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