU_ATTR_ALIAS(urcu_stringify(call_rcu_before_fork))
 966 void alias_call_rcu_before_fork();
 967
 968 /*
 969  * Clean up call_rcu data structures in the parent of a successful fork()
 970  * that is not followed by exec() in the child.  Suitable for
 971  * pthread_atfork() and friends.
 972  */
 973 void call_rcu_after_fork_parent(void)
 974 {
 975         struct call_rcu_data *crdp;
 976         struct urcu_atfork *atfork;
 977
 978         cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
 979                 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
 980         cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
 981                 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
 982                         (void) poll(NULL, 0, 1);
 983         }
 984         atfork = registered_rculfhash_atfork;
 985         if (atfork)
 986                 atfork->after_fork_parent(atfork->priv);
 987         call_rcu_unlock(&call_rcu_mutex);
 988 }
 989 URCU_ATTR_ALIAS(urcu_stringify(call_rcu_after_fork_parent))
 990 void alias_call_rcu_after_fork_parent();
 991
 992 /*
 993  * Clean up call_rcu data structures in the child of a successful fork()
 994  * that is not followed by exec().  Suitable for pthread_atfork() and
 995  * friends.
 996  */
 997 void call_rcu_after_fork_child(void)
 998 {
 999         struct call_rcu_data *crdp, *next;
1000         struct urcu_atfork *atfork;
1001
1002         /* Release the mutex. */
1003         call_rcu_unlock(&call_rcu_mutex);
1004
1005         atfork = registered_rculfhash_atfork;
1006         if (atfork)
1007                 atfork->after_fork_child(atfork->priv);
1008
1009         /* Do nothing when call_rcu() has not been used */
1010         if (cds_list_empty(&call_rcu_data_list))
1011                 return;
1012
1013         /*
1014          * Allocate a new default call_rcu_data structure in order
1015          * to get a working call_rcu thread to go with it.
1016          */
1017         default_call_rcu_data = NULL;
1018         (void)get_default_call_rcu_data();
1019
1020         /* Cleanup call_rcu_data pointers before use */
1021         maxcpus_reset();
1022         free(per_cpu_call_rcu_data);
1023         rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
1024         URCU_TLS(thread_call_rcu_data) = NULL;
1025
1026         /*
1027          * Dispose of all of the rest of the call_rcu_data structures.
1028          * Leftover call_rcu callbacks will be merged into the new
1029          * default call_rcu thread queue.
1030          */
1031         cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
1032                 if (crdp == default_call_rcu_data)
1033                         continue;
1034                 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
1035                 call_rcu_data_free(crdp);
1036         }
1037 }
1038 URCU_ATTR_ALIAS(urcu_stringify(call_rcu_after_fork_child))
1039 void alias_call_rcu_after_fork_child();
1040
1041 void urcu_register_rculfhash_atfork(struct urcu_atfork *atfork)
1042 {
1043         call_rcu_lock(&call_rcu_mutex);
1044         if (registered_rculfhash_atfork_refcount++)
1045                 goto end;
1046         registered_rculfhash_atfork = atfork;
1047 end:
1048         call_rcu_unlock(&call_rcu_mutex);
1049 }
1050 URCU_ATTR_ALIAS(urcu_stringify(urcu_register_rculfhash_atfork))
1051 void alias_urcu_register_rculfhash_atfork();
1052
1053 void urcu_unregister_rculfhash_atfork(struct urcu_atfork *atfork)
1054 {
1055         call_rcu_lock(&call_rcu_mutex);
1056         if (--registered_rculfhash_atfork_refcount)
1057                 goto end;
1058         registered_rculfhash_atfork = NULL;
1059 end:
1060         call_rcu_unlock(&call_rcu_mutex);
1061 }
1062 URCU_ATTR_ALIAS(urcu_stringify(urcu_unregister_rculfhash_atfork))
1063 void alias_urcu_unregister_rculfhash_atfork();