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 #define _GNU_SOURCE
  24 #include <sys/types.h>
  25 #include <sys/socket.h>
  26 #include <sys/mman.h>
  27 #include <sys/stat.h>
  28 #include <sys/types.h>
  29 #include <sys/wait.h>
  30 #include <fcntl.h>
  31 #include <unistd.h>
  32 #include <errno.h>
  33 #include <pthread.h>
  34 #include <semaphore.h>
  35 #include <time.h>
  36 #include <assert.h>
  37 #include <signal.h>
  38 #include <urcu/uatomic.h>
  39 #include <urcu/futex.h>
  40 #include <urcu/compiler.h>
  41
  42 #include <lttng/ust-events.h>
  43 #include <lttng/ust-abi.h>
  44 #include <lttng/ust.h>
  45 #include <lttng/ust-error.h>
  46 #include <lttng/ust-ctl.h>
  47 #include <urcu/tls-compat.h>
  48 #include <ust-comm.h>
  49 #include <usterr-signal-safe.h>
  50 #include <helper.h>
  51 #include "tracepoint-internal.h"
  52 #include "lttng-tracer-core.h"
  53 #include "compat.h"
  54 #include "../libringbuffer/tlsfixup.h"
  55 #include "lttng-ust-baddr.h"
  56 #include "clock.h"
  57 #include "../libringbuffer/getcpu.h"
  58
  59 /*
  60  * Has lttng ust comm constructor been called ?
  61  */
  62 static int initialized;
  63
  64 /*
  65  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  66  * Held when handling a command, also held by fork() to deal with
  67  * removal of threads, and by exit path.
  68  *
  69  * The UST lock is the centralized mutex across UST tracing control and
  70  * probe registration.
  71  *
  72  * ust_exit_mutex must never nest in ust_mutex.
  73  *
  74  * ust_fork_mutex must never nest in ust_mutex.
  75  *
  76  * ust_mutex_nest is a per-thread nesting counter, allowing the perf
  77  * counter lazy initialization called by events within the statedump,
  78  * which traces while the ust_mutex is held.
  79  *
  80  * ust_lock nests within the dynamic loader lock (within glibc) because
  81  * it is taken within the library constructor.
  82  */
  83 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
  84
  85 /* Allow nesting the ust_mutex within the same thread. */
  86 static DEFINE_URCU_TLS(int, ust_mutex_nest);
  87
  88 /*
  89  * ust_exit_mutex protects thread_active variable wrt thread exit. It
  90  * cannot be done by ust_mutex because pthread_cancel(), which takes an
  91  * internal libc lock, cannot nest within ust_mutex.
  92  *
  93  * It never nests within a ust_mutex.
  94  */
  95 static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
  96
  97 /*
  98  * ust_fork_mutex protects base address statedump tracing against forks. It
  99  * prevents the dynamic loader lock to be taken (by base address statedump
 100  * tracing) while a fork is happening, thus preventing deadlock issues with
 101  * the dynamic loader lock.
 102  */
 103 static pthread_mutex_t ust_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
 104
 105 /* Should the ust comm thread quit ? */
 106 static int lttng_ust_comm_should_quit;
 107
 108 /*
 109  * Return 0 on success, -1 if should quit.
 110  * The lock is taken in both cases.
 111  * Signal-safe.
 112  */
 113 int ust_lock(void)
 114 {
 115         sigset_t sig_all_blocked, orig_mask;
 116         int ret;
 117
 118         sigfillset(&sig_all_blocked);
 119         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 120         if (ret) {
 121                 ERR("pthread_sigmask: %s", strerror(ret));
 122         }
 123         if (!URCU_TLS(ust_mutex_nest)++)
 124                 pthread_mutex_lock(&ust_mutex);
 125         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 126         if (ret) {
 127                 ERR("pthread_sigmask: %s", strerror(ret));
 128         }
 129         if (lttng_ust_comm_should_quit) {
 130                 return -1;
 131         } else {
 132                 return 0;
 133         }
 134 }
 135
 136 /*
 137  * ust_lock_nocheck() can be used in constructors/destructors, because
 138  * they are already nested within the dynamic loader lock, and therefore
 139  * have exclusive access against execution of liblttng-ust destructor.
 140  * Signal-safe.
 141  */
 142 void ust_lock_nocheck(void)
 143 {
 144         sigset_t sig_all_blocked, orig_mask;
 145         int ret;
 146
 147         sigfillset(&sig_all_blocked);
 148         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 149         if (ret) {
 150                 ERR("pthread_sigmask: %s", strerror(ret));
 151         }
 152         if (!URCU_TLS(ust_mutex_nest)++)
 153                 pthread_mutex_lock(&ust_mutex);
 154         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 155         if (ret) {
 156                 ERR("pthread_sigmask: %s", strerror(ret));
 157         }
 158 }
 159
 160 /*
 161  * Signal-safe.
 162  */
 163 void ust_unlock(void)
 164 {
 165         sigset_t sig_all_blocked, orig_mask;
 166         int ret;
 167
 168         sigfillset(&sig_all_blocked);
 169         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 170         if (ret) {
 171                 ERR("pthread_sigmask: %s", strerror(ret));
 172         }
 173         if (!--URCU_TLS(ust_mutex_nest))
 174                 pthread_mutex_unlock(&ust_mutex);
 175         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 176         if (ret) {
 177                 ERR("pthread_sigmask: %s", strerror(ret));
 178         }
 179 }
 180
 181 /*
 182  * Wait for either of these before continuing to the main
 183  * program:
 184  * - the register_done message from sessiond daemon
 185  *   (will let the sessiond daemon enable sessions before main
 186  *   starts.)
 187  * - sessiond daemon is not reachable.
 188  * - timeout (ensuring applications are resilient to session
 189  *   daemon problems).
 190  */
 191 static sem_t constructor_wait;
 192 /*
 193  * Doing this for both the global and local sessiond.
 194  */
 195 static int sem_count = { 2 };
 196
 197 /*
 198  * Counting nesting within lttng-ust. Used to ensure that calling fork()
 199  * from liblttng-ust does not execute the pre/post fork handlers.
 200  */
 201 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
 202
 203 /*
 204  * Info about socket and associated listener thread.
 205  */
 206 struct sock_info {
 207         const char *name;
 208         pthread_t ust_listener; /* listener thread */
 209         int root_handle;
 210         int constructor_sem_posted;
 211         int allowed;
 212         int global;
 213         int thread_active;
 214
 215         char sock_path[PATH_MAX];
 216         int socket;
 217         int notify_socket;
 218
 219         char wait_shm_path[PATH_MAX];
 220         char *wait_shm_mmap;
 221         /* Keep track of lazy state dump not performed yet. */
 222         int statedump_pending;
 223 };
 224
 225 /* Socket from app (connect) to session daemon (listen) for communication */
 226 struct sock_info global_apps = {
 227         .name = "global",
 228         .global = 1,
 229
 230         .root_handle = -1,
 231         .allowed = 1,
 232         .thread_active = 0,
 233
 234         .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
 235         .socket = -1,
 236         .notify_socket = -1,
 237
 238         .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 239
 240         .statedump_pending = 0,
 241 };
 242
 243 /* TODO: allow global_apps_sock_path override */
 244
 245 struct sock_info local_apps = {
 246         .name = "local",
 247         .global = 0,
 248         .root_handle = -1,
 249         .allowed = 0,   /* Check setuid bit first */
 250         .thread_active = 0,
 251
 252         .socket = -1,
 253         .notify_socket = -1,
 254
 255         .statedump_pending = 0,
 256 };
 257
 258 static int wait_poll_fallback;
 259
 260 static const char *cmd_name_mapping[] = {
 261         [ LTTNG_UST_RELEASE ] = "Release",
 262         [ LTTNG_UST_SESSION ] = "Create Session",
 263         [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
 264
 265         [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
 266         [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
 267         [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
 268         [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
 269
 270         /* Session FD commands */
 271         [ LTTNG_UST_CHANNEL ] = "Create Channel",
 272         [ LTTNG_UST_SESSION_START ] = "Start Session",
 273         [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
 274
 275         /* Channel FD commands */
 276         [ LTTNG_UST_STREAM ] = "Create Stream",
 277         [ LTTNG_UST_EVENT ] = "Create Event",
 278
 279         /* Event and Channel FD commands */
 280         [ LTTNG_UST_CONTEXT ] = "Create Context",
 281         [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
 282
 283         /* Event, Channel and Session commands */
 284         [ LTTNG_UST_ENABLE ] = "Enable",
 285         [ LTTNG_UST_DISABLE ] = "Disable",
 286
 287         /* Tracepoint list commands */
 288         [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
 289         [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
 290
 291         /* Event FD commands */
 292         [ LTTNG_UST_FILTER ] = "Create Filter",
 293         [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
 294 };
 295
 296 static const char *str_timeout;
 297 static int got_timeout_env;
 298
 299 extern void lttng_ring_buffer_client_overwrite_init(void);
 300 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 301 extern void lttng_ring_buffer_client_discard_init(void);
 302 extern void lttng_ring_buffer_client_discard_rt_init(void);
 303 extern void lttng_ring_buffer_metadata_client_init(void);
 304 extern void lttng_ring_buffer_client_overwrite_exit(void);
 305 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 306 extern void lttng_ring_buffer_client_discard_exit(void);
 307 extern void lttng_ring_buffer_client_discard_rt_exit(void);
 308 extern void lttng_ring_buffer_metadata_client_exit(void);
 309
 310 /*
 311  * Returns the HOME directory path. Caller MUST NOT free(3) the returned
 312  * pointer.
 313  */
 314 static
 315 const char *get_lttng_home_dir(void)
 316 {
 317        const char *val;
 318
 319        val = (const char *) getenv("LTTNG_HOME");
 320        if (val != NULL) {
 321                return val;
 322        }
 323        return (const char *) getenv("HOME");
 324 }
 325
 326 /*
 327  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 328  */
 329 static
 330 void lttng_fixup_nest_count_tls(void)
 331 {
 332         asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 333 }
 334
 335 static
 336 void lttng_fixup_ust_mutex_nest_tls(void)
 337 {
 338         asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest)));
 339 }
 340
 341 /*
 342  * Fixup urcu bp TLS.
 343  */
 344 static
 345 void lttng_fixup_urcu_bp_tls(void)
 346 {
 347         rcu_read_lock();
 348         rcu_read_unlock();
 349 }
 350
 351 int lttng_get_notify_socket(void *owner)
 352 {
 353         struct sock_info *info = owner;
 354
 355         return info->notify_socket;
 356 }
 357
 358 static
 359 void print_cmd(int cmd, int handle)
 360 {
 361         const char *cmd_name = "Unknown";
 362
 363         if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
 364                         && cmd_name_mapping[cmd]) {
 365                 cmd_name = cmd_name_mapping[cmd];
 366         }
 367         DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
 368                 cmd_name, cmd,
 369                 lttng_ust_obj_get_name(handle), handle);
 370 }
 371
 372 static
 373 int setup_local_apps(void)
 374 {
 375         const char *home_dir;
 376         uid_t uid;
 377
 378         uid = getuid();
 379         /*
 380          * Disallow per-user tracing for setuid binaries.
 381          */
 382         if (uid != geteuid()) {
 383                 assert(local_apps.allowed == 0);
 384                 return 0;
 385         }
 386         home_dir = get_lttng_home_dir();
 387         if (!home_dir) {
 388                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 389                 assert(local_apps.allowed == 0);
 390                 return -ENOENT;
 391         }
 392         local_apps.allowed = 1;
 393         snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
 394                 home_dir,
 395                 LTTNG_DEFAULT_HOME_RUNDIR,
 396                 LTTNG_UST_SOCK_FILENAME);
 397         snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
 398                 LTTNG_UST_WAIT_FILENAME,
 399                 uid);
 400         return 0;
 401 }
 402
 403 /*
 404  * Get notify_sock timeout, in ms.
 405  * -1: don't wait. 0: wait forever. >0: timeout, in ms.
 406  */
 407 static
 408 long get_timeout(void)
 409 {
 410         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 411
 412         if (!got_timeout_env) {
 413                 str_timeout = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 414                 got_timeout_env = 1;
 415         }
 416         if (str_timeout)
 417                 constructor_delay_ms = strtol(str_timeout, NULL, 10);
 418         return constructor_delay_ms;
 419 }
 420
 421 static
 422 long get_notify_sock_timeout(void)
 423 {
 424         return get_timeout();
 425 }
 426
 427 /*
 428  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 429  */
 430 static
 431 int get_constructor_timeout(struct timespec *constructor_timeout)
 432 {
 433         long constructor_delay_ms;
 434         int ret;
 435
 436         constructor_delay_ms = get_timeout();
 437
 438         switch (constructor_delay_ms) {
 439         case -1:/* fall-through */
 440         case 0:
 441                 return constructor_delay_ms;
 442         default:
 443                 break;
 444         }
 445
 446         /*
 447          * If we are unable to find the current time, don't wait.
 448          */
 449         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 450         if (ret) {
 451                 return -1;
 452         }
 453         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 454         constructor_timeout->tv_nsec +=
 455                 (constructor_delay_ms % 1000UL) * 1000000UL;
 456         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 457                 constructor_timeout->tv_sec++;
 458                 constructor_timeout->tv_nsec -= 1000000000UL;
 459         }
 460         return 1;
 461 }
 462
 463 static
 464 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 465 {
 466         return ustcomm_send_reg_msg(socket,
 467                 type,
 468                 CAA_BITS_PER_LONG,
 469                 lttng_alignof(uint8_t) * CHAR_BIT,
 470                 lttng_alignof(uint16_t) * CHAR_BIT,
 471                 lttng_alignof(uint32_t) * CHAR_BIT,
 472                 lttng_alignof(uint64_t) * CHAR_BIT,
 473                 lttng_alignof(unsigned long) * CHAR_BIT);
 474 }
 475
 476 static
 477 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 478 {
 479         ssize_t len;
 480
 481         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 482         switch (len) {
 483         case sizeof(*lur):
 484                 DBG("message successfully sent");
 485                 return 0;
 486         default:
 487                 if (len == -ECONNRESET) {
 488                         DBG("remote end closed connection");
 489                         return 0;
 490                 }
 491                 if (len < 0)
 492                         return len;
 493                 DBG("incorrect message size: %zd", len);
 494                 return -EINVAL;
 495         }
 496 }
 497
 498 static
 499 int handle_register_done(struct sock_info *sock_info)
 500 {
 501         int ret;
 502
 503         if (sock_info->constructor_sem_posted)
 504                 return 0;
 505         sock_info->constructor_sem_posted = 1;
 506         if (uatomic_read(&sem_count) <= 0) {
 507                 return 0;
 508         }
 509         ret = uatomic_add_return(&sem_count, -1);
 510         if (ret == 0) {
 511                 ret = sem_post(&constructor_wait);
 512                 assert(!ret);
 513         }
 514         return 0;
 515 }
 516
 517 /*
 518  * Only execute pending statedump after the constructor semaphore has
 519  * been posted by each listener thread. This means statedump will only
 520  * be performed after the "registration done" command is received from
 521  * each session daemon the application is connected to.
 522  *
 523  * This ensures we don't run into deadlock issues with the dynamic
 524  * loader mutex, which is held while the constructor is called and
 525  * waiting on the constructor semaphore. All operations requiring this
 526  * dynamic loader lock need to be postponed using this mechanism.
 527  */
 528 static
 529 void handle_pending_statedump(struct sock_info *sock_info)
 530 {
 531         int ctor_passed = sock_info->constructor_sem_posted;
 532
 533         if (ctor_passed && sock_info->statedump_pending) {
 534                 sock_info->statedump_pending = 0;
 535                 pthread_mutex_lock(&ust_fork_mutex);
 536                 lttng_handle_pending_statedump(sock_info);
 537                 pthread_mutex_unlock(&ust_fork_mutex);
 538         }
 539 }
 540
 541 static
 542 int handle_message(struct sock_info *sock_info,
 543                 int sock, struct ustcomm_ust_msg *lum)
 544 {
 545         int ret = 0;
 546         const struct lttng_ust_objd_ops *ops;
 547         struct ustcomm_ust_reply lur;
 548         union ust_args args;
 549         ssize_t len;
 550
 551         memset(&lur, 0, sizeof(lur));
 552
 553         if (ust_lock()) {
 554                 ret = -LTTNG_UST_ERR_EXITING;
 555                 goto end;
 556         }
 557
 558         ops = objd_ops(lum->handle);
 559         if (!ops) {
 560                 ret = -ENOENT;
 561                 goto end;
 562         }
 563
 564         switch (lum->cmd) {
 565         case LTTNG_UST_REGISTER_DONE:
 566                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 567                         ret = handle_register_done(sock_info);
 568                 else
 569                         ret = -EINVAL;
 570                 break;
 571         case LTTNG_UST_RELEASE:
 572                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 573                         ret = -EPERM;
 574                 else
 575                         ret = lttng_ust_objd_unref(lum->handle, 1);
 576                 break;
 577         case LTTNG_UST_FILTER:
 578         {
 579                 /* Receive filter data */
 580                 struct lttng_ust_filter_bytecode_node *bytecode;
 581
 582                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 583                         ERR("Filter data size is too large: %u bytes",
 584                                 lum->u.filter.data_size);
 585                         ret = -EINVAL;
 586                         goto error;
 587                 }
 588
 589                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 590                         ERR("Filter reloc offset %u is not within data",
 591                                 lum->u.filter.reloc_offset);
 592                         ret = -EINVAL;
 593                         goto error;
 594                 }
 595
 596                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 597                 if (!bytecode) {
 598                         ret = -ENOMEM;
 599                         goto error;
 600                 }
 601                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 602                                 lum->u.filter.data_size);
 603                 switch (len) {
 604                 case 0: /* orderly shutdown */
 605                         ret = 0;
 606                         free(bytecode);
 607                         goto error;
 608                 default:
 609                         if (len == lum->u.filter.data_size) {
 610                                 DBG("filter data received");
 611                                 break;
 612                         } else if (len < 0) {
 613                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 614                                 if (len == -ECONNRESET) {
 615                                         ERR("%s remote end closed connection", sock_info->name);
 616                                         ret = len;
 617                                         free(bytecode);
 618                                         goto error;
 619                                 }
 620                                 ret = len;
 621                                 free(bytecode);
 622                                 goto end;
 623                         } else {
 624                                 DBG("incorrect filter data message size: %zd", len);
 625                                 ret = -EINVAL;
 626                                 free(bytecode);
 627                                 goto end;
 628                         }
 629                 }
 630                 bytecode->bc.len = lum->u.filter.data_size;
 631                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 632                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 633                 if (ops->cmd) {
 634                         ret = ops->cmd(lum->handle, lum->cmd,
 635                                         (unsigned long) bytecode,
 636                                         &args, sock_info);
 637                         if (ret) {
 638                                 free(bytecode);
 639                         }
 640                         /* don't free bytecode if everything went fine. */
 641                 } else {
 642                         ret = -ENOSYS;
 643                         free(bytecode);
 644                 }
 645                 break;
 646         }
 647         case LTTNG_UST_EXCLUSION:
 648         {
 649                 /* Receive exclusion names */
 650                 struct lttng_ust_excluder_node *node;
 651                 unsigned int count;
 652
 653                 count = lum->u.exclusion.count;
 654                 if (count == 0) {
 655                         /* There are no names to read */
 656                         ret = 0;
 657                         goto error;
 658                 }
 659                 node = zmalloc(sizeof(*node) +
 660                                 count * LTTNG_UST_SYM_NAME_LEN);
 661                 if (!node) {
 662                         ret = -ENOMEM;
 663                         goto error;
 664                 }
 665                 node->excluder.count = count;
 666                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 667                                 count * LTTNG_UST_SYM_NAME_LEN);
 668                 switch (len) {
 669                 case 0: /* orderly shutdown */
 670                         ret = 0;
 671                         free(node);
 672                         goto error;
 673                 default:
 674                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 675                                 DBG("Exclusion data received");
 676                                 break;
 677                         } else if (len < 0) {
 678                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 679                                 if (len == -ECONNRESET) {
 680                                         ERR("%s remote end closed connection", sock_info->name);
 681                                         ret = len;
 682                                         free(node);
 683                                         goto error;
 684                                 }
 685                                 ret = len;
 686                                 free(node);
 687                                 goto end;
 688                         } else {
 689                                 DBG("Incorrect exclusion data message size: %zd", len);
 690                                 ret = -EINVAL;
 691                                 free(node);
 692                                 goto end;
 693                         }
 694                 }
 695                 if (ops->cmd) {
 696                         ret = ops->cmd(lum->handle, lum->cmd,
 697                                         (unsigned long) node,
 698                                         &args, sock_info);
 699                         if (ret) {
 700                                 free(node);
 701                         }
 702                         /* Don't free exclusion data if everything went fine. */
 703                 } else {
 704                         ret = -ENOSYS;
 705                         free(node);
 706                 }
 707                 break;
 708         }
 709         case LTTNG_UST_CHANNEL:
 710         {
 711                 void *chan_data;
 712                 int wakeup_fd;
 713
 714                 len = ustcomm_recv_channel_from_sessiond(sock,
 715                                 &chan_data, lum->u.channel.len,
 716                                 &wakeup_fd);
 717                 switch (len) {
 718                 case 0: /* orderly shutdown */
 719                         ret = 0;
 720                         goto error;
 721                 default:
 722                         if (len == lum->u.channel.len) {
 723                                 DBG("channel data received");
 724                                 break;
 725                         } else if (len < 0) {
 726                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 727                                 if (len == -ECONNRESET) {
 728                                         ERR("%s remote end closed connection", sock_info->name);
 729                                         ret = len;
 730                                         goto error;
 731                                 }
 732                                 ret = len;
 733                                 goto end;
 734                         } else {
 735                                 DBG("incorrect channel data message size: %zd", len);
 736                                 ret = -EINVAL;
 737                                 goto end;
 738                         }
 739                 }
 740                 args.channel.chan_data = chan_data;
 741                 args.channel.wakeup_fd = wakeup_fd;
 742                 if (ops->cmd)
 743                         ret = ops->cmd(lum->handle, lum->cmd,
 744                                         (unsigned long) &lum->u,
 745                                         &args, sock_info);
 746                 else
 747                         ret = -ENOSYS;
 748                 break;
 749         }
 750         case LTTNG_UST_STREAM:
 751         {
 752                 /* Receive shm_fd, wakeup_fd */
 753                 ret = ustcomm_recv_stream_from_sessiond(sock,
 754                         &lum->u.stream.len,
 755                         &args.stream.shm_fd,
 756                         &args.stream.wakeup_fd);
 757                 if (ret) {
 758                         goto end;
 759                 }
 760                 if (ops->cmd)
 761                         ret = ops->cmd(lum->handle, lum->cmd,
 762                                         (unsigned long) &lum->u,
 763                                         &args, sock_info);
 764                 else
 765                         ret = -ENOSYS;
 766                 break;
 767         }
 768         default:
 769                 if (ops->cmd)
 770                         ret = ops->cmd(lum->handle, lum->cmd,
 771                                         (unsigned long) &lum->u,
 772                                         &args, sock_info);
 773                 else
 774                         ret = -ENOSYS;
 775                 break;
 776         }
 777
 778 end:
 779         lur.handle = lum->handle;
 780         lur.cmd = lum->cmd;
 781         lur.ret_val = ret;
 782         if (ret >= 0) {
 783                 lur.ret_code = LTTNG_UST_OK;
 784         } else {
 785                 /*
 786                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 787                  * error that are not caused by the transport, except if
 788                  * we already have a more precise error message to
 789                  * report.
 790                  */
 791                 if (ret > -LTTNG_UST_ERR) {
 792                         /* Translate code to UST error. */
 793                         switch (ret) {
 794                         case -EEXIST:
 795                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 796                                 break;
 797                         case -EINVAL:
 798                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 799                                 break;
 800                         case -ENOENT:
 801                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 802                                 break;
 803                         case -EPERM:
 804                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 805                                 break;
 806                         case -ENOSYS:
 807                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 808                                 break;
 809                         default:
 810                                 lur.ret_code = -LTTNG_UST_ERR;
 811                                 break;
 812                         }
 813                 } else {
 814                         lur.ret_code = ret;
 815                 }
 816         }
 817         if (ret >= 0) {
 818                 switch (lum->cmd) {
 819                 case LTTNG_UST_TRACER_VERSION:
 820                         lur.u.version = lum->u.version;
 821                         break;
 822                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 823                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 824                         break;
 825                 }
 826         }
 827         DBG("Return value: %d", lur.ret_val);
 828         ret = send_reply(sock, &lur);
 829         if (ret < 0) {
 830                 DBG("error sending reply");
 831                 goto error;
 832         }
 833
 834         /*
 835          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 836          * after the reply.
 837          */
 838         if (lur.ret_code == LTTNG_UST_OK) {
 839                 switch (lum->cmd) {
 840                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 841                         len = ustcomm_send_unix_sock(sock,
 842                                 &args.field_list.entry,
 843                                 sizeof(args.field_list.entry));
 844                         if (len < 0) {
 845                                 ret = len;
 846                                 goto error;
 847                         }
 848                         if (len != sizeof(args.field_list.entry)) {
 849                                 ret = -EINVAL;
 850                                 goto error;
 851                         }
 852                 }
 853         }
 854
 855 error:
 856         ust_unlock();
 857
 858         /*
 859          * Performed delayed statedump operations outside of the UST
 860          * lock. We need to take the dynamic loader lock before we take
 861          * the UST lock internally within handle_pending_statedump().
 862           */
 863         handle_pending_statedump(sock_info);
 864
 865         return ret;
 866 }
 867
 868 static
 869 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 870 {
 871         int ret;
 872
 873         if (sock_info->root_handle != -1) {
 874                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 875                 if (ret) {
 876                         ERR("Error unref root handle");
 877                 }
 878                 sock_info->root_handle = -1;
 879         }
 880         sock_info->constructor_sem_posted = 0;
 881
 882         /*
 883          * wait_shm_mmap, socket and notify socket are used by listener
 884          * threads outside of the ust lock, so we cannot tear them down
 885          * ourselves, because we cannot join on these threads. Leave
 886          * responsibility of cleaning up these resources to the OS
 887          * process exit.
 888          */
 889         if (exiting)
 890                 return;
 891
 892         if (sock_info->socket != -1) {
 893                 ret = ustcomm_close_unix_sock(sock_info->socket);
 894                 if (ret) {
 895                         ERR("Error closing ust cmd socket");
 896                 }
 897                 sock_info->socket = -1;
 898         }
 899         if (sock_info->notify_socket != -1) {
 900                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 901                 if (ret) {
 902                         ERR("Error closing ust notify socket");
 903                 }
 904                 sock_info->notify_socket = -1;
 905         }
 906         if (sock_info->wait_shm_mmap) {
 907                 long page_size;
 908
 909                 page_size = sysconf(_SC_PAGE_SIZE);
 910                 if (page_size > 0) {
 911                         ret = munmap(sock_info->wait_shm_mmap, page_size);
 912                         if (ret) {
 913                                 ERR("Error unmapping wait shm");
 914                         }
 915                 }
 916                 sock_info->wait_shm_mmap = NULL;
 917         }
 918 }
 919
 920 /*
 921  * Using fork to set umask in the child process (not multi-thread safe).
 922  * We deal with the shm_open vs ftruncate race (happening when the
 923  * sessiond owns the shm and does not let everybody modify it, to ensure
 924  * safety against shm_unlink) by simply letting the mmap fail and
 925  * retrying after a few seconds.
 926  * For global shm, everybody has rw access to it until the sessiond
 927  * starts.
 928  */
 929 static
 930 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 931 {
 932         int wait_shm_fd, ret;
 933         pid_t pid;
 934
 935         /*
 936          * Try to open read-only.
 937          */
 938         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 939         if (wait_shm_fd >= 0) {
 940                 int32_t tmp_read;
 941                 ssize_t len;
 942                 size_t bytes_read = 0;
 943
 944                 /*
 945                  * Try to read the fd. If unable to do so, try opening
 946                  * it in write mode.
 947                  */
 948                 do {
 949                         len = read(wait_shm_fd,
 950                                 &((char *) &tmp_read)[bytes_read],
 951                                 sizeof(tmp_read) - bytes_read);
 952                         if (len > 0) {
 953                                 bytes_read += len;
 954                         }
 955                 } while ((len < 0 && errno == EINTR)
 956                         || (len > 0 && bytes_read < sizeof(tmp_read)));
 957                 if (bytes_read != sizeof(tmp_read)) {
 958                         ret = close(wait_shm_fd);
 959                         if (ret) {
 960                                 ERR("close wait_shm_fd");
 961                         }
 962                         goto open_write;
 963                 }
 964                 goto end;
 965         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 966                 /*
 967                  * Real-only open did not work, and it's not because the
 968                  * entry was not present. It's a failure that prohibits
 969                  * using shm.
 970                  */
 971                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 972                 goto end;
 973         }
 974
 975 open_write:
 976         /*
 977          * If the open failed because the file did not exist, or because
 978          * the file was not truncated yet, try creating it ourself.
 979          */
 980         URCU_TLS(lttng_ust_nest_count)++;
 981         pid = fork();
 982         URCU_TLS(lttng_ust_nest_count)--;
 983         if (pid > 0) {
 984                 int status;
 985
 986                 /*
 987                  * Parent: wait for child to return, in which case the
 988                  * shared memory map will have been created.
 989                  */
 990                 pid = wait(&status);
 991                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 992                         wait_shm_fd = -1;
 993                         goto end;
 994                 }
 995                 /*
 996                  * Try to open read-only again after creation.
 997                  */
 998                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 999                 if (wait_shm_fd < 0) {
1000                         /*
1001                          * Real-only open did not work. It's a failure
1002                          * that prohibits using shm.
1003                          */
1004                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1005                         goto end;
1006                 }
1007                 goto end;
1008         } else if (pid == 0) {
1009                 int create_mode;
1010
1011                 /* Child */
1012                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1013                 if (sock_info->global)
1014                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1015                 /*
1016                  * We're alone in a child process, so we can modify the
1017                  * process-wide umask.
1018                  */
1019                 umask(~create_mode);
1020                 /*
1021                  * Try creating shm (or get rw access).
1022                  * We don't do an exclusive open, because we allow other
1023                  * processes to create+ftruncate it concurrently.
1024                  */
1025                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1026                                 O_RDWR | O_CREAT, create_mode);
1027                 if (wait_shm_fd >= 0) {
1028                         ret = ftruncate(wait_shm_fd, mmap_size);
1029                         if (ret) {
1030                                 PERROR("ftruncate");
1031                                 _exit(EXIT_FAILURE);
1032                         }
1033                         _exit(EXIT_SUCCESS);
1034                 }
1035                 /*
1036                  * For local shm, we need to have rw access to accept
1037                  * opening it: this means the local sessiond will be
1038                  * able to wake us up. For global shm, we open it even
1039                  * if rw access is not granted, because the root.root
1040                  * sessiond will be able to override all rights and wake
1041                  * us up.
1042                  */
1043                 if (!sock_info->global && errno != EACCES) {
1044                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1045                         _exit(EXIT_FAILURE);
1046                 }
1047                 /*
1048                  * The shm exists, but we cannot open it RW. Report
1049                  * success.
1050                  */
1051                 _exit(EXIT_SUCCESS);
1052         } else {
1053                 return -1;
1054         }
1055 end:
1056         if (wait_shm_fd >= 0 && !sock_info->global) {
1057                 struct stat statbuf;
1058
1059                 /*
1060                  * Ensure that our user is the owner of the shm file for
1061                  * local shm. If we do not own the file, it means our
1062                  * sessiond will not have access to wake us up (there is
1063                  * probably a rogue process trying to fake our
1064                  * sessiond). Fallback to polling method in this case.
1065                  */
1066                 ret = fstat(wait_shm_fd, &statbuf);
1067                 if (ret) {
1068                         PERROR("fstat");
1069                         goto error_close;
1070                 }
1071                 if (statbuf.st_uid != getuid())
1072                         goto error_close;
1073         }
1074         return wait_shm_fd;
1075
1076 error_close:
1077         ret = close(wait_shm_fd);
1078         if (ret) {
1079                 PERROR("Error closing fd");
1080         }
1081         return -1;
1082 }
1083
1084 static
1085 char *get_map_shm(struct sock_info *sock_info)
1086 {
1087         long page_size;
1088         int wait_shm_fd, ret;
1089         char *wait_shm_mmap;
1090
1091         page_size = sysconf(_SC_PAGE_SIZE);
1092         if (page_size < 0) {
1093                 goto error;
1094         }
1095
1096         wait_shm_fd = get_wait_shm(sock_info, page_size);
1097         if (wait_shm_fd < 0) {
1098                 goto error;
1099         }
1100         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1101                   MAP_SHARED, wait_shm_fd, 0);
1102         /* close shm fd immediately after taking the mmap reference */
1103         ret = close(wait_shm_fd);
1104         if (ret) {
1105                 PERROR("Error closing fd");
1106         }
1107         if (wait_shm_mmap == MAP_FAILED) {
1108                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1109                 goto error;
1110         }
1111         return wait_shm_mmap;
1112
1113 error:
1114         return NULL;
1115 }
1116
1117 static
1118 void wait_for_sessiond(struct sock_info *sock_info)
1119 {
1120         int ret;
1121
1122         if (ust_lock()) {
1123                 goto quit;
1124         }
1125         if (wait_poll_fallback) {
1126                 goto error;
1127         }
1128         if (!sock_info->wait_shm_mmap) {
1129                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1130                 if (!sock_info->wait_shm_mmap)
1131                         goto error;
1132         }
1133         ust_unlock();
1134
1135         DBG("Waiting for %s apps sessiond", sock_info->name);
1136         /* Wait for futex wakeup */
1137         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
1138                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
1139                         FUTEX_WAIT, 0, NULL, NULL, 0);
1140                 if (ret < 0) {
1141                         if (errno == EFAULT) {
1142                                 wait_poll_fallback = 1;
1143                                 DBG(
1144 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1145 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1146 "Please upgrade your kernel "
1147 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1148 "mainline). LTTng-UST will use polling mode fallback.");
1149                                 if (ust_debug())
1150                                         PERROR("futex");
1151                         }
1152                 }
1153         }
1154         return;
1155
1156 quit:
1157         ust_unlock();
1158         return;
1159
1160 error:
1161         ust_unlock();
1162         return;
1163 }
1164
1165 /*
1166  * This thread does not allocate any resource, except within
1167  * handle_message, within mutex protection. This mutex protects against
1168  * fork and exit.
1169  * The other moment it allocates resources is at socket connection, which
1170  * is also protected by the mutex.
1171  */
1172 static
1173 void *ust_listener_thread(void *arg)
1174 {
1175         struct sock_info *sock_info = arg;
1176         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1177         long timeout;
1178
1179         /* Restart trying to connect to the session daemon */
1180 restart:
1181         if (prev_connect_failed) {
1182                 /* Wait for sessiond availability with pipe */
1183                 wait_for_sessiond(sock_info);
1184                 if (has_waited) {
1185                         has_waited = 0;
1186                         /*
1187                          * Sleep for 5 seconds before retrying after a
1188                          * sequence of failure / wait / failure. This
1189                          * deals with a killed or broken session daemon.
1190                          */
1191                         sleep(5);
1192                 }
1193                 has_waited = 1;
1194                 prev_connect_failed = 0;
1195         }
1196
1197         if (sock_info->socket != -1) {
1198                 ret = ustcomm_close_unix_sock(sock_info->socket);
1199                 if (ret) {
1200                         ERR("Error closing %s ust cmd socket",
1201                                 sock_info->name);
1202                 }
1203                 sock_info->socket = -1;
1204         }
1205         if (sock_info->notify_socket != -1) {
1206                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1207                 if (ret) {
1208                         ERR("Error closing %s ust notify socket",
1209                                 sock_info->name);
1210                 }
1211                 sock_info->notify_socket = -1;
1212         }
1213
1214         /*
1215          * Register. We need to perform both connect and sending
1216          * registration message before doing the next connect otherwise
1217          * we may reach unix socket connect queue max limits and block
1218          * on the 2nd connect while the session daemon is awaiting the
1219          * first connect registration message.
1220          */
1221         /* Connect cmd socket */
1222         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1223         if (ret < 0) {
1224                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1225                 prev_connect_failed = 1;
1226
1227                 if (ust_lock()) {
1228                         goto quit;
1229                 }
1230
1231                 /*
1232                  * If we cannot find the sessiond daemon, don't delay
1233                  * constructor execution.
1234                  */
1235                 ret = handle_register_done(sock_info);
1236                 assert(!ret);
1237                 ust_unlock();
1238                 goto restart;
1239         }
1240         sock_info->socket = ret;
1241
1242         if (ust_lock()) {
1243                 goto quit;
1244         }
1245
1246         /*
1247          * Create only one root handle per listener thread for the whole
1248          * process lifetime, so we ensure we get ID which is statically
1249          * assigned to the root handle.
1250          */
1251         if (sock_info->root_handle == -1) {
1252                 ret = lttng_abi_create_root_handle();
1253                 if (ret < 0) {
1254                         ERR("Error creating root handle");
1255                         goto quit;
1256                 }
1257                 sock_info->root_handle = ret;
1258         }
1259
1260         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1261         if (ret < 0) {
1262                 ERR("Error registering to %s ust cmd socket",
1263                         sock_info->name);
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 restart;
1273         }
1274
1275         ust_unlock();
1276
1277         /* Connect notify socket */
1278         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1279         if (ret < 0) {
1280                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1281                 prev_connect_failed = 1;
1282
1283                 if (ust_lock()) {
1284                         goto quit;
1285                 }
1286
1287                 /*
1288                  * If we cannot find the sessiond daemon, don't delay
1289                  * constructor execution.
1290                  */
1291                 ret = handle_register_done(sock_info);
1292                 assert(!ret);
1293                 ust_unlock();
1294                 goto restart;
1295         }
1296         sock_info->notify_socket = ret;
1297
1298         timeout = get_notify_sock_timeout();
1299         if (timeout >= 0) {
1300                 /*
1301                  * Give at least 10ms to sessiond to reply to
1302                  * notifications.
1303                  */
1304                 if (timeout < 10)
1305                         timeout = 10;
1306                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1307                                 timeout);
1308                 if (ret < 0) {
1309                         WARN("Error setting socket receive timeout");
1310                 }
1311                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1312                                 timeout);
1313                 if (ret < 0) {
1314                         WARN("Error setting socket send timeout");
1315                 }
1316         } else if (timeout < -1) {
1317                 WARN("Unsupported timeout value %ld", timeout);
1318         }
1319
1320         if (ust_lock()) {
1321                 goto quit;
1322         }
1323
1324         ret = register_to_sessiond(sock_info->notify_socket,
1325                         USTCTL_SOCKET_NOTIFY);
1326         if (ret < 0) {
1327                 ERR("Error registering to %s ust notify socket",
1328                         sock_info->name);
1329                 prev_connect_failed = 1;
1330                 /*
1331                  * If we cannot register to the sessiond daemon, don't
1332                  * delay constructor execution.
1333                  */
1334                 ret = handle_register_done(sock_info);
1335                 assert(!ret);
1336                 ust_unlock();
1337                 goto restart;
1338         }
1339         sock = sock_info->socket;
1340
1341         ust_unlock();
1342
1343         for (;;) {
1344                 ssize_t len;
1345                 struct ustcomm_ust_msg lum;
1346
1347                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1348                 switch (len) {
1349                 case 0: /* orderly shutdown */
1350                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1351                         if (ust_lock()) {
1352                                 goto quit;
1353                         }
1354                         /*
1355                          * Either sessiond has shutdown or refused us by closing the socket.
1356                          * In either case, we don't want to delay construction execution,
1357                          * and we need to wait before retry.
1358                          */
1359                         prev_connect_failed = 1;
1360                         /*
1361                          * If we cannot register to the sessiond daemon, don't
1362                          * delay constructor execution.
1363                          */
1364                         ret = handle_register_done(sock_info);
1365                         assert(!ret);
1366                         ust_unlock();
1367                         goto end;
1368                 case sizeof(lum):
1369                         print_cmd(lum.cmd, lum.handle);
1370                         ret = handle_message(sock_info, sock, &lum);
1371                         if (ret) {
1372                                 ERR("Error handling message for %s socket", sock_info->name);
1373                         }
1374                         continue;
1375                 default:
1376                         if (len < 0) {
1377                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1378                         } else {
1379                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1380                         }
1381                         if (len == -ECONNRESET) {
1382                                 DBG("%s remote end closed connection", sock_info->name);
1383                                 goto end;
1384                         }
1385                         goto end;
1386                 }
1387
1388         }
1389 end:
1390         if (ust_lock()) {
1391                 goto quit;
1392         }
1393         /* Cleanup socket handles before trying to reconnect */
1394         lttng_ust_objd_table_owner_cleanup(sock_info);
1395         ust_unlock();
1396         goto restart;   /* try to reconnect */
1397
1398 quit:
1399         ust_unlock();
1400
1401         pthread_mutex_lock(&ust_exit_mutex);
1402         sock_info->thread_active = 0;
1403         pthread_mutex_unlock(&ust_exit_mutex);
1404         return NULL;
1405 }
1406
1407 /*
1408  * Weak symbol to call when the ust malloc wrapper is not loaded.
1409  */
1410 __attribute__((weak))
1411 void lttng_ust_malloc_wrapper_init(void)
1412 {
1413 }
1414
1415 /*
1416  * sessiond monitoring thread: monitor presence of global and per-user
1417  * sessiond by polling the application common named pipe.
1418  */
1419 void __attribute__((constructor)) lttng_ust_init(void)
1420 {
1421         struct timespec constructor_timeout;
1422         sigset_t sig_all_blocked, orig_parent_mask;
1423         pthread_attr_t thread_attr;
1424         int timeout_mode;
1425         int ret;
1426
1427         if (uatomic_xchg(&initialized, 1) == 1)
1428                 return;
1429
1430         /*
1431          * Fixup interdependency between TLS fixup mutex (which happens
1432          * to be the dynamic linker mutex) and ust_lock, taken within
1433          * the ust lock.
1434          */
1435         lttng_fixup_urcu_bp_tls();
1436         lttng_fixup_ringbuffer_tls();
1437         lttng_fixup_vtid_tls();
1438         lttng_fixup_nest_count_tls();
1439         lttng_fixup_procname_tls();
1440         lttng_fixup_ust_mutex_nest_tls();
1441
1442         /*
1443          * We want precise control over the order in which we construct
1444          * our sub-libraries vs starting to receive commands from
1445          * sessiond (otherwise leading to errors when trying to create
1446          * sessiond before the init functions are completed).
1447          */
1448         init_usterr();
1449         init_tracepoint();
1450         lttng_ust_clock_init();
1451         lttng_ust_getcpu_init();
1452         lttng_ust_baddr_statedump_init();
1453         lttng_ring_buffer_metadata_client_init();
1454         lttng_ring_buffer_client_overwrite_init();
1455         lttng_ring_buffer_client_overwrite_rt_init();
1456         lttng_ring_buffer_client_discard_init();
1457         lttng_ring_buffer_client_discard_rt_init();
1458         lttng_perf_counter_init();
1459         lttng_context_init();
1460         /*
1461          * Invoke ust malloc wrapper init before starting other threads.
1462          */
1463         lttng_ust_malloc_wrapper_init();
1464
1465         timeout_mode = get_constructor_timeout(&constructor_timeout);
1466
1467         ret = sem_init(&constructor_wait, 0, 0);
1468         assert(!ret);
1469
1470         ret = setup_local_apps();
1471         if (ret) {
1472                 DBG("local apps setup returned %d", ret);
1473         }
1474
1475         /* A new thread created by pthread_create inherits the signal mask
1476          * from the parent. To avoid any signal being received by the
1477          * listener thread, we block all signals temporarily in the parent,
1478          * while we create the listener thread.
1479          */
1480         sigfillset(&sig_all_blocked);
1481         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1482         if (ret) {
1483                 ERR("pthread_sigmask: %s", strerror(ret));
1484         }
1485
1486         ret = pthread_attr_init(&thread_attr);
1487         if (ret) {
1488                 ERR("pthread_attr_init: %s", strerror(ret));
1489         }
1490         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1491         if (ret) {
1492                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1493         }
1494
1495         pthread_mutex_lock(&ust_exit_mutex);
1496         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1497                         ust_listener_thread, &global_apps);
1498         if (ret) {
1499                 ERR("pthread_create global: %s", strerror(ret));
1500         }
1501         global_apps.thread_active = 1;
1502         pthread_mutex_unlock(&ust_exit_mutex);
1503
1504         if (local_apps.allowed) {
1505                 pthread_mutex_lock(&ust_exit_mutex);
1506                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1507                                 ust_listener_thread, &local_apps);
1508                 if (ret) {
1509                         ERR("pthread_create local: %s", strerror(ret));
1510                 }
1511                 local_apps.thread_active = 1;
1512                 pthread_mutex_unlock(&ust_exit_mutex);
1513         } else {
1514                 handle_register_done(&local_apps);
1515         }
1516         ret = pthread_attr_destroy(&thread_attr);
1517         if (ret) {
1518                 ERR("pthread_attr_destroy: %s", strerror(ret));
1519         }
1520
1521         /* Restore original signal mask in parent */
1522         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1523         if (ret) {
1524                 ERR("pthread_sigmask: %s", strerror(ret));
1525         }
1526
1527         switch (timeout_mode) {
1528         case 1: /* timeout wait */
1529                 do {
1530                         ret = sem_timedwait(&constructor_wait,
1531                                         &constructor_timeout);
1532                 } while (ret < 0 && errno == EINTR);
1533                 if (ret < 0 && errno == ETIMEDOUT) {
1534                         ERR("Timed out waiting for lttng-sessiond");
1535                 } else {
1536                         assert(!ret);
1537                 }
1538                 break;
1539         case -1:/* wait forever */
1540                 do {
1541                         ret = sem_wait(&constructor_wait);
1542                 } while (ret < 0 && errno == EINTR);
1543                 assert(!ret);
1544                 break;
1545         case 0: /* no timeout */
1546                 break;
1547         }
1548 }
1549
1550 static
1551 void lttng_ust_cleanup(int exiting)
1552 {
1553         cleanup_sock_info(&global_apps, exiting);
1554         if (local_apps.allowed) {
1555                 cleanup_sock_info(&local_apps, exiting);
1556         }
1557         /*
1558          * The teardown in this function all affect data structures
1559          * accessed under the UST lock by the listener thread. This
1560          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1561          * that none of these threads are accessing this data at this
1562          * point.
1563          */
1564         lttng_ust_abi_exit();
1565         lttng_ust_events_exit();
1566         lttng_context_exit();
1567         lttng_perf_counter_exit();
1568         lttng_ring_buffer_client_discard_rt_exit();
1569         lttng_ring_buffer_client_discard_exit();
1570         lttng_ring_buffer_client_overwrite_rt_exit();
1571         lttng_ring_buffer_client_overwrite_exit();
1572         lttng_ring_buffer_metadata_client_exit();
1573         lttng_ust_baddr_statedump_destroy();
1574         exit_tracepoint();
1575         if (!exiting) {
1576                 /* Reinitialize values for fork */
1577                 sem_count = 2;
1578                 lttng_ust_comm_should_quit = 0;
1579                 initialized = 0;
1580         }
1581 }
1582
1583 void __attribute__((destructor)) lttng_ust_exit(void)
1584 {
1585         int ret;
1586
1587         /*
1588          * Using pthread_cancel here because:
1589          * A) we don't want to hang application teardown.
1590          * B) the thread is not allocating any resource.
1591          */
1592
1593         /*
1594          * Require the communication thread to quit. Synchronize with
1595          * mutexes to ensure it is not in a mutex critical section when
1596          * pthread_cancel is later called.
1597          */
1598         ust_lock_nocheck();
1599         lttng_ust_comm_should_quit = 1;
1600         ust_unlock();
1601
1602         pthread_mutex_lock(&ust_exit_mutex);
1603         /* cancel threads */
1604         if (global_apps.thread_active) {
1605                 ret = pthread_cancel(global_apps.ust_listener);
1606                 if (ret) {
1607                         ERR("Error cancelling global ust listener thread: %s",
1608                                 strerror(ret));
1609                 } else {
1610                         global_apps.thread_active = 0;
1611                 }
1612         }
1613         if (local_apps.thread_active) {
1614                 ret = pthread_cancel(local_apps.ust_listener);
1615                 if (ret) {
1616                         ERR("Error cancelling local ust listener thread: %s",
1617                                 strerror(ret));
1618                 } else {
1619                         local_apps.thread_active = 0;
1620                 }
1621         }
1622         pthread_mutex_unlock(&ust_exit_mutex);
1623
1624         /*
1625          * Do NOT join threads: use of sys_futex makes it impossible to
1626          * join the threads without using async-cancel, but async-cancel
1627          * is delivered by a signal, which could hit the target thread
1628          * anywhere in its code path, including while the ust_lock() is
1629          * held, causing a deadlock for the other thread. Let the OS
1630          * cleanup the threads if there are stalled in a syscall.
1631          */
1632         lttng_ust_cleanup(1);
1633 }
1634
1635 /*
1636  * We exclude the worker threads across fork and clone (except
1637  * CLONE_VM), because these system calls only keep the forking thread
1638  * running in the child.  Therefore, we don't want to call fork or clone
1639  * in the middle of an tracepoint or ust tracing state modification.
1640  * Holding this mutex protects these structures across fork and clone.
1641  */
1642 void ust_before_fork(sigset_t *save_sigset)
1643 {
1644         /*
1645          * Disable signals. This is to avoid that the child intervenes
1646          * before it is properly setup for tracing. It is safer to
1647          * disable all signals, because then we know we are not breaking
1648          * anything by restoring the original mask.
1649          */
1650         sigset_t all_sigs;
1651         int ret;
1652
1653         if (URCU_TLS(lttng_ust_nest_count))
1654                 return;
1655         /* Disable signals */
1656         sigfillset(&all_sigs);
1657         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1658         if (ret == -1) {
1659                 PERROR("sigprocmask");
1660         }
1661
1662         pthread_mutex_lock(&ust_fork_mutex);
1663
1664         ust_lock_nocheck();
1665         rcu_bp_before_fork();
1666 }
1667
1668 static void ust_after_fork_common(sigset_t *restore_sigset)
1669 {
1670         int ret;
1671
1672         DBG("process %d", getpid());
1673         ust_unlock();
1674
1675         pthread_mutex_unlock(&ust_fork_mutex);
1676
1677         /* Restore signals */
1678         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1679         if (ret == -1) {
1680                 PERROR("sigprocmask");
1681         }
1682 }
1683
1684 void ust_after_fork_parent(sigset_t *restore_sigset)
1685 {
1686         if (URCU_TLS(lttng_ust_nest_count))
1687                 return;
1688         DBG("process %d", getpid());
1689         rcu_bp_after_fork_parent();
1690         /* Release mutexes and reenable signals */
1691         ust_after_fork_common(restore_sigset);
1692 }
1693
1694 /*
1695  * After fork, in the child, we need to cleanup all the leftover state,
1696  * except the worker thread which already magically disappeared thanks
1697  * to the weird Linux fork semantics. After tyding up, we call
1698  * lttng_ust_init() again to start over as a new PID.
1699  *
1700  * This is meant for forks() that have tracing in the child between the
1701  * fork and following exec call (if there is any).
1702  */
1703 void ust_after_fork_child(sigset_t *restore_sigset)
1704 {
1705         if (URCU_TLS(lttng_ust_nest_count))
1706                 return;
1707         DBG("process %d", getpid());
1708         /* Release urcu mutexes */
1709         rcu_bp_after_fork_child();
1710         lttng_ust_cleanup(0);
1711         lttng_context_vtid_reset();
1712         /* Release mutexes and reenable signals */
1713         ust_after_fork_common(restore_sigset);
1714         lttng_ust_init();
1715 }
1716
1717 void lttng_ust_sockinfo_session_enabled(void *owner)
1718 {
1719         struct sock_info *sock_info = owner;
1720         sock_info->statedump_pending = 1;
1721 }