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