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