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 #include "lttng-ust-baddr.h"
  55
  56 /*
  57  * Has lttng ust comm constructor been called ?
  58  */
  59 static int initialized;
  60
  61 /*
  62  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  63  * Held when handling a command, also held by fork() to deal with
  64  * removal of threads, and by exit path.
  65  *
  66  * The UST lock is the centralized mutex across UST tracing control and
  67  * probe registration.
  68  *
  69  * ust_exit_mutex must never nest in ust_mutex.
  70  */
  71 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
  72
  73 /*
  74  * ust_exit_mutex protects thread_active variable wrt thread exit. It
  75  * cannot be done by ust_mutex because pthread_cancel(), which takes an
  76  * internal libc lock, cannot nest within ust_mutex.
  77  *
  78  * It never nests within a ust_mutex.
  79  */
  80 static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
  81
  82 /* Should the ust comm thread quit ? */
  83 static int lttng_ust_comm_should_quit;
  84
  85 /*
  86  * Return 0 on success, -1 if should quilt.
  87  * The lock is taken in both cases.
  88  */
  89 int ust_lock(void)
  90 {
  91         pthread_mutex_lock(&ust_mutex);
  92         if (lttng_ust_comm_should_quit) {
  93                 return -1;
  94         } else {
  95                 return 0;
  96         }
  97 }
  98
  99 /*
 100  * ust_lock_nocheck() can be used in constructors/destructors, because
 101  * they are already nested within the dynamic loader lock, and therefore
 102  * have exclusive access against execution of liblttng-ust destructor.
 103  */
 104 void ust_lock_nocheck(void)
 105 {
 106         pthread_mutex_lock(&ust_mutex);
 107 }
 108
 109 void ust_unlock(void)
 110 {
 111         pthread_mutex_unlock(&ust_mutex);
 112 }
 113
 114 /*
 115  * Wait for either of these before continuing to the main
 116  * program:
 117  * - the register_done message from sessiond daemon
 118  *   (will let the sessiond daemon enable sessions before main
 119  *   starts.)
 120  * - sessiond daemon is not reachable.
 121  * - timeout (ensuring applications are resilient to session
 122  *   daemon problems).
 123  */
 124 static sem_t constructor_wait;
 125 /*
 126  * Doing this for both the global and local sessiond.
 127  */
 128 static int sem_count = { 2 };
 129
 130 /*
 131  * Counting nesting within lttng-ust. Used to ensure that calling fork()
 132  * from liblttng-ust does not execute the pre/post fork handlers.
 133  */
 134 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
 135
 136 /*
 137  * Info about socket and associated listener thread.
 138  */
 139 struct sock_info {
 140         const char *name;
 141         pthread_t ust_listener; /* listener thread */
 142         int root_handle;
 143         int constructor_sem_posted;
 144         int allowed;
 145         int global;
 146         int thread_active;
 147
 148         char sock_path[PATH_MAX];
 149         int socket;
 150         int notify_socket;
 151
 152         char wait_shm_path[PATH_MAX];
 153         char *wait_shm_mmap;
 154         /* Keep track of lazy state dump not performed yet. */
 155         int statedump_pending;
 156 };
 157
 158 /* Socket from app (connect) to session daemon (listen) for communication */
 159 struct sock_info global_apps = {
 160         .name = "global",
 161         .global = 1,
 162
 163         .root_handle = -1,
 164         .allowed = 1,
 165         .thread_active = 0,
 166
 167         .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
 168         .socket = -1,
 169         .notify_socket = -1,
 170
 171         .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 172
 173         .statedump_pending = 0,
 174 };
 175
 176 /* TODO: allow global_apps_sock_path override */
 177
 178 struct sock_info local_apps = {
 179         .name = "local",
 180         .global = 0,
 181         .root_handle = -1,
 182         .allowed = 0,   /* Check setuid bit first */
 183         .thread_active = 0,
 184
 185         .socket = -1,
 186         .notify_socket = -1,
 187
 188         .statedump_pending = 0,
 189 };
 190
 191 static int wait_poll_fallback;
 192
 193 static const char *cmd_name_mapping[] = {
 194         [ LTTNG_UST_RELEASE ] = "Release",
 195         [ LTTNG_UST_SESSION ] = "Create Session",
 196         [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
 197
 198         [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
 199         [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
 200         [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
 201         [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
 202
 203         /* Session FD commands */
 204         [ LTTNG_UST_CHANNEL ] = "Create Channel",
 205         [ LTTNG_UST_SESSION_START ] = "Start Session",
 206         [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
 207
 208         /* Channel FD commands */
 209         [ LTTNG_UST_STREAM ] = "Create Stream",
 210         [ LTTNG_UST_EVENT ] = "Create Event",
 211
 212         /* Event and Channel FD commands */
 213         [ LTTNG_UST_CONTEXT ] = "Create Context",
 214         [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
 215
 216         /* Event, Channel and Session commands */
 217         [ LTTNG_UST_ENABLE ] = "Enable",
 218         [ LTTNG_UST_DISABLE ] = "Disable",
 219
 220         /* Tracepoint list commands */
 221         [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
 222         [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
 223
 224         /* Event FD commands */
 225         [ LTTNG_UST_FILTER ] = "Create Filter",
 226         [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
 227 };
 228
 229 static const char *str_timeout;
 230 static int got_timeout_env;
 231
 232 extern void lttng_ring_buffer_client_overwrite_init(void);
 233 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 234 extern void lttng_ring_buffer_client_discard_init(void);
 235 extern void lttng_ring_buffer_client_discard_rt_init(void);
 236 extern void lttng_ring_buffer_metadata_client_init(void);
 237 extern void lttng_ring_buffer_client_overwrite_exit(void);
 238 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 239 extern void lttng_ring_buffer_client_discard_exit(void);
 240 extern void lttng_ring_buffer_client_discard_rt_exit(void);
 241 extern void lttng_ring_buffer_metadata_client_exit(void);
 242
 243 /*
 244  * Returns the HOME directory path. Caller MUST NOT free(3) the returned
 245  * pointer.
 246  */
 247 static
 248 const char *get_lttng_home_dir(void)
 249 {
 250        const char *val;
 251
 252        val = (const char *) getenv("LTTNG_HOME");
 253        if (val != NULL) {
 254                return val;
 255        }
 256        return (const char *) getenv("HOME");
 257 }
 258
 259 /*
 260  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 261  */
 262 static
 263 void lttng_fixup_nest_count_tls(void)
 264 {
 265         asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 266 }
 267
 268 int lttng_get_notify_socket(void *owner)
 269 {
 270         struct sock_info *info = owner;
 271
 272         return info->notify_socket;
 273 }
 274
 275 static
 276 void print_cmd(int cmd, int handle)
 277 {
 278         const char *cmd_name = "Unknown";
 279
 280         if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
 281                         && cmd_name_mapping[cmd]) {
 282                 cmd_name = cmd_name_mapping[cmd];
 283         }
 284         DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
 285                 cmd_name, cmd,
 286                 lttng_ust_obj_get_name(handle), handle);
 287 }
 288
 289 static
 290 int setup_local_apps(void)
 291 {
 292         const char *home_dir;
 293         uid_t uid;
 294
 295         uid = getuid();
 296         /*
 297          * Disallow per-user tracing for setuid binaries.
 298          */
 299         if (uid != geteuid()) {
 300                 assert(local_apps.allowed == 0);
 301                 return 0;
 302         }
 303         home_dir = get_lttng_home_dir();
 304         if (!home_dir) {
 305                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 306                 assert(local_apps.allowed == 0);
 307                 return -ENOENT;
 308         }
 309         local_apps.allowed = 1;
 310         snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
 311                 home_dir,
 312                 LTTNG_DEFAULT_HOME_RUNDIR,
 313                 LTTNG_UST_SOCK_FILENAME);
 314         snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
 315                 LTTNG_UST_WAIT_FILENAME,
 316                 uid);
 317         return 0;
 318 }
 319
 320 /*
 321  * Get notify_sock timeout, in ms.
 322  * -1: don't wait. 0: wait forever. >0: timeout, in ms.
 323  */
 324 static
 325 long get_timeout(void)
 326 {
 327         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 328
 329         if (!got_timeout_env) {
 330                 str_timeout = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 331                 got_timeout_env = 1;
 332         }
 333         if (str_timeout)
 334                 constructor_delay_ms = strtol(str_timeout, NULL, 10);
 335         return constructor_delay_ms;
 336 }
 337
 338 static
 339 long get_notify_sock_timeout(void)
 340 {
 341         return get_timeout();
 342 }
 343
 344 /*
 345  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 346  */
 347 static
 348 int get_constructor_timeout(struct timespec *constructor_timeout)
 349 {
 350         long constructor_delay_ms;
 351         int ret;
 352
 353         constructor_delay_ms = get_timeout();
 354
 355         switch (constructor_delay_ms) {
 356         case -1:/* fall-through */
 357         case 0:
 358                 return constructor_delay_ms;
 359         default:
 360                 break;
 361         }
 362
 363         /*
 364          * If we are unable to find the current time, don't wait.
 365          */
 366         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 367         if (ret) {
 368                 return -1;
 369         }
 370         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 371         constructor_timeout->tv_nsec +=
 372                 (constructor_delay_ms % 1000UL) * 1000000UL;
 373         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 374                 constructor_timeout->tv_sec++;
 375                 constructor_timeout->tv_nsec -= 1000000000UL;
 376         }
 377         return 1;
 378 }
 379
 380 static
 381 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 382 {
 383         return ustcomm_send_reg_msg(socket,
 384                 type,
 385                 CAA_BITS_PER_LONG,
 386                 lttng_alignof(uint8_t) * CHAR_BIT,
 387                 lttng_alignof(uint16_t) * CHAR_BIT,
 388                 lttng_alignof(uint32_t) * CHAR_BIT,
 389                 lttng_alignof(uint64_t) * CHAR_BIT,
 390                 lttng_alignof(unsigned long) * CHAR_BIT);
 391 }
 392
 393 static
 394 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 395 {
 396         ssize_t len;
 397
 398         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 399         switch (len) {
 400         case sizeof(*lur):
 401                 DBG("message successfully sent");
 402                 return 0;
 403         default:
 404                 if (len == -ECONNRESET) {
 405                         DBG("remote end closed connection");
 406                         return 0;
 407                 }
 408                 if (len < 0)
 409                         return len;
 410                 DBG("incorrect message size: %zd", len);
 411                 return -EINVAL;
 412         }
 413 }
 414
 415 static
 416 int handle_register_done(struct sock_info *sock_info)
 417 {
 418         int ret;
 419
 420         if (sock_info->constructor_sem_posted)
 421                 return 0;
 422         sock_info->constructor_sem_posted = 1;
 423         if (uatomic_read(&sem_count) <= 0) {
 424                 return 0;
 425         }
 426         ret = uatomic_add_return(&sem_count, -1);
 427         if (ret == 0) {
 428                 ret = sem_post(&constructor_wait);
 429                 assert(!ret);
 430         }
 431         return 0;
 432 }
 433
 434 /*
 435  * Only execute pending statedump after the constructor semaphore has
 436  * been posted by each listener thread. This means statedump will only
 437  * be performed after the "registration done" command is received from
 438  * each session daemon the application is connected to.
 439  *
 440  * This ensures we don't run into deadlock issues with the dynamic
 441  * loader mutex, which is held while the constructor is called and
 442  * waiting on the constructor semaphore. All operations requiring this
 443  * dynamic loader lock need to be postponed using this mechanism.
 444  */
 445 static
 446 void handle_pending_statedump(struct sock_info *sock_info)
 447 {
 448         int ctor_passed = sock_info->constructor_sem_posted;
 449
 450         if (ctor_passed && sock_info->statedump_pending) {
 451                 sock_info->statedump_pending = 0;
 452                 lttng_handle_pending_statedump(sock_info);
 453         }
 454 }
 455
 456 static
 457 int handle_message(struct sock_info *sock_info,
 458                 int sock, struct ustcomm_ust_msg *lum)
 459 {
 460         int ret = 0;
 461         const struct lttng_ust_objd_ops *ops;
 462         struct ustcomm_ust_reply lur;
 463         union ust_args args;
 464         ssize_t len;
 465
 466         memset(&lur, 0, sizeof(lur));
 467
 468         if (ust_lock()) {
 469                 ret = -LTTNG_UST_ERR_EXITING;
 470                 goto end;
 471         }
 472
 473         ops = objd_ops(lum->handle);
 474         if (!ops) {
 475                 ret = -ENOENT;
 476                 goto end;
 477         }
 478
 479         switch (lum->cmd) {
 480         case LTTNG_UST_REGISTER_DONE:
 481                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 482                         ret = handle_register_done(sock_info);
 483                 else
 484                         ret = -EINVAL;
 485                 break;
 486         case LTTNG_UST_RELEASE:
 487                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 488                         ret = -EPERM;
 489                 else
 490                         ret = lttng_ust_objd_unref(lum->handle, 1);
 491                 break;
 492         case LTTNG_UST_FILTER:
 493         {
 494                 /* Receive filter data */
 495                 struct lttng_ust_filter_bytecode_node *bytecode;
 496
 497                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 498                         ERR("Filter data size is too large: %u bytes",
 499                                 lum->u.filter.data_size);
 500                         ret = -EINVAL;
 501                         goto error;
 502                 }
 503
 504                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 505                         ERR("Filter reloc offset %u is not within data",
 506                                 lum->u.filter.reloc_offset);
 507                         ret = -EINVAL;
 508                         goto error;
 509                 }
 510
 511                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 512                 if (!bytecode) {
 513                         ret = -ENOMEM;
 514                         goto error;
 515                 }
 516                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 517                                 lum->u.filter.data_size);
 518                 switch (len) {
 519                 case 0: /* orderly shutdown */
 520                         ret = 0;
 521                         free(bytecode);
 522                         goto error;
 523                 default:
 524                         if (len == lum->u.filter.data_size) {
 525                                 DBG("filter data received");
 526                                 break;
 527                         } else if (len < 0) {
 528                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 529                                 if (len == -ECONNRESET) {
 530                                         ERR("%s remote end closed connection", sock_info->name);
 531                                         ret = len;
 532                                         free(bytecode);
 533                                         goto error;
 534                                 }
 535                                 ret = len;
 536                                 free(bytecode);
 537                                 goto end;
 538                         } else {
 539                                 DBG("incorrect filter data message size: %zd", len);
 540                                 ret = -EINVAL;
 541                                 free(bytecode);
 542                                 goto end;
 543                         }
 544                 }
 545                 bytecode->bc.len = lum->u.filter.data_size;
 546                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 547                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 548                 if (ops->cmd) {
 549                         ret = ops->cmd(lum->handle, lum->cmd,
 550                                         (unsigned long) bytecode,
 551                                         &args, sock_info);
 552                         if (ret) {
 553                                 free(bytecode);
 554                         }
 555                         /* don't free bytecode if everything went fine. */
 556                 } else {
 557                         ret = -ENOSYS;
 558                         free(bytecode);
 559                 }
 560                 break;
 561         }
 562         case LTTNG_UST_EXCLUSION:
 563         {
 564                 /* Receive exclusion names */
 565                 struct lttng_ust_excluder_node *node;
 566                 unsigned int count;
 567
 568                 count = lum->u.exclusion.count;
 569                 if (count == 0) {
 570                         /* There are no names to read */
 571                         ret = 0;
 572                         goto error;
 573                 }
 574                 node = zmalloc(sizeof(*node) +
 575                                 count * LTTNG_UST_SYM_NAME_LEN);
 576                 if (!node) {
 577                         ret = -ENOMEM;
 578                         goto error;
 579                 }
 580                 node->excluder.count = count;
 581                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 582                                 count * LTTNG_UST_SYM_NAME_LEN);
 583                 switch (len) {
 584                 case 0: /* orderly shutdown */
 585                         ret = 0;
 586                         free(node);
 587                         goto error;
 588                 default:
 589                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 590                                 DBG("Exclusion data received");
 591                                 break;
 592                         } else if (len < 0) {
 593                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 594                                 if (len == -ECONNRESET) {
 595                                         ERR("%s remote end closed connection", sock_info->name);
 596                                         ret = len;
 597                                         free(node);
 598                                         goto error;
 599                                 }
 600                                 ret = len;
 601                                 free(node);
 602                                 goto end;
 603                         } else {
 604                                 DBG("Incorrect exclusion data message size: %zd", len);
 605                                 ret = -EINVAL;
 606                                 free(node);
 607                                 goto end;
 608                         }
 609                 }
 610                 if (ops->cmd) {
 611                         ret = ops->cmd(lum->handle, lum->cmd,
 612                                         (unsigned long) node,
 613                                         &args, sock_info);
 614                         if (ret) {
 615                                 free(node);
 616                         }
 617                         /* Don't free exclusion data if everything went fine. */
 618                 } else {
 619                         ret = -ENOSYS;
 620                         free(node);
 621                 }
 622                 break;
 623         }
 624         case LTTNG_UST_CHANNEL:
 625         {
 626                 void *chan_data;
 627                 int wakeup_fd;
 628
 629                 len = ustcomm_recv_channel_from_sessiond(sock,
 630                                 &chan_data, lum->u.channel.len,
 631                                 &wakeup_fd);
 632                 switch (len) {
 633                 case 0: /* orderly shutdown */
 634                         ret = 0;
 635                         goto error;
 636                 default:
 637                         if (len == lum->u.channel.len) {
 638                                 DBG("channel data received");
 639                                 break;
 640                         } else if (len < 0) {
 641                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 642                                 if (len == -ECONNRESET) {
 643                                         ERR("%s remote end closed connection", sock_info->name);
 644                                         ret = len;
 645                                         goto error;
 646                                 }
 647                                 ret = len;
 648                                 goto end;
 649                         } else {
 650                                 DBG("incorrect channel data message size: %zd", len);
 651                                 ret = -EINVAL;
 652                                 goto end;
 653                         }
 654                 }
 655                 args.channel.chan_data = chan_data;
 656                 args.channel.wakeup_fd = wakeup_fd;
 657                 if (ops->cmd)
 658                         ret = ops->cmd(lum->handle, lum->cmd,
 659                                         (unsigned long) &lum->u,
 660                                         &args, sock_info);
 661                 else
 662                         ret = -ENOSYS;
 663                 break;
 664         }
 665         case LTTNG_UST_STREAM:
 666         {
 667                 /* Receive shm_fd, wakeup_fd */
 668                 ret = ustcomm_recv_stream_from_sessiond(sock,
 669                         &lum->u.stream.len,
 670                         &args.stream.shm_fd,
 671                         &args.stream.wakeup_fd);
 672                 if (ret) {
 673                         goto end;
 674                 }
 675                 if (ops->cmd)
 676                         ret = ops->cmd(lum->handle, lum->cmd,
 677                                         (unsigned long) &lum->u,
 678                                         &args, sock_info);
 679                 else
 680                         ret = -ENOSYS;
 681                 break;
 682         }
 683         default:
 684                 if (ops->cmd)
 685                         ret = ops->cmd(lum->handle, lum->cmd,
 686                                         (unsigned long) &lum->u,
 687                                         &args, sock_info);
 688                 else
 689                         ret = -ENOSYS;
 690                 break;
 691         }
 692
 693 end:
 694         lur.handle = lum->handle;
 695         lur.cmd = lum->cmd;
 696         lur.ret_val = ret;
 697         if (ret >= 0) {
 698                 lur.ret_code = LTTNG_UST_OK;
 699         } else {
 700                 /*
 701                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 702                  * error that are not caused by the transport, except if
 703                  * we already have a more precise error message to
 704                  * report.
 705                  */
 706                 if (ret > -LTTNG_UST_ERR) {
 707                         /* Translate code to UST error. */
 708                         switch (ret) {
 709                         case -EEXIST:
 710                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 711                                 break;
 712                         case -EINVAL:
 713                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 714                                 break;
 715                         case -ENOENT:
 716                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 717                                 break;
 718                         case -EPERM:
 719                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 720                                 break;
 721                         case -ENOSYS:
 722                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 723                                 break;
 724                         default:
 725                                 lur.ret_code = -LTTNG_UST_ERR;
 726                                 break;
 727                         }
 728                 } else {
 729                         lur.ret_code = ret;
 730                 }
 731         }
 732         if (ret >= 0) {
 733                 switch (lum->cmd) {
 734                 case LTTNG_UST_TRACER_VERSION:
 735                         lur.u.version = lum->u.version;
 736                         break;
 737                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 738                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 739                         break;
 740                 }
 741         }
 742         DBG("Return value: %d", lur.ret_val);
 743         ret = send_reply(sock, &lur);
 744         if (ret < 0) {
 745                 DBG("error sending reply");
 746                 goto error;
 747         }
 748
 749         /*
 750          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 751          * after the reply.
 752          */
 753         if (lur.ret_code == LTTNG_UST_OK) {
 754                 switch (lum->cmd) {
 755                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 756                         len = ustcomm_send_unix_sock(sock,
 757                                 &args.field_list.entry,
 758                                 sizeof(args.field_list.entry));
 759                         if (len < 0) {
 760                                 ret = len;
 761                                 goto error;
 762                         }
 763                         if (len != sizeof(args.field_list.entry)) {
 764                                 ret = -EINVAL;
 765                                 goto error;
 766                         }
 767                 }
 768         }
 769
 770 error:
 771         ust_unlock();
 772
 773         /*
 774          * Performed delayed statedump operations outside of the UST
 775          * lock. We need to take the dynamic loader lock before we take
 776          * the UST lock internally within handle_pending_statedump().
 777           */
 778         handle_pending_statedump(sock_info);
 779
 780         return ret;
 781 }
 782
 783 static
 784 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 785 {
 786         int ret;
 787
 788         if (sock_info->root_handle != -1) {
 789                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 790                 if (ret) {
 791                         ERR("Error unref root handle");
 792                 }
 793                 sock_info->root_handle = -1;
 794         }
 795         sock_info->constructor_sem_posted = 0;
 796
 797         /*
 798          * wait_shm_mmap, socket and notify socket are used by listener
 799          * threads outside of the ust lock, so we cannot tear them down
 800          * ourselves, because we cannot join on these threads. Leave
 801          * responsibility of cleaning up these resources to the OS
 802          * process exit.
 803          */
 804         if (exiting)
 805                 return;
 806
 807         if (sock_info->socket != -1) {
 808                 ret = ustcomm_close_unix_sock(sock_info->socket);
 809                 if (ret) {
 810                         ERR("Error closing ust cmd socket");
 811                 }
 812                 sock_info->socket = -1;
 813         }
 814         if (sock_info->notify_socket != -1) {
 815                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 816                 if (ret) {
 817                         ERR("Error closing ust notify socket");
 818                 }
 819                 sock_info->notify_socket = -1;
 820         }
 821         if (sock_info->wait_shm_mmap) {
 822                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 823                 if (ret) {
 824                         ERR("Error unmapping wait shm");
 825                 }
 826                 sock_info->wait_shm_mmap = NULL;
 827         }
 828 }
 829
 830 /*
 831  * Using fork to set umask in the child process (not multi-thread safe).
 832  * We deal with the shm_open vs ftruncate race (happening when the
 833  * sessiond owns the shm and does not let everybody modify it, to ensure
 834  * safety against shm_unlink) by simply letting the mmap fail and
 835  * retrying after a few seconds.
 836  * For global shm, everybody has rw access to it until the sessiond
 837  * starts.
 838  */
 839 static
 840 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 841 {
 842         int wait_shm_fd, ret;
 843         pid_t pid;
 844
 845         /*
 846          * Try to open read-only.
 847          */
 848         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 849         if (wait_shm_fd >= 0) {
 850                 int32_t tmp_read;
 851                 ssize_t len;
 852                 size_t bytes_read = 0;
 853
 854                 /*
 855                  * Try to read the fd. If unable to do so, try opening
 856                  * it in write mode.
 857                  */
 858                 do {
 859                         len = read(wait_shm_fd,
 860                                 &((char *) &tmp_read)[bytes_read],
 861                                 sizeof(tmp_read) - bytes_read);
 862                         if (len > 0) {
 863                                 bytes_read += len;
 864                         }
 865                 } while ((len < 0 && errno == EINTR)
 866                         || (len > 0 && bytes_read < sizeof(tmp_read)));
 867                 if (bytes_read != sizeof(tmp_read)) {
 868                         ret = close(wait_shm_fd);
 869                         if (ret) {
 870                                 ERR("close wait_shm_fd");
 871                         }
 872                         goto open_write;
 873                 }
 874                 goto end;
 875         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 876                 /*
 877                  * Real-only open did not work, and it's not because the
 878                  * entry was not present. It's a failure that prohibits
 879                  * using shm.
 880                  */
 881                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 882                 goto end;
 883         }
 884
 885 open_write:
 886         /*
 887          * If the open failed because the file did not exist, or because
 888          * the file was not truncated yet, try creating it ourself.
 889          */
 890         URCU_TLS(lttng_ust_nest_count)++;
 891         pid = fork();
 892         URCU_TLS(lttng_ust_nest_count)--;
 893         if (pid > 0) {
 894                 int status;
 895
 896                 /*
 897                  * Parent: wait for child to return, in which case the
 898                  * shared memory map will have been created.
 899                  */
 900                 pid = wait(&status);
 901                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 902                         wait_shm_fd = -1;
 903                         goto end;
 904                 }
 905                 /*
 906                  * Try to open read-only again after creation.
 907                  */
 908                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 909                 if (wait_shm_fd < 0) {
 910                         /*
 911                          * Real-only open did not work. It's a failure
 912                          * that prohibits using shm.
 913                          */
 914                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 915                         goto end;
 916                 }
 917                 goto end;
 918         } else if (pid == 0) {
 919                 int create_mode;
 920
 921                 /* Child */
 922                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 923                 if (sock_info->global)
 924                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 925                 /*
 926                  * We're alone in a child process, so we can modify the
 927                  * process-wide umask.
 928                  */
 929                 umask(~create_mode);
 930                 /*
 931                  * Try creating shm (or get rw access).
 932                  * We don't do an exclusive open, because we allow other
 933                  * processes to create+ftruncate it concurrently.
 934                  */
 935                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 936                                 O_RDWR | O_CREAT, create_mode);
 937                 if (wait_shm_fd >= 0) {
 938                         ret = ftruncate(wait_shm_fd, mmap_size);
 939                         if (ret) {
 940                                 PERROR("ftruncate");
 941                                 _exit(EXIT_FAILURE);
 942                         }
 943                         _exit(EXIT_SUCCESS);
 944                 }
 945                 /*
 946                  * For local shm, we need to have rw access to accept
 947                  * opening it: this means the local sessiond will be
 948                  * able to wake us up. For global shm, we open it even
 949                  * if rw access is not granted, because the root.root
 950                  * sessiond will be able to override all rights and wake
 951                  * us up.
 952                  */
 953                 if (!sock_info->global && errno != EACCES) {
 954                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 955                         _exit(EXIT_FAILURE);
 956                 }
 957                 /*
 958                  * The shm exists, but we cannot open it RW. Report
 959                  * success.
 960                  */
 961                 _exit(EXIT_SUCCESS);
 962         } else {
 963                 return -1;
 964         }
 965 end:
 966         if (wait_shm_fd >= 0 && !sock_info->global) {
 967                 struct stat statbuf;
 968
 969                 /*
 970                  * Ensure that our user is the owner of the shm file for
 971                  * local shm. If we do not own the file, it means our
 972                  * sessiond will not have access to wake us up (there is
 973                  * probably a rogue process trying to fake our
 974                  * sessiond). Fallback to polling method in this case.
 975                  */
 976                 ret = fstat(wait_shm_fd, &statbuf);
 977                 if (ret) {
 978                         PERROR("fstat");
 979                         goto error_close;
 980                 }
 981                 if (statbuf.st_uid != getuid())
 982                         goto error_close;
 983         }
 984         return wait_shm_fd;
 985
 986 error_close:
 987         ret = close(wait_shm_fd);
 988         if (ret) {
 989                 PERROR("Error closing fd");
 990         }
 991         return -1;
 992 }
 993
 994 static
 995 char *get_map_shm(struct sock_info *sock_info)
 996 {
 997         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 998         int wait_shm_fd, ret;
 999         char *wait_shm_mmap;
1000
1001         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
1002         if (wait_shm_fd < 0) {
1003                 goto error;
1004         }
1005         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
1006                   MAP_SHARED, wait_shm_fd, 0);
1007         /* close shm fd immediately after taking the mmap reference */
1008         ret = close(wait_shm_fd);
1009         if (ret) {
1010                 PERROR("Error closing fd");
1011         }
1012         if (wait_shm_mmap == MAP_FAILED) {
1013                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1014                 goto error;
1015         }
1016         return wait_shm_mmap;
1017
1018 error:
1019         return NULL;
1020 }
1021
1022 static
1023 void wait_for_sessiond(struct sock_info *sock_info)
1024 {
1025         int ret;
1026
1027         if (ust_lock()) {
1028                 goto quit;
1029         }
1030         if (wait_poll_fallback) {
1031                 goto error;
1032         }
1033         if (!sock_info->wait_shm_mmap) {
1034                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1035                 if (!sock_info->wait_shm_mmap)
1036                         goto error;
1037         }
1038         ust_unlock();
1039
1040         DBG("Waiting for %s apps sessiond", sock_info->name);
1041         /* Wait for futex wakeup */
1042         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
1043                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
1044                         FUTEX_WAIT, 0, NULL, NULL, 0);
1045                 if (ret < 0) {
1046                         if (errno == EFAULT) {
1047                                 wait_poll_fallback = 1;
1048                                 DBG(
1049 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1050 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1051 "Please upgrade your kernel "
1052 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1053 "mainline). LTTng-UST will use polling mode fallback.");
1054                                 if (ust_debug())
1055                                         PERROR("futex");
1056                         }
1057                 }
1058         }
1059         return;
1060
1061 quit:
1062         ust_unlock();
1063         return;
1064
1065 error:
1066         ust_unlock();
1067         return;
1068 }
1069
1070 /*
1071  * This thread does not allocate any resource, except within
1072  * handle_message, within mutex protection. This mutex protects against
1073  * fork and exit.
1074  * The other moment it allocates resources is at socket connection, which
1075  * is also protected by the mutex.
1076  */
1077 static
1078 void *ust_listener_thread(void *arg)
1079 {
1080         struct sock_info *sock_info = arg;
1081         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1082         long timeout;
1083
1084         /* Restart trying to connect to the session daemon */
1085 restart:
1086         if (prev_connect_failed) {
1087                 /* Wait for sessiond availability with pipe */
1088                 wait_for_sessiond(sock_info);
1089                 if (has_waited) {
1090                         has_waited = 0;
1091                         /*
1092                          * Sleep for 5 seconds before retrying after a
1093                          * sequence of failure / wait / failure. This
1094                          * deals with a killed or broken session daemon.
1095                          */
1096                         sleep(5);
1097                 }
1098                 has_waited = 1;
1099                 prev_connect_failed = 0;
1100         }
1101
1102         if (sock_info->socket != -1) {
1103                 ret = ustcomm_close_unix_sock(sock_info->socket);
1104                 if (ret) {
1105                         ERR("Error closing %s ust cmd socket",
1106                                 sock_info->name);
1107                 }
1108                 sock_info->socket = -1;
1109         }
1110         if (sock_info->notify_socket != -1) {
1111                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1112                 if (ret) {
1113                         ERR("Error closing %s ust notify socket",
1114                                 sock_info->name);
1115                 }
1116                 sock_info->notify_socket = -1;
1117         }
1118
1119         /*
1120          * Register. We need to perform both connect and sending
1121          * registration message before doing the next connect otherwise
1122          * we may reach unix socket connect queue max limits and block
1123          * on the 2nd connect while the session daemon is awaiting the
1124          * first connect registration message.
1125          */
1126         /* Connect cmd socket */
1127         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1128         if (ret < 0) {
1129                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1130                 prev_connect_failed = 1;
1131
1132                 if (ust_lock()) {
1133                         goto quit;
1134                 }
1135
1136                 /*
1137                  * If we cannot find the sessiond daemon, don't delay
1138                  * constructor execution.
1139                  */
1140                 ret = handle_register_done(sock_info);
1141                 assert(!ret);
1142                 ust_unlock();
1143                 goto restart;
1144         }
1145         sock_info->socket = ret;
1146
1147         if (ust_lock()) {
1148                 goto quit;
1149         }
1150
1151         /*
1152          * Create only one root handle per listener thread for the whole
1153          * process lifetime, so we ensure we get ID which is statically
1154          * assigned to the root handle.
1155          */
1156         if (sock_info->root_handle == -1) {
1157                 ret = lttng_abi_create_root_handle();
1158                 if (ret < 0) {
1159                         ERR("Error creating root handle");
1160                         goto quit;
1161                 }
1162                 sock_info->root_handle = ret;
1163         }
1164
1165         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1166         if (ret < 0) {
1167                 ERR("Error registering to %s ust cmd socket",
1168                         sock_info->name);
1169                 prev_connect_failed = 1;
1170                 /*
1171                  * If we cannot register to the sessiond daemon, don't
1172                  * delay constructor execution.
1173                  */
1174                 ret = handle_register_done(sock_info);
1175                 assert(!ret);
1176                 ust_unlock();
1177                 goto restart;
1178         }
1179
1180         ust_unlock();
1181
1182         /* Connect notify socket */
1183         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1184         if (ret < 0) {
1185                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1186                 prev_connect_failed = 1;
1187
1188                 if (ust_lock()) {
1189                         goto quit;
1190                 }
1191
1192                 /*
1193                  * If we cannot find the sessiond daemon, don't delay
1194                  * constructor execution.
1195                  */
1196                 ret = handle_register_done(sock_info);
1197                 assert(!ret);
1198                 ust_unlock();
1199                 goto restart;
1200         }
1201         sock_info->notify_socket = ret;
1202
1203         timeout = get_notify_sock_timeout();
1204         if (timeout >= 0) {
1205                 /*
1206                  * Give at least 10ms to sessiond to reply to
1207                  * notifications.
1208                  */
1209                 if (timeout < 10)
1210                         timeout = 10;
1211                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1212                                 timeout);
1213                 if (ret < 0) {
1214                         WARN("Error setting socket receive timeout");
1215                 }
1216                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1217                                 timeout);
1218                 if (ret < 0) {
1219                         WARN("Error setting socket send timeout");
1220                 }
1221         } else if (timeout < -1) {
1222                 WARN("Unsupported timeout value %ld", timeout);
1223         }
1224
1225         if (ust_lock()) {
1226                 goto quit;
1227         }
1228
1229         ret = register_to_sessiond(sock_info->notify_socket,
1230                         USTCTL_SOCKET_NOTIFY);
1231         if (ret < 0) {
1232                 ERR("Error registering to %s ust notify socket",
1233                         sock_info->name);
1234                 prev_connect_failed = 1;
1235                 /*
1236                  * If we cannot register to the sessiond daemon, don't
1237                  * delay constructor execution.
1238                  */
1239                 ret = handle_register_done(sock_info);
1240                 assert(!ret);
1241                 ust_unlock();
1242                 goto restart;
1243         }
1244         sock = sock_info->socket;
1245
1246         ust_unlock();
1247
1248         for (;;) {
1249                 ssize_t len;
1250                 struct ustcomm_ust_msg lum;
1251
1252                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1253                 switch (len) {
1254                 case 0: /* orderly shutdown */
1255                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1256                         if (ust_lock()) {
1257                                 goto quit;
1258                         }
1259                         /*
1260                          * Either sessiond has shutdown or refused us by closing the socket.
1261                          * In either case, we don't want to delay construction execution,
1262                          * and we need to wait before retry.
1263                          */
1264                         prev_connect_failed = 1;
1265                         /*
1266                          * If we cannot register to the sessiond daemon, don't
1267                          * delay constructor execution.
1268                          */
1269                         ret = handle_register_done(sock_info);
1270                         assert(!ret);
1271                         ust_unlock();
1272                         goto end;
1273                 case sizeof(lum):
1274                         print_cmd(lum.cmd, lum.handle);
1275                         ret = handle_message(sock_info, sock, &lum);
1276                         if (ret) {
1277                                 ERR("Error handling message for %s socket", sock_info->name);
1278                         }
1279                         continue;
1280                 default:
1281                         if (len < 0) {
1282                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1283                         } else {
1284                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1285                         }
1286                         if (len == -ECONNRESET) {
1287                                 DBG("%s remote end closed connection", sock_info->name);
1288                                 goto end;
1289                         }
1290                         goto end;
1291                 }
1292
1293         }
1294 end:
1295         if (ust_lock()) {
1296                 goto quit;
1297         }
1298         /* Cleanup socket handles before trying to reconnect */
1299         lttng_ust_objd_table_owner_cleanup(sock_info);
1300         ust_unlock();
1301         goto restart;   /* try to reconnect */
1302
1303 quit:
1304         ust_unlock();
1305
1306         pthread_mutex_lock(&ust_exit_mutex);
1307         sock_info->thread_active = 0;
1308         pthread_mutex_unlock(&ust_exit_mutex);
1309         return NULL;
1310 }
1311
1312 /*
1313  * Weak symbol to call when the ust malloc wrapper is not loaded.
1314  */
1315 __attribute__((weak))
1316 void lttng_ust_malloc_wrapper_init(void)
1317 {
1318 }
1319
1320 /*
1321  * sessiond monitoring thread: monitor presence of global and per-user
1322  * sessiond by polling the application common named pipe.
1323  */
1324 void __attribute__((constructor)) lttng_ust_init(void)
1325 {
1326         struct timespec constructor_timeout;
1327         sigset_t sig_all_blocked, orig_parent_mask;
1328         pthread_attr_t thread_attr;
1329         int timeout_mode;
1330         int ret;
1331
1332         if (uatomic_xchg(&initialized, 1) == 1)
1333                 return;
1334
1335         /*
1336          * Fixup interdependency between TLS fixup mutex (which happens
1337          * to be the dynamic linker mutex) and ust_lock, taken within
1338          * the ust lock.
1339          */
1340         lttng_fixup_ringbuffer_tls();
1341         lttng_fixup_vtid_tls();
1342         lttng_fixup_nest_count_tls();
1343         lttng_fixup_procname_tls();
1344
1345         /*
1346          * We want precise control over the order in which we construct
1347          * our sub-libraries vs starting to receive commands from
1348          * sessiond (otherwise leading to errors when trying to create
1349          * sessiond before the init functions are completed).
1350          */
1351         init_usterr();
1352         init_tracepoint();
1353         lttng_ust_baddr_statedump_init();
1354         lttng_ring_buffer_metadata_client_init();
1355         lttng_ring_buffer_client_overwrite_init();
1356         lttng_ring_buffer_client_overwrite_rt_init();
1357         lttng_ring_buffer_client_discard_init();
1358         lttng_ring_buffer_client_discard_rt_init();
1359         lttng_context_init();
1360         /*
1361          * Invoke ust malloc wrapper init before starting other threads.
1362          */
1363         lttng_ust_malloc_wrapper_init();
1364
1365         timeout_mode = get_constructor_timeout(&constructor_timeout);
1366
1367         ret = sem_init(&constructor_wait, 0, 0);
1368         assert(!ret);
1369
1370         ret = setup_local_apps();
1371         if (ret) {
1372                 DBG("local apps setup returned %d", ret);
1373         }
1374
1375         /* A new thread created by pthread_create inherits the signal mask
1376          * from the parent. To avoid any signal being received by the
1377          * listener thread, we block all signals temporarily in the parent,
1378          * while we create the listener thread.
1379          */
1380         sigfillset(&sig_all_blocked);
1381         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1382         if (ret) {
1383                 ERR("pthread_sigmask: %s", strerror(ret));
1384         }
1385
1386         ret = pthread_attr_init(&thread_attr);
1387         if (ret) {
1388                 ERR("pthread_attr_init: %s", strerror(ret));
1389         }
1390         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1391         if (ret) {
1392                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1393         }
1394
1395         pthread_mutex_lock(&ust_exit_mutex);
1396         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1397                         ust_listener_thread, &global_apps);
1398         if (ret) {
1399                 ERR("pthread_create global: %s", strerror(ret));
1400         }
1401         global_apps.thread_active = 1;
1402         pthread_mutex_unlock(&ust_exit_mutex);
1403
1404         if (local_apps.allowed) {
1405                 pthread_mutex_lock(&ust_exit_mutex);
1406                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1407                                 ust_listener_thread, &local_apps);
1408                 if (ret) {
1409                         ERR("pthread_create local: %s", strerror(ret));
1410                 }
1411                 local_apps.thread_active = 1;
1412                 pthread_mutex_unlock(&ust_exit_mutex);
1413         } else {
1414                 handle_register_done(&local_apps);
1415         }
1416         ret = pthread_attr_destroy(&thread_attr);
1417         if (ret) {
1418                 ERR("pthread_attr_destroy: %s", strerror(ret));
1419         }
1420
1421         /* Restore original signal mask in parent */
1422         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1423         if (ret) {
1424                 ERR("pthread_sigmask: %s", strerror(ret));
1425         }
1426
1427         switch (timeout_mode) {
1428         case 1: /* timeout wait */
1429                 do {
1430                         ret = sem_timedwait(&constructor_wait,
1431                                         &constructor_timeout);
1432                 } while (ret < 0 && errno == EINTR);
1433                 if (ret < 0 && errno == ETIMEDOUT) {
1434                         ERR("Timed out waiting for lttng-sessiond");
1435                 } else {
1436                         assert(!ret);
1437                 }
1438                 break;
1439         case -1:/* wait forever */
1440                 do {
1441                         ret = sem_wait(&constructor_wait);
1442                 } while (ret < 0 && errno == EINTR);
1443                 assert(!ret);
1444                 break;
1445         case 0: /* no timeout */
1446                 break;
1447         }
1448 }
1449
1450 static
1451 void lttng_ust_cleanup(int exiting)
1452 {
1453         cleanup_sock_info(&global_apps, exiting);
1454         if (local_apps.allowed) {
1455                 cleanup_sock_info(&local_apps, exiting);
1456         }
1457         /*
1458          * The teardown in this function all affect data structures
1459          * accessed under the UST lock by the listener thread. This
1460          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1461          * that none of these threads are accessing this data at this
1462          * point.
1463          */
1464         lttng_ust_abi_exit();
1465         lttng_ust_events_exit();
1466         lttng_context_exit();
1467         lttng_ring_buffer_client_discard_rt_exit();
1468         lttng_ring_buffer_client_discard_exit();
1469         lttng_ring_buffer_client_overwrite_rt_exit();
1470         lttng_ring_buffer_client_overwrite_exit();
1471         lttng_ring_buffer_metadata_client_exit();
1472         lttng_ust_baddr_statedump_destroy();
1473         exit_tracepoint();
1474         if (!exiting) {
1475                 /* Reinitialize values for fork */
1476                 sem_count = 2;
1477                 lttng_ust_comm_should_quit = 0;
1478                 initialized = 0;
1479         }
1480 }
1481
1482 void __attribute__((destructor)) lttng_ust_exit(void)
1483 {
1484         int ret;
1485
1486         /*
1487          * Using pthread_cancel here because:
1488          * A) we don't want to hang application teardown.
1489          * B) the thread is not allocating any resource.
1490          */
1491
1492         /*
1493          * Require the communication thread to quit. Synchronize with
1494          * mutexes to ensure it is not in a mutex critical section when
1495          * pthread_cancel is later called.
1496          */
1497         ust_lock_nocheck();
1498         lttng_ust_comm_should_quit = 1;
1499         ust_unlock();
1500
1501         pthread_mutex_lock(&ust_exit_mutex);
1502         /* cancel threads */
1503         if (global_apps.thread_active) {
1504                 ret = pthread_cancel(global_apps.ust_listener);
1505                 if (ret) {
1506                         ERR("Error cancelling global ust listener thread: %s",
1507                                 strerror(ret));
1508                 } else {
1509                         global_apps.thread_active = 0;
1510                 }
1511         }
1512         if (local_apps.thread_active) {
1513                 ret = pthread_cancel(local_apps.ust_listener);
1514                 if (ret) {
1515                         ERR("Error cancelling local ust listener thread: %s",
1516                                 strerror(ret));
1517                 } else {
1518                         local_apps.thread_active = 0;
1519                 }
1520         }
1521         pthread_mutex_unlock(&ust_exit_mutex);
1522
1523         /*
1524          * Do NOT join threads: use of sys_futex makes it impossible to
1525          * join the threads without using async-cancel, but async-cancel
1526          * is delivered by a signal, which could hit the target thread
1527          * anywhere in its code path, including while the ust_lock() is
1528          * held, causing a deadlock for the other thread. Let the OS
1529          * cleanup the threads if there are stalled in a syscall.
1530          */
1531         lttng_ust_cleanup(1);
1532 }
1533
1534 /*
1535  * We exclude the worker threads across fork and clone (except
1536  * CLONE_VM), because these system calls only keep the forking thread
1537  * running in the child.  Therefore, we don't want to call fork or clone
1538  * in the middle of an tracepoint or ust tracing state modification.
1539  * Holding this mutex protects these structures across fork and clone.
1540  */
1541 void ust_before_fork(sigset_t *save_sigset)
1542 {
1543         /*
1544          * Disable signals. This is to avoid that the child intervenes
1545          * before it is properly setup for tracing. It is safer to
1546          * disable all signals, because then we know we are not breaking
1547          * anything by restoring the original mask.
1548          */
1549         sigset_t all_sigs;
1550         int ret;
1551
1552         if (URCU_TLS(lttng_ust_nest_count))
1553                 return;
1554         /* Disable signals */
1555         sigfillset(&all_sigs);
1556         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1557         if (ret == -1) {
1558                 PERROR("sigprocmask");
1559         }
1560         ust_lock_nocheck();
1561         rcu_bp_before_fork();
1562 }
1563
1564 static void ust_after_fork_common(sigset_t *restore_sigset)
1565 {
1566         int ret;
1567
1568         DBG("process %d", getpid());
1569         ust_unlock();
1570         /* Restore signals */
1571         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1572         if (ret == -1) {
1573                 PERROR("sigprocmask");
1574         }
1575 }
1576
1577 void ust_after_fork_parent(sigset_t *restore_sigset)
1578 {
1579         if (URCU_TLS(lttng_ust_nest_count))
1580                 return;
1581         DBG("process %d", getpid());
1582         rcu_bp_after_fork_parent();
1583         /* Release mutexes and reenable signals */
1584         ust_after_fork_common(restore_sigset);
1585 }
1586
1587 /*
1588  * After fork, in the child, we need to cleanup all the leftover state,
1589  * except the worker thread which already magically disappeared thanks
1590  * to the weird Linux fork semantics. After tyding up, we call
1591  * lttng_ust_init() again to start over as a new PID.
1592  *
1593  * This is meant for forks() that have tracing in the child between the
1594  * fork and following exec call (if there is any).
1595  */
1596 void ust_after_fork_child(sigset_t *restore_sigset)
1597 {
1598         if (URCU_TLS(lttng_ust_nest_count))
1599                 return;
1600         DBG("process %d", getpid());
1601         /* Release urcu mutexes */
1602         rcu_bp_after_fork_child();
1603         lttng_ust_cleanup(0);
1604         lttng_context_vtid_reset();
1605         /* Release mutexes and reenable signals */
1606         ust_after_fork_common(restore_sigset);
1607         lttng_ust_init();
1608 }
1609
1610 void lttng_ust_sockinfo_session_enabled(void *owner)
1611 {
1612         struct sock_info *sock_info = owner;
1613         sock_info->statedump_pending = 1;
1614 }