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