libust/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
  41 #include <lttng-ust-comm.h>
  42 #include <ust/usterr-signal-safe.h>
  43 #include <ust/lttng-ust-abi.h>
  44 #include <ust/tracepoint.h>
  45 #include <ust/tracepoint-internal.h>
  46 #include <ust/ust.h>
  47 #include "ltt-tracer-core.h"
  48
  49 /*
  50  * Has lttng ust comm constructor been called ?
  51  */
  52 static int initialized;
  53
  54 /*
  55  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  56  * Held when handling a command, also held by fork() to deal with
  57  * removal of threads, and by exit path.
  58  */
  59
  60 /* Should the ust comm thread quit ? */
  61 static int lttng_ust_comm_should_quit;
  62
  63 /*
  64  * Wait for either of these before continuing to the main
  65  * program:
  66  * - the register_done message from sessiond daemon
  67  *   (will let the sessiond daemon enable sessions before main
  68  *   starts.)
  69  * - sessiond daemon is not reachable.
  70  * - timeout (ensuring applications are resilient to session
  71  *   daemon problems).
  72  */
  73 static sem_t constructor_wait;
  74 /*
  75  * Doing this for both the global and local sessiond.
  76  */
  77 static int sem_count = { 2 };
  78
  79 /*
  80  * Info about socket and associated listener thread.
  81  */
  82 struct sock_info {
  83         const char *name;
  84         pthread_t ust_listener; /* listener thread */
  85         int root_handle;
  86         int constructor_sem_posted;
  87         int allowed;
  88         int global;
  89
  90         char sock_path[PATH_MAX];
  91         int socket;
  92
  93         char wait_shm_path[PATH_MAX];
  94         char *wait_shm_mmap;
  95 };
  96
  97 /* Socket from app (connect) to session daemon (listen) for communication */
  98 struct sock_info global_apps = {
  99         .name = "global",
 100         .global = 1,
 101
 102         .root_handle = -1,
 103         .allowed = 1,
 104
 105         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 106         .socket = -1,
 107
 108         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 109 };
 110
 111 /* TODO: allow global_apps_sock_path override */
 112
 113 struct sock_info local_apps = {
 114         .name = "local",
 115         .global = 0,
 116         .root_handle = -1,
 117         .allowed = 0,   /* Check setuid bit first */
 118
 119         .socket = -1,
 120 };
 121
 122 static int wait_poll_fallback;
 123
 124 extern void ltt_ring_buffer_client_overwrite_init(void);
 125 extern void ltt_ring_buffer_client_discard_init(void);
 126 extern void ltt_ring_buffer_metadata_client_init(void);
 127 extern void ltt_ring_buffer_client_overwrite_exit(void);
 128 extern void ltt_ring_buffer_client_discard_exit(void);
 129 extern void ltt_ring_buffer_metadata_client_exit(void);
 130
 131 static
 132 int setup_local_apps(void)
 133 {
 134         const char *home_dir;
 135         uid_t uid;
 136
 137         uid = getuid();
 138         /*
 139          * Disallow per-user tracing for setuid binaries.
 140          */
 141         if (uid != geteuid()) {
 142                 local_apps.allowed = 0;
 143                 return 0;
 144         } else {
 145                 local_apps.allowed = 1;
 146         }
 147         home_dir = (const char *) getenv("HOME");
 148         if (!home_dir)
 149                 return -ENOENT;
 150         snprintf(local_apps.sock_path, PATH_MAX,
 151                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 152         snprintf(local_apps.wait_shm_path, PATH_MAX,
 153                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 154         return 0;
 155 }
 156
 157 static
 158 int register_app_to_sessiond(int socket)
 159 {
 160         ssize_t ret;
 161         int prctl_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                 char name[16];  /* process name */
 170         } reg_msg;
 171
 172         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 173         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 174         reg_msg.pid = getpid();
 175         reg_msg.ppid = getppid();
 176         reg_msg.uid = getuid();
 177         reg_msg.gid = getgid();
 178         prctl_ret = prctl(PR_GET_NAME, (unsigned long) reg_msg.name, 0, 0, 0);
 179         if (prctl_ret) {
 180                 ERR("Error executing prctl");
 181                 return -errno;
 182         }
 183
 184         ret = lttcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 185         if (ret >= 0 && ret != sizeof(reg_msg))
 186                 return -EIO;
 187         return ret;
 188 }
 189
 190 static
 191 int send_reply(int sock, struct lttcomm_ust_reply *lur)
 192 {
 193         ssize_t len;
 194
 195         len = lttcomm_send_unix_sock(sock, lur, sizeof(*lur));
 196         switch (len) {
 197         case sizeof(*lur):
 198                 DBG("message successfully sent");
 199                 return 0;
 200         case -1:
 201                 if (errno == ECONNRESET) {
 202                         printf("remote end closed connection\n");
 203                         return 0;
 204                 }
 205                 return -1;
 206         default:
 207                 printf("incorrect message size: %zd\n", len);
 208                 return -1;
 209         }
 210 }
 211
 212 static
 213 int handle_register_done(struct sock_info *sock_info)
 214 {
 215         int ret;
 216
 217         if (sock_info->constructor_sem_posted)
 218                 return 0;
 219         sock_info->constructor_sem_posted = 1;
 220         if (uatomic_read(&sem_count) <= 0) {
 221                 return 0;
 222         }
 223         ret = uatomic_add_return(&sem_count, -1);
 224         if (ret == 0) {
 225                 ret = sem_post(&constructor_wait);
 226                 assert(!ret);
 227         }
 228         return 0;
 229 }
 230
 231 static
 232 int handle_message(struct sock_info *sock_info,
 233                 int sock, struct lttcomm_ust_msg *lum)
 234 {
 235         int ret = 0;
 236         const struct objd_ops *ops;
 237         struct lttcomm_ust_reply lur;
 238         int shm_fd, wait_fd;
 239
 240         ust_lock();
 241
 242         memset(&lur, 0, sizeof(lur));
 243
 244         if (lttng_ust_comm_should_quit) {
 245                 ret = -EPERM;
 246                 goto end;
 247         }
 248
 249         ops = objd_ops(lum->handle);
 250         if (!ops) {
 251                 ret = -ENOENT;
 252                 goto end;
 253         }
 254
 255         switch (lum->cmd) {
 256         case LTTNG_UST_REGISTER_DONE:
 257                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 258                         ret = handle_register_done(sock_info);
 259                 else
 260                         ret = -EINVAL;
 261                 break;
 262         case LTTNG_UST_RELEASE:
 263                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 264                         ret = -EPERM;
 265                 else
 266                         ret = objd_unref(lum->handle);
 267                 break;
 268         default:
 269                 if (ops->cmd)
 270                         ret = ops->cmd(lum->handle, lum->cmd,
 271                                         (unsigned long) &lum->u);
 272                 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 = LTTCOMM_OK;
 283         } else {
 284                 //lur.ret_code = LTTCOMM_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 = lum->u.stream.memory_map_size;
 294                 shm_fd = lum->u.stream.shm_fd;
 295                 wait_fd = lum->u.stream.wait_fd;
 296                 break;
 297         case LTTNG_UST_METADATA:
 298         case LTTNG_UST_CHANNEL:
 299                 lur.u.channel.memory_map_size = lum->u.channel.memory_map_size;
 300                 shm_fd = lum->u.channel.shm_fd;
 301                 wait_fd = lum->u.channel.wait_fd;
 302                 break;
 303         }
 304         ret = send_reply(sock, &lur);
 305         if (ret < 0) {
 306                 perror("error sending reply");
 307                 goto error;
 308         }
 309
 310         if ((lum->cmd == LTTNG_UST_STREAM
 311              || lum->cmd == LTTNG_UST_CHANNEL
 312              || lum->cmd == LTTNG_UST_METADATA)
 313                         && lur.ret_code == LTTCOMM_OK) {
 314                 /* we also need to send the file descriptors. */
 315                 ret = lttcomm_send_fds_unix_sock(sock,
 316                         &shm_fd, &shm_fd,
 317                         1, sizeof(int));
 318                 if (ret < 0) {
 319                         perror("send shm_fd");
 320                         goto error;
 321                 }
 322                 ret = lttcomm_send_fds_unix_sock(sock,
 323                         &wait_fd, &wait_fd,
 324                         1, sizeof(int));
 325                 if (ret < 0) {
 326                         perror("send wait_fd");
 327                         goto error;
 328                 }
 329         }
 330 error:
 331         ust_unlock();
 332         return ret;
 333 }
 334
 335 static
 336 void cleanup_sock_info(struct sock_info *sock_info)
 337 {
 338         int ret;
 339
 340         if (sock_info->socket != -1) {
 341                 ret = close(sock_info->socket);
 342                 if (ret) {
 343                         ERR("Error closing apps socket");
 344                 }
 345                 sock_info->socket = -1;
 346         }
 347         if (sock_info->root_handle != -1) {
 348                 ret = objd_unref(sock_info->root_handle);
 349                 if (ret) {
 350                         ERR("Error unref root handle");
 351                 }
 352                 sock_info->root_handle = -1;
 353         }
 354         sock_info->constructor_sem_posted = 0;
 355         if (sock_info->wait_shm_mmap) {
 356                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 357                 if (ret) {
 358                         ERR("Error unmapping wait shm");
 359                 }
 360                 sock_info->wait_shm_mmap = NULL;
 361         }
 362 }
 363
 364 /*
 365  * Using fork to set umask in the child process (not multi-thread safe).
 366  * We deal with the shm_open vs ftruncate race (happening when the
 367  * sessiond owns the shm and does not let everybody modify it, to ensure
 368  * safety against shm_unlink) by simply letting the mmap fail and
 369  * retrying after a few seconds.
 370  * For global shm, everybody has rw access to it until the sessiond
 371  * starts.
 372  */
 373 static
 374 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 375 {
 376         int wait_shm_fd, ret;
 377         pid_t pid;
 378
 379         /*
 380          * Try to open read-only.
 381          */
 382         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 383         if (wait_shm_fd >= 0) {
 384                 goto end;
 385         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 386                 /*
 387                  * Real-only open did not work, and it's not because the
 388                  * entry was not present. It's a failure that prohibits
 389                  * using shm.
 390                  */
 391                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 392                 goto end;
 393         }
 394         /*
 395          * If the open failed because the file did not exist, try
 396          * creating it ourself.
 397          */
 398         pid = fork();
 399         if (pid > 0) {
 400                 int status;
 401
 402                 /*
 403                  * Parent: wait for child to return, in which case the
 404                  * shared memory map will have been created.
 405                  */
 406                 pid = wait(&status);
 407                 if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 408                         wait_shm_fd = -1;
 409                         goto end;
 410                 }
 411                 /*
 412                  * Try to open read-only again after creation.
 413                  */
 414                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 415                 if (wait_shm_fd < 0) {
 416                         /*
 417                          * Real-only open did not work. It's a failure
 418                          * that prohibits using shm.
 419                          */
 420                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 421                         goto end;
 422                 }
 423                 goto end;
 424         } else if (pid == 0) {
 425                 int create_mode;
 426
 427                 /* Child */
 428                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 429                 if (sock_info->global)
 430                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 431                 /*
 432                  * We're alone in a child process, so we can modify the
 433                  * process-wide umask.
 434                  */
 435                 umask(~create_mode);
 436                 /*
 437                  * Try creating shm (or get rw access).
 438                  * We don't do an exclusive open, because we allow other
 439                  * processes to create+ftruncate it concurrently.
 440                  */
 441                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 442                                 O_RDWR | O_CREAT, create_mode);
 443                 if (wait_shm_fd >= 0) {
 444                         ret = ftruncate(wait_shm_fd, mmap_size);
 445                         if (ret) {
 446                                 PERROR("ftruncate");
 447                                 exit(EXIT_FAILURE);
 448                         }
 449                         exit(EXIT_SUCCESS);
 450                 }
 451                 /*
 452                  * For local shm, we need to have rw access to accept
 453                  * opening it: this means the local sessiond will be
 454                  * able to wake us up. For global shm, we open it even
 455                  * if rw access is not granted, because the root.root
 456                  * sessiond will be able to override all rights and wake
 457                  * us up.
 458                  */
 459                 if (!sock_info->global && errno != EACCES) {
 460                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 461                         exit(EXIT_FAILURE);
 462                 }
 463                 /*
 464                  * The shm exists, but we cannot open it RW. Report
 465                  * success.
 466                  */
 467                 exit(EXIT_SUCCESS);
 468         } else {
 469                 return -1;
 470         }
 471 end:
 472         if (wait_shm_fd >= 0 && !sock_info->global) {
 473                 struct stat statbuf;
 474
 475                 /*
 476                  * Ensure that our user is the owner of the shm file for
 477                  * local shm. If we do not own the file, it means our
 478                  * sessiond will not have access to wake us up (there is
 479                  * probably a rogue process trying to fake our
 480                  * sessiond). Fallback to polling method in this case.
 481                  */
 482                 ret = fstat(wait_shm_fd, &statbuf);
 483                 if (ret) {
 484                         PERROR("fstat");
 485                         goto error_close;
 486                 }
 487                 if (statbuf.st_uid != getuid())
 488                         goto error_close;
 489         }
 490         return wait_shm_fd;
 491
 492 error_close:
 493         ret = close(wait_shm_fd);
 494         if (ret) {
 495                 PERROR("Error closing fd");
 496         }
 497         return -1;
 498 }
 499
 500 static
 501 char *get_map_shm(struct sock_info *sock_info)
 502 {
 503         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 504         int wait_shm_fd, ret;
 505         char *wait_shm_mmap;
 506
 507         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 508         if (wait_shm_fd < 0) {
 509                 goto error;
 510         }
 511         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 512                   MAP_SHARED, wait_shm_fd, 0);
 513         /* close shm fd immediately after taking the mmap reference */
 514         ret = close(wait_shm_fd);
 515         if (ret) {
 516                 PERROR("Error closing fd");
 517         }
 518         if (wait_shm_mmap == MAP_FAILED) {
 519                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 520                 goto error;
 521         }
 522         return wait_shm_mmap;
 523
 524 error:
 525         return NULL;
 526 }
 527
 528 static
 529 void wait_for_sessiond(struct sock_info *sock_info)
 530 {
 531         int ret;
 532
 533         ust_lock();
 534         if (lttng_ust_comm_should_quit) {
 535                 goto quit;
 536         }
 537         if (wait_poll_fallback) {
 538                 goto error;
 539         }
 540         if (!sock_info->wait_shm_mmap) {
 541                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 542                 if (!sock_info->wait_shm_mmap)
 543                         goto error;
 544         }
 545         ust_unlock();
 546
 547         DBG("Waiting for %s apps sessiond", sock_info->name);
 548         /* Wait for futex wakeup */
 549         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 550                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 551                         FUTEX_WAIT, 0, NULL, NULL, 0);
 552                 if (ret < 0) {
 553                         if (errno == EFAULT) {
 554                                 wait_poll_fallback = 1;
 555                                 ERR(
 556 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 557 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 558 "Please upgrade your kernel "
 559 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 560 "mainline). LTTng-UST will use polling mode fallback.");
 561                         }
 562                         PERROR("futex");
 563                 }
 564         }
 565         return;
 566
 567 quit:
 568         ust_unlock();
 569         return;
 570
 571 error:
 572         ust_unlock();
 573         return;
 574 }
 575
 576 /*
 577  * This thread does not allocate any resource, except within
 578  * handle_message, within mutex protection. This mutex protects against
 579  * fork and exit.
 580  * The other moment it allocates resources is at socket connexion, which
 581  * is also protected by the mutex.
 582  */
 583 static
 584 void *ust_listener_thread(void *arg)
 585 {
 586         struct sock_info *sock_info = arg;
 587         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 588
 589         /* Restart trying to connect to the session daemon */
 590 restart:
 591         if (prev_connect_failed) {
 592                 /* Wait for sessiond availability with pipe */
 593                 wait_for_sessiond(sock_info);
 594                 if (has_waited) {
 595                         has_waited = 0;
 596                         /*
 597                          * Sleep for 5 seconds before retrying after a
 598                          * sequence of failure / wait / failure. This
 599                          * deals with a killed or broken session daemon.
 600                          */
 601                         sleep(5);
 602                 }
 603                 has_waited = 1;
 604                 prev_connect_failed = 0;
 605         }
 606         ust_lock();
 607
 608         if (lttng_ust_comm_should_quit) {
 609                 ust_unlock();
 610                 goto quit;
 611         }
 612
 613         if (sock_info->socket != -1) {
 614                 ret = close(sock_info->socket);
 615                 if (ret) {
 616                         ERR("Error closing %s apps socket", sock_info->name);
 617                 }
 618                 sock_info->socket = -1;
 619         }
 620
 621         /* Register */
 622         ret = lttcomm_connect_unix_sock(sock_info->sock_path);
 623         if (ret < 0) {
 624                 ERR("Error connecting to %s apps socket", sock_info->name);
 625                 prev_connect_failed = 1;
 626                 /*
 627                  * If we cannot find the sessiond daemon, don't delay
 628                  * constructor execution.
 629                  */
 630                 ret = handle_register_done(sock_info);
 631                 assert(!ret);
 632                 ust_unlock();
 633                 goto restart;
 634         }
 635
 636         sock_info->socket = sock = ret;
 637
 638         /*
 639          * Create only one root handle per listener thread for the whole
 640          * process lifetime.
 641          */
 642         if (sock_info->root_handle == -1) {
 643                 ret = lttng_abi_create_root_handle();
 644                 if (ret) {
 645                         ERR("Error creating root handle");
 646                         ust_unlock();
 647                         goto quit;
 648                 }
 649                 sock_info->root_handle = ret;
 650         }
 651
 652         ret = register_app_to_sessiond(sock);
 653         if (ret < 0) {
 654                 ERR("Error registering to %s apps socket", sock_info->name);
 655                 prev_connect_failed = 1;
 656                 /*
 657                  * If we cannot register to the sessiond daemon, don't
 658                  * delay constructor execution.
 659                  */
 660                 ret = handle_register_done(sock_info);
 661                 assert(!ret);
 662                 ust_unlock();
 663                 goto restart;
 664         }
 665         ust_unlock();
 666
 667         for (;;) {
 668                 ssize_t len;
 669                 struct lttcomm_ust_msg lum;
 670
 671                 len = lttcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 672                 switch (len) {
 673                 case 0: /* orderly shutdown */
 674                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 675                         goto end;
 676                 case sizeof(lum):
 677                         DBG("message received\n");
 678                         ret = handle_message(sock_info, sock, &lum);
 679                         if (ret < 0) {
 680                                 ERR("Error handling message for %s socket", sock_info->name);
 681                         }
 682                         continue;
 683                 case -1:
 684                         if (errno == ECONNRESET) {
 685                                 ERR("%s remote end closed connection\n", sock_info->name);
 686                                 goto end;
 687                         }
 688                         goto end;
 689                 default:
 690                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 691                         continue;
 692                 }
 693
 694         }
 695 end:
 696         goto restart;   /* try to reconnect */
 697 quit:
 698         return NULL;
 699 }
 700
 701 /*
 702  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 703  */
 704 static
 705 int get_timeout(struct timespec *constructor_timeout)
 706 {
 707         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 708         char *str_delay;
 709         int ret;
 710
 711         str_delay = getenv("UST_REGISTER_TIMEOUT");
 712         if (str_delay) {
 713                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 714         }
 715
 716         switch (constructor_delay_ms) {
 717         case -1:/* fall-through */
 718         case 0:
 719                 return constructor_delay_ms;
 720         default:
 721                 break;
 722         }
 723
 724         /*
 725          * If we are unable to find the current time, don't wait.
 726          */
 727         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 728         if (ret) {
 729                 return -1;
 730         }
 731         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 732         constructor_timeout->tv_nsec +=
 733                 (constructor_delay_ms % 1000UL) * 1000000UL;
 734         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 735                 constructor_timeout->tv_sec++;
 736                 constructor_timeout->tv_nsec -= 1000000000UL;
 737         }
 738         return 1;
 739 }
 740
 741 /*
 742  * sessiond monitoring thread: monitor presence of global and per-user
 743  * sessiond by polling the application common named pipe.
 744  */
 745 /* TODO */
 746
 747 void __attribute__((constructor)) lttng_ust_init(void)
 748 {
 749         struct timespec constructor_timeout;
 750         int timeout_mode;
 751         int ret;
 752
 753         if (uatomic_xchg(&initialized, 1) == 1)
 754                 return;
 755
 756         /*
 757          * We want precise control over the order in which we construct
 758          * our sub-libraries vs starting to receive commands from
 759          * sessiond (otherwise leading to errors when trying to create
 760          * sessiond before the init functions are completed).
 761          */
 762         init_usterr();
 763         init_tracepoint();
 764         ltt_ring_buffer_metadata_client_init();
 765         ltt_ring_buffer_client_overwrite_init();
 766         ltt_ring_buffer_client_discard_init();
 767
 768         timeout_mode = get_timeout(&constructor_timeout);
 769
 770         ret = sem_init(&constructor_wait, 0, 0);
 771         assert(!ret);
 772
 773         ret = setup_local_apps();
 774         if (ret) {
 775                 ERR("Error setting up to local apps");
 776         }
 777         ret = pthread_create(&local_apps.ust_listener, NULL,
 778                         ust_listener_thread, &local_apps);
 779
 780         if (local_apps.allowed) {
 781                 ret = pthread_create(&global_apps.ust_listener, NULL,
 782                                 ust_listener_thread, &global_apps);
 783         } else {
 784                 handle_register_done(&local_apps);
 785         }
 786
 787         switch (timeout_mode) {
 788         case 1: /* timeout wait */
 789                 do {
 790                         ret = sem_timedwait(&constructor_wait,
 791                                         &constructor_timeout);
 792                 } while (ret < 0 && errno == EINTR);
 793                 if (ret < 0 && errno == ETIMEDOUT) {
 794                         ERR("Timed out waiting for ltt-sessiond");
 795                 } else {
 796                         assert(!ret);
 797                 }
 798                 break;
 799         case -1:/* wait forever */
 800                 do {
 801                         ret = sem_wait(&constructor_wait);
 802                 } while (ret < 0 && errno == EINTR);
 803                 assert(!ret);
 804                 break;
 805         case 0: /* no timeout */
 806                 break;
 807         }
 808 }
 809
 810 static
 811 void lttng_ust_cleanup(int exiting)
 812 {
 813         cleanup_sock_info(&global_apps);
 814         if (local_apps.allowed) {
 815                 cleanup_sock_info(&local_apps);
 816         }
 817         lttng_ust_abi_exit();
 818         ltt_events_exit();
 819         ltt_ring_buffer_client_discard_exit();
 820         ltt_ring_buffer_client_overwrite_exit();
 821         ltt_ring_buffer_metadata_client_exit();
 822         exit_tracepoint();
 823         if (!exiting) {
 824                 /* Reinitialize values for fork */
 825                 sem_count = 2;
 826                 lttng_ust_comm_should_quit = 0;
 827                 initialized = 0;
 828         }
 829 }
 830
 831 void __attribute__((destructor)) lttng_ust_exit(void)
 832 {
 833         int ret;
 834
 835         /*
 836          * Using pthread_cancel here because:
 837          * A) we don't want to hang application teardown.
 838          * B) the thread is not allocating any resource.
 839          */
 840
 841         /*
 842          * Require the communication thread to quit. Synchronize with
 843          * mutexes to ensure it is not in a mutex critical section when
 844          * pthread_cancel is later called.
 845          */
 846         ust_lock();
 847         lttng_ust_comm_should_quit = 1;
 848         ust_unlock();
 849
 850         ret = pthread_cancel(global_apps.ust_listener);
 851         if (ret) {
 852                 ERR("Error cancelling global ust listener thread");
 853         }
 854         if (local_apps.allowed) {
 855                 ret = pthread_cancel(local_apps.ust_listener);
 856                 if (ret) {
 857                         ERR("Error cancelling local ust listener thread");
 858                 }
 859         }
 860         lttng_ust_cleanup(1);
 861 }
 862
 863 /*
 864  * We exclude the worker threads across fork and clone (except
 865  * CLONE_VM), because these system calls only keep the forking thread
 866  * running in the child.  Therefore, we don't want to call fork or clone
 867  * in the middle of an tracepoint or ust tracing state modification.
 868  * Holding this mutex protects these structures across fork and clone.
 869  */
 870 void ust_before_fork(ust_fork_info_t *fork_info)
 871 {
 872         /*
 873          * Disable signals. This is to avoid that the child intervenes
 874          * before it is properly setup for tracing. It is safer to
 875          * disable all signals, because then we know we are not breaking
 876          * anything by restoring the original mask.
 877          */
 878         sigset_t all_sigs;
 879         int ret;
 880
 881         /* Disable signals */
 882         sigfillset(&all_sigs);
 883         ret = sigprocmask(SIG_BLOCK, &all_sigs, &fork_info->orig_sigs);
 884         if (ret == -1) {
 885                 PERROR("sigprocmask");
 886         }
 887         ust_lock();
 888         rcu_bp_before_fork();
 889 }
 890
 891 static void ust_after_fork_common(ust_fork_info_t *fork_info)
 892 {
 893         int ret;
 894
 895         DBG("process %d", getpid());
 896         ust_unlock();
 897         /* Restore signals */
 898         ret = sigprocmask(SIG_SETMASK, &fork_info->orig_sigs, NULL);
 899         if (ret == -1) {
 900                 PERROR("sigprocmask");
 901         }
 902 }
 903
 904 void ust_after_fork_parent(ust_fork_info_t *fork_info)
 905 {
 906         DBG("process %d", getpid());
 907         rcu_bp_after_fork_parent();
 908         /* Release mutexes and reenable signals */
 909         ust_after_fork_common(fork_info);
 910 }
 911
 912 /*
 913  * After fork, in the child, we need to cleanup all the leftover state,
 914  * except the worker thread which already magically disappeared thanks
 915  * to the weird Linux fork semantics. After tyding up, we call
 916  * lttng_ust_init() again to start over as a new PID.
 917  *
 918  * This is meant for forks() that have tracing in the child between the
 919  * fork and following exec call (if there is any).
 920  */
 921 void ust_after_fork_child(ust_fork_info_t *fork_info)
 922 {
 923         DBG("process %d", getpid());
 924         /* Release urcu mutexes */
 925         rcu_bp_after_fork_child();
 926         lttng_ust_cleanup(0);
 927         /* Release mutexes and reenable signals */
 928         ust_after_fork_common(fork_info);
 929         lttng_ust_init();
 930 }