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 #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  * Counting nesting within lttng-ust. Used to ensure that calling fork()
  83  * from liblttng-ust does not execute the pre/post fork handlers.
  84  */
  85 static int __thread lttng_ust_nest_count;
  86
  87 /*
  88  * Info about socket and associated listener thread.
  89  */
  90 struct sock_info {
  91         const char *name;
  92         pthread_t ust_listener; /* listener thread */
  93         int root_handle;
  94         int constructor_sem_posted;
  95         int allowed;
  96         int global;
  97
  98         char sock_path[PATH_MAX];
  99         int socket;
 100
 101         char wait_shm_path[PATH_MAX];
 102         char *wait_shm_mmap;
 103 };
 104
 105 /* Socket from app (connect) to session daemon (listen) for communication */
 106 struct sock_info global_apps = {
 107         .name = "global",
 108         .global = 1,
 109
 110         .root_handle = -1,
 111         .allowed = 1,
 112
 113         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 114         .socket = -1,
 115
 116         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 117 };
 118
 119 /* TODO: allow global_apps_sock_path override */
 120
 121 struct sock_info local_apps = {
 122         .name = "local",
 123         .global = 0,
 124         .root_handle = -1,
 125         .allowed = 0,   /* Check setuid bit first */
 126
 127         .socket = -1,
 128 };
 129
 130 static int wait_poll_fallback;
 131
 132 extern void ltt_ring_buffer_client_overwrite_init(void);
 133 extern void ltt_ring_buffer_client_discard_init(void);
 134 extern void ltt_ring_buffer_metadata_client_init(void);
 135 extern void ltt_ring_buffer_client_overwrite_exit(void);
 136 extern void ltt_ring_buffer_client_discard_exit(void);
 137 extern void ltt_ring_buffer_metadata_client_exit(void);
 138
 139 static
 140 int setup_local_apps(void)
 141 {
 142         const char *home_dir;
 143         uid_t uid;
 144
 145         uid = getuid();
 146         /*
 147          * Disallow per-user tracing for setuid binaries.
 148          */
 149         if (uid != geteuid()) {
 150                 local_apps.allowed = 0;
 151                 return 0;
 152         } else {
 153                 local_apps.allowed = 1;
 154         }
 155         home_dir = (const char *) getenv("HOME");
 156         if (!home_dir)
 157                 return -ENOENT;
 158         snprintf(local_apps.sock_path, PATH_MAX,
 159                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 160         snprintf(local_apps.wait_shm_path, PATH_MAX,
 161                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 162         return 0;
 163 }
 164
 165 static
 166 int register_app_to_sessiond(int socket)
 167 {
 168         ssize_t ret;
 169         struct {
 170                 uint32_t major;
 171                 uint32_t minor;
 172                 pid_t pid;
 173                 pid_t ppid;
 174                 uid_t uid;
 175                 gid_t gid;
 176                 uint32_t bits_per_long;
 177                 char name[16];  /* process name */
 178         } reg_msg;
 179
 180         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 181         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 182         reg_msg.pid = getpid();
 183         reg_msg.ppid = getppid();
 184         reg_msg.uid = getuid();
 185         reg_msg.gid = getgid();
 186         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 187         lttng_ust_getprocname(reg_msg.name);
 188
 189         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 190         if (ret >= 0 && ret != sizeof(reg_msg))
 191                 return -EIO;
 192         return ret;
 193 }
 194
 195 static
 196 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 197 {
 198         ssize_t len;
 199
 200         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 201         switch (len) {
 202         case sizeof(*lur):
 203                 DBG("message successfully sent");
 204                 return 0;
 205         case -1:
 206                 if (errno == ECONNRESET) {
 207                         printf("remote end closed connection\n");
 208                         return 0;
 209                 }
 210                 return -1;
 211         default:
 212                 printf("incorrect message size: %zd\n", len);
 213                 return -1;
 214         }
 215 }
 216
 217 static
 218 int handle_register_done(struct sock_info *sock_info)
 219 {
 220         int ret;
 221
 222         if (sock_info->constructor_sem_posted)
 223                 return 0;
 224         sock_info->constructor_sem_posted = 1;
 225         if (uatomic_read(&sem_count) <= 0) {
 226                 return 0;
 227         }
 228         ret = uatomic_add_return(&sem_count, -1);
 229         if (ret == 0) {
 230                 ret = sem_post(&constructor_wait);
 231                 assert(!ret);
 232         }
 233         return 0;
 234 }
 235
 236 static
 237 int handle_message(struct sock_info *sock_info,
 238                 int sock, struct ustcomm_ust_msg *lum)
 239 {
 240         int ret = 0;
 241         const struct lttng_ust_objd_ops *ops;
 242         struct ustcomm_ust_reply lur;
 243         int shm_fd, wait_fd;
 244         union ust_args args;
 245         ssize_t len;
 246
 247         ust_lock();
 248
 249         memset(&lur, 0, sizeof(lur));
 250
 251         if (lttng_ust_comm_should_quit) {
 252                 ret = -EPERM;
 253                 goto end;
 254         }
 255
 256         ops = objd_ops(lum->handle);
 257         if (!ops) {
 258                 ret = -ENOENT;
 259                 goto end;
 260         }
 261
 262         switch (lum->cmd) {
 263         case LTTNG_UST_REGISTER_DONE:
 264                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 265                         ret = handle_register_done(sock_info);
 266                 else
 267                         ret = -EINVAL;
 268                 break;
 269         case LTTNG_UST_RELEASE:
 270                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 271                         ret = -EPERM;
 272                 else
 273                         ret = lttng_ust_objd_unref(lum->handle);
 274                 break;
 275         default:
 276                 if (ops->cmd)
 277                         ret = ops->cmd(lum->handle, lum->cmd,
 278                                         (unsigned long) &lum->u,
 279                                         &args);
 280                 else
 281                         ret = -ENOSYS;
 282                 break;
 283         }
 284
 285 end:
 286         lur.handle = lum->handle;
 287         lur.cmd = lum->cmd;
 288         lur.ret_val = ret;
 289         if (ret >= 0) {
 290                 lur.ret_code = USTCOMM_OK;
 291         } else {
 292                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 293                 lur.ret_code = ret;
 294         }
 295         if (ret >= 0) {
 296                 switch (lum->cmd) {
 297                 case LTTNG_UST_STREAM:
 298                         /*
 299                          * Special-case reply to send stream info.
 300                          * Use lum.u output.
 301                          */
 302                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 303                         shm_fd = *args.stream.shm_fd;
 304                         wait_fd = *args.stream.wait_fd;
 305                         break;
 306                 case LTTNG_UST_METADATA:
 307                 case LTTNG_UST_CHANNEL:
 308                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 309                         shm_fd = *args.channel.shm_fd;
 310                         wait_fd = *args.channel.wait_fd;
 311                         break;
 312                 case LTTNG_UST_TRACER_VERSION:
 313                         lur.u.version = lum->u.version;
 314                         break;
 315                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 316                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 317                         break;
 318                 }
 319         }
 320         ret = send_reply(sock, &lur);
 321         if (ret < 0) {
 322                 perror("error sending reply");
 323                 goto error;
 324         }
 325
 326         if ((lum->cmd == LTTNG_UST_STREAM
 327              || lum->cmd == LTTNG_UST_CHANNEL
 328              || lum->cmd == LTTNG_UST_METADATA)
 329                         && lur.ret_code == USTCOMM_OK) {
 330                 int sendret = 0;
 331
 332                 /* we also need to send the file descriptors. */
 333                 ret = ustcomm_send_fds_unix_sock(sock,
 334                         &shm_fd, &shm_fd,
 335                         1, sizeof(int));
 336                 if (ret < 0) {
 337                         perror("send shm_fd");
 338                         sendret = ret;
 339                 }
 340                 /*
 341                  * The sessiond expects 2 file descriptors, even upon
 342                  * error.
 343                  */
 344                 ret = ustcomm_send_fds_unix_sock(sock,
 345                         &wait_fd, &wait_fd,
 346                         1, sizeof(int));
 347                 if (ret < 0) {
 348                         perror("send wait_fd");
 349                         goto error;
 350                 }
 351                 if (sendret) {
 352                         ret = sendret;
 353                         goto error;
 354                 }
 355         }
 356         /*
 357          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 358          * after the reply.
 359          */
 360         if (lur.ret_code == USTCOMM_OK) {
 361                 switch (lum->cmd) {
 362                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 363                         len = ustcomm_send_unix_sock(sock,
 364                                 &args.field_list.entry,
 365                                 sizeof(args.field_list.entry));
 366                         if (len != sizeof(args.field_list.entry)) {
 367                                 ret = -1;
 368                                 goto error;
 369                         }
 370                 }
 371         }
 372         /*
 373          * We still have the memory map reference, and the fds have been
 374          * sent to the sessiond. We can therefore close those fds. Note
 375          * that we keep the write side of the wait_fd open, but close
 376          * the read side.
 377          */
 378         if (lur.ret_code == USTCOMM_OK) {
 379                 switch (lum->cmd) {
 380                 case LTTNG_UST_STREAM:
 381                         if (shm_fd >= 0) {
 382                                 ret = close(shm_fd);
 383                                 if (ret) {
 384                                         PERROR("Error closing stream shm_fd");
 385                                 }
 386                                 *args.stream.shm_fd = -1;
 387                         }
 388                         if (wait_fd >= 0) {
 389                                 ret = close(wait_fd);
 390                                 if (ret) {
 391                                         PERROR("Error closing stream wait_fd");
 392                                 }
 393                                 *args.stream.wait_fd = -1;
 394                         }
 395                         break;
 396                 case LTTNG_UST_METADATA:
 397                 case LTTNG_UST_CHANNEL:
 398                         if (shm_fd >= 0) {
 399                                 ret = close(shm_fd);
 400                                 if (ret) {
 401                                         PERROR("Error closing channel shm_fd");
 402                                 }
 403                                 *args.channel.shm_fd = -1;
 404                         }
 405                         if (wait_fd >= 0) {
 406                                 ret = close(wait_fd);
 407                                 if (ret) {
 408                                         PERROR("Error closing channel wait_fd");
 409                                 }
 410                                 *args.channel.wait_fd = -1;
 411                         }
 412                         break;
 413                 }
 414         }
 415
 416 error:
 417         ust_unlock();
 418         return ret;
 419 }
 420
 421 static
 422 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 423 {
 424         int ret;
 425
 426         if (sock_info->socket != -1) {
 427                 ret = ustcomm_close_unix_sock(sock_info->socket);
 428                 if (ret) {
 429                         ERR("Error closing apps socket");
 430                 }
 431                 sock_info->socket = -1;
 432         }
 433         if (sock_info->root_handle != -1) {
 434                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 435                 if (ret) {
 436                         ERR("Error unref root handle");
 437                 }
 438                 sock_info->root_handle = -1;
 439         }
 440         sock_info->constructor_sem_posted = 0;
 441         /*
 442          * wait_shm_mmap is used by listener threads outside of the
 443          * ust lock, so we cannot tear it down ourselves, because we
 444          * cannot join on these threads. Leave this task to the OS
 445          * process exit.
 446          */
 447         if (!exiting && sock_info->wait_shm_mmap) {
 448                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 449                 if (ret) {
 450                         ERR("Error unmapping wait shm");
 451                 }
 452                 sock_info->wait_shm_mmap = NULL;
 453         }
 454 }
 455
 456 /*
 457  * Using fork to set umask in the child process (not multi-thread safe).
 458  * We deal with the shm_open vs ftruncate race (happening when the
 459  * sessiond owns the shm and does not let everybody modify it, to ensure
 460  * safety against shm_unlink) by simply letting the mmap fail and
 461  * retrying after a few seconds.
 462  * For global shm, everybody has rw access to it until the sessiond
 463  * starts.
 464  */
 465 static
 466 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 467 {
 468         int wait_shm_fd, ret;
 469         pid_t pid;
 470
 471         /*
 472          * Try to open read-only.
 473          */
 474         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 475         if (wait_shm_fd >= 0) {
 476                 goto end;
 477         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 478                 /*
 479                  * Real-only open did not work, and it's not because the
 480                  * entry was not present. It's a failure that prohibits
 481                  * using shm.
 482                  */
 483                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 484                 goto end;
 485         }
 486         /*
 487          * If the open failed because the file did not exist, try
 488          * creating it ourself.
 489          */
 490         lttng_ust_nest_count++;
 491         pid = fork();
 492         lttng_ust_nest_count--;
 493         if (pid > 0) {
 494                 int status;
 495
 496                 /*
 497                  * Parent: wait for child to return, in which case the
 498                  * shared memory map will have been created.
 499                  */
 500                 pid = wait(&status);
 501                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 502                         wait_shm_fd = -1;
 503                         goto end;
 504                 }
 505                 /*
 506                  * Try to open read-only again after creation.
 507                  */
 508                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 509                 if (wait_shm_fd < 0) {
 510                         /*
 511                          * Real-only open did not work. It's a failure
 512                          * that prohibits using shm.
 513                          */
 514                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 515                         goto end;
 516                 }
 517                 goto end;
 518         } else if (pid == 0) {
 519                 int create_mode;
 520
 521                 /* Child */
 522                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 523                 if (sock_info->global)
 524                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 525                 /*
 526                  * We're alone in a child process, so we can modify the
 527                  * process-wide umask.
 528                  */
 529                 umask(~create_mode);
 530                 /*
 531                  * Try creating shm (or get rw access).
 532                  * We don't do an exclusive open, because we allow other
 533                  * processes to create+ftruncate it concurrently.
 534                  */
 535                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 536                                 O_RDWR | O_CREAT, create_mode);
 537                 if (wait_shm_fd >= 0) {
 538                         ret = ftruncate(wait_shm_fd, mmap_size);
 539                         if (ret) {
 540                                 PERROR("ftruncate");
 541                                 exit(EXIT_FAILURE);
 542                         }
 543                         exit(EXIT_SUCCESS);
 544                 }
 545                 /*
 546                  * For local shm, we need to have rw access to accept
 547                  * opening it: this means the local sessiond will be
 548                  * able to wake us up. For global shm, we open it even
 549                  * if rw access is not granted, because the root.root
 550                  * sessiond will be able to override all rights and wake
 551                  * us up.
 552                  */
 553                 if (!sock_info->global && errno != EACCES) {
 554                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 555                         exit(EXIT_FAILURE);
 556                 }
 557                 /*
 558                  * The shm exists, but we cannot open it RW. Report
 559                  * success.
 560                  */
 561                 exit(EXIT_SUCCESS);
 562         } else {
 563                 return -1;
 564         }
 565 end:
 566         if (wait_shm_fd >= 0 && !sock_info->global) {
 567                 struct stat statbuf;
 568
 569                 /*
 570                  * Ensure that our user is the owner of the shm file for
 571                  * local shm. If we do not own the file, it means our
 572                  * sessiond will not have access to wake us up (there is
 573                  * probably a rogue process trying to fake our
 574                  * sessiond). Fallback to polling method in this case.
 575                  */
 576                 ret = fstat(wait_shm_fd, &statbuf);
 577                 if (ret) {
 578                         PERROR("fstat");
 579                         goto error_close;
 580                 }
 581                 if (statbuf.st_uid != getuid())
 582                         goto error_close;
 583         }
 584         return wait_shm_fd;
 585
 586 error_close:
 587         ret = close(wait_shm_fd);
 588         if (ret) {
 589                 PERROR("Error closing fd");
 590         }
 591         return -1;
 592 }
 593
 594 static
 595 char *get_map_shm(struct sock_info *sock_info)
 596 {
 597         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 598         int wait_shm_fd, ret;
 599         char *wait_shm_mmap;
 600
 601         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 602         if (wait_shm_fd < 0) {
 603                 goto error;
 604         }
 605         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 606                   MAP_SHARED, wait_shm_fd, 0);
 607         /* close shm fd immediately after taking the mmap reference */
 608         ret = close(wait_shm_fd);
 609         if (ret) {
 610                 PERROR("Error closing fd");
 611         }
 612         if (wait_shm_mmap == MAP_FAILED) {
 613                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 614                 goto error;
 615         }
 616         return wait_shm_mmap;
 617
 618 error:
 619         return NULL;
 620 }
 621
 622 static
 623 void wait_for_sessiond(struct sock_info *sock_info)
 624 {
 625         int ret;
 626
 627         ust_lock();
 628         if (lttng_ust_comm_should_quit) {
 629                 goto quit;
 630         }
 631         if (wait_poll_fallback) {
 632                 goto error;
 633         }
 634         if (!sock_info->wait_shm_mmap) {
 635                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 636                 if (!sock_info->wait_shm_mmap)
 637                         goto error;
 638         }
 639         ust_unlock();
 640
 641         DBG("Waiting for %s apps sessiond", sock_info->name);
 642         /* Wait for futex wakeup */
 643         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 644                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 645                         FUTEX_WAIT, 0, NULL, NULL, 0);
 646                 if (ret < 0) {
 647                         if (errno == EFAULT) {
 648                                 wait_poll_fallback = 1;
 649                                 DBG(
 650 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 651 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 652 "Please upgrade your kernel "
 653 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 654 "mainline). LTTng-UST will use polling mode fallback.");
 655                                 if (ust_debug())
 656                                         PERROR("futex");
 657                         }
 658                 }
 659         }
 660         return;
 661
 662 quit:
 663         ust_unlock();
 664         return;
 665
 666 error:
 667         ust_unlock();
 668         return;
 669 }
 670
 671 /*
 672  * This thread does not allocate any resource, except within
 673  * handle_message, within mutex protection. This mutex protects against
 674  * fork and exit.
 675  * The other moment it allocates resources is at socket connexion, which
 676  * is also protected by the mutex.
 677  */
 678 static
 679 void *ust_listener_thread(void *arg)
 680 {
 681         struct sock_info *sock_info = arg;
 682         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 683
 684         /* Restart trying to connect to the session daemon */
 685 restart:
 686         if (prev_connect_failed) {
 687                 /* Wait for sessiond availability with pipe */
 688                 wait_for_sessiond(sock_info);
 689                 if (has_waited) {
 690                         has_waited = 0;
 691                         /*
 692                          * Sleep for 5 seconds before retrying after a
 693                          * sequence of failure / wait / failure. This
 694                          * deals with a killed or broken session daemon.
 695                          */
 696                         sleep(5);
 697                 }
 698                 has_waited = 1;
 699                 prev_connect_failed = 0;
 700         }
 701         ust_lock();
 702
 703         if (lttng_ust_comm_should_quit) {
 704                 ust_unlock();
 705                 goto quit;
 706         }
 707
 708         if (sock_info->socket != -1) {
 709                 ret = ustcomm_close_unix_sock(sock_info->socket);
 710                 if (ret) {
 711                         ERR("Error closing %s apps socket", sock_info->name);
 712                 }
 713                 sock_info->socket = -1;
 714         }
 715
 716         /* Register */
 717         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 718         if (ret < 0) {
 719                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 720                 prev_connect_failed = 1;
 721                 /*
 722                  * If we cannot find the sessiond daemon, don't delay
 723                  * constructor execution.
 724                  */
 725                 ret = handle_register_done(sock_info);
 726                 assert(!ret);
 727                 ust_unlock();
 728                 goto restart;
 729         }
 730
 731         sock_info->socket = sock = ret;
 732
 733         /*
 734          * Create only one root handle per listener thread for the whole
 735          * process lifetime.
 736          */
 737         if (sock_info->root_handle == -1) {
 738                 ret = lttng_abi_create_root_handle();
 739                 if (ret < 0) {
 740                         ERR("Error creating root handle");
 741                         ust_unlock();
 742                         goto quit;
 743                 }
 744                 sock_info->root_handle = ret;
 745         }
 746
 747         ret = register_app_to_sessiond(sock);
 748         if (ret < 0) {
 749                 ERR("Error registering to %s apps socket", sock_info->name);
 750                 prev_connect_failed = 1;
 751                 /*
 752                  * If we cannot register to the sessiond daemon, don't
 753                  * delay constructor execution.
 754                  */
 755                 ret = handle_register_done(sock_info);
 756                 assert(!ret);
 757                 ust_unlock();
 758                 goto restart;
 759         }
 760         ust_unlock();
 761
 762         for (;;) {
 763                 ssize_t len;
 764                 struct ustcomm_ust_msg lum;
 765
 766                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 767                 switch (len) {
 768                 case 0: /* orderly shutdown */
 769                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 770                         ust_lock();
 771                         /*
 772                          * Either sessiond has shutdown or refused us by closing the socket.
 773                          * In either case, we don't want to delay construction execution,
 774                          * and we need to wait before retry.
 775                          */
 776                         prev_connect_failed = 1;
 777                         /*
 778                          * If we cannot register to the sessiond daemon, don't
 779                          * delay constructor execution.
 780                          */
 781                         ret = handle_register_done(sock_info);
 782                         assert(!ret);
 783                         ust_unlock();
 784                         goto end;
 785                 case sizeof(lum):
 786                         DBG("message received\n");
 787                         ret = handle_message(sock_info, sock, &lum);
 788                         if (ret < 0) {
 789                                 ERR("Error handling message for %s socket", sock_info->name);
 790                         }
 791                         continue;
 792                 case -1:
 793                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 794                         if (errno == ECONNRESET) {
 795                                 ERR("%s remote end closed connection\n", sock_info->name);
 796                                 goto end;
 797                         }
 798                         goto end;
 799                 default:
 800                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 801                         continue;
 802                 }
 803
 804         }
 805 end:
 806         goto restart;   /* try to reconnect */
 807 quit:
 808         return NULL;
 809 }
 810
 811 /*
 812  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 813  */
 814 static
 815 int get_timeout(struct timespec *constructor_timeout)
 816 {
 817         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 818         char *str_delay;
 819         int ret;
 820
 821         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 822         if (str_delay) {
 823                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 824         }
 825
 826         switch (constructor_delay_ms) {
 827         case -1:/* fall-through */
 828         case 0:
 829                 return constructor_delay_ms;
 830         default:
 831                 break;
 832         }
 833
 834         /*
 835          * If we are unable to find the current time, don't wait.
 836          */
 837         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 838         if (ret) {
 839                 return -1;
 840         }
 841         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 842         constructor_timeout->tv_nsec +=
 843                 (constructor_delay_ms % 1000UL) * 1000000UL;
 844         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 845                 constructor_timeout->tv_sec++;
 846                 constructor_timeout->tv_nsec -= 1000000000UL;
 847         }
 848         return 1;
 849 }
 850
 851 /*
 852  * sessiond monitoring thread: monitor presence of global and per-user
 853  * sessiond by polling the application common named pipe.
 854  */
 855 /* TODO */
 856
 857 void __attribute__((constructor)) lttng_ust_init(void)
 858 {
 859         struct timespec constructor_timeout;
 860         sigset_t sig_all_blocked, orig_parent_mask;
 861         int timeout_mode;
 862         int ret;
 863
 864         if (uatomic_xchg(&initialized, 1) == 1)
 865                 return;
 866
 867         /*
 868          * Fixup interdependency between TLS fixup mutex (which happens
 869          * to be the dynamic linker mutex) and ust_lock, taken within
 870          * the ust lock.
 871          */
 872         lttng_fixup_event_tls();
 873         lttng_fixup_ringbuffer_tls();
 874         lttng_fixup_vtid_tls();
 875
 876         /*
 877          * We want precise control over the order in which we construct
 878          * our sub-libraries vs starting to receive commands from
 879          * sessiond (otherwise leading to errors when trying to create
 880          * sessiond before the init functions are completed).
 881          */
 882         init_usterr();
 883         init_tracepoint();
 884         ltt_ring_buffer_metadata_client_init();
 885         ltt_ring_buffer_client_overwrite_init();
 886         ltt_ring_buffer_client_discard_init();
 887
 888         timeout_mode = get_timeout(&constructor_timeout);
 889
 890         ret = sem_init(&constructor_wait, 0, 0);
 891         assert(!ret);
 892
 893         ret = setup_local_apps();
 894         if (ret) {
 895                 ERR("Error setting up to local apps");
 896         }
 897
 898         /* A new thread created by pthread_create inherits the signal mask
 899          * from the parent. To avoid any signal being received by the
 900          * listener thread, we block all signals temporarily in the parent,
 901          * while we create the listener thread.
 902          */
 903         sigfillset(&sig_all_blocked);
 904         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 905         if (ret) {
 906                 ERR("pthread_sigmask: %s", strerror(ret));
 907         }
 908
 909         ret = pthread_create(&global_apps.ust_listener, NULL,
 910                         ust_listener_thread, &global_apps);
 911         if (ret) {
 912                 ERR("pthread_create global: %s", strerror(ret));
 913         }
 914         if (local_apps.allowed) {
 915                 ret = pthread_create(&local_apps.ust_listener, NULL,
 916                                 ust_listener_thread, &local_apps);
 917                 if (ret) {
 918                         ERR("pthread_create local: %s", strerror(ret));
 919                 }
 920         } else {
 921                 handle_register_done(&local_apps);
 922         }
 923
 924         /* Restore original signal mask in parent */
 925         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
 926         if (ret) {
 927                 ERR("pthread_sigmask: %s", strerror(ret));
 928         }
 929
 930         switch (timeout_mode) {
 931         case 1: /* timeout wait */
 932                 do {
 933                         ret = sem_timedwait(&constructor_wait,
 934                                         &constructor_timeout);
 935                 } while (ret < 0 && errno == EINTR);
 936                 if (ret < 0 && errno == ETIMEDOUT) {
 937                         ERR("Timed out waiting for ltt-sessiond");
 938                 } else {
 939                         assert(!ret);
 940                 }
 941                 break;
 942         case -1:/* wait forever */
 943                 do {
 944                         ret = sem_wait(&constructor_wait);
 945                 } while (ret < 0 && errno == EINTR);
 946                 assert(!ret);
 947                 break;
 948         case 0: /* no timeout */
 949                 break;
 950         }
 951 }
 952
 953 static
 954 void lttng_ust_cleanup(int exiting)
 955 {
 956         cleanup_sock_info(&global_apps, exiting);
 957         if (local_apps.allowed) {
 958                 cleanup_sock_info(&local_apps, exiting);
 959         }
 960         /*
 961          * The teardown in this function all affect data structures
 962          * accessed under the UST lock by the listener thread. This
 963          * lock, along with the lttng_ust_comm_should_quit flag, ensure
 964          * that none of these threads are accessing this data at this
 965          * point.
 966          */
 967         lttng_ust_abi_exit();
 968         lttng_ust_events_exit();
 969         ltt_ring_buffer_client_discard_exit();
 970         ltt_ring_buffer_client_overwrite_exit();
 971         ltt_ring_buffer_metadata_client_exit();
 972         exit_tracepoint();
 973         if (!exiting) {
 974                 /* Reinitialize values for fork */
 975                 sem_count = 2;
 976                 lttng_ust_comm_should_quit = 0;
 977                 initialized = 0;
 978         }
 979 }
 980
 981 void __attribute__((destructor)) lttng_ust_exit(void)
 982 {
 983         int ret;
 984
 985         /*
 986          * Using pthread_cancel here because:
 987          * A) we don't want to hang application teardown.
 988          * B) the thread is not allocating any resource.
 989          */
 990
 991         /*
 992          * Require the communication thread to quit. Synchronize with
 993          * mutexes to ensure it is not in a mutex critical section when
 994          * pthread_cancel is later called.
 995          */
 996         ust_lock();
 997         lttng_ust_comm_should_quit = 1;
 998         ust_unlock();
 999
1000         /* cancel threads */
1001         ret = pthread_cancel(global_apps.ust_listener);
1002         if (ret) {
1003                 ERR("Error cancelling global ust listener thread: %s",
1004                         strerror(ret));
1005         }
1006         if (local_apps.allowed) {
1007                 ret = pthread_cancel(local_apps.ust_listener);
1008                 if (ret) {
1009                         ERR("Error cancelling local ust listener thread: %s",
1010                                 strerror(ret));
1011                 }
1012         }
1013         /*
1014          * Do NOT join threads: use of sys_futex makes it impossible to
1015          * join the threads without using async-cancel, but async-cancel
1016          * is delivered by a signal, which could hit the target thread
1017          * anywhere in its code path, including while the ust_lock() is
1018          * held, causing a deadlock for the other thread. Let the OS
1019          * cleanup the threads if there are stalled in a syscall.
1020          */
1021         lttng_ust_cleanup(1);
1022 }
1023
1024 /*
1025  * We exclude the worker threads across fork and clone (except
1026  * CLONE_VM), because these system calls only keep the forking thread
1027  * running in the child.  Therefore, we don't want to call fork or clone
1028  * in the middle of an tracepoint or ust tracing state modification.
1029  * Holding this mutex protects these structures across fork and clone.
1030  */
1031 void ust_before_fork(sigset_t *save_sigset)
1032 {
1033         /*
1034          * Disable signals. This is to avoid that the child intervenes
1035          * before it is properly setup for tracing. It is safer to
1036          * disable all signals, because then we know we are not breaking
1037          * anything by restoring the original mask.
1038          */
1039         sigset_t all_sigs;
1040         int ret;
1041
1042         if (lttng_ust_nest_count)
1043                 return;
1044         /* Disable signals */
1045         sigfillset(&all_sigs);
1046         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1047         if (ret == -1) {
1048                 PERROR("sigprocmask");
1049         }
1050         ust_lock();
1051         rcu_bp_before_fork();
1052 }
1053
1054 static void ust_after_fork_common(sigset_t *restore_sigset)
1055 {
1056         int ret;
1057
1058         DBG("process %d", getpid());
1059         ust_unlock();
1060         /* Restore signals */
1061         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1062         if (ret == -1) {
1063                 PERROR("sigprocmask");
1064         }
1065 }
1066
1067 void ust_after_fork_parent(sigset_t *restore_sigset)
1068 {
1069         if (lttng_ust_nest_count)
1070                 return;
1071         DBG("process %d", getpid());
1072         rcu_bp_after_fork_parent();
1073         /* Release mutexes and reenable signals */
1074         ust_after_fork_common(restore_sigset);
1075 }
1076
1077 /*
1078  * After fork, in the child, we need to cleanup all the leftover state,
1079  * except the worker thread which already magically disappeared thanks
1080  * to the weird Linux fork semantics. After tyding up, we call
1081  * lttng_ust_init() again to start over as a new PID.
1082  *
1083  * This is meant for forks() that have tracing in the child between the
1084  * fork and following exec call (if there is any).
1085  */
1086 void ust_after_fork_child(sigset_t *restore_sigset)
1087 {
1088         if (lttng_ust_nest_count)
1089                 return;
1090         DBG("process %d", getpid());
1091         /* Release urcu mutexes */
1092         rcu_bp_after_fork_child();
1093         lttng_ust_cleanup(0);
1094         lttng_context_vtid_reset();
1095         /* Release mutexes and reenable signals */
1096         ust_after_fork_common(restore_sigset);
1097         lttng_ust_init();
1098 }