liblttng-ust/lttng-ust-comm.c

   1 /*
   2  * lttng-ust-comm.c
   3  *
   4  * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
   5  * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   6  *
   7  * This library is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; only
  10  * version 2.1 of the License.
  11  *
  12  * This library is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with this library; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #define _LGPL_SOURCE
  23 #include <sys/types.h>
  24 #include <sys/socket.h>
  25 #include <sys/mman.h>
  26 #include <sys/stat.h>
  27 #include <sys/types.h>
  28 #include <sys/wait.h>
  29 #include <fcntl.h>
  30 #include <unistd.h>
  31 #include <errno.h>
  32 #include <pthread.h>
  33 #include <semaphore.h>
  34 #include <time.h>
  35 #include <assert.h>
  36 #include <signal.h>
  37 #include <urcu/uatomic.h>
  38 #include <urcu/futex.h>
  39 #include <urcu/compiler.h>
  40
  41 #include <lttng/ust-events.h>
  42 #include <lttng/ust-abi.h>
  43 #include <lttng/ust.h>
  44 #include <ust-comm.h>
  45 #include <usterr-signal-safe.h>
  46 #include <helper.h>
  47 #include "tracepoint-internal.h"
  48 #include "ltt-tracer-core.h"
  49 #include "compat.h"
  50 #include "../libringbuffer/tlsfixup.h"
  51
  52 /*
  53  * Has lttng ust comm constructor been called ?
  54  */
  55 static int initialized;
  56
  57 /*
  58  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  59  * Held when handling a command, also held by fork() to deal with
  60  * removal of threads, and by exit path.
  61  */
  62
  63 /* Should the ust comm thread quit ? */
  64 static int lttng_ust_comm_should_quit;
  65
  66 /*
  67  * Wait for either of these before continuing to the main
  68  * program:
  69  * - the register_done message from sessiond daemon
  70  *   (will let the sessiond daemon enable sessions before main
  71  *   starts.)
  72  * - sessiond daemon is not reachable.
  73  * - timeout (ensuring applications are resilient to session
  74  *   daemon problems).
  75  */
  76 static sem_t constructor_wait;
  77 /*
  78  * Doing this for both the global and local sessiond.
  79  */
  80 static int sem_count = { 2 };
  81
  82 /*
  83  * Counting nesting within lttng-ust. Used to ensure that calling fork()
  84  * from liblttng-ust does not execute the pre/post fork handlers.
  85  */
  86 static int __thread lttng_ust_nest_count;
  87
  88 /*
  89  * Info about socket and associated listener thread.
  90  */
  91 struct sock_info {
  92         const char *name;
  93         pthread_t ust_listener; /* listener thread */
  94         int root_handle;
  95         int constructor_sem_posted;
  96         int allowed;
  97         int global;
  98
  99         char sock_path[PATH_MAX];
 100         int socket;
 101
 102         char wait_shm_path[PATH_MAX];
 103         char *wait_shm_mmap;
 104 };
 105
 106 /* Socket from app (connect) to session daemon (listen) for communication */
 107 struct sock_info global_apps = {
 108         .name = "global",
 109         .global = 1,
 110
 111         .root_handle = -1,
 112         .allowed = 1,
 113
 114         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 115         .socket = -1,
 116
 117         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 118 };
 119
 120 /* TODO: allow global_apps_sock_path override */
 121
 122 struct sock_info local_apps = {
 123         .name = "local",
 124         .global = 0,
 125         .root_handle = -1,
 126         .allowed = 0,   /* Check setuid bit first */
 127
 128         .socket = -1,
 129 };
 130
 131 static int wait_poll_fallback;
 132
 133 extern void ltt_ring_buffer_client_overwrite_init(void);
 134 extern void ltt_ring_buffer_client_discard_init(void);
 135 extern void ltt_ring_buffer_metadata_client_init(void);
 136 extern void ltt_ring_buffer_client_overwrite_exit(void);
 137 extern void ltt_ring_buffer_client_discard_exit(void);
 138 extern void ltt_ring_buffer_metadata_client_exit(void);
 139
 140 /*
 141  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 142  */
 143 static
 144 void lttng_fixup_nest_count_tls(void)
 145 {
 146         asm volatile ("" : : "m" (lttng_ust_nest_count));
 147 }
 148
 149 static
 150 int setup_local_apps(void)
 151 {
 152         const char *home_dir;
 153         uid_t uid;
 154
 155         uid = getuid();
 156         /*
 157          * Disallow per-user tracing for setuid binaries.
 158          */
 159         if (uid != geteuid()) {
 160                 local_apps.allowed = 0;
 161                 return 0;
 162         } else {
 163                 local_apps.allowed = 1;
 164         }
 165         home_dir = (const char *) getenv("HOME");
 166         if (!home_dir)
 167                 return -ENOENT;
 168         snprintf(local_apps.sock_path, PATH_MAX,
 169                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 170         snprintf(local_apps.wait_shm_path, PATH_MAX,
 171                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 172         return 0;
 173 }
 174
 175 static
 176 int register_app_to_sessiond(int socket)
 177 {
 178         ssize_t ret;
 179         struct {
 180                 uint32_t major;
 181                 uint32_t minor;
 182                 pid_t pid;
 183                 pid_t ppid;
 184                 uid_t uid;
 185                 gid_t gid;
 186                 uint32_t bits_per_long;
 187                 char name[16];  /* process name */
 188         } reg_msg;
 189
 190         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 191         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 192         reg_msg.pid = getpid();
 193         reg_msg.ppid = getppid();
 194         reg_msg.uid = getuid();
 195         reg_msg.gid = getgid();
 196         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 197         lttng_ust_getprocname(reg_msg.name);
 198
 199         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 200         if (ret >= 0 && ret != sizeof(reg_msg))
 201                 return -EIO;
 202         return ret;
 203 }
 204
 205 static
 206 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 207 {
 208         ssize_t len;
 209
 210         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 211         switch (len) {
 212         case sizeof(*lur):
 213                 DBG("message successfully sent");
 214                 return 0;
 215         case -1:
 216                 if (errno == ECONNRESET) {
 217                         printf("remote end closed connection\n");
 218                         return 0;
 219                 }
 220                 return -1;
 221         default:
 222                 printf("incorrect message size: %zd\n", len);
 223                 return -1;
 224         }
 225 }
 226
 227 static
 228 int handle_register_done(struct sock_info *sock_info)
 229 {
 230         int ret;
 231
 232         if (sock_info->constructor_sem_posted)
 233                 return 0;
 234         sock_info->constructor_sem_posted = 1;
 235         if (uatomic_read(&sem_count) <= 0) {
 236                 return 0;
 237         }
 238         ret = uatomic_add_return(&sem_count, -1);
 239         if (ret == 0) {
 240                 ret = sem_post(&constructor_wait);
 241                 assert(!ret);
 242         }
 243         return 0;
 244 }
 245
 246 static
 247 int handle_message(struct sock_info *sock_info,
 248                 int sock, struct ustcomm_ust_msg *lum)
 249 {
 250         int ret = 0;
 251         const struct lttng_ust_objd_ops *ops;
 252         struct ustcomm_ust_reply lur;
 253         int shm_fd, wait_fd;
 254         union ust_args args;
 255         ssize_t len;
 256
 257         ust_lock();
 258
 259         memset(&lur, 0, sizeof(lur));
 260
 261         if (lttng_ust_comm_should_quit) {
 262                 ret = -EPERM;
 263                 goto end;
 264         }
 265
 266         ops = objd_ops(lum->handle);
 267         if (!ops) {
 268                 ret = -ENOENT;
 269                 goto end;
 270         }
 271
 272         switch (lum->cmd) {
 273         case LTTNG_UST_REGISTER_DONE:
 274                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 275                         ret = handle_register_done(sock_info);
 276                 else
 277                         ret = -EINVAL;
 278                 break;
 279         case LTTNG_UST_RELEASE:
 280                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 281                         ret = -EPERM;
 282                 else
 283                         ret = lttng_ust_objd_unref(lum->handle);
 284                 break;
 285         case LTTNG_UST_FILTER:
 286         {
 287                 /* Receive filter data */
 288                 struct lttng_ust_filter_bytecode *bytecode;
 289
 290                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 291                         ERR("Filter data size is too large: %u bytes\n",
 292                                 lum->u.filter.data_size);
 293                         ret = -EINVAL;
 294                         goto error;
 295                 }
 296                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 297                 if (!bytecode) {
 298                         ret = -ENOMEM;
 299                         goto error;
 300                 }
 301                 len = ustcomm_recv_unix_sock(sock, bytecode->data,
 302                                 lum->u.filter.data_size);
 303                 switch (len) {
 304                 case 0: /* orderly shutdown */
 305                         ret = 0;
 306                         free(bytecode);
 307                         goto error;
 308                 case -1:
 309                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 310                         if (errno == ECONNRESET) {
 311                                 ERR("%s remote end closed connection\n", sock_info->name);
 312                                 ret = -EINVAL;
 313                                 free(bytecode);
 314                                 goto error;
 315                         }
 316                         ret = -EINVAL;
 317                         goto end;
 318                 default:
 319                         if (len == lum->u.filter.data_size) {
 320                                 DBG("filter data received\n");
 321                                 break;
 322                         } else {
 323                                 ERR("incorrect filter data message size: %zd\n", len);
 324                                 ret = -EINVAL;
 325                                 free(bytecode);
 326                                 goto end;
 327                         }
 328                 }
 329                 bytecode->len = lum->u.filter.data_size;
 330                 bytecode->reloc_offset = lum->u.filter.reloc_offset;
 331                 if (ops->cmd) {
 332                         ret = ops->cmd(lum->handle, lum->cmd,
 333                                         (unsigned long) bytecode,
 334                                         &args);
 335                         if (ret) {
 336                                 free(bytecode);
 337                         }
 338                         /* don't free bytecode if everything went fine. */
 339                 } else {
 340                         ret = -ENOSYS;
 341                         free(bytecode);
 342                 }
 343                 break;
 344         }
 345         default:
 346                 if (ops->cmd)
 347                         ret = ops->cmd(lum->handle, lum->cmd,
 348                                         (unsigned long) &lum->u,
 349                                         &args);
 350                 else
 351                         ret = -ENOSYS;
 352                 break;
 353         }
 354
 355 end:
 356         lur.handle = lum->handle;
 357         lur.cmd = lum->cmd;
 358         lur.ret_val = ret;
 359         if (ret >= 0) {
 360                 lur.ret_code = USTCOMM_OK;
 361         } else {
 362                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 363                 lur.ret_code = ret;
 364         }
 365         if (ret >= 0) {
 366                 switch (lum->cmd) {
 367                 case LTTNG_UST_STREAM:
 368                         /*
 369                          * Special-case reply to send stream info.
 370                          * Use lum.u output.
 371                          */
 372                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 373                         shm_fd = *args.stream.shm_fd;
 374                         wait_fd = *args.stream.wait_fd;
 375                         break;
 376                 case LTTNG_UST_METADATA:
 377                 case LTTNG_UST_CHANNEL:
 378                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 379                         shm_fd = *args.channel.shm_fd;
 380                         wait_fd = *args.channel.wait_fd;
 381                         break;
 382                 case LTTNG_UST_TRACER_VERSION:
 383                         lur.u.version = lum->u.version;
 384                         break;
 385                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 386                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 387                         break;
 388                 }
 389         }
 390         ret = send_reply(sock, &lur);
 391         if (ret < 0) {
 392                 perror("error sending reply");
 393                 goto error;
 394         }
 395
 396         if ((lum->cmd == LTTNG_UST_STREAM
 397              || lum->cmd == LTTNG_UST_CHANNEL
 398              || lum->cmd == LTTNG_UST_METADATA)
 399                         && lur.ret_code == USTCOMM_OK) {
 400                 int sendret = 0;
 401
 402                 /* we also need to send the file descriptors. */
 403                 ret = ustcomm_send_fds_unix_sock(sock,
 404                         &shm_fd, &shm_fd,
 405                         1, sizeof(int));
 406                 if (ret < 0) {
 407                         perror("send shm_fd");
 408                         sendret = ret;
 409                 }
 410                 /*
 411                  * The sessiond expects 2 file descriptors, even upon
 412                  * error.
 413                  */
 414                 ret = ustcomm_send_fds_unix_sock(sock,
 415                         &wait_fd, &wait_fd,
 416                         1, sizeof(int));
 417                 if (ret < 0) {
 418                         perror("send wait_fd");
 419                         goto error;
 420                 }
 421                 if (sendret) {
 422                         ret = sendret;
 423                         goto error;
 424                 }
 425         }
 426         /*
 427          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 428          * after the reply.
 429          */
 430         if (lur.ret_code == USTCOMM_OK) {
 431                 switch (lum->cmd) {
 432                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 433                         len = ustcomm_send_unix_sock(sock,
 434                                 &args.field_list.entry,
 435                                 sizeof(args.field_list.entry));
 436                         if (len != sizeof(args.field_list.entry)) {
 437                                 ret = -1;
 438                                 goto error;
 439                         }
 440                 }
 441         }
 442         /*
 443          * We still have the memory map reference, and the fds have been
 444          * sent to the sessiond. We can therefore close those fds. Note
 445          * that we keep the write side of the wait_fd open, but close
 446          * the read side.
 447          */
 448         if (lur.ret_code == USTCOMM_OK) {
 449                 switch (lum->cmd) {
 450                 case LTTNG_UST_STREAM:
 451                         if (shm_fd >= 0) {
 452                                 ret = close(shm_fd);
 453                                 if (ret) {
 454                                         PERROR("Error closing stream shm_fd");
 455                                 }
 456                                 *args.stream.shm_fd = -1;
 457                         }
 458                         if (wait_fd >= 0) {
 459                                 ret = close(wait_fd);
 460                                 if (ret) {
 461                                         PERROR("Error closing stream wait_fd");
 462                                 }
 463                                 *args.stream.wait_fd = -1;
 464                         }
 465                         break;
 466                 case LTTNG_UST_METADATA:
 467                 case LTTNG_UST_CHANNEL:
 468                         if (shm_fd >= 0) {
 469                                 ret = close(shm_fd);
 470                                 if (ret) {
 471                                         PERROR("Error closing channel shm_fd");
 472                                 }
 473                                 *args.channel.shm_fd = -1;
 474                         }
 475                         if (wait_fd >= 0) {
 476                                 ret = close(wait_fd);
 477                                 if (ret) {
 478                                         PERROR("Error closing channel wait_fd");
 479                                 }
 480                                 *args.channel.wait_fd = -1;
 481                         }
 482                         break;
 483                 }
 484         }
 485
 486 error:
 487         ust_unlock();
 488         return ret;
 489 }
 490
 491 static
 492 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 493 {
 494         int ret;
 495
 496         if (sock_info->socket != -1) {
 497                 ret = ustcomm_close_unix_sock(sock_info->socket);
 498                 if (ret) {
 499                         ERR("Error closing apps socket");
 500                 }
 501                 sock_info->socket = -1;
 502         }
 503         if (sock_info->root_handle != -1) {
 504                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 505                 if (ret) {
 506                         ERR("Error unref root handle");
 507                 }
 508                 sock_info->root_handle = -1;
 509         }
 510         sock_info->constructor_sem_posted = 0;
 511         /*
 512          * wait_shm_mmap is used by listener threads outside of the
 513          * ust lock, so we cannot tear it down ourselves, because we
 514          * cannot join on these threads. Leave this task to the OS
 515          * process exit.
 516          */
 517         if (!exiting && sock_info->wait_shm_mmap) {
 518                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 519                 if (ret) {
 520                         ERR("Error unmapping wait shm");
 521                 }
 522                 sock_info->wait_shm_mmap = NULL;
 523         }
 524 }
 525
 526 /*
 527  * Using fork to set umask in the child process (not multi-thread safe).
 528  * We deal with the shm_open vs ftruncate race (happening when the
 529  * sessiond owns the shm and does not let everybody modify it, to ensure
 530  * safety against shm_unlink) by simply letting the mmap fail and
 531  * retrying after a few seconds.
 532  * For global shm, everybody has rw access to it until the sessiond
 533  * starts.
 534  */
 535 static
 536 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 537 {
 538         int wait_shm_fd, ret;
 539         pid_t pid;
 540
 541         /*
 542          * Try to open read-only.
 543          */
 544         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 545         if (wait_shm_fd >= 0) {
 546                 goto end;
 547         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 548                 /*
 549                  * Real-only open did not work, and it's not because the
 550                  * entry was not present. It's a failure that prohibits
 551                  * using shm.
 552                  */
 553                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 554                 goto end;
 555         }
 556         /*
 557          * If the open failed because the file did not exist, try
 558          * creating it ourself.
 559          */
 560         lttng_ust_nest_count++;
 561         pid = fork();
 562         lttng_ust_nest_count--;
 563         if (pid > 0) {
 564                 int status;
 565
 566                 /*
 567                  * Parent: wait for child to return, in which case the
 568                  * shared memory map will have been created.
 569                  */
 570                 pid = wait(&status);
 571                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 572                         wait_shm_fd = -1;
 573                         goto end;
 574                 }
 575                 /*
 576                  * Try to open read-only again after creation.
 577                  */
 578                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 579                 if (wait_shm_fd < 0) {
 580                         /*
 581                          * Real-only open did not work. It's a failure
 582                          * that prohibits using shm.
 583                          */
 584                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 585                         goto end;
 586                 }
 587                 goto end;
 588         } else if (pid == 0) {
 589                 int create_mode;
 590
 591                 /* Child */
 592                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 593                 if (sock_info->global)
 594                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 595                 /*
 596                  * We're alone in a child process, so we can modify the
 597                  * process-wide umask.
 598                  */
 599                 umask(~create_mode);
 600                 /*
 601                  * Try creating shm (or get rw access).
 602                  * We don't do an exclusive open, because we allow other
 603                  * processes to create+ftruncate it concurrently.
 604                  */
 605                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 606                                 O_RDWR | O_CREAT, create_mode);
 607                 if (wait_shm_fd >= 0) {
 608                         ret = ftruncate(wait_shm_fd, mmap_size);
 609                         if (ret) {
 610                                 PERROR("ftruncate");
 611                                 exit(EXIT_FAILURE);
 612                         }
 613                         exit(EXIT_SUCCESS);
 614                 }
 615                 /*
 616                  * For local shm, we need to have rw access to accept
 617                  * opening it: this means the local sessiond will be
 618                  * able to wake us up. For global shm, we open it even
 619                  * if rw access is not granted, because the root.root
 620                  * sessiond will be able to override all rights and wake
 621                  * us up.
 622                  */
 623                 if (!sock_info->global && errno != EACCES) {
 624                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 625                         exit(EXIT_FAILURE);
 626                 }
 627                 /*
 628                  * The shm exists, but we cannot open it RW. Report
 629                  * success.
 630                  */
 631                 exit(EXIT_SUCCESS);
 632         } else {
 633                 return -1;
 634         }
 635 end:
 636         if (wait_shm_fd >= 0 && !sock_info->global) {
 637                 struct stat statbuf;
 638
 639                 /*
 640                  * Ensure that our user is the owner of the shm file for
 641                  * local shm. If we do not own the file, it means our
 642                  * sessiond will not have access to wake us up (there is
 643                  * probably a rogue process trying to fake our
 644                  * sessiond). Fallback to polling method in this case.
 645                  */
 646                 ret = fstat(wait_shm_fd, &statbuf);
 647                 if (ret) {
 648                         PERROR("fstat");
 649                         goto error_close;
 650                 }
 651                 if (statbuf.st_uid != getuid())
 652                         goto error_close;
 653         }
 654         return wait_shm_fd;
 655
 656 error_close:
 657         ret = close(wait_shm_fd);
 658         if (ret) {
 659                 PERROR("Error closing fd");
 660         }
 661         return -1;
 662 }
 663
 664 static
 665 char *get_map_shm(struct sock_info *sock_info)
 666 {
 667         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 668         int wait_shm_fd, ret;
 669         char *wait_shm_mmap;
 670
 671         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 672         if (wait_shm_fd < 0) {
 673                 goto error;
 674         }
 675         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 676                   MAP_SHARED, wait_shm_fd, 0);
 677         /* close shm fd immediately after taking the mmap reference */
 678         ret = close(wait_shm_fd);
 679         if (ret) {
 680                 PERROR("Error closing fd");
 681         }
 682         if (wait_shm_mmap == MAP_FAILED) {
 683                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 684                 goto error;
 685         }
 686         return wait_shm_mmap;
 687
 688 error:
 689         return NULL;
 690 }
 691
 692 static
 693 void wait_for_sessiond(struct sock_info *sock_info)
 694 {
 695         int ret;
 696
 697         ust_lock();
 698         if (lttng_ust_comm_should_quit) {
 699                 goto quit;
 700         }
 701         if (wait_poll_fallback) {
 702                 goto error;
 703         }
 704         if (!sock_info->wait_shm_mmap) {
 705                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 706                 if (!sock_info->wait_shm_mmap)
 707                         goto error;
 708         }
 709         ust_unlock();
 710
 711         DBG("Waiting for %s apps sessiond", sock_info->name);
 712         /* Wait for futex wakeup */
 713         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 714                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 715                         FUTEX_WAIT, 0, NULL, NULL, 0);
 716                 if (ret < 0) {
 717                         if (errno == EFAULT) {
 718                                 wait_poll_fallback = 1;
 719                                 DBG(
 720 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 721 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 722 "Please upgrade your kernel "
 723 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 724 "mainline). LTTng-UST will use polling mode fallback.");
 725                                 if (ust_debug())
 726                                         PERROR("futex");
 727                         }
 728                 }
 729         }
 730         return;
 731
 732 quit:
 733         ust_unlock();
 734         return;
 735
 736 error:
 737         ust_unlock();
 738         return;
 739 }
 740
 741 /*
 742  * This thread does not allocate any resource, except within
 743  * handle_message, within mutex protection. This mutex protects against
 744  * fork and exit.
 745  * The other moment it allocates resources is at socket connection, which
 746  * is also protected by the mutex.
 747  */
 748 static
 749 void *ust_listener_thread(void *arg)
 750 {
 751         struct sock_info *sock_info = arg;
 752         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 753
 754         /* Restart trying to connect to the session daemon */
 755 restart:
 756         if (prev_connect_failed) {
 757                 /* Wait for sessiond availability with pipe */
 758                 wait_for_sessiond(sock_info);
 759                 if (has_waited) {
 760                         has_waited = 0;
 761                         /*
 762                          * Sleep for 5 seconds before retrying after a
 763                          * sequence of failure / wait / failure. This
 764                          * deals with a killed or broken session daemon.
 765                          */
 766                         sleep(5);
 767                 }
 768                 has_waited = 1;
 769                 prev_connect_failed = 0;
 770         }
 771         ust_lock();
 772
 773         if (lttng_ust_comm_should_quit) {
 774                 ust_unlock();
 775                 goto quit;
 776         }
 777
 778         if (sock_info->socket != -1) {
 779                 ret = ustcomm_close_unix_sock(sock_info->socket);
 780                 if (ret) {
 781                         ERR("Error closing %s apps socket", sock_info->name);
 782                 }
 783                 sock_info->socket = -1;
 784         }
 785
 786         /* Register */
 787         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 788         if (ret < 0) {
 789                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 790                 prev_connect_failed = 1;
 791                 /*
 792                  * If we cannot find the sessiond daemon, don't delay
 793                  * constructor execution.
 794                  */
 795                 ret = handle_register_done(sock_info);
 796                 assert(!ret);
 797                 ust_unlock();
 798                 goto restart;
 799         }
 800
 801         sock_info->socket = sock = ret;
 802
 803         /*
 804          * Create only one root handle per listener thread for the whole
 805          * process lifetime.
 806          */
 807         if (sock_info->root_handle == -1) {
 808                 ret = lttng_abi_create_root_handle();
 809                 if (ret < 0) {
 810                         ERR("Error creating root handle");
 811                         ust_unlock();
 812                         goto quit;
 813                 }
 814                 sock_info->root_handle = ret;
 815         }
 816
 817         ret = register_app_to_sessiond(sock);
 818         if (ret < 0) {
 819                 ERR("Error registering to %s apps socket", sock_info->name);
 820                 prev_connect_failed = 1;
 821                 /*
 822                  * If we cannot register to the sessiond daemon, don't
 823                  * delay constructor execution.
 824                  */
 825                 ret = handle_register_done(sock_info);
 826                 assert(!ret);
 827                 ust_unlock();
 828                 goto restart;
 829         }
 830         ust_unlock();
 831
 832         for (;;) {
 833                 ssize_t len;
 834                 struct ustcomm_ust_msg lum;
 835
 836                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 837                 switch (len) {
 838                 case 0: /* orderly shutdown */
 839                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 840                         ust_lock();
 841                         /*
 842                          * Either sessiond has shutdown or refused us by closing the socket.
 843                          * In either case, we don't want to delay construction execution,
 844                          * and we need to wait before retry.
 845                          */
 846                         prev_connect_failed = 1;
 847                         /*
 848                          * If we cannot register to the sessiond daemon, don't
 849                          * delay constructor execution.
 850                          */
 851                         ret = handle_register_done(sock_info);
 852                         assert(!ret);
 853                         ust_unlock();
 854                         goto end;
 855                 case sizeof(lum):
 856                         DBG("message received\n");
 857                         ret = handle_message(sock_info, sock, &lum);
 858                         if (ret < 0) {
 859                                 ERR("Error handling message for %s socket", sock_info->name);
 860                         }
 861                         continue;
 862                 case -1:
 863                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 864                         if (errno == ECONNRESET) {
 865                                 ERR("%s remote end closed connection\n", sock_info->name);
 866                                 goto end;
 867                         }
 868                         goto end;
 869                 default:
 870                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 871                         continue;
 872                 }
 873
 874         }
 875 end:
 876         goto restart;   /* try to reconnect */
 877 quit:
 878         return NULL;
 879 }
 880
 881 /*
 882  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 883  */
 884 static
 885 int get_timeout(struct timespec *constructor_timeout)
 886 {
 887         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 888         char *str_delay;
 889         int ret;
 890
 891         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 892         if (str_delay) {
 893                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 894         }
 895
 896         switch (constructor_delay_ms) {
 897         case -1:/* fall-through */
 898         case 0:
 899                 return constructor_delay_ms;
 900         default:
 901                 break;
 902         }
 903
 904         /*
 905          * If we are unable to find the current time, don't wait.
 906          */
 907         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 908         if (ret) {
 909                 return -1;
 910         }
 911         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 912         constructor_timeout->tv_nsec +=
 913                 (constructor_delay_ms % 1000UL) * 1000000UL;
 914         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 915                 constructor_timeout->tv_sec++;
 916                 constructor_timeout->tv_nsec -= 1000000000UL;
 917         }
 918         return 1;
 919 }
 920
 921 /*
 922  * sessiond monitoring thread: monitor presence of global and per-user
 923  * sessiond by polling the application common named pipe.
 924  */
 925 /* TODO */
 926
 927 void __attribute__((constructor)) lttng_ust_init(void)
 928 {
 929         struct timespec constructor_timeout;
 930         sigset_t sig_all_blocked, orig_parent_mask;
 931         pthread_attr_t thread_attr;
 932         int timeout_mode;
 933         int ret;
 934
 935         if (uatomic_xchg(&initialized, 1) == 1)
 936                 return;
 937
 938         /*
 939          * Fixup interdependency between TLS fixup mutex (which happens
 940          * to be the dynamic linker mutex) and ust_lock, taken within
 941          * the ust lock.
 942          */
 943         lttng_fixup_event_tls();
 944         lttng_fixup_ringbuffer_tls();
 945         lttng_fixup_vtid_tls();
 946         lttng_fixup_nest_count_tls();
 947
 948         /*
 949          * We want precise control over the order in which we construct
 950          * our sub-libraries vs starting to receive commands from
 951          * sessiond (otherwise leading to errors when trying to create
 952          * sessiond before the init functions are completed).
 953          */
 954         init_usterr();
 955         init_tracepoint();
 956         ltt_ring_buffer_metadata_client_init();
 957         ltt_ring_buffer_client_overwrite_init();
 958         ltt_ring_buffer_client_discard_init();
 959
 960         timeout_mode = get_timeout(&constructor_timeout);
 961
 962         ret = sem_init(&constructor_wait, 0, 0);
 963         assert(!ret);
 964
 965         ret = setup_local_apps();
 966         if (ret) {
 967                 ERR("Error setting up to local apps");
 968         }
 969
 970         /* A new thread created by pthread_create inherits the signal mask
 971          * from the parent. To avoid any signal being received by the
 972          * listener thread, we block all signals temporarily in the parent,
 973          * while we create the listener thread.
 974          */
 975         sigfillset(&sig_all_blocked);
 976         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 977         if (ret) {
 978                 ERR("pthread_sigmask: %s", strerror(ret));
 979         }
 980
 981         ret = pthread_attr_init(&thread_attr);
 982         if (ret) {
 983                 ERR("pthread_attr_init: %s", strerror(ret));
 984         }
 985         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
 986         if (ret) {
 987                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
 988         }
 989
 990         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
 991                         ust_listener_thread, &global_apps);
 992         if (ret) {
 993                 ERR("pthread_create global: %s", strerror(ret));
 994         }
 995         if (local_apps.allowed) {
 996                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
 997                                 ust_listener_thread, &local_apps);
 998                 if (ret) {
 999                         ERR("pthread_create local: %s", strerror(ret));
1000                 }
1001         } else {
1002                 handle_register_done(&local_apps);
1003         }
1004         ret = pthread_attr_destroy(&thread_attr);
1005         if (ret) {
1006                 ERR("pthread_attr_destroy: %s", strerror(ret));
1007         }
1008
1009         /* Restore original signal mask in parent */
1010         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1011         if (ret) {
1012                 ERR("pthread_sigmask: %s", strerror(ret));
1013         }
1014
1015         switch (timeout_mode) {
1016         case 1: /* timeout wait */
1017                 do {
1018                         ret = sem_timedwait(&constructor_wait,
1019                                         &constructor_timeout);
1020                 } while (ret < 0 && errno == EINTR);
1021                 if (ret < 0 && errno == ETIMEDOUT) {
1022                         ERR("Timed out waiting for ltt-sessiond");
1023                 } else {
1024                         assert(!ret);
1025                 }
1026                 break;
1027         case -1:/* wait forever */
1028                 do {
1029                         ret = sem_wait(&constructor_wait);
1030                 } while (ret < 0 && errno == EINTR);
1031                 assert(!ret);
1032                 break;
1033         case 0: /* no timeout */
1034                 break;
1035         }
1036 }
1037
1038 static
1039 void lttng_ust_cleanup(int exiting)
1040 {
1041         cleanup_sock_info(&global_apps, exiting);
1042         if (local_apps.allowed) {
1043                 cleanup_sock_info(&local_apps, exiting);
1044         }
1045         /*
1046          * The teardown in this function all affect data structures
1047          * accessed under the UST lock by the listener thread. This
1048          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1049          * that none of these threads are accessing this data at this
1050          * point.
1051          */
1052         lttng_ust_abi_exit();
1053         lttng_ust_events_exit();
1054         ltt_ring_buffer_client_discard_exit();
1055         ltt_ring_buffer_client_overwrite_exit();
1056         ltt_ring_buffer_metadata_client_exit();
1057         exit_tracepoint();
1058         if (!exiting) {
1059                 /* Reinitialize values for fork */
1060                 sem_count = 2;
1061                 lttng_ust_comm_should_quit = 0;
1062                 initialized = 0;
1063         }
1064 }
1065
1066 void __attribute__((destructor)) lttng_ust_exit(void)
1067 {
1068         int ret;
1069
1070         /*
1071          * Using pthread_cancel here because:
1072          * A) we don't want to hang application teardown.
1073          * B) the thread is not allocating any resource.
1074          */
1075
1076         /*
1077          * Require the communication thread to quit. Synchronize with
1078          * mutexes to ensure it is not in a mutex critical section when
1079          * pthread_cancel is later called.
1080          */
1081         ust_lock();
1082         lttng_ust_comm_should_quit = 1;
1083         ust_unlock();
1084
1085         /* cancel threads */
1086         ret = pthread_cancel(global_apps.ust_listener);
1087         if (ret) {
1088                 ERR("Error cancelling global ust listener thread: %s",
1089                         strerror(ret));
1090         }
1091         if (local_apps.allowed) {
1092                 ret = pthread_cancel(local_apps.ust_listener);
1093                 if (ret) {
1094                         ERR("Error cancelling local ust listener thread: %s",
1095                                 strerror(ret));
1096                 }
1097         }
1098         /*
1099          * Do NOT join threads: use of sys_futex makes it impossible to
1100          * join the threads without using async-cancel, but async-cancel
1101          * is delivered by a signal, which could hit the target thread
1102          * anywhere in its code path, including while the ust_lock() is
1103          * held, causing a deadlock for the other thread. Let the OS
1104          * cleanup the threads if there are stalled in a syscall.
1105          */
1106         lttng_ust_cleanup(1);
1107 }
1108
1109 /*
1110  * We exclude the worker threads across fork and clone (except
1111  * CLONE_VM), because these system calls only keep the forking thread
1112  * running in the child.  Therefore, we don't want to call fork or clone
1113  * in the middle of an tracepoint or ust tracing state modification.
1114  * Holding this mutex protects these structures across fork and clone.
1115  */
1116 void ust_before_fork(sigset_t *save_sigset)
1117 {
1118         /*
1119          * Disable signals. This is to avoid that the child intervenes
1120          * before it is properly setup for tracing. It is safer to
1121          * disable all signals, because then we know we are not breaking
1122          * anything by restoring the original mask.
1123          */
1124         sigset_t all_sigs;
1125         int ret;
1126
1127         if (lttng_ust_nest_count)
1128                 return;
1129         /* Disable signals */
1130         sigfillset(&all_sigs);
1131         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1132         if (ret == -1) {
1133                 PERROR("sigprocmask");
1134         }
1135         ust_lock();
1136         rcu_bp_before_fork();
1137 }
1138
1139 static void ust_after_fork_common(sigset_t *restore_sigset)
1140 {
1141         int ret;
1142
1143         DBG("process %d", getpid());
1144         ust_unlock();
1145         /* Restore signals */
1146         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1147         if (ret == -1) {
1148                 PERROR("sigprocmask");
1149         }
1150 }
1151
1152 void ust_after_fork_parent(sigset_t *restore_sigset)
1153 {
1154         if (lttng_ust_nest_count)
1155                 return;
1156         DBG("process %d", getpid());
1157         rcu_bp_after_fork_parent();
1158         /* Release mutexes and reenable signals */
1159         ust_after_fork_common(restore_sigset);
1160 }
1161
1162 /*
1163  * After fork, in the child, we need to cleanup all the leftover state,
1164  * except the worker thread which already magically disappeared thanks
1165  * to the weird Linux fork semantics. After tyding up, we call
1166  * lttng_ust_init() again to start over as a new PID.
1167  *
1168  * This is meant for forks() that have tracing in the child between the
1169  * fork and following exec call (if there is any).
1170  */
1171 void ust_after_fork_child(sigset_t *restore_sigset)
1172 {
1173         if (lttng_ust_nest_count)
1174                 return;
1175         DBG("process %d", getpid());
1176         /* Release urcu mutexes */
1177         rcu_bp_after_fork_child();
1178         lttng_ust_cleanup(0);
1179         lttng_context_vtid_reset();
1180         /* Release mutexes and reenable signals */
1181         ust_after_fork_common(restore_sigset);
1182         lttng_ust_init();
1183 }