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