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 #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  * Counting nesting within lttng-ust. Used to ensure that calling fork()
  83  * from liblttng-ust does not execute the pre/post fork handlers.
  84  */
  85 static int __thread lttng_ust_nest_count;
  86
  87 /*
  88  * Info about socket and associated listener thread.
  89  */
  90 struct sock_info {
  91         const char *name;
  92         pthread_t ust_listener; /* listener thread */
  93         int root_handle;
  94         int constructor_sem_posted;
  95         int allowed;
  96         int global;
  97
  98         char sock_path[PATH_MAX];
  99         int socket;
 100
 101         char wait_shm_path[PATH_MAX];
 102         char *wait_shm_mmap;
 103 };
 104
 105 /* Socket from app (connect) to session daemon (listen) for communication */
 106 struct sock_info global_apps = {
 107         .name = "global",
 108         .global = 1,
 109
 110         .root_handle = -1,
 111         .allowed = 1,
 112
 113         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 114         .socket = -1,
 115
 116         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 117 };
 118
 119 /* TODO: allow global_apps_sock_path override */
 120
 121 struct sock_info local_apps = {
 122         .name = "local",
 123         .global = 0,
 124         .root_handle = -1,
 125         .allowed = 0,   /* Check setuid bit first */
 126
 127         .socket = -1,
 128 };
 129
 130 static int wait_poll_fallback;
 131
 132 extern void ltt_ring_buffer_client_overwrite_init(void);
 133 extern void ltt_ring_buffer_client_discard_init(void);
 134 extern void ltt_ring_buffer_metadata_client_init(void);
 135 extern void ltt_ring_buffer_client_overwrite_exit(void);
 136 extern void ltt_ring_buffer_client_discard_exit(void);
 137 extern void ltt_ring_buffer_metadata_client_exit(void);
 138
 139 static
 140 int setup_local_apps(void)
 141 {
 142         const char *home_dir;
 143         uid_t uid;
 144
 145         uid = getuid();
 146         /*
 147          * Disallow per-user tracing for setuid binaries.
 148          */
 149         if (uid != geteuid()) {
 150                 local_apps.allowed = 0;
 151                 return 0;
 152         } else {
 153                 local_apps.allowed = 1;
 154         }
 155         home_dir = (const char *) getenv("HOME");
 156         if (!home_dir)
 157                 return -ENOENT;
 158         snprintf(local_apps.sock_path, PATH_MAX,
 159                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 160         snprintf(local_apps.wait_shm_path, PATH_MAX,
 161                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 162         return 0;
 163 }
 164
 165 static
 166 int register_app_to_sessiond(int socket)
 167 {
 168         ssize_t ret;
 169         int prctl_ret;
 170         struct {
 171                 uint32_t major;
 172                 uint32_t minor;
 173                 pid_t pid;
 174                 pid_t ppid;
 175                 uid_t uid;
 176                 gid_t gid;
 177                 uint32_t bits_per_long;
 178                 char name[16];  /* process name */
 179         } reg_msg;
 180
 181         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 182         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 183         reg_msg.pid = getpid();
 184         reg_msg.ppid = getppid();
 185         reg_msg.uid = getuid();
 186         reg_msg.gid = getgid();
 187         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 188         prctl_ret = prctl(PR_GET_NAME, (unsigned long) reg_msg.name, 0, 0, 0);
 189         if (prctl_ret) {
 190                 ERR("Error executing prctl");
 191                 return -errno;
 192         }
 193
 194         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 195         if (ret >= 0 && ret != sizeof(reg_msg))
 196                 return -EIO;
 197         return ret;
 198 }
 199
 200 static
 201 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 202 {
 203         ssize_t len;
 204
 205         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 206         switch (len) {
 207         case sizeof(*lur):
 208                 DBG("message successfully sent");
 209                 return 0;
 210         case -1:
 211                 if (errno == ECONNRESET) {
 212                         printf("remote end closed connection\n");
 213                         return 0;
 214                 }
 215                 return -1;
 216         default:
 217                 printf("incorrect message size: %zd\n", len);
 218                 return -1;
 219         }
 220 }
 221
 222 static
 223 int handle_register_done(struct sock_info *sock_info)
 224 {
 225         int ret;
 226
 227         if (sock_info->constructor_sem_posted)
 228                 return 0;
 229         sock_info->constructor_sem_posted = 1;
 230         if (uatomic_read(&sem_count) <= 0) {
 231                 return 0;
 232         }
 233         ret = uatomic_add_return(&sem_count, -1);
 234         if (ret == 0) {
 235                 ret = sem_post(&constructor_wait);
 236                 assert(!ret);
 237         }
 238         return 0;
 239 }
 240
 241 static
 242 int handle_message(struct sock_info *sock_info,
 243                 int sock, struct ustcomm_ust_msg *lum)
 244 {
 245         int ret = 0;
 246         const struct lttng_ust_objd_ops *ops;
 247         struct ustcomm_ust_reply lur;
 248         int shm_fd, wait_fd;
 249         union ust_args args;
 250
 251         ust_lock();
 252
 253         memset(&lur, 0, sizeof(lur));
 254
 255         if (lttng_ust_comm_should_quit) {
 256                 ret = -EPERM;
 257                 goto end;
 258         }
 259
 260         ops = objd_ops(lum->handle);
 261         if (!ops) {
 262                 ret = -ENOENT;
 263                 goto end;
 264         }
 265
 266         switch (lum->cmd) {
 267         case LTTNG_UST_REGISTER_DONE:
 268                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 269                         ret = handle_register_done(sock_info);
 270                 else
 271                         ret = -EINVAL;
 272                 break;
 273         case LTTNG_UST_RELEASE:
 274                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 275                         ret = -EPERM;
 276                 else
 277                         ret = lttng_ust_objd_unref(lum->handle);
 278                 break;
 279         default:
 280                 if (ops->cmd)
 281                         ret = ops->cmd(lum->handle, lum->cmd,
 282                                         (unsigned long) &lum->u,
 283                                         &args);
 284                 else
 285                         ret = -ENOSYS;
 286                 break;
 287         }
 288
 289 end:
 290         lur.handle = lum->handle;
 291         lur.cmd = lum->cmd;
 292         lur.ret_val = ret;
 293         if (ret >= 0) {
 294                 lur.ret_code = USTCOMM_OK;
 295         } else {
 296                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 297                 lur.ret_code = ret;
 298         }
 299         if (ret >= 0) {
 300                 switch (lum->cmd) {
 301                 case LTTNG_UST_STREAM:
 302                         /*
 303                          * Special-case reply to send stream info.
 304                          * Use lum.u output.
 305                          */
 306                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 307                         shm_fd = *args.stream.shm_fd;
 308                         wait_fd = *args.stream.wait_fd;
 309                         break;
 310                 case LTTNG_UST_METADATA:
 311                 case LTTNG_UST_CHANNEL:
 312                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 313                         shm_fd = *args.channel.shm_fd;
 314                         wait_fd = *args.channel.wait_fd;
 315                         break;
 316                 case LTTNG_UST_TRACER_VERSION:
 317                         lur.u.version = lum->u.version;
 318                         break;
 319                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 320                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 321                         break;
 322                 }
 323         }
 324         ret = send_reply(sock, &lur);
 325         if (ret < 0) {
 326                 perror("error sending reply");
 327                 goto error;
 328         }
 329
 330         if ((lum->cmd == LTTNG_UST_STREAM
 331              || lum->cmd == LTTNG_UST_CHANNEL
 332              || lum->cmd == LTTNG_UST_METADATA)
 333                         && lur.ret_code == USTCOMM_OK) {
 334                 int sendret = 0;
 335
 336                 /* we also need to send the file descriptors. */
 337                 ret = ustcomm_send_fds_unix_sock(sock,
 338                         &shm_fd, &shm_fd,
 339                         1, sizeof(int));
 340                 if (ret < 0) {
 341                         perror("send shm_fd");
 342                         sendret = ret;
 343                 }
 344                 /*
 345                  * The sessiond expects 2 file descriptors, even upon
 346                  * error.
 347                  */
 348                 ret = ustcomm_send_fds_unix_sock(sock,
 349                         &wait_fd, &wait_fd,
 350                         1, sizeof(int));
 351                 if (ret < 0) {
 352                         perror("send wait_fd");
 353                         goto error;
 354                 }
 355                 if (sendret) {
 356                         ret = sendret;
 357                         goto error;
 358                 }
 359         }
 360         /*
 361          * We still have the memory map reference, and the fds have been
 362          * sent to the sessiond. We can therefore close those fds. Note
 363          * that we keep the write side of the wait_fd open, but close
 364          * the read side.
 365          */
 366         if (lur.ret_code == USTCOMM_OK) {
 367                 switch (lum->cmd) {
 368                 case LTTNG_UST_STREAM:
 369                         if (shm_fd >= 0) {
 370                                 ret = close(shm_fd);
 371                                 if (ret) {
 372                                         PERROR("Error closing stream shm_fd");
 373                                 }
 374                                 *args.stream.shm_fd = -1;
 375                         }
 376                         if (wait_fd >= 0) {
 377                                 ret = close(wait_fd);
 378                                 if (ret) {
 379                                         PERROR("Error closing stream wait_fd");
 380                                 }
 381                                 *args.stream.wait_fd = -1;
 382                         }
 383                         break;
 384                 case LTTNG_UST_METADATA:
 385                 case LTTNG_UST_CHANNEL:
 386                         if (shm_fd >= 0) {
 387                                 ret = close(shm_fd);
 388                                 if (ret) {
 389                                         PERROR("Error closing channel shm_fd");
 390                                 }
 391                                 *args.channel.shm_fd = -1;
 392                         }
 393                         if (wait_fd >= 0) {
 394                                 ret = close(wait_fd);
 395                                 if (ret) {
 396                                         PERROR("Error closing channel wait_fd");
 397                                 }
 398                                 *args.channel.wait_fd = -1;
 399                         }
 400                         break;
 401                 }
 402         }
 403
 404 error:
 405         ust_unlock();
 406         return ret;
 407 }
 408
 409 static
 410 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 411 {
 412         int ret;
 413
 414         if (sock_info->socket != -1) {
 415                 ret = ustcomm_close_unix_sock(sock_info->socket);
 416                 if (ret) {
 417                         ERR("Error closing apps socket");
 418                 }
 419                 sock_info->socket = -1;
 420         }
 421         if (sock_info->root_handle != -1) {
 422                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 423                 if (ret) {
 424                         ERR("Error unref root handle");
 425                 }
 426                 sock_info->root_handle = -1;
 427         }
 428         sock_info->constructor_sem_posted = 0;
 429         /*
 430          * wait_shm_mmap is used by listener threads outside of the
 431          * ust lock, so we cannot tear it down ourselves, because we
 432          * cannot join on these threads. Leave this task to the OS
 433          * process exit.
 434          */
 435         if (!exiting && sock_info->wait_shm_mmap) {
 436                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 437                 if (ret) {
 438                         ERR("Error unmapping wait shm");
 439                 }
 440                 sock_info->wait_shm_mmap = NULL;
 441         }
 442 }
 443
 444 /*
 445  * Using fork to set umask in the child process (not multi-thread safe).
 446  * We deal with the shm_open vs ftruncate race (happening when the
 447  * sessiond owns the shm and does not let everybody modify it, to ensure
 448  * safety against shm_unlink) by simply letting the mmap fail and
 449  * retrying after a few seconds.
 450  * For global shm, everybody has rw access to it until the sessiond
 451  * starts.
 452  */
 453 static
 454 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 455 {
 456         int wait_shm_fd, ret;
 457         pid_t pid;
 458
 459         /*
 460          * Try to open read-only.
 461          */
 462         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 463         if (wait_shm_fd >= 0) {
 464                 goto end;
 465         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 466                 /*
 467                  * Real-only open did not work, and it's not because the
 468                  * entry was not present. It's a failure that prohibits
 469                  * using shm.
 470                  */
 471                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 472                 goto end;
 473         }
 474         /*
 475          * If the open failed because the file did not exist, try
 476          * creating it ourself.
 477          */
 478         lttng_ust_nest_count++;
 479         pid = fork();
 480         lttng_ust_nest_count--;
 481         if (pid > 0) {
 482                 int status;
 483
 484                 /*
 485                  * Parent: wait for child to return, in which case the
 486                  * shared memory map will have been created.
 487                  */
 488                 pid = wait(&status);
 489                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 490                         wait_shm_fd = -1;
 491                         goto end;
 492                 }
 493                 /*
 494                  * Try to open read-only again after creation.
 495                  */
 496                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 497                 if (wait_shm_fd < 0) {
 498                         /*
 499                          * Real-only open did not work. It's a failure
 500                          * that prohibits using shm.
 501                          */
 502                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 503                         goto end;
 504                 }
 505                 goto end;
 506         } else if (pid == 0) {
 507                 int create_mode;
 508
 509                 /* Child */
 510                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 511                 if (sock_info->global)
 512                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 513                 /*
 514                  * We're alone in a child process, so we can modify the
 515                  * process-wide umask.
 516                  */
 517                 umask(~create_mode);
 518                 /*
 519                  * Try creating shm (or get rw access).
 520                  * We don't do an exclusive open, because we allow other
 521                  * processes to create+ftruncate it concurrently.
 522                  */
 523                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 524                                 O_RDWR | O_CREAT, create_mode);
 525                 if (wait_shm_fd >= 0) {
 526                         ret = ftruncate(wait_shm_fd, mmap_size);
 527                         if (ret) {
 528                                 PERROR("ftruncate");
 529                                 exit(EXIT_FAILURE);
 530                         }
 531                         exit(EXIT_SUCCESS);
 532                 }
 533                 /*
 534                  * For local shm, we need to have rw access to accept
 535                  * opening it: this means the local sessiond will be
 536                  * able to wake us up. For global shm, we open it even
 537                  * if rw access is not granted, because the root.root
 538                  * sessiond will be able to override all rights and wake
 539                  * us up.
 540                  */
 541                 if (!sock_info->global && errno != EACCES) {
 542                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 543                         exit(EXIT_FAILURE);
 544                 }
 545                 /*
 546                  * The shm exists, but we cannot open it RW. Report
 547                  * success.
 548                  */
 549                 exit(EXIT_SUCCESS);
 550         } else {
 551                 return -1;
 552         }
 553 end:
 554         if (wait_shm_fd >= 0 && !sock_info->global) {
 555                 struct stat statbuf;
 556
 557                 /*
 558                  * Ensure that our user is the owner of the shm file for
 559                  * local shm. If we do not own the file, it means our
 560                  * sessiond will not have access to wake us up (there is
 561                  * probably a rogue process trying to fake our
 562                  * sessiond). Fallback to polling method in this case.
 563                  */
 564                 ret = fstat(wait_shm_fd, &statbuf);
 565                 if (ret) {
 566                         PERROR("fstat");
 567                         goto error_close;
 568                 }
 569                 if (statbuf.st_uid != getuid())
 570                         goto error_close;
 571         }
 572         return wait_shm_fd;
 573
 574 error_close:
 575         ret = close(wait_shm_fd);
 576         if (ret) {
 577                 PERROR("Error closing fd");
 578         }
 579         return -1;
 580 }
 581
 582 static
 583 char *get_map_shm(struct sock_info *sock_info)
 584 {
 585         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 586         int wait_shm_fd, ret;
 587         char *wait_shm_mmap;
 588
 589         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 590         if (wait_shm_fd < 0) {
 591                 goto error;
 592         }
 593         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 594                   MAP_SHARED, wait_shm_fd, 0);
 595         /* close shm fd immediately after taking the mmap reference */
 596         ret = close(wait_shm_fd);
 597         if (ret) {
 598                 PERROR("Error closing fd");
 599         }
 600         if (wait_shm_mmap == MAP_FAILED) {
 601                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 602                 goto error;
 603         }
 604         return wait_shm_mmap;
 605
 606 error:
 607         return NULL;
 608 }
 609
 610 static
 611 void wait_for_sessiond(struct sock_info *sock_info)
 612 {
 613         int ret;
 614
 615         ust_lock();
 616         if (lttng_ust_comm_should_quit) {
 617                 goto quit;
 618         }
 619         if (wait_poll_fallback) {
 620                 goto error;
 621         }
 622         if (!sock_info->wait_shm_mmap) {
 623                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 624                 if (!sock_info->wait_shm_mmap)
 625                         goto error;
 626         }
 627         ust_unlock();
 628
 629         DBG("Waiting for %s apps sessiond", sock_info->name);
 630         /* Wait for futex wakeup */
 631         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 632                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 633                         FUTEX_WAIT, 0, NULL, NULL, 0);
 634                 if (ret < 0) {
 635                         if (errno == EFAULT) {
 636                                 wait_poll_fallback = 1;
 637                                 DBG(
 638 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 639 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 640 "Please upgrade your kernel "
 641 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 642 "mainline). LTTng-UST will use polling mode fallback.");
 643                                 if (ust_debug())
 644                                         PERROR("futex");
 645                         }
 646                 }
 647         }
 648         return;
 649
 650 quit:
 651         ust_unlock();
 652         return;
 653
 654 error:
 655         ust_unlock();
 656         return;
 657 }
 658
 659 /*
 660  * This thread does not allocate any resource, except within
 661  * handle_message, within mutex protection. This mutex protects against
 662  * fork and exit.
 663  * The other moment it allocates resources is at socket connexion, which
 664  * is also protected by the mutex.
 665  */
 666 static
 667 void *ust_listener_thread(void *arg)
 668 {
 669         struct sock_info *sock_info = arg;
 670         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 671
 672         /* Restart trying to connect to the session daemon */
 673 restart:
 674         if (prev_connect_failed) {
 675                 /* Wait for sessiond availability with pipe */
 676                 wait_for_sessiond(sock_info);
 677                 if (has_waited) {
 678                         has_waited = 0;
 679                         /*
 680                          * Sleep for 5 seconds before retrying after a
 681                          * sequence of failure / wait / failure. This
 682                          * deals with a killed or broken session daemon.
 683                          */
 684                         sleep(5);
 685                 }
 686                 has_waited = 1;
 687                 prev_connect_failed = 0;
 688         }
 689         ust_lock();
 690
 691         if (lttng_ust_comm_should_quit) {
 692                 ust_unlock();
 693                 goto quit;
 694         }
 695
 696         if (sock_info->socket != -1) {
 697                 ret = ustcomm_close_unix_sock(sock_info->socket);
 698                 if (ret) {
 699                         ERR("Error closing %s apps socket", sock_info->name);
 700                 }
 701                 sock_info->socket = -1;
 702         }
 703
 704         /* Register */
 705         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 706         if (ret < 0) {
 707                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 708                 prev_connect_failed = 1;
 709                 /*
 710                  * If we cannot find the sessiond daemon, don't delay
 711                  * constructor execution.
 712                  */
 713                 ret = handle_register_done(sock_info);
 714                 assert(!ret);
 715                 ust_unlock();
 716                 goto restart;
 717         }
 718
 719         sock_info->socket = sock = ret;
 720
 721         /*
 722          * Create only one root handle per listener thread for the whole
 723          * process lifetime.
 724          */
 725         if (sock_info->root_handle == -1) {
 726                 ret = lttng_abi_create_root_handle();
 727                 if (ret < 0) {
 728                         ERR("Error creating root handle");
 729                         ust_unlock();
 730                         goto quit;
 731                 }
 732                 sock_info->root_handle = ret;
 733         }
 734
 735         ret = register_app_to_sessiond(sock);
 736         if (ret < 0) {
 737                 ERR("Error registering to %s apps socket", sock_info->name);
 738                 prev_connect_failed = 1;
 739                 /*
 740                  * If we cannot register to the sessiond daemon, don't
 741                  * delay constructor execution.
 742                  */
 743                 ret = handle_register_done(sock_info);
 744                 assert(!ret);
 745                 ust_unlock();
 746                 goto restart;
 747         }
 748         ust_unlock();
 749
 750         for (;;) {
 751                 ssize_t len;
 752                 struct ustcomm_ust_msg lum;
 753
 754                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 755                 switch (len) {
 756                 case 0: /* orderly shutdown */
 757                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 758                         ust_lock();
 759                         /*
 760                          * Either sessiond has shutdown or refused us by closing the socket.
 761                          * In either case, we don't want to delay construction execution,
 762                          * and we need to wait before retry.
 763                          */
 764                         prev_connect_failed = 1;
 765                         /*
 766                          * If we cannot register to the sessiond daemon, don't
 767                          * delay constructor execution.
 768                          */
 769                         ret = handle_register_done(sock_info);
 770                         assert(!ret);
 771                         ust_unlock();
 772                         goto end;
 773                 case sizeof(lum):
 774                         DBG("message received\n");
 775                         ret = handle_message(sock_info, sock, &lum);
 776                         if (ret < 0) {
 777                                 ERR("Error handling message for %s socket", sock_info->name);
 778                         }
 779                         continue;
 780                 case -1:
 781                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 782                         if (errno == ECONNRESET) {
 783                                 ERR("%s remote end closed connection\n", sock_info->name);
 784                                 goto end;
 785                         }
 786                         goto end;
 787                 default:
 788                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 789                         continue;
 790                 }
 791
 792         }
 793 end:
 794         goto restart;   /* try to reconnect */
 795 quit:
 796         return NULL;
 797 }
 798
 799 /*
 800  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 801  */
 802 static
 803 int get_timeout(struct timespec *constructor_timeout)
 804 {
 805         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 806         char *str_delay;
 807         int ret;
 808
 809         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 810         if (str_delay) {
 811                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 812         }
 813
 814         switch (constructor_delay_ms) {
 815         case -1:/* fall-through */
 816         case 0:
 817                 return constructor_delay_ms;
 818         default:
 819                 break;
 820         }
 821
 822         /*
 823          * If we are unable to find the current time, don't wait.
 824          */
 825         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 826         if (ret) {
 827                 return -1;
 828         }
 829         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 830         constructor_timeout->tv_nsec +=
 831                 (constructor_delay_ms % 1000UL) * 1000000UL;
 832         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 833                 constructor_timeout->tv_sec++;
 834                 constructor_timeout->tv_nsec -= 1000000000UL;
 835         }
 836         return 1;
 837 }
 838
 839 /*
 840  * sessiond monitoring thread: monitor presence of global and per-user
 841  * sessiond by polling the application common named pipe.
 842  */
 843 /* TODO */
 844
 845 void __attribute__((constructor)) lttng_ust_init(void)
 846 {
 847         struct timespec constructor_timeout;
 848         sigset_t sig_all_blocked, orig_parent_mask;
 849         int timeout_mode;
 850         int ret;
 851
 852         if (uatomic_xchg(&initialized, 1) == 1)
 853                 return;
 854
 855         /*
 856          * Fixup interdependency between TLS fixup mutex (which happens
 857          * to be the dynamic linker mutex) and ust_lock, taken within
 858          * the ust lock.
 859          */
 860         lttng_fixup_event_tls();
 861         lttng_fixup_ringbuffer_tls();
 862         lttng_fixup_vtid_tls();
 863
 864         /*
 865          * We want precise control over the order in which we construct
 866          * our sub-libraries vs starting to receive commands from
 867          * sessiond (otherwise leading to errors when trying to create
 868          * sessiond before the init functions are completed).
 869          */
 870         init_usterr();
 871         init_tracepoint();
 872         ltt_ring_buffer_metadata_client_init();
 873         ltt_ring_buffer_client_overwrite_init();
 874         ltt_ring_buffer_client_discard_init();
 875
 876         timeout_mode = get_timeout(&constructor_timeout);
 877
 878         ret = sem_init(&constructor_wait, 0, 0);
 879         assert(!ret);
 880
 881         ret = setup_local_apps();
 882         if (ret) {
 883                 ERR("Error setting up to local apps");
 884         }
 885
 886         /* A new thread created by pthread_create inherits the signal mask
 887          * from the parent. To avoid any signal being received by the
 888          * listener thread, we block all signals temporarily in the parent,
 889          * while we create the listener thread.
 890          */
 891         sigfillset(&sig_all_blocked);
 892         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 893         if (ret) {
 894                 ERR("pthread_sigmask: %s", strerror(ret));
 895         }
 896
 897         ret = pthread_create(&global_apps.ust_listener, NULL,
 898                         ust_listener_thread, &global_apps);
 899         if (ret) {
 900                 ERR("pthread_create global: %s", strerror(ret));
 901         }
 902         if (local_apps.allowed) {
 903                 ret = pthread_create(&local_apps.ust_listener, NULL,
 904                                 ust_listener_thread, &local_apps);
 905                 if (ret) {
 906                         ERR("pthread_create local: %s", strerror(ret));
 907                 }
 908         } else {
 909                 handle_register_done(&local_apps);
 910         }
 911
 912         /* Restore original signal mask in parent */
 913         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
 914         if (ret) {
 915                 ERR("pthread_sigmask: %s", strerror(ret));
 916         }
 917
 918         switch (timeout_mode) {
 919         case 1: /* timeout wait */
 920                 do {
 921                         ret = sem_timedwait(&constructor_wait,
 922                                         &constructor_timeout);
 923                 } while (ret < 0 && errno == EINTR);
 924                 if (ret < 0 && errno == ETIMEDOUT) {
 925                         ERR("Timed out waiting for ltt-sessiond");
 926                 } else {
 927                         assert(!ret);
 928                 }
 929                 break;
 930         case -1:/* wait forever */
 931                 do {
 932                         ret = sem_wait(&constructor_wait);
 933                 } while (ret < 0 && errno == EINTR);
 934                 assert(!ret);
 935                 break;
 936         case 0: /* no timeout */
 937                 break;
 938         }
 939 }
 940
 941 static
 942 void lttng_ust_cleanup(int exiting)
 943 {
 944         cleanup_sock_info(&global_apps, exiting);
 945         if (local_apps.allowed) {
 946                 cleanup_sock_info(&local_apps, exiting);
 947         }
 948         /*
 949          * The teardown in this function all affect data structures
 950          * accessed under the UST lock by the listener thread. This
 951          * lock, along with the lttng_ust_comm_should_quit flag, ensure
 952          * that none of these threads are accessing this data at this
 953          * point.
 954          */
 955         lttng_ust_abi_exit();
 956         lttng_ust_events_exit();
 957         ltt_ring_buffer_client_discard_exit();
 958         ltt_ring_buffer_client_overwrite_exit();
 959         ltt_ring_buffer_metadata_client_exit();
 960         exit_tracepoint();
 961         if (!exiting) {
 962                 /* Reinitialize values for fork */
 963                 sem_count = 2;
 964                 lttng_ust_comm_should_quit = 0;
 965                 initialized = 0;
 966         }
 967 }
 968
 969 void __attribute__((destructor)) lttng_ust_exit(void)
 970 {
 971         int ret;
 972
 973         /*
 974          * Using pthread_cancel here because:
 975          * A) we don't want to hang application teardown.
 976          * B) the thread is not allocating any resource.
 977          */
 978
 979         /*
 980          * Require the communication thread to quit. Synchronize with
 981          * mutexes to ensure it is not in a mutex critical section when
 982          * pthread_cancel is later called.
 983          */
 984         ust_lock();
 985         lttng_ust_comm_should_quit = 1;
 986         ust_unlock();
 987
 988         /* cancel threads */
 989         ret = pthread_cancel(global_apps.ust_listener);
 990         if (ret) {
 991                 ERR("Error cancelling global ust listener thread: %s",
 992                         strerror(ret));
 993         }
 994         if (local_apps.allowed) {
 995                 ret = pthread_cancel(local_apps.ust_listener);
 996                 if (ret) {
 997                         ERR("Error cancelling local ust listener thread: %s",
 998                                 strerror(ret));
 999                 }
1000         }
1001         /*
1002          * Do NOT join threads: use of sys_futex makes it impossible to
1003          * join the threads without using async-cancel, but async-cancel
1004          * is delivered by a signal, which could hit the target thread
1005          * anywhere in its code path, including while the ust_lock() is
1006          * held, causing a deadlock for the other thread. Let the OS
1007          * cleanup the threads if there are stalled in a syscall.
1008          */
1009         lttng_ust_cleanup(1);
1010 }
1011
1012 /*
1013  * We exclude the worker threads across fork and clone (except
1014  * CLONE_VM), because these system calls only keep the forking thread
1015  * running in the child.  Therefore, we don't want to call fork or clone
1016  * in the middle of an tracepoint or ust tracing state modification.
1017  * Holding this mutex protects these structures across fork and clone.
1018  */
1019 void ust_before_fork(sigset_t *save_sigset)
1020 {
1021         /*
1022          * Disable signals. This is to avoid that the child intervenes
1023          * before it is properly setup for tracing. It is safer to
1024          * disable all signals, because then we know we are not breaking
1025          * anything by restoring the original mask.
1026          */
1027         sigset_t all_sigs;
1028         int ret;
1029
1030         if (lttng_ust_nest_count)
1031                 return;
1032         /* Disable signals */
1033         sigfillset(&all_sigs);
1034         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1035         if (ret == -1) {
1036                 PERROR("sigprocmask");
1037         }
1038         ust_lock();
1039         rcu_bp_before_fork();
1040 }
1041
1042 static void ust_after_fork_common(sigset_t *restore_sigset)
1043 {
1044         int ret;
1045
1046         DBG("process %d", getpid());
1047         ust_unlock();
1048         /* Restore signals */
1049         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1050         if (ret == -1) {
1051                 PERROR("sigprocmask");
1052         }
1053 }
1054
1055 void ust_after_fork_parent(sigset_t *restore_sigset)
1056 {
1057         if (lttng_ust_nest_count)
1058                 return;
1059         DBG("process %d", getpid());
1060         rcu_bp_after_fork_parent();
1061         /* Release mutexes and reenable signals */
1062         ust_after_fork_common(restore_sigset);
1063 }
1064
1065 /*
1066  * After fork, in the child, we need to cleanup all the leftover state,
1067  * except the worker thread which already magically disappeared thanks
1068  * to the weird Linux fork semantics. After tyding up, we call
1069  * lttng_ust_init() again to start over as a new PID.
1070  *
1071  * This is meant for forks() that have tracing in the child between the
1072  * fork and following exec call (if there is any).
1073  */
1074 void ust_after_fork_child(sigset_t *restore_sigset)
1075 {
1076         if (lttng_ust_nest_count)
1077                 return;
1078         DBG("process %d", getpid());
1079         /* Release urcu mutexes */
1080         rcu_bp_after_fork_child();
1081         lttng_ust_cleanup(0);
1082         lttng_context_vtid_reset();
1083         /* Release mutexes and reenable signals */
1084         ust_after_fork_common(restore_sigset);
1085         lttng_ust_init();
1086 }