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 <helper.h>
  47 #include "tracepoint-internal.h"
  48 #include "ltt-tracer-core.h"
  49 #include "compat.h"
  50 #include "../libringbuffer/tlsfixup.h"
  51
  52 /*
  53  * Has lttng ust comm constructor been called ?
  54  */
  55 static int initialized;
  56
  57 /*
  58  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  59  * Held when handling a command, also held by fork() to deal with
  60  * removal of threads, and by exit path.
  61  */
  62
  63 /* Should the ust comm thread quit ? */
  64 static int lttng_ust_comm_should_quit;
  65
  66 /*
  67  * Wait for either of these before continuing to the main
  68  * program:
  69  * - the register_done message from sessiond daemon
  70  *   (will let the sessiond daemon enable sessions before main
  71  *   starts.)
  72  * - sessiond daemon is not reachable.
  73  * - timeout (ensuring applications are resilient to session
  74  *   daemon problems).
  75  */
  76 static sem_t constructor_wait;
  77 /*
  78  * Doing this for both the global and local sessiond.
  79  */
  80 static int sem_count = { 2 };
  81
  82 /*
  83  * Counting nesting within lttng-ust. Used to ensure that calling fork()
  84  * from liblttng-ust does not execute the pre/post fork handlers.
  85  */
  86 static int __thread lttng_ust_nest_count;
  87
  88 /*
  89  * Info about socket and associated listener thread.
  90  */
  91 struct sock_info {
  92         const char *name;
  93         pthread_t ust_listener; /* listener thread */
  94         int root_handle;
  95         int constructor_sem_posted;
  96         int allowed;
  97         int global;
  98
  99         char sock_path[PATH_MAX];
 100         int socket;
 101
 102         char wait_shm_path[PATH_MAX];
 103         char *wait_shm_mmap;
 104 };
 105
 106 /* Socket from app (connect) to session daemon (listen) for communication */
 107 struct sock_info global_apps = {
 108         .name = "global",
 109         .global = 1,
 110
 111         .root_handle = -1,
 112         .allowed = 1,
 113
 114         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 115         .socket = -1,
 116
 117         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 118 };
 119
 120 /* TODO: allow global_apps_sock_path override */
 121
 122 struct sock_info local_apps = {
 123         .name = "local",
 124         .global = 0,
 125         .root_handle = -1,
 126         .allowed = 0,   /* Check setuid bit first */
 127
 128         .socket = -1,
 129 };
 130
 131 static int wait_poll_fallback;
 132
 133 extern void ltt_ring_buffer_client_overwrite_init(void);
 134 extern void ltt_ring_buffer_client_discard_init(void);
 135 extern void ltt_ring_buffer_metadata_client_init(void);
 136 extern void ltt_ring_buffer_client_overwrite_exit(void);
 137 extern void ltt_ring_buffer_client_discard_exit(void);
 138 extern void ltt_ring_buffer_metadata_client_exit(void);
 139
 140 /*
 141  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 142  */
 143 static
 144 void lttng_fixup_nest_count_tls(void)
 145 {
 146         asm volatile ("" : : "m" (lttng_ust_nest_count));
 147 }
 148
 149 static
 150 int setup_local_apps(void)
 151 {
 152         const char *home_dir;
 153         uid_t uid;
 154
 155         uid = getuid();
 156         /*
 157          * Disallow per-user tracing for setuid binaries.
 158          */
 159         if (uid != geteuid()) {
 160                 local_apps.allowed = 0;
 161                 return 0;
 162         } else {
 163                 local_apps.allowed = 1;
 164         }
 165         home_dir = (const char *) getenv("HOME");
 166         if (!home_dir)
 167                 return -ENOENT;
 168         snprintf(local_apps.sock_path, PATH_MAX,
 169                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 170         snprintf(local_apps.wait_shm_path, PATH_MAX,
 171                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 172         return 0;
 173 }
 174
 175 static
 176 int register_app_to_sessiond(int socket)
 177 {
 178         ssize_t ret;
 179         struct {
 180                 uint32_t major;
 181                 uint32_t minor;
 182                 pid_t pid;
 183                 pid_t ppid;
 184                 uid_t uid;
 185                 gid_t gid;
 186                 uint32_t bits_per_long;
 187                 char name[16];  /* process name */
 188         } reg_msg;
 189
 190         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 191         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 192         reg_msg.pid = getpid();
 193         reg_msg.ppid = getppid();
 194         reg_msg.uid = getuid();
 195         reg_msg.gid = getgid();
 196         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 197         lttng_ust_getprocname(reg_msg.name);
 198
 199         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 200         if (ret >= 0 && ret != sizeof(reg_msg))
 201                 return -EIO;
 202         return ret;
 203 }
 204
 205 static
 206 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 207 {
 208         ssize_t len;
 209
 210         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 211         switch (len) {
 212         case sizeof(*lur):
 213                 DBG("message successfully sent");
 214                 return 0;
 215         case -1:
 216                 if (errno == ECONNRESET) {
 217                         printf("remote end closed connection\n");
 218                         return 0;
 219                 }
 220                 return -1;
 221         default:
 222                 printf("incorrect message size: %zd\n", len);
 223                 return -1;
 224         }
 225 }
 226
 227 static
 228 int handle_register_done(struct sock_info *sock_info)
 229 {
 230         int ret;
 231
 232         if (sock_info->constructor_sem_posted)
 233                 return 0;
 234         sock_info->constructor_sem_posted = 1;
 235         if (uatomic_read(&sem_count) <= 0) {
 236                 return 0;
 237         }
 238         ret = uatomic_add_return(&sem_count, -1);
 239         if (ret == 0) {
 240                 ret = sem_post(&constructor_wait);
 241                 assert(!ret);
 242         }
 243         return 0;
 244 }
 245
 246 static
 247 int handle_message(struct sock_info *sock_info,
 248                 int sock, struct ustcomm_ust_msg *lum)
 249 {
 250         int ret = 0;
 251         const struct lttng_ust_objd_ops *ops;
 252         struct ustcomm_ust_reply lur;
 253         int shm_fd, wait_fd;
 254         union ust_args args;
 255         ssize_t len;
 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         case LTTNG_UST_FILTER:
 286         {
 287                 /* Receive filter data */
 288                 struct lttng_ust_filter_bytecode *bytecode;
 289
 290                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 291                         ERR("Filter data size is too large: %u bytes\n",
 292                                 lum->u.filter.data_size);
 293                         ret = -EINVAL;
 294                         goto error;
 295                 }
 296                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 297                 if (!bytecode) {
 298                         ret = -ENOMEM;
 299                         goto error;
 300                 }
 301                 len = ustcomm_recv_unix_sock(sock, bytecode->data,
 302                                 lum->u.filter.data_size);
 303                 switch (len) {
 304                 case 0: /* orderly shutdown */
 305                         ret = 0;
 306                         free(bytecode);
 307                         goto error;
 308                 case -1:
 309                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 310                         if (errno == ECONNRESET) {
 311                                 ERR("%s remote end closed connection\n", sock_info->name);
 312                                 ret = -EINVAL;
 313                                 free(bytecode);
 314                                 goto error;
 315                         }
 316                         ret = -EINVAL;
 317                         goto end;
 318                 default:
 319                         if (len == lum->u.filter.data_size) {
 320                                 DBG("filter data received\n");
 321                                 break;
 322                         } else {
 323                                 ERR("incorrect filter data message size: %zd\n", len);
 324                                 ret = -EINVAL;
 325                                 free(bytecode);
 326                                 goto end;
 327                         }
 328                 }
 329                 bytecode->len = lum->u.filter.data_size;
 330                 bytecode->reloc_offset = lum->u.filter.reloc_offset;
 331                 fprintf(stderr, "RECV len %d rel %d\n",
 332                         lum->u.filter.data_size,
 333                         lum->u.filter.reloc_offset);
 334                 if (ops->cmd) {
 335                         ret = ops->cmd(lum->handle, lum->cmd,
 336                                         (unsigned long) bytecode,
 337                                         &args);
 338                         if (ret) {
 339                                 free(bytecode);
 340                         }
 341                         /* don't free bytecode if everything went fine. */
 342                 } else {
 343                         ret = -ENOSYS;
 344                         free(bytecode);
 345                 }
 346                 break;
 347         }
 348         default:
 349                 if (ops->cmd)
 350                         ret = ops->cmd(lum->handle, lum->cmd,
 351                                         (unsigned long) &lum->u,
 352                                         &args);
 353                 else
 354                         ret = -ENOSYS;
 355                 break;
 356         }
 357
 358 end:
 359         lur.handle = lum->handle;
 360         lur.cmd = lum->cmd;
 361         lur.ret_val = ret;
 362         if (ret >= 0) {
 363                 lur.ret_code = USTCOMM_OK;
 364         } else {
 365                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 366                 lur.ret_code = ret;
 367         }
 368         if (ret >= 0) {
 369                 switch (lum->cmd) {
 370                 case LTTNG_UST_STREAM:
 371                         /*
 372                          * Special-case reply to send stream info.
 373                          * Use lum.u output.
 374                          */
 375                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 376                         shm_fd = *args.stream.shm_fd;
 377                         wait_fd = *args.stream.wait_fd;
 378                         break;
 379                 case LTTNG_UST_METADATA:
 380                 case LTTNG_UST_CHANNEL:
 381                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 382                         shm_fd = *args.channel.shm_fd;
 383                         wait_fd = *args.channel.wait_fd;
 384                         break;
 385                 case LTTNG_UST_TRACER_VERSION:
 386                         lur.u.version = lum->u.version;
 387                         break;
 388                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 389                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 390                         break;
 391                 }
 392         }
 393         ret = send_reply(sock, &lur);
 394         if (ret < 0) {
 395                 perror("error sending reply");
 396                 goto error;
 397         }
 398
 399         if ((lum->cmd == LTTNG_UST_STREAM
 400              || lum->cmd == LTTNG_UST_CHANNEL
 401              || lum->cmd == LTTNG_UST_METADATA)
 402                         && lur.ret_code == USTCOMM_OK) {
 403                 int sendret = 0;
 404
 405                 /* we also need to send the file descriptors. */
 406                 ret = ustcomm_send_fds_unix_sock(sock,
 407                         &shm_fd, &shm_fd,
 408                         1, sizeof(int));
 409                 if (ret < 0) {
 410                         perror("send shm_fd");
 411                         sendret = ret;
 412                 }
 413                 /*
 414                  * The sessiond expects 2 file descriptors, even upon
 415                  * error.
 416                  */
 417                 ret = ustcomm_send_fds_unix_sock(sock,
 418                         &wait_fd, &wait_fd,
 419                         1, sizeof(int));
 420                 if (ret < 0) {
 421                         perror("send wait_fd");
 422                         goto error;
 423                 }
 424                 if (sendret) {
 425                         ret = sendret;
 426                         goto error;
 427                 }
 428         }
 429         /*
 430          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 431          * after the reply.
 432          */
 433         if (lur.ret_code == USTCOMM_OK) {
 434                 switch (lum->cmd) {
 435                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 436                         len = ustcomm_send_unix_sock(sock,
 437                                 &args.field_list.entry,
 438                                 sizeof(args.field_list.entry));
 439                         if (len != sizeof(args.field_list.entry)) {
 440                                 ret = -1;
 441                                 goto error;
 442                         }
 443                 }
 444         }
 445         /*
 446          * We still have the memory map reference, and the fds have been
 447          * sent to the sessiond. We can therefore close those fds. Note
 448          * that we keep the write side of the wait_fd open, but close
 449          * the read side.
 450          */
 451         if (lur.ret_code == USTCOMM_OK) {
 452                 switch (lum->cmd) {
 453                 case LTTNG_UST_STREAM:
 454                         if (shm_fd >= 0) {
 455                                 ret = close(shm_fd);
 456                                 if (ret) {
 457                                         PERROR("Error closing stream shm_fd");
 458                                 }
 459                                 *args.stream.shm_fd = -1;
 460                         }
 461                         if (wait_fd >= 0) {
 462                                 ret = close(wait_fd);
 463                                 if (ret) {
 464                                         PERROR("Error closing stream wait_fd");
 465                                 }
 466                                 *args.stream.wait_fd = -1;
 467                         }
 468                         break;
 469                 case LTTNG_UST_METADATA:
 470                 case LTTNG_UST_CHANNEL:
 471                         if (shm_fd >= 0) {
 472                                 ret = close(shm_fd);
 473                                 if (ret) {
 474                                         PERROR("Error closing channel shm_fd");
 475                                 }
 476                                 *args.channel.shm_fd = -1;
 477                         }
 478                         if (wait_fd >= 0) {
 479                                 ret = close(wait_fd);
 480                                 if (ret) {
 481                                         PERROR("Error closing channel wait_fd");
 482                                 }
 483                                 *args.channel.wait_fd = -1;
 484                         }
 485                         break;
 486                 }
 487         }
 488
 489 error:
 490         ust_unlock();
 491         return ret;
 492 }
 493
 494 static
 495 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 496 {
 497         int ret;
 498
 499         if (sock_info->socket != -1) {
 500                 ret = ustcomm_close_unix_sock(sock_info->socket);
 501                 if (ret) {
 502                         ERR("Error closing apps socket");
 503                 }
 504                 sock_info->socket = -1;
 505         }
 506         if (sock_info->root_handle != -1) {
 507                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 508                 if (ret) {
 509                         ERR("Error unref root handle");
 510                 }
 511                 sock_info->root_handle = -1;
 512         }
 513         sock_info->constructor_sem_posted = 0;
 514         /*
 515          * wait_shm_mmap is used by listener threads outside of the
 516          * ust lock, so we cannot tear it down ourselves, because we
 517          * cannot join on these threads. Leave this task to the OS
 518          * process exit.
 519          */
 520         if (!exiting && sock_info->wait_shm_mmap) {
 521                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 522                 if (ret) {
 523                         ERR("Error unmapping wait shm");
 524                 }
 525                 sock_info->wait_shm_mmap = NULL;
 526         }
 527 }
 528
 529 /*
 530  * Using fork to set umask in the child process (not multi-thread safe).
 531  * We deal with the shm_open vs ftruncate race (happening when the
 532  * sessiond owns the shm and does not let everybody modify it, to ensure
 533  * safety against shm_unlink) by simply letting the mmap fail and
 534  * retrying after a few seconds.
 535  * For global shm, everybody has rw access to it until the sessiond
 536  * starts.
 537  */
 538 static
 539 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 540 {
 541         int wait_shm_fd, ret;
 542         pid_t pid;
 543
 544         /*
 545          * Try to open read-only.
 546          */
 547         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 548         if (wait_shm_fd >= 0) {
 549                 goto end;
 550         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 551                 /*
 552                  * Real-only open did not work, and it's not because the
 553                  * entry was not present. It's a failure that prohibits
 554                  * using shm.
 555                  */
 556                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 557                 goto end;
 558         }
 559         /*
 560          * If the open failed because the file did not exist, try
 561          * creating it ourself.
 562          */
 563         lttng_ust_nest_count++;
 564         pid = fork();
 565         lttng_ust_nest_count--;
 566         if (pid > 0) {
 567                 int status;
 568
 569                 /*
 570                  * Parent: wait for child to return, in which case the
 571                  * shared memory map will have been created.
 572                  */
 573                 pid = wait(&status);
 574                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 575                         wait_shm_fd = -1;
 576                         goto end;
 577                 }
 578                 /*
 579                  * Try to open read-only again after creation.
 580                  */
 581                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 582                 if (wait_shm_fd < 0) {
 583                         /*
 584                          * Real-only open did not work. It's a failure
 585                          * that prohibits using shm.
 586                          */
 587                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 588                         goto end;
 589                 }
 590                 goto end;
 591         } else if (pid == 0) {
 592                 int create_mode;
 593
 594                 /* Child */
 595                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 596                 if (sock_info->global)
 597                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 598                 /*
 599                  * We're alone in a child process, so we can modify the
 600                  * process-wide umask.
 601                  */
 602                 umask(~create_mode);
 603                 /*
 604                  * Try creating shm (or get rw access).
 605                  * We don't do an exclusive open, because we allow other
 606                  * processes to create+ftruncate it concurrently.
 607                  */
 608                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 609                                 O_RDWR | O_CREAT, create_mode);
 610                 if (wait_shm_fd >= 0) {
 611                         ret = ftruncate(wait_shm_fd, mmap_size);
 612                         if (ret) {
 613                                 PERROR("ftruncate");
 614                                 exit(EXIT_FAILURE);
 615                         }
 616                         exit(EXIT_SUCCESS);
 617                 }
 618                 /*
 619                  * For local shm, we need to have rw access to accept
 620                  * opening it: this means the local sessiond will be
 621                  * able to wake us up. For global shm, we open it even
 622                  * if rw access is not granted, because the root.root
 623                  * sessiond will be able to override all rights and wake
 624                  * us up.
 625                  */
 626                 if (!sock_info->global && errno != EACCES) {
 627                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 628                         exit(EXIT_FAILURE);
 629                 }
 630                 /*
 631                  * The shm exists, but we cannot open it RW. Report
 632                  * success.
 633                  */
 634                 exit(EXIT_SUCCESS);
 635         } else {
 636                 return -1;
 637         }
 638 end:
 639         if (wait_shm_fd >= 0 && !sock_info->global) {
 640                 struct stat statbuf;
 641
 642                 /*
 643                  * Ensure that our user is the owner of the shm file for
 644                  * local shm. If we do not own the file, it means our
 645                  * sessiond will not have access to wake us up (there is
 646                  * probably a rogue process trying to fake our
 647                  * sessiond). Fallback to polling method in this case.
 648                  */
 649                 ret = fstat(wait_shm_fd, &statbuf);
 650                 if (ret) {
 651                         PERROR("fstat");
 652                         goto error_close;
 653                 }
 654                 if (statbuf.st_uid != getuid())
 655                         goto error_close;
 656         }
 657         return wait_shm_fd;
 658
 659 error_close:
 660         ret = close(wait_shm_fd);
 661         if (ret) {
 662                 PERROR("Error closing fd");
 663         }
 664         return -1;
 665 }
 666
 667 static
 668 char *get_map_shm(struct sock_info *sock_info)
 669 {
 670         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 671         int wait_shm_fd, ret;
 672         char *wait_shm_mmap;
 673
 674         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 675         if (wait_shm_fd < 0) {
 676                 goto error;
 677         }
 678         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 679                   MAP_SHARED, wait_shm_fd, 0);
 680         /* close shm fd immediately after taking the mmap reference */
 681         ret = close(wait_shm_fd);
 682         if (ret) {
 683                 PERROR("Error closing fd");
 684         }
 685         if (wait_shm_mmap == MAP_FAILED) {
 686                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 687                 goto error;
 688         }
 689         return wait_shm_mmap;
 690
 691 error:
 692         return NULL;
 693 }
 694
 695 static
 696 void wait_for_sessiond(struct sock_info *sock_info)
 697 {
 698         int ret;
 699
 700         ust_lock();
 701         if (lttng_ust_comm_should_quit) {
 702                 goto quit;
 703         }
 704         if (wait_poll_fallback) {
 705                 goto error;
 706         }
 707         if (!sock_info->wait_shm_mmap) {
 708                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 709                 if (!sock_info->wait_shm_mmap)
 710                         goto error;
 711         }
 712         ust_unlock();
 713
 714         DBG("Waiting for %s apps sessiond", sock_info->name);
 715         /* Wait for futex wakeup */
 716         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 717                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 718                         FUTEX_WAIT, 0, NULL, NULL, 0);
 719                 if (ret < 0) {
 720                         if (errno == EFAULT) {
 721                                 wait_poll_fallback = 1;
 722                                 DBG(
 723 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 724 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 725 "Please upgrade your kernel "
 726 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 727 "mainline). LTTng-UST will use polling mode fallback.");
 728                                 if (ust_debug())
 729                                         PERROR("futex");
 730                         }
 731                 }
 732         }
 733         return;
 734
 735 quit:
 736         ust_unlock();
 737         return;
 738
 739 error:
 740         ust_unlock();
 741         return;
 742 }
 743
 744 /*
 745  * This thread does not allocate any resource, except within
 746  * handle_message, within mutex protection. This mutex protects against
 747  * fork and exit.
 748  * The other moment it allocates resources is at socket connection, which
 749  * is also protected by the mutex.
 750  */
 751 static
 752 void *ust_listener_thread(void *arg)
 753 {
 754         struct sock_info *sock_info = arg;
 755         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 756
 757         /* Restart trying to connect to the session daemon */
 758 restart:
 759         if (prev_connect_failed) {
 760                 /* Wait for sessiond availability with pipe */
 761                 wait_for_sessiond(sock_info);
 762                 if (has_waited) {
 763                         has_waited = 0;
 764                         /*
 765                          * Sleep for 5 seconds before retrying after a
 766                          * sequence of failure / wait / failure. This
 767                          * deals with a killed or broken session daemon.
 768                          */
 769                         sleep(5);
 770                 }
 771                 has_waited = 1;
 772                 prev_connect_failed = 0;
 773         }
 774         ust_lock();
 775
 776         if (lttng_ust_comm_should_quit) {
 777                 ust_unlock();
 778                 goto quit;
 779         }
 780
 781         if (sock_info->socket != -1) {
 782                 ret = ustcomm_close_unix_sock(sock_info->socket);
 783                 if (ret) {
 784                         ERR("Error closing %s apps socket", sock_info->name);
 785                 }
 786                 sock_info->socket = -1;
 787         }
 788
 789         /* Register */
 790         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 791         if (ret < 0) {
 792                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 793                 prev_connect_failed = 1;
 794                 /*
 795                  * If we cannot find the sessiond daemon, don't delay
 796                  * constructor execution.
 797                  */
 798                 ret = handle_register_done(sock_info);
 799                 assert(!ret);
 800                 ust_unlock();
 801                 goto restart;
 802         }
 803
 804         sock_info->socket = sock = ret;
 805
 806         /*
 807          * Create only one root handle per listener thread for the whole
 808          * process lifetime.
 809          */
 810         if (sock_info->root_handle == -1) {
 811                 ret = lttng_abi_create_root_handle();
 812                 if (ret < 0) {
 813                         ERR("Error creating root handle");
 814                         ust_unlock();
 815                         goto quit;
 816                 }
 817                 sock_info->root_handle = ret;
 818         }
 819
 820         ret = register_app_to_sessiond(sock);
 821         if (ret < 0) {
 822                 ERR("Error registering to %s apps socket", sock_info->name);
 823                 prev_connect_failed = 1;
 824                 /*
 825                  * If we cannot register to the sessiond daemon, don't
 826                  * delay constructor execution.
 827                  */
 828                 ret = handle_register_done(sock_info);
 829                 assert(!ret);
 830                 ust_unlock();
 831                 goto restart;
 832         }
 833         ust_unlock();
 834
 835         for (;;) {
 836                 ssize_t len;
 837                 struct ustcomm_ust_msg lum;
 838
 839                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 840                 switch (len) {
 841                 case 0: /* orderly shutdown */
 842                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 843                         ust_lock();
 844                         /*
 845                          * Either sessiond has shutdown or refused us by closing the socket.
 846                          * In either case, we don't want to delay construction execution,
 847                          * and we need to wait before retry.
 848                          */
 849                         prev_connect_failed = 1;
 850                         /*
 851                          * If we cannot register to the sessiond daemon, don't
 852                          * delay constructor execution.
 853                          */
 854                         ret = handle_register_done(sock_info);
 855                         assert(!ret);
 856                         ust_unlock();
 857                         goto end;
 858                 case sizeof(lum):
 859                         DBG("message received\n");
 860                         ret = handle_message(sock_info, sock, &lum);
 861                         if (ret < 0) {
 862                                 ERR("Error handling message for %s socket", sock_info->name);
 863                         }
 864                         continue;
 865                 case -1:
 866                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 867                         if (errno == ECONNRESET) {
 868                                 ERR("%s remote end closed connection\n", sock_info->name);
 869                                 goto end;
 870                         }
 871                         goto end;
 872                 default:
 873                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 874                         continue;
 875                 }
 876
 877         }
 878 end:
 879         goto restart;   /* try to reconnect */
 880 quit:
 881         return NULL;
 882 }
 883
 884 /*
 885  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 886  */
 887 static
 888 int get_timeout(struct timespec *constructor_timeout)
 889 {
 890         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 891         char *str_delay;
 892         int ret;
 893
 894         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 895         if (str_delay) {
 896                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 897         }
 898
 899         switch (constructor_delay_ms) {
 900         case -1:/* fall-through */
 901         case 0:
 902                 return constructor_delay_ms;
 903         default:
 904                 break;
 905         }
 906
 907         /*
 908          * If we are unable to find the current time, don't wait.
 909          */
 910         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 911         if (ret) {
 912                 return -1;
 913         }
 914         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 915         constructor_timeout->tv_nsec +=
 916                 (constructor_delay_ms % 1000UL) * 1000000UL;
 917         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 918                 constructor_timeout->tv_sec++;
 919                 constructor_timeout->tv_nsec -= 1000000000UL;
 920         }
 921         return 1;
 922 }
 923
 924 /*
 925  * sessiond monitoring thread: monitor presence of global and per-user
 926  * sessiond by polling the application common named pipe.
 927  */
 928 /* TODO */
 929
 930 void __attribute__((constructor)) lttng_ust_init(void)
 931 {
 932         struct timespec constructor_timeout;
 933         sigset_t sig_all_blocked, orig_parent_mask;
 934         int timeout_mode;
 935         int ret;
 936
 937         if (uatomic_xchg(&initialized, 1) == 1)
 938                 return;
 939
 940         /*
 941          * Fixup interdependency between TLS fixup mutex (which happens
 942          * to be the dynamic linker mutex) and ust_lock, taken within
 943          * the ust lock.
 944          */
 945         lttng_fixup_event_tls();
 946         lttng_fixup_ringbuffer_tls();
 947         lttng_fixup_vtid_tls();
 948         lttng_fixup_nest_count_tls();
 949
 950         /*
 951          * We want precise control over the order in which we construct
 952          * our sub-libraries vs starting to receive commands from
 953          * sessiond (otherwise leading to errors when trying to create
 954          * sessiond before the init functions are completed).
 955          */
 956         init_usterr();
 957         init_tracepoint();
 958         ltt_ring_buffer_metadata_client_init();
 959         ltt_ring_buffer_client_overwrite_init();
 960         ltt_ring_buffer_client_discard_init();
 961
 962         timeout_mode = get_timeout(&constructor_timeout);
 963
 964         ret = sem_init(&constructor_wait, 0, 0);
 965         assert(!ret);
 966
 967         ret = setup_local_apps();
 968         if (ret) {
 969                 ERR("Error setting up to local apps");
 970         }
 971
 972         /* A new thread created by pthread_create inherits the signal mask
 973          * from the parent. To avoid any signal being received by the
 974          * listener thread, we block all signals temporarily in the parent,
 975          * while we create the listener thread.
 976          */
 977         sigfillset(&sig_all_blocked);
 978         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 979         if (ret) {
 980                 ERR("pthread_sigmask: %s", strerror(ret));
 981         }
 982
 983         ret = pthread_create(&global_apps.ust_listener, NULL,
 984                         ust_listener_thread, &global_apps);
 985         if (ret) {
 986                 ERR("pthread_create global: %s", strerror(ret));
 987         }
 988         if (local_apps.allowed) {
 989                 ret = pthread_create(&local_apps.ust_listener, NULL,
 990                                 ust_listener_thread, &local_apps);
 991                 if (ret) {
 992                         ERR("pthread_create local: %s", strerror(ret));
 993                 }
 994         } else {
 995                 handle_register_done(&local_apps);
 996         }
 997
 998         /* Restore original signal mask in parent */
 999         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1000         if (ret) {
1001                 ERR("pthread_sigmask: %s", strerror(ret));
1002         }
1003
1004         switch (timeout_mode) {
1005         case 1: /* timeout wait */
1006                 do {
1007                         ret = sem_timedwait(&constructor_wait,
1008                                         &constructor_timeout);
1009                 } while (ret < 0 && errno == EINTR);
1010                 if (ret < 0 && errno == ETIMEDOUT) {
1011                         ERR("Timed out waiting for ltt-sessiond");
1012                 } else {
1013                         assert(!ret);
1014                 }
1015                 break;
1016         case -1:/* wait forever */
1017                 do {
1018                         ret = sem_wait(&constructor_wait);
1019                 } while (ret < 0 && errno == EINTR);
1020                 assert(!ret);
1021                 break;
1022         case 0: /* no timeout */
1023                 break;
1024         }
1025 }
1026
1027 static
1028 void lttng_ust_cleanup(int exiting)
1029 {
1030         cleanup_sock_info(&global_apps, exiting);
1031         if (local_apps.allowed) {
1032                 cleanup_sock_info(&local_apps, exiting);
1033         }
1034         /*
1035          * The teardown in this function all affect data structures
1036          * accessed under the UST lock by the listener thread. This
1037          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1038          * that none of these threads are accessing this data at this
1039          * point.
1040          */
1041         lttng_ust_abi_exit();
1042         lttng_ust_events_exit();
1043         ltt_ring_buffer_client_discard_exit();
1044         ltt_ring_buffer_client_overwrite_exit();
1045         ltt_ring_buffer_metadata_client_exit();
1046         exit_tracepoint();
1047         if (!exiting) {
1048                 /* Reinitialize values for fork */
1049                 sem_count = 2;
1050                 lttng_ust_comm_should_quit = 0;
1051                 initialized = 0;
1052         }
1053 }
1054
1055 void __attribute__((destructor)) lttng_ust_exit(void)
1056 {
1057         int ret;
1058
1059         /*
1060          * Using pthread_cancel here because:
1061          * A) we don't want to hang application teardown.
1062          * B) the thread is not allocating any resource.
1063          */
1064
1065         /*
1066          * Require the communication thread to quit. Synchronize with
1067          * mutexes to ensure it is not in a mutex critical section when
1068          * pthread_cancel is later called.
1069          */
1070         ust_lock();
1071         lttng_ust_comm_should_quit = 1;
1072         ust_unlock();
1073
1074         /* cancel threads */
1075         ret = pthread_cancel(global_apps.ust_listener);
1076         if (ret) {
1077                 ERR("Error cancelling global ust listener thread: %s",
1078                         strerror(ret));
1079         }
1080         if (local_apps.allowed) {
1081                 ret = pthread_cancel(local_apps.ust_listener);
1082                 if (ret) {
1083                         ERR("Error cancelling local ust listener thread: %s",
1084                                 strerror(ret));
1085                 }
1086         }
1087         /*
1088          * Do NOT join threads: use of sys_futex makes it impossible to
1089          * join the threads without using async-cancel, but async-cancel
1090          * is delivered by a signal, which could hit the target thread
1091          * anywhere in its code path, including while the ust_lock() is
1092          * held, causing a deadlock for the other thread. Let the OS
1093          * cleanup the threads if there are stalled in a syscall.
1094          */
1095         lttng_ust_cleanup(1);
1096 }
1097
1098 /*
1099  * We exclude the worker threads across fork and clone (except
1100  * CLONE_VM), because these system calls only keep the forking thread
1101  * running in the child.  Therefore, we don't want to call fork or clone
1102  * in the middle of an tracepoint or ust tracing state modification.
1103  * Holding this mutex protects these structures across fork and clone.
1104  */
1105 void ust_before_fork(sigset_t *save_sigset)
1106 {
1107         /*
1108          * Disable signals. This is to avoid that the child intervenes
1109          * before it is properly setup for tracing. It is safer to
1110          * disable all signals, because then we know we are not breaking
1111          * anything by restoring the original mask.
1112          */
1113         sigset_t all_sigs;
1114         int ret;
1115
1116         if (lttng_ust_nest_count)
1117                 return;
1118         /* Disable signals */
1119         sigfillset(&all_sigs);
1120         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1121         if (ret == -1) {
1122                 PERROR("sigprocmask");
1123         }
1124         ust_lock();
1125         rcu_bp_before_fork();
1126 }
1127
1128 static void ust_after_fork_common(sigset_t *restore_sigset)
1129 {
1130         int ret;
1131
1132         DBG("process %d", getpid());
1133         ust_unlock();
1134         /* Restore signals */
1135         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1136         if (ret == -1) {
1137                 PERROR("sigprocmask");
1138         }
1139 }
1140
1141 void ust_after_fork_parent(sigset_t *restore_sigset)
1142 {
1143         if (lttng_ust_nest_count)
1144                 return;
1145         DBG("process %d", getpid());
1146         rcu_bp_after_fork_parent();
1147         /* Release mutexes and reenable signals */
1148         ust_after_fork_common(restore_sigset);
1149 }
1150
1151 /*
1152  * After fork, in the child, we need to cleanup all the leftover state,
1153  * except the worker thread which already magically disappeared thanks
1154  * to the weird Linux fork semantics. After tyding up, we call
1155  * lttng_ust_init() again to start over as a new PID.
1156  *
1157  * This is meant for forks() that have tracing in the child between the
1158  * fork and following exec call (if there is any).
1159  */
1160 void ust_after_fork_child(sigset_t *restore_sigset)
1161 {
1162         if (lttng_ust_nest_count)
1163                 return;
1164         DBG("process %d", getpid());
1165         /* Release urcu mutexes */
1166         rcu_bp_after_fork_child();
1167         lttng_ust_cleanup(0);
1168         lttng_context_vtid_reset();
1169         /* Release mutexes and reenable signals */
1170         ust_after_fork_common(restore_sigset);
1171         lttng_ust_init();
1172 }