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