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