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/mman.h>
  26 #include <sys/stat.h>
  27 #include <sys/types.h>
  28 #include <sys/wait.h>
  29 #include <fcntl.h>
  30 #include <unistd.h>
  31 #include <errno.h>
  32 #include <pthread.h>
  33 #include <semaphore.h>
  34 #include <time.h>
  35 #include <assert.h>
  36 #include <signal.h>
  37 #include <urcu/uatomic.h>
  38 #include <urcu/futex.h>
  39 #include <urcu/compiler.h>
  40
  41 #include <lttng/ust-events.h>
  42 #include <lttng/ust-abi.h>
  43 #include <lttng/ust.h>
  44 #include <ust-comm.h>
  45 #include <usterr-signal-safe.h>
  46 #include "tracepoint-internal.h"
  47 #include "ltt-tracer-core.h"
  48 #include "compat.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         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                 uint32_t bits_per_long;
 170                 char name[16];  /* process name */
 171         } reg_msg;
 172
 173         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 174         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 175         reg_msg.pid = getpid();
 176         reg_msg.ppid = getppid();
 177         reg_msg.uid = getuid();
 178         reg_msg.gid = getgid();
 179         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 180         lttng_ust_getprocname(reg_msg.name);
 181
 182         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 183         if (ret >= 0 && ret != sizeof(reg_msg))
 184                 return -EIO;
 185         return ret;
 186 }
 187
 188 static
 189 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 190 {
 191         ssize_t len;
 192
 193         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 194         switch (len) {
 195         case sizeof(*lur):
 196                 DBG("message successfully sent");
 197                 return 0;
 198         case -1:
 199                 if (errno == ECONNRESET) {
 200                         printf("remote end closed connection\n");
 201                         return 0;
 202                 }
 203                 return -1;
 204         default:
 205                 printf("incorrect message size: %zd\n", len);
 206                 return -1;
 207         }
 208 }
 209
 210 static
 211 int handle_register_done(struct sock_info *sock_info)
 212 {
 213         int ret;
 214
 215         if (sock_info->constructor_sem_posted)
 216                 return 0;
 217         sock_info->constructor_sem_posted = 1;
 218         if (uatomic_read(&sem_count) <= 0) {
 219                 return 0;
 220         }
 221         ret = uatomic_add_return(&sem_count, -1);
 222         if (ret == 0) {
 223                 ret = sem_post(&constructor_wait);
 224                 assert(!ret);
 225         }
 226         return 0;
 227 }
 228
 229 static
 230 int handle_message(struct sock_info *sock_info,
 231                 int sock, struct ustcomm_ust_msg *lum)
 232 {
 233         int ret = 0;
 234         const struct lttng_ust_objd_ops *ops;
 235         struct ustcomm_ust_reply lur;
 236         int shm_fd, wait_fd;
 237         union ust_args args;
 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 = lttng_ust_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                                         &args);
 272                 else
 273                         ret = -ENOSYS;
 274                 break;
 275         }
 276
 277 end:
 278         lur.handle = lum->handle;
 279         lur.cmd = lum->cmd;
 280         lur.ret_val = ret;
 281         if (ret >= 0) {
 282                 lur.ret_code = USTCOMM_OK;
 283         } else {
 284                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 285                 lur.ret_code = ret;
 286         }
 287         switch (lum->cmd) {
 288         case LTTNG_UST_STREAM:
 289                 /*
 290                  * Special-case reply to send stream info.
 291                  * Use lum.u output.
 292                  */
 293                 lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 294                 shm_fd = *args.stream.shm_fd;
 295                 wait_fd = *args.stream.wait_fd;
 296                 break;
 297         case LTTNG_UST_METADATA:
 298         case LTTNG_UST_CHANNEL:
 299                 lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 300                 shm_fd = *args.channel.shm_fd;
 301                 wait_fd = *args.channel.wait_fd;
 302                 break;
 303         case LTTNG_UST_TRACER_VERSION:
 304                 lur.u.version = lum->u.version;
 305                 break;
 306         case LTTNG_UST_TRACEPOINT_LIST_GET:
 307                 memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 308                 break;
 309         }
 310         ret = send_reply(sock, &lur);
 311         if (ret < 0) {
 312                 perror("error sending reply");
 313                 goto error;
 314         }
 315
 316         if ((lum->cmd == LTTNG_UST_STREAM
 317              || lum->cmd == LTTNG_UST_CHANNEL
 318              || lum->cmd == LTTNG_UST_METADATA)
 319                         && lur.ret_code == USTCOMM_OK) {
 320                 /* we also need to send the file descriptors. */
 321                 ret = ustcomm_send_fds_unix_sock(sock,
 322                         &shm_fd, &shm_fd,
 323                         1, sizeof(int));
 324                 if (ret < 0) {
 325                         perror("send shm_fd");
 326                         goto error;
 327                 }
 328                 ret = ustcomm_send_fds_unix_sock(sock,
 329                         &wait_fd, &wait_fd,
 330                         1, sizeof(int));
 331                 if (ret < 0) {
 332                         perror("send wait_fd");
 333                         goto error;
 334                 }
 335         }
 336         /*
 337          * We still have the memory map reference, and the fds have been
 338          * sent to the sessiond. We can therefore close those fds. Note
 339          * that we keep the write side of the wait_fd open, but close
 340          * the read side.
 341          */
 342         if (lur.ret_code == USTCOMM_OK) {
 343                 switch (lum->cmd) {
 344                 case LTTNG_UST_STREAM:
 345                         if (shm_fd >= 0) {
 346                                 ret = close(shm_fd);
 347                                 if (ret) {
 348                                         PERROR("Error closing stream shm_fd");
 349                                 }
 350                                 *args.stream.shm_fd = -1;
 351                         }
 352                         if (wait_fd >= 0) {
 353                                 ret = close(wait_fd);
 354                                 if (ret) {
 355                                         PERROR("Error closing stream wait_fd");
 356                                 }
 357                                 *args.stream.wait_fd = -1;
 358                         }
 359                         break;
 360                 case LTTNG_UST_METADATA:
 361                 case LTTNG_UST_CHANNEL:
 362                         if (shm_fd >= 0) {
 363                                 ret = close(shm_fd);
 364                                 if (ret) {
 365                                         PERROR("Error closing channel shm_fd");
 366                                 }
 367                                 *args.channel.shm_fd = -1;
 368                         }
 369                         if (wait_fd >= 0) {
 370                                 ret = close(wait_fd);
 371                                 if (ret) {
 372                                         PERROR("Error closing channel wait_fd");
 373                                 }
 374                                 *args.channel.wait_fd = -1;
 375                         }
 376                         break;
 377                 }
 378         }
 379
 380 error:
 381         ust_unlock();
 382         return ret;
 383 }
 384
 385 static
 386 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 387 {
 388         int ret;
 389
 390         if (sock_info->socket != -1) {
 391                 ret = close(sock_info->socket);
 392                 if (ret) {
 393                         ERR("Error closing apps socket");
 394                 }
 395                 sock_info->socket = -1;
 396         }
 397         if (sock_info->root_handle != -1) {
 398                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 399                 if (ret) {
 400                         ERR("Error unref root handle");
 401                 }
 402                 sock_info->root_handle = -1;
 403         }
 404         sock_info->constructor_sem_posted = 0;
 405         /*
 406          * When called from process exit, we allow this memory map to be
 407          * released by the OS at exit(), because removing it prior to
 408          * this can cause a segmentation fault when using the
 409          * futex_async timer-based fallback. And we cannot join those
 410          * threads because sys_futex does not react to the cancellation
 411          * request.
 412          *
 413          * So we actually _do_ release it only after a fork, since all
 414          * threads have vanished anyway.
 415          */
 416         if (!exiting && sock_info->wait_shm_mmap) {
 417                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 418                 if (ret) {
 419                         ERR("Error unmapping wait shm");
 420                 }
 421                 sock_info->wait_shm_mmap = NULL;
 422         }
 423 }
 424
 425 /*
 426  * Using fork to set umask in the child process (not multi-thread safe).
 427  * We deal with the shm_open vs ftruncate race (happening when the
 428  * sessiond owns the shm and does not let everybody modify it, to ensure
 429  * safety against shm_unlink) by simply letting the mmap fail and
 430  * retrying after a few seconds.
 431  * For global shm, everybody has rw access to it until the sessiond
 432  * starts.
 433  */
 434 static
 435 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 436 {
 437         int wait_shm_fd, ret;
 438         pid_t pid;
 439
 440         /*
 441          * Try to open read-only.
 442          */
 443         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 444         if (wait_shm_fd >= 0) {
 445                 goto end;
 446         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 447                 /*
 448                  * Real-only open did not work, and it's not because the
 449                  * entry was not present. It's a failure that prohibits
 450                  * using shm.
 451                  */
 452                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 453                 goto end;
 454         }
 455         /*
 456          * If the open failed because the file did not exist, try
 457          * creating it ourself.
 458          */
 459         pid = fork();
 460         if (pid > 0) {
 461                 int status;
 462
 463                 /*
 464                  * Parent: wait for child to return, in which case the
 465                  * shared memory map will have been created.
 466                  */
 467                 pid = wait(&status);
 468                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 469                         wait_shm_fd = -1;
 470                         goto end;
 471                 }
 472                 /*
 473                  * Try to open read-only again after creation.
 474                  */
 475                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 476                 if (wait_shm_fd < 0) {
 477                         /*
 478                          * Real-only open did not work. It's a failure
 479                          * that prohibits using shm.
 480                          */
 481                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 482                         goto end;
 483                 }
 484                 goto end;
 485         } else if (pid == 0) {
 486                 int create_mode;
 487
 488                 /* Child */
 489                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 490                 if (sock_info->global)
 491                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 492                 /*
 493                  * We're alone in a child process, so we can modify the
 494                  * process-wide umask.
 495                  */
 496                 umask(~create_mode);
 497                 /*
 498                  * Try creating shm (or get rw access).
 499                  * We don't do an exclusive open, because we allow other
 500                  * processes to create+ftruncate it concurrently.
 501                  */
 502                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 503                                 O_RDWR | O_CREAT, create_mode);
 504                 if (wait_shm_fd >= 0) {
 505                         ret = ftruncate(wait_shm_fd, mmap_size);
 506                         if (ret) {
 507                                 PERROR("ftruncate");
 508                                 exit(EXIT_FAILURE);
 509                         }
 510                         exit(EXIT_SUCCESS);
 511                 }
 512                 /*
 513                  * For local shm, we need to have rw access to accept
 514                  * opening it: this means the local sessiond will be
 515                  * able to wake us up. For global shm, we open it even
 516                  * if rw access is not granted, because the root.root
 517                  * sessiond will be able to override all rights and wake
 518                  * us up.
 519                  */
 520                 if (!sock_info->global && errno != EACCES) {
 521                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 522                         exit(EXIT_FAILURE);
 523                 }
 524                 /*
 525                  * The shm exists, but we cannot open it RW. Report
 526                  * success.
 527                  */
 528                 exit(EXIT_SUCCESS);
 529         } else {
 530                 return -1;
 531         }
 532 end:
 533         if (wait_shm_fd >= 0 && !sock_info->global) {
 534                 struct stat statbuf;
 535
 536                 /*
 537                  * Ensure that our user is the owner of the shm file for
 538                  * local shm. If we do not own the file, it means our
 539                  * sessiond will not have access to wake us up (there is
 540                  * probably a rogue process trying to fake our
 541                  * sessiond). Fallback to polling method in this case.
 542                  */
 543                 ret = fstat(wait_shm_fd, &statbuf);
 544                 if (ret) {
 545                         PERROR("fstat");
 546                         goto error_close;
 547                 }
 548                 if (statbuf.st_uid != getuid())
 549                         goto error_close;
 550         }
 551         return wait_shm_fd;
 552
 553 error_close:
 554         ret = close(wait_shm_fd);
 555         if (ret) {
 556                 PERROR("Error closing fd");
 557         }
 558         return -1;
 559 }
 560
 561 static
 562 char *get_map_shm(struct sock_info *sock_info)
 563 {
 564         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 565         int wait_shm_fd, ret;
 566         char *wait_shm_mmap;
 567
 568         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 569         if (wait_shm_fd < 0) {
 570                 goto error;
 571         }
 572         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 573                   MAP_SHARED, wait_shm_fd, 0);
 574         /* close shm fd immediately after taking the mmap reference */
 575         ret = close(wait_shm_fd);
 576         if (ret) {
 577                 PERROR("Error closing fd");
 578         }
 579         if (wait_shm_mmap == MAP_FAILED) {
 580                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 581                 goto error;
 582         }
 583         return wait_shm_mmap;
 584
 585 error:
 586         return NULL;
 587 }
 588
 589 static
 590 void wait_for_sessiond(struct sock_info *sock_info)
 591 {
 592         int ret;
 593
 594         ust_lock();
 595         if (lttng_ust_comm_should_quit) {
 596                 goto quit;
 597         }
 598         if (wait_poll_fallback) {
 599                 goto error;
 600         }
 601         if (!sock_info->wait_shm_mmap) {
 602                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 603                 if (!sock_info->wait_shm_mmap)
 604                         goto error;
 605         }
 606         ust_unlock();
 607
 608         DBG("Waiting for %s apps sessiond", sock_info->name);
 609         /* Wait for futex wakeup */
 610         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 611                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 612                         FUTEX_WAIT, 0, NULL, NULL, 0);
 613                 if (ret < 0) {
 614                         if (errno == EFAULT) {
 615                                 wait_poll_fallback = 1;
 616                                 DBG(
 617 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 618 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 619 "Please upgrade your kernel "
 620 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 621 "mainline). LTTng-UST will use polling mode fallback.");
 622                                 if (ust_debug())
 623                                         PERROR("futex");
 624                         }
 625                 }
 626         }
 627         return;
 628
 629 quit:
 630         ust_unlock();
 631         return;
 632
 633 error:
 634         ust_unlock();
 635         return;
 636 }
 637
 638 /*
 639  * This thread does not allocate any resource, except within
 640  * handle_message, within mutex protection. This mutex protects against
 641  * fork and exit.
 642  * The other moment it allocates resources is at socket connexion, which
 643  * is also protected by the mutex.
 644  */
 645 static
 646 void *ust_listener_thread(void *arg)
 647 {
 648         struct sock_info *sock_info = arg;
 649         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 650
 651         /* Restart trying to connect to the session daemon */
 652 restart:
 653         if (prev_connect_failed) {
 654                 /* Wait for sessiond availability with pipe */
 655                 wait_for_sessiond(sock_info);
 656                 if (has_waited) {
 657                         has_waited = 0;
 658                         /*
 659                          * Sleep for 5 seconds before retrying after a
 660                          * sequence of failure / wait / failure. This
 661                          * deals with a killed or broken session daemon.
 662                          */
 663                         sleep(5);
 664                 }
 665                 has_waited = 1;
 666                 prev_connect_failed = 0;
 667         }
 668         ust_lock();
 669
 670         if (lttng_ust_comm_should_quit) {
 671                 ust_unlock();
 672                 goto quit;
 673         }
 674
 675         if (sock_info->socket != -1) {
 676                 ret = close(sock_info->socket);
 677                 if (ret) {
 678                         ERR("Error closing %s apps socket", sock_info->name);
 679                 }
 680                 sock_info->socket = -1;
 681         }
 682
 683         /* Register */
 684         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 685         if (ret < 0) {
 686                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 687                 prev_connect_failed = 1;
 688                 /*
 689                  * If we cannot find the sessiond daemon, don't delay
 690                  * constructor execution.
 691                  */
 692                 ret = handle_register_done(sock_info);
 693                 assert(!ret);
 694                 ust_unlock();
 695                 goto restart;
 696         }
 697
 698         sock_info->socket = sock = ret;
 699
 700         /*
 701          * Create only one root handle per listener thread for the whole
 702          * process lifetime.
 703          */
 704         if (sock_info->root_handle == -1) {
 705                 ret = lttng_abi_create_root_handle();
 706                 if (ret < 0) {
 707                         ERR("Error creating root handle");
 708                         ust_unlock();
 709                         goto quit;
 710                 }
 711                 sock_info->root_handle = ret;
 712         }
 713
 714         ret = register_app_to_sessiond(sock);
 715         if (ret < 0) {
 716                 ERR("Error registering to %s apps socket", sock_info->name);
 717                 prev_connect_failed = 1;
 718                 /*
 719                  * If we cannot register to the sessiond daemon, don't
 720                  * delay constructor execution.
 721                  */
 722                 ret = handle_register_done(sock_info);
 723                 assert(!ret);
 724                 ust_unlock();
 725                 goto restart;
 726         }
 727         ust_unlock();
 728
 729         for (;;) {
 730                 ssize_t len;
 731                 struct ustcomm_ust_msg lum;
 732
 733                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 734                 switch (len) {
 735                 case 0: /* orderly shutdown */
 736                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 737                         ust_lock();
 738                         /*
 739                          * Either sessiond has shutdown or refused us by closing the socket.
 740                          * In either case, we don't want to delay construction execution,
 741                          * and we need to wait before retry.
 742                          */
 743                         prev_connect_failed = 1;
 744                         /*
 745                          * If we cannot register to the sessiond daemon, don't
 746                          * delay constructor execution.
 747                          */
 748                         ret = handle_register_done(sock_info);
 749                         assert(!ret);
 750                         ust_unlock();
 751                         goto end;
 752                 case sizeof(lum):
 753                         DBG("message received\n");
 754                         ret = handle_message(sock_info, sock, &lum);
 755                         if (ret < 0) {
 756                                 ERR("Error handling message for %s socket", sock_info->name);
 757                         }
 758                         continue;
 759                 case -1:
 760                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 761                         if (errno == ECONNRESET) {
 762                                 ERR("%s remote end closed connection\n", sock_info->name);
 763                                 goto end;
 764                         }
 765                         goto end;
 766                 default:
 767                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 768                         continue;
 769                 }
 770
 771         }
 772 end:
 773         goto restart;   /* try to reconnect */
 774 quit:
 775         return NULL;
 776 }
 777
 778 /*
 779  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 780  */
 781 static
 782 int get_timeout(struct timespec *constructor_timeout)
 783 {
 784         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 785         char *str_delay;
 786         int ret;
 787
 788         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 789         if (str_delay) {
 790                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 791         }
 792
 793         switch (constructor_delay_ms) {
 794         case -1:/* fall-through */
 795         case 0:
 796                 return constructor_delay_ms;
 797         default:
 798                 break;
 799         }
 800
 801         /*
 802          * If we are unable to find the current time, don't wait.
 803          */
 804         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 805         if (ret) {
 806                 return -1;
 807         }
 808         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 809         constructor_timeout->tv_nsec +=
 810                 (constructor_delay_ms % 1000UL) * 1000000UL;
 811         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 812                 constructor_timeout->tv_sec++;
 813                 constructor_timeout->tv_nsec -= 1000000000UL;
 814         }
 815         return 1;
 816 }
 817
 818 /*
 819  * sessiond monitoring thread: monitor presence of global and per-user
 820  * sessiond by polling the application common named pipe.
 821  */
 822 /* TODO */
 823
 824 void __attribute__((constructor)) lttng_ust_init(void)
 825 {
 826         struct timespec constructor_timeout;
 827         int timeout_mode;
 828         int ret;
 829
 830         if (uatomic_xchg(&initialized, 1) == 1)
 831                 return;
 832
 833         /*
 834          * We want precise control over the order in which we construct
 835          * our sub-libraries vs starting to receive commands from
 836          * sessiond (otherwise leading to errors when trying to create
 837          * sessiond before the init functions are completed).
 838          */
 839         init_usterr();
 840         init_tracepoint();
 841         ltt_ring_buffer_metadata_client_init();
 842         ltt_ring_buffer_client_overwrite_init();
 843         ltt_ring_buffer_client_discard_init();
 844
 845         timeout_mode = get_timeout(&constructor_timeout);
 846
 847         ret = sem_init(&constructor_wait, 0, 0);
 848         assert(!ret);
 849
 850         ret = setup_local_apps();
 851         if (ret) {
 852                 ERR("Error setting up to local apps");
 853         }
 854         ret = pthread_create(&local_apps.ust_listener, NULL,
 855                         ust_listener_thread, &local_apps);
 856
 857         if (local_apps.allowed) {
 858                 ret = pthread_create(&global_apps.ust_listener, NULL,
 859                                 ust_listener_thread, &global_apps);
 860         } else {
 861                 handle_register_done(&local_apps);
 862         }
 863
 864         switch (timeout_mode) {
 865         case 1: /* timeout wait */
 866                 do {
 867                         ret = sem_timedwait(&constructor_wait,
 868                                         &constructor_timeout);
 869                 } while (ret < 0 && errno == EINTR);
 870                 if (ret < 0 && errno == ETIMEDOUT) {
 871                         ERR("Timed out waiting for ltt-sessiond");
 872                 } else {
 873                         assert(!ret);
 874                 }
 875                 break;
 876         case -1:/* wait forever */
 877                 do {
 878                         ret = sem_wait(&constructor_wait);
 879                 } while (ret < 0 && errno == EINTR);
 880                 assert(!ret);
 881                 break;
 882         case 0: /* no timeout */
 883                 break;
 884         }
 885 }
 886
 887 static
 888 void lttng_ust_cleanup(int exiting)
 889 {
 890         cleanup_sock_info(&global_apps, exiting);
 891         if (local_apps.allowed) {
 892                 cleanup_sock_info(&local_apps, exiting);
 893         }
 894         lttng_ust_abi_exit();
 895         lttng_ust_events_exit();
 896         ltt_ring_buffer_client_discard_exit();
 897         ltt_ring_buffer_client_overwrite_exit();
 898         ltt_ring_buffer_metadata_client_exit();
 899         exit_tracepoint();
 900         if (!exiting) {
 901                 /* Reinitialize values for fork */
 902                 sem_count = 2;
 903                 lttng_ust_comm_should_quit = 0;
 904                 initialized = 0;
 905         }
 906 }
 907
 908 void __attribute__((destructor)) lttng_ust_exit(void)
 909 {
 910         int ret;
 911
 912         /*
 913          * Using pthread_cancel here because:
 914          * A) we don't want to hang application teardown.
 915          * B) the thread is not allocating any resource.
 916          */
 917
 918         /*
 919          * Require the communication thread to quit. Synchronize with
 920          * mutexes to ensure it is not in a mutex critical section when
 921          * pthread_cancel is later called.
 922          */
 923         ust_lock();
 924         lttng_ust_comm_should_quit = 1;
 925         ust_unlock();
 926
 927         ret = pthread_cancel(global_apps.ust_listener);
 928         if (ret) {
 929                 ERR("Error cancelling global ust listener thread");
 930         }
 931         if (local_apps.allowed) {
 932                 ret = pthread_cancel(local_apps.ust_listener);
 933                 if (ret) {
 934                         ERR("Error cancelling local ust listener thread");
 935                 }
 936         }
 937         /*
 938          * We cannot join the threads because they might be waiting on
 939          * sys_futex. Simply let the OS exit() clean up those threads.
 940          */
 941         lttng_ust_cleanup(1);
 942 }
 943
 944 /*
 945  * We exclude the worker threads across fork and clone (except
 946  * CLONE_VM), because these system calls only keep the forking thread
 947  * running in the child.  Therefore, we don't want to call fork or clone
 948  * in the middle of an tracepoint or ust tracing state modification.
 949  * Holding this mutex protects these structures across fork and clone.
 950  */
 951 void ust_before_fork(sigset_t *save_sigset)
 952 {
 953         /*
 954          * Disable signals. This is to avoid that the child intervenes
 955          * before it is properly setup for tracing. It is safer to
 956          * disable all signals, because then we know we are not breaking
 957          * anything by restoring the original mask.
 958          */
 959         sigset_t all_sigs;
 960         int ret;
 961
 962         /* Disable signals */
 963         sigfillset(&all_sigs);
 964         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
 965         if (ret == -1) {
 966                 PERROR("sigprocmask");
 967         }
 968         ust_lock();
 969         rcu_bp_before_fork();
 970 }
 971
 972 static void ust_after_fork_common(sigset_t *restore_sigset)
 973 {
 974         int ret;
 975
 976         DBG("process %d", getpid());
 977         ust_unlock();
 978         /* Restore signals */
 979         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
 980         if (ret == -1) {
 981                 PERROR("sigprocmask");
 982         }
 983 }
 984
 985 void ust_after_fork_parent(sigset_t *restore_sigset)
 986 {
 987         DBG("process %d", getpid());
 988         rcu_bp_after_fork_parent();
 989         /* Release mutexes and reenable signals */
 990         ust_after_fork_common(restore_sigset);
 991 }
 992
 993 /*
 994  * After fork, in the child, we need to cleanup all the leftover state,
 995  * except the worker thread which already magically disappeared thanks
 996  * to the weird Linux fork semantics. After tyding up, we call
 997  * lttng_ust_init() again to start over as a new PID.
 998  *
 999  * This is meant for forks() that have tracing in the child between the
1000  * fork and following exec call (if there is any).
1001  */
1002 void ust_after_fork_child(sigset_t *restore_sigset)
1003 {
1004         DBG("process %d", getpid());
1005         /* Release urcu mutexes */
1006         rcu_bp_after_fork_child();
1007         lttng_ust_cleanup(0);
1008         lttng_context_vtid_reset();
1009         /* Release mutexes and reenable signals */
1010         ust_after_fork_common(restore_sigset);
1011         lttng_ust_init();
1012 }