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                 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         int prctl_ret;
 181         struct {
 182                 uint32_t major;
 183                 uint32_t minor;
 184                 pid_t pid;
 185                 pid_t ppid;
 186                 uid_t uid;
 187                 gid_t gid;
 188                 uint32_t bits_per_long;
 189                 char name[16];  /* process name */
 190         } reg_msg;
 191
 192         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 193         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 194         reg_msg.pid = getpid();
 195         reg_msg.ppid = getppid();
 196         reg_msg.uid = getuid();
 197         reg_msg.gid = getgid();
 198         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 199         prctl_ret = prctl(PR_GET_NAME, (unsigned long) reg_msg.name, 0, 0, 0);
 200         if (prctl_ret) {
 201                 ERR("Error executing prctl");
 202                 return -errno;
 203         }
 204
 205         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 206         if (ret >= 0 && ret != sizeof(reg_msg))
 207                 return -EIO;
 208         return ret;
 209 }
 210
 211 static
 212 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 213 {
 214         ssize_t len;
 215
 216         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 217         switch (len) {
 218         case sizeof(*lur):
 219                 DBG("message successfully sent");
 220                 return 0;
 221         case -1:
 222                 if (errno == ECONNRESET) {
 223                         printf("remote end closed connection\n");
 224                         return 0;
 225                 }
 226                 return -1;
 227         default:
 228                 printf("incorrect message size: %zd\n", len);
 229                 return -1;
 230         }
 231 }
 232
 233 static
 234 int handle_register_done(struct sock_info *sock_info)
 235 {
 236         int ret;
 237
 238         if (sock_info->constructor_sem_posted)
 239                 return 0;
 240         sock_info->constructor_sem_posted = 1;
 241         if (uatomic_read(&sem_count) <= 0) {
 242                 return 0;
 243         }
 244         ret = uatomic_add_return(&sem_count, -1);
 245         if (ret == 0) {
 246                 ret = sem_post(&constructor_wait);
 247                 assert(!ret);
 248         }
 249         return 0;
 250 }
 251
 252 static
 253 int handle_message(struct sock_info *sock_info,
 254                 int sock, struct ustcomm_ust_msg *lum)
 255 {
 256         int ret = 0;
 257         const struct lttng_ust_objd_ops *ops;
 258         struct ustcomm_ust_reply lur;
 259         int shm_fd, wait_fd;
 260         union ust_args args;
 261
 262         ust_lock();
 263
 264         memset(&lur, 0, sizeof(lur));
 265
 266         if (lttng_ust_comm_should_quit) {
 267                 ret = -EPERM;
 268                 goto end;
 269         }
 270
 271         ops = objd_ops(lum->handle);
 272         if (!ops) {
 273                 ret = -ENOENT;
 274                 goto end;
 275         }
 276
 277         switch (lum->cmd) {
 278         case LTTNG_UST_REGISTER_DONE:
 279                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 280                         ret = handle_register_done(sock_info);
 281                 else
 282                         ret = -EINVAL;
 283                 break;
 284         case LTTNG_UST_RELEASE:
 285                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 286                         ret = -EPERM;
 287                 else
 288                         ret = lttng_ust_objd_unref(lum->handle);
 289                 break;
 290         default:
 291                 if (ops->cmd)
 292                         ret = ops->cmd(lum->handle, lum->cmd,
 293                                         (unsigned long) &lum->u,
 294                                         &args);
 295                 else
 296                         ret = -ENOSYS;
 297                 break;
 298         }
 299
 300 end:
 301         lur.handle = lum->handle;
 302         lur.cmd = lum->cmd;
 303         lur.ret_val = ret;
 304         if (ret >= 0) {
 305                 lur.ret_code = USTCOMM_OK;
 306         } else {
 307                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 308                 lur.ret_code = ret;
 309         }
 310         if (ret >= 0) {
 311                 switch (lum->cmd) {
 312                 case LTTNG_UST_STREAM:
 313                         /*
 314                          * Special-case reply to send stream info.
 315                          * Use lum.u output.
 316                          */
 317                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 318                         shm_fd = *args.stream.shm_fd;
 319                         wait_fd = *args.stream.wait_fd;
 320                         break;
 321                 case LTTNG_UST_METADATA:
 322                 case LTTNG_UST_CHANNEL:
 323                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 324                         shm_fd = *args.channel.shm_fd;
 325                         wait_fd = *args.channel.wait_fd;
 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                 /* we also need to send the file descriptors. */
 348                 ret = ustcomm_send_fds_unix_sock(sock,
 349                         &shm_fd, &shm_fd,
 350                         1, sizeof(int));
 351                 if (ret < 0) {
 352                         perror("send shm_fd");
 353                         sendret = ret;
 354                 }
 355                 /*
 356                  * The sessiond expects 2 file descriptors, even upon
 357                  * error.
 358                  */
 359                 ret = ustcomm_send_fds_unix_sock(sock,
 360                         &wait_fd, &wait_fd,
 361                         1, sizeof(int));
 362                 if (ret < 0) {
 363                         perror("send wait_fd");
 364                         goto error;
 365                 }
 366                 if (sendret) {
 367                         ret = sendret;
 368                         goto error;
 369                 }
 370         }
 371         /*
 372          * We still have the memory map reference, and the fds have been
 373          * sent to the sessiond. We can therefore close those fds. Note
 374          * that we keep the write side of the wait_fd open, but close
 375          * the read side.
 376          */
 377         if (lur.ret_code == USTCOMM_OK) {
 378                 switch (lum->cmd) {
 379                 case LTTNG_UST_STREAM:
 380                         if (shm_fd >= 0) {
 381                                 ret = close(shm_fd);
 382                                 if (ret) {
 383                                         PERROR("Error closing stream shm_fd");
 384                                 }
 385                                 *args.stream.shm_fd = -1;
 386                         }
 387                         if (wait_fd >= 0) {
 388                                 ret = close(wait_fd);
 389                                 if (ret) {
 390                                         PERROR("Error closing stream wait_fd");
 391                                 }
 392                                 *args.stream.wait_fd = -1;
 393                         }
 394                         break;
 395                 case LTTNG_UST_METADATA:
 396                 case LTTNG_UST_CHANNEL:
 397                         if (shm_fd >= 0) {
 398                                 ret = close(shm_fd);
 399                                 if (ret) {
 400                                         PERROR("Error closing channel shm_fd");
 401                                 }
 402                                 *args.channel.shm_fd = -1;
 403                         }
 404                         if (wait_fd >= 0) {
 405                                 ret = close(wait_fd);
 406                                 if (ret) {
 407                                         PERROR("Error closing channel wait_fd");
 408                                 }
 409                                 *args.channel.wait_fd = -1;
 410                         }
 411                         break;
 412                 }
 413         }
 414
 415 error:
 416         ust_unlock();
 417         return ret;
 418 }
 419
 420 static
 421 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 422 {
 423         int ret;
 424
 425         if (sock_info->socket != -1) {
 426                 ret = ustcomm_close_unix_sock(sock_info->socket);
 427                 if (ret) {
 428                         ERR("Error closing apps socket");
 429                 }
 430                 sock_info->socket = -1;
 431         }
 432         if (sock_info->root_handle != -1) {
 433                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 434                 if (ret) {
 435                         ERR("Error unref root handle");
 436                 }
 437                 sock_info->root_handle = -1;
 438         }
 439         sock_info->constructor_sem_posted = 0;
 440         /*
 441          * wait_shm_mmap is used by listener threads outside of the
 442          * ust lock, so we cannot tear it down ourselves, because we
 443          * cannot join on these threads. Leave this task to the OS
 444          * process exit.
 445          */
 446         if (!exiting && sock_info->wait_shm_mmap) {
 447                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 448                 if (ret) {
 449                         ERR("Error unmapping wait shm");
 450                 }
 451                 sock_info->wait_shm_mmap = NULL;
 452         }
 453 }
 454
 455 /*
 456  * Using fork to set umask in the child process (not multi-thread safe).
 457  * We deal with the shm_open vs ftruncate race (happening when the
 458  * sessiond owns the shm and does not let everybody modify it, to ensure
 459  * safety against shm_unlink) by simply letting the mmap fail and
 460  * retrying after a few seconds.
 461  * For global shm, everybody has rw access to it until the sessiond
 462  * starts.
 463  */
 464 static
 465 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 466 {
 467         int wait_shm_fd, ret;
 468         pid_t pid;
 469
 470         /*
 471          * Try to open read-only.
 472          */
 473         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 474         if (wait_shm_fd >= 0) {
 475                 goto end;
 476         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 477                 /*
 478                  * Real-only open did not work, and it's not because the
 479                  * entry was not present. It's a failure that prohibits
 480                  * using shm.
 481                  */
 482                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 483                 goto end;
 484         }
 485         /*
 486          * If the open failed because the file did not exist, try
 487          * creating it ourself.
 488          */
 489         lttng_ust_nest_count++;
 490         pid = fork();
 491         lttng_ust_nest_count--;
 492         if (pid > 0) {
 493                 int status;
 494
 495                 /*
 496                  * Parent: wait for child to return, in which case the
 497                  * shared memory map will have been created.
 498                  */
 499                 pid = wait(&status);
 500                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 501                         wait_shm_fd = -1;
 502                         goto end;
 503                 }
 504                 /*
 505                  * Try to open read-only again after creation.
 506                  */
 507                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 508                 if (wait_shm_fd < 0) {
 509                         /*
 510                          * Real-only open did not work. It's a failure
 511                          * that prohibits using shm.
 512                          */
 513                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 514                         goto end;
 515                 }
 516                 goto end;
 517         } else if (pid == 0) {
 518                 int create_mode;
 519
 520                 /* Child */
 521                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 522                 if (sock_info->global)
 523                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 524                 /*
 525                  * We're alone in a child process, so we can modify the
 526                  * process-wide umask.
 527                  */
 528                 umask(~create_mode);
 529                 /*
 530                  * Try creating shm (or get rw access).
 531                  * We don't do an exclusive open, because we allow other
 532                  * processes to create+ftruncate it concurrently.
 533                  */
 534                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 535                                 O_RDWR | O_CREAT, create_mode);
 536                 if (wait_shm_fd >= 0) {
 537                         ret = ftruncate(wait_shm_fd, mmap_size);
 538                         if (ret) {
 539                                 PERROR("ftruncate");
 540                                 _exit(EXIT_FAILURE);
 541                         }
 542                         _exit(EXIT_SUCCESS);
 543                 }
 544                 /*
 545                  * For local shm, we need to have rw access to accept
 546                  * opening it: this means the local sessiond will be
 547                  * able to wake us up. For global shm, we open it even
 548                  * if rw access is not granted, because the root.root
 549                  * sessiond will be able to override all rights and wake
 550                  * us up.
 551                  */
 552                 if (!sock_info->global && errno != EACCES) {
 553                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 554                         _exit(EXIT_FAILURE);
 555                 }
 556                 /*
 557                  * The shm exists, but we cannot open it RW. Report
 558                  * success.
 559                  */
 560                 _exit(EXIT_SUCCESS);
 561         } else {
 562                 return -1;
 563         }
 564 end:
 565         if (wait_shm_fd >= 0 && !sock_info->global) {
 566                 struct stat statbuf;
 567
 568                 /*
 569                  * Ensure that our user is the owner of the shm file for
 570                  * local shm. If we do not own the file, it means our
 571                  * sessiond will not have access to wake us up (there is
 572                  * probably a rogue process trying to fake our
 573                  * sessiond). Fallback to polling method in this case.
 574                  */
 575                 ret = fstat(wait_shm_fd, &statbuf);
 576                 if (ret) {
 577                         PERROR("fstat");
 578                         goto error_close;
 579                 }
 580                 if (statbuf.st_uid != getuid())
 581                         goto error_close;
 582         }
 583         return wait_shm_fd;
 584
 585 error_close:
 586         ret = close(wait_shm_fd);
 587         if (ret) {
 588                 PERROR("Error closing fd");
 589         }
 590         return -1;
 591 }
 592
 593 static
 594 char *get_map_shm(struct sock_info *sock_info)
 595 {
 596         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 597         int wait_shm_fd, ret;
 598         char *wait_shm_mmap;
 599
 600         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 601         if (wait_shm_fd < 0) {
 602                 goto error;
 603         }
 604         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 605                   MAP_SHARED, wait_shm_fd, 0);
 606         /* close shm fd immediately after taking the mmap reference */
 607         ret = close(wait_shm_fd);
 608         if (ret) {
 609                 PERROR("Error closing fd");
 610         }
 611         if (wait_shm_mmap == MAP_FAILED) {
 612                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 613                 goto error;
 614         }
 615         return wait_shm_mmap;
 616
 617 error:
 618         return NULL;
 619 }
 620
 621 static
 622 void wait_for_sessiond(struct sock_info *sock_info)
 623 {
 624         int ret;
 625
 626         ust_lock();
 627         if (lttng_ust_comm_should_quit) {
 628                 goto quit;
 629         }
 630         if (wait_poll_fallback) {
 631                 goto error;
 632         }
 633         if (!sock_info->wait_shm_mmap) {
 634                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 635                 if (!sock_info->wait_shm_mmap)
 636                         goto error;
 637         }
 638         ust_unlock();
 639
 640         DBG("Waiting for %s apps sessiond", sock_info->name);
 641         /* Wait for futex wakeup */
 642         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 643                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 644                         FUTEX_WAIT, 0, NULL, NULL, 0);
 645                 if (ret < 0) {
 646                         if (errno == EFAULT) {
 647                                 wait_poll_fallback = 1;
 648                                 DBG(
 649 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 650 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 651 "Please upgrade your kernel "
 652 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 653 "mainline). LTTng-UST will use polling mode fallback.");
 654                                 if (ust_debug())
 655                                         PERROR("futex");
 656                         }
 657                 }
 658         }
 659         return;
 660
 661 quit:
 662         ust_unlock();
 663         return;
 664
 665 error:
 666         ust_unlock();
 667         return;
 668 }
 669
 670 /*
 671  * This thread does not allocate any resource, except within
 672  * handle_message, within mutex protection. This mutex protects against
 673  * fork and exit.
 674  * The other moment it allocates resources is at socket connexion, which
 675  * is also protected by the mutex.
 676  */
 677 static
 678 void *ust_listener_thread(void *arg)
 679 {
 680         struct sock_info *sock_info = arg;
 681         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 682
 683         /* Restart trying to connect to the session daemon */
 684 restart:
 685         if (prev_connect_failed) {
 686                 /* Wait for sessiond availability with pipe */
 687                 wait_for_sessiond(sock_info);
 688                 if (has_waited) {
 689                         has_waited = 0;
 690                         /*
 691                          * Sleep for 5 seconds before retrying after a
 692                          * sequence of failure / wait / failure. This
 693                          * deals with a killed or broken session daemon.
 694                          */
 695                         sleep(5);
 696                 }
 697                 has_waited = 1;
 698                 prev_connect_failed = 0;
 699         }
 700         ust_lock();
 701
 702         if (lttng_ust_comm_should_quit) {
 703                 ust_unlock();
 704                 goto quit;
 705         }
 706
 707         if (sock_info->socket != -1) {
 708                 ret = ustcomm_close_unix_sock(sock_info->socket);
 709                 if (ret) {
 710                         ERR("Error closing %s apps socket", sock_info->name);
 711                 }
 712                 sock_info->socket = -1;
 713         }
 714
 715         /* Register */
 716         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 717         if (ret < 0) {
 718                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 719                 prev_connect_failed = 1;
 720                 /*
 721                  * If we cannot find the sessiond daemon, don't delay
 722                  * constructor execution.
 723                  */
 724                 ret = handle_register_done(sock_info);
 725                 assert(!ret);
 726                 ust_unlock();
 727                 goto restart;
 728         }
 729
 730         sock_info->socket = sock = ret;
 731
 732         /*
 733          * Create only one root handle per listener thread for the whole
 734          * process lifetime.
 735          */
 736         if (sock_info->root_handle == -1) {
 737                 ret = lttng_abi_create_root_handle();
 738                 if (ret < 0) {
 739                         ERR("Error creating root handle");
 740                         ust_unlock();
 741                         goto quit;
 742                 }
 743                 sock_info->root_handle = ret;
 744         }
 745
 746         ret = register_app_to_sessiond(sock);
 747         if (ret < 0) {
 748                 ERR("Error registering to %s apps socket", sock_info->name);
 749                 prev_connect_failed = 1;
 750                 /*
 751                  * If we cannot register to the sessiond daemon, don't
 752                  * delay constructor execution.
 753                  */
 754                 ret = handle_register_done(sock_info);
 755                 assert(!ret);
 756                 ust_unlock();
 757                 goto restart;
 758         }
 759         ust_unlock();
 760
 761         for (;;) {
 762                 ssize_t len;
 763                 struct ustcomm_ust_msg lum;
 764
 765                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 766                 switch (len) {
 767                 case 0: /* orderly shutdown */
 768                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 769                         ust_lock();
 770                         /*
 771                          * Either sessiond has shutdown or refused us by closing the socket.
 772                          * In either case, we don't want to delay construction execution,
 773                          * and we need to wait before retry.
 774                          */
 775                         prev_connect_failed = 1;
 776                         /*
 777                          * If we cannot register to the sessiond daemon, don't
 778                          * delay constructor execution.
 779                          */
 780                         ret = handle_register_done(sock_info);
 781                         assert(!ret);
 782                         ust_unlock();
 783                         goto end;
 784                 case sizeof(lum):
 785                         DBG("message received\n");
 786                         ret = handle_message(sock_info, sock, &lum);
 787                         if (ret < 0) {
 788                                 ERR("Error handling message for %s socket", sock_info->name);
 789                         }
 790                         continue;
 791                 case -1:
 792                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 793                         if (errno == ECONNRESET) {
 794                                 ERR("%s remote end closed connection\n", sock_info->name);
 795                                 goto end;
 796                         }
 797                         goto end;
 798                 default:
 799                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 800                         continue;
 801                 }
 802
 803         }
 804 end:
 805         goto restart;   /* try to reconnect */
 806 quit:
 807         return NULL;
 808 }
 809
 810 /*
 811  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 812  */
 813 static
 814 int get_timeout(struct timespec *constructor_timeout)
 815 {
 816         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 817         char *str_delay;
 818         int ret;
 819
 820         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 821         if (str_delay) {
 822                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 823         }
 824
 825         switch (constructor_delay_ms) {
 826         case -1:/* fall-through */
 827         case 0:
 828                 return constructor_delay_ms;
 829         default:
 830                 break;
 831         }
 832
 833         /*
 834          * If we are unable to find the current time, don't wait.
 835          */
 836         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 837         if (ret) {
 838                 return -1;
 839         }
 840         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 841         constructor_timeout->tv_nsec +=
 842                 (constructor_delay_ms % 1000UL) * 1000000UL;
 843         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 844                 constructor_timeout->tv_sec++;
 845                 constructor_timeout->tv_nsec -= 1000000000UL;
 846         }
 847         return 1;
 848 }
 849
 850 /*
 851  * sessiond monitoring thread: monitor presence of global and per-user
 852  * sessiond by polling the application common named pipe.
 853  */
 854 /* TODO */
 855
 856 void __attribute__((constructor)) lttng_ust_init(void)
 857 {
 858         struct timespec constructor_timeout;
 859         sigset_t sig_all_blocked, orig_parent_mask;
 860         pthread_attr_t thread_attr;
 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         lttng_fixup_nest_count_tls();
 876
 877         /*
 878          * We want precise control over the order in which we construct
 879          * our sub-libraries vs starting to receive commands from
 880          * sessiond (otherwise leading to errors when trying to create
 881          * sessiond before the init functions are completed).
 882          */
 883         init_usterr();
 884         init_tracepoint();
 885         ltt_ring_buffer_metadata_client_init();
 886         ltt_ring_buffer_client_overwrite_init();
 887         ltt_ring_buffer_client_discard_init();
 888
 889         timeout_mode = get_timeout(&constructor_timeout);
 890
 891         ret = sem_init(&constructor_wait, 0, 0);
 892         assert(!ret);
 893
 894         ret = setup_local_apps();
 895         if (ret) {
 896                 DBG("local apps setup returned %d", ret);
 897         }
 898
 899         /* A new thread created by pthread_create inherits the signal mask
 900          * from the parent. To avoid any signal being received by the
 901          * listener thread, we block all signals temporarily in the parent,
 902          * while we create the listener thread.
 903          */
 904         sigfillset(&sig_all_blocked);
 905         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 906         if (ret) {
 907                 ERR("pthread_sigmask: %s", strerror(ret));
 908         }
 909
 910         ret = pthread_attr_init(&thread_attr);
 911         if (ret) {
 912                 ERR("pthread_attr_init: %s", strerror(ret));
 913         }
 914         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
 915         if (ret) {
 916                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
 917         }
 918
 919         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
 920                         ust_listener_thread, &global_apps);
 921         if (ret) {
 922                 ERR("pthread_create global: %s", strerror(ret));
 923         }
 924         if (local_apps.allowed) {
 925                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
 926                                 ust_listener_thread, &local_apps);
 927                 if (ret) {
 928                         ERR("pthread_create local: %s", strerror(ret));
 929                 }
 930         } else {
 931                 handle_register_done(&local_apps);
 932         }
 933         ret = pthread_attr_destroy(&thread_attr);
 934         if (ret) {
 935                 ERR("pthread_attr_destroy: %s", strerror(ret));
 936         }
 937
 938         /* Restore original signal mask in parent */
 939         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
 940         if (ret) {
 941                 ERR("pthread_sigmask: %s", strerror(ret));
 942         }
 943
 944         switch (timeout_mode) {
 945         case 1: /* timeout wait */
 946                 do {
 947                         ret = sem_timedwait(&constructor_wait,
 948                                         &constructor_timeout);
 949                 } while (ret < 0 && errno == EINTR);
 950                 if (ret < 0 && errno == ETIMEDOUT) {
 951                         ERR("Timed out waiting for ltt-sessiond");
 952                 } else {
 953                         assert(!ret);
 954                 }
 955                 break;
 956         case -1:/* wait forever */
 957                 do {
 958                         ret = sem_wait(&constructor_wait);
 959                 } while (ret < 0 && errno == EINTR);
 960                 assert(!ret);
 961                 break;
 962         case 0: /* no timeout */
 963                 break;
 964         }
 965 }
 966
 967 static
 968 void lttng_ust_cleanup(int exiting)
 969 {
 970         cleanup_sock_info(&global_apps, exiting);
 971         if (local_apps.allowed) {
 972                 cleanup_sock_info(&local_apps, exiting);
 973         }
 974         /*
 975          * The teardown in this function all affect data structures
 976          * accessed under the UST lock by the listener thread. This
 977          * lock, along with the lttng_ust_comm_should_quit flag, ensure
 978          * that none of these threads are accessing this data at this
 979          * point.
 980          */
 981         lttng_ust_abi_exit();
 982         lttng_ust_events_exit();
 983         ltt_ring_buffer_client_discard_exit();
 984         ltt_ring_buffer_client_overwrite_exit();
 985         ltt_ring_buffer_metadata_client_exit();
 986         exit_tracepoint();
 987         if (!exiting) {
 988                 /* Reinitialize values for fork */
 989                 sem_count = 2;
 990                 lttng_ust_comm_should_quit = 0;
 991                 initialized = 0;
 992         }
 993 }
 994
 995 void __attribute__((destructor)) lttng_ust_exit(void)
 996 {
 997         int ret;
 998
 999         /*
1000          * Using pthread_cancel here because:
1001          * A) we don't want to hang application teardown.
1002          * B) the thread is not allocating any resource.
1003          */
1004
1005         /*
1006          * Require the communication thread to quit. Synchronize with
1007          * mutexes to ensure it is not in a mutex critical section when
1008          * pthread_cancel is later called.
1009          */
1010         ust_lock();
1011         lttng_ust_comm_should_quit = 1;
1012         ust_unlock();
1013
1014         /* cancel threads */
1015         ret = pthread_cancel(global_apps.ust_listener);
1016         if (ret) {
1017                 ERR("Error cancelling global ust listener thread: %s",
1018                         strerror(ret));
1019         }
1020         if (local_apps.allowed) {
1021                 ret = pthread_cancel(local_apps.ust_listener);
1022                 if (ret) {
1023                         ERR("Error cancelling local ust listener thread: %s",
1024                                 strerror(ret));
1025                 }
1026         }
1027         /*
1028          * Do NOT join threads: use of sys_futex makes it impossible to
1029          * join the threads without using async-cancel, but async-cancel
1030          * is delivered by a signal, which could hit the target thread
1031          * anywhere in its code path, including while the ust_lock() is
1032          * held, causing a deadlock for the other thread. Let the OS
1033          * cleanup the threads if there are stalled in a syscall.
1034          */
1035         lttng_ust_cleanup(1);
1036 }
1037
1038 /*
1039  * We exclude the worker threads across fork and clone (except
1040  * CLONE_VM), because these system calls only keep the forking thread
1041  * running in the child.  Therefore, we don't want to call fork or clone
1042  * in the middle of an tracepoint or ust tracing state modification.
1043  * Holding this mutex protects these structures across fork and clone.
1044  */
1045 void ust_before_fork(sigset_t *save_sigset)
1046 {
1047         /*
1048          * Disable signals. This is to avoid that the child intervenes
1049          * before it is properly setup for tracing. It is safer to
1050          * disable all signals, because then we know we are not breaking
1051          * anything by restoring the original mask.
1052          */
1053         sigset_t all_sigs;
1054         int ret;
1055
1056         if (lttng_ust_nest_count)
1057                 return;
1058         /* Disable signals */
1059         sigfillset(&all_sigs);
1060         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1061         if (ret == -1) {
1062                 PERROR("sigprocmask");
1063         }
1064         ust_lock();
1065         rcu_bp_before_fork();
1066 }
1067
1068 static void ust_after_fork_common(sigset_t *restore_sigset)
1069 {
1070         int ret;
1071
1072         DBG("process %d", getpid());
1073         ust_unlock();
1074         /* Restore signals */
1075         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1076         if (ret == -1) {
1077                 PERROR("sigprocmask");
1078         }
1079 }
1080
1081 void ust_after_fork_parent(sigset_t *restore_sigset)
1082 {
1083         if (lttng_ust_nest_count)
1084                 return;
1085         DBG("process %d", getpid());
1086         rcu_bp_after_fork_parent();
1087         /* Release mutexes and reenable signals */
1088         ust_after_fork_common(restore_sigset);
1089 }
1090
1091 /*
1092  * After fork, in the child, we need to cleanup all the leftover state,
1093  * except the worker thread which already magically disappeared thanks
1094  * to the weird Linux fork semantics. After tyding up, we call
1095  * lttng_ust_init() again to start over as a new PID.
1096  *
1097  * This is meant for forks() that have tracing in the child between the
1098  * fork and following exec call (if there is any).
1099  */
1100 void ust_after_fork_child(sigset_t *restore_sigset)
1101 {
1102         if (lttng_ust_nest_count)
1103                 return;
1104         DBG("process %d", getpid());
1105         /* Release urcu mutexes */
1106         rcu_bp_after_fork_child();
1107         lttng_ust_cleanup(0);
1108         lttng_context_vtid_reset();
1109         /* Release mutexes and reenable signals */
1110         ust_after_fork_common(restore_sigset);
1111         lttng_ust_init();
1112 }