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  * ust_fork_mutex must never nest in ust_mutex.
  72  *
  73  * ust_mutex_nest is a per-thread nesting counter, allowing the perf
  74  * counter lazy initialization called by events within the statedump,
  75  * which traces while the ust_mutex is held.
  76  *
  77  * ust_lock nests within the dynamic loader lock (within glibc) because
  78  * it is taken within the library constructor.
  79  */
  80 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
  81
  82 /* Allow nesting the ust_mutex within the same thread. */
  83 static DEFINE_URCU_TLS(int, ust_mutex_nest);
  84
  85 /*
  86  * ust_exit_mutex protects thread_active variable wrt thread exit. It
  87  * cannot be done by ust_mutex because pthread_cancel(), which takes an
  88  * internal libc lock, cannot nest within ust_mutex.
  89  *
  90  * It never nests within a ust_mutex.
  91  */
  92 static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
  93
  94 /*
  95  * ust_fork_mutex protects base address statedump tracing against forks. It
  96  * prevents the dynamic loader lock to be taken (by base address statedump
  97  * tracing) while a fork is happening, thus preventing deadlock issues with
  98  * the dynamic loader lock.
  99  */
 100 static pthread_mutex_t ust_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
 101
 102 /* Should the ust comm thread quit ? */
 103 static int lttng_ust_comm_should_quit;
 104
 105 /*
 106  * Return 0 on success, -1 if should quit.
 107  * The lock is taken in both cases.
 108  * Signal-safe.
 109  */
 110 int ust_lock(void)
 111 {
 112         sigset_t sig_all_blocked, orig_mask;
 113         int ret;
 114
 115         sigfillset(&sig_all_blocked);
 116         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 117         if (ret) {
 118                 ERR("pthread_sigmask: %s", strerror(ret));
 119         }
 120         if (!URCU_TLS(ust_mutex_nest)++)
 121                 pthread_mutex_lock(&ust_mutex);
 122         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 123         if (ret) {
 124                 ERR("pthread_sigmask: %s", strerror(ret));
 125         }
 126         if (lttng_ust_comm_should_quit) {
 127                 return -1;
 128         } else {
 129                 return 0;
 130         }
 131 }
 132
 133 /*
 134  * ust_lock_nocheck() can be used in constructors/destructors, because
 135  * they are already nested within the dynamic loader lock, and therefore
 136  * have exclusive access against execution of liblttng-ust destructor.
 137  * Signal-safe.
 138  */
 139 void ust_lock_nocheck(void)
 140 {
 141         sigset_t sig_all_blocked, orig_mask;
 142         int ret;
 143
 144         sigfillset(&sig_all_blocked);
 145         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 146         if (ret) {
 147                 ERR("pthread_sigmask: %s", strerror(ret));
 148         }
 149         if (!URCU_TLS(ust_mutex_nest)++)
 150                 pthread_mutex_lock(&ust_mutex);
 151         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 152         if (ret) {
 153                 ERR("pthread_sigmask: %s", strerror(ret));
 154         }
 155 }
 156
 157 /*
 158  * Signal-safe.
 159  */
 160 void ust_unlock(void)
 161 {
 162         sigset_t sig_all_blocked, orig_mask;
 163         int ret;
 164
 165         sigfillset(&sig_all_blocked);
 166         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 167         if (ret) {
 168                 ERR("pthread_sigmask: %s", strerror(ret));
 169         }
 170         if (!--URCU_TLS(ust_mutex_nest))
 171                 pthread_mutex_unlock(&ust_mutex);
 172         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 173         if (ret) {
 174                 ERR("pthread_sigmask: %s", strerror(ret));
 175         }
 176 }
 177
 178 /*
 179  * Wait for either of these before continuing to the main
 180  * program:
 181  * - the register_done message from sessiond daemon
 182  *   (will let the sessiond daemon enable sessions before main
 183  *   starts.)
 184  * - sessiond daemon is not reachable.
 185  * - timeout (ensuring applications are resilient to session
 186  *   daemon problems).
 187  */
 188 static sem_t constructor_wait;
 189 /*
 190  * Doing this for both the global and local sessiond.
 191  */
 192 static int sem_count = { 2 };
 193
 194 /*
 195  * Counting nesting within lttng-ust. Used to ensure that calling fork()
 196  * from liblttng-ust does not execute the pre/post fork handlers.
 197  */
 198 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
 199
 200 /*
 201  * Info about socket and associated listener thread.
 202  */
 203 struct sock_info {
 204         const char *name;
 205         pthread_t ust_listener; /* listener thread */
 206         int root_handle;
 207         int constructor_sem_posted;
 208         int allowed;
 209         int global;
 210         int thread_active;
 211
 212         char sock_path[PATH_MAX];
 213         int socket;
 214         int notify_socket;
 215
 216         char wait_shm_path[PATH_MAX];
 217         char *wait_shm_mmap;
 218         /* Keep track of lazy state dump not performed yet. */
 219         int statedump_pending;
 220 };
 221
 222 /* Socket from app (connect) to session daemon (listen) for communication */
 223 struct sock_info global_apps = {
 224         .name = "global",
 225         .global = 1,
 226
 227         .root_handle = -1,
 228         .allowed = 1,
 229         .thread_active = 0,
 230
 231         .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
 232         .socket = -1,
 233         .notify_socket = -1,
 234
 235         .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 236
 237         .statedump_pending = 0,
 238 };
 239
 240 /* TODO: allow global_apps_sock_path override */
 241
 242 struct sock_info local_apps = {
 243         .name = "local",
 244         .global = 0,
 245         .root_handle = -1,
 246         .allowed = 0,   /* Check setuid bit first */
 247         .thread_active = 0,
 248
 249         .socket = -1,
 250         .notify_socket = -1,
 251
 252         .statedump_pending = 0,
 253 };
 254
 255 static int wait_poll_fallback;
 256
 257 static const char *cmd_name_mapping[] = {
 258         [ LTTNG_UST_RELEASE ] = "Release",
 259         [ LTTNG_UST_SESSION ] = "Create Session",
 260         [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
 261
 262         [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
 263         [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
 264         [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
 265         [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
 266
 267         /* Session FD commands */
 268         [ LTTNG_UST_CHANNEL ] = "Create Channel",
 269         [ LTTNG_UST_SESSION_START ] = "Start Session",
 270         [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
 271
 272         /* Channel FD commands */
 273         [ LTTNG_UST_STREAM ] = "Create Stream",
 274         [ LTTNG_UST_EVENT ] = "Create Event",
 275
 276         /* Event and Channel FD commands */
 277         [ LTTNG_UST_CONTEXT ] = "Create Context",
 278         [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
 279
 280         /* Event, Channel and Session commands */
 281         [ LTTNG_UST_ENABLE ] = "Enable",
 282         [ LTTNG_UST_DISABLE ] = "Disable",
 283
 284         /* Tracepoint list commands */
 285         [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
 286         [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
 287
 288         /* Event FD commands */
 289         [ LTTNG_UST_FILTER ] = "Create Filter",
 290         [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
 291 };
 292
 293 static const char *str_timeout;
 294 static int got_timeout_env;
 295
 296 extern void lttng_ring_buffer_client_overwrite_init(void);
 297 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 298 extern void lttng_ring_buffer_client_discard_init(void);
 299 extern void lttng_ring_buffer_client_discard_rt_init(void);
 300 extern void lttng_ring_buffer_metadata_client_init(void);
 301 extern void lttng_ring_buffer_client_overwrite_exit(void);
 302 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 303 extern void lttng_ring_buffer_client_discard_exit(void);
 304 extern void lttng_ring_buffer_client_discard_rt_exit(void);
 305 extern void lttng_ring_buffer_metadata_client_exit(void);
 306
 307 /*
 308  * Returns the HOME directory path. Caller MUST NOT free(3) the returned
 309  * pointer.
 310  */
 311 static
 312 const char *get_lttng_home_dir(void)
 313 {
 314        const char *val;
 315
 316        val = (const char *) getenv("LTTNG_HOME");
 317        if (val != NULL) {
 318                return val;
 319        }
 320        return (const char *) getenv("HOME");
 321 }
 322
 323 /*
 324  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 325  */
 326 static
 327 void lttng_fixup_nest_count_tls(void)
 328 {
 329         asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 330 }
 331
 332 static
 333 void lttng_fixup_ust_mutex_nest_tls(void)
 334 {
 335         asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest)));
 336 }
 337
 338 /*
 339  * Fixup urcu bp TLS.
 340  */
 341 static
 342 void lttng_fixup_urcu_bp_tls(void)
 343 {
 344         rcu_read_lock();
 345         rcu_read_unlock();
 346 }
 347
 348 int lttng_get_notify_socket(void *owner)
 349 {
 350         struct sock_info *info = owner;
 351
 352         return info->notify_socket;
 353 }
 354
 355 static
 356 void print_cmd(int cmd, int handle)
 357 {
 358         const char *cmd_name = "Unknown";
 359
 360         if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
 361                         && cmd_name_mapping[cmd]) {
 362                 cmd_name = cmd_name_mapping[cmd];
 363         }
 364         DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
 365                 cmd_name, cmd,
 366                 lttng_ust_obj_get_name(handle), handle);
 367 }
 368
 369 static
 370 int setup_local_apps(void)
 371 {
 372         const char *home_dir;
 373         uid_t uid;
 374
 375         uid = getuid();
 376         /*
 377          * Disallow per-user tracing for setuid binaries.
 378          */
 379         if (uid != geteuid()) {
 380                 assert(local_apps.allowed == 0);
 381                 return 0;
 382         }
 383         home_dir = get_lttng_home_dir();
 384         if (!home_dir) {
 385                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 386                 assert(local_apps.allowed == 0);
 387                 return -ENOENT;
 388         }
 389         local_apps.allowed = 1;
 390         snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
 391                 home_dir,
 392                 LTTNG_DEFAULT_HOME_RUNDIR,
 393                 LTTNG_UST_SOCK_FILENAME);
 394         snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
 395                 LTTNG_UST_WAIT_FILENAME,
 396                 uid);
 397         return 0;
 398 }
 399
 400 /*
 401  * Get notify_sock timeout, in ms.
 402  * -1: wait forever. 0: don't wait. >0: timeout, in ms.
 403  */
 404 static
 405 long get_timeout(void)
 406 {
 407         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 408
 409         if (!got_timeout_env) {
 410                 str_timeout = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 411                 got_timeout_env = 1;
 412         }
 413         if (str_timeout)
 414                 constructor_delay_ms = strtol(str_timeout, NULL, 10);
 415         return constructor_delay_ms;
 416 }
 417
 418 static
 419 long get_notify_sock_timeout(void)
 420 {
 421         return get_timeout();
 422 }
 423
 424 /*
 425  * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
 426  */
 427 static
 428 int get_constructor_timeout(struct timespec *constructor_timeout)
 429 {
 430         long constructor_delay_ms;
 431         int ret;
 432
 433         constructor_delay_ms = get_timeout();
 434
 435         switch (constructor_delay_ms) {
 436         case -1:/* fall-through */
 437         case 0:
 438                 return constructor_delay_ms;
 439         default:
 440                 break;
 441         }
 442
 443         /*
 444          * If we are unable to find the current time, don't wait.
 445          */
 446         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 447         if (ret) {
 448                 /* Don't wait. */
 449                 return 0;
 450         }
 451         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 452         constructor_timeout->tv_nsec +=
 453                 (constructor_delay_ms % 1000UL) * 1000000UL;
 454         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 455                 constructor_timeout->tv_sec++;
 456                 constructor_timeout->tv_nsec -= 1000000000UL;
 457         }
 458         /* Timeout wait (constructor_delay_ms). */
 459         return 1;
 460 }
 461
 462 static
 463 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 464 {
 465         return ustcomm_send_reg_msg(socket,
 466                 type,
 467                 CAA_BITS_PER_LONG,
 468                 lttng_alignof(uint8_t) * CHAR_BIT,
 469                 lttng_alignof(uint16_t) * CHAR_BIT,
 470                 lttng_alignof(uint32_t) * CHAR_BIT,
 471                 lttng_alignof(uint64_t) * CHAR_BIT,
 472                 lttng_alignof(unsigned long) * CHAR_BIT);
 473 }
 474
 475 static
 476 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 477 {
 478         ssize_t len;
 479
 480         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 481         switch (len) {
 482         case sizeof(*lur):
 483                 DBG("message successfully sent");
 484                 return 0;
 485         default:
 486                 if (len == -ECONNRESET) {
 487                         DBG("remote end closed connection");
 488                         return 0;
 489                 }
 490                 if (len < 0)
 491                         return len;
 492                 DBG("incorrect message size: %zd", len);
 493                 return -EINVAL;
 494         }
 495 }
 496
 497 static
 498 int handle_register_done(struct sock_info *sock_info)
 499 {
 500         int ret;
 501
 502         if (sock_info->constructor_sem_posted)
 503                 return 0;
 504         sock_info->constructor_sem_posted = 1;
 505         if (uatomic_read(&sem_count) <= 0) {
 506                 return 0;
 507         }
 508         ret = uatomic_add_return(&sem_count, -1);
 509         if (ret == 0) {
 510                 ret = sem_post(&constructor_wait);
 511                 assert(!ret);
 512         }
 513         return 0;
 514 }
 515
 516 /*
 517  * Only execute pending statedump after the constructor semaphore has
 518  * been posted by each listener thread. This means statedump will only
 519  * be performed after the "registration done" command is received from
 520  * each session daemon the application is connected to.
 521  *
 522  * This ensures we don't run into deadlock issues with the dynamic
 523  * loader mutex, which is held while the constructor is called and
 524  * waiting on the constructor semaphore. All operations requiring this
 525  * dynamic loader lock need to be postponed using this mechanism.
 526  */
 527 static
 528 void handle_pending_statedump(struct sock_info *sock_info)
 529 {
 530         int ctor_passed = sock_info->constructor_sem_posted;
 531
 532         if (ctor_passed && sock_info->statedump_pending) {
 533                 sock_info->statedump_pending = 0;
 534                 pthread_mutex_lock(&ust_fork_mutex);
 535                 lttng_handle_pending_statedump(sock_info);
 536                 pthread_mutex_unlock(&ust_fork_mutex);
 537         }
 538 }
 539
 540 static
 541 int handle_message(struct sock_info *sock_info,
 542                 int sock, struct ustcomm_ust_msg *lum)
 543 {
 544         int ret = 0;
 545         const struct lttng_ust_objd_ops *ops;
 546         struct ustcomm_ust_reply lur;
 547         union ust_args args;
 548         ssize_t len;
 549
 550         memset(&lur, 0, sizeof(lur));
 551
 552         if (ust_lock()) {
 553                 ret = -LTTNG_UST_ERR_EXITING;
 554                 goto end;
 555         }
 556
 557         ops = objd_ops(lum->handle);
 558         if (!ops) {
 559                 ret = -ENOENT;
 560                 goto end;
 561         }
 562
 563         switch (lum->cmd) {
 564         case LTTNG_UST_REGISTER_DONE:
 565                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 566                         ret = handle_register_done(sock_info);
 567                 else
 568                         ret = -EINVAL;
 569                 break;
 570         case LTTNG_UST_RELEASE:
 571                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 572                         ret = -EPERM;
 573                 else
 574                         ret = lttng_ust_objd_unref(lum->handle, 1);
 575                 break;
 576         case LTTNG_UST_FILTER:
 577         {
 578                 /* Receive filter data */
 579                 struct lttng_ust_filter_bytecode_node *bytecode;
 580
 581                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 582                         ERR("Filter data size is too large: %u bytes",
 583                                 lum->u.filter.data_size);
 584                         ret = -EINVAL;
 585                         goto error;
 586                 }
 587
 588                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 589                         ERR("Filter reloc offset %u is not within data",
 590                                 lum->u.filter.reloc_offset);
 591                         ret = -EINVAL;
 592                         goto error;
 593                 }
 594
 595                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 596                 if (!bytecode) {
 597                         ret = -ENOMEM;
 598                         goto error;
 599                 }
 600                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 601                                 lum->u.filter.data_size);
 602                 switch (len) {
 603                 case 0: /* orderly shutdown */
 604                         ret = 0;
 605                         free(bytecode);
 606                         goto error;
 607                 default:
 608                         if (len == lum->u.filter.data_size) {
 609                                 DBG("filter data received");
 610                                 break;
 611                         } else if (len < 0) {
 612                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 613                                 if (len == -ECONNRESET) {
 614                                         ERR("%s remote end closed connection", sock_info->name);
 615                                         ret = len;
 616                                         free(bytecode);
 617                                         goto error;
 618                                 }
 619                                 ret = len;
 620                                 free(bytecode);
 621                                 goto end;
 622                         } else {
 623                                 DBG("incorrect filter data message size: %zd", len);
 624                                 ret = -EINVAL;
 625                                 free(bytecode);
 626                                 goto end;
 627                         }
 628                 }
 629                 bytecode->bc.len = lum->u.filter.data_size;
 630                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 631                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 632                 if (ops->cmd) {
 633                         ret = ops->cmd(lum->handle, lum->cmd,
 634                                         (unsigned long) bytecode,
 635                                         &args, sock_info);
 636                         if (ret) {
 637                                 free(bytecode);
 638                         }
 639                         /* don't free bytecode if everything went fine. */
 640                 } else {
 641                         ret = -ENOSYS;
 642                         free(bytecode);
 643                 }
 644                 break;
 645         }
 646         case LTTNG_UST_EXCLUSION:
 647         {
 648                 /* Receive exclusion names */
 649                 struct lttng_ust_excluder_node *node;
 650                 unsigned int count;
 651
 652                 count = lum->u.exclusion.count;
 653                 if (count == 0) {
 654                         /* There are no names to read */
 655                         ret = 0;
 656                         goto error;
 657                 }
 658                 node = zmalloc(sizeof(*node) +
 659                                 count * LTTNG_UST_SYM_NAME_LEN);
 660                 if (!node) {
 661                         ret = -ENOMEM;
 662                         goto error;
 663                 }
 664                 node->excluder.count = count;
 665                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 666                                 count * LTTNG_UST_SYM_NAME_LEN);
 667                 switch (len) {
 668                 case 0: /* orderly shutdown */
 669                         ret = 0;
 670                         free(node);
 671                         goto error;
 672                 default:
 673                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 674                                 DBG("Exclusion data received");
 675                                 break;
 676                         } else if (len < 0) {
 677                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 678                                 if (len == -ECONNRESET) {
 679                                         ERR("%s remote end closed connection", sock_info->name);
 680                                         ret = len;
 681                                         free(node);
 682                                         goto error;
 683                                 }
 684                                 ret = len;
 685                                 free(node);
 686                                 goto end;
 687                         } else {
 688                                 DBG("Incorrect exclusion data message size: %zd", len);
 689                                 ret = -EINVAL;
 690                                 free(node);
 691                                 goto end;
 692                         }
 693                 }
 694                 if (ops->cmd) {
 695                         ret = ops->cmd(lum->handle, lum->cmd,
 696                                         (unsigned long) node,
 697                                         &args, sock_info);
 698                         if (ret) {
 699                                 free(node);
 700                         }
 701                         /* Don't free exclusion data if everything went fine. */
 702                 } else {
 703                         ret = -ENOSYS;
 704                         free(node);
 705                 }
 706                 break;
 707         }
 708         case LTTNG_UST_CHANNEL:
 709         {
 710                 void *chan_data;
 711                 int wakeup_fd;
 712
 713                 len = ustcomm_recv_channel_from_sessiond(sock,
 714                                 &chan_data, lum->u.channel.len,
 715                                 &wakeup_fd);
 716                 switch (len) {
 717                 case 0: /* orderly shutdown */
 718                         ret = 0;
 719                         goto error;
 720                 default:
 721                         if (len == lum->u.channel.len) {
 722                                 DBG("channel data received");
 723                                 break;
 724                         } else if (len < 0) {
 725                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 726                                 if (len == -ECONNRESET) {
 727                                         ERR("%s remote end closed connection", sock_info->name);
 728                                         ret = len;
 729                                         goto error;
 730                                 }
 731                                 ret = len;
 732                                 goto end;
 733                         } else {
 734                                 DBG("incorrect channel data message size: %zd", len);
 735                                 ret = -EINVAL;
 736                                 goto end;
 737                         }
 738                 }
 739                 args.channel.chan_data = chan_data;
 740                 args.channel.wakeup_fd = wakeup_fd;
 741                 if (ops->cmd)
 742                         ret = ops->cmd(lum->handle, lum->cmd,
 743                                         (unsigned long) &lum->u,
 744                                         &args, sock_info);
 745                 else
 746                         ret = -ENOSYS;
 747                 break;
 748         }
 749         case LTTNG_UST_STREAM:
 750         {
 751                 /* Receive shm_fd, wakeup_fd */
 752                 ret = ustcomm_recv_stream_from_sessiond(sock,
 753                         &lum->u.stream.len,
 754                         &args.stream.shm_fd,
 755                         &args.stream.wakeup_fd);
 756                 if (ret) {
 757                         goto end;
 758                 }
 759                 if (ops->cmd)
 760                         ret = ops->cmd(lum->handle, lum->cmd,
 761                                         (unsigned long) &lum->u,
 762                                         &args, sock_info);
 763                 else
 764                         ret = -ENOSYS;
 765                 break;
 766         }
 767         default:
 768                 if (ops->cmd)
 769                         ret = ops->cmd(lum->handle, lum->cmd,
 770                                         (unsigned long) &lum->u,
 771                                         &args, sock_info);
 772                 else
 773                         ret = -ENOSYS;
 774                 break;
 775         }
 776
 777 end:
 778         lur.handle = lum->handle;
 779         lur.cmd = lum->cmd;
 780         lur.ret_val = ret;
 781         if (ret >= 0) {
 782                 lur.ret_code = LTTNG_UST_OK;
 783         } else {
 784                 /*
 785                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 786                  * error that are not caused by the transport, except if
 787                  * we already have a more precise error message to
 788                  * report.
 789                  */
 790                 if (ret > -LTTNG_UST_ERR) {
 791                         /* Translate code to UST error. */
 792                         switch (ret) {
 793                         case -EEXIST:
 794                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 795                                 break;
 796                         case -EINVAL:
 797                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 798                                 break;
 799                         case -ENOENT:
 800                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 801                                 break;
 802                         case -EPERM:
 803                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 804                                 break;
 805                         case -ENOSYS:
 806                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 807                                 break;
 808                         default:
 809                                 lur.ret_code = -LTTNG_UST_ERR;
 810                                 break;
 811                         }
 812                 } else {
 813                         lur.ret_code = ret;
 814                 }
 815         }
 816         if (ret >= 0) {
 817                 switch (lum->cmd) {
 818                 case LTTNG_UST_TRACER_VERSION:
 819                         lur.u.version = lum->u.version;
 820                         break;
 821                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 822                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 823                         break;
 824                 }
 825         }
 826         DBG("Return value: %d", lur.ret_val);
 827         ret = send_reply(sock, &lur);
 828         if (ret < 0) {
 829                 DBG("error sending reply");
 830                 goto error;
 831         }
 832
 833         /*
 834          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 835          * after the reply.
 836          */
 837         if (lur.ret_code == LTTNG_UST_OK) {
 838                 switch (lum->cmd) {
 839                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 840                         len = ustcomm_send_unix_sock(sock,
 841                                 &args.field_list.entry,
 842                                 sizeof(args.field_list.entry));
 843                         if (len < 0) {
 844                                 ret = len;
 845                                 goto error;
 846                         }
 847                         if (len != sizeof(args.field_list.entry)) {
 848                                 ret = -EINVAL;
 849                                 goto error;
 850                         }
 851                 }
 852         }
 853
 854 error:
 855         ust_unlock();
 856
 857         /*
 858          * Performed delayed statedump operations outside of the UST
 859          * lock. We need to take the dynamic loader lock before we take
 860          * the UST lock internally within handle_pending_statedump().
 861           */
 862         handle_pending_statedump(sock_info);
 863
 864         return ret;
 865 }
 866
 867 static
 868 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 869 {
 870         int ret;
 871
 872         if (sock_info->root_handle != -1) {
 873                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 874                 if (ret) {
 875                         ERR("Error unref root handle");
 876                 }
 877                 sock_info->root_handle = -1;
 878         }
 879         sock_info->constructor_sem_posted = 0;
 880
 881         /*
 882          * wait_shm_mmap, socket and notify socket are used by listener
 883          * threads outside of the ust lock, so we cannot tear them down
 884          * ourselves, because we cannot join on these threads. Leave
 885          * responsibility of cleaning up these resources to the OS
 886          * process exit.
 887          */
 888         if (exiting)
 889                 return;
 890
 891         if (sock_info->socket != -1) {
 892                 ret = ustcomm_close_unix_sock(sock_info->socket);
 893                 if (ret) {
 894                         ERR("Error closing ust cmd socket");
 895                 }
 896                 sock_info->socket = -1;
 897         }
 898         if (sock_info->notify_socket != -1) {
 899                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 900                 if (ret) {
 901                         ERR("Error closing ust notify socket");
 902                 }
 903                 sock_info->notify_socket = -1;
 904         }
 905         if (sock_info->wait_shm_mmap) {
 906                 long page_size;
 907
 908                 page_size = sysconf(_SC_PAGE_SIZE);
 909                 if (page_size > 0) {
 910                         ret = munmap(sock_info->wait_shm_mmap, page_size);
 911                         if (ret) {
 912                                 ERR("Error unmapping wait shm");
 913                         }
 914                 }
 915                 sock_info->wait_shm_mmap = NULL;
 916         }
 917 }
 918
 919 /*
 920  * Using fork to set umask in the child process (not multi-thread safe).
 921  * We deal with the shm_open vs ftruncate race (happening when the
 922  * sessiond owns the shm and does not let everybody modify it, to ensure
 923  * safety against shm_unlink) by simply letting the mmap fail and
 924  * retrying after a few seconds.
 925  * For global shm, everybody has rw access to it until the sessiond
 926  * starts.
 927  */
 928 static
 929 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 930 {
 931         int wait_shm_fd, ret;
 932         pid_t pid;
 933
 934         /*
 935          * Try to open read-only.
 936          */
 937         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 938         if (wait_shm_fd >= 0) {
 939                 int32_t tmp_read;
 940                 ssize_t len;
 941                 size_t bytes_read = 0;
 942
 943                 /*
 944                  * Try to read the fd. If unable to do so, try opening
 945                  * it in write mode.
 946                  */
 947                 do {
 948                         len = read(wait_shm_fd,
 949                                 &((char *) &tmp_read)[bytes_read],
 950                                 sizeof(tmp_read) - bytes_read);
 951                         if (len > 0) {
 952                                 bytes_read += len;
 953                         }
 954                 } while ((len < 0 && errno == EINTR)
 955                         || (len > 0 && bytes_read < sizeof(tmp_read)));
 956                 if (bytes_read != sizeof(tmp_read)) {
 957                         ret = close(wait_shm_fd);
 958                         if (ret) {
 959                                 ERR("close wait_shm_fd");
 960                         }
 961                         goto open_write;
 962                 }
 963                 goto end;
 964         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 965                 /*
 966                  * Real-only open did not work, and it's not because the
 967                  * entry was not present. It's a failure that prohibits
 968                  * using shm.
 969                  */
 970                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 971                 goto end;
 972         }
 973
 974 open_write:
 975         /*
 976          * If the open failed because the file did not exist, or because
 977          * the file was not truncated yet, try creating it ourself.
 978          */
 979         URCU_TLS(lttng_ust_nest_count)++;
 980         pid = fork();
 981         URCU_TLS(lttng_ust_nest_count)--;
 982         if (pid > 0) {
 983                 int status;
 984
 985                 /*
 986                  * Parent: wait for child to return, in which case the
 987                  * shared memory map will have been created.
 988                  */
 989                 pid = wait(&status);
 990                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 991                         wait_shm_fd = -1;
 992                         goto end;
 993                 }
 994                 /*
 995                  * Try to open read-only again after creation.
 996                  */
 997                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 998                 if (wait_shm_fd < 0) {
 999                         /*
1000                          * Real-only open did not work. It's a failure
1001                          * that prohibits using shm.
1002                          */
1003                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1004                         goto end;
1005                 }
1006                 goto end;
1007         } else if (pid == 0) {
1008                 int create_mode;
1009
1010                 /* Child */
1011                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1012                 if (sock_info->global)
1013                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1014                 /*
1015                  * We're alone in a child process, so we can modify the
1016                  * process-wide umask.
1017                  */
1018                 umask(~create_mode);
1019                 /*
1020                  * Try creating shm (or get rw access).
1021                  * We don't do an exclusive open, because we allow other
1022                  * processes to create+ftruncate it concurrently.
1023                  */
1024                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1025                                 O_RDWR | O_CREAT, create_mode);
1026                 if (wait_shm_fd >= 0) {
1027                         ret = ftruncate(wait_shm_fd, mmap_size);
1028                         if (ret) {
1029                                 PERROR("ftruncate");
1030                                 _exit(EXIT_FAILURE);
1031                         }
1032                         _exit(EXIT_SUCCESS);
1033                 }
1034                 /*
1035                  * For local shm, we need to have rw access to accept
1036                  * opening it: this means the local sessiond will be
1037                  * able to wake us up. For global shm, we open it even
1038                  * if rw access is not granted, because the root.root
1039                  * sessiond will be able to override all rights and wake
1040                  * us up.
1041                  */
1042                 if (!sock_info->global && errno != EACCES) {
1043                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1044                         _exit(EXIT_FAILURE);
1045                 }
1046                 /*
1047                  * The shm exists, but we cannot open it RW. Report
1048                  * success.
1049                  */
1050                 _exit(EXIT_SUCCESS);
1051         } else {
1052                 return -1;
1053         }
1054 end:
1055         if (wait_shm_fd >= 0 && !sock_info->global) {
1056                 struct stat statbuf;
1057
1058                 /*
1059                  * Ensure that our user is the owner of the shm file for
1060                  * local shm. If we do not own the file, it means our
1061                  * sessiond will not have access to wake us up (there is
1062                  * probably a rogue process trying to fake our
1063                  * sessiond). Fallback to polling method in this case.
1064                  */
1065                 ret = fstat(wait_shm_fd, &statbuf);
1066                 if (ret) {
1067                         PERROR("fstat");
1068                         goto error_close;
1069                 }
1070                 if (statbuf.st_uid != getuid())
1071                         goto error_close;
1072         }
1073         return wait_shm_fd;
1074
1075 error_close:
1076         ret = close(wait_shm_fd);
1077         if (ret) {
1078                 PERROR("Error closing fd");
1079         }
1080         return -1;
1081 }
1082
1083 static
1084 char *get_map_shm(struct sock_info *sock_info)
1085 {
1086         long page_size;
1087         int wait_shm_fd, ret;
1088         char *wait_shm_mmap;
1089
1090         page_size = sysconf(_SC_PAGE_SIZE);
1091         if (page_size < 0) {
1092                 goto error;
1093         }
1094
1095         wait_shm_fd = get_wait_shm(sock_info, page_size);
1096         if (wait_shm_fd < 0) {
1097                 goto error;
1098         }
1099         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1100                   MAP_SHARED, wait_shm_fd, 0);
1101         /* close shm fd immediately after taking the mmap reference */
1102         ret = close(wait_shm_fd);
1103         if (ret) {
1104                 PERROR("Error closing fd");
1105         }
1106         if (wait_shm_mmap == MAP_FAILED) {
1107                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1108                 goto error;
1109         }
1110         return wait_shm_mmap;
1111
1112 error:
1113         return NULL;
1114 }
1115
1116 static
1117 void wait_for_sessiond(struct sock_info *sock_info)
1118 {
1119         int ret;
1120
1121         if (ust_lock()) {
1122                 goto quit;
1123         }
1124         if (wait_poll_fallback) {
1125                 goto error;
1126         }
1127         if (!sock_info->wait_shm_mmap) {
1128                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1129                 if (!sock_info->wait_shm_mmap)
1130                         goto error;
1131         }
1132         ust_unlock();
1133
1134         DBG("Waiting for %s apps sessiond", sock_info->name);
1135         /* Wait for futex wakeup */
1136         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
1137                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
1138                         FUTEX_WAIT, 0, NULL, NULL, 0);
1139                 if (ret < 0) {
1140                         if (errno == EFAULT) {
1141                                 wait_poll_fallback = 1;
1142                                 DBG(
1143 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1144 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1145 "Please upgrade your kernel "
1146 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1147 "mainline). LTTng-UST will use polling mode fallback.");
1148                                 if (ust_debug())
1149                                         PERROR("futex");
1150                         }
1151                 }
1152         }
1153         return;
1154
1155 quit:
1156         ust_unlock();
1157         return;
1158
1159 error:
1160         ust_unlock();
1161         return;
1162 }
1163
1164 /*
1165  * This thread does not allocate any resource, except within
1166  * handle_message, within mutex protection. This mutex protects against
1167  * fork and exit.
1168  * The other moment it allocates resources is at socket connection, which
1169  * is also protected by the mutex.
1170  */
1171 static
1172 void *ust_listener_thread(void *arg)
1173 {
1174         struct sock_info *sock_info = arg;
1175         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1176         long timeout;
1177
1178         /* Restart trying to connect to the session daemon */
1179 restart:
1180         if (prev_connect_failed) {
1181                 /* Wait for sessiond availability with pipe */
1182                 wait_for_sessiond(sock_info);
1183                 if (has_waited) {
1184                         has_waited = 0;
1185                         /*
1186                          * Sleep for 5 seconds before retrying after a
1187                          * sequence of failure / wait / failure. This
1188                          * deals with a killed or broken session daemon.
1189                          */
1190                         sleep(5);
1191                 }
1192                 has_waited = 1;
1193                 prev_connect_failed = 0;
1194         }
1195
1196         if (sock_info->socket != -1) {
1197                 ret = ustcomm_close_unix_sock(sock_info->socket);
1198                 if (ret) {
1199                         ERR("Error closing %s ust cmd socket",
1200                                 sock_info->name);
1201                 }
1202                 sock_info->socket = -1;
1203         }
1204         if (sock_info->notify_socket != -1) {
1205                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1206                 if (ret) {
1207                         ERR("Error closing %s ust notify socket",
1208                                 sock_info->name);
1209                 }
1210                 sock_info->notify_socket = -1;
1211         }
1212
1213         /*
1214          * Register. We need to perform both connect and sending
1215          * registration message before doing the next connect otherwise
1216          * we may reach unix socket connect queue max limits and block
1217          * on the 2nd connect while the session daemon is awaiting the
1218          * first connect registration message.
1219          */
1220         /* Connect cmd socket */
1221         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1222         if (ret < 0) {
1223                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1224                 prev_connect_failed = 1;
1225
1226                 if (ust_lock()) {
1227                         goto quit;
1228                 }
1229
1230                 /*
1231                  * If we cannot find the sessiond daemon, don't delay
1232                  * constructor execution.
1233                  */
1234                 ret = handle_register_done(sock_info);
1235                 assert(!ret);
1236                 ust_unlock();
1237                 goto restart;
1238         }
1239         sock_info->socket = ret;
1240
1241         if (ust_lock()) {
1242                 goto quit;
1243         }
1244
1245         /*
1246          * Create only one root handle per listener thread for the whole
1247          * process lifetime, so we ensure we get ID which is statically
1248          * assigned to the root handle.
1249          */
1250         if (sock_info->root_handle == -1) {
1251                 ret = lttng_abi_create_root_handle();
1252                 if (ret < 0) {
1253                         ERR("Error creating root handle");
1254                         goto quit;
1255                 }
1256                 sock_info->root_handle = ret;
1257         }
1258
1259         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1260         if (ret < 0) {
1261                 ERR("Error registering to %s ust cmd socket",
1262                         sock_info->name);
1263                 prev_connect_failed = 1;
1264                 /*
1265                  * If we cannot register to the sessiond daemon, don't
1266                  * delay constructor execution.
1267                  */
1268                 ret = handle_register_done(sock_info);
1269                 assert(!ret);
1270                 ust_unlock();
1271                 goto restart;
1272         }
1273
1274         ust_unlock();
1275
1276         /* Connect notify socket */
1277         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1278         if (ret < 0) {
1279                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1280                 prev_connect_failed = 1;
1281
1282                 if (ust_lock()) {
1283                         goto quit;
1284                 }
1285
1286                 /*
1287                  * If we cannot find the sessiond daemon, don't delay
1288                  * constructor execution.
1289                  */
1290                 ret = handle_register_done(sock_info);
1291                 assert(!ret);
1292                 ust_unlock();
1293                 goto restart;
1294         }
1295         sock_info->notify_socket = ret;
1296
1297         timeout = get_notify_sock_timeout();
1298         if (timeout >= 0) {
1299                 /*
1300                  * Give at least 10ms to sessiond to reply to
1301                  * notifications.
1302                  */
1303                 if (timeout < 10)
1304                         timeout = 10;
1305                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1306                                 timeout);
1307                 if (ret < 0) {
1308                         WARN("Error setting socket receive timeout");
1309                 }
1310                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1311                                 timeout);
1312                 if (ret < 0) {
1313                         WARN("Error setting socket send timeout");
1314                 }
1315         } else if (timeout < -1) {
1316                 WARN("Unsupported timeout value %ld", timeout);
1317         }
1318
1319         if (ust_lock()) {
1320                 goto quit;
1321         }
1322
1323         ret = register_to_sessiond(sock_info->notify_socket,
1324                         USTCTL_SOCKET_NOTIFY);
1325         if (ret < 0) {
1326                 ERR("Error registering to %s ust notify socket",
1327                         sock_info->name);
1328                 prev_connect_failed = 1;
1329                 /*
1330                  * If we cannot register to the sessiond daemon, don't
1331                  * delay constructor execution.
1332                  */
1333                 ret = handle_register_done(sock_info);
1334                 assert(!ret);
1335                 ust_unlock();
1336                 goto restart;
1337         }
1338         sock = sock_info->socket;
1339
1340         ust_unlock();
1341
1342         for (;;) {
1343                 ssize_t len;
1344                 struct ustcomm_ust_msg lum;
1345
1346                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1347                 switch (len) {
1348                 case 0: /* orderly shutdown */
1349                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1350                         if (ust_lock()) {
1351                                 goto quit;
1352                         }
1353                         /*
1354                          * Either sessiond has shutdown or refused us by closing the socket.
1355                          * In either case, we don't want to delay construction execution,
1356                          * and we need to wait before retry.
1357                          */
1358                         prev_connect_failed = 1;
1359                         /*
1360                          * If we cannot register to the sessiond daemon, don't
1361                          * delay constructor execution.
1362                          */
1363                         ret = handle_register_done(sock_info);
1364                         assert(!ret);
1365                         ust_unlock();
1366                         goto end;
1367                 case sizeof(lum):
1368                         print_cmd(lum.cmd, lum.handle);
1369                         ret = handle_message(sock_info, sock, &lum);
1370                         if (ret) {
1371                                 ERR("Error handling message for %s socket", sock_info->name);
1372                         }
1373                         continue;
1374                 default:
1375                         if (len < 0) {
1376                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1377                         } else {
1378                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1379                         }
1380                         if (len == -ECONNRESET) {
1381                                 DBG("%s remote end closed connection", sock_info->name);
1382                                 goto end;
1383                         }
1384                         goto end;
1385                 }
1386
1387         }
1388 end:
1389         if (ust_lock()) {
1390                 goto quit;
1391         }
1392         /* Cleanup socket handles before trying to reconnect */
1393         lttng_ust_objd_table_owner_cleanup(sock_info);
1394         ust_unlock();
1395         goto restart;   /* try to reconnect */
1396
1397 quit:
1398         ust_unlock();
1399
1400         pthread_mutex_lock(&ust_exit_mutex);
1401         sock_info->thread_active = 0;
1402         pthread_mutex_unlock(&ust_exit_mutex);
1403         return NULL;
1404 }
1405
1406 /*
1407  * Weak symbol to call when the ust malloc wrapper is not loaded.
1408  */
1409 __attribute__((weak))
1410 void lttng_ust_malloc_wrapper_init(void)
1411 {
1412 }
1413
1414 /*
1415  * sessiond monitoring thread: monitor presence of global and per-user
1416  * sessiond by polling the application common named pipe.
1417  */
1418 void __attribute__((constructor)) lttng_ust_init(void)
1419 {
1420         struct timespec constructor_timeout;
1421         sigset_t sig_all_blocked, orig_parent_mask;
1422         pthread_attr_t thread_attr;
1423         int timeout_mode;
1424         int ret;
1425
1426         if (uatomic_xchg(&initialized, 1) == 1)
1427                 return;
1428
1429         /*
1430          * Fixup interdependency between TLS fixup mutex (which happens
1431          * to be the dynamic linker mutex) and ust_lock, taken within
1432          * the ust lock.
1433          */
1434         lttng_fixup_urcu_bp_tls();
1435         lttng_fixup_ringbuffer_tls();
1436         lttng_fixup_vtid_tls();
1437         lttng_fixup_nest_count_tls();
1438         lttng_fixup_procname_tls();
1439         lttng_fixup_ust_mutex_nest_tls();
1440
1441         /*
1442          * We want precise control over the order in which we construct
1443          * our sub-libraries vs starting to receive commands from
1444          * sessiond (otherwise leading to errors when trying to create
1445          * sessiond before the init functions are completed).
1446          */
1447         init_usterr();
1448         init_tracepoint();
1449         lttng_ust_baddr_statedump_init();
1450         lttng_ring_buffer_metadata_client_init();
1451         lttng_ring_buffer_client_overwrite_init();
1452         lttng_ring_buffer_client_overwrite_rt_init();
1453         lttng_ring_buffer_client_discard_init();
1454         lttng_ring_buffer_client_discard_rt_init();
1455         lttng_perf_counter_init();
1456         lttng_context_init();
1457         /*
1458          * Invoke ust malloc wrapper init before starting other threads.
1459          */
1460         lttng_ust_malloc_wrapper_init();
1461
1462         timeout_mode = get_constructor_timeout(&constructor_timeout);
1463
1464         ret = sem_init(&constructor_wait, 0, 0);
1465         assert(!ret);
1466
1467         ret = setup_local_apps();
1468         if (ret) {
1469                 DBG("local apps setup returned %d", ret);
1470         }
1471
1472         /* A new thread created by pthread_create inherits the signal mask
1473          * from the parent. To avoid any signal being received by the
1474          * listener thread, we block all signals temporarily in the parent,
1475          * while we create the listener thread.
1476          */
1477         sigfillset(&sig_all_blocked);
1478         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1479         if (ret) {
1480                 ERR("pthread_sigmask: %s", strerror(ret));
1481         }
1482
1483         ret = pthread_attr_init(&thread_attr);
1484         if (ret) {
1485                 ERR("pthread_attr_init: %s", strerror(ret));
1486         }
1487         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1488         if (ret) {
1489                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1490         }
1491
1492         pthread_mutex_lock(&ust_exit_mutex);
1493         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1494                         ust_listener_thread, &global_apps);
1495         if (ret) {
1496                 ERR("pthread_create global: %s", strerror(ret));
1497         }
1498         global_apps.thread_active = 1;
1499         pthread_mutex_unlock(&ust_exit_mutex);
1500
1501         if (local_apps.allowed) {
1502                 pthread_mutex_lock(&ust_exit_mutex);
1503                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1504                                 ust_listener_thread, &local_apps);
1505                 if (ret) {
1506                         ERR("pthread_create local: %s", strerror(ret));
1507                 }
1508                 local_apps.thread_active = 1;
1509                 pthread_mutex_unlock(&ust_exit_mutex);
1510         } else {
1511                 handle_register_done(&local_apps);
1512         }
1513         ret = pthread_attr_destroy(&thread_attr);
1514         if (ret) {
1515                 ERR("pthread_attr_destroy: %s", strerror(ret));
1516         }
1517
1518         /* Restore original signal mask in parent */
1519         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1520         if (ret) {
1521                 ERR("pthread_sigmask: %s", strerror(ret));
1522         }
1523
1524         switch (timeout_mode) {
1525         case 1: /* timeout wait */
1526                 do {
1527                         ret = sem_timedwait(&constructor_wait,
1528                                         &constructor_timeout);
1529                 } while (ret < 0 && errno == EINTR);
1530                 if (ret < 0 && errno == ETIMEDOUT) {
1531                         ERR("Timed out waiting for lttng-sessiond");
1532                 } else {
1533                         assert(!ret);
1534                 }
1535                 break;
1536         case -1:/* wait forever */
1537                 do {
1538                         ret = sem_wait(&constructor_wait);
1539                 } while (ret < 0 && errno == EINTR);
1540                 assert(!ret);
1541                 break;
1542         case 0: /* no timeout */
1543                 break;
1544         }
1545 }
1546
1547 static
1548 void lttng_ust_cleanup(int exiting)
1549 {
1550         cleanup_sock_info(&global_apps, exiting);
1551         if (local_apps.allowed) {
1552                 cleanup_sock_info(&local_apps, exiting);
1553         }
1554         /*
1555          * The teardown in this function all affect data structures
1556          * accessed under the UST lock by the listener thread. This
1557          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1558          * that none of these threads are accessing this data at this
1559          * point.
1560          */
1561         lttng_ust_abi_exit();
1562         lttng_ust_events_exit();
1563         lttng_context_exit();
1564         lttng_perf_counter_exit();
1565         lttng_ring_buffer_client_discard_rt_exit();
1566         lttng_ring_buffer_client_discard_exit();
1567         lttng_ring_buffer_client_overwrite_rt_exit();
1568         lttng_ring_buffer_client_overwrite_exit();
1569         lttng_ring_buffer_metadata_client_exit();
1570         lttng_ust_baddr_statedump_destroy();
1571         exit_tracepoint();
1572         if (!exiting) {
1573                 /* Reinitialize values for fork */
1574                 sem_count = 2;
1575                 lttng_ust_comm_should_quit = 0;
1576                 initialized = 0;
1577         }
1578 }
1579
1580 void __attribute__((destructor)) lttng_ust_exit(void)
1581 {
1582         int ret;
1583
1584         /*
1585          * Using pthread_cancel here because:
1586          * A) we don't want to hang application teardown.
1587          * B) the thread is not allocating any resource.
1588          */
1589
1590         /*
1591          * Require the communication thread to quit. Synchronize with
1592          * mutexes to ensure it is not in a mutex critical section when
1593          * pthread_cancel is later called.
1594          */
1595         ust_lock_nocheck();
1596         lttng_ust_comm_should_quit = 1;
1597         ust_unlock();
1598
1599         pthread_mutex_lock(&ust_exit_mutex);
1600         /* cancel threads */
1601         if (global_apps.thread_active) {
1602                 ret = pthread_cancel(global_apps.ust_listener);
1603                 if (ret) {
1604                         ERR("Error cancelling global ust listener thread: %s",
1605                                 strerror(ret));
1606                 } else {
1607                         global_apps.thread_active = 0;
1608                 }
1609         }
1610         if (local_apps.thread_active) {
1611                 ret = pthread_cancel(local_apps.ust_listener);
1612                 if (ret) {
1613                         ERR("Error cancelling local ust listener thread: %s",
1614                                 strerror(ret));
1615                 } else {
1616                         local_apps.thread_active = 0;
1617                 }
1618         }
1619         pthread_mutex_unlock(&ust_exit_mutex);
1620
1621         /*
1622          * Do NOT join threads: use of sys_futex makes it impossible to
1623          * join the threads without using async-cancel, but async-cancel
1624          * is delivered by a signal, which could hit the target thread
1625          * anywhere in its code path, including while the ust_lock() is
1626          * held, causing a deadlock for the other thread. Let the OS
1627          * cleanup the threads if there are stalled in a syscall.
1628          */
1629         lttng_ust_cleanup(1);
1630 }
1631
1632 /*
1633  * We exclude the worker threads across fork and clone (except
1634  * CLONE_VM), because these system calls only keep the forking thread
1635  * running in the child.  Therefore, we don't want to call fork or clone
1636  * in the middle of an tracepoint or ust tracing state modification.
1637  * Holding this mutex protects these structures across fork and clone.
1638  */
1639 void ust_before_fork(sigset_t *save_sigset)
1640 {
1641         /*
1642          * Disable signals. This is to avoid that the child intervenes
1643          * before it is properly setup for tracing. It is safer to
1644          * disable all signals, because then we know we are not breaking
1645          * anything by restoring the original mask.
1646          */
1647         sigset_t all_sigs;
1648         int ret;
1649
1650         if (URCU_TLS(lttng_ust_nest_count))
1651                 return;
1652         /* Disable signals */
1653         sigfillset(&all_sigs);
1654         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1655         if (ret == -1) {
1656                 PERROR("sigprocmask");
1657         }
1658
1659         pthread_mutex_lock(&ust_fork_mutex);
1660
1661         ust_lock_nocheck();
1662         rcu_bp_before_fork();
1663 }
1664
1665 static void ust_after_fork_common(sigset_t *restore_sigset)
1666 {
1667         int ret;
1668
1669         DBG("process %d", getpid());
1670         ust_unlock();
1671
1672         pthread_mutex_unlock(&ust_fork_mutex);
1673
1674         /* Restore signals */
1675         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1676         if (ret == -1) {
1677                 PERROR("sigprocmask");
1678         }
1679 }
1680
1681 void ust_after_fork_parent(sigset_t *restore_sigset)
1682 {
1683         if (URCU_TLS(lttng_ust_nest_count))
1684                 return;
1685         DBG("process %d", getpid());
1686         rcu_bp_after_fork_parent();
1687         /* Release mutexes and reenable signals */
1688         ust_after_fork_common(restore_sigset);
1689 }
1690
1691 /*
1692  * After fork, in the child, we need to cleanup all the leftover state,
1693  * except the worker thread which already magically disappeared thanks
1694  * to the weird Linux fork semantics. After tyding up, we call
1695  * lttng_ust_init() again to start over as a new PID.
1696  *
1697  * This is meant for forks() that have tracing in the child between the
1698  * fork and following exec call (if there is any).
1699  */
1700 void ust_after_fork_child(sigset_t *restore_sigset)
1701 {
1702         if (URCU_TLS(lttng_ust_nest_count))
1703                 return;
1704         DBG("process %d", getpid());
1705         /* Release urcu mutexes */
1706         rcu_bp_after_fork_child();
1707         lttng_ust_cleanup(0);
1708         lttng_context_vtid_reset();
1709         /* Release mutexes and reenable signals */
1710         ust_after_fork_common(restore_sigset);
1711         lttng_ust_init();
1712 }
1713
1714 void lttng_ust_sockinfo_session_enabled(void *owner)
1715 {
1716         struct sock_info *sock_info = owner;
1717         sock_info->statedump_pending = 1;
1718 }