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 <lttng/ust-error.h>
  45 #include <ust-comm.h>
  46 #include <usterr-signal-safe.h>
  47 #include <helper.h>
  48 #include "tracepoint-internal.h"
  49 #include "ltt-tracer-core.h"
  50 #include "compat.h"
  51 #include "../libringbuffer/tlsfixup.h"
  52
  53 /*
  54  * Has lttng ust comm constructor been called ?
  55  */
  56 static int initialized;
  57
  58 /*
  59  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  60  * Held when handling a command, also held by fork() to deal with
  61  * removal of threads, and by exit path.
  62  */
  63
  64 /* Should the ust comm thread quit ? */
  65 static int lttng_ust_comm_should_quit;
  66
  67 /*
  68  * Wait for either of these before continuing to the main
  69  * program:
  70  * - the register_done message from sessiond daemon
  71  *   (will let the sessiond daemon enable sessions before main
  72  *   starts.)
  73  * - sessiond daemon is not reachable.
  74  * - timeout (ensuring applications are resilient to session
  75  *   daemon problems).
  76  */
  77 static sem_t constructor_wait;
  78 /*
  79  * Doing this for both the global and local sessiond.
  80  */
  81 static int sem_count = { 2 };
  82
  83 /*
  84  * Counting nesting within lttng-ust. Used to ensure that calling fork()
  85  * from liblttng-ust does not execute the pre/post fork handlers.
  86  */
  87 static int __thread lttng_ust_nest_count;
  88
  89 /*
  90  * Info about socket and associated listener thread.
  91  */
  92 struct sock_info {
  93         const char *name;
  94         pthread_t ust_listener; /* listener thread */
  95         int root_handle;
  96         int constructor_sem_posted;
  97         int allowed;
  98         int global;
  99
 100         char sock_path[PATH_MAX];
 101         int socket;
 102
 103         char wait_shm_path[PATH_MAX];
 104         char *wait_shm_mmap;
 105 };
 106
 107 /* Socket from app (connect) to session daemon (listen) for communication */
 108 struct sock_info global_apps = {
 109         .name = "global",
 110         .global = 1,
 111
 112         .root_handle = -1,
 113         .allowed = 1,
 114
 115         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 116         .socket = -1,
 117
 118         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 119 };
 120
 121 /* TODO: allow global_apps_sock_path override */
 122
 123 struct sock_info local_apps = {
 124         .name = "local",
 125         .global = 0,
 126         .root_handle = -1,
 127         .allowed = 0,   /* Check setuid bit first */
 128
 129         .socket = -1,
 130 };
 131
 132 static int wait_poll_fallback;
 133
 134 extern void ltt_ring_buffer_client_overwrite_init(void);
 135 extern void ltt_ring_buffer_client_discard_init(void);
 136 extern void ltt_ring_buffer_metadata_client_init(void);
 137 extern void ltt_ring_buffer_client_overwrite_exit(void);
 138 extern void ltt_ring_buffer_client_discard_exit(void);
 139 extern void ltt_ring_buffer_metadata_client_exit(void);
 140
 141 /*
 142  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 143  */
 144 static
 145 void lttng_fixup_nest_count_tls(void)
 146 {
 147         asm volatile ("" : : "m" (lttng_ust_nest_count));
 148 }
 149
 150 static
 151 int setup_local_apps(void)
 152 {
 153         const char *home_dir;
 154         uid_t uid;
 155
 156         uid = getuid();
 157         /*
 158          * Disallow per-user tracing for setuid binaries.
 159          */
 160         if (uid != geteuid()) {
 161                 assert(local_apps.allowed == 0);
 162                 return 0;
 163         }
 164         home_dir = (const char *) getenv("HOME");
 165         if (!home_dir) {
 166                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 167                 assert(local_apps.allowed == 0);
 168                 return -ENOENT;
 169         }
 170         local_apps.allowed = 1;
 171         snprintf(local_apps.sock_path, PATH_MAX,
 172                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 173         snprintf(local_apps.wait_shm_path, PATH_MAX,
 174                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 175         return 0;
 176 }
 177
 178 static
 179 int register_app_to_sessiond(int socket)
 180 {
 181         ssize_t ret;
 182         struct {
 183                 uint32_t major;
 184                 uint32_t minor;
 185                 pid_t pid;
 186                 pid_t ppid;
 187                 uid_t uid;
 188                 gid_t gid;
 189                 uint32_t bits_per_long;
 190                 char name[16];  /* process name */
 191         } reg_msg;
 192
 193         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 194         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 195         reg_msg.pid = getpid();
 196         reg_msg.ppid = getppid();
 197         reg_msg.uid = getuid();
 198         reg_msg.gid = getgid();
 199         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 200         lttng_ust_getprocname(reg_msg.name);
 201
 202         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 203         if (ret >= 0 && ret != sizeof(reg_msg))
 204                 return -EIO;
 205         return ret;
 206 }
 207
 208 static
 209 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 210 {
 211         ssize_t len;
 212
 213         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 214         switch (len) {
 215         case sizeof(*lur):
 216                 DBG("message successfully sent");
 217                 return 0;
 218         default:
 219                 if (len == -ECONNRESET) {
 220                         DBG("remote end closed connection");
 221                         return 0;
 222                 }
 223                 if (len < 0)
 224                         return len;
 225                 DBG("incorrect message size: %zd", len);
 226                 return -EINVAL;
 227         }
 228 }
 229
 230 static
 231 int handle_register_done(struct sock_info *sock_info)
 232 {
 233         int ret;
 234
 235         if (sock_info->constructor_sem_posted)
 236                 return 0;
 237         sock_info->constructor_sem_posted = 1;
 238         if (uatomic_read(&sem_count) <= 0) {
 239                 return 0;
 240         }
 241         ret = uatomic_add_return(&sem_count, -1);
 242         if (ret == 0) {
 243                 ret = sem_post(&constructor_wait);
 244                 assert(!ret);
 245         }
 246         return 0;
 247 }
 248
 249 static
 250 int handle_message(struct sock_info *sock_info,
 251                 int sock, struct ustcomm_ust_msg *lum)
 252 {
 253         int ret = 0;
 254         const struct lttng_ust_objd_ops *ops;
 255         struct ustcomm_ust_reply lur;
 256         int shm_fd, wait_fd;
 257         union ust_args args;
 258         ssize_t len;
 259
 260         ust_lock();
 261
 262         memset(&lur, 0, sizeof(lur));
 263
 264         if (lttng_ust_comm_should_quit) {
 265                 ret = -EPERM;
 266                 goto end;
 267         }
 268
 269         ops = objd_ops(lum->handle);
 270         if (!ops) {
 271                 ret = -ENOENT;
 272                 goto end;
 273         }
 274
 275         switch (lum->cmd) {
 276         case LTTNG_UST_REGISTER_DONE:
 277                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 278                         ret = handle_register_done(sock_info);
 279                 else
 280                         ret = -EINVAL;
 281                 break;
 282         case LTTNG_UST_RELEASE:
 283                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 284                         ret = -EPERM;
 285                 else
 286                         ret = lttng_ust_objd_unref(lum->handle);
 287                 break;
 288         case LTTNG_UST_FILTER:
 289         {
 290                 /* Receive filter data */
 291                 struct lttng_ust_filter_bytecode *bytecode;
 292
 293                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 294                         ERR("Filter data size is too large: %u bytes",
 295                                 lum->u.filter.data_size);
 296                         ret = -EINVAL;
 297                         goto error;
 298                 }
 299
 300                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 301                         ERR("Filter reloc offset %u is not within data",
 302                                 lum->u.filter.reloc_offset);
 303                         ret = -EINVAL;
 304                         goto error;
 305                 }
 306
 307                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 308                 if (!bytecode) {
 309                         ret = -ENOMEM;
 310                         goto error;
 311                 }
 312                 len = ustcomm_recv_unix_sock(sock, bytecode->data,
 313                                 lum->u.filter.data_size);
 314                 switch (len) {
 315                 case 0: /* orderly shutdown */
 316                         ret = 0;
 317                         free(bytecode);
 318                         goto error;
 319                 default:
 320                         if (len == lum->u.filter.data_size) {
 321                                 DBG("filter data received");
 322                                 break;
 323                         } else if (len < 0) {
 324                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 325                                 if (len == -ECONNRESET) {
 326                                         ERR("%s remote end closed connection", sock_info->name);
 327                                         ret = len;
 328                                         free(bytecode);
 329                                         goto error;
 330                                 }
 331                                 ret = len;
 332                                 goto end;
 333                         } else {
 334                                 DBG("incorrect filter data message size: %zd", len);
 335                                 ret = -EINVAL;
 336                                 free(bytecode);
 337                                 goto end;
 338                         }
 339                 }
 340                 bytecode->len = lum->u.filter.data_size;
 341                 bytecode->reloc_offset = lum->u.filter.reloc_offset;
 342                 if (ops->cmd) {
 343                         ret = ops->cmd(lum->handle, lum->cmd,
 344                                         (unsigned long) bytecode,
 345                                         &args);
 346                         if (ret) {
 347                                 free(bytecode);
 348                         }
 349                         /* don't free bytecode if everything went fine. */
 350                 } else {
 351                         ret = -ENOSYS;
 352                         free(bytecode);
 353                 }
 354                 break;
 355         }
 356         default:
 357                 if (ops->cmd)
 358                         ret = ops->cmd(lum->handle, lum->cmd,
 359                                         (unsigned long) &lum->u,
 360                                         &args);
 361                 else
 362                         ret = -ENOSYS;
 363                 break;
 364         }
 365
 366 end:
 367         lur.handle = lum->handle;
 368         lur.cmd = lum->cmd;
 369         lur.ret_val = ret;
 370         if (ret >= 0) {
 371                 lur.ret_code = LTTNG_UST_OK;
 372         } else {
 373                 /*
 374                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 375                  * error that are not caused by the transport, except if
 376                  * we already have a more precise error message to
 377                  * report.
 378                  */
 379                 if (ret > -LTTNG_UST_ERR)
 380                         lur.ret_code = -LTTNG_UST_ERR;
 381                 else
 382                         lur.ret_code = ret;
 383         }
 384         if (ret >= 0) {
 385                 switch (lum->cmd) {
 386                 case LTTNG_UST_STREAM:
 387                         /*
 388                          * Special-case reply to send stream info.
 389                          * Use lum.u output.
 390                          */
 391                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 392                         shm_fd = *args.stream.shm_fd;
 393                         wait_fd = *args.stream.wait_fd;
 394                         break;
 395                 case LTTNG_UST_METADATA:
 396                 case LTTNG_UST_CHANNEL:
 397                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 398                         shm_fd = *args.channel.shm_fd;
 399                         wait_fd = *args.channel.wait_fd;
 400                         break;
 401                 case LTTNG_UST_TRACER_VERSION:
 402                         lur.u.version = lum->u.version;
 403                         break;
 404                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 405                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 406                         break;
 407                 }
 408         }
 409         ret = send_reply(sock, &lur);
 410         if (ret < 0) {
 411                 DBG("error sending reply");
 412                 goto error;
 413         }
 414
 415         if ((lum->cmd == LTTNG_UST_STREAM
 416              || lum->cmd == LTTNG_UST_CHANNEL
 417              || lum->cmd == LTTNG_UST_METADATA)
 418                         && lur.ret_code == LTTNG_UST_OK) {
 419                 int sendret = 0;
 420
 421                 /* we also need to send the file descriptors. */
 422                 ret = ustcomm_send_fds_unix_sock(sock,
 423                         &shm_fd, &shm_fd,
 424                         1, sizeof(int));
 425                 if (ret < 0) {
 426                         ERR("send shm_fd");
 427                         sendret = ret;
 428                 }
 429                 /*
 430                  * The sessiond expects 2 file descriptors, even upon
 431                  * error.
 432                  */
 433                 ret = ustcomm_send_fds_unix_sock(sock,
 434                         &wait_fd, &wait_fd,
 435                         1, sizeof(int));
 436                 if (ret < 0) {
 437                         perror("send wait_fd");
 438                         goto error;
 439                 }
 440                 if (sendret) {
 441                         ret = sendret;
 442                         goto error;
 443                 }
 444         }
 445         /*
 446          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 447          * after the reply.
 448          */
 449         if (lur.ret_code == LTTNG_UST_OK) {
 450                 switch (lum->cmd) {
 451                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 452                         len = ustcomm_send_unix_sock(sock,
 453                                 &args.field_list.entry,
 454                                 sizeof(args.field_list.entry));
 455                         if (len < 0) {
 456                                 ret = len;
 457                                 goto error;
 458                         }
 459                         if (len != sizeof(args.field_list.entry)) {
 460                                 ret = -EINVAL;
 461                                 goto error;
 462                         }
 463                 }
 464         }
 465         /*
 466          * We still have the memory map reference, and the fds have been
 467          * sent to the sessiond. We can therefore close those fds. Note
 468          * that we keep the write side of the wait_fd open, but close
 469          * the read side.
 470          */
 471         if (lur.ret_code == LTTNG_UST_OK) {
 472                 switch (lum->cmd) {
 473                 case LTTNG_UST_STREAM:
 474                         if (shm_fd >= 0) {
 475                                 ret = close(shm_fd);
 476                                 if (ret) {
 477                                         PERROR("Error closing stream shm_fd");
 478                                 }
 479                                 *args.stream.shm_fd = -1;
 480                         }
 481                         if (wait_fd >= 0) {
 482                                 ret = close(wait_fd);
 483                                 if (ret) {
 484                                         PERROR("Error closing stream wait_fd");
 485                                 }
 486                                 *args.stream.wait_fd = -1;
 487                         }
 488                         break;
 489                 case LTTNG_UST_METADATA:
 490                 case LTTNG_UST_CHANNEL:
 491                         if (shm_fd >= 0) {
 492                                 ret = close(shm_fd);
 493                                 if (ret) {
 494                                         PERROR("Error closing channel shm_fd");
 495                                 }
 496                                 *args.channel.shm_fd = -1;
 497                         }
 498                         if (wait_fd >= 0) {
 499                                 ret = close(wait_fd);
 500                                 if (ret) {
 501                                         PERROR("Error closing channel wait_fd");
 502                                 }
 503                                 *args.channel.wait_fd = -1;
 504                         }
 505                         break;
 506                 }
 507         }
 508
 509 error:
 510         ust_unlock();
 511         return ret;
 512 }
 513
 514 static
 515 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 516 {
 517         int ret;
 518
 519         if (sock_info->socket != -1) {
 520                 ret = ustcomm_close_unix_sock(sock_info->socket);
 521                 if (ret) {
 522                         ERR("Error closing apps socket");
 523                 }
 524                 sock_info->socket = -1;
 525         }
 526         if (sock_info->root_handle != -1) {
 527                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 528                 if (ret) {
 529                         ERR("Error unref root handle");
 530                 }
 531                 sock_info->root_handle = -1;
 532         }
 533         sock_info->constructor_sem_posted = 0;
 534         /*
 535          * wait_shm_mmap is used by listener threads outside of the
 536          * ust lock, so we cannot tear it down ourselves, because we
 537          * cannot join on these threads. Leave this task to the OS
 538          * process exit.
 539          */
 540         if (!exiting && sock_info->wait_shm_mmap) {
 541                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 542                 if (ret) {
 543                         ERR("Error unmapping wait shm");
 544                 }
 545                 sock_info->wait_shm_mmap = NULL;
 546         }
 547 }
 548
 549 /*
 550  * Using fork to set umask in the child process (not multi-thread safe).
 551  * We deal with the shm_open vs ftruncate race (happening when the
 552  * sessiond owns the shm and does not let everybody modify it, to ensure
 553  * safety against shm_unlink) by simply letting the mmap fail and
 554  * retrying after a few seconds.
 555  * For global shm, everybody has rw access to it until the sessiond
 556  * starts.
 557  */
 558 static
 559 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 560 {
 561         int wait_shm_fd, ret;
 562         pid_t pid;
 563
 564         /*
 565          * Try to open read-only.
 566          */
 567         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 568         if (wait_shm_fd >= 0) {
 569                 goto end;
 570         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 571                 /*
 572                  * Real-only open did not work, and it's not because the
 573                  * entry was not present. It's a failure that prohibits
 574                  * using shm.
 575                  */
 576                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 577                 goto end;
 578         }
 579         /*
 580          * If the open failed because the file did not exist, try
 581          * creating it ourself.
 582          */
 583         lttng_ust_nest_count++;
 584         pid = fork();
 585         lttng_ust_nest_count--;
 586         if (pid > 0) {
 587                 int status;
 588
 589                 /*
 590                  * Parent: wait for child to return, in which case the
 591                  * shared memory map will have been created.
 592                  */
 593                 pid = wait(&status);
 594                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 595                         wait_shm_fd = -1;
 596                         goto end;
 597                 }
 598                 /*
 599                  * Try to open read-only again after creation.
 600                  */
 601                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 602                 if (wait_shm_fd < 0) {
 603                         /*
 604                          * Real-only open did not work. It's a failure
 605                          * that prohibits using shm.
 606                          */
 607                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 608                         goto end;
 609                 }
 610                 goto end;
 611         } else if (pid == 0) {
 612                 int create_mode;
 613
 614                 /* Child */
 615                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 616                 if (sock_info->global)
 617                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 618                 /*
 619                  * We're alone in a child process, so we can modify the
 620                  * process-wide umask.
 621                  */
 622                 umask(~create_mode);
 623                 /*
 624                  * Try creating shm (or get rw access).
 625                  * We don't do an exclusive open, because we allow other
 626                  * processes to create+ftruncate it concurrently.
 627                  */
 628                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 629                                 O_RDWR | O_CREAT, create_mode);
 630                 if (wait_shm_fd >= 0) {
 631                         ret = ftruncate(wait_shm_fd, mmap_size);
 632                         if (ret) {
 633                                 PERROR("ftruncate");
 634                                 _exit(EXIT_FAILURE);
 635                         }
 636                         _exit(EXIT_SUCCESS);
 637                 }
 638                 /*
 639                  * For local shm, we need to have rw access to accept
 640                  * opening it: this means the local sessiond will be
 641                  * able to wake us up. For global shm, we open it even
 642                  * if rw access is not granted, because the root.root
 643                  * sessiond will be able to override all rights and wake
 644                  * us up.
 645                  */
 646                 if (!sock_info->global && errno != EACCES) {
 647                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 648                         _exit(EXIT_FAILURE);
 649                 }
 650                 /*
 651                  * The shm exists, but we cannot open it RW. Report
 652                  * success.
 653                  */
 654                 _exit(EXIT_SUCCESS);
 655         } else {
 656                 return -1;
 657         }
 658 end:
 659         if (wait_shm_fd >= 0 && !sock_info->global) {
 660                 struct stat statbuf;
 661
 662                 /*
 663                  * Ensure that our user is the owner of the shm file for
 664                  * local shm. If we do not own the file, it means our
 665                  * sessiond will not have access to wake us up (there is
 666                  * probably a rogue process trying to fake our
 667                  * sessiond). Fallback to polling method in this case.
 668                  */
 669                 ret = fstat(wait_shm_fd, &statbuf);
 670                 if (ret) {
 671                         PERROR("fstat");
 672                         goto error_close;
 673                 }
 674                 if (statbuf.st_uid != getuid())
 675                         goto error_close;
 676         }
 677         return wait_shm_fd;
 678
 679 error_close:
 680         ret = close(wait_shm_fd);
 681         if (ret) {
 682                 PERROR("Error closing fd");
 683         }
 684         return -1;
 685 }
 686
 687 static
 688 char *get_map_shm(struct sock_info *sock_info)
 689 {
 690         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 691         int wait_shm_fd, ret;
 692         char *wait_shm_mmap;
 693
 694         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 695         if (wait_shm_fd < 0) {
 696                 goto error;
 697         }
 698         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 699                   MAP_SHARED, wait_shm_fd, 0);
 700         /* close shm fd immediately after taking the mmap reference */
 701         ret = close(wait_shm_fd);
 702         if (ret) {
 703                 PERROR("Error closing fd");
 704         }
 705         if (wait_shm_mmap == MAP_FAILED) {
 706                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 707                 goto error;
 708         }
 709         return wait_shm_mmap;
 710
 711 error:
 712         return NULL;
 713 }
 714
 715 static
 716 void wait_for_sessiond(struct sock_info *sock_info)
 717 {
 718         int ret;
 719
 720         ust_lock();
 721         if (lttng_ust_comm_should_quit) {
 722                 goto quit;
 723         }
 724         if (wait_poll_fallback) {
 725                 goto error;
 726         }
 727         if (!sock_info->wait_shm_mmap) {
 728                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 729                 if (!sock_info->wait_shm_mmap)
 730                         goto error;
 731         }
 732         ust_unlock();
 733
 734         DBG("Waiting for %s apps sessiond", sock_info->name);
 735         /* Wait for futex wakeup */
 736         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 737                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 738                         FUTEX_WAIT, 0, NULL, NULL, 0);
 739                 if (ret < 0) {
 740                         if (errno == EFAULT) {
 741                                 wait_poll_fallback = 1;
 742                                 DBG(
 743 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 744 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 745 "Please upgrade your kernel "
 746 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 747 "mainline). LTTng-UST will use polling mode fallback.");
 748                                 if (ust_debug())
 749                                         PERROR("futex");
 750                         }
 751                 }
 752         }
 753         return;
 754
 755 quit:
 756         ust_unlock();
 757         return;
 758
 759 error:
 760         ust_unlock();
 761         return;
 762 }
 763
 764 /*
 765  * This thread does not allocate any resource, except within
 766  * handle_message, within mutex protection. This mutex protects against
 767  * fork and exit.
 768  * The other moment it allocates resources is at socket connection, which
 769  * is also protected by the mutex.
 770  */
 771 static
 772 void *ust_listener_thread(void *arg)
 773 {
 774         struct sock_info *sock_info = arg;
 775         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 776
 777         /* Restart trying to connect to the session daemon */
 778 restart:
 779         if (prev_connect_failed) {
 780                 /* Wait for sessiond availability with pipe */
 781                 wait_for_sessiond(sock_info);
 782                 if (has_waited) {
 783                         has_waited = 0;
 784                         /*
 785                          * Sleep for 5 seconds before retrying after a
 786                          * sequence of failure / wait / failure. This
 787                          * deals with a killed or broken session daemon.
 788                          */
 789                         sleep(5);
 790                 }
 791                 has_waited = 1;
 792                 prev_connect_failed = 0;
 793         }
 794         ust_lock();
 795
 796         if (lttng_ust_comm_should_quit) {
 797                 ust_unlock();
 798                 goto quit;
 799         }
 800
 801         if (sock_info->socket != -1) {
 802                 ret = ustcomm_close_unix_sock(sock_info->socket);
 803                 if (ret) {
 804                         ERR("Error closing %s apps socket", sock_info->name);
 805                 }
 806                 sock_info->socket = -1;
 807         }
 808
 809         /* Register */
 810         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 811         if (ret < 0) {
 812                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 813                 prev_connect_failed = 1;
 814                 /*
 815                  * If we cannot find the sessiond daemon, don't delay
 816                  * constructor execution.
 817                  */
 818                 ret = handle_register_done(sock_info);
 819                 assert(!ret);
 820                 ust_unlock();
 821                 goto restart;
 822         }
 823
 824         sock_info->socket = sock = ret;
 825
 826         /*
 827          * Create only one root handle per listener thread for the whole
 828          * process lifetime.
 829          */
 830         if (sock_info->root_handle == -1) {
 831                 ret = lttng_abi_create_root_handle();
 832                 if (ret < 0) {
 833                         ERR("Error creating root handle");
 834                         ust_unlock();
 835                         goto quit;
 836                 }
 837                 sock_info->root_handle = ret;
 838         }
 839
 840         ret = register_app_to_sessiond(sock);
 841         if (ret < 0) {
 842                 ERR("Error registering to %s apps socket", sock_info->name);
 843                 prev_connect_failed = 1;
 844                 /*
 845                  * If we cannot register to the sessiond daemon, don't
 846                  * delay constructor execution.
 847                  */
 848                 ret = handle_register_done(sock_info);
 849                 assert(!ret);
 850                 ust_unlock();
 851                 goto restart;
 852         }
 853         ust_unlock();
 854
 855         for (;;) {
 856                 ssize_t len;
 857                 struct ustcomm_ust_msg lum;
 858
 859                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 860                 switch (len) {
 861                 case 0: /* orderly shutdown */
 862                         DBG("%s ltt-sessiond has performed an orderly shutdown", sock_info->name);
 863                         ust_lock();
 864                         /*
 865                          * Either sessiond has shutdown or refused us by closing the socket.
 866                          * In either case, we don't want to delay construction execution,
 867                          * and we need to wait before retry.
 868                          */
 869                         prev_connect_failed = 1;
 870                         /*
 871                          * If we cannot register to the sessiond daemon, don't
 872                          * delay constructor execution.
 873                          */
 874                         ret = handle_register_done(sock_info);
 875                         assert(!ret);
 876                         ust_unlock();
 877                         goto end;
 878                 case sizeof(lum):
 879                         DBG("message received");
 880                         ret = handle_message(sock_info, sock, &lum);
 881                         if (ret) {
 882                                 ERR("Error handling message for %s socket", sock_info->name);
 883                         }
 884                         continue;
 885                 default:
 886                         if (len < 0) {
 887                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 888                         } else {
 889                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
 890                         }
 891                         if (len == -ECONNRESET) {
 892                                 DBG("%s remote end closed connection", sock_info->name);
 893                                 goto end;
 894                         }
 895                         goto end;
 896                 }
 897
 898         }
 899 end:
 900         goto restart;   /* try to reconnect */
 901 quit:
 902         return NULL;
 903 }
 904
 905 /*
 906  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 907  */
 908 static
 909 int get_timeout(struct timespec *constructor_timeout)
 910 {
 911         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 912         char *str_delay;
 913         int ret;
 914
 915         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 916         if (str_delay) {
 917                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 918         }
 919
 920         switch (constructor_delay_ms) {
 921         case -1:/* fall-through */
 922         case 0:
 923                 return constructor_delay_ms;
 924         default:
 925                 break;
 926         }
 927
 928         /*
 929          * If we are unable to find the current time, don't wait.
 930          */
 931         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 932         if (ret) {
 933                 return -1;
 934         }
 935         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 936         constructor_timeout->tv_nsec +=
 937                 (constructor_delay_ms % 1000UL) * 1000000UL;
 938         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 939                 constructor_timeout->tv_sec++;
 940                 constructor_timeout->tv_nsec -= 1000000000UL;
 941         }
 942         return 1;
 943 }
 944
 945 /*
 946  * sessiond monitoring thread: monitor presence of global and per-user
 947  * sessiond by polling the application common named pipe.
 948  */
 949 void __attribute__((constructor)) lttng_ust_init(void)
 950 {
 951         struct timespec constructor_timeout;
 952         sigset_t sig_all_blocked, orig_parent_mask;
 953         pthread_attr_t thread_attr;
 954         int timeout_mode;
 955         int ret;
 956
 957         if (uatomic_xchg(&initialized, 1) == 1)
 958                 return;
 959
 960         /*
 961          * Fixup interdependency between TLS fixup mutex (which happens
 962          * to be the dynamic linker mutex) and ust_lock, taken within
 963          * the ust lock.
 964          */
 965         lttng_fixup_event_tls();
 966         lttng_fixup_ringbuffer_tls();
 967         lttng_fixup_vtid_tls();
 968         lttng_fixup_nest_count_tls();
 969         lttng_fixup_procname_tls();
 970
 971         /*
 972          * We want precise control over the order in which we construct
 973          * our sub-libraries vs starting to receive commands from
 974          * sessiond (otherwise leading to errors when trying to create
 975          * sessiond before the init functions are completed).
 976          */
 977         init_usterr();
 978         init_tracepoint();
 979         ltt_ring_buffer_metadata_client_init();
 980         ltt_ring_buffer_client_overwrite_init();
 981         ltt_ring_buffer_client_discard_init();
 982
 983         timeout_mode = get_timeout(&constructor_timeout);
 984
 985         ret = sem_init(&constructor_wait, 0, 0);
 986         assert(!ret);
 987
 988         ret = setup_local_apps();
 989         if (ret) {
 990                 DBG("local apps setup returned %d", ret);
 991         }
 992
 993         /* A new thread created by pthread_create inherits the signal mask
 994          * from the parent. To avoid any signal being received by the
 995          * listener thread, we block all signals temporarily in the parent,
 996          * while we create the listener thread.
 997          */
 998         sigfillset(&sig_all_blocked);
 999         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1000         if (ret) {
1001                 ERR("pthread_sigmask: %s", strerror(ret));
1002         }
1003
1004         ret = pthread_attr_init(&thread_attr);
1005         if (ret) {
1006                 ERR("pthread_attr_init: %s", strerror(ret));
1007         }
1008         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1009         if (ret) {
1010                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1011         }
1012
1013         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1014                         ust_listener_thread, &global_apps);
1015         if (ret) {
1016                 ERR("pthread_create global: %s", strerror(ret));
1017         }
1018         if (local_apps.allowed) {
1019                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1020                                 ust_listener_thread, &local_apps);
1021                 if (ret) {
1022                         ERR("pthread_create local: %s", strerror(ret));
1023                 }
1024         } else {
1025                 handle_register_done(&local_apps);
1026         }
1027         ret = pthread_attr_destroy(&thread_attr);
1028         if (ret) {
1029                 ERR("pthread_attr_destroy: %s", strerror(ret));
1030         }
1031
1032         /* Restore original signal mask in parent */
1033         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1034         if (ret) {
1035                 ERR("pthread_sigmask: %s", strerror(ret));
1036         }
1037
1038         switch (timeout_mode) {
1039         case 1: /* timeout wait */
1040                 do {
1041                         ret = sem_timedwait(&constructor_wait,
1042                                         &constructor_timeout);
1043                 } while (ret < 0 && errno == EINTR);
1044                 if (ret < 0 && errno == ETIMEDOUT) {
1045                         ERR("Timed out waiting for ltt-sessiond");
1046                 } else {
1047                         assert(!ret);
1048                 }
1049                 break;
1050         case -1:/* wait forever */
1051                 do {
1052                         ret = sem_wait(&constructor_wait);
1053                 } while (ret < 0 && errno == EINTR);
1054                 assert(!ret);
1055                 break;
1056         case 0: /* no timeout */
1057                 break;
1058         }
1059 }
1060
1061 static
1062 void lttng_ust_cleanup(int exiting)
1063 {
1064         cleanup_sock_info(&global_apps, exiting);
1065         if (local_apps.allowed) {
1066                 cleanup_sock_info(&local_apps, exiting);
1067         }
1068         /*
1069          * The teardown in this function all affect data structures
1070          * accessed under the UST lock by the listener thread. This
1071          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1072          * that none of these threads are accessing this data at this
1073          * point.
1074          */
1075         lttng_ust_abi_exit();
1076         lttng_ust_events_exit();
1077         ltt_ring_buffer_client_discard_exit();
1078         ltt_ring_buffer_client_overwrite_exit();
1079         ltt_ring_buffer_metadata_client_exit();
1080         exit_tracepoint();
1081         if (!exiting) {
1082                 /* Reinitialize values for fork */
1083                 sem_count = 2;
1084                 lttng_ust_comm_should_quit = 0;
1085                 initialized = 0;
1086         }
1087 }
1088
1089 void __attribute__((destructor)) lttng_ust_exit(void)
1090 {
1091         int ret;
1092
1093         /*
1094          * Using pthread_cancel here because:
1095          * A) we don't want to hang application teardown.
1096          * B) the thread is not allocating any resource.
1097          */
1098
1099         /*
1100          * Require the communication thread to quit. Synchronize with
1101          * mutexes to ensure it is not in a mutex critical section when
1102          * pthread_cancel is later called.
1103          */
1104         ust_lock();
1105         lttng_ust_comm_should_quit = 1;
1106         ust_unlock();
1107
1108         /* cancel threads */
1109         ret = pthread_cancel(global_apps.ust_listener);
1110         if (ret) {
1111                 ERR("Error cancelling global ust listener thread: %s",
1112                         strerror(ret));
1113         }
1114         if (local_apps.allowed) {
1115                 ret = pthread_cancel(local_apps.ust_listener);
1116                 if (ret) {
1117                         ERR("Error cancelling local ust listener thread: %s",
1118                                 strerror(ret));
1119                 }
1120         }
1121         /*
1122          * Do NOT join threads: use of sys_futex makes it impossible to
1123          * join the threads without using async-cancel, but async-cancel
1124          * is delivered by a signal, which could hit the target thread
1125          * anywhere in its code path, including while the ust_lock() is
1126          * held, causing a deadlock for the other thread. Let the OS
1127          * cleanup the threads if there are stalled in a syscall.
1128          */
1129         lttng_ust_cleanup(1);
1130 }
1131
1132 /*
1133  * We exclude the worker threads across fork and clone (except
1134  * CLONE_VM), because these system calls only keep the forking thread
1135  * running in the child.  Therefore, we don't want to call fork or clone
1136  * in the middle of an tracepoint or ust tracing state modification.
1137  * Holding this mutex protects these structures across fork and clone.
1138  */
1139 void ust_before_fork(sigset_t *save_sigset)
1140 {
1141         /*
1142          * Disable signals. This is to avoid that the child intervenes
1143          * before it is properly setup for tracing. It is safer to
1144          * disable all signals, because then we know we are not breaking
1145          * anything by restoring the original mask.
1146          */
1147         sigset_t all_sigs;
1148         int ret;
1149
1150         if (lttng_ust_nest_count)
1151                 return;
1152         /* Disable signals */
1153         sigfillset(&all_sigs);
1154         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1155         if (ret == -1) {
1156                 PERROR("sigprocmask");
1157         }
1158         ust_lock();
1159         rcu_bp_before_fork();
1160 }
1161
1162 static void ust_after_fork_common(sigset_t *restore_sigset)
1163 {
1164         int ret;
1165
1166         DBG("process %d", getpid());
1167         ust_unlock();
1168         /* Restore signals */
1169         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1170         if (ret == -1) {
1171                 PERROR("sigprocmask");
1172         }
1173 }
1174
1175 void ust_after_fork_parent(sigset_t *restore_sigset)
1176 {
1177         if (lttng_ust_nest_count)
1178                 return;
1179         DBG("process %d", getpid());
1180         rcu_bp_after_fork_parent();
1181         /* Release mutexes and reenable signals */
1182         ust_after_fork_common(restore_sigset);
1183 }
1184
1185 /*
1186  * After fork, in the child, we need to cleanup all the leftover state,
1187  * except the worker thread which already magically disappeared thanks
1188  * to the weird Linux fork semantics. After tyding up, we call
1189  * lttng_ust_init() again to start over as a new PID.
1190  *
1191  * This is meant for forks() that have tracing in the child between the
1192  * fork and following exec call (if there is any).
1193  */
1194 void ust_after_fork_child(sigset_t *restore_sigset)
1195 {
1196         if (lttng_ust_nest_count)
1197                 return;
1198         DBG("process %d", getpid());
1199         /* Release urcu mutexes */
1200         rcu_bp_after_fork_child();
1201         lttng_ust_cleanup(0);
1202         lttng_context_vtid_reset();
1203         /* Release mutexes and reenable signals */
1204         ust_after_fork_common(restore_sigset);
1205         lttng_ust_init();
1206 }