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