liblttng-ust/lttng-ust-urcu.c

   1 /*
   2  * SPDX-License-Identifier: LGPL-2.1-or-later
   3  *
   4  * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   5  * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
   6  *
   7  * Userspace RCU library for LTTng-UST, derived from liburcu "bulletproof" version.
   8  */
   9
  10 #define _LGPL_SOURCE
  11 #include <stdio.h>
  12 #include <pthread.h>
  13 #include <signal.h>
  14 #include <assert.h>
  15 #include <stdlib.h>
  16 #include <string.h>
  17 #include <errno.h>
  18 #include <poll.h>
  19 #include <unistd.h>
  20 #include <stdbool.h>
  21 #include <sys/mman.h>
  22
  23 #include <urcu/arch.h>
  24 #include <urcu/wfcqueue.h>
  25 #include <lttng/urcu/static/urcu-ust.h>
  26 #include <lttng/urcu/pointer.h>
  27 #include <urcu/tls-compat.h>
  28
  29 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  30 #undef _LGPL_SOURCE
  31 #include <lttng/urcu/urcu-ust.h>
  32 #define _LGPL_SOURCE
  33
  34 #ifndef MAP_ANONYMOUS
  35 #define MAP_ANONYMOUS MAP_ANON
  36 #endif
  37
  38 #ifdef __linux__
  39 static
  40 void *mremap_wrapper(void *old_address, size_t old_size,
  41                 size_t new_size, int flags)
  42 {
  43         return mremap(old_address, old_size, new_size, flags);
  44 }
  45 #else
  46
  47 #define MREMAP_MAYMOVE  1
  48 #define MREMAP_FIXED    2
  49
  50 /*
  51  * mremap wrapper for non-Linux systems not allowing MAYMOVE.
  52  * This is not generic.
  53 */
  54 static
  55 void *mremap_wrapper(void *old_address, size_t old_size,
  56                 size_t new_size, int flags)
  57 {
  58         assert(!(flags & MREMAP_MAYMOVE));
  59
  60         return MAP_FAILED;
  61 }
  62 #endif
  63
  64 /* Sleep delay in ms */
  65 #define RCU_SLEEP_DELAY_MS      10
  66 #define INIT_NR_THREADS         8
  67 #define ARENA_INIT_ALLOC                \
  68         sizeof(struct registry_chunk)   \
  69         + INIT_NR_THREADS * sizeof(struct lttng_ust_urcu_reader)
  70
  71 /*
  72  * Active attempts to check for reader Q.S. before calling sleep().
  73  */
  74 #define RCU_QS_ACTIVE_ATTEMPTS 100
  75
  76 static
  77 int lttng_ust_urcu_refcount;
  78
  79 /* If the headers do not support membarrier system call, fall back smp_mb. */
  80 #ifdef __NR_membarrier
  81 # define membarrier(...)                syscall(__NR_membarrier, __VA_ARGS__)
  82 #else
  83 # define membarrier(...)                -ENOSYS
  84 #endif
  85
  86 enum membarrier_cmd {
  87         MEMBARRIER_CMD_QUERY                            = 0,
  88         MEMBARRIER_CMD_SHARED                           = (1 << 0),
  89         /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
  90         /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
  91         MEMBARRIER_CMD_PRIVATE_EXPEDITED                = (1 << 3),
  92         MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED       = (1 << 4),
  93 };
  94
  95 static
  96 void _lttng_ust_urcu_init(void)
  97         __attribute__((constructor));
  98 static
  99 void lttng_ust_urcu_exit(void)
 100         __attribute__((destructor));
 101
 102 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 103 int lttng_ust_urcu_has_sys_membarrier;
 104 #endif
 105
 106 /*
 107  * rcu_gp_lock ensures mutual exclusion between threads calling
 108  * synchronize_rcu().
 109  */
 110 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 111 /*
 112  * rcu_registry_lock ensures mutual exclusion between threads
 113  * registering and unregistering themselves to/from the registry, and
 114  * with threads reading that registry from synchronize_rcu(). However,
 115  * this lock is not held all the way through the completion of awaiting
 116  * for the grace period. It is sporadically released between iterations
 117  * on the registry.
 118  * rcu_registry_lock may nest inside rcu_gp_lock.
 119  */
 120 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 121
 122 static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
 123 static int initialized;
 124
 125 static pthread_key_t lttng_ust_urcu_key;
 126
 127 struct lttng_ust_urcu_gp lttng_ust_urcu_gp = { .ctr = LTTNG_UST_URCU_GP_COUNT };
 128
 129 /*
 130  * Pointer to registry elements. Written to only by each individual reader. Read
 131  * by both the reader and the writers.
 132  */
 133 DEFINE_URCU_TLS(struct lttng_ust_urcu_reader *, lttng_ust_urcu_reader);
 134
 135 static CDS_LIST_HEAD(registry);
 136
 137 struct registry_chunk {
 138         size_t data_len;                /* data length */
 139         size_t used;                    /* amount of data used */
 140         struct cds_list_head node;      /* chunk_list node */
 141         char data[];
 142 };
 143
 144 struct registry_arena {
 145         struct cds_list_head chunk_list;
 146 };
 147
 148 static struct registry_arena registry_arena = {
 149         .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list),
 150 };
 151
 152 /* Saved fork signal mask, protected by rcu_gp_lock */
 153 static sigset_t saved_fork_signal_mask;
 154
 155 static void mutex_lock(pthread_mutex_t *mutex)
 156 {
 157         int ret;
 158
 159 #ifndef DISTRUST_SIGNALS_EXTREME
 160         ret = pthread_mutex_lock(mutex);
 161         if (ret)
 162                 abort();
 163 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 164         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 165                 if (ret != EBUSY && ret != EINTR)
 166                         abort();
 167                 poll(NULL,0,10);
 168         }
 169 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 170 }
 171
 172 static void mutex_unlock(pthread_mutex_t *mutex)
 173 {
 174         int ret;
 175
 176         ret = pthread_mutex_unlock(mutex);
 177         if (ret)
 178                 abort();
 179 }
 180
 181 static void smp_mb_master(void)
 182 {
 183         if (caa_likely(lttng_ust_urcu_has_sys_membarrier)) {
 184                 if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0))
 185                         abort();
 186         } else {
 187                 cmm_smp_mb();
 188         }
 189 }
 190
 191 /*
 192  * Always called with rcu_registry lock held. Releases this lock between
 193  * iterations and grabs it again. Holds the lock when it returns.
 194  */
 195 static void wait_for_readers(struct cds_list_head *input_readers,
 196                         struct cds_list_head *cur_snap_readers,
 197                         struct cds_list_head *qsreaders)
 198 {
 199         unsigned int wait_loops = 0;
 200         struct lttng_ust_urcu_reader *index, *tmp;
 201
 202         /*
 203          * Wait for each thread URCU_TLS(lttng_ust_urcu_reader).ctr to either
 204          * indicate quiescence (not nested), or observe the current
 205          * rcu_gp.ctr value.
 206          */
 207         for (;;) {
 208                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 209                         wait_loops++;
 210
 211                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 212                         switch (lttng_ust_urcu_reader_state(&index->ctr)) {
 213                         case LTTNG_UST_URCU_READER_ACTIVE_CURRENT:
 214                                 if (cur_snap_readers) {
 215                                         cds_list_move(&index->node,
 216                                                 cur_snap_readers);
 217                                         break;
 218                                 }
 219                                 /* Fall-through */
 220                         case LTTNG_UST_URCU_READER_INACTIVE:
 221                                 cds_list_move(&index->node, qsreaders);
 222                                 break;
 223                         case LTTNG_UST_URCU_READER_ACTIVE_OLD:
 224                                 /*
 225                                  * Old snapshot. Leaving node in
 226                                  * input_readers will make us busy-loop
 227                                  * until the snapshot becomes current or
 228                                  * the reader becomes inactive.
 229                                  */
 230                                 break;
 231                         }
 232                 }
 233
 234                 if (cds_list_empty(input_readers)) {
 235                         break;
 236                 } else {
 237                         /* Temporarily unlock the registry lock. */
 238                         mutex_unlock(&rcu_registry_lock);
 239                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
 240                                 (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS);
 241                         else
 242                                 caa_cpu_relax();
 243                         /* Re-lock the registry lock before the next loop. */
 244                         mutex_lock(&rcu_registry_lock);
 245                 }
 246         }
 247 }
 248
 249 void lttng_ust_urcu_synchronize_rcu(void)
 250 {
 251         CDS_LIST_HEAD(cur_snap_readers);
 252         CDS_LIST_HEAD(qsreaders);
 253         sigset_t newmask, oldmask;
 254         int ret;
 255
 256         ret = sigfillset(&newmask);
 257         assert(!ret);
 258         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 259         assert(!ret);
 260
 261         mutex_lock(&rcu_gp_lock);
 262
 263         mutex_lock(&rcu_registry_lock);
 264
 265         if (cds_list_empty(&registry))
 266                 goto out;
 267
 268         /* All threads should read qparity before accessing data structure
 269          * where new ptr points to. */
 270         /* Write new ptr before changing the qparity */
 271         smp_mb_master();
 272
 273         /*
 274          * Wait for readers to observe original parity or be quiescent.
 275          * wait_for_readers() can release and grab again rcu_registry_lock
 276          * interally.
 277          */
 278         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 279
 280         /*
 281          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 282          * model easier to understand. It does not have a big performance impact
 283          * anyway, given this is the write-side.
 284          */
 285         cmm_smp_mb();
 286
 287         /* Switch parity: 0 -> 1, 1 -> 0 */
 288         CMM_STORE_SHARED(lttng_ust_urcu_gp.ctr, lttng_ust_urcu_gp.ctr ^ LTTNG_UST_URCU_GP_CTR_PHASE);
 289
 290         /*
 291          * Must commit qparity update to memory before waiting for other parity
 292          * quiescent state. Failure to do so could result in the writer waiting
 293          * forever while new readers are always accessing data (no progress).
 294          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 295          */
 296
 297         /*
 298          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 299          * model easier to understand. It does not have a big performance impact
 300          * anyway, given this is the write-side.
 301          */
 302         cmm_smp_mb();
 303
 304         /*
 305          * Wait for readers to observe new parity or be quiescent.
 306          * wait_for_readers() can release and grab again rcu_registry_lock
 307          * interally.
 308          */
 309         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 310
 311         /*
 312          * Put quiescent reader list back into registry.
 313          */
 314         cds_list_splice(&qsreaders, &registry);
 315
 316         /*
 317          * Finish waiting for reader threads before letting the old ptr being
 318          * freed.
 319          */
 320         smp_mb_master();
 321 out:
 322         mutex_unlock(&rcu_registry_lock);
 323         mutex_unlock(&rcu_gp_lock);
 324         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 325         assert(!ret);
 326 }
 327
 328 /*
 329  * library wrappers to be used by non-LGPL compatible source code.
 330  */
 331
 332 void lttng_ust_urcu_read_lock(void)
 333 {
 334         _lttng_ust_urcu_read_lock();
 335 }
 336
 337 void lttng_ust_urcu_read_unlock(void)
 338 {
 339         _lttng_ust_urcu_read_unlock();
 340 }
 341
 342 int lttng_ust_urcu_read_ongoing(void)
 343 {
 344         return _lttng_ust_urcu_read_ongoing();
 345 }
 346
 347 /*
 348  * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
 349  * Else, try expanding the last chunk. If this fails, allocate a new
 350  * chunk twice as big as the last chunk.
 351  * Memory used by chunks _never_ moves. A chunk could theoretically be
 352  * freed when all "used" slots are released, but we don't do it at this
 353  * point.
 354  */
 355 static
 356 void expand_arena(struct registry_arena *arena)
 357 {
 358         struct registry_chunk *new_chunk, *last_chunk;
 359         size_t old_chunk_len, new_chunk_len;
 360
 361         /* No chunk. */
 362         if (cds_list_empty(&arena->chunk_list)) {
 363                 assert(ARENA_INIT_ALLOC >=
 364                         sizeof(struct registry_chunk)
 365                         + sizeof(struct lttng_ust_urcu_reader));
 366                 new_chunk_len = ARENA_INIT_ALLOC;
 367                 new_chunk = (struct registry_chunk *) mmap(NULL,
 368                         new_chunk_len,
 369                         PROT_READ | PROT_WRITE,
 370                         MAP_ANONYMOUS | MAP_PRIVATE,
 371                         -1, 0);
 372                 if (new_chunk == MAP_FAILED)
 373                         abort();
 374                 memset(new_chunk, 0, new_chunk_len);
 375                 new_chunk->data_len =
 376                         new_chunk_len - sizeof(struct registry_chunk);
 377                 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 378                 return;         /* We're done. */
 379         }
 380
 381         /* Try expanding last chunk. */
 382         last_chunk = cds_list_entry(arena->chunk_list.prev,
 383                 struct registry_chunk, node);
 384         old_chunk_len =
 385                 last_chunk->data_len + sizeof(struct registry_chunk);
 386         new_chunk_len = old_chunk_len << 1;
 387
 388         /* Don't allow memory mapping to move, just expand. */
 389         new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
 390                 new_chunk_len, 0);
 391         if (new_chunk != MAP_FAILED) {
 392                 /* Should not have moved. */
 393                 assert(new_chunk == last_chunk);
 394                 memset((char *) last_chunk + old_chunk_len, 0,
 395                         new_chunk_len - old_chunk_len);
 396                 last_chunk->data_len =
 397                         new_chunk_len - sizeof(struct registry_chunk);
 398                 return;         /* We're done. */
 399         }
 400
 401         /* Remap did not succeed, we need to add a new chunk. */
 402         new_chunk = (struct registry_chunk *) mmap(NULL,
 403                 new_chunk_len,
 404                 PROT_READ | PROT_WRITE,
 405                 MAP_ANONYMOUS | MAP_PRIVATE,
 406                 -1, 0);
 407         if (new_chunk == MAP_FAILED)
 408                 abort();
 409         memset(new_chunk, 0, new_chunk_len);
 410         new_chunk->data_len =
 411                 new_chunk_len - sizeof(struct registry_chunk);
 412         cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 413 }
 414
 415 static
 416 struct lttng_ust_urcu_reader *arena_alloc(struct registry_arena *arena)
 417 {
 418         struct registry_chunk *chunk;
 419         struct lttng_ust_urcu_reader *rcu_reader_reg;
 420         int expand_done = 0;    /* Only allow to expand once per alloc */
 421         size_t len = sizeof(struct lttng_ust_urcu_reader);
 422
 423 retry:
 424         cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
 425                 if (chunk->data_len - chunk->used < len)
 426                         continue;
 427                 /* Find spot */
 428                 for (rcu_reader_reg = (struct lttng_ust_urcu_reader *) &chunk->data[0];
 429                                 rcu_reader_reg < (struct lttng_ust_urcu_reader *) &chunk->data[chunk->data_len];
 430                                 rcu_reader_reg++) {
 431                         if (!rcu_reader_reg->alloc) {
 432                                 rcu_reader_reg->alloc = 1;
 433                                 chunk->used += len;
 434                                 return rcu_reader_reg;
 435                         }
 436                 }
 437         }
 438
 439         if (!expand_done) {
 440                 expand_arena(arena);
 441                 expand_done = 1;
 442                 goto retry;
 443         }
 444
 445         return NULL;
 446 }
 447
 448 /* Called with signals off and mutex locked */
 449 static
 450 void add_thread(void)
 451 {
 452         struct lttng_ust_urcu_reader *rcu_reader_reg;
 453         int ret;
 454
 455         rcu_reader_reg = arena_alloc(&registry_arena);
 456         if (!rcu_reader_reg)
 457                 abort();
 458         ret = pthread_setspecific(lttng_ust_urcu_key, rcu_reader_reg);
 459         if (ret)
 460                 abort();
 461
 462         /* Add to registry */
 463         rcu_reader_reg->tid = pthread_self();
 464         assert(rcu_reader_reg->ctr == 0);
 465         cds_list_add(&rcu_reader_reg->node, &registry);
 466         /*
 467          * Reader threads are pointing to the reader registry. This is
 468          * why its memory should never be relocated.
 469          */
 470         URCU_TLS(lttng_ust_urcu_reader) = rcu_reader_reg;
 471 }
 472
 473 /* Called with mutex locked */
 474 static
 475 void cleanup_thread(struct registry_chunk *chunk,
 476                 struct lttng_ust_urcu_reader *rcu_reader_reg)
 477 {
 478         rcu_reader_reg->ctr = 0;
 479         cds_list_del(&rcu_reader_reg->node);
 480         rcu_reader_reg->tid = 0;
 481         rcu_reader_reg->alloc = 0;
 482         chunk->used -= sizeof(struct lttng_ust_urcu_reader);
 483 }
 484
 485 static
 486 struct registry_chunk *find_chunk(struct lttng_ust_urcu_reader *rcu_reader_reg)
 487 {
 488         struct registry_chunk *chunk;
 489
 490         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 491                 if (rcu_reader_reg < (struct lttng_ust_urcu_reader *) &chunk->data[0])
 492                         continue;
 493                 if (rcu_reader_reg >= (struct lttng_ust_urcu_reader *) &chunk->data[chunk->data_len])
 494                         continue;
 495                 return chunk;
 496         }
 497         return NULL;
 498 }
 499
 500 /* Called with signals off and mutex locked */
 501 static
 502 void remove_thread(struct lttng_ust_urcu_reader *rcu_reader_reg)
 503 {
 504         cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 505         URCU_TLS(lttng_ust_urcu_reader) = NULL;
 506 }
 507
 508 /* Disable signals, take mutex, add to registry */
 509 void lttng_ust_urcu_register(void)
 510 {
 511         sigset_t newmask, oldmask;
 512         int ret;
 513
 514         ret = sigfillset(&newmask);
 515         if (ret)
 516                 abort();
 517         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 518         if (ret)
 519                 abort();
 520
 521         /*
 522          * Check if a signal concurrently registered our thread since
 523          * the check in rcu_read_lock().
 524          */
 525         if (URCU_TLS(lttng_ust_urcu_reader))
 526                 goto end;
 527
 528         /*
 529          * Take care of early registration before lttng_ust_urcu constructor.
 530          */
 531         _lttng_ust_urcu_init();
 532
 533         mutex_lock(&rcu_registry_lock);
 534         add_thread();
 535         mutex_unlock(&rcu_registry_lock);
 536 end:
 537         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 538         if (ret)
 539                 abort();
 540 }
 541
 542 void lttng_ust_urcu_register_thread(void)
 543 {
 544         if (caa_unlikely(!URCU_TLS(lttng_ust_urcu_reader)))
 545                 lttng_ust_urcu_register(); /* If not yet registered. */
 546 }
 547
 548 /* Disable signals, take mutex, remove from registry */
 549 static
 550 void lttng_ust_urcu_unregister(struct lttng_ust_urcu_reader *rcu_reader_reg)
 551 {
 552         sigset_t newmask, oldmask;
 553         int ret;
 554
 555         ret = sigfillset(&newmask);
 556         if (ret)
 557                 abort();
 558         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 559         if (ret)
 560                 abort();
 561
 562         mutex_lock(&rcu_registry_lock);
 563         remove_thread(rcu_reader_reg);
 564         mutex_unlock(&rcu_registry_lock);
 565         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 566         if (ret)
 567                 abort();
 568         lttng_ust_urcu_exit();
 569 }
 570
 571 /*
 572  * Remove thread from the registry when it exits, and flag it as
 573  * destroyed so garbage collection can take care of it.
 574  */
 575 static
 576 void lttng_ust_urcu_thread_exit_notifier(void *rcu_key)
 577 {
 578         lttng_ust_urcu_unregister(rcu_key);
 579 }
 580
 581 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 582 static
 583 void lttng_ust_urcu_sys_membarrier_status(bool available)
 584 {
 585         if (!available)
 586                 abort();
 587 }
 588 #else
 589 static
 590 void lttng_ust_urcu_sys_membarrier_status(bool available)
 591 {
 592         if (!available)
 593                 return;
 594         lttng_ust_urcu_has_sys_membarrier = 1;
 595 }
 596 #endif
 597
 598 static
 599 void lttng_ust_urcu_sys_membarrier_init(void)
 600 {
 601         bool available = false;
 602         int mask;
 603
 604         mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
 605         if (mask >= 0) {
 606                 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
 607                         if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
 608                                 abort();
 609                         available = true;
 610                 }
 611         }
 612         lttng_ust_urcu_sys_membarrier_status(available);
 613 }
 614
 615 static
 616 void _lttng_ust_urcu_init(void)
 617 {
 618         mutex_lock(&init_lock);
 619         if (!lttng_ust_urcu_refcount++) {
 620                 int ret;
 621
 622                 ret = pthread_key_create(&lttng_ust_urcu_key,
 623                                 lttng_ust_urcu_thread_exit_notifier);
 624                 if (ret)
 625                         abort();
 626                 lttng_ust_urcu_sys_membarrier_init();
 627                 initialized = 1;
 628         }
 629         mutex_unlock(&init_lock);
 630 }
 631
 632 static
 633 void lttng_ust_urcu_exit(void)
 634 {
 635         mutex_lock(&init_lock);
 636         if (!--lttng_ust_urcu_refcount) {
 637                 struct registry_chunk *chunk, *tmp;
 638                 int ret;
 639
 640                 cds_list_for_each_entry_safe(chunk, tmp,
 641                                 &registry_arena.chunk_list, node) {
 642                         munmap((void *) chunk, chunk->data_len
 643                                         + sizeof(struct registry_chunk));
 644                 }
 645                 CDS_INIT_LIST_HEAD(&registry_arena.chunk_list);
 646                 ret = pthread_key_delete(lttng_ust_urcu_key);
 647                 if (ret)
 648                         abort();
 649         }
 650         mutex_unlock(&init_lock);
 651 }
 652
 653 /*
 654  * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
 655  * sure we fork() don't race with a concurrent thread executing with
 656  * any of those locks held. This ensures that the registry and data
 657  * protected by rcu_gp_lock are in a coherent state in the child.
 658  */
 659 void lttng_ust_urcu_before_fork(void)
 660 {
 661         sigset_t newmask, oldmask;
 662         int ret;
 663
 664         ret = sigfillset(&newmask);
 665         assert(!ret);
 666         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 667         assert(!ret);
 668         mutex_lock(&rcu_gp_lock);
 669         mutex_lock(&rcu_registry_lock);
 670         saved_fork_signal_mask = oldmask;
 671 }
 672
 673 void lttng_ust_urcu_after_fork_parent(void)
 674 {
 675         sigset_t oldmask;
 676         int ret;
 677
 678         oldmask = saved_fork_signal_mask;
 679         mutex_unlock(&rcu_registry_lock);
 680         mutex_unlock(&rcu_gp_lock);
 681         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 682         assert(!ret);
 683 }
 684
 685 /*
 686  * Prune all entries from registry except our own thread. Fits the Linux
 687  * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
 688  */
 689 static
 690 void lttng_ust_urcu_prune_registry(void)
 691 {
 692         struct registry_chunk *chunk;
 693         struct lttng_ust_urcu_reader *rcu_reader_reg;
 694
 695         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 696                 for (rcu_reader_reg = (struct lttng_ust_urcu_reader *) &chunk->data[0];
 697                                 rcu_reader_reg < (struct lttng_ust_urcu_reader *) &chunk->data[chunk->data_len];
 698                                 rcu_reader_reg++) {
 699                         if (!rcu_reader_reg->alloc)
 700                                 continue;
 701                         if (rcu_reader_reg->tid == pthread_self())
 702                                 continue;
 703                         cleanup_thread(chunk, rcu_reader_reg);
 704                 }
 705         }
 706 }
 707
 708 void lttng_ust_urcu_after_fork_child(void)
 709 {
 710         sigset_t oldmask;
 711         int ret;
 712
 713         lttng_ust_urcu_prune_registry();
 714         oldmask = saved_fork_signal_mask;
 715         mutex_unlock(&rcu_registry_lock);
 716         mutex_unlock(&rcu_gp_lock);
 717         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 718         assert(!ret);
 719 }