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/mman.h>
  26 #include <sys/stat.h>
  27 #include <sys/types.h>
  28 #include <sys/wait.h>
  29 #include <fcntl.h>
  30 #include <unistd.h>
  31 #include <errno.h>
  32 #include <pthread.h>
  33 #include <semaphore.h>
  34 #include <time.h>
  35 #include <assert.h>
  36 #include <signal.h>
  37 #include <urcu/uatomic.h>
  38 #include <urcu/futex.h>
  39 #include <urcu/compiler.h>
  40
  41 #include <lttng/ust-events.h>
  42 #include <lttng/ust-abi.h>
  43 #include <lttng/ust.h>
  44 #include <ust-comm.h>
  45 #include <usterr-signal-safe.h>
  46 #include "tracepoint-internal.h"
  47 #include "ltt-tracer-core.h"
  48 #include "compat.h"
  49 #include "../libringbuffer/tlsfixup.h"
  50
  51 /*
  52  * Has lttng ust comm constructor been called ?
  53  */
  54 static int initialized;
  55
  56 /*
  57  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  58  * Held when handling a command, also held by fork() to deal with
  59  * removal of threads, and by exit path.
  60  */
  61
  62 /* Should the ust comm thread quit ? */
  63 static int lttng_ust_comm_should_quit;
  64
  65 /*
  66  * Wait for either of these before continuing to the main
  67  * program:
  68  * - the register_done message from sessiond daemon
  69  *   (will let the sessiond daemon enable sessions before main
  70  *   starts.)
  71  * - sessiond daemon is not reachable.
  72  * - timeout (ensuring applications are resilient to session
  73  *   daemon problems).
  74  */
  75 static sem_t constructor_wait;
  76 /*
  77  * Doing this for both the global and local sessiond.
  78  */
  79 static int sem_count = { 2 };
  80
  81 /*
  82  * Info about socket and associated listener thread.
  83  */
  84 struct sock_info {
  85         const char *name;
  86         pthread_t ust_listener; /* listener thread */
  87         int root_handle;
  88         int constructor_sem_posted;
  89         int allowed;
  90         int global;
  91
  92         char sock_path[PATH_MAX];
  93         int socket;
  94
  95         char wait_shm_path[PATH_MAX];
  96         char *wait_shm_mmap;
  97 };
  98
  99 /* Socket from app (connect) to session daemon (listen) for communication */
 100 struct sock_info global_apps = {
 101         .name = "global",
 102         .global = 1,
 103
 104         .root_handle = -1,
 105         .allowed = 1,
 106
 107         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 108         .socket = -1,
 109
 110         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 111 };
 112
 113 /* TODO: allow global_apps_sock_path override */
 114
 115 struct sock_info local_apps = {
 116         .name = "local",
 117         .global = 0,
 118         .root_handle = -1,
 119         .allowed = 0,   /* Check setuid bit first */
 120
 121         .socket = -1,
 122 };
 123
 124 static int wait_poll_fallback;
 125
 126 extern void ltt_ring_buffer_client_overwrite_init(void);
 127 extern void ltt_ring_buffer_client_discard_init(void);
 128 extern void ltt_ring_buffer_metadata_client_init(void);
 129 extern void ltt_ring_buffer_client_overwrite_exit(void);
 130 extern void ltt_ring_buffer_client_discard_exit(void);
 131 extern void ltt_ring_buffer_metadata_client_exit(void);
 132
 133 static
 134 int setup_local_apps(void)
 135 {
 136         const char *home_dir;
 137         uid_t uid;
 138
 139         uid = getuid();
 140         /*
 141          * Disallow per-user tracing for setuid binaries.
 142          */
 143         if (uid != geteuid()) {
 144                 local_apps.allowed = 0;
 145                 return 0;
 146         } else {
 147                 local_apps.allowed = 1;
 148         }
 149         home_dir = (const char *) getenv("HOME");
 150         if (!home_dir)
 151                 return -ENOENT;
 152         snprintf(local_apps.sock_path, PATH_MAX,
 153                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 154         snprintf(local_apps.wait_shm_path, PATH_MAX,
 155                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 156         return 0;
 157 }
 158
 159 static
 160 int register_app_to_sessiond(int socket)
 161 {
 162         ssize_t 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         lttng_ust_getprocname(reg_msg.name);
 182
 183         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 184         if (ret >= 0 && ret != sizeof(reg_msg))
 185                 return -EIO;
 186         return ret;
 187 }
 188
 189 static
 190 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 191 {
 192         ssize_t len;
 193
 194         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 195         switch (len) {
 196         case sizeof(*lur):
 197                 DBG("message successfully sent");
 198                 return 0;
 199         case -1:
 200                 if (errno == ECONNRESET) {
 201                         printf("remote end closed connection\n");
 202                         return 0;
 203                 }
 204                 return -1;
 205         default:
 206                 printf("incorrect message size: %zd\n", len);
 207                 return -1;
 208         }
 209 }
 210
 211 static
 212 int handle_register_done(struct sock_info *sock_info)
 213 {
 214         int ret;
 215
 216         if (sock_info->constructor_sem_posted)
 217                 return 0;
 218         sock_info->constructor_sem_posted = 1;
 219         if (uatomic_read(&sem_count) <= 0) {
 220                 return 0;
 221         }
 222         ret = uatomic_add_return(&sem_count, -1);
 223         if (ret == 0) {
 224                 ret = sem_post(&constructor_wait);
 225                 assert(!ret);
 226         }
 227         return 0;
 228 }
 229
 230 static
 231 int handle_message(struct sock_info *sock_info,
 232                 int sock, struct ustcomm_ust_msg *lum)
 233 {
 234         int ret = 0;
 235         const struct lttng_ust_objd_ops *ops;
 236         struct ustcomm_ust_reply lur;
 237         int shm_fd, wait_fd;
 238         union ust_args args;
 239
 240         ust_lock();
 241
 242         memset(&lur, 0, sizeof(lur));
 243
 244         if (lttng_ust_comm_should_quit) {
 245                 ret = -EPERM;
 246                 goto end;
 247         }
 248
 249         ops = objd_ops(lum->handle);
 250         if (!ops) {
 251                 ret = -ENOENT;
 252                 goto end;
 253         }
 254
 255         switch (lum->cmd) {
 256         case LTTNG_UST_REGISTER_DONE:
 257                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 258                         ret = handle_register_done(sock_info);
 259                 else
 260                         ret = -EINVAL;
 261                 break;
 262         case LTTNG_UST_RELEASE:
 263                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 264                         ret = -EPERM;
 265                 else
 266                         ret = lttng_ust_objd_unref(lum->handle);
 267                 break;
 268         default:
 269                 if (ops->cmd)
 270                         ret = ops->cmd(lum->handle, lum->cmd,
 271                                         (unsigned long) &lum->u,
 272                                         &args);
 273                 else
 274                         ret = -ENOSYS;
 275                 break;
 276         }
 277
 278 end:
 279         lur.handle = lum->handle;
 280         lur.cmd = lum->cmd;
 281         lur.ret_val = ret;
 282         if (ret >= 0) {
 283                 lur.ret_code = USTCOMM_OK;
 284         } else {
 285                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 286                 lur.ret_code = ret;
 287         }
 288         if (ret >= 0) {
 289                 switch (lum->cmd) {
 290                 case LTTNG_UST_STREAM:
 291                         /*
 292                          * Special-case reply to send stream info.
 293                          * Use lum.u output.
 294                          */
 295                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 296                         shm_fd = *args.stream.shm_fd;
 297                         wait_fd = *args.stream.wait_fd;
 298                         break;
 299                 case LTTNG_UST_METADATA:
 300                 case LTTNG_UST_CHANNEL:
 301                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 302                         shm_fd = *args.channel.shm_fd;
 303                         wait_fd = *args.channel.wait_fd;
 304                         break;
 305                 case LTTNG_UST_TRACER_VERSION:
 306                         lur.u.version = lum->u.version;
 307                         break;
 308                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 309                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 310                         break;
 311                 }
 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                 int sendret = 0;
 324
 325                 /* we also need to send the file descriptors. */
 326                 ret = ustcomm_send_fds_unix_sock(sock,
 327                         &shm_fd, &shm_fd,
 328                         1, sizeof(int));
 329                 if (ret < 0) {
 330                         perror("send shm_fd");
 331                         sendret = ret;
 332                 }
 333                 /*
 334                  * The sessiond expects 2 file descriptors, even upon
 335                  * error.
 336                  */
 337                 ret = ustcomm_send_fds_unix_sock(sock,
 338                         &wait_fd, &wait_fd,
 339                         1, sizeof(int));
 340                 if (ret < 0) {
 341                         perror("send wait_fd");
 342                         goto error;
 343                 }
 344                 if (sendret) {
 345                         ret = sendret;
 346                         goto error;
 347                 }
 348         }
 349         /*
 350          * We still have the memory map reference, and the fds have been
 351          * sent to the sessiond. We can therefore close those fds. Note
 352          * that we keep the write side of the wait_fd open, but close
 353          * the read side.
 354          */
 355         if (lur.ret_code == USTCOMM_OK) {
 356                 switch (lum->cmd) {
 357                 case LTTNG_UST_STREAM:
 358                         if (shm_fd >= 0) {
 359                                 ret = close(shm_fd);
 360                                 if (ret) {
 361                                         PERROR("Error closing stream shm_fd");
 362                                 }
 363                                 *args.stream.shm_fd = -1;
 364                         }
 365                         if (wait_fd >= 0) {
 366                                 ret = close(wait_fd);
 367                                 if (ret) {
 368                                         PERROR("Error closing stream wait_fd");
 369                                 }
 370                                 *args.stream.wait_fd = -1;
 371                         }
 372                         break;
 373                 case LTTNG_UST_METADATA:
 374                 case LTTNG_UST_CHANNEL:
 375                         if (shm_fd >= 0) {
 376                                 ret = close(shm_fd);
 377                                 if (ret) {
 378                                         PERROR("Error closing channel shm_fd");
 379                                 }
 380                                 *args.channel.shm_fd = -1;
 381                         }
 382                         if (wait_fd >= 0) {
 383                                 ret = close(wait_fd);
 384                                 if (ret) {
 385                                         PERROR("Error closing channel wait_fd");
 386                                 }
 387                                 *args.channel.wait_fd = -1;
 388                         }
 389                         break;
 390                 }
 391         }
 392
 393 error:
 394         ust_unlock();
 395         return ret;
 396 }
 397
 398 static
 399 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 400 {
 401         int ret;
 402
 403         if (sock_info->socket != -1) {
 404                 ret = ustcomm_close_unix_sock(sock_info->socket);
 405                 if (ret) {
 406                         ERR("Error closing apps socket");
 407                 }
 408                 sock_info->socket = -1;
 409         }
 410         if (sock_info->root_handle != -1) {
 411                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 412                 if (ret) {
 413                         ERR("Error unref root handle");
 414                 }
 415                 sock_info->root_handle = -1;
 416         }
 417         sock_info->constructor_sem_posted = 0;
 418         /*
 419          * wait_shm_mmap is used by listener threads outside of the
 420          * ust lock, so we cannot tear it down ourselves, because we
 421          * cannot join on these threads. Leave this task to the OS
 422          * process exit.
 423          */
 424         if (!exiting && sock_info->wait_shm_mmap) {
 425                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 426                 if (ret) {
 427                         ERR("Error unmapping wait shm");
 428                 }
 429                 sock_info->wait_shm_mmap = NULL;
 430         }
 431 }
 432
 433 /*
 434  * Using fork to set umask in the child process (not multi-thread safe).
 435  * We deal with the shm_open vs ftruncate race (happening when the
 436  * sessiond owns the shm and does not let everybody modify it, to ensure
 437  * safety against shm_unlink) by simply letting the mmap fail and
 438  * retrying after a few seconds.
 439  * For global shm, everybody has rw access to it until the sessiond
 440  * starts.
 441  */
 442 static
 443 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 444 {
 445         int wait_shm_fd, ret;
 446         pid_t pid;
 447
 448         /*
 449          * Try to open read-only.
 450          */
 451         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 452         if (wait_shm_fd >= 0) {
 453                 goto end;
 454         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 455                 /*
 456                  * Real-only open did not work, and it's not because the
 457                  * entry was not present. It's a failure that prohibits
 458                  * using shm.
 459                  */
 460                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 461                 goto end;
 462         }
 463         /*
 464          * If the open failed because the file did not exist, try
 465          * creating it ourself.
 466          */
 467         pid = fork();
 468         if (pid > 0) {
 469                 int status;
 470
 471                 /*
 472                  * Parent: wait for child to return, in which case the
 473                  * shared memory map will have been created.
 474                  */
 475                 pid = wait(&status);
 476                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 477                         wait_shm_fd = -1;
 478                         goto end;
 479                 }
 480                 /*
 481                  * Try to open read-only again after creation.
 482                  */
 483                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 484                 if (wait_shm_fd < 0) {
 485                         /*
 486                          * Real-only open did not work. It's a failure
 487                          * that prohibits using shm.
 488                          */
 489                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 490                         goto end;
 491                 }
 492                 goto end;
 493         } else if (pid == 0) {
 494                 int create_mode;
 495
 496                 /* Child */
 497                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 498                 if (sock_info->global)
 499                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 500                 /*
 501                  * We're alone in a child process, so we can modify the
 502                  * process-wide umask.
 503                  */
 504                 umask(~create_mode);
 505                 /*
 506                  * Try creating shm (or get rw access).
 507                  * We don't do an exclusive open, because we allow other
 508                  * processes to create+ftruncate it concurrently.
 509                  */
 510                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 511                                 O_RDWR | O_CREAT, create_mode);
 512                 if (wait_shm_fd >= 0) {
 513                         ret = ftruncate(wait_shm_fd, mmap_size);
 514                         if (ret) {
 515                                 PERROR("ftruncate");
 516                                 exit(EXIT_FAILURE);
 517                         }
 518                         exit(EXIT_SUCCESS);
 519                 }
 520                 /*
 521                  * For local shm, we need to have rw access to accept
 522                  * opening it: this means the local sessiond will be
 523                  * able to wake us up. For global shm, we open it even
 524                  * if rw access is not granted, because the root.root
 525                  * sessiond will be able to override all rights and wake
 526                  * us up.
 527                  */
 528                 if (!sock_info->global && errno != EACCES) {
 529                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 530                         exit(EXIT_FAILURE);
 531                 }
 532                 /*
 533                  * The shm exists, but we cannot open it RW. Report
 534                  * success.
 535                  */
 536                 exit(EXIT_SUCCESS);
 537         } else {
 538                 return -1;
 539         }
 540 end:
 541         if (wait_shm_fd >= 0 && !sock_info->global) {
 542                 struct stat statbuf;
 543
 544                 /*
 545                  * Ensure that our user is the owner of the shm file for
 546                  * local shm. If we do not own the file, it means our
 547                  * sessiond will not have access to wake us up (there is
 548                  * probably a rogue process trying to fake our
 549                  * sessiond). Fallback to polling method in this case.
 550                  */
 551                 ret = fstat(wait_shm_fd, &statbuf);
 552                 if (ret) {
 553                         PERROR("fstat");
 554                         goto error_close;
 555                 }
 556                 if (statbuf.st_uid != getuid())
 557                         goto error_close;
 558         }
 559         return wait_shm_fd;
 560
 561 error_close:
 562         ret = close(wait_shm_fd);
 563         if (ret) {
 564                 PERROR("Error closing fd");
 565         }
 566         return -1;
 567 }
 568
 569 static
 570 char *get_map_shm(struct sock_info *sock_info)
 571 {
 572         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 573         int wait_shm_fd, ret;
 574         char *wait_shm_mmap;
 575
 576         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 577         if (wait_shm_fd < 0) {
 578                 goto error;
 579         }
 580         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 581                   MAP_SHARED, wait_shm_fd, 0);
 582         /* close shm fd immediately after taking the mmap reference */
 583         ret = close(wait_shm_fd);
 584         if (ret) {
 585                 PERROR("Error closing fd");
 586         }
 587         if (wait_shm_mmap == MAP_FAILED) {
 588                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 589                 goto error;
 590         }
 591         return wait_shm_mmap;
 592
 593 error:
 594         return NULL;
 595 }
 596
 597 static
 598 void wait_for_sessiond(struct sock_info *sock_info)
 599 {
 600         int ret;
 601
 602         ust_lock();
 603         if (lttng_ust_comm_should_quit) {
 604                 goto quit;
 605         }
 606         if (wait_poll_fallback) {
 607                 goto error;
 608         }
 609         if (!sock_info->wait_shm_mmap) {
 610                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 611                 if (!sock_info->wait_shm_mmap)
 612                         goto error;
 613         }
 614         ust_unlock();
 615
 616         DBG("Waiting for %s apps sessiond", sock_info->name);
 617         /* Wait for futex wakeup */
 618         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 619                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 620                         FUTEX_WAIT, 0, NULL, NULL, 0);
 621                 if (ret < 0) {
 622                         if (errno == EFAULT) {
 623                                 wait_poll_fallback = 1;
 624                                 DBG(
 625 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 626 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 627 "Please upgrade your kernel "
 628 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 629 "mainline). LTTng-UST will use polling mode fallback.");
 630                                 if (ust_debug())
 631                                         PERROR("futex");
 632                         }
 633                 }
 634         }
 635         return;
 636
 637 quit:
 638         ust_unlock();
 639         return;
 640
 641 error:
 642         ust_unlock();
 643         return;
 644 }
 645
 646 /*
 647  * This thread does not allocate any resource, except within
 648  * handle_message, within mutex protection. This mutex protects against
 649  * fork and exit.
 650  * The other moment it allocates resources is at socket connexion, which
 651  * is also protected by the mutex.
 652  */
 653 static
 654 void *ust_listener_thread(void *arg)
 655 {
 656         struct sock_info *sock_info = arg;
 657         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 658
 659         /* Restart trying to connect to the session daemon */
 660 restart:
 661         if (prev_connect_failed) {
 662                 /* Wait for sessiond availability with pipe */
 663                 wait_for_sessiond(sock_info);
 664                 if (has_waited) {
 665                         has_waited = 0;
 666                         /*
 667                          * Sleep for 5 seconds before retrying after a
 668                          * sequence of failure / wait / failure. This
 669                          * deals with a killed or broken session daemon.
 670                          */
 671                         sleep(5);
 672                 }
 673                 has_waited = 1;
 674                 prev_connect_failed = 0;
 675         }
 676         ust_lock();
 677
 678         if (lttng_ust_comm_should_quit) {
 679                 ust_unlock();
 680                 goto quit;
 681         }
 682
 683         if (sock_info->socket != -1) {
 684                 ret = ustcomm_close_unix_sock(sock_info->socket);
 685                 if (ret) {
 686                         ERR("Error closing %s apps socket", sock_info->name);
 687                 }
 688                 sock_info->socket = -1;
 689         }
 690
 691         /* Register */
 692         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 693         if (ret < 0) {
 694                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 695                 prev_connect_failed = 1;
 696                 /*
 697                  * If we cannot find the sessiond daemon, don't delay
 698                  * constructor execution.
 699                  */
 700                 ret = handle_register_done(sock_info);
 701                 assert(!ret);
 702                 ust_unlock();
 703                 goto restart;
 704         }
 705
 706         sock_info->socket = sock = ret;
 707
 708         /*
 709          * Create only one root handle per listener thread for the whole
 710          * process lifetime.
 711          */
 712         if (sock_info->root_handle == -1) {
 713                 ret = lttng_abi_create_root_handle();
 714                 if (ret < 0) {
 715                         ERR("Error creating root handle");
 716                         ust_unlock();
 717                         goto quit;
 718                 }
 719                 sock_info->root_handle = ret;
 720         }
 721
 722         ret = register_app_to_sessiond(sock);
 723         if (ret < 0) {
 724                 ERR("Error registering to %s apps socket", sock_info->name);
 725                 prev_connect_failed = 1;
 726                 /*
 727                  * If we cannot register to the sessiond daemon, don't
 728                  * delay constructor execution.
 729                  */
 730                 ret = handle_register_done(sock_info);
 731                 assert(!ret);
 732                 ust_unlock();
 733                 goto restart;
 734         }
 735         ust_unlock();
 736
 737         for (;;) {
 738                 ssize_t len;
 739                 struct ustcomm_ust_msg lum;
 740
 741                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 742                 switch (len) {
 743                 case 0: /* orderly shutdown */
 744                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 745                         ust_lock();
 746                         /*
 747                          * Either sessiond has shutdown or refused us by closing the socket.
 748                          * In either case, we don't want to delay construction execution,
 749                          * and we need to wait before retry.
 750                          */
 751                         prev_connect_failed = 1;
 752                         /*
 753                          * If we cannot register to the sessiond daemon, don't
 754                          * delay constructor execution.
 755                          */
 756                         ret = handle_register_done(sock_info);
 757                         assert(!ret);
 758                         ust_unlock();
 759                         goto end;
 760                 case sizeof(lum):
 761                         DBG("message received\n");
 762                         ret = handle_message(sock_info, sock, &lum);
 763                         if (ret < 0) {
 764                                 ERR("Error handling message for %s socket", sock_info->name);
 765                         }
 766                         continue;
 767                 case -1:
 768                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 769                         if (errno == ECONNRESET) {
 770                                 ERR("%s remote end closed connection\n", sock_info->name);
 771                                 goto end;
 772                         }
 773                         goto end;
 774                 default:
 775                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 776                         continue;
 777                 }
 778
 779         }
 780 end:
 781         goto restart;   /* try to reconnect */
 782 quit:
 783         return NULL;
 784 }
 785
 786 /*
 787  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 788  */
 789 static
 790 int get_timeout(struct timespec *constructor_timeout)
 791 {
 792         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 793         char *str_delay;
 794         int ret;
 795
 796         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 797         if (str_delay) {
 798                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 799         }
 800
 801         switch (constructor_delay_ms) {
 802         case -1:/* fall-through */
 803         case 0:
 804                 return constructor_delay_ms;
 805         default:
 806                 break;
 807         }
 808
 809         /*
 810          * If we are unable to find the current time, don't wait.
 811          */
 812         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 813         if (ret) {
 814                 return -1;
 815         }
 816         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 817         constructor_timeout->tv_nsec +=
 818                 (constructor_delay_ms % 1000UL) * 1000000UL;
 819         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 820                 constructor_timeout->tv_sec++;
 821                 constructor_timeout->tv_nsec -= 1000000000UL;
 822         }
 823         return 1;
 824 }
 825
 826 /*
 827  * sessiond monitoring thread: monitor presence of global and per-user
 828  * sessiond by polling the application common named pipe.
 829  */
 830 /* TODO */
 831
 832 void __attribute__((constructor)) lttng_ust_init(void)
 833 {
 834         struct timespec constructor_timeout;
 835         sigset_t sig_all_blocked, orig_parent_mask;
 836         int timeout_mode;
 837         int ret;
 838
 839         if (uatomic_xchg(&initialized, 1) == 1)
 840                 return;
 841
 842         /*
 843          * Fixup interdependency between TLS fixup mutex (which happens
 844          * to be the dynamic linker mutex) and ust_lock, taken within
 845          * the ust lock.
 846          */
 847         lttng_fixup_event_tls();
 848         lttng_fixup_ringbuffer_tls();
 849         lttng_fixup_vtid_tls();
 850
 851         /*
 852          * We want precise control over the order in which we construct
 853          * our sub-libraries vs starting to receive commands from
 854          * sessiond (otherwise leading to errors when trying to create
 855          * sessiond before the init functions are completed).
 856          */
 857         init_usterr();
 858         init_tracepoint();
 859         ltt_ring_buffer_metadata_client_init();
 860         ltt_ring_buffer_client_overwrite_init();
 861         ltt_ring_buffer_client_discard_init();
 862
 863         timeout_mode = get_timeout(&constructor_timeout);
 864
 865         ret = sem_init(&constructor_wait, 0, 0);
 866         assert(!ret);
 867
 868         ret = setup_local_apps();
 869         if (ret) {
 870                 ERR("Error setting up to local apps");
 871         }
 872
 873         /* A new thread created by pthread_create inherits the signal mask
 874          * from the parent. To avoid any signal being received by the
 875          * listener thread, we block all signals temporarily in the parent,
 876          * while we create the listener thread.
 877          */
 878         sigfillset(&sig_all_blocked);
 879         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 880         if (ret) {
 881                 PERROR("pthread_sigmask: %s", strerror(ret));
 882         }
 883
 884         ret = pthread_create(&local_apps.ust_listener, NULL,
 885                         ust_listener_thread, &local_apps);
 886
 887         if (local_apps.allowed) {
 888                 ret = pthread_create(&global_apps.ust_listener, NULL,
 889                                 ust_listener_thread, &global_apps);
 890         } else {
 891                 handle_register_done(&local_apps);
 892         }
 893
 894         /* Restore original signal mask in parent */
 895         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
 896         if (ret) {
 897                 PERROR("pthread_sigmask: %s", strerror(ret));
 898         }
 899
 900         switch (timeout_mode) {
 901         case 1: /* timeout wait */
 902                 do {
 903                         ret = sem_timedwait(&constructor_wait,
 904                                         &constructor_timeout);
 905                 } while (ret < 0 && errno == EINTR);
 906                 if (ret < 0 && errno == ETIMEDOUT) {
 907                         ERR("Timed out waiting for ltt-sessiond");
 908                 } else {
 909                         assert(!ret);
 910                 }
 911                 break;
 912         case -1:/* wait forever */
 913                 do {
 914                         ret = sem_wait(&constructor_wait);
 915                 } while (ret < 0 && errno == EINTR);
 916                 assert(!ret);
 917                 break;
 918         case 0: /* no timeout */
 919                 break;
 920         }
 921 }
 922
 923 static
 924 void lttng_ust_cleanup(int exiting)
 925 {
 926         cleanup_sock_info(&global_apps, exiting);
 927         if (local_apps.allowed) {
 928                 cleanup_sock_info(&local_apps, exiting);
 929         }
 930         /*
 931          * The teardown in this function all affect data structures
 932          * accessed under the UST lock by the listener thread. This
 933          * lock, along with the lttng_ust_comm_should_quit flag, ensure
 934          * that none of these threads are accessing this data at this
 935          * point.
 936          */
 937         lttng_ust_abi_exit();
 938         lttng_ust_events_exit();
 939         ltt_ring_buffer_client_discard_exit();
 940         ltt_ring_buffer_client_overwrite_exit();
 941         ltt_ring_buffer_metadata_client_exit();
 942         exit_tracepoint();
 943         if (!exiting) {
 944                 /* Reinitialize values for fork */
 945                 sem_count = 2;
 946                 lttng_ust_comm_should_quit = 0;
 947                 initialized = 0;
 948         }
 949 }
 950
 951 void __attribute__((destructor)) lttng_ust_exit(void)
 952 {
 953         int ret;
 954
 955         /*
 956          * Using pthread_cancel here because:
 957          * A) we don't want to hang application teardown.
 958          * B) the thread is not allocating any resource.
 959          */
 960
 961         /*
 962          * Require the communication thread to quit. Synchronize with
 963          * mutexes to ensure it is not in a mutex critical section when
 964          * pthread_cancel is later called.
 965          */
 966         ust_lock();
 967         lttng_ust_comm_should_quit = 1;
 968         ust_unlock();
 969
 970         /* cancel threads */
 971         ret = pthread_cancel(global_apps.ust_listener);
 972         if (ret) {
 973                 ERR("Error cancelling global ust listener thread");
 974         }
 975         if (local_apps.allowed) {
 976                 ret = pthread_cancel(local_apps.ust_listener);
 977                 if (ret) {
 978                         ERR("Error cancelling local ust listener thread");
 979                 }
 980         }
 981         /*
 982          * Do NOT join threads: use of sys_futex makes it impossible to
 983          * join the threads without using async-cancel, but async-cancel
 984          * is delivered by a signal, which could hit the target thread
 985          * anywhere in its code path, including while the ust_lock() is
 986          * held, causing a deadlock for the other thread. Let the OS
 987          * cleanup the threads if there are stalled in a syscall.
 988          */
 989         lttng_ust_cleanup(1);
 990 }
 991
 992 /*
 993  * We exclude the worker threads across fork and clone (except
 994  * CLONE_VM), because these system calls only keep the forking thread
 995  * running in the child.  Therefore, we don't want to call fork or clone
 996  * in the middle of an tracepoint or ust tracing state modification.
 997  * Holding this mutex protects these structures across fork and clone.
 998  */
 999 void ust_before_fork(sigset_t *save_sigset)
1000 {
1001         /*
1002          * Disable signals. This is to avoid that the child intervenes
1003          * before it is properly setup for tracing. It is safer to
1004          * disable all signals, because then we know we are not breaking
1005          * anything by restoring the original mask.
1006          */
1007         sigset_t all_sigs;
1008         int ret;
1009
1010         /* Disable signals */
1011         sigfillset(&all_sigs);
1012         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1013         if (ret == -1) {
1014                 PERROR("sigprocmask");
1015         }
1016         ust_lock();
1017         rcu_bp_before_fork();
1018 }
1019
1020 static void ust_after_fork_common(sigset_t *restore_sigset)
1021 {
1022         int ret;
1023
1024         DBG("process %d", getpid());
1025         ust_unlock();
1026         /* Restore signals */
1027         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1028         if (ret == -1) {
1029                 PERROR("sigprocmask");
1030         }
1031 }
1032
1033 void ust_after_fork_parent(sigset_t *restore_sigset)
1034 {
1035         DBG("process %d", getpid());
1036         rcu_bp_after_fork_parent();
1037         /* Release mutexes and reenable signals */
1038         ust_after_fork_common(restore_sigset);
1039 }
1040
1041 /*
1042  * After fork, in the child, we need to cleanup all the leftover state,
1043  * except the worker thread which already magically disappeared thanks
1044  * to the weird Linux fork semantics. After tyding up, we call
1045  * lttng_ust_init() again to start over as a new PID.
1046  *
1047  * This is meant for forks() that have tracing in the child between the
1048  * fork and following exec call (if there is any).
1049  */
1050 void ust_after_fork_child(sigset_t *restore_sigset)
1051 {
1052         DBG("process %d", getpid());
1053         /* Release urcu mutexes */
1054         rcu_bp_after_fork_child();
1055         lttng_ust_cleanup(0);
1056         lttng_context_vtid_reset();
1057         /* Release mutexes and reenable signals */
1058         ust_after_fork_common(restore_sigset);
1059         lttng_ust_init();
1060 }