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                 /* 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         /*
 340          * We still have the memory map reference, and the fds have been
 341          * sent to the sessiond. We can therefore close those fds. Note
 342          * that we keep the write side of the wait_fd open, but close
 343          * the read side.
 344          */
 345         if (lur.ret_code == USTCOMM_OK) {
 346                 switch (lum->cmd) {
 347                 case LTTNG_UST_STREAM:
 348                         if (shm_fd >= 0) {
 349                                 ret = close(shm_fd);
 350                                 if (ret) {
 351                                         PERROR("Error closing stream shm_fd");
 352                                 }
 353                                 *args.stream.shm_fd = -1;
 354                         }
 355                         if (wait_fd >= 0) {
 356                                 ret = close(wait_fd);
 357                                 if (ret) {
 358                                         PERROR("Error closing stream wait_fd");
 359                                 }
 360                                 *args.stream.wait_fd = -1;
 361                         }
 362                         break;
 363                 case LTTNG_UST_METADATA:
 364                 case LTTNG_UST_CHANNEL:
 365                         if (shm_fd >= 0) {
 366                                 ret = close(shm_fd);
 367                                 if (ret) {
 368                                         PERROR("Error closing channel shm_fd");
 369                                 }
 370                                 *args.channel.shm_fd = -1;
 371                         }
 372                         if (wait_fd >= 0) {
 373                                 ret = close(wait_fd);
 374                                 if (ret) {
 375                                         PERROR("Error closing channel wait_fd");
 376                                 }
 377                                 *args.channel.wait_fd = -1;
 378                         }
 379                         break;
 380                 }
 381         }
 382
 383 error:
 384         ust_unlock();
 385         return ret;
 386 }
 387
 388 static
 389 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 390 {
 391         int ret;
 392
 393         if (sock_info->socket != -1) {
 394                 ret = close(sock_info->socket);
 395                 if (ret) {
 396                         ERR("Error closing apps socket");
 397                 }
 398                 sock_info->socket = -1;
 399         }
 400         if (sock_info->root_handle != -1) {
 401                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 402                 if (ret) {
 403                         ERR("Error unref root handle");
 404                 }
 405                 sock_info->root_handle = -1;
 406         }
 407         sock_info->constructor_sem_posted = 0;
 408         /*
 409          * wait_shm_mmap is used by listener threads outside of the
 410          * ust lock, so we cannot tear it down ourselves, because we
 411          * cannot join on these threads. Leave this task to the OS
 412          * process exit.
 413          */
 414         if (!exiting && sock_info->wait_shm_mmap) {
 415                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 416                 if (ret) {
 417                         ERR("Error unmapping wait shm");
 418                 }
 419                 sock_info->wait_shm_mmap = NULL;
 420         }
 421 }
 422
 423 /*
 424  * Using fork to set umask in the child process (not multi-thread safe).
 425  * We deal with the shm_open vs ftruncate race (happening when the
 426  * sessiond owns the shm and does not let everybody modify it, to ensure
 427  * safety against shm_unlink) by simply letting the mmap fail and
 428  * retrying after a few seconds.
 429  * For global shm, everybody has rw access to it until the sessiond
 430  * starts.
 431  */
 432 static
 433 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 434 {
 435         int wait_shm_fd, ret;
 436         pid_t pid;
 437
 438         /*
 439          * Try to open read-only.
 440          */
 441         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 442         if (wait_shm_fd >= 0) {
 443                 goto end;
 444         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 445                 /*
 446                  * Real-only open did not work, and it's not because the
 447                  * entry was not present. It's a failure that prohibits
 448                  * using shm.
 449                  */
 450                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 451                 goto end;
 452         }
 453         /*
 454          * If the open failed because the file did not exist, try
 455          * creating it ourself.
 456          */
 457         pid = fork();
 458         if (pid > 0) {
 459                 int status;
 460
 461                 /*
 462                  * Parent: wait for child to return, in which case the
 463                  * shared memory map will have been created.
 464                  */
 465                 pid = wait(&status);
 466                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 467                         wait_shm_fd = -1;
 468                         goto end;
 469                 }
 470                 /*
 471                  * Try to open read-only again after creation.
 472                  */
 473                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 474                 if (wait_shm_fd < 0) {
 475                         /*
 476                          * Real-only open did not work. It's a failure
 477                          * that prohibits using shm.
 478                          */
 479                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 480                         goto end;
 481                 }
 482                 goto end;
 483         } else if (pid == 0) {
 484                 int create_mode;
 485
 486                 /* Child */
 487                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 488                 if (sock_info->global)
 489                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 490                 /*
 491                  * We're alone in a child process, so we can modify the
 492                  * process-wide umask.
 493                  */
 494                 umask(~create_mode);
 495                 /*
 496                  * Try creating shm (or get rw access).
 497                  * We don't do an exclusive open, because we allow other
 498                  * processes to create+ftruncate it concurrently.
 499                  */
 500                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 501                                 O_RDWR | O_CREAT, create_mode);
 502                 if (wait_shm_fd >= 0) {
 503                         ret = ftruncate(wait_shm_fd, mmap_size);
 504                         if (ret) {
 505                                 PERROR("ftruncate");
 506                                 exit(EXIT_FAILURE);
 507                         }
 508                         exit(EXIT_SUCCESS);
 509                 }
 510                 /*
 511                  * For local shm, we need to have rw access to accept
 512                  * opening it: this means the local sessiond will be
 513                  * able to wake us up. For global shm, we open it even
 514                  * if rw access is not granted, because the root.root
 515                  * sessiond will be able to override all rights and wake
 516                  * us up.
 517                  */
 518                 if (!sock_info->global && errno != EACCES) {
 519                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 520                         exit(EXIT_FAILURE);
 521                 }
 522                 /*
 523                  * The shm exists, but we cannot open it RW. Report
 524                  * success.
 525                  */
 526                 exit(EXIT_SUCCESS);
 527         } else {
 528                 return -1;
 529         }
 530 end:
 531         if (wait_shm_fd >= 0 && !sock_info->global) {
 532                 struct stat statbuf;
 533
 534                 /*
 535                  * Ensure that our user is the owner of the shm file for
 536                  * local shm. If we do not own the file, it means our
 537                  * sessiond will not have access to wake us up (there is
 538                  * probably a rogue process trying to fake our
 539                  * sessiond). Fallback to polling method in this case.
 540                  */
 541                 ret = fstat(wait_shm_fd, &statbuf);
 542                 if (ret) {
 543                         PERROR("fstat");
 544                         goto error_close;
 545                 }
 546                 if (statbuf.st_uid != getuid())
 547                         goto error_close;
 548         }
 549         return wait_shm_fd;
 550
 551 error_close:
 552         ret = close(wait_shm_fd);
 553         if (ret) {
 554                 PERROR("Error closing fd");
 555         }
 556         return -1;
 557 }
 558
 559 static
 560 char *get_map_shm(struct sock_info *sock_info)
 561 {
 562         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 563         int wait_shm_fd, ret;
 564         char *wait_shm_mmap;
 565
 566         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 567         if (wait_shm_fd < 0) {
 568                 goto error;
 569         }
 570         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 571                   MAP_SHARED, wait_shm_fd, 0);
 572         /* close shm fd immediately after taking the mmap reference */
 573         ret = close(wait_shm_fd);
 574         if (ret) {
 575                 PERROR("Error closing fd");
 576         }
 577         if (wait_shm_mmap == MAP_FAILED) {
 578                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 579                 goto error;
 580         }
 581         return wait_shm_mmap;
 582
 583 error:
 584         return NULL;
 585 }
 586
 587 static
 588 void wait_for_sessiond(struct sock_info *sock_info)
 589 {
 590         int ret;
 591
 592         ust_lock();
 593         if (lttng_ust_comm_should_quit) {
 594                 goto quit;
 595         }
 596         if (wait_poll_fallback) {
 597                 goto error;
 598         }
 599         if (!sock_info->wait_shm_mmap) {
 600                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 601                 if (!sock_info->wait_shm_mmap)
 602                         goto error;
 603         }
 604         ust_unlock();
 605
 606         DBG("Waiting for %s apps sessiond", sock_info->name);
 607         /* Wait for futex wakeup */
 608         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 609                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 610                         FUTEX_WAIT, 0, NULL, NULL, 0);
 611                 if (ret < 0) {
 612                         if (errno == EFAULT) {
 613                                 wait_poll_fallback = 1;
 614                                 DBG(
 615 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 616 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 617 "Please upgrade your kernel "
 618 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 619 "mainline). LTTng-UST will use polling mode fallback.");
 620                                 if (ust_debug())
 621                                         PERROR("futex");
 622                         }
 623                 }
 624         }
 625         return;
 626
 627 quit:
 628         ust_unlock();
 629         return;
 630
 631 error:
 632         ust_unlock();
 633         return;
 634 }
 635
 636 /*
 637  * This thread does not allocate any resource, except within
 638  * handle_message, within mutex protection. This mutex protects against
 639  * fork and exit.
 640  * The other moment it allocates resources is at socket connexion, which
 641  * is also protected by the mutex.
 642  */
 643 static
 644 void *ust_listener_thread(void *arg)
 645 {
 646         struct sock_info *sock_info = arg;
 647         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 648
 649         /* Restart trying to connect to the session daemon */
 650 restart:
 651         if (prev_connect_failed) {
 652                 /* Wait for sessiond availability with pipe */
 653                 wait_for_sessiond(sock_info);
 654                 if (has_waited) {
 655                         has_waited = 0;
 656                         /*
 657                          * Sleep for 5 seconds before retrying after a
 658                          * sequence of failure / wait / failure. This
 659                          * deals with a killed or broken session daemon.
 660                          */
 661                         sleep(5);
 662                 }
 663                 has_waited = 1;
 664                 prev_connect_failed = 0;
 665         }
 666         ust_lock();
 667
 668         if (lttng_ust_comm_should_quit) {
 669                 ust_unlock();
 670                 goto quit;
 671         }
 672
 673         if (sock_info->socket != -1) {
 674                 ret = close(sock_info->socket);
 675                 if (ret) {
 676                         ERR("Error closing %s apps socket", sock_info->name);
 677                 }
 678                 sock_info->socket = -1;
 679         }
 680
 681         /* Register */
 682         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 683         if (ret < 0) {
 684                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 685                 prev_connect_failed = 1;
 686                 /*
 687                  * If we cannot find the sessiond daemon, don't delay
 688                  * constructor execution.
 689                  */
 690                 ret = handle_register_done(sock_info);
 691                 assert(!ret);
 692                 ust_unlock();
 693                 goto restart;
 694         }
 695
 696         sock_info->socket = sock = ret;
 697
 698         /*
 699          * Create only one root handle per listener thread for the whole
 700          * process lifetime.
 701          */
 702         if (sock_info->root_handle == -1) {
 703                 ret = lttng_abi_create_root_handle();
 704                 if (ret < 0) {
 705                         ERR("Error creating root handle");
 706                         ust_unlock();
 707                         goto quit;
 708                 }
 709                 sock_info->root_handle = ret;
 710         }
 711
 712         ret = register_app_to_sessiond(sock);
 713         if (ret < 0) {
 714                 ERR("Error registering to %s apps socket", sock_info->name);
 715                 prev_connect_failed = 1;
 716                 /*
 717                  * If we cannot register to the sessiond daemon, don't
 718                  * delay constructor execution.
 719                  */
 720                 ret = handle_register_done(sock_info);
 721                 assert(!ret);
 722                 ust_unlock();
 723                 goto restart;
 724         }
 725         ust_unlock();
 726
 727         for (;;) {
 728                 ssize_t len;
 729                 struct ustcomm_ust_msg lum;
 730
 731                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 732                 switch (len) {
 733                 case 0: /* orderly shutdown */
 734                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 735                         ust_lock();
 736                         /*
 737                          * Either sessiond has shutdown or refused us by closing the socket.
 738                          * In either case, we don't want to delay construction execution,
 739                          * and we need to wait before retry.
 740                          */
 741                         prev_connect_failed = 1;
 742                         /*
 743                          * If we cannot register to the sessiond daemon, don't
 744                          * delay constructor execution.
 745                          */
 746                         ret = handle_register_done(sock_info);
 747                         assert(!ret);
 748                         ust_unlock();
 749                         goto end;
 750                 case sizeof(lum):
 751                         DBG("message received\n");
 752                         ret = handle_message(sock_info, sock, &lum);
 753                         if (ret < 0) {
 754                                 ERR("Error handling message for %s socket", sock_info->name);
 755                         }
 756                         continue;
 757                 case -1:
 758                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 759                         if (errno == ECONNRESET) {
 760                                 ERR("%s remote end closed connection\n", sock_info->name);
 761                                 goto end;
 762                         }
 763                         goto end;
 764                 default:
 765                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 766                         continue;
 767                 }
 768
 769         }
 770 end:
 771         goto restart;   /* try to reconnect */
 772 quit:
 773         return NULL;
 774 }
 775
 776 /*
 777  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 778  */
 779 static
 780 int get_timeout(struct timespec *constructor_timeout)
 781 {
 782         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 783         char *str_delay;
 784         int ret;
 785
 786         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 787         if (str_delay) {
 788                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 789         }
 790
 791         switch (constructor_delay_ms) {
 792         case -1:/* fall-through */
 793         case 0:
 794                 return constructor_delay_ms;
 795         default:
 796                 break;
 797         }
 798
 799         /*
 800          * If we are unable to find the current time, don't wait.
 801          */
 802         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 803         if (ret) {
 804                 return -1;
 805         }
 806         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 807         constructor_timeout->tv_nsec +=
 808                 (constructor_delay_ms % 1000UL) * 1000000UL;
 809         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 810                 constructor_timeout->tv_sec++;
 811                 constructor_timeout->tv_nsec -= 1000000000UL;
 812         }
 813         return 1;
 814 }
 815
 816 /*
 817  * sessiond monitoring thread: monitor presence of global and per-user
 818  * sessiond by polling the application common named pipe.
 819  */
 820 /* TODO */
 821
 822 void __attribute__((constructor)) lttng_ust_init(void)
 823 {
 824         struct timespec constructor_timeout;
 825         int timeout_mode;
 826         int ret;
 827
 828         if (uatomic_xchg(&initialized, 1) == 1)
 829                 return;
 830
 831         /*
 832          * Fixup interdependency between TLS fixup mutex (which happens
 833          * to be the dynamic linker mutex) and ust_lock, taken within
 834          * the ust lock.
 835          */
 836         lttng_fixup_event_tls();
 837         lttng_fixup_ringbuffer_tls();
 838         lttng_fixup_vtid_tls();
 839
 840         /*
 841          * We want precise control over the order in which we construct
 842          * our sub-libraries vs starting to receive commands from
 843          * sessiond (otherwise leading to errors when trying to create
 844          * sessiond before the init functions are completed).
 845          */
 846         init_usterr();
 847         init_tracepoint();
 848         ltt_ring_buffer_metadata_client_init();
 849         ltt_ring_buffer_client_overwrite_init();
 850         ltt_ring_buffer_client_discard_init();
 851
 852         timeout_mode = get_timeout(&constructor_timeout);
 853
 854         ret = sem_init(&constructor_wait, 0, 0);
 855         assert(!ret);
 856
 857         ret = setup_local_apps();
 858         if (ret) {
 859                 ERR("Error setting up to local apps");
 860         }
 861         ret = pthread_create(&local_apps.ust_listener, NULL,
 862                         ust_listener_thread, &local_apps);
 863
 864         if (local_apps.allowed) {
 865                 ret = pthread_create(&global_apps.ust_listener, NULL,
 866                                 ust_listener_thread, &global_apps);
 867         } else {
 868                 handle_register_done(&local_apps);
 869         }
 870
 871         switch (timeout_mode) {
 872         case 1: /* timeout wait */
 873                 do {
 874                         ret = sem_timedwait(&constructor_wait,
 875                                         &constructor_timeout);
 876                 } while (ret < 0 && errno == EINTR);
 877                 if (ret < 0 && errno == ETIMEDOUT) {
 878                         ERR("Timed out waiting for ltt-sessiond");
 879                 } else {
 880                         assert(!ret);
 881                 }
 882                 break;
 883         case -1:/* wait forever */
 884                 do {
 885                         ret = sem_wait(&constructor_wait);
 886                 } while (ret < 0 && errno == EINTR);
 887                 assert(!ret);
 888                 break;
 889         case 0: /* no timeout */
 890                 break;
 891         }
 892 }
 893
 894 static
 895 void lttng_ust_cleanup(int exiting)
 896 {
 897         cleanup_sock_info(&global_apps, exiting);
 898         if (local_apps.allowed) {
 899                 cleanup_sock_info(&local_apps, exiting);
 900         }
 901         /*
 902          * The teardown in this function all affect data structures
 903          * accessed under the UST lock by the listener thread. This
 904          * lock, along with the lttng_ust_comm_should_quit flag, ensure
 905          * that none of these threads are accessing this data at this
 906          * point.
 907          */
 908         lttng_ust_abi_exit();
 909         lttng_ust_events_exit();
 910         ltt_ring_buffer_client_discard_exit();
 911         ltt_ring_buffer_client_overwrite_exit();
 912         ltt_ring_buffer_metadata_client_exit();
 913         exit_tracepoint();
 914         if (!exiting) {
 915                 /* Reinitialize values for fork */
 916                 sem_count = 2;
 917                 lttng_ust_comm_should_quit = 0;
 918                 initialized = 0;
 919         }
 920 }
 921
 922 void __attribute__((destructor)) lttng_ust_exit(void)
 923 {
 924         int ret;
 925
 926         /*
 927          * Using pthread_cancel here because:
 928          * A) we don't want to hang application teardown.
 929          * B) the thread is not allocating any resource.
 930          */
 931
 932         /*
 933          * Require the communication thread to quit. Synchronize with
 934          * mutexes to ensure it is not in a mutex critical section when
 935          * pthread_cancel is later called.
 936          */
 937         ust_lock();
 938         lttng_ust_comm_should_quit = 1;
 939         ust_unlock();
 940
 941         /* cancel threads */
 942         ret = pthread_cancel(global_apps.ust_listener);
 943         if (ret) {
 944                 ERR("Error cancelling global ust listener thread");
 945         }
 946         if (local_apps.allowed) {
 947                 ret = pthread_cancel(local_apps.ust_listener);
 948                 if (ret) {
 949                         ERR("Error cancelling local ust listener thread");
 950                 }
 951         }
 952         /*
 953          * Do NOT join threads: use of sys_futex makes it impossible to
 954          * join the threads without using async-cancel, but async-cancel
 955          * is delivered by a signal, which could hit the target thread
 956          * anywhere in its code path, including while the ust_lock() is
 957          * held, causing a deadlock for the other thread. Let the OS
 958          * cleanup the threads if there are stalled in a syscall.
 959          */
 960         lttng_ust_cleanup(1);
 961 }
 962
 963 /*
 964  * We exclude the worker threads across fork and clone (except
 965  * CLONE_VM), because these system calls only keep the forking thread
 966  * running in the child.  Therefore, we don't want to call fork or clone
 967  * in the middle of an tracepoint or ust tracing state modification.
 968  * Holding this mutex protects these structures across fork and clone.
 969  */
 970 void ust_before_fork(sigset_t *save_sigset)
 971 {
 972         /*
 973          * Disable signals. This is to avoid that the child intervenes
 974          * before it is properly setup for tracing. It is safer to
 975          * disable all signals, because then we know we are not breaking
 976          * anything by restoring the original mask.
 977          */
 978         sigset_t all_sigs;
 979         int ret;
 980
 981         /* Disable signals */
 982         sigfillset(&all_sigs);
 983         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
 984         if (ret == -1) {
 985                 PERROR("sigprocmask");
 986         }
 987         ust_lock();
 988         rcu_bp_before_fork();
 989 }
 990
 991 static void ust_after_fork_common(sigset_t *restore_sigset)
 992 {
 993         int ret;
 994
 995         DBG("process %d", getpid());
 996         ust_unlock();
 997         /* Restore signals */
 998         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
 999         if (ret == -1) {
1000                 PERROR("sigprocmask");
1001         }
1002 }
1003
1004 void ust_after_fork_parent(sigset_t *restore_sigset)
1005 {
1006         DBG("process %d", getpid());
1007         rcu_bp_after_fork_parent();
1008         /* Release mutexes and reenable signals */
1009         ust_after_fork_common(restore_sigset);
1010 }
1011
1012 /*
1013  * After fork, in the child, we need to cleanup all the leftover state,
1014  * except the worker thread which already magically disappeared thanks
1015  * to the weird Linux fork semantics. After tyding up, we call
1016  * lttng_ust_init() again to start over as a new PID.
1017  *
1018  * This is meant for forks() that have tracing in the child between the
1019  * fork and following exec call (if there is any).
1020  */
1021 void ust_after_fork_child(sigset_t *restore_sigset)
1022 {
1023         DBG("process %d", getpid());
1024         /* Release urcu mutexes */
1025         rcu_bp_after_fork_child();
1026         lttng_ust_cleanup(0);
1027         lttng_context_vtid_reset();
1028         /* Release mutexes and reenable signals */
1029         ust_after_fork_common(restore_sigset);
1030         lttng_ust_init();
1031 }