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
 243         ust_lock();
 244
 245         memset(&lur, 0, sizeof(lur));
 246
 247         if (lttng_ust_comm_should_quit) {
 248                 ret = -EPERM;
 249                 goto end;
 250         }
 251
 252         ops = objd_ops(lum->handle);
 253         if (!ops) {
 254                 ret = -ENOENT;
 255                 goto end;
 256         }
 257
 258         switch (lum->cmd) {
 259         case LTTNG_UST_REGISTER_DONE:
 260                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 261                         ret = handle_register_done(sock_info);
 262                 else
 263                         ret = -EINVAL;
 264                 break;
 265         case LTTNG_UST_RELEASE:
 266                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 267                         ret = -EPERM;
 268                 else
 269                         ret = lttng_ust_objd_unref(lum->handle);
 270                 break;
 271         default:
 272                 if (ops->cmd)
 273                         ret = ops->cmd(lum->handle, lum->cmd,
 274                                         (unsigned long) &lum->u);
 275                 else
 276                         ret = -ENOSYS;
 277                 break;
 278         }
 279
 280 end:
 281         lur.handle = lum->handle;
 282         lur.cmd = lum->cmd;
 283         lur.ret_val = ret;
 284         if (ret >= 0) {
 285                 lur.ret_code = USTCOMM_OK;
 286         } else {
 287                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 288                 lur.ret_code = ret;
 289         }
 290         switch (lum->cmd) {
 291         case LTTNG_UST_STREAM:
 292                 /*
 293                  * Special-case reply to send stream info.
 294                  * Use lum.u output.
 295                  */
 296                 lur.u.stream.memory_map_size = lum->u.stream.memory_map_size;
 297                 shm_fd = lum->u.stream.shm_fd;
 298                 wait_fd = lum->u.stream.wait_fd;
 299                 break;
 300         case LTTNG_UST_METADATA:
 301         case LTTNG_UST_CHANNEL:
 302                 lur.u.channel.memory_map_size = lum->u.channel.memory_map_size;
 303                 shm_fd = lum->u.channel.shm_fd;
 304                 wait_fd = lum->u.channel.wait_fd;
 305                 break;
 306         case LTTNG_UST_TRACER_VERSION:
 307                 lur.u.version = lum->u.version;
 308                 break;
 309         case LTTNG_UST_TRACEPOINT_LIST_GET:
 310                 memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 311                 break;
 312         }
 313         ret = send_reply(sock, &lur);
 314         if (ret < 0) {
 315                 perror("error sending reply");
 316                 goto error;
 317         }
 318
 319         if ((lum->cmd == LTTNG_UST_STREAM
 320              || lum->cmd == LTTNG_UST_CHANNEL
 321              || lum->cmd == LTTNG_UST_METADATA)
 322                         && lur.ret_code == USTCOMM_OK) {
 323                 /* we also need to send the file descriptors. */
 324                 ret = ustcomm_send_fds_unix_sock(sock,
 325                         &shm_fd, &shm_fd,
 326                         1, sizeof(int));
 327                 if (ret < 0) {
 328                         perror("send shm_fd");
 329                         goto error;
 330                 }
 331                 ret = ustcomm_send_fds_unix_sock(sock,
 332                         &wait_fd, &wait_fd,
 333                         1, sizeof(int));
 334                 if (ret < 0) {
 335                         perror("send wait_fd");
 336                         goto error;
 337                 }
 338         }
 339 error:
 340         ust_unlock();
 341         return ret;
 342 }
 343
 344 static
 345 void cleanup_sock_info(struct sock_info *sock_info)
 346 {
 347         int ret;
 348
 349         if (sock_info->socket != -1) {
 350                 ret = close(sock_info->socket);
 351                 if (ret) {
 352                         ERR("Error closing apps socket");
 353                 }
 354                 sock_info->socket = -1;
 355         }
 356         if (sock_info->root_handle != -1) {
 357                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 358                 if (ret) {
 359                         ERR("Error unref root handle");
 360                 }
 361                 sock_info->root_handle = -1;
 362         }
 363         sock_info->constructor_sem_posted = 0;
 364         if (sock_info->wait_shm_mmap) {
 365                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 366                 if (ret) {
 367                         ERR("Error unmapping wait shm");
 368                 }
 369                 sock_info->wait_shm_mmap = NULL;
 370         }
 371 }
 372
 373 /*
 374  * Using fork to set umask in the child process (not multi-thread safe).
 375  * We deal with the shm_open vs ftruncate race (happening when the
 376  * sessiond owns the shm and does not let everybody modify it, to ensure
 377  * safety against shm_unlink) by simply letting the mmap fail and
 378  * retrying after a few seconds.
 379  * For global shm, everybody has rw access to it until the sessiond
 380  * starts.
 381  */
 382 static
 383 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 384 {
 385         int wait_shm_fd, ret;
 386         pid_t pid;
 387
 388         /*
 389          * Try to open read-only.
 390          */
 391         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 392         if (wait_shm_fd >= 0) {
 393                 goto end;
 394         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 395                 /*
 396                  * Real-only open did not work, and it's not because the
 397                  * entry was not present. It's a failure that prohibits
 398                  * using shm.
 399                  */
 400                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 401                 goto end;
 402         }
 403         /*
 404          * If the open failed because the file did not exist, try
 405          * creating it ourself.
 406          */
 407         pid = fork();
 408         if (pid > 0) {
 409                 int status;
 410
 411                 /*
 412                  * Parent: wait for child to return, in which case the
 413                  * shared memory map will have been created.
 414                  */
 415                 pid = wait(&status);
 416                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 417                         wait_shm_fd = -1;
 418                         goto end;
 419                 }
 420                 /*
 421                  * Try to open read-only again after creation.
 422                  */
 423                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 424                 if (wait_shm_fd < 0) {
 425                         /*
 426                          * Real-only open did not work. It's a failure
 427                          * that prohibits using shm.
 428                          */
 429                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 430                         goto end;
 431                 }
 432                 goto end;
 433         } else if (pid == 0) {
 434                 int create_mode;
 435
 436                 /* Child */
 437                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 438                 if (sock_info->global)
 439                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 440                 /*
 441                  * We're alone in a child process, so we can modify the
 442                  * process-wide umask.
 443                  */
 444                 umask(~create_mode);
 445                 /*
 446                  * Try creating shm (or get rw access).
 447                  * We don't do an exclusive open, because we allow other
 448                  * processes to create+ftruncate it concurrently.
 449                  */
 450                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 451                                 O_RDWR | O_CREAT, create_mode);
 452                 if (wait_shm_fd >= 0) {
 453                         ret = ftruncate(wait_shm_fd, mmap_size);
 454                         if (ret) {
 455                                 PERROR("ftruncate");
 456                                 exit(EXIT_FAILURE);
 457                         }
 458                         exit(EXIT_SUCCESS);
 459                 }
 460                 /*
 461                  * For local shm, we need to have rw access to accept
 462                  * opening it: this means the local sessiond will be
 463                  * able to wake us up. For global shm, we open it even
 464                  * if rw access is not granted, because the root.root
 465                  * sessiond will be able to override all rights and wake
 466                  * us up.
 467                  */
 468                 if (!sock_info->global && errno != EACCES) {
 469                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 470                         exit(EXIT_FAILURE);
 471                 }
 472                 /*
 473                  * The shm exists, but we cannot open it RW. Report
 474                  * success.
 475                  */
 476                 exit(EXIT_SUCCESS);
 477         } else {
 478                 return -1;
 479         }
 480 end:
 481         if (wait_shm_fd >= 0 && !sock_info->global) {
 482                 struct stat statbuf;
 483
 484                 /*
 485                  * Ensure that our user is the owner of the shm file for
 486                  * local shm. If we do not own the file, it means our
 487                  * sessiond will not have access to wake us up (there is
 488                  * probably a rogue process trying to fake our
 489                  * sessiond). Fallback to polling method in this case.
 490                  */
 491                 ret = fstat(wait_shm_fd, &statbuf);
 492                 if (ret) {
 493                         PERROR("fstat");
 494                         goto error_close;
 495                 }
 496                 if (statbuf.st_uid != getuid())
 497                         goto error_close;
 498         }
 499         return wait_shm_fd;
 500
 501 error_close:
 502         ret = close(wait_shm_fd);
 503         if (ret) {
 504                 PERROR("Error closing fd");
 505         }
 506         return -1;
 507 }
 508
 509 static
 510 char *get_map_shm(struct sock_info *sock_info)
 511 {
 512         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 513         int wait_shm_fd, ret;
 514         char *wait_shm_mmap;
 515
 516         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 517         if (wait_shm_fd < 0) {
 518                 goto error;
 519         }
 520         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 521                   MAP_SHARED, wait_shm_fd, 0);
 522         /* close shm fd immediately after taking the mmap reference */
 523         ret = close(wait_shm_fd);
 524         if (ret) {
 525                 PERROR("Error closing fd");
 526         }
 527         if (wait_shm_mmap == MAP_FAILED) {
 528                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 529                 goto error;
 530         }
 531         return wait_shm_mmap;
 532
 533 error:
 534         return NULL;
 535 }
 536
 537 static
 538 void wait_for_sessiond(struct sock_info *sock_info)
 539 {
 540         int ret;
 541
 542         ust_lock();
 543         if (lttng_ust_comm_should_quit) {
 544                 goto quit;
 545         }
 546         if (wait_poll_fallback) {
 547                 goto error;
 548         }
 549         if (!sock_info->wait_shm_mmap) {
 550                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 551                 if (!sock_info->wait_shm_mmap)
 552                         goto error;
 553         }
 554         ust_unlock();
 555
 556         DBG("Waiting for %s apps sessiond", sock_info->name);
 557         /* Wait for futex wakeup */
 558         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 559                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 560                         FUTEX_WAIT, 0, NULL, NULL, 0);
 561                 if (ret < 0) {
 562                         if (errno == EFAULT) {
 563                                 wait_poll_fallback = 1;
 564                                 DBG(
 565 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 566 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 567 "Please upgrade your kernel "
 568 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 569 "mainline). LTTng-UST will use polling mode fallback.");
 570                                 if (ust_debug())
 571                                         PERROR("futex");
 572                         }
 573                 }
 574         }
 575         return;
 576
 577 quit:
 578         ust_unlock();
 579         return;
 580
 581 error:
 582         ust_unlock();
 583         return;
 584 }
 585
 586 /*
 587  * This thread does not allocate any resource, except within
 588  * handle_message, within mutex protection. This mutex protects against
 589  * fork and exit.
 590  * The other moment it allocates resources is at socket connexion, which
 591  * is also protected by the mutex.
 592  */
 593 static
 594 void *ust_listener_thread(void *arg)
 595 {
 596         struct sock_info *sock_info = arg;
 597         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 598
 599         /* Restart trying to connect to the session daemon */
 600 restart:
 601         if (prev_connect_failed) {
 602                 /* Wait for sessiond availability with pipe */
 603                 wait_for_sessiond(sock_info);
 604                 if (has_waited) {
 605                         has_waited = 0;
 606                         /*
 607                          * Sleep for 5 seconds before retrying after a
 608                          * sequence of failure / wait / failure. This
 609                          * deals with a killed or broken session daemon.
 610                          */
 611                         sleep(5);
 612                 }
 613                 has_waited = 1;
 614                 prev_connect_failed = 0;
 615         }
 616         ust_lock();
 617
 618         if (lttng_ust_comm_should_quit) {
 619                 ust_unlock();
 620                 goto quit;
 621         }
 622
 623         if (sock_info->socket != -1) {
 624                 ret = close(sock_info->socket);
 625                 if (ret) {
 626                         ERR("Error closing %s apps socket", sock_info->name);
 627                 }
 628                 sock_info->socket = -1;
 629         }
 630
 631         /* Register */
 632         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 633         if (ret < 0) {
 634                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 635                 prev_connect_failed = 1;
 636                 /*
 637                  * If we cannot find the sessiond daemon, don't delay
 638                  * constructor execution.
 639                  */
 640                 ret = handle_register_done(sock_info);
 641                 assert(!ret);
 642                 ust_unlock();
 643                 goto restart;
 644         }
 645
 646         sock_info->socket = sock = ret;
 647
 648         /*
 649          * Create only one root handle per listener thread for the whole
 650          * process lifetime.
 651          */
 652         if (sock_info->root_handle == -1) {
 653                 ret = lttng_abi_create_root_handle();
 654                 if (ret < 0) {
 655                         ERR("Error creating root handle");
 656                         ust_unlock();
 657                         goto quit;
 658                 }
 659                 sock_info->root_handle = ret;
 660         }
 661
 662         ret = register_app_to_sessiond(sock);
 663         if (ret < 0) {
 664                 ERR("Error registering to %s apps socket", sock_info->name);
 665                 prev_connect_failed = 1;
 666                 /*
 667                  * If we cannot register to the sessiond daemon, don't
 668                  * delay constructor execution.
 669                  */
 670                 ret = handle_register_done(sock_info);
 671                 assert(!ret);
 672                 ust_unlock();
 673                 goto restart;
 674         }
 675         ust_unlock();
 676
 677         for (;;) {
 678                 ssize_t len;
 679                 struct ustcomm_ust_msg lum;
 680
 681                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 682                 switch (len) {
 683                 case 0: /* orderly shutdown */
 684                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 685                         ust_lock();
 686                         /*
 687                          * Either sessiond has shutdown or refused us by closing the socket.
 688                          * In either case, we don't want to delay construction execution,
 689                          * and we need to wait before retry.
 690                          */
 691                         prev_connect_failed = 1;
 692                         /*
 693                          * If we cannot register to the sessiond daemon, don't
 694                          * delay constructor execution.
 695                          */
 696                         ret = handle_register_done(sock_info);
 697                         assert(!ret);
 698                         ust_unlock();
 699                         goto end;
 700                 case sizeof(lum):
 701                         DBG("message received\n");
 702                         ret = handle_message(sock_info, sock, &lum);
 703                         if (ret < 0) {
 704                                 ERR("Error handling message for %s socket", sock_info->name);
 705                         }
 706                         continue;
 707                 case -1:
 708                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 709                         if (errno == ECONNRESET) {
 710                                 ERR("%s remote end closed connection\n", sock_info->name);
 711                                 goto end;
 712                         }
 713                         goto end;
 714                 default:
 715                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 716                         continue;
 717                 }
 718
 719         }
 720 end:
 721         goto restart;   /* try to reconnect */
 722 quit:
 723         return NULL;
 724 }
 725
 726 /*
 727  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 728  */
 729 static
 730 int get_timeout(struct timespec *constructor_timeout)
 731 {
 732         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 733         char *str_delay;
 734         int ret;
 735
 736         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 737         if (str_delay) {
 738                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 739         }
 740
 741         switch (constructor_delay_ms) {
 742         case -1:/* fall-through */
 743         case 0:
 744                 return constructor_delay_ms;
 745         default:
 746                 break;
 747         }
 748
 749         /*
 750          * If we are unable to find the current time, don't wait.
 751          */
 752         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 753         if (ret) {
 754                 return -1;
 755         }
 756         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 757         constructor_timeout->tv_nsec +=
 758                 (constructor_delay_ms % 1000UL) * 1000000UL;
 759         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 760                 constructor_timeout->tv_sec++;
 761                 constructor_timeout->tv_nsec -= 1000000000UL;
 762         }
 763         return 1;
 764 }
 765
 766 /*
 767  * sessiond monitoring thread: monitor presence of global and per-user
 768  * sessiond by polling the application common named pipe.
 769  */
 770 /* TODO */
 771
 772 void __attribute__((constructor)) lttng_ust_init(void)
 773 {
 774         struct timespec constructor_timeout;
 775         int timeout_mode;
 776         int ret;
 777
 778         if (uatomic_xchg(&initialized, 1) == 1)
 779                 return;
 780
 781         /*
 782          * We want precise control over the order in which we construct
 783          * our sub-libraries vs starting to receive commands from
 784          * sessiond (otherwise leading to errors when trying to create
 785          * sessiond before the init functions are completed).
 786          */
 787         init_usterr();
 788         init_tracepoint();
 789         ltt_ring_buffer_metadata_client_init();
 790         ltt_ring_buffer_client_overwrite_init();
 791         ltt_ring_buffer_client_discard_init();
 792
 793         timeout_mode = get_timeout(&constructor_timeout);
 794
 795         ret = sem_init(&constructor_wait, 0, 0);
 796         assert(!ret);
 797
 798         ret = setup_local_apps();
 799         if (ret) {
 800                 ERR("Error setting up to local apps");
 801         }
 802         ret = pthread_create(&local_apps.ust_listener, NULL,
 803                         ust_listener_thread, &local_apps);
 804
 805         if (local_apps.allowed) {
 806                 ret = pthread_create(&global_apps.ust_listener, NULL,
 807                                 ust_listener_thread, &global_apps);
 808         } else {
 809                 handle_register_done(&local_apps);
 810         }
 811
 812         switch (timeout_mode) {
 813         case 1: /* timeout wait */
 814                 do {
 815                         ret = sem_timedwait(&constructor_wait,
 816                                         &constructor_timeout);
 817                 } while (ret < 0 && errno == EINTR);
 818                 if (ret < 0 && errno == ETIMEDOUT) {
 819                         ERR("Timed out waiting for ltt-sessiond");
 820                 } else {
 821                         assert(!ret);
 822                 }
 823                 break;
 824         case -1:/* wait forever */
 825                 do {
 826                         ret = sem_wait(&constructor_wait);
 827                 } while (ret < 0 && errno == EINTR);
 828                 assert(!ret);
 829                 break;
 830         case 0: /* no timeout */
 831                 break;
 832         }
 833 }
 834
 835 static
 836 void lttng_ust_cleanup(int exiting)
 837 {
 838         cleanup_sock_info(&global_apps);
 839         if (local_apps.allowed) {
 840                 cleanup_sock_info(&local_apps);
 841         }
 842         lttng_ust_abi_exit();
 843         lttng_ust_events_exit();
 844         ltt_ring_buffer_client_discard_exit();
 845         ltt_ring_buffer_client_overwrite_exit();
 846         ltt_ring_buffer_metadata_client_exit();
 847         exit_tracepoint();
 848         if (!exiting) {
 849                 /* Reinitialize values for fork */
 850                 sem_count = 2;
 851                 lttng_ust_comm_should_quit = 0;
 852                 initialized = 0;
 853         }
 854 }
 855
 856 void __attribute__((destructor)) lttng_ust_exit(void)
 857 {
 858         int ret;
 859
 860         /*
 861          * Using pthread_cancel here because:
 862          * A) we don't want to hang application teardown.
 863          * B) the thread is not allocating any resource.
 864          */
 865
 866         /*
 867          * Require the communication thread to quit. Synchronize with
 868          * mutexes to ensure it is not in a mutex critical section when
 869          * pthread_cancel is later called.
 870          */
 871         ust_lock();
 872         lttng_ust_comm_should_quit = 1;
 873         ust_unlock();
 874
 875         ret = pthread_cancel(global_apps.ust_listener);
 876         if (ret) {
 877                 ERR("Error cancelling global ust listener thread");
 878         }
 879         if (local_apps.allowed) {
 880                 ret = pthread_cancel(local_apps.ust_listener);
 881                 if (ret) {
 882                         ERR("Error cancelling local ust listener thread");
 883                 }
 884         }
 885         lttng_ust_cleanup(1);
 886 }
 887
 888 /*
 889  * We exclude the worker threads across fork and clone (except
 890  * CLONE_VM), because these system calls only keep the forking thread
 891  * running in the child.  Therefore, we don't want to call fork or clone
 892  * in the middle of an tracepoint or ust tracing state modification.
 893  * Holding this mutex protects these structures across fork and clone.
 894  */
 895 void ust_before_fork(sigset_t *save_sigset)
 896 {
 897         /*
 898          * Disable signals. This is to avoid that the child intervenes
 899          * before it is properly setup for tracing. It is safer to
 900          * disable all signals, because then we know we are not breaking
 901          * anything by restoring the original mask.
 902          */
 903         sigset_t all_sigs;
 904         int ret;
 905
 906         /* Disable signals */
 907         sigfillset(&all_sigs);
 908         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
 909         if (ret == -1) {
 910                 PERROR("sigprocmask");
 911         }
 912         ust_lock();
 913         rcu_bp_before_fork();
 914 }
 915
 916 static void ust_after_fork_common(sigset_t *restore_sigset)
 917 {
 918         int ret;
 919
 920         DBG("process %d", getpid());
 921         ust_unlock();
 922         /* Restore signals */
 923         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
 924         if (ret == -1) {
 925                 PERROR("sigprocmask");
 926         }
 927 }
 928
 929 void ust_after_fork_parent(sigset_t *restore_sigset)
 930 {
 931         DBG("process %d", getpid());
 932         rcu_bp_after_fork_parent();
 933         /* Release mutexes and reenable signals */
 934         ust_after_fork_common(restore_sigset);
 935 }
 936
 937 /*
 938  * After fork, in the child, we need to cleanup all the leftover state,
 939  * except the worker thread which already magically disappeared thanks
 940  * to the weird Linux fork semantics. After tyding up, we call
 941  * lttng_ust_init() again to start over as a new PID.
 942  *
 943  * This is meant for forks() that have tracing in the child between the
 944  * fork and following exec call (if there is any).
 945  */
 946 void ust_after_fork_child(sigset_t *restore_sigset)
 947 {
 948         DBG("process %d", getpid());
 949         /* Release urcu mutexes */
 950         rcu_bp_after_fork_child();
 951         lttng_ust_cleanup(0);
 952         lttng_context_vtid_reset();
 953         /* Release mutexes and reenable signals */
 954         ust_after_fork_common(restore_sigset);
 955         lttng_ust_init();
 956 }