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
 297                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size - 1) {
 298                         ERR("Filter reloc offset %u is not within data\n",
 299                                 lum->u.filter.reloc_offset);
 300                         ret = -EINVAL;
 301                         goto error;
 302                 }
 303
 304                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 305                 if (!bytecode) {
 306                         ret = -ENOMEM;
 307                         goto error;
 308                 }
 309                 len = ustcomm_recv_unix_sock(sock, bytecode->data,
 310                                 lum->u.filter.data_size);
 311                 switch (len) {
 312                 case 0: /* orderly shutdown */
 313                         ret = 0;
 314                         free(bytecode);
 315                         goto error;
 316                 case -1:
 317                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 318                         if (errno == ECONNRESET) {
 319                                 ERR("%s remote end closed connection\n", sock_info->name);
 320                                 ret = -EINVAL;
 321                                 free(bytecode);
 322                                 goto error;
 323                         }
 324                         ret = -EINVAL;
 325                         goto end;
 326                 default:
 327                         if (len == lum->u.filter.data_size) {
 328                                 DBG("filter data received\n");
 329                                 break;
 330                         } else {
 331                                 ERR("incorrect filter data message size: %zd\n", len);
 332                                 ret = -EINVAL;
 333                                 free(bytecode);
 334                                 goto end;
 335                         }
 336                 }
 337                 bytecode->len = lum->u.filter.data_size;
 338                 bytecode->reloc_offset = lum->u.filter.reloc_offset;
 339                 if (ops->cmd) {
 340                         ret = ops->cmd(lum->handle, lum->cmd,
 341                                         (unsigned long) bytecode,
 342                                         &args);
 343                         if (ret) {
 344                                 free(bytecode);
 345                         }
 346                         /* don't free bytecode if everything went fine. */
 347                 } else {
 348                         ret = -ENOSYS;
 349                         free(bytecode);
 350                 }
 351                 break;
 352         }
 353         default:
 354                 if (ops->cmd)
 355                         ret = ops->cmd(lum->handle, lum->cmd,
 356                                         (unsigned long) &lum->u,
 357                                         &args);
 358                 else
 359                         ret = -ENOSYS;
 360                 break;
 361         }
 362
 363 end:
 364         lur.handle = lum->handle;
 365         lur.cmd = lum->cmd;
 366         lur.ret_val = ret;
 367         if (ret >= 0) {
 368                 lur.ret_code = USTCOMM_OK;
 369         } else {
 370                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 371                 lur.ret_code = ret;
 372         }
 373         if (ret >= 0) {
 374                 switch (lum->cmd) {
 375                 case LTTNG_UST_STREAM:
 376                         /*
 377                          * Special-case reply to send stream info.
 378                          * Use lum.u output.
 379                          */
 380                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 381                         shm_fd = *args.stream.shm_fd;
 382                         wait_fd = *args.stream.wait_fd;
 383                         break;
 384                 case LTTNG_UST_METADATA:
 385                 case LTTNG_UST_CHANNEL:
 386                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 387                         shm_fd = *args.channel.shm_fd;
 388                         wait_fd = *args.channel.wait_fd;
 389                         break;
 390                 case LTTNG_UST_TRACER_VERSION:
 391                         lur.u.version = lum->u.version;
 392                         break;
 393                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 394                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 395                         break;
 396                 }
 397         }
 398         ret = send_reply(sock, &lur);
 399         if (ret < 0) {
 400                 perror("error sending reply");
 401                 goto error;
 402         }
 403
 404         if ((lum->cmd == LTTNG_UST_STREAM
 405              || lum->cmd == LTTNG_UST_CHANNEL
 406              || lum->cmd == LTTNG_UST_METADATA)
 407                         && lur.ret_code == USTCOMM_OK) {
 408                 int sendret = 0;
 409
 410                 /* we also need to send the file descriptors. */
 411                 ret = ustcomm_send_fds_unix_sock(sock,
 412                         &shm_fd, &shm_fd,
 413                         1, sizeof(int));
 414                 if (ret < 0) {
 415                         perror("send shm_fd");
 416                         sendret = ret;
 417                 }
 418                 /*
 419                  * The sessiond expects 2 file descriptors, even upon
 420                  * error.
 421                  */
 422                 ret = ustcomm_send_fds_unix_sock(sock,
 423                         &wait_fd, &wait_fd,
 424                         1, sizeof(int));
 425                 if (ret < 0) {
 426                         perror("send wait_fd");
 427                         goto error;
 428                 }
 429                 if (sendret) {
 430                         ret = sendret;
 431                         goto error;
 432                 }
 433         }
 434         /*
 435          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 436          * after the reply.
 437          */
 438         if (lur.ret_code == USTCOMM_OK) {
 439                 switch (lum->cmd) {
 440                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 441                         len = ustcomm_send_unix_sock(sock,
 442                                 &args.field_list.entry,
 443                                 sizeof(args.field_list.entry));
 444                         if (len != sizeof(args.field_list.entry)) {
 445                                 ret = -1;
 446                                 goto error;
 447                         }
 448                 }
 449         }
 450         /*
 451          * We still have the memory map reference, and the fds have been
 452          * sent to the sessiond. We can therefore close those fds. Note
 453          * that we keep the write side of the wait_fd open, but close
 454          * the read side.
 455          */
 456         if (lur.ret_code == USTCOMM_OK) {
 457                 switch (lum->cmd) {
 458                 case LTTNG_UST_STREAM:
 459                         if (shm_fd >= 0) {
 460                                 ret = close(shm_fd);
 461                                 if (ret) {
 462                                         PERROR("Error closing stream shm_fd");
 463                                 }
 464                                 *args.stream.shm_fd = -1;
 465                         }
 466                         if (wait_fd >= 0) {
 467                                 ret = close(wait_fd);
 468                                 if (ret) {
 469                                         PERROR("Error closing stream wait_fd");
 470                                 }
 471                                 *args.stream.wait_fd = -1;
 472                         }
 473                         break;
 474                 case LTTNG_UST_METADATA:
 475                 case LTTNG_UST_CHANNEL:
 476                         if (shm_fd >= 0) {
 477                                 ret = close(shm_fd);
 478                                 if (ret) {
 479                                         PERROR("Error closing channel shm_fd");
 480                                 }
 481                                 *args.channel.shm_fd = -1;
 482                         }
 483                         if (wait_fd >= 0) {
 484                                 ret = close(wait_fd);
 485                                 if (ret) {
 486                                         PERROR("Error closing channel wait_fd");
 487                                 }
 488                                 *args.channel.wait_fd = -1;
 489                         }
 490                         break;
 491                 }
 492         }
 493
 494 error:
 495         ust_unlock();
 496         return ret;
 497 }
 498
 499 static
 500 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 501 {
 502         int ret;
 503
 504         if (sock_info->socket != -1) {
 505                 ret = ustcomm_close_unix_sock(sock_info->socket);
 506                 if (ret) {
 507                         ERR("Error closing apps socket");
 508                 }
 509                 sock_info->socket = -1;
 510         }
 511         if (sock_info->root_handle != -1) {
 512                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 513                 if (ret) {
 514                         ERR("Error unref root handle");
 515                 }
 516                 sock_info->root_handle = -1;
 517         }
 518         sock_info->constructor_sem_posted = 0;
 519         /*
 520          * wait_shm_mmap is used by listener threads outside of the
 521          * ust lock, so we cannot tear it down ourselves, because we
 522          * cannot join on these threads. Leave this task to the OS
 523          * process exit.
 524          */
 525         if (!exiting && sock_info->wait_shm_mmap) {
 526                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 527                 if (ret) {
 528                         ERR("Error unmapping wait shm");
 529                 }
 530                 sock_info->wait_shm_mmap = NULL;
 531         }
 532 }
 533
 534 /*
 535  * Using fork to set umask in the child process (not multi-thread safe).
 536  * We deal with the shm_open vs ftruncate race (happening when the
 537  * sessiond owns the shm and does not let everybody modify it, to ensure
 538  * safety against shm_unlink) by simply letting the mmap fail and
 539  * retrying after a few seconds.
 540  * For global shm, everybody has rw access to it until the sessiond
 541  * starts.
 542  */
 543 static
 544 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 545 {
 546         int wait_shm_fd, ret;
 547         pid_t pid;
 548
 549         /*
 550          * Try to open read-only.
 551          */
 552         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 553         if (wait_shm_fd >= 0) {
 554                 goto end;
 555         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 556                 /*
 557                  * Real-only open did not work, and it's not because the
 558                  * entry was not present. It's a failure that prohibits
 559                  * using shm.
 560                  */
 561                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 562                 goto end;
 563         }
 564         /*
 565          * If the open failed because the file did not exist, try
 566          * creating it ourself.
 567          */
 568         lttng_ust_nest_count++;
 569         pid = fork();
 570         lttng_ust_nest_count--;
 571         if (pid > 0) {
 572                 int status;
 573
 574                 /*
 575                  * Parent: wait for child to return, in which case the
 576                  * shared memory map will have been created.
 577                  */
 578                 pid = wait(&status);
 579                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 580                         wait_shm_fd = -1;
 581                         goto end;
 582                 }
 583                 /*
 584                  * Try to open read-only again after creation.
 585                  */
 586                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 587                 if (wait_shm_fd < 0) {
 588                         /*
 589                          * Real-only open did not work. It's a failure
 590                          * that prohibits using shm.
 591                          */
 592                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 593                         goto end;
 594                 }
 595                 goto end;
 596         } else if (pid == 0) {
 597                 int create_mode;
 598
 599                 /* Child */
 600                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 601                 if (sock_info->global)
 602                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 603                 /*
 604                  * We're alone in a child process, so we can modify the
 605                  * process-wide umask.
 606                  */
 607                 umask(~create_mode);
 608                 /*
 609                  * Try creating shm (or get rw access).
 610                  * We don't do an exclusive open, because we allow other
 611                  * processes to create+ftruncate it concurrently.
 612                  */
 613                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 614                                 O_RDWR | O_CREAT, create_mode);
 615                 if (wait_shm_fd >= 0) {
 616                         ret = ftruncate(wait_shm_fd, mmap_size);
 617                         if (ret) {
 618                                 PERROR("ftruncate");
 619                                 _exit(EXIT_FAILURE);
 620                         }
 621                         _exit(EXIT_SUCCESS);
 622                 }
 623                 /*
 624                  * For local shm, we need to have rw access to accept
 625                  * opening it: this means the local sessiond will be
 626                  * able to wake us up. For global shm, we open it even
 627                  * if rw access is not granted, because the root.root
 628                  * sessiond will be able to override all rights and wake
 629                  * us up.
 630                  */
 631                 if (!sock_info->global && errno != EACCES) {
 632                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 633                         _exit(EXIT_FAILURE);
 634                 }
 635                 /*
 636                  * The shm exists, but we cannot open it RW. Report
 637                  * success.
 638                  */
 639                 _exit(EXIT_SUCCESS);
 640         } else {
 641                 return -1;
 642         }
 643 end:
 644         if (wait_shm_fd >= 0 && !sock_info->global) {
 645                 struct stat statbuf;
 646
 647                 /*
 648                  * Ensure that our user is the owner of the shm file for
 649                  * local shm. If we do not own the file, it means our
 650                  * sessiond will not have access to wake us up (there is
 651                  * probably a rogue process trying to fake our
 652                  * sessiond). Fallback to polling method in this case.
 653                  */
 654                 ret = fstat(wait_shm_fd, &statbuf);
 655                 if (ret) {
 656                         PERROR("fstat");
 657                         goto error_close;
 658                 }
 659                 if (statbuf.st_uid != getuid())
 660                         goto error_close;
 661         }
 662         return wait_shm_fd;
 663
 664 error_close:
 665         ret = close(wait_shm_fd);
 666         if (ret) {
 667                 PERROR("Error closing fd");
 668         }
 669         return -1;
 670 }
 671
 672 static
 673 char *get_map_shm(struct sock_info *sock_info)
 674 {
 675         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 676         int wait_shm_fd, ret;
 677         char *wait_shm_mmap;
 678
 679         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 680         if (wait_shm_fd < 0) {
 681                 goto error;
 682         }
 683         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 684                   MAP_SHARED, wait_shm_fd, 0);
 685         /* close shm fd immediately after taking the mmap reference */
 686         ret = close(wait_shm_fd);
 687         if (ret) {
 688                 PERROR("Error closing fd");
 689         }
 690         if (wait_shm_mmap == MAP_FAILED) {
 691                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 692                 goto error;
 693         }
 694         return wait_shm_mmap;
 695
 696 error:
 697         return NULL;
 698 }
 699
 700 static
 701 void wait_for_sessiond(struct sock_info *sock_info)
 702 {
 703         int ret;
 704
 705         ust_lock();
 706         if (lttng_ust_comm_should_quit) {
 707                 goto quit;
 708         }
 709         if (wait_poll_fallback) {
 710                 goto error;
 711         }
 712         if (!sock_info->wait_shm_mmap) {
 713                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 714                 if (!sock_info->wait_shm_mmap)
 715                         goto error;
 716         }
 717         ust_unlock();
 718
 719         DBG("Waiting for %s apps sessiond", sock_info->name);
 720         /* Wait for futex wakeup */
 721         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 722                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 723                         FUTEX_WAIT, 0, NULL, NULL, 0);
 724                 if (ret < 0) {
 725                         if (errno == EFAULT) {
 726                                 wait_poll_fallback = 1;
 727                                 DBG(
 728 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 729 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 730 "Please upgrade your kernel "
 731 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 732 "mainline). LTTng-UST will use polling mode fallback.");
 733                                 if (ust_debug())
 734                                         PERROR("futex");
 735                         }
 736                 }
 737         }
 738         return;
 739
 740 quit:
 741         ust_unlock();
 742         return;
 743
 744 error:
 745         ust_unlock();
 746         return;
 747 }
 748
 749 /*
 750  * This thread does not allocate any resource, except within
 751  * handle_message, within mutex protection. This mutex protects against
 752  * fork and exit.
 753  * The other moment it allocates resources is at socket connection, which
 754  * is also protected by the mutex.
 755  */
 756 static
 757 void *ust_listener_thread(void *arg)
 758 {
 759         struct sock_info *sock_info = arg;
 760         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 761
 762         /* Restart trying to connect to the session daemon */
 763 restart:
 764         if (prev_connect_failed) {
 765                 /* Wait for sessiond availability with pipe */
 766                 wait_for_sessiond(sock_info);
 767                 if (has_waited) {
 768                         has_waited = 0;
 769                         /*
 770                          * Sleep for 5 seconds before retrying after a
 771                          * sequence of failure / wait / failure. This
 772                          * deals with a killed or broken session daemon.
 773                          */
 774                         sleep(5);
 775                 }
 776                 has_waited = 1;
 777                 prev_connect_failed = 0;
 778         }
 779         ust_lock();
 780
 781         if (lttng_ust_comm_should_quit) {
 782                 ust_unlock();
 783                 goto quit;
 784         }
 785
 786         if (sock_info->socket != -1) {
 787                 ret = ustcomm_close_unix_sock(sock_info->socket);
 788                 if (ret) {
 789                         ERR("Error closing %s apps socket", sock_info->name);
 790                 }
 791                 sock_info->socket = -1;
 792         }
 793
 794         /* Register */
 795         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 796         if (ret < 0) {
 797                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 798                 prev_connect_failed = 1;
 799                 /*
 800                  * If we cannot find the sessiond daemon, don't delay
 801                  * constructor execution.
 802                  */
 803                 ret = handle_register_done(sock_info);
 804                 assert(!ret);
 805                 ust_unlock();
 806                 goto restart;
 807         }
 808
 809         sock_info->socket = sock = ret;
 810
 811         /*
 812          * Create only one root handle per listener thread for the whole
 813          * process lifetime.
 814          */
 815         if (sock_info->root_handle == -1) {
 816                 ret = lttng_abi_create_root_handle();
 817                 if (ret < 0) {
 818                         ERR("Error creating root handle");
 819                         ust_unlock();
 820                         goto quit;
 821                 }
 822                 sock_info->root_handle = ret;
 823         }
 824
 825         ret = register_app_to_sessiond(sock);
 826         if (ret < 0) {
 827                 ERR("Error registering to %s apps socket", sock_info->name);
 828                 prev_connect_failed = 1;
 829                 /*
 830                  * If we cannot register to the sessiond daemon, don't
 831                  * delay constructor execution.
 832                  */
 833                 ret = handle_register_done(sock_info);
 834                 assert(!ret);
 835                 ust_unlock();
 836                 goto restart;
 837         }
 838         ust_unlock();
 839
 840         for (;;) {
 841                 ssize_t len;
 842                 struct ustcomm_ust_msg lum;
 843
 844                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 845                 switch (len) {
 846                 case 0: /* orderly shutdown */
 847                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 848                         ust_lock();
 849                         /*
 850                          * Either sessiond has shutdown or refused us by closing the socket.
 851                          * In either case, we don't want to delay construction execution,
 852                          * and we need to wait before retry.
 853                          */
 854                         prev_connect_failed = 1;
 855                         /*
 856                          * If we cannot register to the sessiond daemon, don't
 857                          * delay constructor execution.
 858                          */
 859                         ret = handle_register_done(sock_info);
 860                         assert(!ret);
 861                         ust_unlock();
 862                         goto end;
 863                 case sizeof(lum):
 864                         DBG("message received\n");
 865                         ret = handle_message(sock_info, sock, &lum);
 866                         if (ret < 0) {
 867                                 ERR("Error handling message for %s socket", sock_info->name);
 868                         }
 869                         continue;
 870                 case -1:
 871                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 872                         if (errno == ECONNRESET) {
 873                                 ERR("%s remote end closed connection\n", sock_info->name);
 874                                 goto end;
 875                         }
 876                         goto end;
 877                 default:
 878                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 879                         continue;
 880                 }
 881
 882         }
 883 end:
 884         goto restart;   /* try to reconnect */
 885 quit:
 886         return NULL;
 887 }
 888
 889 /*
 890  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 891  */
 892 static
 893 int get_timeout(struct timespec *constructor_timeout)
 894 {
 895         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 896         char *str_delay;
 897         int ret;
 898
 899         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 900         if (str_delay) {
 901                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 902         }
 903
 904         switch (constructor_delay_ms) {
 905         case -1:/* fall-through */
 906         case 0:
 907                 return constructor_delay_ms;
 908         default:
 909                 break;
 910         }
 911
 912         /*
 913          * If we are unable to find the current time, don't wait.
 914          */
 915         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 916         if (ret) {
 917                 return -1;
 918         }
 919         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 920         constructor_timeout->tv_nsec +=
 921                 (constructor_delay_ms % 1000UL) * 1000000UL;
 922         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 923                 constructor_timeout->tv_sec++;
 924                 constructor_timeout->tv_nsec -= 1000000000UL;
 925         }
 926         return 1;
 927 }
 928
 929 /*
 930  * sessiond monitoring thread: monitor presence of global and per-user
 931  * sessiond by polling the application common named pipe.
 932  */
 933 /* TODO */
 934
 935 void __attribute__((constructor)) lttng_ust_init(void)
 936 {
 937         struct timespec constructor_timeout;
 938         sigset_t sig_all_blocked, orig_parent_mask;
 939         pthread_attr_t thread_attr;
 940         int timeout_mode;
 941         int ret;
 942
 943         if (uatomic_xchg(&initialized, 1) == 1)
 944                 return;
 945
 946         /*
 947          * Fixup interdependency between TLS fixup mutex (which happens
 948          * to be the dynamic linker mutex) and ust_lock, taken within
 949          * the ust lock.
 950          */
 951         lttng_fixup_event_tls();
 952         lttng_fixup_ringbuffer_tls();
 953         lttng_fixup_vtid_tls();
 954         lttng_fixup_nest_count_tls();
 955
 956         /*
 957          * We want precise control over the order in which we construct
 958          * our sub-libraries vs starting to receive commands from
 959          * sessiond (otherwise leading to errors when trying to create
 960          * sessiond before the init functions are completed).
 961          */
 962         init_usterr();
 963         init_tracepoint();
 964         ltt_ring_buffer_metadata_client_init();
 965         ltt_ring_buffer_client_overwrite_init();
 966         ltt_ring_buffer_client_discard_init();
 967
 968         timeout_mode = get_timeout(&constructor_timeout);
 969
 970         ret = sem_init(&constructor_wait, 0, 0);
 971         assert(!ret);
 972
 973         ret = setup_local_apps();
 974         if (ret) {
 975                 ERR("Error setting up to local apps");
 976         }
 977
 978         /* A new thread created by pthread_create inherits the signal mask
 979          * from the parent. To avoid any signal being received by the
 980          * listener thread, we block all signals temporarily in the parent,
 981          * while we create the listener thread.
 982          */
 983         sigfillset(&sig_all_blocked);
 984         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 985         if (ret) {
 986                 ERR("pthread_sigmask: %s", strerror(ret));
 987         }
 988
 989         ret = pthread_attr_init(&thread_attr);
 990         if (ret) {
 991                 ERR("pthread_attr_init: %s", strerror(ret));
 992         }
 993         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
 994         if (ret) {
 995                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
 996         }
 997
 998         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
 999                         ust_listener_thread, &global_apps);
1000         if (ret) {
1001                 ERR("pthread_create global: %s", strerror(ret));
1002         }
1003         if (local_apps.allowed) {
1004                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1005                                 ust_listener_thread, &local_apps);
1006                 if (ret) {
1007                         ERR("pthread_create local: %s", strerror(ret));
1008                 }
1009         } else {
1010                 handle_register_done(&local_apps);
1011         }
1012         ret = pthread_attr_destroy(&thread_attr);
1013         if (ret) {
1014                 ERR("pthread_attr_destroy: %s", strerror(ret));
1015         }
1016
1017         /* Restore original signal mask in parent */
1018         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1019         if (ret) {
1020                 ERR("pthread_sigmask: %s", strerror(ret));
1021         }
1022
1023         switch (timeout_mode) {
1024         case 1: /* timeout wait */
1025                 do {
1026                         ret = sem_timedwait(&constructor_wait,
1027                                         &constructor_timeout);
1028                 } while (ret < 0 && errno == EINTR);
1029                 if (ret < 0 && errno == ETIMEDOUT) {
1030                         ERR("Timed out waiting for ltt-sessiond");
1031                 } else {
1032                         assert(!ret);
1033                 }
1034                 break;
1035         case -1:/* wait forever */
1036                 do {
1037                         ret = sem_wait(&constructor_wait);
1038                 } while (ret < 0 && errno == EINTR);
1039                 assert(!ret);
1040                 break;
1041         case 0: /* no timeout */
1042                 break;
1043         }
1044 }
1045
1046 static
1047 void lttng_ust_cleanup(int exiting)
1048 {
1049         cleanup_sock_info(&global_apps, exiting);
1050         if (local_apps.allowed) {
1051                 cleanup_sock_info(&local_apps, exiting);
1052         }
1053         /*
1054          * The teardown in this function all affect data structures
1055          * accessed under the UST lock by the listener thread. This
1056          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1057          * that none of these threads are accessing this data at this
1058          * point.
1059          */
1060         lttng_ust_abi_exit();
1061         lttng_ust_events_exit();
1062         ltt_ring_buffer_client_discard_exit();
1063         ltt_ring_buffer_client_overwrite_exit();
1064         ltt_ring_buffer_metadata_client_exit();
1065         exit_tracepoint();
1066         if (!exiting) {
1067                 /* Reinitialize values for fork */
1068                 sem_count = 2;
1069                 lttng_ust_comm_should_quit = 0;
1070                 initialized = 0;
1071         }
1072 }
1073
1074 void __attribute__((destructor)) lttng_ust_exit(void)
1075 {
1076         int ret;
1077
1078         /*
1079          * Using pthread_cancel here because:
1080          * A) we don't want to hang application teardown.
1081          * B) the thread is not allocating any resource.
1082          */
1083
1084         /*
1085          * Require the communication thread to quit. Synchronize with
1086          * mutexes to ensure it is not in a mutex critical section when
1087          * pthread_cancel is later called.
1088          */
1089         ust_lock();
1090         lttng_ust_comm_should_quit = 1;
1091         ust_unlock();
1092
1093         /* cancel threads */
1094         ret = pthread_cancel(global_apps.ust_listener);
1095         if (ret) {
1096                 ERR("Error cancelling global ust listener thread: %s",
1097                         strerror(ret));
1098         }
1099         if (local_apps.allowed) {
1100                 ret = pthread_cancel(local_apps.ust_listener);
1101                 if (ret) {
1102                         ERR("Error cancelling local ust listener thread: %s",
1103                                 strerror(ret));
1104                 }
1105         }
1106         /*
1107          * Do NOT join threads: use of sys_futex makes it impossible to
1108          * join the threads without using async-cancel, but async-cancel
1109          * is delivered by a signal, which could hit the target thread
1110          * anywhere in its code path, including while the ust_lock() is
1111          * held, causing a deadlock for the other thread. Let the OS
1112          * cleanup the threads if there are stalled in a syscall.
1113          */
1114         lttng_ust_cleanup(1);
1115 }
1116
1117 /*
1118  * We exclude the worker threads across fork and clone (except
1119  * CLONE_VM), because these system calls only keep the forking thread
1120  * running in the child.  Therefore, we don't want to call fork or clone
1121  * in the middle of an tracepoint or ust tracing state modification.
1122  * Holding this mutex protects these structures across fork and clone.
1123  */
1124 void ust_before_fork(sigset_t *save_sigset)
1125 {
1126         /*
1127          * Disable signals. This is to avoid that the child intervenes
1128          * before it is properly setup for tracing. It is safer to
1129          * disable all signals, because then we know we are not breaking
1130          * anything by restoring the original mask.
1131          */
1132         sigset_t all_sigs;
1133         int ret;
1134
1135         if (lttng_ust_nest_count)
1136                 return;
1137         /* Disable signals */
1138         sigfillset(&all_sigs);
1139         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1140         if (ret == -1) {
1141                 PERROR("sigprocmask");
1142         }
1143         ust_lock();
1144         rcu_bp_before_fork();
1145 }
1146
1147 static void ust_after_fork_common(sigset_t *restore_sigset)
1148 {
1149         int ret;
1150
1151         DBG("process %d", getpid());
1152         ust_unlock();
1153         /* Restore signals */
1154         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1155         if (ret == -1) {
1156                 PERROR("sigprocmask");
1157         }
1158 }
1159
1160 void ust_after_fork_parent(sigset_t *restore_sigset)
1161 {
1162         if (lttng_ust_nest_count)
1163                 return;
1164         DBG("process %d", getpid());
1165         rcu_bp_after_fork_parent();
1166         /* Release mutexes and reenable signals */
1167         ust_after_fork_common(restore_sigset);
1168 }
1169
1170 /*
1171  * After fork, in the child, we need to cleanup all the leftover state,
1172  * except the worker thread which already magically disappeared thanks
1173  * to the weird Linux fork semantics. After tyding up, we call
1174  * lttng_ust_init() again to start over as a new PID.
1175  *
1176  * This is meant for forks() that have tracing in the child between the
1177  * fork and following exec call (if there is any).
1178  */
1179 void ust_after_fork_child(sigset_t *restore_sigset)
1180 {
1181         if (lttng_ust_nest_count)
1182                 return;
1183         DBG("process %d", getpid());
1184         /* Release urcu mutexes */
1185         rcu_bp_after_fork_child();
1186         lttng_ust_cleanup(0);
1187         lttng_context_vtid_reset();
1188         /* Release mutexes and reenable signals */
1189         ust_after_fork_common(restore_sigset);
1190         lttng_ust_init();
1191 }