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)
 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         if (sock_info->wait_shm_mmap) {
 406                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 407                 if (ret) {
 408                         ERR("Error unmapping wait shm");
 409                 }
 410                 sock_info->wait_shm_mmap = NULL;
 411         }
 412 }
 413
 414 /*
 415  * Using fork to set umask in the child process (not multi-thread safe).
 416  * We deal with the shm_open vs ftruncate race (happening when the
 417  * sessiond owns the shm and does not let everybody modify it, to ensure
 418  * safety against shm_unlink) by simply letting the mmap fail and
 419  * retrying after a few seconds.
 420  * For global shm, everybody has rw access to it until the sessiond
 421  * starts.
 422  */
 423 static
 424 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 425 {
 426         int wait_shm_fd, ret;
 427         pid_t pid;
 428
 429         /*
 430          * Try to open read-only.
 431          */
 432         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 433         if (wait_shm_fd >= 0) {
 434                 goto end;
 435         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 436                 /*
 437                  * Real-only open did not work, and it's not because the
 438                  * entry was not present. It's a failure that prohibits
 439                  * using shm.
 440                  */
 441                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 442                 goto end;
 443         }
 444         /*
 445          * If the open failed because the file did not exist, try
 446          * creating it ourself.
 447          */
 448         pid = fork();
 449         if (pid > 0) {
 450                 int status;
 451
 452                 /*
 453                  * Parent: wait for child to return, in which case the
 454                  * shared memory map will have been created.
 455                  */
 456                 pid = wait(&status);
 457                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 458                         wait_shm_fd = -1;
 459                         goto end;
 460                 }
 461                 /*
 462                  * Try to open read-only again after creation.
 463                  */
 464                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 465                 if (wait_shm_fd < 0) {
 466                         /*
 467                          * Real-only open did not work. It's a failure
 468                          * that prohibits using shm.
 469                          */
 470                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 471                         goto end;
 472                 }
 473                 goto end;
 474         } else if (pid == 0) {
 475                 int create_mode;
 476
 477                 /* Child */
 478                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 479                 if (sock_info->global)
 480                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 481                 /*
 482                  * We're alone in a child process, so we can modify the
 483                  * process-wide umask.
 484                  */
 485                 umask(~create_mode);
 486                 /*
 487                  * Try creating shm (or get rw access).
 488                  * We don't do an exclusive open, because we allow other
 489                  * processes to create+ftruncate it concurrently.
 490                  */
 491                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 492                                 O_RDWR | O_CREAT, create_mode);
 493                 if (wait_shm_fd >= 0) {
 494                         ret = ftruncate(wait_shm_fd, mmap_size);
 495                         if (ret) {
 496                                 PERROR("ftruncate");
 497                                 exit(EXIT_FAILURE);
 498                         }
 499                         exit(EXIT_SUCCESS);
 500                 }
 501                 /*
 502                  * For local shm, we need to have rw access to accept
 503                  * opening it: this means the local sessiond will be
 504                  * able to wake us up. For global shm, we open it even
 505                  * if rw access is not granted, because the root.root
 506                  * sessiond will be able to override all rights and wake
 507                  * us up.
 508                  */
 509                 if (!sock_info->global && errno != EACCES) {
 510                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 511                         exit(EXIT_FAILURE);
 512                 }
 513                 /*
 514                  * The shm exists, but we cannot open it RW. Report
 515                  * success.
 516                  */
 517                 exit(EXIT_SUCCESS);
 518         } else {
 519                 return -1;
 520         }
 521 end:
 522         if (wait_shm_fd >= 0 && !sock_info->global) {
 523                 struct stat statbuf;
 524
 525                 /*
 526                  * Ensure that our user is the owner of the shm file for
 527                  * local shm. If we do not own the file, it means our
 528                  * sessiond will not have access to wake us up (there is
 529                  * probably a rogue process trying to fake our
 530                  * sessiond). Fallback to polling method in this case.
 531                  */
 532                 ret = fstat(wait_shm_fd, &statbuf);
 533                 if (ret) {
 534                         PERROR("fstat");
 535                         goto error_close;
 536                 }
 537                 if (statbuf.st_uid != getuid())
 538                         goto error_close;
 539         }
 540         return wait_shm_fd;
 541
 542 error_close:
 543         ret = close(wait_shm_fd);
 544         if (ret) {
 545                 PERROR("Error closing fd");
 546         }
 547         return -1;
 548 }
 549
 550 static
 551 char *get_map_shm(struct sock_info *sock_info)
 552 {
 553         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 554         int wait_shm_fd, ret;
 555         char *wait_shm_mmap;
 556
 557         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 558         if (wait_shm_fd < 0) {
 559                 goto error;
 560         }
 561         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 562                   MAP_SHARED, wait_shm_fd, 0);
 563         /* close shm fd immediately after taking the mmap reference */
 564         ret = close(wait_shm_fd);
 565         if (ret) {
 566                 PERROR("Error closing fd");
 567         }
 568         if (wait_shm_mmap == MAP_FAILED) {
 569                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 570                 goto error;
 571         }
 572         return wait_shm_mmap;
 573
 574 error:
 575         return NULL;
 576 }
 577
 578 static
 579 void wait_for_sessiond(struct sock_info *sock_info)
 580 {
 581         int ret, oldtype;
 582
 583         ust_lock();
 584         if (lttng_ust_comm_should_quit) {
 585                 goto quit;
 586         }
 587         if (wait_poll_fallback) {
 588                 goto error;
 589         }
 590         if (!sock_info->wait_shm_mmap) {
 591                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 592                 if (!sock_info->wait_shm_mmap)
 593                         goto error;
 594         }
 595         ust_unlock();
 596
 597         DBG("Waiting for %s apps sessiond", sock_info->name);
 598         /*
 599          * sys_futex does not honor pthread cancel requests. Set to
 600          * async.
 601          */
 602         ret = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &oldtype);
 603         if (ret) {
 604                 ERR("Error setting thread cancel type");
 605         }
 606         /* Wait for futex wakeup */
 607         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 608                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 609                         FUTEX_WAIT, 0, NULL, NULL, 0);
 610                 if (ret < 0) {
 611                         if (errno == EFAULT) {
 612                                 wait_poll_fallback = 1;
 613                                 DBG(
 614 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 615 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 616 "Please upgrade your kernel "
 617 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 618 "mainline). LTTng-UST will use polling mode fallback.");
 619                                 if (ust_debug())
 620                                         PERROR("futex");
 621                         }
 622                 }
 623         }
 624         ret = pthread_setcanceltype(oldtype, &oldtype);
 625         if (ret) {
 626                 ERR("Error setting thread cancel type");
 627         }
 628         return;
 629
 630 quit:
 631         ust_unlock();
 632         return;
 633
 634 error:
 635         ust_unlock();
 636         return;
 637 }
 638
 639 /*
 640  * This thread does not allocate any resource, except within
 641  * handle_message, within mutex protection. This mutex protects against
 642  * fork and exit.
 643  * The other moment it allocates resources is at socket connexion, which
 644  * is also protected by the mutex.
 645  */
 646 static
 647 void *ust_listener_thread(void *arg)
 648 {
 649         struct sock_info *sock_info = arg;
 650         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 651
 652         /* Restart trying to connect to the session daemon */
 653 restart:
 654         if (prev_connect_failed) {
 655                 /* Wait for sessiond availability with pipe */
 656                 wait_for_sessiond(sock_info);
 657                 if (has_waited) {
 658                         has_waited = 0;
 659                         /*
 660                          * Sleep for 5 seconds before retrying after a
 661                          * sequence of failure / wait / failure. This
 662                          * deals with a killed or broken session daemon.
 663                          */
 664                         sleep(5);
 665                 }
 666                 has_waited = 1;
 667                 prev_connect_failed = 0;
 668         }
 669         ust_lock();
 670
 671         if (lttng_ust_comm_should_quit) {
 672                 ust_unlock();
 673                 goto quit;
 674         }
 675
 676         if (sock_info->socket != -1) {
 677                 ret = close(sock_info->socket);
 678                 if (ret) {
 679                         ERR("Error closing %s apps socket", sock_info->name);
 680                 }
 681                 sock_info->socket = -1;
 682         }
 683
 684         /* Register */
 685         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 686         if (ret < 0) {
 687                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 688                 prev_connect_failed = 1;
 689                 /*
 690                  * If we cannot find the sessiond daemon, don't delay
 691                  * constructor execution.
 692                  */
 693                 ret = handle_register_done(sock_info);
 694                 assert(!ret);
 695                 ust_unlock();
 696                 goto restart;
 697         }
 698
 699         sock_info->socket = sock = ret;
 700
 701         /*
 702          * Create only one root handle per listener thread for the whole
 703          * process lifetime.
 704          */
 705         if (sock_info->root_handle == -1) {
 706                 ret = lttng_abi_create_root_handle();
 707                 if (ret < 0) {
 708                         ERR("Error creating root handle");
 709                         ust_unlock();
 710                         goto quit;
 711                 }
 712                 sock_info->root_handle = ret;
 713         }
 714
 715         ret = register_app_to_sessiond(sock);
 716         if (ret < 0) {
 717                 ERR("Error registering to %s apps socket", sock_info->name);
 718                 prev_connect_failed = 1;
 719                 /*
 720                  * If we cannot register to the sessiond daemon, don't
 721                  * delay constructor execution.
 722                  */
 723                 ret = handle_register_done(sock_info);
 724                 assert(!ret);
 725                 ust_unlock();
 726                 goto restart;
 727         }
 728         ust_unlock();
 729
 730         for (;;) {
 731                 ssize_t len;
 732                 struct ustcomm_ust_msg lum;
 733
 734                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 735                 switch (len) {
 736                 case 0: /* orderly shutdown */
 737                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 738                         ust_lock();
 739                         /*
 740                          * Either sessiond has shutdown or refused us by closing the socket.
 741                          * In either case, we don't want to delay construction execution,
 742                          * and we need to wait before retry.
 743                          */
 744                         prev_connect_failed = 1;
 745                         /*
 746                          * If we cannot register to the sessiond daemon, don't
 747                          * delay constructor execution.
 748                          */
 749                         ret = handle_register_done(sock_info);
 750                         assert(!ret);
 751                         ust_unlock();
 752                         goto end;
 753                 case sizeof(lum):
 754                         DBG("message received\n");
 755                         ret = handle_message(sock_info, sock, &lum);
 756                         if (ret < 0) {
 757                                 ERR("Error handling message for %s socket", sock_info->name);
 758                         }
 759                         continue;
 760                 case -1:
 761                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 762                         if (errno == ECONNRESET) {
 763                                 ERR("%s remote end closed connection\n", sock_info->name);
 764                                 goto end;
 765                         }
 766                         goto end;
 767                 default:
 768                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 769                         continue;
 770                 }
 771
 772         }
 773 end:
 774         goto restart;   /* try to reconnect */
 775 quit:
 776         return NULL;
 777 }
 778
 779 /*
 780  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 781  */
 782 static
 783 int get_timeout(struct timespec *constructor_timeout)
 784 {
 785         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 786         char *str_delay;
 787         int ret;
 788
 789         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 790         if (str_delay) {
 791                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 792         }
 793
 794         switch (constructor_delay_ms) {
 795         case -1:/* fall-through */
 796         case 0:
 797                 return constructor_delay_ms;
 798         default:
 799                 break;
 800         }
 801
 802         /*
 803          * If we are unable to find the current time, don't wait.
 804          */
 805         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 806         if (ret) {
 807                 return -1;
 808         }
 809         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 810         constructor_timeout->tv_nsec +=
 811                 (constructor_delay_ms % 1000UL) * 1000000UL;
 812         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 813                 constructor_timeout->tv_sec++;
 814                 constructor_timeout->tv_nsec -= 1000000000UL;
 815         }
 816         return 1;
 817 }
 818
 819 /*
 820  * sessiond monitoring thread: monitor presence of global and per-user
 821  * sessiond by polling the application common named pipe.
 822  */
 823 /* TODO */
 824
 825 void __attribute__((constructor)) lttng_ust_init(void)
 826 {
 827         struct timespec constructor_timeout;
 828         int timeout_mode;
 829         int ret;
 830
 831         if (uatomic_xchg(&initialized, 1) == 1)
 832                 return;
 833
 834         /*
 835          * We want precise control over the order in which we construct
 836          * our sub-libraries vs starting to receive commands from
 837          * sessiond (otherwise leading to errors when trying to create
 838          * sessiond before the init functions are completed).
 839          */
 840         init_usterr();
 841         init_tracepoint();
 842         ltt_ring_buffer_metadata_client_init();
 843         ltt_ring_buffer_client_overwrite_init();
 844         ltt_ring_buffer_client_discard_init();
 845
 846         timeout_mode = get_timeout(&constructor_timeout);
 847
 848         ret = sem_init(&constructor_wait, 0, 0);
 849         assert(!ret);
 850
 851         ret = setup_local_apps();
 852         if (ret) {
 853                 ERR("Error setting up to local apps");
 854         }
 855         ret = pthread_create(&local_apps.ust_listener, NULL,
 856                         ust_listener_thread, &local_apps);
 857
 858         if (local_apps.allowed) {
 859                 ret = pthread_create(&global_apps.ust_listener, NULL,
 860                                 ust_listener_thread, &global_apps);
 861         } else {
 862                 handle_register_done(&local_apps);
 863         }
 864
 865         switch (timeout_mode) {
 866         case 1: /* timeout wait */
 867                 do {
 868                         ret = sem_timedwait(&constructor_wait,
 869                                         &constructor_timeout);
 870                 } while (ret < 0 && errno == EINTR);
 871                 if (ret < 0 && errno == ETIMEDOUT) {
 872                         ERR("Timed out waiting for ltt-sessiond");
 873                 } else {
 874                         assert(!ret);
 875                 }
 876                 break;
 877         case -1:/* wait forever */
 878                 do {
 879                         ret = sem_wait(&constructor_wait);
 880                 } while (ret < 0 && errno == EINTR);
 881                 assert(!ret);
 882                 break;
 883         case 0: /* no timeout */
 884                 break;
 885         }
 886 }
 887
 888 static
 889 void lttng_ust_cleanup(int exiting)
 890 {
 891         cleanup_sock_info(&global_apps);
 892         if (local_apps.allowed) {
 893                 cleanup_sock_info(&local_apps);
 894         }
 895         lttng_ust_abi_exit();
 896         lttng_ust_events_exit();
 897         ltt_ring_buffer_client_discard_exit();
 898         ltt_ring_buffer_client_overwrite_exit();
 899         ltt_ring_buffer_metadata_client_exit();
 900         exit_tracepoint();
 901         if (!exiting) {
 902                 /* Reinitialize values for fork */
 903                 sem_count = 2;
 904                 lttng_ust_comm_should_quit = 0;
 905                 initialized = 0;
 906         }
 907 }
 908
 909 void __attribute__((destructor)) lttng_ust_exit(void)
 910 {
 911         int ret;
 912
 913         /*
 914          * Using pthread_cancel here because:
 915          * A) we don't want to hang application teardown.
 916          * B) the thread is not allocating any resource.
 917          */
 918
 919         /*
 920          * Require the communication thread to quit. Synchronize with
 921          * mutexes to ensure it is not in a mutex critical section when
 922          * pthread_cancel is later called.
 923          */
 924         ust_lock();
 925         lttng_ust_comm_should_quit = 1;
 926         ust_unlock();
 927
 928         /* cancel threads */
 929         ret = pthread_cancel(global_apps.ust_listener);
 930         if (ret) {
 931                 ERR("Error cancelling global ust listener thread");
 932         }
 933         if (local_apps.allowed) {
 934                 ret = pthread_cancel(local_apps.ust_listener);
 935                 if (ret) {
 936                         ERR("Error cancelling local ust listener thread");
 937                 }
 938         }
 939         /* join threads */
 940         ret = pthread_join(global_apps.ust_listener, NULL);
 941         if (ret) {
 942                 ERR("Error joining global ust listener thread");
 943         }
 944         if (local_apps.allowed) {
 945                 ret = pthread_join(local_apps.ust_listener, NULL);
 946                 if (ret) {
 947                         ERR("Error joining local ust listener thread");
 948                 }
 949         }
 950         lttng_ust_cleanup(1);
 951 }
 952
 953 /*
 954  * We exclude the worker threads across fork and clone (except
 955  * CLONE_VM), because these system calls only keep the forking thread
 956  * running in the child.  Therefore, we don't want to call fork or clone
 957  * in the middle of an tracepoint or ust tracing state modification.
 958  * Holding this mutex protects these structures across fork and clone.
 959  */
 960 void ust_before_fork(sigset_t *save_sigset)
 961 {
 962         /*
 963          * Disable signals. This is to avoid that the child intervenes
 964          * before it is properly setup for tracing. It is safer to
 965          * disable all signals, because then we know we are not breaking
 966          * anything by restoring the original mask.
 967          */
 968         sigset_t all_sigs;
 969         int ret;
 970
 971         /* Disable signals */
 972         sigfillset(&all_sigs);
 973         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
 974         if (ret == -1) {
 975                 PERROR("sigprocmask");
 976         }
 977         ust_lock();
 978         rcu_bp_before_fork();
 979 }
 980
 981 static void ust_after_fork_common(sigset_t *restore_sigset)
 982 {
 983         int ret;
 984
 985         DBG("process %d", getpid());
 986         ust_unlock();
 987         /* Restore signals */
 988         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
 989         if (ret == -1) {
 990                 PERROR("sigprocmask");
 991         }
 992 }
 993
 994 void ust_after_fork_parent(sigset_t *restore_sigset)
 995 {
 996         DBG("process %d", getpid());
 997         rcu_bp_after_fork_parent();
 998         /* Release mutexes and reenable signals */
 999         ust_after_fork_common(restore_sigset);
1000 }
1001
1002 /*
1003  * After fork, in the child, we need to cleanup all the leftover state,
1004  * except the worker thread which already magically disappeared thanks
1005  * to the weird Linux fork semantics. After tyding up, we call
1006  * lttng_ust_init() again to start over as a new PID.
1007  *
1008  * This is meant for forks() that have tracing in the child between the
1009  * fork and following exec call (if there is any).
1010  */
1011 void ust_after_fork_child(sigset_t *restore_sigset)
1012 {
1013         DBG("process %d", getpid());
1014         /* Release urcu mutexes */
1015         rcu_bp_after_fork_child();
1016         lttng_ust_cleanup(0);
1017         lttng_context_vtid_reset();
1018         /* Release mutexes and reenable signals */
1019         ust_after_fork_common(restore_sigset);
1020         lttng_ust_init();
1021 }