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 #include "../libringbuffer/tlsfixup.h"
  50
  51 /*
  52  * Has lttng ust comm constructor been called ?
  53  */
  54 static int initialized;
  55
  56 /*
  57  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  58  * Held when handling a command, also held by fork() to deal with
  59  * removal of threads, and by exit path.
  60  */
  61
  62 /* Should the ust comm thread quit ? */
  63 static int lttng_ust_comm_should_quit;
  64
  65 /*
  66  * Wait for either of these before continuing to the main
  67  * program:
  68  * - the register_done message from sessiond daemon
  69  *   (will let the sessiond daemon enable sessions before main
  70  *   starts.)
  71  * - sessiond daemon is not reachable.
  72  * - timeout (ensuring applications are resilient to session
  73  *   daemon problems).
  74  */
  75 static sem_t constructor_wait;
  76 /*
  77  * Doing this for both the global and local sessiond.
  78  */
  79 static int sem_count = { 2 };
  80
  81 /*
  82  * Info about socket and associated listener thread.
  83  */
  84 struct sock_info {
  85         const char *name;
  86         pthread_t ust_listener; /* listener thread */
  87         int root_handle;
  88         int constructor_sem_posted;
  89         int allowed;
  90         int global;
  91
  92         char sock_path[PATH_MAX];
  93         int socket;
  94
  95         char wait_shm_path[PATH_MAX];
  96         char *wait_shm_mmap;
  97 };
  98
  99 /* Socket from app (connect) to session daemon (listen) for communication */
 100 struct sock_info global_apps = {
 101         .name = "global",
 102         .global = 1,
 103
 104         .root_handle = -1,
 105         .allowed = 1,
 106
 107         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 108         .socket = -1,
 109
 110         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 111 };
 112
 113 /* TODO: allow global_apps_sock_path override */
 114
 115 struct sock_info local_apps = {
 116         .name = "local",
 117         .global = 0,
 118         .root_handle = -1,
 119         .allowed = 0,   /* Check setuid bit first */
 120
 121         .socket = -1,
 122 };
 123
 124 static int wait_poll_fallback;
 125
 126 extern void ltt_ring_buffer_client_overwrite_init(void);
 127 extern void ltt_ring_buffer_client_discard_init(void);
 128 extern void ltt_ring_buffer_metadata_client_init(void);
 129 extern void ltt_ring_buffer_client_overwrite_exit(void);
 130 extern void ltt_ring_buffer_client_discard_exit(void);
 131 extern void ltt_ring_buffer_metadata_client_exit(void);
 132
 133 static
 134 int setup_local_apps(void)
 135 {
 136         const char *home_dir;
 137         uid_t uid;
 138
 139         uid = getuid();
 140         /*
 141          * Disallow per-user tracing for setuid binaries.
 142          */
 143         if (uid != geteuid()) {
 144                 local_apps.allowed = 0;
 145                 return 0;
 146         } else {
 147                 local_apps.allowed = 1;
 148         }
 149         home_dir = (const char *) getenv("HOME");
 150         if (!home_dir)
 151                 return -ENOENT;
 152         snprintf(local_apps.sock_path, PATH_MAX,
 153                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 154         snprintf(local_apps.wait_shm_path, PATH_MAX,
 155                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 156         return 0;
 157 }
 158
 159 static
 160 int register_app_to_sessiond(int socket)
 161 {
 162         ssize_t ret;
 163         int prctl_ret;
 164         struct {
 165                 uint32_t major;
 166                 uint32_t minor;
 167                 pid_t pid;
 168                 pid_t ppid;
 169                 uid_t uid;
 170                 gid_t gid;
 171                 uint32_t bits_per_long;
 172                 char name[16];  /* process name */
 173         } reg_msg;
 174
 175         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 176         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 177         reg_msg.pid = getpid();
 178         reg_msg.ppid = getppid();
 179         reg_msg.uid = getuid();
 180         reg_msg.gid = getgid();
 181         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 182         prctl_ret = prctl(PR_GET_NAME, (unsigned long) reg_msg.name, 0, 0, 0);
 183         if (prctl_ret) {
 184                 ERR("Error executing prctl");
 185                 return -errno;
 186         }
 187
 188         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 189         if (ret >= 0 && ret != sizeof(reg_msg))
 190                 return -EIO;
 191         return ret;
 192 }
 193
 194 static
 195 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 196 {
 197         ssize_t len;
 198
 199         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 200         switch (len) {
 201         case sizeof(*lur):
 202                 DBG("message successfully sent");
 203                 return 0;
 204         case -1:
 205                 if (errno == ECONNRESET) {
 206                         printf("remote end closed connection\n");
 207                         return 0;
 208                 }
 209                 return -1;
 210         default:
 211                 printf("incorrect message size: %zd\n", len);
 212                 return -1;
 213         }
 214 }
 215
 216 static
 217 int handle_register_done(struct sock_info *sock_info)
 218 {
 219         int ret;
 220
 221         if (sock_info->constructor_sem_posted)
 222                 return 0;
 223         sock_info->constructor_sem_posted = 1;
 224         if (uatomic_read(&sem_count) <= 0) {
 225                 return 0;
 226         }
 227         ret = uatomic_add_return(&sem_count, -1);
 228         if (ret == 0) {
 229                 ret = sem_post(&constructor_wait);
 230                 assert(!ret);
 231         }
 232         return 0;
 233 }
 234
 235 static
 236 int handle_message(struct sock_info *sock_info,
 237                 int sock, struct ustcomm_ust_msg *lum)
 238 {
 239         int ret = 0;
 240         const struct lttng_ust_objd_ops *ops;
 241         struct ustcomm_ust_reply lur;
 242         int shm_fd, wait_fd;
 243         union ust_args args;
 244
 245         ust_lock();
 246
 247         memset(&lur, 0, sizeof(lur));
 248
 249         if (lttng_ust_comm_should_quit) {
 250                 ret = -EPERM;
 251                 goto end;
 252         }
 253
 254         ops = objd_ops(lum->handle);
 255         if (!ops) {
 256                 ret = -ENOENT;
 257                 goto end;
 258         }
 259
 260         switch (lum->cmd) {
 261         case LTTNG_UST_REGISTER_DONE:
 262                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 263                         ret = handle_register_done(sock_info);
 264                 else
 265                         ret = -EINVAL;
 266                 break;
 267         case LTTNG_UST_RELEASE:
 268                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 269                         ret = -EPERM;
 270                 else
 271                         ret = lttng_ust_objd_unref(lum->handle);
 272                 break;
 273         default:
 274                 if (ops->cmd)
 275                         ret = ops->cmd(lum->handle, lum->cmd,
 276                                         (unsigned long) &lum->u,
 277                                         &args);
 278                 else
 279                         ret = -ENOSYS;
 280                 break;
 281         }
 282
 283 end:
 284         lur.handle = lum->handle;
 285         lur.cmd = lum->cmd;
 286         lur.ret_val = ret;
 287         if (ret >= 0) {
 288                 lur.ret_code = USTCOMM_OK;
 289         } else {
 290                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 291                 lur.ret_code = ret;
 292         }
 293         if (ret >= 0) {
 294                 switch (lum->cmd) {
 295                 case LTTNG_UST_STREAM:
 296                         /*
 297                          * Special-case reply to send stream info.
 298                          * Use lum.u output.
 299                          */
 300                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 301                         shm_fd = *args.stream.shm_fd;
 302                         wait_fd = *args.stream.wait_fd;
 303                         break;
 304                 case LTTNG_UST_METADATA:
 305                 case LTTNG_UST_CHANNEL:
 306                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 307                         shm_fd = *args.channel.shm_fd;
 308                         wait_fd = *args.channel.wait_fd;
 309                         break;
 310                 case LTTNG_UST_TRACER_VERSION:
 311                         lur.u.version = lum->u.version;
 312                         break;
 313                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 314                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 315                         break;
 316                 }
 317         }
 318         ret = send_reply(sock, &lur);
 319         if (ret < 0) {
 320                 perror("error sending reply");
 321                 goto error;
 322         }
 323
 324         if ((lum->cmd == LTTNG_UST_STREAM
 325              || lum->cmd == LTTNG_UST_CHANNEL
 326              || lum->cmd == LTTNG_UST_METADATA)
 327                         && lur.ret_code == USTCOMM_OK) {
 328                 int sendret = 0;
 329
 330                 /* we also need to send the file descriptors. */
 331                 ret = ustcomm_send_fds_unix_sock(sock,
 332                         &shm_fd, &shm_fd,
 333                         1, sizeof(int));
 334                 if (ret < 0) {
 335                         perror("send shm_fd");
 336                         sendret = ret;
 337                 }
 338                 /*
 339                  * The sessiond expects 2 file descriptors, even upon
 340                  * error.
 341                  */
 342                 ret = ustcomm_send_fds_unix_sock(sock,
 343                         &wait_fd, &wait_fd,
 344                         1, sizeof(int));
 345                 if (ret < 0) {
 346                         perror("send wait_fd");
 347                         goto error;
 348                 }
 349                 if (sendret) {
 350                         ret = sendret;
 351                         goto error;
 352                 }
 353         }
 354         /*
 355          * We still have the memory map reference, and the fds have been
 356          * sent to the sessiond. We can therefore close those fds. Note
 357          * that we keep the write side of the wait_fd open, but close
 358          * the read side.
 359          */
 360         if (lur.ret_code == USTCOMM_OK) {
 361                 switch (lum->cmd) {
 362                 case LTTNG_UST_STREAM:
 363                         if (shm_fd >= 0) {
 364                                 ret = close(shm_fd);
 365                                 if (ret) {
 366                                         PERROR("Error closing stream shm_fd");
 367                                 }
 368                                 *args.stream.shm_fd = -1;
 369                         }
 370                         if (wait_fd >= 0) {
 371                                 ret = close(wait_fd);
 372                                 if (ret) {
 373                                         PERROR("Error closing stream wait_fd");
 374                                 }
 375                                 *args.stream.wait_fd = -1;
 376                         }
 377                         break;
 378                 case LTTNG_UST_METADATA:
 379                 case LTTNG_UST_CHANNEL:
 380                         if (shm_fd >= 0) {
 381                                 ret = close(shm_fd);
 382                                 if (ret) {
 383                                         PERROR("Error closing channel shm_fd");
 384                                 }
 385                                 *args.channel.shm_fd = -1;
 386                         }
 387                         if (wait_fd >= 0) {
 388                                 ret = close(wait_fd);
 389                                 if (ret) {
 390                                         PERROR("Error closing channel wait_fd");
 391                                 }
 392                                 *args.channel.wait_fd = -1;
 393                         }
 394                         break;
 395                 }
 396         }
 397
 398 error:
 399         ust_unlock();
 400         return ret;
 401 }
 402
 403 static
 404 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 405 {
 406         int ret;
 407
 408         if (sock_info->socket != -1) {
 409                 ret = ustcomm_close_unix_sock(sock_info->socket);
 410                 if (ret) {
 411                         ERR("Error closing apps socket");
 412                 }
 413                 sock_info->socket = -1;
 414         }
 415         if (sock_info->root_handle != -1) {
 416                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 417                 if (ret) {
 418                         ERR("Error unref root handle");
 419                 }
 420                 sock_info->root_handle = -1;
 421         }
 422         sock_info->constructor_sem_posted = 0;
 423         /*
 424          * wait_shm_mmap is used by listener threads outside of the
 425          * ust lock, so we cannot tear it down ourselves, because we
 426          * cannot join on these threads. Leave this task to the OS
 427          * process exit.
 428          */
 429         if (!exiting && sock_info->wait_shm_mmap) {
 430                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 431                 if (ret) {
 432                         ERR("Error unmapping wait shm");
 433                 }
 434                 sock_info->wait_shm_mmap = NULL;
 435         }
 436 }
 437
 438 /*
 439  * Using fork to set umask in the child process (not multi-thread safe).
 440  * We deal with the shm_open vs ftruncate race (happening when the
 441  * sessiond owns the shm and does not let everybody modify it, to ensure
 442  * safety against shm_unlink) by simply letting the mmap fail and
 443  * retrying after a few seconds.
 444  * For global shm, everybody has rw access to it until the sessiond
 445  * starts.
 446  */
 447 static
 448 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 449 {
 450         int wait_shm_fd, ret;
 451         pid_t pid;
 452
 453         /*
 454          * Try to open read-only.
 455          */
 456         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 457         if (wait_shm_fd >= 0) {
 458                 goto end;
 459         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 460                 /*
 461                  * Real-only open did not work, and it's not because the
 462                  * entry was not present. It's a failure that prohibits
 463                  * using shm.
 464                  */
 465                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 466                 goto end;
 467         }
 468         /*
 469          * If the open failed because the file did not exist, try
 470          * creating it ourself.
 471          */
 472         pid = fork();
 473         if (pid > 0) {
 474                 int status;
 475
 476                 /*
 477                  * Parent: wait for child to return, in which case the
 478                  * shared memory map will have been created.
 479                  */
 480                 pid = wait(&status);
 481                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 482                         wait_shm_fd = -1;
 483                         goto end;
 484                 }
 485                 /*
 486                  * Try to open read-only again after creation.
 487                  */
 488                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 489                 if (wait_shm_fd < 0) {
 490                         /*
 491                          * Real-only open did not work. It's a failure
 492                          * that prohibits using shm.
 493                          */
 494                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 495                         goto end;
 496                 }
 497                 goto end;
 498         } else if (pid == 0) {
 499                 int create_mode;
 500
 501                 /* Child */
 502                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 503                 if (sock_info->global)
 504                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 505                 /*
 506                  * We're alone in a child process, so we can modify the
 507                  * process-wide umask.
 508                  */
 509                 umask(~create_mode);
 510                 /*
 511                  * Try creating shm (or get rw access).
 512                  * We don't do an exclusive open, because we allow other
 513                  * processes to create+ftruncate it concurrently.
 514                  */
 515                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 516                                 O_RDWR | O_CREAT, create_mode);
 517                 if (wait_shm_fd >= 0) {
 518                         ret = ftruncate(wait_shm_fd, mmap_size);
 519                         if (ret) {
 520                                 PERROR("ftruncate");
 521                                 exit(EXIT_FAILURE);
 522                         }
 523                         exit(EXIT_SUCCESS);
 524                 }
 525                 /*
 526                  * For local shm, we need to have rw access to accept
 527                  * opening it: this means the local sessiond will be
 528                  * able to wake us up. For global shm, we open it even
 529                  * if rw access is not granted, because the root.root
 530                  * sessiond will be able to override all rights and wake
 531                  * us up.
 532                  */
 533                 if (!sock_info->global && errno != EACCES) {
 534                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 535                         exit(EXIT_FAILURE);
 536                 }
 537                 /*
 538                  * The shm exists, but we cannot open it RW. Report
 539                  * success.
 540                  */
 541                 exit(EXIT_SUCCESS);
 542         } else {
 543                 return -1;
 544         }
 545 end:
 546         if (wait_shm_fd >= 0 && !sock_info->global) {
 547                 struct stat statbuf;
 548
 549                 /*
 550                  * Ensure that our user is the owner of the shm file for
 551                  * local shm. If we do not own the file, it means our
 552                  * sessiond will not have access to wake us up (there is
 553                  * probably a rogue process trying to fake our
 554                  * sessiond). Fallback to polling method in this case.
 555                  */
 556                 ret = fstat(wait_shm_fd, &statbuf);
 557                 if (ret) {
 558                         PERROR("fstat");
 559                         goto error_close;
 560                 }
 561                 if (statbuf.st_uid != getuid())
 562                         goto error_close;
 563         }
 564         return wait_shm_fd;
 565
 566 error_close:
 567         ret = close(wait_shm_fd);
 568         if (ret) {
 569                 PERROR("Error closing fd");
 570         }
 571         return -1;
 572 }
 573
 574 static
 575 char *get_map_shm(struct sock_info *sock_info)
 576 {
 577         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 578         int wait_shm_fd, ret;
 579         char *wait_shm_mmap;
 580
 581         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 582         if (wait_shm_fd < 0) {
 583                 goto error;
 584         }
 585         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 586                   MAP_SHARED, wait_shm_fd, 0);
 587         /* close shm fd immediately after taking the mmap reference */
 588         ret = close(wait_shm_fd);
 589         if (ret) {
 590                 PERROR("Error closing fd");
 591         }
 592         if (wait_shm_mmap == MAP_FAILED) {
 593                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 594                 goto error;
 595         }
 596         return wait_shm_mmap;
 597
 598 error:
 599         return NULL;
 600 }
 601
 602 static
 603 void wait_for_sessiond(struct sock_info *sock_info)
 604 {
 605         int ret;
 606
 607         ust_lock();
 608         if (lttng_ust_comm_should_quit) {
 609                 goto quit;
 610         }
 611         if (wait_poll_fallback) {
 612                 goto error;
 613         }
 614         if (!sock_info->wait_shm_mmap) {
 615                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 616                 if (!sock_info->wait_shm_mmap)
 617                         goto error;
 618         }
 619         ust_unlock();
 620
 621         DBG("Waiting for %s apps sessiond", sock_info->name);
 622         /* Wait for futex wakeup */
 623         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 624                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 625                         FUTEX_WAIT, 0, NULL, NULL, 0);
 626                 if (ret < 0) {
 627                         if (errno == EFAULT) {
 628                                 wait_poll_fallback = 1;
 629                                 DBG(
 630 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 631 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 632 "Please upgrade your kernel "
 633 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 634 "mainline). LTTng-UST will use polling mode fallback.");
 635                                 if (ust_debug())
 636                                         PERROR("futex");
 637                         }
 638                 }
 639         }
 640         return;
 641
 642 quit:
 643         ust_unlock();
 644         return;
 645
 646 error:
 647         ust_unlock();
 648         return;
 649 }
 650
 651 /*
 652  * This thread does not allocate any resource, except within
 653  * handle_message, within mutex protection. This mutex protects against
 654  * fork and exit.
 655  * The other moment it allocates resources is at socket connexion, which
 656  * is also protected by the mutex.
 657  */
 658 static
 659 void *ust_listener_thread(void *arg)
 660 {
 661         struct sock_info *sock_info = arg;
 662         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 663
 664         /* Restart trying to connect to the session daemon */
 665 restart:
 666         if (prev_connect_failed) {
 667                 /* Wait for sessiond availability with pipe */
 668                 wait_for_sessiond(sock_info);
 669                 if (has_waited) {
 670                         has_waited = 0;
 671                         /*
 672                          * Sleep for 5 seconds before retrying after a
 673                          * sequence of failure / wait / failure. This
 674                          * deals with a killed or broken session daemon.
 675                          */
 676                         sleep(5);
 677                 }
 678                 has_waited = 1;
 679                 prev_connect_failed = 0;
 680         }
 681         ust_lock();
 682
 683         if (lttng_ust_comm_should_quit) {
 684                 ust_unlock();
 685                 goto quit;
 686         }
 687
 688         if (sock_info->socket != -1) {
 689                 ret = ustcomm_close_unix_sock(sock_info->socket);
 690                 if (ret) {
 691                         ERR("Error closing %s apps socket", sock_info->name);
 692                 }
 693                 sock_info->socket = -1;
 694         }
 695
 696         /* Register */
 697         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 698         if (ret < 0) {
 699                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 700                 prev_connect_failed = 1;
 701                 /*
 702                  * If we cannot find the sessiond daemon, don't delay
 703                  * constructor execution.
 704                  */
 705                 ret = handle_register_done(sock_info);
 706                 assert(!ret);
 707                 ust_unlock();
 708                 goto restart;
 709         }
 710
 711         sock_info->socket = sock = ret;
 712
 713         /*
 714          * Create only one root handle per listener thread for the whole
 715          * process lifetime.
 716          */
 717         if (sock_info->root_handle == -1) {
 718                 ret = lttng_abi_create_root_handle();
 719                 if (ret < 0) {
 720                         ERR("Error creating root handle");
 721                         ust_unlock();
 722                         goto quit;
 723                 }
 724                 sock_info->root_handle = ret;
 725         }
 726
 727         ret = register_app_to_sessiond(sock);
 728         if (ret < 0) {
 729                 ERR("Error registering to %s apps socket", sock_info->name);
 730                 prev_connect_failed = 1;
 731                 /*
 732                  * If we cannot register to the sessiond daemon, don't
 733                  * delay constructor execution.
 734                  */
 735                 ret = handle_register_done(sock_info);
 736                 assert(!ret);
 737                 ust_unlock();
 738                 goto restart;
 739         }
 740         ust_unlock();
 741
 742         for (;;) {
 743                 ssize_t len;
 744                 struct ustcomm_ust_msg lum;
 745
 746                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 747                 switch (len) {
 748                 case 0: /* orderly shutdown */
 749                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 750                         ust_lock();
 751                         /*
 752                          * Either sessiond has shutdown or refused us by closing the socket.
 753                          * In either case, we don't want to delay construction execution,
 754                          * and we need to wait before retry.
 755                          */
 756                         prev_connect_failed = 1;
 757                         /*
 758                          * If we cannot register to the sessiond daemon, don't
 759                          * delay constructor execution.
 760                          */
 761                         ret = handle_register_done(sock_info);
 762                         assert(!ret);
 763                         ust_unlock();
 764                         goto end;
 765                 case sizeof(lum):
 766                         DBG("message received\n");
 767                         ret = handle_message(sock_info, sock, &lum);
 768                         if (ret < 0) {
 769                                 ERR("Error handling message for %s socket", sock_info->name);
 770                         }
 771                         continue;
 772                 case -1:
 773                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 774                         if (errno == ECONNRESET) {
 775                                 ERR("%s remote end closed connection\n", sock_info->name);
 776                                 goto end;
 777                         }
 778                         goto end;
 779                 default:
 780                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 781                         continue;
 782                 }
 783
 784         }
 785 end:
 786         goto restart;   /* try to reconnect */
 787 quit:
 788         return NULL;
 789 }
 790
 791 /*
 792  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 793  */
 794 static
 795 int get_timeout(struct timespec *constructor_timeout)
 796 {
 797         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 798         char *str_delay;
 799         int ret;
 800
 801         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 802         if (str_delay) {
 803                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 804         }
 805
 806         switch (constructor_delay_ms) {
 807         case -1:/* fall-through */
 808         case 0:
 809                 return constructor_delay_ms;
 810         default:
 811                 break;
 812         }
 813
 814         /*
 815          * If we are unable to find the current time, don't wait.
 816          */
 817         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 818         if (ret) {
 819                 return -1;
 820         }
 821         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 822         constructor_timeout->tv_nsec +=
 823                 (constructor_delay_ms % 1000UL) * 1000000UL;
 824         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 825                 constructor_timeout->tv_sec++;
 826                 constructor_timeout->tv_nsec -= 1000000000UL;
 827         }
 828         return 1;
 829 }
 830
 831 /*
 832  * sessiond monitoring thread: monitor presence of global and per-user
 833  * sessiond by polling the application common named pipe.
 834  */
 835 /* TODO */
 836
 837 void __attribute__((constructor)) lttng_ust_init(void)
 838 {
 839         struct timespec constructor_timeout;
 840         sigset_t sig_all_blocked, orig_parent_mask;
 841         int timeout_mode;
 842         int ret;
 843
 844         if (uatomic_xchg(&initialized, 1) == 1)
 845                 return;
 846
 847         /*
 848          * Fixup interdependency between TLS fixup mutex (which happens
 849          * to be the dynamic linker mutex) and ust_lock, taken within
 850          * the ust lock.
 851          */
 852         lttng_fixup_event_tls();
 853         lttng_fixup_ringbuffer_tls();
 854         lttng_fixup_vtid_tls();
 855
 856         /*
 857          * We want precise control over the order in which we construct
 858          * our sub-libraries vs starting to receive commands from
 859          * sessiond (otherwise leading to errors when trying to create
 860          * sessiond before the init functions are completed).
 861          */
 862         init_usterr();
 863         init_tracepoint();
 864         ltt_ring_buffer_metadata_client_init();
 865         ltt_ring_buffer_client_overwrite_init();
 866         ltt_ring_buffer_client_discard_init();
 867
 868         timeout_mode = get_timeout(&constructor_timeout);
 869
 870         ret = sem_init(&constructor_wait, 0, 0);
 871         assert(!ret);
 872
 873         ret = setup_local_apps();
 874         if (ret) {
 875                 ERR("Error setting up to local apps");
 876         }
 877
 878         /* A new thread created by pthread_create inherits the signal mask
 879          * from the parent. To avoid any signal being received by the
 880          * listener thread, we block all signals temporarily in the parent,
 881          * while we create the listener thread.
 882          */
 883         sigfillset(&sig_all_blocked);
 884         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 885         if (ret) {
 886                 ERR("pthread_sigmask: %s", strerror(ret));
 887         }
 888
 889         ret = pthread_create(&global_apps.ust_listener, NULL,
 890                         ust_listener_thread, &global_apps);
 891         if (ret) {
 892                 ERR("pthread_create global: %s", strerror(ret));
 893         }
 894         if (local_apps.allowed) {
 895                 ret = pthread_create(&local_apps.ust_listener, NULL,
 896                                 ust_listener_thread, &local_apps);
 897                 if (ret) {
 898                         ERR("pthread_create local: %s", strerror(ret));
 899                 }
 900         } else {
 901                 handle_register_done(&local_apps);
 902         }
 903
 904         /* Restore original signal mask in parent */
 905         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
 906         if (ret) {
 907                 ERR("pthread_sigmask: %s", strerror(ret));
 908         }
 909
 910         switch (timeout_mode) {
 911         case 1: /* timeout wait */
 912                 do {
 913                         ret = sem_timedwait(&constructor_wait,
 914                                         &constructor_timeout);
 915                 } while (ret < 0 && errno == EINTR);
 916                 if (ret < 0 && errno == ETIMEDOUT) {
 917                         ERR("Timed out waiting for ltt-sessiond");
 918                 } else {
 919                         assert(!ret);
 920                 }
 921                 break;
 922         case -1:/* wait forever */
 923                 do {
 924                         ret = sem_wait(&constructor_wait);
 925                 } while (ret < 0 && errno == EINTR);
 926                 assert(!ret);
 927                 break;
 928         case 0: /* no timeout */
 929                 break;
 930         }
 931 }
 932
 933 static
 934 void lttng_ust_cleanup(int exiting)
 935 {
 936         cleanup_sock_info(&global_apps, exiting);
 937         if (local_apps.allowed) {
 938                 cleanup_sock_info(&local_apps, exiting);
 939         }
 940         /*
 941          * The teardown in this function all affect data structures
 942          * accessed under the UST lock by the listener thread. This
 943          * lock, along with the lttng_ust_comm_should_quit flag, ensure
 944          * that none of these threads are accessing this data at this
 945          * point.
 946          */
 947         lttng_ust_abi_exit();
 948         lttng_ust_events_exit();
 949         ltt_ring_buffer_client_discard_exit();
 950         ltt_ring_buffer_client_overwrite_exit();
 951         ltt_ring_buffer_metadata_client_exit();
 952         exit_tracepoint();
 953         if (!exiting) {
 954                 /* Reinitialize values for fork */
 955                 sem_count = 2;
 956                 lttng_ust_comm_should_quit = 0;
 957                 initialized = 0;
 958         }
 959 }
 960
 961 void __attribute__((destructor)) lttng_ust_exit(void)
 962 {
 963         int ret;
 964
 965         /*
 966          * Using pthread_cancel here because:
 967          * A) we don't want to hang application teardown.
 968          * B) the thread is not allocating any resource.
 969          */
 970
 971         /*
 972          * Require the communication thread to quit. Synchronize with
 973          * mutexes to ensure it is not in a mutex critical section when
 974          * pthread_cancel is later called.
 975          */
 976         ust_lock();
 977         lttng_ust_comm_should_quit = 1;
 978         ust_unlock();
 979
 980         /* cancel threads */
 981         ret = pthread_cancel(global_apps.ust_listener);
 982         if (ret) {
 983                 ERR("Error cancelling global ust listener thread: %s",
 984                         strerror(ret));
 985         }
 986         if (local_apps.allowed) {
 987                 ret = pthread_cancel(local_apps.ust_listener);
 988                 if (ret) {
 989                         ERR("Error cancelling local ust listener thread: %s",
 990                                 strerror(ret));
 991                 }
 992         }
 993         /*
 994          * Do NOT join threads: use of sys_futex makes it impossible to
 995          * join the threads without using async-cancel, but async-cancel
 996          * is delivered by a signal, which could hit the target thread
 997          * anywhere in its code path, including while the ust_lock() is
 998          * held, causing a deadlock for the other thread. Let the OS
 999          * cleanup the threads if there are stalled in a syscall.
1000          */
1001         lttng_ust_cleanup(1);
1002 }
1003
1004 /*
1005  * We exclude the worker threads across fork and clone (except
1006  * CLONE_VM), because these system calls only keep the forking thread
1007  * running in the child.  Therefore, we don't want to call fork or clone
1008  * in the middle of an tracepoint or ust tracing state modification.
1009  * Holding this mutex protects these structures across fork and clone.
1010  */
1011 void ust_before_fork(sigset_t *save_sigset)
1012 {
1013         /*
1014          * Disable signals. This is to avoid that the child intervenes
1015          * before it is properly setup for tracing. It is safer to
1016          * disable all signals, because then we know we are not breaking
1017          * anything by restoring the original mask.
1018          */
1019         sigset_t all_sigs;
1020         int ret;
1021
1022         /* Disable signals */
1023         sigfillset(&all_sigs);
1024         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1025         if (ret == -1) {
1026                 PERROR("sigprocmask");
1027         }
1028         ust_lock();
1029         rcu_bp_before_fork();
1030 }
1031
1032 static void ust_after_fork_common(sigset_t *restore_sigset)
1033 {
1034         int ret;
1035
1036         DBG("process %d", getpid());
1037         ust_unlock();
1038         /* Restore signals */
1039         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1040         if (ret == -1) {
1041                 PERROR("sigprocmask");
1042         }
1043 }
1044
1045 void ust_after_fork_parent(sigset_t *restore_sigset)
1046 {
1047         DBG("process %d", getpid());
1048         rcu_bp_after_fork_parent();
1049         /* Release mutexes and reenable signals */
1050         ust_after_fork_common(restore_sigset);
1051 }
1052
1053 /*
1054  * After fork, in the child, we need to cleanup all the leftover state,
1055  * except the worker thread which already magically disappeared thanks
1056  * to the weird Linux fork semantics. After tyding up, we call
1057  * lttng_ust_init() again to start over as a new PID.
1058  *
1059  * This is meant for forks() that have tracing in the child between the
1060  * fork and following exec call (if there is any).
1061  */
1062 void ust_after_fork_child(sigset_t *restore_sigset)
1063 {
1064         DBG("process %d", getpid());
1065         /* Release urcu mutexes */
1066         rcu_bp_after_fork_child();
1067         lttng_ust_cleanup(0);
1068         lttng_context_vtid_reset();
1069         /* Release mutexes and reenable signals */
1070         ust_after_fork_common(restore_sigset);
1071         lttng_ust_init();
1072 }