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