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 /*
 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                 assert(local_apps.allowed == 0);
 160                 return 0;
 161         }
 162         home_dir = (const char *) getenv("HOME");
 163         if (!home_dir) {
 164                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 165                 assert(local_apps.allowed == 0);
 166                 return -ENOENT;
 167         }
 168         local_apps.allowed = 1;
 169         snprintf(local_apps.sock_path, PATH_MAX,
 170                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 171         snprintf(local_apps.wait_shm_path, PATH_MAX,
 172                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 173         return 0;
 174 }
 175
 176 static
 177 int register_app_to_sessiond(int socket)
 178 {
 179         ssize_t ret;
 180         struct {
 181                 uint32_t major;
 182                 uint32_t minor;
 183                 pid_t pid;
 184                 pid_t ppid;
 185                 uid_t uid;
 186                 gid_t gid;
 187                 uint32_t bits_per_long;
 188                 char name[16];  /* process name */
 189         } reg_msg;
 190
 191         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 192         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 193         reg_msg.pid = getpid();
 194         reg_msg.ppid = getppid();
 195         reg_msg.uid = getuid();
 196         reg_msg.gid = getgid();
 197         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 198         lttng_ust_getprocname(reg_msg.name);
 199
 200         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 201         if (ret >= 0 && ret != sizeof(reg_msg))
 202                 return -EIO;
 203         return ret;
 204 }
 205
 206 static
 207 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 208 {
 209         ssize_t len;
 210
 211         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 212         switch (len) {
 213         case sizeof(*lur):
 214                 DBG("message successfully sent");
 215                 return 0;
 216         case -1:
 217                 if (errno == ECONNRESET) {
 218                         printf("remote end closed connection\n");
 219                         return 0;
 220                 }
 221                 return -1;
 222         default:
 223                 printf("incorrect message size: %zd\n", len);
 224                 return -1;
 225         }
 226 }
 227
 228 static
 229 int handle_register_done(struct sock_info *sock_info)
 230 {
 231         int ret;
 232
 233         if (sock_info->constructor_sem_posted)
 234                 return 0;
 235         sock_info->constructor_sem_posted = 1;
 236         if (uatomic_read(&sem_count) <= 0) {
 237                 return 0;
 238         }
 239         ret = uatomic_add_return(&sem_count, -1);
 240         if (ret == 0) {
 241                 ret = sem_post(&constructor_wait);
 242                 assert(!ret);
 243         }
 244         return 0;
 245 }
 246
 247 static
 248 int handle_message(struct sock_info *sock_info,
 249                 int sock, struct ustcomm_ust_msg *lum)
 250 {
 251         int ret = 0;
 252         const struct lttng_ust_objd_ops *ops;
 253         struct ustcomm_ust_reply lur;
 254         int shm_fd, wait_fd;
 255         char *shm_path, *wait_pipe_path;
 256         union ust_args args;
 257
 258         ust_lock();
 259
 260         memset(&lur, 0, sizeof(lur));
 261
 262         if (lttng_ust_comm_should_quit) {
 263                 ret = -EPERM;
 264                 goto end;
 265         }
 266
 267         ops = objd_ops(lum->handle);
 268         if (!ops) {
 269                 ret = -ENOENT;
 270                 goto end;
 271         }
 272
 273         switch (lum->cmd) {
 274         case LTTNG_UST_REGISTER_DONE:
 275                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 276                         ret = handle_register_done(sock_info);
 277                 else
 278                         ret = -EINVAL;
 279                 break;
 280         case LTTNG_UST_RELEASE:
 281                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 282                         ret = -EPERM;
 283                 else
 284                         ret = lttng_ust_objd_unref(lum->handle);
 285                 break;
 286         default:
 287                 if (ops->cmd)
 288                         ret = ops->cmd(lum->handle, lum->cmd,
 289                                         (unsigned long) &lum->u,
 290                                         &args);
 291                 else
 292                         ret = -ENOSYS;
 293                 break;
 294         }
 295
 296 end:
 297         lur.handle = lum->handle;
 298         lur.cmd = lum->cmd;
 299         lur.ret_val = ret;
 300         if (ret >= 0) {
 301                 lur.ret_code = USTCOMM_OK;
 302         } else {
 303                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 304                 lur.ret_code = ret;
 305         }
 306         if (ret >= 0) {
 307                 switch (lum->cmd) {
 308                 case LTTNG_UST_STREAM:
 309                         /*
 310                          * Special-case reply to send stream info.
 311                          * Use lum.u output.
 312                          */
 313                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 314                         shm_fd         = *args.stream.shm_fd;
 315                         shm_path       = args.stream.shm_path;
 316                         wait_fd        = *args.stream.wait_fd;
 317                         wait_pipe_path = args.stream.wait_pipe_path;
 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                         shm_path       = args.channel.shm_path;
 324                         wait_fd        = *args.channel.wait_fd;
 325                         wait_pipe_path = args.channel.wait_pipe_path;
 326                         break;
 327                 case LTTNG_UST_TRACER_VERSION:
 328                         lur.u.version = lum->u.version;
 329                         break;
 330                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 331                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 332                         break;
 333                 }
 334         }
 335         ret = send_reply(sock, &lur);
 336         if (ret < 0) {
 337                 perror("error sending reply");
 338                 goto error;
 339         }
 340
 341         if ((lum->cmd == LTTNG_UST_STREAM
 342              || lum->cmd == LTTNG_UST_CHANNEL
 343              || lum->cmd == LTTNG_UST_METADATA)
 344                         && lur.ret_code == USTCOMM_OK) {
 345                 int sendret = 0;
 346
 347                 /* send the shm path */
 348                 ret = ustcomm_send_string(sock, shm_path, strlen(shm_path));
 349                 if (ret < 0) {
 350                         perror("send shm_path");
 351                         sendret = ret;
 352                 }
 353                 /*
 354                  * The sessiond expects 2 file descriptors, even upon
 355                  * error.
 356                  */
 357                 ret = ustcomm_send_string(sock, wait_pipe_path, strlen(wait_pipe_path));
 358                 if (ret < 0) {
 359                         perror("send wait_pipe_path");
 360                         goto error;
 361                 }
 362                 if (sendret) {
 363                         ret = sendret;
 364                         goto error;
 365                 }
 366         }
 367         /*
 368          * We still have the memory map reference, and the fds have been
 369          * sent to the sessiond. We can therefore close those fds. Note
 370          * that we keep the write side of the wait_fd open, but close
 371          * the read side.
 372          */
 373         if (lur.ret_code == USTCOMM_OK) {
 374                 switch (lum->cmd) {
 375                 case LTTNG_UST_STREAM:
 376                         if (shm_fd >= 0) {
 377                                 ret = close(shm_fd);
 378                                 if (ret) {
 379                                         PERROR("Error closing stream shm_fd");
 380                                 }
 381                                 *args.stream.shm_fd = -1;
 382                         }
 383                         if (wait_fd >= 0) {
 384                                 ret = close(wait_fd);
 385                                 if (ret) {
 386                                         PERROR("Error closing stream wait_fd");
 387                                 }
 388                                 *args.stream.wait_fd = -1;
 389                         }
 390                         break;
 391                 case LTTNG_UST_METADATA:
 392                 case LTTNG_UST_CHANNEL:
 393                         if (shm_fd >= 0) {
 394                                 ret = close(shm_fd);
 395                                 if (ret) {
 396                                         PERROR("Error closing channel shm_fd");
 397                                 }
 398                                 *args.channel.shm_fd = -1;
 399                         }
 400                         if (wait_fd >= 0) {
 401                                 ret = close(wait_fd);
 402                                 if (ret) {
 403                                         PERROR("Error closing channel wait_fd");
 404                                 }
 405                                 *args.channel.wait_fd = -1;
 406                         }
 407                         break;
 408                 }
 409         }
 410
 411 error:
 412         ust_unlock();
 413         return ret;
 414 }
 415
 416 static
 417 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 418 {
 419         int ret;
 420
 421         if (sock_info->socket != -1) {
 422                 ret = ustcomm_close_unix_sock(sock_info->socket);
 423                 if (ret) {
 424                         ERR("Error closing apps socket");
 425                 }
 426                 sock_info->socket = -1;
 427         }
 428         if (sock_info->root_handle != -1) {
 429                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 430                 if (ret) {
 431                         ERR("Error unref root handle");
 432                 }
 433                 sock_info->root_handle = -1;
 434         }
 435         sock_info->constructor_sem_posted = 0;
 436         /*
 437          * wait_shm_mmap is used by listener threads outside of the
 438          * ust lock, so we cannot tear it down ourselves, because we
 439          * cannot join on these threads. Leave this task to the OS
 440          * process exit.
 441          */
 442         if (!exiting && sock_info->wait_shm_mmap) {
 443                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 444                 if (ret) {
 445                         ERR("Error unmapping wait shm");
 446                 }
 447                 sock_info->wait_shm_mmap = NULL;
 448         }
 449 }
 450
 451 /*
 452  * Using fork to set umask in the child process (not multi-thread safe).
 453  * We deal with the shm_open vs ftruncate race (happening when the
 454  * sessiond owns the shm and does not let everybody modify it, to ensure
 455  * safety against shm_unlink) by simply letting the mmap fail and
 456  * retrying after a few seconds.
 457  * For global shm, everybody has rw access to it until the sessiond
 458  * starts.
 459  */
 460 static
 461 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 462 {
 463         int wait_shm_fd, ret;
 464         pid_t pid;
 465
 466         /*
 467          * Try to open read-only.
 468          */
 469         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 470         if (wait_shm_fd >= 0) {
 471                 goto end;
 472         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 473                 /*
 474                  * Real-only open did not work, and it's not because the
 475                  * entry was not present. It's a failure that prohibits
 476                  * using shm.
 477                  */
 478                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 479                 goto end;
 480         }
 481         /*
 482          * If the open failed because the file did not exist, try
 483          * creating it ourself.
 484          */
 485         lttng_ust_nest_count++;
 486         pid = fork();
 487         lttng_ust_nest_count--;
 488         if (pid > 0) {
 489                 int status;
 490
 491                 /*
 492                  * Parent: wait for child to return, in which case the
 493                  * shared memory map will have been created.
 494                  */
 495                 pid = wait(&status);
 496                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 497                         wait_shm_fd = -1;
 498                         goto end;
 499                 }
 500                 /*
 501                  * Try to open read-only again after creation.
 502                  */
 503                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 504                 if (wait_shm_fd < 0) {
 505                         /*
 506                          * Real-only open did not work. It's a failure
 507                          * that prohibits using shm.
 508                          */
 509                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 510                         goto end;
 511                 }
 512                 goto end;
 513         } else if (pid == 0) {
 514                 int create_mode;
 515
 516                 /* Child */
 517                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 518                 if (sock_info->global)
 519                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 520                 /*
 521                  * We're alone in a child process, so we can modify the
 522                  * process-wide umask.
 523                  */
 524                 umask(~create_mode);
 525                 /*
 526                  * Try creating shm (or get rw access).
 527                  * We don't do an exclusive open, because we allow other
 528                  * processes to create+ftruncate it concurrently.
 529                  */
 530                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 531                                 O_RDWR | O_CREAT, create_mode);
 532                 if (wait_shm_fd >= 0) {
 533                         ret = ftruncate(wait_shm_fd, mmap_size);
 534                         if (ret) {
 535                                 PERROR("ftruncate");
 536                                 _exit(EXIT_FAILURE);
 537                         }
 538                         _exit(EXIT_SUCCESS);
 539                 }
 540                 /*
 541                  * For local shm, we need to have rw access to accept
 542                  * opening it: this means the local sessiond will be
 543                  * able to wake us up. For global shm, we open it even
 544                  * if rw access is not granted, because the root.root
 545                  * sessiond will be able to override all rights and wake
 546                  * us up.
 547                  */
 548                 if (!sock_info->global && errno != EACCES) {
 549                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 550                         _exit(EXIT_FAILURE);
 551                 }
 552                 /*
 553                  * The shm exists, but we cannot open it RW. Report
 554                  * success.
 555                  */
 556                 _exit(EXIT_SUCCESS);
 557         } else {
 558                 return -1;
 559         }
 560 end:
 561         if (wait_shm_fd >= 0 && !sock_info->global) {
 562                 struct stat statbuf;
 563
 564                 /*
 565                  * Ensure that our user is the owner of the shm file for
 566                  * local shm. If we do not own the file, it means our
 567                  * sessiond will not have access to wake us up (there is
 568                  * probably a rogue process trying to fake our
 569                  * sessiond). Fallback to polling method in this case.
 570                  */
 571                 ret = fstat(wait_shm_fd, &statbuf);
 572                 if (ret) {
 573                         PERROR("fstat");
 574                         goto error_close;
 575                 }
 576                 if (statbuf.st_uid != getuid())
 577                         goto error_close;
 578         }
 579         return wait_shm_fd;
 580
 581 error_close:
 582         ret = close(wait_shm_fd);
 583         if (ret) {
 584                 PERROR("Error closing fd");
 585         }
 586         return -1;
 587 }
 588
 589 static
 590 char *get_map_shm(struct sock_info *sock_info)
 591 {
 592         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 593         int wait_shm_fd, ret;
 594         char *wait_shm_mmap;
 595
 596         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 597         if (wait_shm_fd < 0) {
 598                 goto error;
 599         }
 600         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 601                   MAP_SHARED, wait_shm_fd, 0);
 602         /* close shm fd immediately after taking the mmap reference */
 603         ret = close(wait_shm_fd);
 604         if (ret) {
 605                 PERROR("Error closing fd");
 606         }
 607         if (wait_shm_mmap == MAP_FAILED) {
 608                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 609                 goto error;
 610         }
 611         return wait_shm_mmap;
 612
 613 error:
 614         return NULL;
 615 }
 616
 617 static
 618 void wait_for_sessiond(struct sock_info *sock_info)
 619 {
 620         int ret;
 621
 622         ust_lock();
 623         if (lttng_ust_comm_should_quit) {
 624                 goto quit;
 625         }
 626         if (wait_poll_fallback) {
 627                 goto error;
 628         }
 629         if (!sock_info->wait_shm_mmap) {
 630                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 631                 if (!sock_info->wait_shm_mmap)
 632                         goto error;
 633         }
 634         ust_unlock();
 635
 636         DBG("Waiting for %s apps sessiond", sock_info->name);
 637         /* Wait for futex wakeup */
 638         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 639                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 640                         FUTEX_WAIT, 0, NULL, NULL, 0);
 641                 if (ret < 0) {
 642                         if (errno == EFAULT) {
 643                                 wait_poll_fallback = 1;
 644                                 DBG(
 645 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 646 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 647 "Please upgrade your kernel "
 648 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 649 "mainline). LTTng-UST will use polling mode fallback.");
 650                                 if (ust_debug())
 651                                         PERROR("futex");
 652                         }
 653                 }
 654         }
 655         return;
 656
 657 quit:
 658         ust_unlock();
 659         return;
 660
 661 error:
 662         ust_unlock();
 663         return;
 664 }
 665
 666 /*
 667  * This thread does not allocate any resource, except within
 668  * handle_message, within mutex protection. This mutex protects against
 669  * fork and exit.
 670  * The other moment it allocates resources is at socket connexion, which
 671  * is also protected by the mutex.
 672  */
 673 static
 674 void *ust_listener_thread(void *arg)
 675 {
 676         struct sock_info *sock_info = arg;
 677         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 678
 679         /* Restart trying to connect to the session daemon */
 680 restart:
 681         if (prev_connect_failed) {
 682                 /* Wait for sessiond availability with pipe */
 683                 wait_for_sessiond(sock_info);
 684                 if (has_waited) {
 685                         has_waited = 0;
 686                         /*
 687                          * Sleep for 5 seconds before retrying after a
 688                          * sequence of failure / wait / failure. This
 689                          * deals with a killed or broken session daemon.
 690                          */
 691                         sleep(5);
 692                 }
 693                 has_waited = 1;
 694                 prev_connect_failed = 0;
 695         }
 696         ust_lock();
 697
 698         if (lttng_ust_comm_should_quit) {
 699                 ust_unlock();
 700                 goto quit;
 701         }
 702
 703         if (sock_info->socket != -1) {
 704                 ret = ustcomm_close_unix_sock(sock_info->socket);
 705                 if (ret) {
 706                         ERR("Error closing %s apps socket", sock_info->name);
 707                 }
 708                 sock_info->socket = -1;
 709         }
 710
 711         /* Register */
 712         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 713         if (ret < 0) {
 714                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 715                 prev_connect_failed = 1;
 716                 /*
 717                  * If we cannot find the sessiond daemon, don't delay
 718                  * constructor execution.
 719                  */
 720                 ret = handle_register_done(sock_info);
 721                 assert(!ret);
 722                 ust_unlock();
 723                 goto restart;
 724         }
 725
 726         sock_info->socket = sock = ret;
 727
 728         /*
 729          * Create only one root handle per listener thread for the whole
 730          * process lifetime.
 731          */
 732         if (sock_info->root_handle == -1) {
 733                 ret = lttng_abi_create_root_handle();
 734                 if (ret < 0) {
 735                         ERR("Error creating root handle");
 736                         ust_unlock();
 737                         goto quit;
 738                 }
 739                 sock_info->root_handle = ret;
 740         }
 741
 742         ret = register_app_to_sessiond(sock);
 743         if (ret < 0) {
 744                 ERR("Error registering to %s apps socket", sock_info->name);
 745                 prev_connect_failed = 1;
 746                 /*
 747                  * If we cannot register to the sessiond daemon, don't
 748                  * delay constructor execution.
 749                  */
 750                 ret = handle_register_done(sock_info);
 751                 assert(!ret);
 752                 ust_unlock();
 753                 goto restart;
 754         }
 755         ust_unlock();
 756
 757         for (;;) {
 758                 ssize_t len;
 759                 struct ustcomm_ust_msg lum;
 760
 761                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 762                 switch (len) {
 763                 case 0: /* orderly shutdown */
 764                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 765                         ust_lock();
 766                         /*
 767                          * Either sessiond has shutdown or refused us by closing the socket.
 768                          * In either case, we don't want to delay construction execution,
 769                          * and we need to wait before retry.
 770                          */
 771                         prev_connect_failed = 1;
 772                         /*
 773                          * If we cannot register to the sessiond daemon, don't
 774                          * delay constructor execution.
 775                          */
 776                         ret = handle_register_done(sock_info);
 777                         assert(!ret);
 778                         ust_unlock();
 779                         goto end;
 780                 case sizeof(lum):
 781                         DBG("message received\n");
 782                         ret = handle_message(sock_info, sock, &lum);
 783                         if (ret < 0) {
 784                                 ERR("Error handling message for %s socket", sock_info->name);
 785                         }
 786                         continue;
 787                 case -1:
 788                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 789                         if (errno == ECONNRESET) {
 790                                 ERR("%s remote end closed connection\n", sock_info->name);
 791                                 goto end;
 792                         }
 793                         goto end;
 794                 default:
 795                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 796                         continue;
 797                 }
 798
 799         }
 800 end:
 801         goto restart;   /* try to reconnect */
 802 quit:
 803         return NULL;
 804 }
 805
 806 /*
 807  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 808  */
 809 static
 810 int get_timeout(struct timespec *constructor_timeout)
 811 {
 812         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 813         char *str_delay;
 814         int ret;
 815
 816         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 817         if (str_delay) {
 818                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 819         }
 820
 821         switch (constructor_delay_ms) {
 822         case -1:/* fall-through */
 823         case 0:
 824                 return constructor_delay_ms;
 825         default:
 826                 break;
 827         }
 828
 829         /*
 830          * If we are unable to find the current time, don't wait.
 831          */
 832         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 833         if (ret) {
 834                 return -1;
 835         }
 836         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 837         constructor_timeout->tv_nsec +=
 838                 (constructor_delay_ms % 1000UL) * 1000000UL;
 839         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 840                 constructor_timeout->tv_sec++;
 841                 constructor_timeout->tv_nsec -= 1000000000UL;
 842         }
 843         return 1;
 844 }
 845
 846 /*
 847  * sessiond monitoring thread: monitor presence of global and per-user
 848  * sessiond by polling the application common named pipe.
 849  */
 850 /* TODO */
 851
 852 void __attribute__((constructor)) lttng_ust_init(void)
 853 {
 854         struct timespec constructor_timeout;
 855         sigset_t sig_all_blocked, orig_parent_mask;
 856         pthread_attr_t thread_attr;
 857         int timeout_mode;
 858         int ret;
 859
 860         if (uatomic_xchg(&initialized, 1) == 1)
 861                 return;
 862
 863         /*
 864          * Fixup interdependency between TLS fixup mutex (which happens
 865          * to be the dynamic linker mutex) and ust_lock, taken within
 866          * the ust lock.
 867          */
 868         lttng_fixup_event_tls();
 869         lttng_fixup_ringbuffer_tls();
 870         lttng_fixup_vtid_tls();
 871         lttng_fixup_nest_count_tls();
 872
 873         /*
 874          * We want precise control over the order in which we construct
 875          * our sub-libraries vs starting to receive commands from
 876          * sessiond (otherwise leading to errors when trying to create
 877          * sessiond before the init functions are completed).
 878          */
 879         init_usterr();
 880         init_tracepoint();
 881         ltt_ring_buffer_metadata_client_init();
 882         ltt_ring_buffer_client_overwrite_init();
 883         ltt_ring_buffer_client_discard_init();
 884
 885         timeout_mode = get_timeout(&constructor_timeout);
 886
 887         ret = sem_init(&constructor_wait, 0, 0);
 888         assert(!ret);
 889
 890         ret = setup_local_apps();
 891         if (ret) {
 892                 DBG("local apps setup returned %d", ret);
 893         }
 894
 895         /* A new thread created by pthread_create inherits the signal mask
 896          * from the parent. To avoid any signal being received by the
 897          * listener thread, we block all signals temporarily in the parent,
 898          * while we create the listener thread.
 899          */
 900         sigfillset(&sig_all_blocked);
 901         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 902         if (ret) {
 903                 ERR("pthread_sigmask: %s", strerror(ret));
 904         }
 905
 906         ret = pthread_attr_init(&thread_attr);
 907         if (ret) {
 908                 ERR("pthread_attr_init: %s", strerror(ret));
 909         }
 910         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
 911         if (ret) {
 912                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
 913         }
 914
 915         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
 916                         ust_listener_thread, &global_apps);
 917         if (ret) {
 918                 ERR("pthread_create global: %s", strerror(ret));
 919         }
 920         if (local_apps.allowed) {
 921                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
 922                                 ust_listener_thread, &local_apps);
 923                 if (ret) {
 924                         ERR("pthread_create local: %s", strerror(ret));
 925                 }
 926         } else {
 927                 handle_register_done(&local_apps);
 928         }
 929         ret = pthread_attr_destroy(&thread_attr);
 930         if (ret) {
 931                 ERR("pthread_attr_destroy: %s", strerror(ret));
 932         }
 933
 934         /* Restore original signal mask in parent */
 935         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
 936         if (ret) {
 937                 ERR("pthread_sigmask: %s", strerror(ret));
 938         }
 939
 940         switch (timeout_mode) {
 941         case 1: /* timeout wait */
 942                 do {
 943                         ret = sem_timedwait(&constructor_wait,
 944                                         &constructor_timeout);
 945                 } while (ret < 0 && errno == EINTR);
 946                 if (ret < 0 && errno == ETIMEDOUT) {
 947                         ERR("Timed out waiting for ltt-sessiond");
 948                 } else {
 949                         assert(!ret);
 950                 }
 951                 break;
 952         case -1:/* wait forever */
 953                 do {
 954                         ret = sem_wait(&constructor_wait);
 955                 } while (ret < 0 && errno == EINTR);
 956                 assert(!ret);
 957                 break;
 958         case 0: /* no timeout */
 959                 break;
 960         }
 961 }
 962
 963 static
 964 void lttng_ust_cleanup(int exiting)
 965 {
 966         cleanup_sock_info(&global_apps, exiting);
 967         if (local_apps.allowed) {
 968                 cleanup_sock_info(&local_apps, exiting);
 969         }
 970         /*
 971          * The teardown in this function all affect data structures
 972          * accessed under the UST lock by the listener thread. This
 973          * lock, along with the lttng_ust_comm_should_quit flag, ensure
 974          * that none of these threads are accessing this data at this
 975          * point.
 976          */
 977         lttng_ust_abi_exit();
 978         lttng_ust_events_exit();
 979         ltt_ring_buffer_client_discard_exit();
 980         ltt_ring_buffer_client_overwrite_exit();
 981         ltt_ring_buffer_metadata_client_exit();
 982         exit_tracepoint();
 983         if (!exiting) {
 984                 /* Reinitialize values for fork */
 985                 sem_count = 2;
 986                 lttng_ust_comm_should_quit = 0;
 987                 initialized = 0;
 988         }
 989 }
 990
 991 void __attribute__((destructor)) lttng_ust_exit(void)
 992 {
 993         int ret;
 994
 995         /*
 996          * Using pthread_cancel here because:
 997          * A) we don't want to hang application teardown.
 998          * B) the thread is not allocating any resource.
 999          */
1000
1001         /*
1002          * Require the communication thread to quit. Synchronize with
1003          * mutexes to ensure it is not in a mutex critical section when
1004          * pthread_cancel is later called.
1005          */
1006         ust_lock();
1007         lttng_ust_comm_should_quit = 1;
1008         ust_unlock();
1009
1010         /* cancel threads */
1011         ret = pthread_cancel(global_apps.ust_listener);
1012         if (ret) {
1013                 ERR("Error cancelling global ust listener thread: %s",
1014                         strerror(ret));
1015         }
1016         if (local_apps.allowed) {
1017                 ret = pthread_cancel(local_apps.ust_listener);
1018                 if (ret) {
1019                         ERR("Error cancelling local ust listener thread: %s",
1020                                 strerror(ret));
1021                 }
1022         }
1023         /*
1024          * Do NOT join threads: use of sys_futex makes it impossible to
1025          * join the threads without using async-cancel, but async-cancel
1026          * is delivered by a signal, which could hit the target thread
1027          * anywhere in its code path, including while the ust_lock() is
1028          * held, causing a deadlock for the other thread. Let the OS
1029          * cleanup the threads if there are stalled in a syscall.
1030          */
1031         lttng_ust_cleanup(1);
1032 }
1033
1034 /*
1035  * We exclude the worker threads across fork and clone (except
1036  * CLONE_VM), because these system calls only keep the forking thread
1037  * running in the child.  Therefore, we don't want to call fork or clone
1038  * in the middle of an tracepoint or ust tracing state modification.
1039  * Holding this mutex protects these structures across fork and clone.
1040  */
1041 void ust_before_fork(sigset_t *save_sigset)
1042 {
1043         /*
1044          * Disable signals. This is to avoid that the child intervenes
1045          * before it is properly setup for tracing. It is safer to
1046          * disable all signals, because then we know we are not breaking
1047          * anything by restoring the original mask.
1048          */
1049         sigset_t all_sigs;
1050         int ret;
1051
1052         if (lttng_ust_nest_count)
1053                 return;
1054         /* Disable signals */
1055         sigfillset(&all_sigs);
1056         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1057         if (ret == -1) {
1058                 PERROR("sigprocmask");
1059         }
1060         ust_lock();
1061         rcu_bp_before_fork();
1062 }
1063
1064 static void ust_after_fork_common(sigset_t *restore_sigset)
1065 {
1066         int ret;
1067
1068         DBG("process %d", getpid());
1069         ust_unlock();
1070         /* Restore signals */
1071         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1072         if (ret == -1) {
1073                 PERROR("sigprocmask");
1074         }
1075 }
1076
1077 void ust_after_fork_parent(sigset_t *restore_sigset)
1078 {
1079         if (lttng_ust_nest_count)
1080                 return;
1081         DBG("process %d", getpid());
1082         rcu_bp_after_fork_parent();
1083         /* Release mutexes and reenable signals */
1084         ust_after_fork_common(restore_sigset);
1085 }
1086
1087 /*
1088  * After fork, in the child, we need to cleanup all the leftover state,
1089  * except the worker thread which already magically disappeared thanks
1090  * to the weird Linux fork semantics. After tyding up, we call
1091  * lttng_ust_init() again to start over as a new PID.
1092  *
1093  * This is meant for forks() that have tracing in the child between the
1094  * fork and following exec call (if there is any).
1095  */
1096 void ust_after_fork_child(sigset_t *restore_sigset)
1097 {
1098         if (lttng_ust_nest_count)
1099                 return;
1100         DBG("process %d", getpid());
1101         /* Release urcu mutexes */
1102         rcu_bp_after_fork_child();
1103         lttng_ust_cleanup(0);
1104         lttng_context_vtid_reset();
1105         /* Release mutexes and reenable signals */
1106         ust_after_fork_common(restore_sigset);
1107         lttng_ust_init();
1108 }