liblttng-ust/lttng-ust-comm.c

   1 /*
   2  * lttng-ust-comm.c
   3  *
   4  * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
   5  * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   6  *
   7  * This library is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; only
  10  * version 2.1 of the License.
  11  *
  12  * This library is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with this library; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #define _LGPL_SOURCE
  23 #include <sys/types.h>
  24 #include <sys/socket.h>
  25 #include <sys/prctl.h>
  26 #include <sys/mman.h>
  27 #include <sys/stat.h>
  28 #include <sys/types.h>
  29 #include <sys/wait.h>
  30 #include <fcntl.h>
  31 #include <unistd.h>
  32 #include <errno.h>
  33 #include <pthread.h>
  34 #include <semaphore.h>
  35 #include <time.h>
  36 #include <assert.h>
  37 #include <signal.h>
  38 #include <urcu/uatomic.h>
  39 #include <urcu/futex.h>
  40 #include <urcu/compiler.h>
  41
  42 #include <lttng/ust-events.h>
  43 #include <lttng/ust-abi.h>
  44 #include <lttng/ust.h>
  45 #include <ust-comm.h>
  46 #include <usterr-signal-safe.h>
  47 #include "tracepoint-internal.h"
  48 #include "ltt-tracer-core.h"
  49
  50 /*
  51  * Has lttng ust comm constructor been called ?
  52  */
  53 static int initialized;
  54
  55 /*
  56  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  57  * Held when handling a command, also held by fork() to deal with
  58  * removal of threads, and by exit path.
  59  */
  60
  61 /* Should the ust comm thread quit ? */
  62 static int lttng_ust_comm_should_quit;
  63
  64 /*
  65  * Wait for either of these before continuing to the main
  66  * program:
  67  * - the register_done message from sessiond daemon
  68  *   (will let the sessiond daemon enable sessions before main
  69  *   starts.)
  70  * - sessiond daemon is not reachable.
  71  * - timeout (ensuring applications are resilient to session
  72  *   daemon problems).
  73  */
  74 static sem_t constructor_wait;
  75 /*
  76  * Doing this for both the global and local sessiond.
  77  */
  78 static int sem_count = { 2 };
  79
  80 /*
  81  * Info about socket and associated listener thread.
  82  */
  83 struct sock_info {
  84         const char *name;
  85         pthread_t ust_listener; /* listener thread */
  86         int root_handle;
  87         int constructor_sem_posted;
  88         int allowed;
  89         int global;
  90
  91         char sock_path[PATH_MAX];
  92         int socket;
  93
  94         char wait_shm_path[PATH_MAX];
  95         char *wait_shm_mmap;
  96 };
  97
  98 /* Socket from app (connect) to session daemon (listen) for communication */
  99 struct sock_info global_apps = {
 100         .name = "global",
 101         .global = 1,
 102
 103         .root_handle = -1,
 104         .allowed = 1,
 105
 106         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 107         .socket = -1,
 108
 109         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 110 };
 111
 112 /* TODO: allow global_apps_sock_path override */
 113
 114 struct sock_info local_apps = {
 115         .name = "local",
 116         .global = 0,
 117         .root_handle = -1,
 118         .allowed = 0,   /* Check setuid bit first */
 119
 120         .socket = -1,
 121 };
 122
 123 static int wait_poll_fallback;
 124
 125 extern void ltt_ring_buffer_client_overwrite_init(void);
 126 extern void ltt_ring_buffer_client_discard_init(void);
 127 extern void ltt_ring_buffer_metadata_client_init(void);
 128 extern void ltt_ring_buffer_client_overwrite_exit(void);
 129 extern void ltt_ring_buffer_client_discard_exit(void);
 130 extern void ltt_ring_buffer_metadata_client_exit(void);
 131
 132 static
 133 int setup_local_apps(void)
 134 {
 135         const char *home_dir;
 136         uid_t uid;
 137
 138         uid = getuid();
 139         /*
 140          * Disallow per-user tracing for setuid binaries.
 141          */
 142         if (uid != geteuid()) {
 143                 local_apps.allowed = 0;
 144                 return 0;
 145         } else {
 146                 local_apps.allowed = 1;
 147         }
 148         home_dir = (const char *) getenv("HOME");
 149         if (!home_dir)
 150                 return -ENOENT;
 151         snprintf(local_apps.sock_path, PATH_MAX,
 152                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 153         snprintf(local_apps.wait_shm_path, PATH_MAX,
 154                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 155         return 0;
 156 }
 157
 158 static
 159 int register_app_to_sessiond(int socket)
 160 {
 161         ssize_t ret;
 162         int prctl_ret;
 163         struct {
 164                 uint32_t major;
 165                 uint32_t minor;
 166                 pid_t pid;
 167                 pid_t ppid;
 168                 uid_t uid;
 169                 gid_t gid;
 170                 uint32_t bits_per_long;
 171                 char name[16];  /* process name */
 172         } reg_msg;
 173
 174         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 175         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 176         reg_msg.pid = getpid();
 177         reg_msg.ppid = getppid();
 178         reg_msg.uid = getuid();
 179         reg_msg.gid = getgid();
 180         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 181         prctl_ret = prctl(PR_GET_NAME, (unsigned long) reg_msg.name, 0, 0, 0);
 182         if (prctl_ret) {
 183                 ERR("Error executing prctl");
 184                 return -errno;
 185         }
 186
 187         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 188         if (ret >= 0 && ret != sizeof(reg_msg))
 189                 return -EIO;
 190         return ret;
 191 }
 192
 193 static
 194 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 195 {
 196         ssize_t len;
 197
 198         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 199         switch (len) {
 200         case sizeof(*lur):
 201                 DBG("message successfully sent");
 202                 return 0;
 203         case -1:
 204                 if (errno == ECONNRESET) {
 205                         printf("remote end closed connection\n");
 206                         return 0;
 207                 }
 208                 return -1;
 209         default:
 210                 printf("incorrect message size: %zd\n", len);
 211                 return -1;
 212         }
 213 }
 214
 215 static
 216 int handle_register_done(struct sock_info *sock_info)
 217 {
 218         int ret;
 219
 220         if (sock_info->constructor_sem_posted)
 221                 return 0;
 222         sock_info->constructor_sem_posted = 1;
 223         if (uatomic_read(&sem_count) <= 0) {
 224                 return 0;
 225         }
 226         ret = uatomic_add_return(&sem_count, -1);
 227         if (ret == 0) {
 228                 ret = sem_post(&constructor_wait);
 229                 assert(!ret);
 230         }
 231         return 0;
 232 }
 233
 234 static
 235 int handle_message(struct sock_info *sock_info,
 236                 int sock, struct ustcomm_ust_msg *lum)
 237 {
 238         int ret = 0;
 239         const struct lttng_ust_objd_ops *ops;
 240         struct ustcomm_ust_reply lur;
 241         int shm_fd, wait_fd;
 242         union ust_args args;
 243
 244         ust_lock();
 245
 246         memset(&lur, 0, sizeof(lur));
 247
 248         if (lttng_ust_comm_should_quit) {
 249                 ret = -EPERM;
 250                 goto end;
 251         }
 252
 253         ops = objd_ops(lum->handle);
 254         if (!ops) {
 255                 ret = -ENOENT;
 256                 goto end;
 257         }
 258
 259         switch (lum->cmd) {
 260         case LTTNG_UST_REGISTER_DONE:
 261                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 262                         ret = handle_register_done(sock_info);
 263                 else
 264                         ret = -EINVAL;
 265                 break;
 266         case LTTNG_UST_RELEASE:
 267                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 268                         ret = -EPERM;
 269                 else
 270                         ret = lttng_ust_objd_unref(lum->handle);
 271                 break;
 272         default:
 273                 if (ops->cmd)
 274                         ret = ops->cmd(lum->handle, lum->cmd,
 275                                         (unsigned long) &lum->u,
 276                                         &args);
 277                 else
 278                         ret = -ENOSYS;
 279                 break;
 280         }
 281
 282 end:
 283         lur.handle = lum->handle;
 284         lur.cmd = lum->cmd;
 285         lur.ret_val = ret;
 286         if (ret >= 0) {
 287                 lur.ret_code = USTCOMM_OK;
 288         } else {
 289                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 290                 lur.ret_code = ret;
 291         }
 292         if (ret >= 0) {
 293                 switch (lum->cmd) {
 294                 case LTTNG_UST_STREAM:
 295                         /*
 296                          * Special-case reply to send stream info.
 297                          * Use lum.u output.
 298                          */
 299                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 300                         shm_fd = *args.stream.shm_fd;
 301                         wait_fd = *args.stream.wait_fd;
 302                         break;
 303                 case LTTNG_UST_METADATA:
 304                 case LTTNG_UST_CHANNEL:
 305                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 306                         shm_fd = *args.channel.shm_fd;
 307                         wait_fd = *args.channel.wait_fd;
 308                         break;
 309                 case LTTNG_UST_TRACER_VERSION:
 310                         lur.u.version = lum->u.version;
 311                         break;
 312                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 313                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 314                         break;
 315                 }
 316         }
 317         ret = send_reply(sock, &lur);
 318         if (ret < 0) {
 319                 perror("error sending reply");
 320                 goto error;
 321         }
 322
 323         if ((lum->cmd == LTTNG_UST_STREAM
 324              || lum->cmd == LTTNG_UST_CHANNEL
 325              || lum->cmd == LTTNG_UST_METADATA)
 326                         && lur.ret_code == USTCOMM_OK) {
 327                 /* we also need to send the file descriptors. */
 328                 ret = ustcomm_send_fds_unix_sock(sock,
 329                         &shm_fd, &shm_fd,
 330                         1, sizeof(int));
 331                 if (ret < 0) {
 332                         perror("send shm_fd");
 333                         goto error;
 334                 }
 335                 ret = ustcomm_send_fds_unix_sock(sock,
 336                         &wait_fd, &wait_fd,
 337                         1, sizeof(int));
 338                 if (ret < 0) {
 339                         perror("send wait_fd");
 340                         goto error;
 341                 }
 342         }
 343         /*
 344          * We still have the memory map reference, and the fds have been
 345          * sent to the sessiond. We can therefore close those fds. Note
 346          * that we keep the write side of the wait_fd open, but close
 347          * the read side.
 348          */
 349         if (lur.ret_code == USTCOMM_OK) {
 350                 switch (lum->cmd) {
 351                 case LTTNG_UST_STREAM:
 352                         if (shm_fd >= 0) {
 353                                 ret = close(shm_fd);
 354                                 if (ret) {
 355                                         PERROR("Error closing stream shm_fd");
 356                                 }
 357                                 *args.stream.shm_fd = -1;
 358                         }
 359                         if (wait_fd >= 0) {
 360                                 ret = close(wait_fd);
 361                                 if (ret) {
 362                                         PERROR("Error closing stream wait_fd");
 363                                 }
 364                                 *args.stream.wait_fd = -1;
 365                         }
 366                         break;
 367                 case LTTNG_UST_METADATA:
 368                 case LTTNG_UST_CHANNEL:
 369                         if (shm_fd >= 0) {
 370                                 ret = close(shm_fd);
 371                                 if (ret) {
 372                                         PERROR("Error closing channel shm_fd");
 373                                 }
 374                                 *args.channel.shm_fd = -1;
 375                         }
 376                         if (wait_fd >= 0) {
 377                                 ret = close(wait_fd);
 378                                 if (ret) {
 379                                         PERROR("Error closing channel wait_fd");
 380                                 }
 381                                 *args.channel.wait_fd = -1;
 382                         }
 383                         break;
 384                 }
 385         }
 386
 387 error:
 388         ust_unlock();
 389         return ret;
 390 }
 391
 392 static
 393 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 394 {
 395         int ret;
 396
 397         if (sock_info->socket != -1) {
 398                 ret = close(sock_info->socket);
 399                 if (ret) {
 400                         ERR("Error closing apps socket");
 401                 }
 402                 sock_info->socket = -1;
 403         }
 404         if (sock_info->root_handle != -1) {
 405                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 406                 if (ret) {
 407                         ERR("Error unref root handle");
 408                 }
 409                 sock_info->root_handle = -1;
 410         }
 411         sock_info->constructor_sem_posted = 0;
 412         /*
 413          * wait_shm_mmap is used by listener threads outside of the
 414          * ust lock, so we cannot tear it down ourselves, because we
 415          * cannot join on these threads. Leave this task to the OS
 416          * process exit.
 417          */
 418         if (!exiting && sock_info->wait_shm_mmap) {
 419                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 420                 if (ret) {
 421                         ERR("Error unmapping wait shm");
 422                 }
 423                 sock_info->wait_shm_mmap = NULL;
 424         }
 425 }
 426
 427 /*
 428  * Using fork to set umask in the child process (not multi-thread safe).
 429  * We deal with the shm_open vs ftruncate race (happening when the
 430  * sessiond owns the shm and does not let everybody modify it, to ensure
 431  * safety against shm_unlink) by simply letting the mmap fail and
 432  * retrying after a few seconds.
 433  * For global shm, everybody has rw access to it until the sessiond
 434  * starts.
 435  */
 436 static
 437 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 438 {
 439         int wait_shm_fd, ret;
 440         pid_t pid;
 441
 442         /*
 443          * Try to open read-only.
 444          */
 445         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 446         if (wait_shm_fd >= 0) {
 447                 goto end;
 448         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 449                 /*
 450                  * Real-only open did not work, and it's not because the
 451                  * entry was not present. It's a failure that prohibits
 452                  * using shm.
 453                  */
 454                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 455                 goto end;
 456         }
 457         /*
 458          * If the open failed because the file did not exist, try
 459          * creating it ourself.
 460          */
 461         pid = fork();
 462         if (pid > 0) {
 463                 int status;
 464
 465                 /*
 466                  * Parent: wait for child to return, in which case the
 467                  * shared memory map will have been created.
 468                  */
 469                 pid = wait(&status);
 470                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 471                         wait_shm_fd = -1;
 472                         goto end;
 473                 }
 474                 /*
 475                  * Try to open read-only again after creation.
 476                  */
 477                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 478                 if (wait_shm_fd < 0) {
 479                         /*
 480                          * Real-only open did not work. It's a failure
 481                          * that prohibits using shm.
 482                          */
 483                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 484                         goto end;
 485                 }
 486                 goto end;
 487         } else if (pid == 0) {
 488                 int create_mode;
 489
 490                 /* Child */
 491                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 492                 if (sock_info->global)
 493                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 494                 /*
 495                  * We're alone in a child process, so we can modify the
 496                  * process-wide umask.
 497                  */
 498                 umask(~create_mode);
 499                 /*
 500                  * Try creating shm (or get rw access).
 501                  * We don't do an exclusive open, because we allow other
 502                  * processes to create+ftruncate it concurrently.
 503                  */
 504                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 505                                 O_RDWR | O_CREAT, create_mode);
 506                 if (wait_shm_fd >= 0) {
 507                         ret = ftruncate(wait_shm_fd, mmap_size);
 508                         if (ret) {
 509                                 PERROR("ftruncate");
 510                                 exit(EXIT_FAILURE);
 511                         }
 512                         exit(EXIT_SUCCESS);
 513                 }
 514                 /*
 515                  * For local shm, we need to have rw access to accept
 516                  * opening it: this means the local sessiond will be
 517                  * able to wake us up. For global shm, we open it even
 518                  * if rw access is not granted, because the root.root
 519                  * sessiond will be able to override all rights and wake
 520                  * us up.
 521                  */
 522                 if (!sock_info->global && errno != EACCES) {
 523                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 524                         exit(EXIT_FAILURE);
 525                 }
 526                 /*
 527                  * The shm exists, but we cannot open it RW. Report
 528                  * success.
 529                  */
 530                 exit(EXIT_SUCCESS);
 531         } else {
 532                 return -1;
 533         }
 534 end:
 535         if (wait_shm_fd >= 0 && !sock_info->global) {
 536                 struct stat statbuf;
 537
 538                 /*
 539                  * Ensure that our user is the owner of the shm file for
 540                  * local shm. If we do not own the file, it means our
 541                  * sessiond will not have access to wake us up (there is
 542                  * probably a rogue process trying to fake our
 543                  * sessiond). Fallback to polling method in this case.
 544                  */
 545                 ret = fstat(wait_shm_fd, &statbuf);
 546                 if (ret) {
 547                         PERROR("fstat");
 548                         goto error_close;
 549                 }
 550                 if (statbuf.st_uid != getuid())
 551                         goto error_close;
 552         }
 553         return wait_shm_fd;
 554
 555 error_close:
 556         ret = close(wait_shm_fd);
 557         if (ret) {
 558                 PERROR("Error closing fd");
 559         }
 560         return -1;
 561 }
 562
 563 static
 564 char *get_map_shm(struct sock_info *sock_info)
 565 {
 566         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 567         int wait_shm_fd, ret;
 568         char *wait_shm_mmap;
 569
 570         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 571         if (wait_shm_fd < 0) {
 572                 goto error;
 573         }
 574         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 575                   MAP_SHARED, wait_shm_fd, 0);
 576         /* close shm fd immediately after taking the mmap reference */
 577         ret = close(wait_shm_fd);
 578         if (ret) {
 579                 PERROR("Error closing fd");
 580         }
 581         if (wait_shm_mmap == MAP_FAILED) {
 582                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 583                 goto error;
 584         }
 585         return wait_shm_mmap;
 586
 587 error:
 588         return NULL;
 589 }
 590
 591 static
 592 void wait_for_sessiond(struct sock_info *sock_info)
 593 {
 594         int ret;
 595
 596         ust_lock();
 597         if (lttng_ust_comm_should_quit) {
 598                 goto quit;
 599         }
 600         if (wait_poll_fallback) {
 601                 goto error;
 602         }
 603         if (!sock_info->wait_shm_mmap) {
 604                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 605                 if (!sock_info->wait_shm_mmap)
 606                         goto error;
 607         }
 608         ust_unlock();
 609
 610         DBG("Waiting for %s apps sessiond", sock_info->name);
 611         /* Wait for futex wakeup */
 612         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 613                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 614                         FUTEX_WAIT, 0, NULL, NULL, 0);
 615                 if (ret < 0) {
 616                         if (errno == EFAULT) {
 617                                 wait_poll_fallback = 1;
 618                                 DBG(
 619 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 620 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 621 "Please upgrade your kernel "
 622 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 623 "mainline). LTTng-UST will use polling mode fallback.");
 624                                 if (ust_debug())
 625                                         PERROR("futex");
 626                         }
 627                 }
 628         }
 629         return;
 630
 631 quit:
 632         ust_unlock();
 633         return;
 634
 635 error:
 636         ust_unlock();
 637         return;
 638 }
 639
 640 /*
 641  * This thread does not allocate any resource, except within
 642  * handle_message, within mutex protection. This mutex protects against
 643  * fork and exit.
 644  * The other moment it allocates resources is at socket connexion, which
 645  * is also protected by the mutex.
 646  */
 647 static
 648 void *ust_listener_thread(void *arg)
 649 {
 650         struct sock_info *sock_info = arg;
 651         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 652
 653         /* Restart trying to connect to the session daemon */
 654 restart:
 655         if (prev_connect_failed) {
 656                 /* Wait for sessiond availability with pipe */
 657                 wait_for_sessiond(sock_info);
 658                 if (has_waited) {
 659                         has_waited = 0;
 660                         /*
 661                          * Sleep for 5 seconds before retrying after a
 662                          * sequence of failure / wait / failure. This
 663                          * deals with a killed or broken session daemon.
 664                          */
 665                         sleep(5);
 666                 }
 667                 has_waited = 1;
 668                 prev_connect_failed = 0;
 669         }
 670         ust_lock();
 671
 672         if (lttng_ust_comm_should_quit) {
 673                 ust_unlock();
 674                 goto quit;
 675         }
 676
 677         if (sock_info->socket != -1) {
 678                 ret = close(sock_info->socket);
 679                 if (ret) {
 680                         ERR("Error closing %s apps socket", sock_info->name);
 681                 }
 682                 sock_info->socket = -1;
 683         }
 684
 685         /* Register */
 686         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 687         if (ret < 0) {
 688                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 689                 prev_connect_failed = 1;
 690                 /*
 691                  * If we cannot find the sessiond daemon, don't delay
 692                  * constructor execution.
 693                  */
 694                 ret = handle_register_done(sock_info);
 695                 assert(!ret);
 696                 ust_unlock();
 697                 goto restart;
 698         }
 699
 700         sock_info->socket = sock = ret;
 701
 702         /*
 703          * Create only one root handle per listener thread for the whole
 704          * process lifetime.
 705          */
 706         if (sock_info->root_handle == -1) {
 707                 ret = lttng_abi_create_root_handle();
 708                 if (ret < 0) {
 709                         ERR("Error creating root handle");
 710                         ust_unlock();
 711                         goto quit;
 712                 }
 713                 sock_info->root_handle = ret;
 714         }
 715
 716         ret = register_app_to_sessiond(sock);
 717         if (ret < 0) {
 718                 ERR("Error registering to %s apps socket", sock_info->name);
 719                 prev_connect_failed = 1;
 720                 /*
 721                  * If we cannot register to the sessiond daemon, don't
 722                  * delay constructor execution.
 723                  */
 724                 ret = handle_register_done(sock_info);
 725                 assert(!ret);
 726                 ust_unlock();
 727                 goto restart;
 728         }
 729         ust_unlock();
 730
 731         for (;;) {
 732                 ssize_t len;
 733                 struct ustcomm_ust_msg lum;
 734
 735                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 736                 switch (len) {
 737                 case 0: /* orderly shutdown */
 738                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 739                         ust_lock();
 740                         /*
 741                          * Either sessiond has shutdown or refused us by closing the socket.
 742                          * In either case, we don't want to delay construction execution,
 743                          * and we need to wait before retry.
 744                          */
 745                         prev_connect_failed = 1;
 746                         /*
 747                          * If we cannot register to the sessiond daemon, don't
 748                          * delay constructor execution.
 749                          */
 750                         ret = handle_register_done(sock_info);
 751                         assert(!ret);
 752                         ust_unlock();
 753                         goto end;
 754                 case sizeof(lum):
 755                         DBG("message received\n");
 756                         ret = handle_message(sock_info, sock, &lum);
 757                         if (ret < 0) {
 758                                 ERR("Error handling message for %s socket", sock_info->name);
 759                         }
 760                         continue;
 761                 case -1:
 762                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 763                         if (errno == ECONNRESET) {
 764                                 ERR("%s remote end closed connection\n", sock_info->name);
 765                                 goto end;
 766                         }
 767                         goto end;
 768                 default:
 769                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 770                         continue;
 771                 }
 772
 773         }
 774 end:
 775         goto restart;   /* try to reconnect */
 776 quit:
 777         return NULL;
 778 }
 779
 780 /*
 781  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 782  */
 783 static
 784 int get_timeout(struct timespec *constructor_timeout)
 785 {
 786         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 787         char *str_delay;
 788         int ret;
 789
 790         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 791         if (str_delay) {
 792                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 793         }
 794
 795         switch (constructor_delay_ms) {
 796         case -1:/* fall-through */
 797         case 0:
 798                 return constructor_delay_ms;
 799         default:
 800                 break;
 801         }
 802
 803         /*
 804          * If we are unable to find the current time, don't wait.
 805          */
 806         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 807         if (ret) {
 808                 return -1;
 809         }
 810         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 811         constructor_timeout->tv_nsec +=
 812                 (constructor_delay_ms % 1000UL) * 1000000UL;
 813         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 814                 constructor_timeout->tv_sec++;
 815                 constructor_timeout->tv_nsec -= 1000000000UL;
 816         }
 817         return 1;
 818 }
 819
 820 /*
 821  * sessiond monitoring thread: monitor presence of global and per-user
 822  * sessiond by polling the application common named pipe.
 823  */
 824 /* TODO */
 825
 826 void __attribute__((constructor)) lttng_ust_init(void)
 827 {
 828         struct timespec constructor_timeout;
 829         int timeout_mode;
 830         int ret;
 831
 832         if (uatomic_xchg(&initialized, 1) == 1)
 833                 return;
 834
 835         /*
 836          * We want precise control over the order in which we construct
 837          * our sub-libraries vs starting to receive commands from
 838          * sessiond (otherwise leading to errors when trying to create
 839          * sessiond before the init functions are completed).
 840          */
 841         init_usterr();
 842         init_tracepoint();
 843         ltt_ring_buffer_metadata_client_init();
 844         ltt_ring_buffer_client_overwrite_init();
 845         ltt_ring_buffer_client_discard_init();
 846
 847         timeout_mode = get_timeout(&constructor_timeout);
 848
 849         ret = sem_init(&constructor_wait, 0, 0);
 850         assert(!ret);
 851
 852         ret = setup_local_apps();
 853         if (ret) {
 854                 ERR("Error setting up to local apps");
 855         }
 856         ret = pthread_create(&local_apps.ust_listener, NULL,
 857                         ust_listener_thread, &local_apps);
 858
 859         if (local_apps.allowed) {
 860                 ret = pthread_create(&global_apps.ust_listener, NULL,
 861                                 ust_listener_thread, &global_apps);
 862         } else {
 863                 handle_register_done(&local_apps);
 864         }
 865
 866         switch (timeout_mode) {
 867         case 1: /* timeout wait */
 868                 do {
 869                         ret = sem_timedwait(&constructor_wait,
 870                                         &constructor_timeout);
 871                 } while (ret < 0 && errno == EINTR);
 872                 if (ret < 0 && errno == ETIMEDOUT) {
 873                         ERR("Timed out waiting for ltt-sessiond");
 874                 } else {
 875                         assert(!ret);
 876                 }
 877                 break;
 878         case -1:/* wait forever */
 879                 do {
 880                         ret = sem_wait(&constructor_wait);
 881                 } while (ret < 0 && errno == EINTR);
 882                 assert(!ret);
 883                 break;
 884         case 0: /* no timeout */
 885                 break;
 886         }
 887 }
 888
 889 static
 890 void lttng_ust_cleanup(int exiting)
 891 {
 892         cleanup_sock_info(&global_apps, exiting);
 893         if (local_apps.allowed) {
 894                 cleanup_sock_info(&local_apps, exiting);
 895         }
 896         /*
 897          * The teardown in this function all affect data structures
 898          * accessed under the UST lock by the listener thread. This
 899          * lock, along with the lttng_ust_comm_should_quit flag, ensure
 900          * that none of these threads are accessing this data at this
 901          * point.
 902          */
 903         lttng_ust_abi_exit();
 904         lttng_ust_events_exit();
 905         ltt_ring_buffer_client_discard_exit();
 906         ltt_ring_buffer_client_overwrite_exit();
 907         ltt_ring_buffer_metadata_client_exit();
 908         exit_tracepoint();
 909         if (!exiting) {
 910                 /* Reinitialize values for fork */
 911                 sem_count = 2;
 912                 lttng_ust_comm_should_quit = 0;
 913                 initialized = 0;
 914         }
 915 }
 916
 917 void __attribute__((destructor)) lttng_ust_exit(void)
 918 {
 919         int ret;
 920
 921         /*
 922          * Using pthread_cancel here because:
 923          * A) we don't want to hang application teardown.
 924          * B) the thread is not allocating any resource.
 925          */
 926
 927         /*
 928          * Require the communication thread to quit. Synchronize with
 929          * mutexes to ensure it is not in a mutex critical section when
 930          * pthread_cancel is later called.
 931          */
 932         ust_lock();
 933         lttng_ust_comm_should_quit = 1;
 934         ust_unlock();
 935
 936         /* cancel threads */
 937         ret = pthread_cancel(global_apps.ust_listener);
 938         if (ret) {
 939                 ERR("Error cancelling global ust listener thread");
 940         }
 941         if (local_apps.allowed) {
 942                 ret = pthread_cancel(local_apps.ust_listener);
 943                 if (ret) {
 944                         ERR("Error cancelling local ust listener thread");
 945                 }
 946         }
 947         /*
 948          * Do NOT join threads: use of sys_futex makes it impossible to
 949          * join the threads without using async-cancel, but async-cancel
 950          * is delivered by a signal, which could hit the target thread
 951          * anywhere in its code path, including while the ust_lock() is
 952          * held, causing a deadlock for the other thread. Let the OS
 953          * cleanup the threads if there are stalled in a syscall.
 954          */
 955         lttng_ust_cleanup(1);
 956 }
 957
 958 /*
 959  * We exclude the worker threads across fork and clone (except
 960  * CLONE_VM), because these system calls only keep the forking thread
 961  * running in the child.  Therefore, we don't want to call fork or clone
 962  * in the middle of an tracepoint or ust tracing state modification.
 963  * Holding this mutex protects these structures across fork and clone.
 964  */
 965 void ust_before_fork(sigset_t *save_sigset)
 966 {
 967         /*
 968          * Disable signals. This is to avoid that the child intervenes
 969          * before it is properly setup for tracing. It is safer to
 970          * disable all signals, because then we know we are not breaking
 971          * anything by restoring the original mask.
 972          */
 973         sigset_t all_sigs;
 974         int ret;
 975
 976         /* Disable signals */
 977         sigfillset(&all_sigs);
 978         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
 979         if (ret == -1) {
 980                 PERROR("sigprocmask");
 981         }
 982         ust_lock();
 983         rcu_bp_before_fork();
 984 }
 985
 986 static void ust_after_fork_common(sigset_t *restore_sigset)
 987 {
 988         int ret;
 989
 990         DBG("process %d", getpid());
 991         ust_unlock();
 992         /* Restore signals */
 993         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
 994         if (ret == -1) {
 995                 PERROR("sigprocmask");
 996         }
 997 }
 998
 999 void ust_after_fork_parent(sigset_t *restore_sigset)
1000 {
1001         DBG("process %d", getpid());
1002         rcu_bp_after_fork_parent();
1003         /* Release mutexes and reenable signals */
1004         ust_after_fork_common(restore_sigset);
1005 }
1006
1007 /*
1008  * After fork, in the child, we need to cleanup all the leftover state,
1009  * except the worker thread which already magically disappeared thanks
1010  * to the weird Linux fork semantics. After tyding up, we call
1011  * lttng_ust_init() again to start over as a new PID.
1012  *
1013  * This is meant for forks() that have tracing in the child between the
1014  * fork and following exec call (if there is any).
1015  */
1016 void ust_after_fork_child(sigset_t *restore_sigset)
1017 {
1018         DBG("process %d", getpid());
1019         /* Release urcu mutexes */
1020         rcu_bp_after_fork_child();
1021         lttng_ust_cleanup(0);
1022         lttng_context_vtid_reset();
1023         /* Release mutexes and reenable signals */
1024         ust_after_fork_common(restore_sigset);
1025         lttng_ust_init();
1026 }