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
  50 /*
  51  * Has lttng ust comm constructor been called ?
  52  */
  53 static int initialized;
  54
  55 /*
  56  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  57  * Held when handling a command, also held by fork() to deal with
  58  * removal of threads, and by exit path.
  59  */
  60
  61 /* Should the ust comm thread quit ? */
  62 static int lttng_ust_comm_should_quit;
  63
  64 /*
  65  * Wait for either of these before continuing to the main
  66  * program:
  67  * - the register_done message from sessiond daemon
  68  *   (will let the sessiond daemon enable sessions before main
  69  *   starts.)
  70  * - sessiond daemon is not reachable.
  71  * - timeout (ensuring applications are resilient to session
  72  *   daemon problems).
  73  */
  74 static sem_t constructor_wait;
  75 /*
  76  * Doing this for both the global and local sessiond.
  77  */
  78 static int sem_count = { 2 };
  79
  80 /*
  81  * Info about socket and associated listener thread.
  82  */
  83 struct sock_info {
  84         const char *name;
  85         pthread_t ust_listener; /* listener thread */
  86         int root_handle;
  87         int constructor_sem_posted;
  88         int allowed;
  89         int global;
  90
  91         char sock_path[PATH_MAX];
  92         int socket;
  93
  94         char wait_shm_path[PATH_MAX];
  95         char *wait_shm_mmap;
  96 };
  97
  98 /* Socket from app (connect) to session daemon (listen) for communication */
  99 struct sock_info global_apps = {
 100         .name = "global",
 101         .global = 1,
 102
 103         .root_handle = -1,
 104         .allowed = 1,
 105
 106         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 107         .socket = -1,
 108
 109         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 110 };
 111
 112 /* TODO: allow global_apps_sock_path override */
 113
 114 struct sock_info local_apps = {
 115         .name = "local",
 116         .global = 0,
 117         .root_handle = -1,
 118         .allowed = 0,   /* Check setuid bit first */
 119
 120         .socket = -1,
 121 };
 122
 123 static int wait_poll_fallback;
 124
 125 extern void ltt_ring_buffer_client_overwrite_init(void);
 126 extern void ltt_ring_buffer_client_discard_init(void);
 127 extern void ltt_ring_buffer_metadata_client_init(void);
 128 extern void ltt_ring_buffer_client_overwrite_exit(void);
 129 extern void ltt_ring_buffer_client_discard_exit(void);
 130 extern void ltt_ring_buffer_metadata_client_exit(void);
 131
 132 static
 133 int setup_local_apps(void)
 134 {
 135         const char *home_dir;
 136         uid_t uid;
 137
 138         uid = getuid();
 139         /*
 140          * Disallow per-user tracing for setuid binaries.
 141          */
 142         if (uid != geteuid()) {
 143                 local_apps.allowed = 0;
 144                 return 0;
 145         } else {
 146                 local_apps.allowed = 1;
 147         }
 148         home_dir = (const char *) getenv("HOME");
 149         if (!home_dir)
 150                 return -ENOENT;
 151         snprintf(local_apps.sock_path, PATH_MAX,
 152                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 153         snprintf(local_apps.wait_shm_path, PATH_MAX,
 154                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 155         return 0;
 156 }
 157
 158 static
 159 int register_app_to_sessiond(int socket)
 160 {
 161         ssize_t ret;
 162         int prctl_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         prctl_ret = prctl(PR_GET_NAME, (unsigned long) reg_msg.name, 0, 0, 0);
 182         if (prctl_ret) {
 183                 ERR("Error executing prctl");
 184                 return -errno;
 185         }
 186
 187         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 188         if (ret >= 0 && ret != sizeof(reg_msg))
 189                 return -EIO;
 190         return ret;
 191 }
 192
 193 static
 194 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 195 {
 196         ssize_t len;
 197
 198         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 199         switch (len) {
 200         case sizeof(*lur):
 201                 DBG("message successfully sent");
 202                 return 0;
 203         case -1:
 204                 if (errno == ECONNRESET) {
 205                         printf("remote end closed connection\n");
 206                         return 0;
 207                 }
 208                 return -1;
 209         default:
 210                 printf("incorrect message size: %zd\n", len);
 211                 return -1;
 212         }
 213 }
 214
 215 static
 216 int handle_register_done(struct sock_info *sock_info)
 217 {
 218         int ret;
 219
 220         if (sock_info->constructor_sem_posted)
 221                 return 0;
 222         sock_info->constructor_sem_posted = 1;
 223         if (uatomic_read(&sem_count) <= 0) {
 224                 return 0;
 225         }
 226         ret = uatomic_add_return(&sem_count, -1);
 227         if (ret == 0) {
 228                 ret = sem_post(&constructor_wait);
 229                 assert(!ret);
 230         }
 231         return 0;
 232 }
 233
 234 static
 235 int handle_message(struct sock_info *sock_info,
 236                 int sock, struct ustcomm_ust_msg *lum)
 237 {
 238         int ret = 0;
 239         const struct lttng_ust_objd_ops *ops;
 240         struct ustcomm_ust_reply lur;
 241         int shm_fd, wait_fd;
 242
 243         ust_lock();
 244
 245         memset(&lur, 0, sizeof(lur));
 246
 247         if (lttng_ust_comm_should_quit) {
 248                 ret = -EPERM;
 249                 goto end;
 250         }
 251
 252         ops = objd_ops(lum->handle);
 253         if (!ops) {
 254                 ret = -ENOENT;
 255                 goto end;
 256         }
 257
 258         switch (lum->cmd) {
 259         case LTTNG_UST_REGISTER_DONE:
 260                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 261                         ret = handle_register_done(sock_info);
 262                 else
 263                         ret = -EINVAL;
 264                 break;
 265         case LTTNG_UST_RELEASE:
 266                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 267                         ret = -EPERM;
 268                 else
 269                         ret = lttng_ust_objd_unref(lum->handle);
 270                 break;
 271         default:
 272                 if (ops->cmd)
 273                         ret = ops->cmd(lum->handle, lum->cmd,
 274                                         (unsigned long) &lum->u);
 275                 else
 276                         ret = -ENOSYS;
 277                 break;
 278         }
 279
 280 end:
 281         lur.handle = lum->handle;
 282         lur.cmd = lum->cmd;
 283         lur.ret_val = ret;
 284         if (ret >= 0) {
 285                 lur.ret_code = USTCOMM_OK;
 286         } else {
 287                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 288                 lur.ret_code = ret;
 289         }
 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 = lum->u.stream.memory_map_size;
 297                 shm_fd = lum->u.stream.shm_fd;
 298                 wait_fd = lum->u.stream.wait_fd;
 299                 break;
 300         case LTTNG_UST_METADATA:
 301         case LTTNG_UST_CHANNEL:
 302                 lur.u.channel.memory_map_size = lum->u.channel.memory_map_size;
 303                 shm_fd = lum->u.channel.shm_fd;
 304                 wait_fd = lum->u.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_list_entry,
 311                         lum->u.tracepoint_list_entry,
 312                         LTTNG_UST_SYM_NAME_LEN);
 313                 break;
 314         }
 315         ret = send_reply(sock, &lur);
 316         if (ret < 0) {
 317                 perror("error sending reply");
 318                 goto error;
 319         }
 320
 321         if ((lum->cmd == LTTNG_UST_STREAM
 322              || lum->cmd == LTTNG_UST_CHANNEL
 323              || lum->cmd == LTTNG_UST_METADATA)
 324                         && lur.ret_code == USTCOMM_OK) {
 325                 /* we also need to send the file descriptors. */
 326                 ret = ustcomm_send_fds_unix_sock(sock,
 327                         &shm_fd, &shm_fd,
 328                         1, sizeof(int));
 329                 if (ret < 0) {
 330                         perror("send shm_fd");
 331                         goto error;
 332                 }
 333                 ret = ustcomm_send_fds_unix_sock(sock,
 334                         &wait_fd, &wait_fd,
 335                         1, sizeof(int));
 336                 if (ret < 0) {
 337                         perror("send wait_fd");
 338                         goto error;
 339                 }
 340         }
 341 error:
 342         ust_unlock();
 343         return ret;
 344 }
 345
 346 static
 347 void cleanup_sock_info(struct sock_info *sock_info)
 348 {
 349         int ret;
 350
 351         if (sock_info->socket != -1) {
 352                 ret = close(sock_info->socket);
 353                 if (ret) {
 354                         ERR("Error closing apps socket");
 355                 }
 356                 sock_info->socket = -1;
 357         }
 358         if (sock_info->root_handle != -1) {
 359                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 360                 if (ret) {
 361                         ERR("Error unref root handle");
 362                 }
 363                 sock_info->root_handle = -1;
 364         }
 365         sock_info->constructor_sem_posted = 0;
 366         if (sock_info->wait_shm_mmap) {
 367                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 368                 if (ret) {
 369                         ERR("Error unmapping wait shm");
 370                 }
 371                 sock_info->wait_shm_mmap = NULL;
 372         }
 373 }
 374
 375 /*
 376  * Using fork to set umask in the child process (not multi-thread safe).
 377  * We deal with the shm_open vs ftruncate race (happening when the
 378  * sessiond owns the shm and does not let everybody modify it, to ensure
 379  * safety against shm_unlink) by simply letting the mmap fail and
 380  * retrying after a few seconds.
 381  * For global shm, everybody has rw access to it until the sessiond
 382  * starts.
 383  */
 384 static
 385 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 386 {
 387         int wait_shm_fd, ret;
 388         pid_t pid;
 389
 390         /*
 391          * Try to open read-only.
 392          */
 393         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 394         if (wait_shm_fd >= 0) {
 395                 goto end;
 396         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 397                 /*
 398                  * Real-only open did not work, and it's not because the
 399                  * entry was not present. It's a failure that prohibits
 400                  * using shm.
 401                  */
 402                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 403                 goto end;
 404         }
 405         /*
 406          * If the open failed because the file did not exist, try
 407          * creating it ourself.
 408          */
 409         pid = fork();
 410         if (pid > 0) {
 411                 int status;
 412
 413                 /*
 414                  * Parent: wait for child to return, in which case the
 415                  * shared memory map will have been created.
 416                  */
 417                 pid = wait(&status);
 418                 if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 419                         wait_shm_fd = -1;
 420                         goto end;
 421                 }
 422                 /*
 423                  * Try to open read-only again after creation.
 424                  */
 425                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 426                 if (wait_shm_fd < 0) {
 427                         /*
 428                          * Real-only open did not work. It's a failure
 429                          * that prohibits using shm.
 430                          */
 431                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 432                         goto end;
 433                 }
 434                 goto end;
 435         } else if (pid == 0) {
 436                 int create_mode;
 437
 438                 /* Child */
 439                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 440                 if (sock_info->global)
 441                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 442                 /*
 443                  * We're alone in a child process, so we can modify the
 444                  * process-wide umask.
 445                  */
 446                 umask(~create_mode);
 447                 /*
 448                  * Try creating shm (or get rw access).
 449                  * We don't do an exclusive open, because we allow other
 450                  * processes to create+ftruncate it concurrently.
 451                  */
 452                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 453                                 O_RDWR | O_CREAT, create_mode);
 454                 if (wait_shm_fd >= 0) {
 455                         ret = ftruncate(wait_shm_fd, mmap_size);
 456                         if (ret) {
 457                                 PERROR("ftruncate");
 458                                 exit(EXIT_FAILURE);
 459                         }
 460                         exit(EXIT_SUCCESS);
 461                 }
 462                 /*
 463                  * For local shm, we need to have rw access to accept
 464                  * opening it: this means the local sessiond will be
 465                  * able to wake us up. For global shm, we open it even
 466                  * if rw access is not granted, because the root.root
 467                  * sessiond will be able to override all rights and wake
 468                  * us up.
 469                  */
 470                 if (!sock_info->global && errno != EACCES) {
 471                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 472                         exit(EXIT_FAILURE);
 473                 }
 474                 /*
 475                  * The shm exists, but we cannot open it RW. Report
 476                  * success.
 477                  */
 478                 exit(EXIT_SUCCESS);
 479         } else {
 480                 return -1;
 481         }
 482 end:
 483         if (wait_shm_fd >= 0 && !sock_info->global) {
 484                 struct stat statbuf;
 485
 486                 /*
 487                  * Ensure that our user is the owner of the shm file for
 488                  * local shm. If we do not own the file, it means our
 489                  * sessiond will not have access to wake us up (there is
 490                  * probably a rogue process trying to fake our
 491                  * sessiond). Fallback to polling method in this case.
 492                  */
 493                 ret = fstat(wait_shm_fd, &statbuf);
 494                 if (ret) {
 495                         PERROR("fstat");
 496                         goto error_close;
 497                 }
 498                 if (statbuf.st_uid != getuid())
 499                         goto error_close;
 500         }
 501         return wait_shm_fd;
 502
 503 error_close:
 504         ret = close(wait_shm_fd);
 505         if (ret) {
 506                 PERROR("Error closing fd");
 507         }
 508         return -1;
 509 }
 510
 511 static
 512 char *get_map_shm(struct sock_info *sock_info)
 513 {
 514         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 515         int wait_shm_fd, ret;
 516         char *wait_shm_mmap;
 517
 518         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 519         if (wait_shm_fd < 0) {
 520                 goto error;
 521         }
 522         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 523                   MAP_SHARED, wait_shm_fd, 0);
 524         /* close shm fd immediately after taking the mmap reference */
 525         ret = close(wait_shm_fd);
 526         if (ret) {
 527                 PERROR("Error closing fd");
 528         }
 529         if (wait_shm_mmap == MAP_FAILED) {
 530                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 531                 goto error;
 532         }
 533         return wait_shm_mmap;
 534
 535 error:
 536         return NULL;
 537 }
 538
 539 static
 540 void wait_for_sessiond(struct sock_info *sock_info)
 541 {
 542         int ret;
 543
 544         ust_lock();
 545         if (lttng_ust_comm_should_quit) {
 546                 goto quit;
 547         }
 548         if (wait_poll_fallback) {
 549                 goto error;
 550         }
 551         if (!sock_info->wait_shm_mmap) {
 552                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 553                 if (!sock_info->wait_shm_mmap)
 554                         goto error;
 555         }
 556         ust_unlock();
 557
 558         DBG("Waiting for %s apps sessiond", sock_info->name);
 559         /* Wait for futex wakeup */
 560         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 561                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 562                         FUTEX_WAIT, 0, NULL, NULL, 0);
 563                 if (ret < 0) {
 564                         if (errno == EFAULT) {
 565                                 wait_poll_fallback = 1;
 566                                 WARN(
 567 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 568 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 569 "Please upgrade your kernel "
 570 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 571 "mainline). LTTng-UST will use polling mode fallback.");
 572                         }
 573                         PERROR("futex");
 574                 }
 575         }
 576         return;
 577
 578 quit:
 579         ust_unlock();
 580         return;
 581
 582 error:
 583         ust_unlock();
 584         return;
 585 }
 586
 587 /*
 588  * This thread does not allocate any resource, except within
 589  * handle_message, within mutex protection. This mutex protects against
 590  * fork and exit.
 591  * The other moment it allocates resources is at socket connexion, which
 592  * is also protected by the mutex.
 593  */
 594 static
 595 void *ust_listener_thread(void *arg)
 596 {
 597         struct sock_info *sock_info = arg;
 598         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 599
 600         /* Restart trying to connect to the session daemon */
 601 restart:
 602         if (prev_connect_failed) {
 603                 /* Wait for sessiond availability with pipe */
 604                 wait_for_sessiond(sock_info);
 605                 if (has_waited) {
 606                         has_waited = 0;
 607                         /*
 608                          * Sleep for 5 seconds before retrying after a
 609                          * sequence of failure / wait / failure. This
 610                          * deals with a killed or broken session daemon.
 611                          */
 612                         sleep(5);
 613                 }
 614                 has_waited = 1;
 615                 prev_connect_failed = 0;
 616         }
 617         ust_lock();
 618
 619         if (lttng_ust_comm_should_quit) {
 620                 ust_unlock();
 621                 goto quit;
 622         }
 623
 624         if (sock_info->socket != -1) {
 625                 ret = close(sock_info->socket);
 626                 if (ret) {
 627                         ERR("Error closing %s apps socket", sock_info->name);
 628                 }
 629                 sock_info->socket = -1;
 630         }
 631
 632         /* Register */
 633         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 634         if (ret < 0) {
 635                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 636                 prev_connect_failed = 1;
 637                 /*
 638                  * If we cannot find the sessiond daemon, don't delay
 639                  * constructor execution.
 640                  */
 641                 ret = handle_register_done(sock_info);
 642                 assert(!ret);
 643                 ust_unlock();
 644                 goto restart;
 645         }
 646
 647         sock_info->socket = sock = ret;
 648
 649         /*
 650          * Create only one root handle per listener thread for the whole
 651          * process lifetime.
 652          */
 653         if (sock_info->root_handle == -1) {
 654                 ret = lttng_abi_create_root_handle();
 655                 if (ret < 0) {
 656                         ERR("Error creating root handle");
 657                         ust_unlock();
 658                         goto quit;
 659                 }
 660                 sock_info->root_handle = ret;
 661         }
 662
 663         ret = register_app_to_sessiond(sock);
 664         if (ret < 0) {
 665                 ERR("Error registering to %s apps socket", sock_info->name);
 666                 prev_connect_failed = 1;
 667                 /*
 668                  * If we cannot register to the sessiond daemon, don't
 669                  * delay constructor execution.
 670                  */
 671                 ret = handle_register_done(sock_info);
 672                 assert(!ret);
 673                 ust_unlock();
 674                 goto restart;
 675         }
 676         ust_unlock();
 677
 678         for (;;) {
 679                 ssize_t len;
 680                 struct ustcomm_ust_msg lum;
 681
 682                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 683                 switch (len) {
 684                 case 0: /* orderly shutdown */
 685                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 686                         ust_lock();
 687                         /*
 688                          * Either sessiond has shutdown or refused us by closing the socket.
 689                          * In either case, we don't want to delay construction execution,
 690                          * and we need to wait before retry.
 691                          */
 692                         prev_connect_failed = 1;
 693                         /*
 694                          * If we cannot register to the sessiond daemon, don't
 695                          * delay constructor execution.
 696                          */
 697                         ret = handle_register_done(sock_info);
 698                         assert(!ret);
 699                         ust_unlock();
 700                         goto end;
 701                 case sizeof(lum):
 702                         DBG("message received\n");
 703                         ret = handle_message(sock_info, sock, &lum);
 704                         if (ret < 0) {
 705                                 ERR("Error handling message for %s socket", sock_info->name);
 706                         }
 707                         continue;
 708                 case -1:
 709                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 710                         if (errno == ECONNRESET) {
 711                                 ERR("%s remote end closed connection\n", sock_info->name);
 712                                 goto end;
 713                         }
 714                         goto end;
 715                 default:
 716                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 717                         continue;
 718                 }
 719
 720         }
 721 end:
 722         goto restart;   /* try to reconnect */
 723 quit:
 724         return NULL;
 725 }
 726
 727 /*
 728  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 729  */
 730 static
 731 int get_timeout(struct timespec *constructor_timeout)
 732 {
 733         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 734         char *str_delay;
 735         int ret;
 736
 737         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 738         if (str_delay) {
 739                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 740         }
 741
 742         switch (constructor_delay_ms) {
 743         case -1:/* fall-through */
 744         case 0:
 745                 return constructor_delay_ms;
 746         default:
 747                 break;
 748         }
 749
 750         /*
 751          * If we are unable to find the current time, don't wait.
 752          */
 753         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 754         if (ret) {
 755                 return -1;
 756         }
 757         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 758         constructor_timeout->tv_nsec +=
 759                 (constructor_delay_ms % 1000UL) * 1000000UL;
 760         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 761                 constructor_timeout->tv_sec++;
 762                 constructor_timeout->tv_nsec -= 1000000000UL;
 763         }
 764         return 1;
 765 }
 766
 767 /*
 768  * sessiond monitoring thread: monitor presence of global and per-user
 769  * sessiond by polling the application common named pipe.
 770  */
 771 /* TODO */
 772
 773 void __attribute__((constructor)) lttng_ust_init(void)
 774 {
 775         struct timespec constructor_timeout;
 776         int timeout_mode;
 777         int ret;
 778
 779         if (uatomic_xchg(&initialized, 1) == 1)
 780                 return;
 781
 782         /*
 783          * We want precise control over the order in which we construct
 784          * our sub-libraries vs starting to receive commands from
 785          * sessiond (otherwise leading to errors when trying to create
 786          * sessiond before the init functions are completed).
 787          */
 788         init_usterr();
 789         init_tracepoint();
 790         ltt_ring_buffer_metadata_client_init();
 791         ltt_ring_buffer_client_overwrite_init();
 792         ltt_ring_buffer_client_discard_init();
 793
 794         timeout_mode = get_timeout(&constructor_timeout);
 795
 796         ret = sem_init(&constructor_wait, 0, 0);
 797         assert(!ret);
 798
 799         ret = setup_local_apps();
 800         if (ret) {
 801                 ERR("Error setting up to local apps");
 802         }
 803         ret = pthread_create(&local_apps.ust_listener, NULL,
 804                         ust_listener_thread, &local_apps);
 805
 806         if (local_apps.allowed) {
 807                 ret = pthread_create(&global_apps.ust_listener, NULL,
 808                                 ust_listener_thread, &global_apps);
 809         } else {
 810                 handle_register_done(&local_apps);
 811         }
 812
 813         switch (timeout_mode) {
 814         case 1: /* timeout wait */
 815                 do {
 816                         ret = sem_timedwait(&constructor_wait,
 817                                         &constructor_timeout);
 818                 } while (ret < 0 && errno == EINTR);
 819                 if (ret < 0 && errno == ETIMEDOUT) {
 820                         ERR("Timed out waiting for ltt-sessiond");
 821                 } else {
 822                         assert(!ret);
 823                 }
 824                 break;
 825         case -1:/* wait forever */
 826                 do {
 827                         ret = sem_wait(&constructor_wait);
 828                 } while (ret < 0 && errno == EINTR);
 829                 assert(!ret);
 830                 break;
 831         case 0: /* no timeout */
 832                 break;
 833         }
 834 }
 835
 836 static
 837 void lttng_ust_cleanup(int exiting)
 838 {
 839         cleanup_sock_info(&global_apps);
 840         if (local_apps.allowed) {
 841                 cleanup_sock_info(&local_apps);
 842         }
 843         lttng_ust_abi_exit();
 844         lttng_ust_events_exit();
 845         ltt_ring_buffer_client_discard_exit();
 846         ltt_ring_buffer_client_overwrite_exit();
 847         ltt_ring_buffer_metadata_client_exit();
 848         exit_tracepoint();
 849         if (!exiting) {
 850                 /* Reinitialize values for fork */
 851                 sem_count = 2;
 852                 lttng_ust_comm_should_quit = 0;
 853                 initialized = 0;
 854         }
 855 }
 856
 857 void __attribute__((destructor)) lttng_ust_exit(void)
 858 {
 859         int ret;
 860
 861         /*
 862          * Using pthread_cancel here because:
 863          * A) we don't want to hang application teardown.
 864          * B) the thread is not allocating any resource.
 865          */
 866
 867         /*
 868          * Require the communication thread to quit. Synchronize with
 869          * mutexes to ensure it is not in a mutex critical section when
 870          * pthread_cancel is later called.
 871          */
 872         ust_lock();
 873         lttng_ust_comm_should_quit = 1;
 874         ust_unlock();
 875
 876         ret = pthread_cancel(global_apps.ust_listener);
 877         if (ret) {
 878                 ERR("Error cancelling global ust listener thread");
 879         }
 880         if (local_apps.allowed) {
 881                 ret = pthread_cancel(local_apps.ust_listener);
 882                 if (ret) {
 883                         ERR("Error cancelling local ust listener thread");
 884                 }
 885         }
 886         lttng_ust_cleanup(1);
 887 }
 888
 889 /*
 890  * We exclude the worker threads across fork and clone (except
 891  * CLONE_VM), because these system calls only keep the forking thread
 892  * running in the child.  Therefore, we don't want to call fork or clone
 893  * in the middle of an tracepoint or ust tracing state modification.
 894  * Holding this mutex protects these structures across fork and clone.
 895  */
 896 void ust_before_fork(sigset_t *save_sigset)
 897 {
 898         /*
 899          * Disable signals. This is to avoid that the child intervenes
 900          * before it is properly setup for tracing. It is safer to
 901          * disable all signals, because then we know we are not breaking
 902          * anything by restoring the original mask.
 903          */
 904         sigset_t all_sigs;
 905         int ret;
 906
 907         /* Disable signals */
 908         sigfillset(&all_sigs);
 909         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
 910         if (ret == -1) {
 911                 PERROR("sigprocmask");
 912         }
 913         ust_lock();
 914         rcu_bp_before_fork();
 915 }
 916
 917 static void ust_after_fork_common(sigset_t *restore_sigset)
 918 {
 919         int ret;
 920
 921         DBG("process %d", getpid());
 922         ust_unlock();
 923         /* Restore signals */
 924         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
 925         if (ret == -1) {
 926                 PERROR("sigprocmask");
 927         }
 928 }
 929
 930 void ust_after_fork_parent(sigset_t *restore_sigset)
 931 {
 932         DBG("process %d", getpid());
 933         rcu_bp_after_fork_parent();
 934         /* Release mutexes and reenable signals */
 935         ust_after_fork_common(restore_sigset);
 936 }
 937
 938 /*
 939  * After fork, in the child, we need to cleanup all the leftover state,
 940  * except the worker thread which already magically disappeared thanks
 941  * to the weird Linux fork semantics. After tyding up, we call
 942  * lttng_ust_init() again to start over as a new PID.
 943  *
 944  * This is meant for forks() that have tracing in the child between the
 945  * fork and following exec call (if there is any).
 946  */
 947 void ust_after_fork_child(sigset_t *restore_sigset)
 948 {
 949         DBG("process %d", getpid());
 950         /* Release urcu mutexes */
 951         rcu_bp_after_fork_child();
 952         lttng_ust_cleanup(0);
 953         lttng_context_vtid_reset();
 954         /* Release mutexes and reenable signals */
 955         ust_after_fork_common(restore_sigset);
 956         lttng_ust_init();
 957 }