liblttng-ust/lttng-ust-comm.c

   1 /*
   2  * lttng-ust-comm.c
   3  *
   4  * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
   5  * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   6  *
   7  * This library is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; only
  10  * version 2.1 of the License.
  11  *
  12  * This library is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with this library; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #define _LGPL_SOURCE
  23 #include <sys/types.h>
  24 #include <sys/socket.h>
  25 #include <sys/mman.h>
  26 #include <sys/stat.h>
  27 #include <sys/types.h>
  28 #include <sys/wait.h>
  29 #include <fcntl.h>
  30 #include <unistd.h>
  31 #include <errno.h>
  32 #include <pthread.h>
  33 #include <semaphore.h>
  34 #include <time.h>
  35 #include <assert.h>
  36 #include <signal.h>
  37 #include <urcu/uatomic.h>
  38 #include <urcu/futex.h>
  39 #include <urcu/compiler.h>
  40
  41 #include <lttng/ust-events.h>
  42 #include <lttng/ust-abi.h>
  43 #include <lttng/ust.h>
  44 #include <lttng/ust-error.h>
  45 #include <lttng/ust-ctl.h>
  46 #include <urcu/tls-compat.h>
  47 #include <ust-comm.h>
  48 #include <usterr-signal-safe.h>
  49 #include <helper.h>
  50 #include "tracepoint-internal.h"
  51 #include "lttng-tracer-core.h"
  52 #include "compat.h"
  53 #include "../libringbuffer/tlsfixup.h"
  54
  55 /*
  56  * Has lttng ust comm constructor been called ?
  57  */
  58 static int initialized;
  59
  60 /*
  61  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  62  * Held when handling a command, also held by fork() to deal with
  63  * removal of threads, and by exit path.
  64  */
  65
  66 /* Should the ust comm thread quit ? */
  67 static int lttng_ust_comm_should_quit;
  68
  69 /*
  70  * Wait for either of these before continuing to the main
  71  * program:
  72  * - the register_done message from sessiond daemon
  73  *   (will let the sessiond daemon enable sessions before main
  74  *   starts.)
  75  * - sessiond daemon is not reachable.
  76  * - timeout (ensuring applications are resilient to session
  77  *   daemon problems).
  78  */
  79 static sem_t constructor_wait;
  80 /*
  81  * Doing this for both the global and local sessiond.
  82  */
  83 static int sem_count = { 2 };
  84
  85 /*
  86  * Counting nesting within lttng-ust. Used to ensure that calling fork()
  87  * from liblttng-ust does not execute the pre/post fork handlers.
  88  */
  89 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
  90
  91 /*
  92  * Info about socket and associated listener thread.
  93  */
  94 struct sock_info {
  95         const char *name;
  96         pthread_t ust_listener; /* listener thread */
  97         int root_handle;
  98         int constructor_sem_posted;
  99         int allowed;
 100         int global;
 101         int thread_active;
 102
 103         char sock_path[PATH_MAX];
 104         int socket;
 105         int notify_socket;
 106
 107         char wait_shm_path[PATH_MAX];
 108         char *wait_shm_mmap;
 109 };
 110
 111 /* Socket from app (connect) to session daemon (listen) for communication */
 112 struct sock_info global_apps = {
 113         .name = "global",
 114         .global = 1,
 115
 116         .root_handle = -1,
 117         .allowed = 1,
 118         .thread_active = 0,
 119
 120         .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
 121         .socket = -1,
 122         .notify_socket = -1,
 123
 124         .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 125 };
 126
 127 /* TODO: allow global_apps_sock_path override */
 128
 129 struct sock_info local_apps = {
 130         .name = "local",
 131         .global = 0,
 132         .root_handle = -1,
 133         .allowed = 0,   /* Check setuid bit first */
 134         .thread_active = 0,
 135
 136         .socket = -1,
 137         .notify_socket = -1,
 138 };
 139
 140 static int wait_poll_fallback;
 141
 142 static const char *cmd_name_mapping[] = {
 143         [ LTTNG_UST_RELEASE ] = "Release",
 144         [ LTTNG_UST_SESSION ] = "Create Session",
 145         [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
 146
 147         [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
 148         [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
 149         [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
 150         [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
 151
 152         /* Session FD commands */
 153         [ LTTNG_UST_CHANNEL ] = "Create Channel",
 154         [ LTTNG_UST_SESSION_START ] = "Start Session",
 155         [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
 156
 157         /* Channel FD commands */
 158         [ LTTNG_UST_STREAM ] = "Create Stream",
 159         [ LTTNG_UST_EVENT ] = "Create Event",
 160
 161         /* Event and Channel FD commands */
 162         [ LTTNG_UST_CONTEXT ] = "Create Context",
 163         [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
 164
 165         /* Event, Channel and Session commands */
 166         [ LTTNG_UST_ENABLE ] = "Enable",
 167         [ LTTNG_UST_DISABLE ] = "Disable",
 168
 169         /* Tracepoint list commands */
 170         [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
 171         [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
 172
 173         /* Event FD commands */
 174         [ LTTNG_UST_FILTER ] = "Create Filter",
 175 };
 176
 177 static const char *str_timeout;
 178 static int got_timeout_env;
 179
 180 extern void lttng_ring_buffer_client_overwrite_init(void);
 181 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 182 extern void lttng_ring_buffer_client_discard_init(void);
 183 extern void lttng_ring_buffer_client_discard_rt_init(void);
 184 extern void lttng_ring_buffer_metadata_client_init(void);
 185 extern void lttng_ring_buffer_client_overwrite_exit(void);
 186 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 187 extern void lttng_ring_buffer_client_discard_exit(void);
 188 extern void lttng_ring_buffer_client_discard_rt_exit(void);
 189 extern void lttng_ring_buffer_metadata_client_exit(void);
 190
 191 /*
 192  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 193  */
 194 static
 195 void lttng_fixup_nest_count_tls(void)
 196 {
 197         asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 198 }
 199
 200 int lttng_get_notify_socket(void *owner)
 201 {
 202         struct sock_info *info = owner;
 203
 204         return info->notify_socket;
 205 }
 206
 207 static
 208 void print_cmd(int cmd, int handle)
 209 {
 210         const char *cmd_name = "Unknown";
 211
 212         if (cmd_name_mapping[cmd]) {
 213                 cmd_name = cmd_name_mapping[cmd];
 214         }
 215         DBG("Message Received \"%s\", Handle \"%s\" (%d)", cmd_name,
 216                 lttng_ust_obj_get_name(handle), handle);
 217 }
 218
 219 static
 220 int setup_local_apps(void)
 221 {
 222         const char *home_dir;
 223         uid_t uid;
 224
 225         uid = getuid();
 226         /*
 227          * Disallow per-user tracing for setuid binaries.
 228          */
 229         if (uid != geteuid()) {
 230                 assert(local_apps.allowed == 0);
 231                 return 0;
 232         }
 233         home_dir = (const char *) getenv("HOME");
 234         if (!home_dir) {
 235                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 236                 assert(local_apps.allowed == 0);
 237                 return -ENOENT;
 238         }
 239         local_apps.allowed = 1;
 240         snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
 241                 home_dir,
 242                 LTTNG_DEFAULT_HOME_RUNDIR,
 243                 LTTNG_UST_SOCK_FILENAME);
 244         snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
 245                 LTTNG_UST_WAIT_FILENAME,
 246                 uid);
 247         return 0;
 248 }
 249
 250 /*
 251  * Get notify_sock timeout, in ms.
 252  * -1: don't wait. 0: wait forever. >0: timeout, in ms.
 253  */
 254 static
 255 long get_timeout(void)
 256 {
 257         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 258
 259         if (!got_timeout_env) {
 260                 str_timeout = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 261                 got_timeout_env = 1;
 262         }
 263         if (str_timeout)
 264                 constructor_delay_ms = strtol(str_timeout, NULL, 10);
 265         return constructor_delay_ms;
 266 }
 267
 268 static
 269 long get_notify_sock_timeout(void)
 270 {
 271         return get_timeout();
 272 }
 273
 274 /*
 275  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 276  */
 277 static
 278 int get_constructor_timeout(struct timespec *constructor_timeout)
 279 {
 280         long constructor_delay_ms;
 281         int ret;
 282
 283         constructor_delay_ms = get_timeout();
 284
 285         switch (constructor_delay_ms) {
 286         case -1:/* fall-through */
 287         case 0:
 288                 return constructor_delay_ms;
 289         default:
 290                 break;
 291         }
 292
 293         /*
 294          * If we are unable to find the current time, don't wait.
 295          */
 296         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 297         if (ret) {
 298                 return -1;
 299         }
 300         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 301         constructor_timeout->tv_nsec +=
 302                 (constructor_delay_ms % 1000UL) * 1000000UL;
 303         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 304                 constructor_timeout->tv_sec++;
 305                 constructor_timeout->tv_nsec -= 1000000000UL;
 306         }
 307         return 1;
 308 }
 309
 310 static
 311 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 312 {
 313         return ustcomm_send_reg_msg(socket,
 314                 type,
 315                 CAA_BITS_PER_LONG,
 316                 lttng_alignof(uint8_t) * CHAR_BIT,
 317                 lttng_alignof(uint16_t) * CHAR_BIT,
 318                 lttng_alignof(uint32_t) * CHAR_BIT,
 319                 lttng_alignof(uint64_t) * CHAR_BIT,
 320                 lttng_alignof(unsigned long) * CHAR_BIT);
 321 }
 322
 323 static
 324 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 325 {
 326         ssize_t len;
 327
 328         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 329         switch (len) {
 330         case sizeof(*lur):
 331                 DBG("message successfully sent");
 332                 return 0;
 333         default:
 334                 if (len == -ECONNRESET) {
 335                         DBG("remote end closed connection");
 336                         return 0;
 337                 }
 338                 if (len < 0)
 339                         return len;
 340                 DBG("incorrect message size: %zd", len);
 341                 return -EINVAL;
 342         }
 343 }
 344
 345 static
 346 int handle_register_done(struct sock_info *sock_info)
 347 {
 348         int ret;
 349
 350         if (sock_info->constructor_sem_posted)
 351                 return 0;
 352         sock_info->constructor_sem_posted = 1;
 353         if (uatomic_read(&sem_count) <= 0) {
 354                 return 0;
 355         }
 356         ret = uatomic_add_return(&sem_count, -1);
 357         if (ret == 0) {
 358                 ret = sem_post(&constructor_wait);
 359                 assert(!ret);
 360         }
 361         return 0;
 362 }
 363
 364 static
 365 int handle_message(struct sock_info *sock_info,
 366                 int sock, struct ustcomm_ust_msg *lum)
 367 {
 368         int ret = 0;
 369         const struct lttng_ust_objd_ops *ops;
 370         struct ustcomm_ust_reply lur;
 371         union ust_args args;
 372         ssize_t len;
 373
 374         ust_lock();
 375
 376         memset(&lur, 0, sizeof(lur));
 377
 378         if (lttng_ust_comm_should_quit) {
 379                 ret = -LTTNG_UST_ERR_EXITING;
 380                 goto end;
 381         }
 382
 383         ops = objd_ops(lum->handle);
 384         if (!ops) {
 385                 ret = -ENOENT;
 386                 goto end;
 387         }
 388
 389         switch (lum->cmd) {
 390         case LTTNG_UST_REGISTER_DONE:
 391                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 392                         ret = handle_register_done(sock_info);
 393                 else
 394                         ret = -EINVAL;
 395                 break;
 396         case LTTNG_UST_RELEASE:
 397                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 398                         ret = -EPERM;
 399                 else
 400                         ret = lttng_ust_objd_unref(lum->handle, 1);
 401                 break;
 402         case LTTNG_UST_FILTER:
 403         {
 404                 /* Receive filter data */
 405                 struct lttng_ust_filter_bytecode_node *bytecode;
 406
 407                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 408                         ERR("Filter data size is too large: %u bytes",
 409                                 lum->u.filter.data_size);
 410                         ret = -EINVAL;
 411                         goto error;
 412                 }
 413
 414                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 415                         ERR("Filter reloc offset %u is not within data",
 416                                 lum->u.filter.reloc_offset);
 417                         ret = -EINVAL;
 418                         goto error;
 419                 }
 420
 421                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 422                 if (!bytecode) {
 423                         ret = -ENOMEM;
 424                         goto error;
 425                 }
 426                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 427                                 lum->u.filter.data_size);
 428                 switch (len) {
 429                 case 0: /* orderly shutdown */
 430                         ret = 0;
 431                         free(bytecode);
 432                         goto error;
 433                 default:
 434                         if (len == lum->u.filter.data_size) {
 435                                 DBG("filter data received");
 436                                 break;
 437                         } else if (len < 0) {
 438                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 439                                 if (len == -ECONNRESET) {
 440                                         ERR("%s remote end closed connection", sock_info->name);
 441                                         ret = len;
 442                                         free(bytecode);
 443                                         goto error;
 444                                 }
 445                                 ret = len;
 446                                 goto end;
 447                         } else {
 448                                 DBG("incorrect filter data message size: %zd", len);
 449                                 ret = -EINVAL;
 450                                 free(bytecode);
 451                                 goto end;
 452                         }
 453                 }
 454                 bytecode->bc.len = lum->u.filter.data_size;
 455                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 456                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 457                 if (ops->cmd) {
 458                         ret = ops->cmd(lum->handle, lum->cmd,
 459                                         (unsigned long) bytecode,
 460                                         &args, sock_info);
 461                         if (ret) {
 462                                 free(bytecode);
 463                         }
 464                         /* don't free bytecode if everything went fine. */
 465                 } else {
 466                         ret = -ENOSYS;
 467                         free(bytecode);
 468                 }
 469                 break;
 470         }
 471         case LTTNG_UST_CHANNEL:
 472         {
 473                 void *chan_data;
 474                 int wakeup_fd;
 475
 476                 len = ustcomm_recv_channel_from_sessiond(sock,
 477                                 &chan_data, lum->u.channel.len,
 478                                 &wakeup_fd);
 479                 switch (len) {
 480                 case 0: /* orderly shutdown */
 481                         ret = 0;
 482                         goto error;
 483                 default:
 484                         if (len == lum->u.channel.len) {
 485                                 DBG("channel data received");
 486                                 break;
 487                         } else if (len < 0) {
 488                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 489                                 if (len == -ECONNRESET) {
 490                                         ERR("%s remote end closed connection", sock_info->name);
 491                                         ret = len;
 492                                         goto error;
 493                                 }
 494                                 ret = len;
 495                                 goto end;
 496                         } else {
 497                                 DBG("incorrect channel data message size: %zd", len);
 498                                 ret = -EINVAL;
 499                                 goto end;
 500                         }
 501                 }
 502                 args.channel.chan_data = chan_data;
 503                 args.channel.wakeup_fd = wakeup_fd;
 504                 if (ops->cmd)
 505                         ret = ops->cmd(lum->handle, lum->cmd,
 506                                         (unsigned long) &lum->u,
 507                                         &args, sock_info);
 508                 else
 509                         ret = -ENOSYS;
 510                 break;
 511         }
 512         case LTTNG_UST_STREAM:
 513         {
 514                 /* Receive shm_fd, wakeup_fd */
 515                 ret = ustcomm_recv_stream_from_sessiond(sock,
 516                         &lum->u.stream.len,
 517                         &args.stream.shm_fd,
 518                         &args.stream.wakeup_fd);
 519                 if (ret) {
 520                         goto end;
 521                 }
 522                 if (ops->cmd)
 523                         ret = ops->cmd(lum->handle, lum->cmd,
 524                                         (unsigned long) &lum->u,
 525                                         &args, sock_info);
 526                 else
 527                         ret = -ENOSYS;
 528                 break;
 529         }
 530         default:
 531                 if (ops->cmd)
 532                         ret = ops->cmd(lum->handle, lum->cmd,
 533                                         (unsigned long) &lum->u,
 534                                         &args, sock_info);
 535                 else
 536                         ret = -ENOSYS;
 537                 break;
 538         }
 539
 540 end:
 541         lur.handle = lum->handle;
 542         lur.cmd = lum->cmd;
 543         lur.ret_val = ret;
 544         if (ret >= 0) {
 545                 lur.ret_code = LTTNG_UST_OK;
 546         } else {
 547                 /*
 548                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 549                  * error that are not caused by the transport, except if
 550                  * we already have a more precise error message to
 551                  * report.
 552                  */
 553                 if (ret > -LTTNG_UST_ERR) {
 554                         /* Translate code to UST error. */
 555                         switch (ret) {
 556                         case -EEXIST:
 557                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 558                                 break;
 559                         case -EINVAL:
 560                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 561                                 break;
 562                         case -ENOENT:
 563                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 564                                 break;
 565                         case -EPERM:
 566                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 567                                 break;
 568                         case -ENOSYS:
 569                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 570                                 break;
 571                         default:
 572                                 lur.ret_code = -LTTNG_UST_ERR;
 573                                 break;
 574                         }
 575                 } else {
 576                         lur.ret_code = ret;
 577                 }
 578         }
 579         if (ret >= 0) {
 580                 switch (lum->cmd) {
 581                 case LTTNG_UST_TRACER_VERSION:
 582                         lur.u.version = lum->u.version;
 583                         break;
 584                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 585                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 586                         break;
 587                 }
 588         }
 589         DBG("Return value: %d", lur.ret_val);
 590         ret = send_reply(sock, &lur);
 591         if (ret < 0) {
 592                 DBG("error sending reply");
 593                 goto error;
 594         }
 595
 596         /*
 597          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 598          * after the reply.
 599          */
 600         if (lur.ret_code == LTTNG_UST_OK) {
 601                 switch (lum->cmd) {
 602                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 603                         len = ustcomm_send_unix_sock(sock,
 604                                 &args.field_list.entry,
 605                                 sizeof(args.field_list.entry));
 606                         if (len < 0) {
 607                                 ret = len;
 608                                 goto error;
 609                         }
 610                         if (len != sizeof(args.field_list.entry)) {
 611                                 ret = -EINVAL;
 612                                 goto error;
 613                         }
 614                 }
 615         }
 616
 617 error:
 618         ust_unlock();
 619         return ret;
 620 }
 621
 622 static
 623 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 624 {
 625         int ret;
 626
 627         if (sock_info->socket != -1) {
 628                 ret = ustcomm_close_unix_sock(sock_info->socket);
 629                 if (ret) {
 630                         ERR("Error closing ust cmd socket");
 631                 }
 632                 sock_info->socket = -1;
 633         }
 634         if (sock_info->notify_socket != -1) {
 635                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 636                 if (ret) {
 637                         ERR("Error closing ust notify socket");
 638                 }
 639                 sock_info->notify_socket = -1;
 640         }
 641         if (sock_info->root_handle != -1) {
 642                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 643                 if (ret) {
 644                         ERR("Error unref root handle");
 645                 }
 646                 sock_info->root_handle = -1;
 647         }
 648         sock_info->constructor_sem_posted = 0;
 649         /*
 650          * wait_shm_mmap is used by listener threads outside of the
 651          * ust lock, so we cannot tear it down ourselves, because we
 652          * cannot join on these threads. Leave this task to the OS
 653          * process exit.
 654          */
 655         if (!exiting && sock_info->wait_shm_mmap) {
 656                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 657                 if (ret) {
 658                         ERR("Error unmapping wait shm");
 659                 }
 660                 sock_info->wait_shm_mmap = NULL;
 661         }
 662 }
 663
 664 /*
 665  * Using fork to set umask in the child process (not multi-thread safe).
 666  * We deal with the shm_open vs ftruncate race (happening when the
 667  * sessiond owns the shm and does not let everybody modify it, to ensure
 668  * safety against shm_unlink) by simply letting the mmap fail and
 669  * retrying after a few seconds.
 670  * For global shm, everybody has rw access to it until the sessiond
 671  * starts.
 672  */
 673 static
 674 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 675 {
 676         int wait_shm_fd, ret;
 677         pid_t pid;
 678
 679         /*
 680          * Try to open read-only.
 681          */
 682         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 683         if (wait_shm_fd >= 0) {
 684                 goto end;
 685         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 686                 /*
 687                  * Real-only open did not work, and it's not because the
 688                  * entry was not present. It's a failure that prohibits
 689                  * using shm.
 690                  */
 691                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 692                 goto end;
 693         }
 694         /*
 695          * If the open failed because the file did not exist, try
 696          * creating it ourself.
 697          */
 698         URCU_TLS(lttng_ust_nest_count)++;
 699         pid = fork();
 700         URCU_TLS(lttng_ust_nest_count)--;
 701         if (pid > 0) {
 702                 int status;
 703
 704                 /*
 705                  * Parent: wait for child to return, in which case the
 706                  * shared memory map will have been created.
 707                  */
 708                 pid = wait(&status);
 709                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 710                         wait_shm_fd = -1;
 711                         goto end;
 712                 }
 713                 /*
 714                  * Try to open read-only again after creation.
 715                  */
 716                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 717                 if (wait_shm_fd < 0) {
 718                         /*
 719                          * Real-only open did not work. It's a failure
 720                          * that prohibits using shm.
 721                          */
 722                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 723                         goto end;
 724                 }
 725                 goto end;
 726         } else if (pid == 0) {
 727                 int create_mode;
 728
 729                 /* Child */
 730                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 731                 if (sock_info->global)
 732                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 733                 /*
 734                  * We're alone in a child process, so we can modify the
 735                  * process-wide umask.
 736                  */
 737                 umask(~create_mode);
 738                 /*
 739                  * Try creating shm (or get rw access).
 740                  * We don't do an exclusive open, because we allow other
 741                  * processes to create+ftruncate it concurrently.
 742                  */
 743                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 744                                 O_RDWR | O_CREAT, create_mode);
 745                 if (wait_shm_fd >= 0) {
 746                         ret = ftruncate(wait_shm_fd, mmap_size);
 747                         if (ret) {
 748                                 PERROR("ftruncate");
 749                                 _exit(EXIT_FAILURE);
 750                         }
 751                         _exit(EXIT_SUCCESS);
 752                 }
 753                 /*
 754                  * For local shm, we need to have rw access to accept
 755                  * opening it: this means the local sessiond will be
 756                  * able to wake us up. For global shm, we open it even
 757                  * if rw access is not granted, because the root.root
 758                  * sessiond will be able to override all rights and wake
 759                  * us up.
 760                  */
 761                 if (!sock_info->global && errno != EACCES) {
 762                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 763                         _exit(EXIT_FAILURE);
 764                 }
 765                 /*
 766                  * The shm exists, but we cannot open it RW. Report
 767                  * success.
 768                  */
 769                 _exit(EXIT_SUCCESS);
 770         } else {
 771                 return -1;
 772         }
 773 end:
 774         if (wait_shm_fd >= 0 && !sock_info->global) {
 775                 struct stat statbuf;
 776
 777                 /*
 778                  * Ensure that our user is the owner of the shm file for
 779                  * local shm. If we do not own the file, it means our
 780                  * sessiond will not have access to wake us up (there is
 781                  * probably a rogue process trying to fake our
 782                  * sessiond). Fallback to polling method in this case.
 783                  */
 784                 ret = fstat(wait_shm_fd, &statbuf);
 785                 if (ret) {
 786                         PERROR("fstat");
 787                         goto error_close;
 788                 }
 789                 if (statbuf.st_uid != getuid())
 790                         goto error_close;
 791         }
 792         return wait_shm_fd;
 793
 794 error_close:
 795         ret = close(wait_shm_fd);
 796         if (ret) {
 797                 PERROR("Error closing fd");
 798         }
 799         return -1;
 800 }
 801
 802 static
 803 char *get_map_shm(struct sock_info *sock_info)
 804 {
 805         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 806         int wait_shm_fd, ret;
 807         char *wait_shm_mmap;
 808
 809         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 810         if (wait_shm_fd < 0) {
 811                 goto error;
 812         }
 813         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 814                   MAP_SHARED, wait_shm_fd, 0);
 815         /* close shm fd immediately after taking the mmap reference */
 816         ret = close(wait_shm_fd);
 817         if (ret) {
 818                 PERROR("Error closing fd");
 819         }
 820         if (wait_shm_mmap == MAP_FAILED) {
 821                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 822                 goto error;
 823         }
 824         return wait_shm_mmap;
 825
 826 error:
 827         return NULL;
 828 }
 829
 830 static
 831 void wait_for_sessiond(struct sock_info *sock_info)
 832 {
 833         int ret;
 834
 835         ust_lock();
 836         if (lttng_ust_comm_should_quit) {
 837                 goto quit;
 838         }
 839         if (wait_poll_fallback) {
 840                 goto error;
 841         }
 842         if (!sock_info->wait_shm_mmap) {
 843                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 844                 if (!sock_info->wait_shm_mmap)
 845                         goto error;
 846         }
 847         ust_unlock();
 848
 849         DBG("Waiting for %s apps sessiond", sock_info->name);
 850         /* Wait for futex wakeup */
 851         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 852                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 853                         FUTEX_WAIT, 0, NULL, NULL, 0);
 854                 if (ret < 0) {
 855                         if (errno == EFAULT) {
 856                                 wait_poll_fallback = 1;
 857                                 DBG(
 858 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 859 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 860 "Please upgrade your kernel "
 861 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 862 "mainline). LTTng-UST will use polling mode fallback.");
 863                                 if (ust_debug())
 864                                         PERROR("futex");
 865                         }
 866                 }
 867         }
 868         return;
 869
 870 quit:
 871         ust_unlock();
 872         return;
 873
 874 error:
 875         ust_unlock();
 876         return;
 877 }
 878
 879 /*
 880  * This thread does not allocate any resource, except within
 881  * handle_message, within mutex protection. This mutex protects against
 882  * fork and exit.
 883  * The other moment it allocates resources is at socket connection, which
 884  * is also protected by the mutex.
 885  */
 886 static
 887 void *ust_listener_thread(void *arg)
 888 {
 889         struct sock_info *sock_info = arg;
 890         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 891         int open_sock[2];
 892         int i;
 893         long timeout;
 894
 895         /* Restart trying to connect to the session daemon */
 896 restart:
 897         if (prev_connect_failed) {
 898                 /* Wait for sessiond availability with pipe */
 899                 wait_for_sessiond(sock_info);
 900                 if (has_waited) {
 901                         has_waited = 0;
 902                         /*
 903                          * Sleep for 5 seconds before retrying after a
 904                          * sequence of failure / wait / failure. This
 905                          * deals with a killed or broken session daemon.
 906                          */
 907                         sleep(5);
 908                 }
 909                 has_waited = 1;
 910                 prev_connect_failed = 0;
 911         }
 912         ust_lock();
 913
 914         if (lttng_ust_comm_should_quit) {
 915                 goto quit;
 916         }
 917
 918         if (sock_info->socket != -1) {
 919                 ret = ustcomm_close_unix_sock(sock_info->socket);
 920                 if (ret) {
 921                         ERR("Error closing %s ust cmd socket",
 922                                 sock_info->name);
 923                 }
 924                 sock_info->socket = -1;
 925         }
 926         if (sock_info->notify_socket != -1) {
 927                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 928                 if (ret) {
 929                         ERR("Error closing %s ust notify socket",
 930                                 sock_info->name);
 931                 }
 932                 sock_info->notify_socket = -1;
 933         }
 934
 935         /* Register */
 936         for (i = 0; i < 2; i++) {
 937                 ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 938                 if (ret < 0) {
 939                         DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 940                         prev_connect_failed = 1;
 941                         /*
 942                          * If we cannot find the sessiond daemon, don't delay
 943                          * constructor execution.
 944                          */
 945                         ret = handle_register_done(sock_info);
 946                         assert(!ret);
 947                         ust_unlock();
 948                         goto restart;
 949                 }
 950                 open_sock[i] = ret;
 951         }
 952
 953         sock_info->socket = open_sock[0];
 954         sock_info->notify_socket = open_sock[1];
 955
 956         timeout = get_notify_sock_timeout();
 957         if (timeout >= 0) {
 958                 /*
 959                  * Give at least 10ms to sessiond to reply to
 960                  * notifications.
 961                  */
 962                 if (timeout < 10)
 963                         timeout = 10;
 964                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
 965                                 timeout);
 966                 if (ret < 0) {
 967                         WARN("Error setting socket receive timeout");
 968                 }
 969                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
 970                                 timeout);
 971                 if (ret < 0) {
 972                         WARN("Error setting socket send timeout");
 973                 }
 974         } else if (timeout < -1) {
 975                 WARN("Unsuppoorted timeout value %ld", timeout);
 976         }
 977
 978         /*
 979          * Create only one root handle per listener thread for the whole
 980          * process lifetime, so we ensure we get ID which is statically
 981          * assigned to the root handle.
 982          */
 983         if (sock_info->root_handle == -1) {
 984                 ret = lttng_abi_create_root_handle();
 985                 if (ret < 0) {
 986                         ERR("Error creating root handle");
 987                         goto quit;
 988                 }
 989                 sock_info->root_handle = ret;
 990         }
 991
 992         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
 993         if (ret < 0) {
 994                 ERR("Error registering to %s ust cmd socket",
 995                         sock_info->name);
 996                 prev_connect_failed = 1;
 997                 /*
 998                  * If we cannot register to the sessiond daemon, don't
 999                  * delay constructor execution.
1000                  */
1001                 ret = handle_register_done(sock_info);
1002                 assert(!ret);
1003                 ust_unlock();
1004                 goto restart;
1005         }
1006         ret = register_to_sessiond(sock_info->notify_socket,
1007                         USTCTL_SOCKET_NOTIFY);
1008         if (ret < 0) {
1009                 ERR("Error registering to %s ust notify socket",
1010                         sock_info->name);
1011                 prev_connect_failed = 1;
1012                 /*
1013                  * If we cannot register to the sessiond daemon, don't
1014                  * delay constructor execution.
1015                  */
1016                 ret = handle_register_done(sock_info);
1017                 assert(!ret);
1018                 ust_unlock();
1019                 goto restart;
1020         }
1021         sock = sock_info->socket;
1022
1023         ust_unlock();
1024
1025         for (;;) {
1026                 ssize_t len;
1027                 struct ustcomm_ust_msg lum;
1028
1029                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1030                 switch (len) {
1031                 case 0: /* orderly shutdown */
1032                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1033                         ust_lock();
1034                         if (lttng_ust_comm_should_quit) {
1035                                 goto quit;
1036                         }
1037                         /*
1038                          * Either sessiond has shutdown or refused us by closing the socket.
1039                          * In either case, we don't want to delay construction execution,
1040                          * and we need to wait before retry.
1041                          */
1042                         prev_connect_failed = 1;
1043                         /*
1044                          * If we cannot register to the sessiond daemon, don't
1045                          * delay constructor execution.
1046                          */
1047                         ret = handle_register_done(sock_info);
1048                         assert(!ret);
1049                         ust_unlock();
1050                         goto end;
1051                 case sizeof(lum):
1052                         print_cmd(lum.cmd, lum.handle);
1053                         ret = handle_message(sock_info, sock, &lum);
1054                         if (ret) {
1055                                 ERR("Error handling message for %s socket", sock_info->name);
1056                         }
1057                         continue;
1058                 default:
1059                         if (len < 0) {
1060                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1061                         } else {
1062                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1063                         }
1064                         if (len == -ECONNRESET) {
1065                                 DBG("%s remote end closed connection", sock_info->name);
1066                                 goto end;
1067                         }
1068                         goto end;
1069                 }
1070
1071         }
1072 end:
1073         ust_lock();
1074         if (lttng_ust_comm_should_quit) {
1075                 goto quit;
1076         }
1077         /* Cleanup socket handles before trying to reconnect */
1078         lttng_ust_objd_table_owner_cleanup(sock_info);
1079         ust_unlock();
1080         goto restart;   /* try to reconnect */
1081
1082 quit:
1083         sock_info->thread_active = 0;
1084         ust_unlock();
1085         return NULL;
1086 }
1087
1088 /*
1089  * sessiond monitoring thread: monitor presence of global and per-user
1090  * sessiond by polling the application common named pipe.
1091  */
1092 void __attribute__((constructor)) lttng_ust_init(void)
1093 {
1094         struct timespec constructor_timeout;
1095         sigset_t sig_all_blocked, orig_parent_mask;
1096         pthread_attr_t thread_attr;
1097         int timeout_mode;
1098         int ret;
1099
1100         if (uatomic_xchg(&initialized, 1) == 1)
1101                 return;
1102
1103         /*
1104          * Fixup interdependency between TLS fixup mutex (which happens
1105          * to be the dynamic linker mutex) and ust_lock, taken within
1106          * the ust lock.
1107          */
1108         lttng_fixup_ringbuffer_tls();
1109         lttng_fixup_vtid_tls();
1110         lttng_fixup_nest_count_tls();
1111         lttng_fixup_procname_tls();
1112
1113         /*
1114          * We want precise control over the order in which we construct
1115          * our sub-libraries vs starting to receive commands from
1116          * sessiond (otherwise leading to errors when trying to create
1117          * sessiond before the init functions are completed).
1118          */
1119         init_usterr();
1120         init_tracepoint();
1121         lttng_ring_buffer_metadata_client_init();
1122         lttng_ring_buffer_client_overwrite_init();
1123         lttng_ring_buffer_client_overwrite_rt_init();
1124         lttng_ring_buffer_client_discard_init();
1125         lttng_ring_buffer_client_discard_rt_init();
1126         lttng_context_init();
1127
1128         timeout_mode = get_constructor_timeout(&constructor_timeout);
1129
1130         ret = sem_init(&constructor_wait, 0, 0);
1131         assert(!ret);
1132
1133         ret = setup_local_apps();
1134         if (ret) {
1135                 DBG("local apps setup returned %d", ret);
1136         }
1137
1138         /* A new thread created by pthread_create inherits the signal mask
1139          * from the parent. To avoid any signal being received by the
1140          * listener thread, we block all signals temporarily in the parent,
1141          * while we create the listener thread.
1142          */
1143         sigfillset(&sig_all_blocked);
1144         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1145         if (ret) {
1146                 ERR("pthread_sigmask: %s", strerror(ret));
1147         }
1148
1149         ret = pthread_attr_init(&thread_attr);
1150         if (ret) {
1151                 ERR("pthread_attr_init: %s", strerror(ret));
1152         }
1153         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1154         if (ret) {
1155                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1156         }
1157
1158         ust_lock();
1159         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1160                         ust_listener_thread, &global_apps);
1161         if (ret) {
1162                 ERR("pthread_create global: %s", strerror(ret));
1163         }
1164         global_apps.thread_active = 1;
1165         ust_unlock();
1166
1167         if (local_apps.allowed) {
1168                 ust_lock();
1169                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1170                                 ust_listener_thread, &local_apps);
1171                 if (ret) {
1172                         ERR("pthread_create local: %s", strerror(ret));
1173                 }
1174                 local_apps.thread_active = 1;
1175                 ust_unlock();
1176         } else {
1177                 handle_register_done(&local_apps);
1178         }
1179         ret = pthread_attr_destroy(&thread_attr);
1180         if (ret) {
1181                 ERR("pthread_attr_destroy: %s", strerror(ret));
1182         }
1183
1184         /* Restore original signal mask in parent */
1185         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1186         if (ret) {
1187                 ERR("pthread_sigmask: %s", strerror(ret));
1188         }
1189
1190         switch (timeout_mode) {
1191         case 1: /* timeout wait */
1192                 do {
1193                         ret = sem_timedwait(&constructor_wait,
1194                                         &constructor_timeout);
1195                 } while (ret < 0 && errno == EINTR);
1196                 if (ret < 0 && errno == ETIMEDOUT) {
1197                         ERR("Timed out waiting for lttng-sessiond");
1198                 } else {
1199                         assert(!ret);
1200                 }
1201                 break;
1202         case -1:/* wait forever */
1203                 do {
1204                         ret = sem_wait(&constructor_wait);
1205                 } while (ret < 0 && errno == EINTR);
1206                 assert(!ret);
1207                 break;
1208         case 0: /* no timeout */
1209                 break;
1210         }
1211 }
1212
1213 static
1214 void lttng_ust_cleanup(int exiting)
1215 {
1216         cleanup_sock_info(&global_apps, exiting);
1217         if (local_apps.allowed) {
1218                 cleanup_sock_info(&local_apps, exiting);
1219         }
1220         /*
1221          * The teardown in this function all affect data structures
1222          * accessed under the UST lock by the listener thread. This
1223          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1224          * that none of these threads are accessing this data at this
1225          * point.
1226          */
1227         lttng_ust_abi_exit();
1228         lttng_ust_events_exit();
1229         lttng_context_exit();
1230         lttng_ring_buffer_client_discard_rt_exit();
1231         lttng_ring_buffer_client_discard_exit();
1232         lttng_ring_buffer_client_overwrite_rt_exit();
1233         lttng_ring_buffer_client_overwrite_exit();
1234         lttng_ring_buffer_metadata_client_exit();
1235         exit_tracepoint();
1236         if (!exiting) {
1237                 /* Reinitialize values for fork */
1238                 sem_count = 2;
1239                 lttng_ust_comm_should_quit = 0;
1240                 initialized = 0;
1241         }
1242 }
1243
1244 void __attribute__((destructor)) lttng_ust_exit(void)
1245 {
1246         int ret;
1247
1248         /*
1249          * Using pthread_cancel here because:
1250          * A) we don't want to hang application teardown.
1251          * B) the thread is not allocating any resource.
1252          */
1253
1254         /*
1255          * Require the communication thread to quit. Synchronize with
1256          * mutexes to ensure it is not in a mutex critical section when
1257          * pthread_cancel is later called.
1258          */
1259         ust_lock();
1260         lttng_ust_comm_should_quit = 1;
1261
1262         /* cancel threads */
1263         if (global_apps.thread_active) {
1264                 ret = pthread_cancel(global_apps.ust_listener);
1265                 if (ret) {
1266                         ERR("Error cancelling global ust listener thread: %s",
1267                                 strerror(ret));
1268                 } else {
1269                         global_apps.thread_active = 0;
1270                 }
1271         }
1272         if (local_apps.thread_active) {
1273                 ret = pthread_cancel(local_apps.ust_listener);
1274                 if (ret) {
1275                         ERR("Error cancelling local ust listener thread: %s",
1276                                 strerror(ret));
1277                 } else {
1278                         local_apps.thread_active = 0;
1279                 }
1280         }
1281         ust_unlock();
1282
1283         /*
1284          * Do NOT join threads: use of sys_futex makes it impossible to
1285          * join the threads without using async-cancel, but async-cancel
1286          * is delivered by a signal, which could hit the target thread
1287          * anywhere in its code path, including while the ust_lock() is
1288          * held, causing a deadlock for the other thread. Let the OS
1289          * cleanup the threads if there are stalled in a syscall.
1290          */
1291         lttng_ust_cleanup(1);
1292 }
1293
1294 /*
1295  * We exclude the worker threads across fork and clone (except
1296  * CLONE_VM), because these system calls only keep the forking thread
1297  * running in the child.  Therefore, we don't want to call fork or clone
1298  * in the middle of an tracepoint or ust tracing state modification.
1299  * Holding this mutex protects these structures across fork and clone.
1300  */
1301 void ust_before_fork(sigset_t *save_sigset)
1302 {
1303         /*
1304          * Disable signals. This is to avoid that the child intervenes
1305          * before it is properly setup for tracing. It is safer to
1306          * disable all signals, because then we know we are not breaking
1307          * anything by restoring the original mask.
1308          */
1309         sigset_t all_sigs;
1310         int ret;
1311
1312         if (URCU_TLS(lttng_ust_nest_count))
1313                 return;
1314         /* Disable signals */
1315         sigfillset(&all_sigs);
1316         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1317         if (ret == -1) {
1318                 PERROR("sigprocmask");
1319         }
1320         ust_lock();
1321         rcu_bp_before_fork();
1322 }
1323
1324 static void ust_after_fork_common(sigset_t *restore_sigset)
1325 {
1326         int ret;
1327
1328         DBG("process %d", getpid());
1329         ust_unlock();
1330         /* Restore signals */
1331         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1332         if (ret == -1) {
1333                 PERROR("sigprocmask");
1334         }
1335 }
1336
1337 void ust_after_fork_parent(sigset_t *restore_sigset)
1338 {
1339         if (URCU_TLS(lttng_ust_nest_count))
1340                 return;
1341         DBG("process %d", getpid());
1342         rcu_bp_after_fork_parent();
1343         /* Release mutexes and reenable signals */
1344         ust_after_fork_common(restore_sigset);
1345 }
1346
1347 /*
1348  * After fork, in the child, we need to cleanup all the leftover state,
1349  * except the worker thread which already magically disappeared thanks
1350  * to the weird Linux fork semantics. After tyding up, we call
1351  * lttng_ust_init() again to start over as a new PID.
1352  *
1353  * This is meant for forks() that have tracing in the child between the
1354  * fork and following exec call (if there is any).
1355  */
1356 void ust_after_fork_child(sigset_t *restore_sigset)
1357 {
1358         if (URCU_TLS(lttng_ust_nest_count))
1359                 return;
1360         DBG("process %d", getpid());
1361         /* Release urcu mutexes */
1362         rcu_bp_after_fork_child();
1363         lttng_ust_cleanup(0);
1364         lttng_context_vtid_reset();
1365         /* Release mutexes and reenable signals */
1366         ust_after_fork_common(restore_sigset);
1367         lttng_ust_init();
1368 }