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: wait forever. 0: don't wait. >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: wait forever. 0: don't wait. 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                 /* Don't wait. */
 452                 return 0;
 453         }
 454         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 455         constructor_timeout->tv_nsec +=
 456                 (constructor_delay_ms % 1000UL) * 1000000UL;
 457         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 458                 constructor_timeout->tv_sec++;
 459                 constructor_timeout->tv_nsec -= 1000000000UL;
 460         }
 461         /* Timeout wait (constructor_delay_ms). */
 462         return 1;
 463 }
 464
 465 static
 466 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 467 {
 468         return ustcomm_send_reg_msg(socket,
 469                 type,
 470                 CAA_BITS_PER_LONG,
 471                 lttng_alignof(uint8_t) * CHAR_BIT,
 472                 lttng_alignof(uint16_t) * CHAR_BIT,
 473                 lttng_alignof(uint32_t) * CHAR_BIT,
 474                 lttng_alignof(uint64_t) * CHAR_BIT,
 475                 lttng_alignof(unsigned long) * CHAR_BIT);
 476 }
 477
 478 static
 479 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 480 {
 481         ssize_t len;
 482
 483         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 484         switch (len) {
 485         case sizeof(*lur):
 486                 DBG("message successfully sent");
 487                 return 0;
 488         default:
 489                 if (len == -ECONNRESET) {
 490                         DBG("remote end closed connection");
 491                         return 0;
 492                 }
 493                 if (len < 0)
 494                         return len;
 495                 DBG("incorrect message size: %zd", len);
 496                 return -EINVAL;
 497         }
 498 }
 499
 500 static
 501 int handle_register_done(struct sock_info *sock_info)
 502 {
 503         int ret;
 504
 505         if (sock_info->constructor_sem_posted)
 506                 return 0;
 507         sock_info->constructor_sem_posted = 1;
 508         if (uatomic_read(&sem_count) <= 0) {
 509                 return 0;
 510         }
 511         ret = uatomic_add_return(&sem_count, -1);
 512         if (ret == 0) {
 513                 ret = sem_post(&constructor_wait);
 514                 assert(!ret);
 515         }
 516         return 0;
 517 }
 518
 519 /*
 520  * Only execute pending statedump after the constructor semaphore has
 521  * been posted by each listener thread. This means statedump will only
 522  * be performed after the "registration done" command is received from
 523  * each session daemon the application is connected to.
 524  *
 525  * This ensures we don't run into deadlock issues with the dynamic
 526  * loader mutex, which is held while the constructor is called and
 527  * waiting on the constructor semaphore. All operations requiring this
 528  * dynamic loader lock need to be postponed using this mechanism.
 529  */
 530 static
 531 void handle_pending_statedump(struct sock_info *sock_info)
 532 {
 533         int ctor_passed = sock_info->constructor_sem_posted;
 534
 535         if (ctor_passed && sock_info->statedump_pending) {
 536                 sock_info->statedump_pending = 0;
 537                 pthread_mutex_lock(&ust_fork_mutex);
 538                 lttng_handle_pending_statedump(sock_info);
 539                 pthread_mutex_unlock(&ust_fork_mutex);
 540         }
 541 }
 542
 543 static
 544 int handle_message(struct sock_info *sock_info,
 545                 int sock, struct ustcomm_ust_msg *lum)
 546 {
 547         int ret = 0;
 548         const struct lttng_ust_objd_ops *ops;
 549         struct ustcomm_ust_reply lur;
 550         union ust_args args;
 551         ssize_t len;
 552
 553         memset(&lur, 0, sizeof(lur));
 554
 555         if (ust_lock()) {
 556                 ret = -LTTNG_UST_ERR_EXITING;
 557                 goto end;
 558         }
 559
 560         ops = objd_ops(lum->handle);
 561         if (!ops) {
 562                 ret = -ENOENT;
 563                 goto end;
 564         }
 565
 566         switch (lum->cmd) {
 567         case LTTNG_UST_REGISTER_DONE:
 568                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 569                         ret = handle_register_done(sock_info);
 570                 else
 571                         ret = -EINVAL;
 572                 break;
 573         case LTTNG_UST_RELEASE:
 574                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 575                         ret = -EPERM;
 576                 else
 577                         ret = lttng_ust_objd_unref(lum->handle, 1);
 578                 break;
 579         case LTTNG_UST_FILTER:
 580         {
 581                 /* Receive filter data */
 582                 struct lttng_ust_filter_bytecode_node *bytecode;
 583
 584                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 585                         ERR("Filter data size is too large: %u bytes",
 586                                 lum->u.filter.data_size);
 587                         ret = -EINVAL;
 588                         goto error;
 589                 }
 590
 591                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 592                         ERR("Filter reloc offset %u is not within data",
 593                                 lum->u.filter.reloc_offset);
 594                         ret = -EINVAL;
 595                         goto error;
 596                 }
 597
 598                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 599                 if (!bytecode) {
 600                         ret = -ENOMEM;
 601                         goto error;
 602                 }
 603                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 604                                 lum->u.filter.data_size);
 605                 switch (len) {
 606                 case 0: /* orderly shutdown */
 607                         ret = 0;
 608                         free(bytecode);
 609                         goto error;
 610                 default:
 611                         if (len == lum->u.filter.data_size) {
 612                                 DBG("filter data received");
 613                                 break;
 614                         } else if (len < 0) {
 615                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 616                                 if (len == -ECONNRESET) {
 617                                         ERR("%s remote end closed connection", sock_info->name);
 618                                         ret = len;
 619                                         free(bytecode);
 620                                         goto error;
 621                                 }
 622                                 ret = len;
 623                                 free(bytecode);
 624                                 goto end;
 625                         } else {
 626                                 DBG("incorrect filter data message size: %zd", len);
 627                                 ret = -EINVAL;
 628                                 free(bytecode);
 629                                 goto end;
 630                         }
 631                 }
 632                 bytecode->bc.len = lum->u.filter.data_size;
 633                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 634                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 635                 if (ops->cmd) {
 636                         ret = ops->cmd(lum->handle, lum->cmd,
 637                                         (unsigned long) bytecode,
 638                                         &args, sock_info);
 639                         if (ret) {
 640                                 free(bytecode);
 641                         }
 642                         /* don't free bytecode if everything went fine. */
 643                 } else {
 644                         ret = -ENOSYS;
 645                         free(bytecode);
 646                 }
 647                 break;
 648         }
 649         case LTTNG_UST_EXCLUSION:
 650         {
 651                 /* Receive exclusion names */
 652                 struct lttng_ust_excluder_node *node;
 653                 unsigned int count;
 654
 655                 count = lum->u.exclusion.count;
 656                 if (count == 0) {
 657                         /* There are no names to read */
 658                         ret = 0;
 659                         goto error;
 660                 }
 661                 node = zmalloc(sizeof(*node) +
 662                                 count * LTTNG_UST_SYM_NAME_LEN);
 663                 if (!node) {
 664                         ret = -ENOMEM;
 665                         goto error;
 666                 }
 667                 node->excluder.count = count;
 668                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 669                                 count * LTTNG_UST_SYM_NAME_LEN);
 670                 switch (len) {
 671                 case 0: /* orderly shutdown */
 672                         ret = 0;
 673                         free(node);
 674                         goto error;
 675                 default:
 676                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 677                                 DBG("Exclusion data received");
 678                                 break;
 679                         } else if (len < 0) {
 680                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 681                                 if (len == -ECONNRESET) {
 682                                         ERR("%s remote end closed connection", sock_info->name);
 683                                         ret = len;
 684                                         free(node);
 685                                         goto error;
 686                                 }
 687                                 ret = len;
 688                                 free(node);
 689                                 goto end;
 690                         } else {
 691                                 DBG("Incorrect exclusion data message size: %zd", len);
 692                                 ret = -EINVAL;
 693                                 free(node);
 694                                 goto end;
 695                         }
 696                 }
 697                 if (ops->cmd) {
 698                         ret = ops->cmd(lum->handle, lum->cmd,
 699                                         (unsigned long) node,
 700                                         &args, sock_info);
 701                         if (ret) {
 702                                 free(node);
 703                         }
 704                         /* Don't free exclusion data if everything went fine. */
 705                 } else {
 706                         ret = -ENOSYS;
 707                         free(node);
 708                 }
 709                 break;
 710         }
 711         case LTTNG_UST_CHANNEL:
 712         {
 713                 void *chan_data;
 714                 int wakeup_fd;
 715
 716                 len = ustcomm_recv_channel_from_sessiond(sock,
 717                                 &chan_data, lum->u.channel.len,
 718                                 &wakeup_fd);
 719                 switch (len) {
 720                 case 0: /* orderly shutdown */
 721                         ret = 0;
 722                         goto error;
 723                 default:
 724                         if (len == lum->u.channel.len) {
 725                                 DBG("channel data received");
 726                                 break;
 727                         } else if (len < 0) {
 728                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 729                                 if (len == -ECONNRESET) {
 730                                         ERR("%s remote end closed connection", sock_info->name);
 731                                         ret = len;
 732                                         goto error;
 733                                 }
 734                                 ret = len;
 735                                 goto end;
 736                         } else {
 737                                 DBG("incorrect channel data message size: %zd", len);
 738                                 ret = -EINVAL;
 739                                 goto end;
 740                         }
 741                 }
 742                 args.channel.chan_data = chan_data;
 743                 args.channel.wakeup_fd = wakeup_fd;
 744                 if (ops->cmd)
 745                         ret = ops->cmd(lum->handle, lum->cmd,
 746                                         (unsigned long) &lum->u,
 747                                         &args, sock_info);
 748                 else
 749                         ret = -ENOSYS;
 750                 break;
 751         }
 752         case LTTNG_UST_STREAM:
 753         {
 754                 /* Receive shm_fd, wakeup_fd */
 755                 ret = ustcomm_recv_stream_from_sessiond(sock,
 756                         &lum->u.stream.len,
 757                         &args.stream.shm_fd,
 758                         &args.stream.wakeup_fd);
 759                 if (ret) {
 760                         goto end;
 761                 }
 762                 if (ops->cmd)
 763                         ret = ops->cmd(lum->handle, lum->cmd,
 764                                         (unsigned long) &lum->u,
 765                                         &args, sock_info);
 766                 else
 767                         ret = -ENOSYS;
 768                 break;
 769         }
 770         default:
 771                 if (ops->cmd)
 772                         ret = ops->cmd(lum->handle, lum->cmd,
 773                                         (unsigned long) &lum->u,
 774                                         &args, sock_info);
 775                 else
 776                         ret = -ENOSYS;
 777                 break;
 778         }
 779
 780 end:
 781         lur.handle = lum->handle;
 782         lur.cmd = lum->cmd;
 783         lur.ret_val = ret;
 784         if (ret >= 0) {
 785                 lur.ret_code = LTTNG_UST_OK;
 786         } else {
 787                 /*
 788                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 789                  * error that are not caused by the transport, except if
 790                  * we already have a more precise error message to
 791                  * report.
 792                  */
 793                 if (ret > -LTTNG_UST_ERR) {
 794                         /* Translate code to UST error. */
 795                         switch (ret) {
 796                         case -EEXIST:
 797                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 798                                 break;
 799                         case -EINVAL:
 800                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 801                                 break;
 802                         case -ENOENT:
 803                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 804                                 break;
 805                         case -EPERM:
 806                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 807                                 break;
 808                         case -ENOSYS:
 809                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 810                                 break;
 811                         default:
 812                                 lur.ret_code = -LTTNG_UST_ERR;
 813                                 break;
 814                         }
 815                 } else {
 816                         lur.ret_code = ret;
 817                 }
 818         }
 819         if (ret >= 0) {
 820                 switch (lum->cmd) {
 821                 case LTTNG_UST_TRACER_VERSION:
 822                         lur.u.version = lum->u.version;
 823                         break;
 824                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 825                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 826                         break;
 827                 }
 828         }
 829         DBG("Return value: %d", lur.ret_val);
 830         ret = send_reply(sock, &lur);
 831         if (ret < 0) {
 832                 DBG("error sending reply");
 833                 goto error;
 834         }
 835
 836         /*
 837          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 838          * after the reply.
 839          */
 840         if (lur.ret_code == LTTNG_UST_OK) {
 841                 switch (lum->cmd) {
 842                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 843                         len = ustcomm_send_unix_sock(sock,
 844                                 &args.field_list.entry,
 845                                 sizeof(args.field_list.entry));
 846                         if (len < 0) {
 847                                 ret = len;
 848                                 goto error;
 849                         }
 850                         if (len != sizeof(args.field_list.entry)) {
 851                                 ret = -EINVAL;
 852                                 goto error;
 853                         }
 854                 }
 855         }
 856
 857 error:
 858         ust_unlock();
 859
 860         /*
 861          * Performed delayed statedump operations outside of the UST
 862          * lock. We need to take the dynamic loader lock before we take
 863          * the UST lock internally within handle_pending_statedump().
 864           */
 865         handle_pending_statedump(sock_info);
 866
 867         return ret;
 868 }
 869
 870 static
 871 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 872 {
 873         int ret;
 874
 875         if (sock_info->root_handle != -1) {
 876                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 877                 if (ret) {
 878                         ERR("Error unref root handle");
 879                 }
 880                 sock_info->root_handle = -1;
 881         }
 882         sock_info->constructor_sem_posted = 0;
 883
 884         /*
 885          * wait_shm_mmap, socket and notify socket are used by listener
 886          * threads outside of the ust lock, so we cannot tear them down
 887          * ourselves, because we cannot join on these threads. Leave
 888          * responsibility of cleaning up these resources to the OS
 889          * process exit.
 890          */
 891         if (exiting)
 892                 return;
 893
 894         if (sock_info->socket != -1) {
 895                 ret = ustcomm_close_unix_sock(sock_info->socket);
 896                 if (ret) {
 897                         ERR("Error closing ust cmd socket");
 898                 }
 899                 sock_info->socket = -1;
 900         }
 901         if (sock_info->notify_socket != -1) {
 902                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 903                 if (ret) {
 904                         ERR("Error closing ust notify socket");
 905                 }
 906                 sock_info->notify_socket = -1;
 907         }
 908         if (sock_info->wait_shm_mmap) {
 909                 long page_size;
 910
 911                 page_size = sysconf(_SC_PAGE_SIZE);
 912                 if (page_size > 0) {
 913                         ret = munmap(sock_info->wait_shm_mmap, page_size);
 914                         if (ret) {
 915                                 ERR("Error unmapping wait shm");
 916                         }
 917                 }
 918                 sock_info->wait_shm_mmap = NULL;
 919         }
 920 }
 921
 922 /*
 923  * Using fork to set umask in the child process (not multi-thread safe).
 924  * We deal with the shm_open vs ftruncate race (happening when the
 925  * sessiond owns the shm and does not let everybody modify it, to ensure
 926  * safety against shm_unlink) by simply letting the mmap fail and
 927  * retrying after a few seconds.
 928  * For global shm, everybody has rw access to it until the sessiond
 929  * starts.
 930  */
 931 static
 932 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 933 {
 934         int wait_shm_fd, ret;
 935         pid_t pid;
 936
 937         /*
 938          * Try to open read-only.
 939          */
 940         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 941         if (wait_shm_fd >= 0) {
 942                 int32_t tmp_read;
 943                 ssize_t len;
 944                 size_t bytes_read = 0;
 945
 946                 /*
 947                  * Try to read the fd. If unable to do so, try opening
 948                  * it in write mode.
 949                  */
 950                 do {
 951                         len = read(wait_shm_fd,
 952                                 &((char *) &tmp_read)[bytes_read],
 953                                 sizeof(tmp_read) - bytes_read);
 954                         if (len > 0) {
 955                                 bytes_read += len;
 956                         }
 957                 } while ((len < 0 && errno == EINTR)
 958                         || (len > 0 && bytes_read < sizeof(tmp_read)));
 959                 if (bytes_read != sizeof(tmp_read)) {
 960                         ret = close(wait_shm_fd);
 961                         if (ret) {
 962                                 ERR("close wait_shm_fd");
 963                         }
 964                         goto open_write;
 965                 }
 966                 goto end;
 967         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 968                 /*
 969                  * Real-only open did not work, and it's not because the
 970                  * entry was not present. It's a failure that prohibits
 971                  * using shm.
 972                  */
 973                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 974                 goto end;
 975         }
 976
 977 open_write:
 978         /*
 979          * If the open failed because the file did not exist, or because
 980          * the file was not truncated yet, try creating it ourself.
 981          */
 982         URCU_TLS(lttng_ust_nest_count)++;
 983         pid = fork();
 984         URCU_TLS(lttng_ust_nest_count)--;
 985         if (pid > 0) {
 986                 int status;
 987
 988                 /*
 989                  * Parent: wait for child to return, in which case the
 990                  * shared memory map will have been created.
 991                  */
 992                 pid = wait(&status);
 993                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 994                         wait_shm_fd = -1;
 995                         goto end;
 996                 }
 997                 /*
 998                  * Try to open read-only again after creation.
 999                  */
1000                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1001                 if (wait_shm_fd < 0) {
1002                         /*
1003                          * Real-only open did not work. It's a failure
1004                          * that prohibits using shm.
1005                          */
1006                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1007                         goto end;
1008                 }
1009                 goto end;
1010         } else if (pid == 0) {
1011                 int create_mode;
1012
1013                 /* Child */
1014                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1015                 if (sock_info->global)
1016                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1017                 /*
1018                  * We're alone in a child process, so we can modify the
1019                  * process-wide umask.
1020                  */
1021                 umask(~create_mode);
1022                 /*
1023                  * Try creating shm (or get rw access).
1024                  * We don't do an exclusive open, because we allow other
1025                  * processes to create+ftruncate it concurrently.
1026                  */
1027                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1028                                 O_RDWR | O_CREAT, create_mode);
1029                 if (wait_shm_fd >= 0) {
1030                         ret = ftruncate(wait_shm_fd, mmap_size);
1031                         if (ret) {
1032                                 PERROR("ftruncate");
1033                                 _exit(EXIT_FAILURE);
1034                         }
1035                         _exit(EXIT_SUCCESS);
1036                 }
1037                 /*
1038                  * For local shm, we need to have rw access to accept
1039                  * opening it: this means the local sessiond will be
1040                  * able to wake us up. For global shm, we open it even
1041                  * if rw access is not granted, because the root.root
1042                  * sessiond will be able to override all rights and wake
1043                  * us up.
1044                  */
1045                 if (!sock_info->global && errno != EACCES) {
1046                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1047                         _exit(EXIT_FAILURE);
1048                 }
1049                 /*
1050                  * The shm exists, but we cannot open it RW. Report
1051                  * success.
1052                  */
1053                 _exit(EXIT_SUCCESS);
1054         } else {
1055                 return -1;
1056         }
1057 end:
1058         if (wait_shm_fd >= 0 && !sock_info->global) {
1059                 struct stat statbuf;
1060
1061                 /*
1062                  * Ensure that our user is the owner of the shm file for
1063                  * local shm. If we do not own the file, it means our
1064                  * sessiond will not have access to wake us up (there is
1065                  * probably a rogue process trying to fake our
1066                  * sessiond). Fallback to polling method in this case.
1067                  */
1068                 ret = fstat(wait_shm_fd, &statbuf);
1069                 if (ret) {
1070                         PERROR("fstat");
1071                         goto error_close;
1072                 }
1073                 if (statbuf.st_uid != getuid())
1074                         goto error_close;
1075         }
1076         return wait_shm_fd;
1077
1078 error_close:
1079         ret = close(wait_shm_fd);
1080         if (ret) {
1081                 PERROR("Error closing fd");
1082         }
1083         return -1;
1084 }
1085
1086 static
1087 char *get_map_shm(struct sock_info *sock_info)
1088 {
1089         long page_size;
1090         int wait_shm_fd, ret;
1091         char *wait_shm_mmap;
1092
1093         page_size = sysconf(_SC_PAGE_SIZE);
1094         if (page_size < 0) {
1095                 goto error;
1096         }
1097
1098         wait_shm_fd = get_wait_shm(sock_info, page_size);
1099         if (wait_shm_fd < 0) {
1100                 goto error;
1101         }
1102         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1103                   MAP_SHARED, wait_shm_fd, 0);
1104         /* close shm fd immediately after taking the mmap reference */
1105         ret = close(wait_shm_fd);
1106         if (ret) {
1107                 PERROR("Error closing fd");
1108         }
1109         if (wait_shm_mmap == MAP_FAILED) {
1110                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1111                 goto error;
1112         }
1113         return wait_shm_mmap;
1114
1115 error:
1116         return NULL;
1117 }
1118
1119 static
1120 void wait_for_sessiond(struct sock_info *sock_info)
1121 {
1122         int ret;
1123
1124         if (ust_lock()) {
1125                 goto quit;
1126         }
1127         if (wait_poll_fallback) {
1128                 goto error;
1129         }
1130         if (!sock_info->wait_shm_mmap) {
1131                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1132                 if (!sock_info->wait_shm_mmap)
1133                         goto error;
1134         }
1135         ust_unlock();
1136
1137         DBG("Waiting for %s apps sessiond", sock_info->name);
1138         /* Wait for futex wakeup */
1139         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
1140                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
1141                         FUTEX_WAIT, 0, NULL, NULL, 0);
1142                 if (ret < 0) {
1143                         if (errno == EFAULT) {
1144                                 wait_poll_fallback = 1;
1145                                 DBG(
1146 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1147 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1148 "Please upgrade your kernel "
1149 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1150 "mainline). LTTng-UST will use polling mode fallback.");
1151                                 if (ust_debug())
1152                                         PERROR("futex");
1153                         }
1154                 }
1155         }
1156         return;
1157
1158 quit:
1159         ust_unlock();
1160         return;
1161
1162 error:
1163         ust_unlock();
1164         return;
1165 }
1166
1167 /*
1168  * This thread does not allocate any resource, except within
1169  * handle_message, within mutex protection. This mutex protects against
1170  * fork and exit.
1171  * The other moment it allocates resources is at socket connection, which
1172  * is also protected by the mutex.
1173  */
1174 static
1175 void *ust_listener_thread(void *arg)
1176 {
1177         struct sock_info *sock_info = arg;
1178         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1179         long timeout;
1180
1181         /* Restart trying to connect to the session daemon */
1182 restart:
1183         if (prev_connect_failed) {
1184                 /* Wait for sessiond availability with pipe */
1185                 wait_for_sessiond(sock_info);
1186                 if (has_waited) {
1187                         has_waited = 0;
1188                         /*
1189                          * Sleep for 5 seconds before retrying after a
1190                          * sequence of failure / wait / failure. This
1191                          * deals with a killed or broken session daemon.
1192                          */
1193                         sleep(5);
1194                 }
1195                 has_waited = 1;
1196                 prev_connect_failed = 0;
1197         }
1198
1199         if (sock_info->socket != -1) {
1200                 ret = ustcomm_close_unix_sock(sock_info->socket);
1201                 if (ret) {
1202                         ERR("Error closing %s ust cmd socket",
1203                                 sock_info->name);
1204                 }
1205                 sock_info->socket = -1;
1206         }
1207         if (sock_info->notify_socket != -1) {
1208                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1209                 if (ret) {
1210                         ERR("Error closing %s ust notify socket",
1211                                 sock_info->name);
1212                 }
1213                 sock_info->notify_socket = -1;
1214         }
1215
1216         /*
1217          * Register. We need to perform both connect and sending
1218          * registration message before doing the next connect otherwise
1219          * we may reach unix socket connect queue max limits and block
1220          * on the 2nd connect while the session daemon is awaiting the
1221          * first connect registration message.
1222          */
1223         /* Connect cmd socket */
1224         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1225         if (ret < 0) {
1226                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1227                 prev_connect_failed = 1;
1228
1229                 if (ust_lock()) {
1230                         goto quit;
1231                 }
1232
1233                 /*
1234                  * If we cannot find the sessiond daemon, don't delay
1235                  * constructor execution.
1236                  */
1237                 ret = handle_register_done(sock_info);
1238                 assert(!ret);
1239                 ust_unlock();
1240                 goto restart;
1241         }
1242         sock_info->socket = ret;
1243
1244         if (ust_lock()) {
1245                 goto quit;
1246         }
1247
1248         /*
1249          * Create only one root handle per listener thread for the whole
1250          * process lifetime, so we ensure we get ID which is statically
1251          * assigned to the root handle.
1252          */
1253         if (sock_info->root_handle == -1) {
1254                 ret = lttng_abi_create_root_handle();
1255                 if (ret < 0) {
1256                         ERR("Error creating root handle");
1257                         goto quit;
1258                 }
1259                 sock_info->root_handle = ret;
1260         }
1261
1262         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1263         if (ret < 0) {
1264                 ERR("Error registering to %s ust cmd socket",
1265                         sock_info->name);
1266                 prev_connect_failed = 1;
1267                 /*
1268                  * If we cannot register to the sessiond daemon, don't
1269                  * delay constructor execution.
1270                  */
1271                 ret = handle_register_done(sock_info);
1272                 assert(!ret);
1273                 ust_unlock();
1274                 goto restart;
1275         }
1276
1277         ust_unlock();
1278
1279         /* Connect notify socket */
1280         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1281         if (ret < 0) {
1282                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1283                 prev_connect_failed = 1;
1284
1285                 if (ust_lock()) {
1286                         goto quit;
1287                 }
1288
1289                 /*
1290                  * If we cannot find the sessiond daemon, don't delay
1291                  * constructor execution.
1292                  */
1293                 ret = handle_register_done(sock_info);
1294                 assert(!ret);
1295                 ust_unlock();
1296                 goto restart;
1297         }
1298         sock_info->notify_socket = ret;
1299
1300         timeout = get_notify_sock_timeout();
1301         if (timeout >= 0) {
1302                 /*
1303                  * Give at least 10ms to sessiond to reply to
1304                  * notifications.
1305                  */
1306                 if (timeout < 10)
1307                         timeout = 10;
1308                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1309                                 timeout);
1310                 if (ret < 0) {
1311                         WARN("Error setting socket receive timeout");
1312                 }
1313                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1314                                 timeout);
1315                 if (ret < 0) {
1316                         WARN("Error setting socket send timeout");
1317                 }
1318         } else if (timeout < -1) {
1319                 WARN("Unsupported timeout value %ld", timeout);
1320         }
1321
1322         if (ust_lock()) {
1323                 goto quit;
1324         }
1325
1326         ret = register_to_sessiond(sock_info->notify_socket,
1327                         USTCTL_SOCKET_NOTIFY);
1328         if (ret < 0) {
1329                 ERR("Error registering to %s ust notify socket",
1330                         sock_info->name);
1331                 prev_connect_failed = 1;
1332                 /*
1333                  * If we cannot register to the sessiond daemon, don't
1334                  * delay constructor execution.
1335                  */
1336                 ret = handle_register_done(sock_info);
1337                 assert(!ret);
1338                 ust_unlock();
1339                 goto restart;
1340         }
1341         sock = sock_info->socket;
1342
1343         ust_unlock();
1344
1345         for (;;) {
1346                 ssize_t len;
1347                 struct ustcomm_ust_msg lum;
1348
1349                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1350                 switch (len) {
1351                 case 0: /* orderly shutdown */
1352                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1353                         if (ust_lock()) {
1354                                 goto quit;
1355                         }
1356                         /*
1357                          * Either sessiond has shutdown or refused us by closing the socket.
1358                          * In either case, we don't want to delay construction execution,
1359                          * and we need to wait before retry.
1360                          */
1361                         prev_connect_failed = 1;
1362                         /*
1363                          * If we cannot register to the sessiond daemon, don't
1364                          * delay constructor execution.
1365                          */
1366                         ret = handle_register_done(sock_info);
1367                         assert(!ret);
1368                         ust_unlock();
1369                         goto end;
1370                 case sizeof(lum):
1371                         print_cmd(lum.cmd, lum.handle);
1372                         ret = handle_message(sock_info, sock, &lum);
1373                         if (ret) {
1374                                 ERR("Error handling message for %s socket", sock_info->name);
1375                         }
1376                         continue;
1377                 default:
1378                         if (len < 0) {
1379                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1380                         } else {
1381                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1382                         }
1383                         if (len == -ECONNRESET) {
1384                                 DBG("%s remote end closed connection", sock_info->name);
1385                                 goto end;
1386                         }
1387                         goto end;
1388                 }
1389
1390         }
1391 end:
1392         if (ust_lock()) {
1393                 goto quit;
1394         }
1395         /* Cleanup socket handles before trying to reconnect */
1396         lttng_ust_objd_table_owner_cleanup(sock_info);
1397         ust_unlock();
1398         goto restart;   /* try to reconnect */
1399
1400 quit:
1401         ust_unlock();
1402
1403         pthread_mutex_lock(&ust_exit_mutex);
1404         sock_info->thread_active = 0;
1405         pthread_mutex_unlock(&ust_exit_mutex);
1406         return NULL;
1407 }
1408
1409 /*
1410  * Weak symbol to call when the ust malloc wrapper is not loaded.
1411  */
1412 __attribute__((weak))
1413 void lttng_ust_malloc_wrapper_init(void)
1414 {
1415 }
1416
1417 /*
1418  * sessiond monitoring thread: monitor presence of global and per-user
1419  * sessiond by polling the application common named pipe.
1420  */
1421 void __attribute__((constructor)) lttng_ust_init(void)
1422 {
1423         struct timespec constructor_timeout;
1424         sigset_t sig_all_blocked, orig_parent_mask;
1425         pthread_attr_t thread_attr;
1426         int timeout_mode;
1427         int ret;
1428
1429         if (uatomic_xchg(&initialized, 1) == 1)
1430                 return;
1431
1432         /*
1433          * Fixup interdependency between TLS fixup mutex (which happens
1434          * to be the dynamic linker mutex) and ust_lock, taken within
1435          * the ust lock.
1436          */
1437         lttng_fixup_urcu_bp_tls();
1438         lttng_fixup_ringbuffer_tls();
1439         lttng_fixup_vtid_tls();
1440         lttng_fixup_nest_count_tls();
1441         lttng_fixup_procname_tls();
1442         lttng_fixup_ust_mutex_nest_tls();
1443
1444         /*
1445          * We want precise control over the order in which we construct
1446          * our sub-libraries vs starting to receive commands from
1447          * sessiond (otherwise leading to errors when trying to create
1448          * sessiond before the init functions are completed).
1449          */
1450         init_usterr();
1451         init_tracepoint();
1452         lttng_ust_clock_init();
1453         lttng_ust_getcpu_init();
1454         lttng_ust_baddr_statedump_init();
1455         lttng_ring_buffer_metadata_client_init();
1456         lttng_ring_buffer_client_overwrite_init();
1457         lttng_ring_buffer_client_overwrite_rt_init();
1458         lttng_ring_buffer_client_discard_init();
1459         lttng_ring_buffer_client_discard_rt_init();
1460         lttng_perf_counter_init();
1461         lttng_context_init();
1462         /*
1463          * Invoke ust malloc wrapper init before starting other threads.
1464          */
1465         lttng_ust_malloc_wrapper_init();
1466
1467         timeout_mode = get_constructor_timeout(&constructor_timeout);
1468
1469         ret = sem_init(&constructor_wait, 0, 0);
1470         assert(!ret);
1471
1472         ret = setup_local_apps();
1473         if (ret) {
1474                 DBG("local apps setup returned %d", ret);
1475         }
1476
1477         /* A new thread created by pthread_create inherits the signal mask
1478          * from the parent. To avoid any signal being received by the
1479          * listener thread, we block all signals temporarily in the parent,
1480          * while we create the listener thread.
1481          */
1482         sigfillset(&sig_all_blocked);
1483         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1484         if (ret) {
1485                 ERR("pthread_sigmask: %s", strerror(ret));
1486         }
1487
1488         ret = pthread_attr_init(&thread_attr);
1489         if (ret) {
1490                 ERR("pthread_attr_init: %s", strerror(ret));
1491         }
1492         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1493         if (ret) {
1494                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1495         }
1496
1497         pthread_mutex_lock(&ust_exit_mutex);
1498         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1499                         ust_listener_thread, &global_apps);
1500         if (ret) {
1501                 ERR("pthread_create global: %s", strerror(ret));
1502         }
1503         global_apps.thread_active = 1;
1504         pthread_mutex_unlock(&ust_exit_mutex);
1505
1506         if (local_apps.allowed) {
1507                 pthread_mutex_lock(&ust_exit_mutex);
1508                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1509                                 ust_listener_thread, &local_apps);
1510                 if (ret) {
1511                         ERR("pthread_create local: %s", strerror(ret));
1512                 }
1513                 local_apps.thread_active = 1;
1514                 pthread_mutex_unlock(&ust_exit_mutex);
1515         } else {
1516                 handle_register_done(&local_apps);
1517         }
1518         ret = pthread_attr_destroy(&thread_attr);
1519         if (ret) {
1520                 ERR("pthread_attr_destroy: %s", strerror(ret));
1521         }
1522
1523         /* Restore original signal mask in parent */
1524         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1525         if (ret) {
1526                 ERR("pthread_sigmask: %s", strerror(ret));
1527         }
1528
1529         switch (timeout_mode) {
1530         case 1: /* timeout wait */
1531                 do {
1532                         ret = sem_timedwait(&constructor_wait,
1533                                         &constructor_timeout);
1534                 } while (ret < 0 && errno == EINTR);
1535                 if (ret < 0 && errno == ETIMEDOUT) {
1536                         ERR("Timed out waiting for lttng-sessiond");
1537                 } else {
1538                         assert(!ret);
1539                 }
1540                 break;
1541         case -1:/* wait forever */
1542                 do {
1543                         ret = sem_wait(&constructor_wait);
1544                 } while (ret < 0 && errno == EINTR);
1545                 assert(!ret);
1546                 break;
1547         case 0: /* no timeout */
1548                 break;
1549         }
1550 }
1551
1552 static
1553 void lttng_ust_cleanup(int exiting)
1554 {
1555         cleanup_sock_info(&global_apps, exiting);
1556         if (local_apps.allowed) {
1557                 cleanup_sock_info(&local_apps, exiting);
1558         }
1559         /*
1560          * The teardown in this function all affect data structures
1561          * accessed under the UST lock by the listener thread. This
1562          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1563          * that none of these threads are accessing this data at this
1564          * point.
1565          */
1566         lttng_ust_abi_exit();
1567         lttng_ust_events_exit();
1568         lttng_context_exit();
1569         lttng_perf_counter_exit();
1570         lttng_ring_buffer_client_discard_rt_exit();
1571         lttng_ring_buffer_client_discard_exit();
1572         lttng_ring_buffer_client_overwrite_rt_exit();
1573         lttng_ring_buffer_client_overwrite_exit();
1574         lttng_ring_buffer_metadata_client_exit();
1575         lttng_ust_baddr_statedump_destroy();
1576         exit_tracepoint();
1577         if (!exiting) {
1578                 /* Reinitialize values for fork */
1579                 sem_count = 2;
1580                 lttng_ust_comm_should_quit = 0;
1581                 initialized = 0;
1582         }
1583 }
1584
1585 void __attribute__((destructor)) lttng_ust_exit(void)
1586 {
1587         int ret;
1588
1589         /*
1590          * Using pthread_cancel here because:
1591          * A) we don't want to hang application teardown.
1592          * B) the thread is not allocating any resource.
1593          */
1594
1595         /*
1596          * Require the communication thread to quit. Synchronize with
1597          * mutexes to ensure it is not in a mutex critical section when
1598          * pthread_cancel is later called.
1599          */
1600         ust_lock_nocheck();
1601         lttng_ust_comm_should_quit = 1;
1602         ust_unlock();
1603
1604         pthread_mutex_lock(&ust_exit_mutex);
1605         /* cancel threads */
1606         if (global_apps.thread_active) {
1607                 ret = pthread_cancel(global_apps.ust_listener);
1608                 if (ret) {
1609                         ERR("Error cancelling global ust listener thread: %s",
1610                                 strerror(ret));
1611                 } else {
1612                         global_apps.thread_active = 0;
1613                 }
1614         }
1615         if (local_apps.thread_active) {
1616                 ret = pthread_cancel(local_apps.ust_listener);
1617                 if (ret) {
1618                         ERR("Error cancelling local ust listener thread: %s",
1619                                 strerror(ret));
1620                 } else {
1621                         local_apps.thread_active = 0;
1622                 }
1623         }
1624         pthread_mutex_unlock(&ust_exit_mutex);
1625
1626         /*
1627          * Do NOT join threads: use of sys_futex makes it impossible to
1628          * join the threads without using async-cancel, but async-cancel
1629          * is delivered by a signal, which could hit the target thread
1630          * anywhere in its code path, including while the ust_lock() is
1631          * held, causing a deadlock for the other thread. Let the OS
1632          * cleanup the threads if there are stalled in a syscall.
1633          */
1634         lttng_ust_cleanup(1);
1635 }
1636
1637 /*
1638  * We exclude the worker threads across fork and clone (except
1639  * CLONE_VM), because these system calls only keep the forking thread
1640  * running in the child.  Therefore, we don't want to call fork or clone
1641  * in the middle of an tracepoint or ust tracing state modification.
1642  * Holding this mutex protects these structures across fork and clone.
1643  */
1644 void ust_before_fork(sigset_t *save_sigset)
1645 {
1646         /*
1647          * Disable signals. This is to avoid that the child intervenes
1648          * before it is properly setup for tracing. It is safer to
1649          * disable all signals, because then we know we are not breaking
1650          * anything by restoring the original mask.
1651          */
1652         sigset_t all_sigs;
1653         int ret;
1654
1655         if (URCU_TLS(lttng_ust_nest_count))
1656                 return;
1657         /* Disable signals */
1658         sigfillset(&all_sigs);
1659         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1660         if (ret == -1) {
1661                 PERROR("sigprocmask");
1662         }
1663
1664         pthread_mutex_lock(&ust_fork_mutex);
1665
1666         ust_lock_nocheck();
1667         rcu_bp_before_fork();
1668 }
1669
1670 static void ust_after_fork_common(sigset_t *restore_sigset)
1671 {
1672         int ret;
1673
1674         DBG("process %d", getpid());
1675         ust_unlock();
1676
1677         pthread_mutex_unlock(&ust_fork_mutex);
1678
1679         /* Restore signals */
1680         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1681         if (ret == -1) {
1682                 PERROR("sigprocmask");
1683         }
1684 }
1685
1686 void ust_after_fork_parent(sigset_t *restore_sigset)
1687 {
1688         if (URCU_TLS(lttng_ust_nest_count))
1689                 return;
1690         DBG("process %d", getpid());
1691         rcu_bp_after_fork_parent();
1692         /* Release mutexes and reenable signals */
1693         ust_after_fork_common(restore_sigset);
1694 }
1695
1696 /*
1697  * After fork, in the child, we need to cleanup all the leftover state,
1698  * except the worker thread which already magically disappeared thanks
1699  * to the weird Linux fork semantics. After tyding up, we call
1700  * lttng_ust_init() again to start over as a new PID.
1701  *
1702  * This is meant for forks() that have tracing in the child between the
1703  * fork and following exec call (if there is any).
1704  */
1705 void ust_after_fork_child(sigset_t *restore_sigset)
1706 {
1707         if (URCU_TLS(lttng_ust_nest_count))
1708                 return;
1709         DBG("process %d", getpid());
1710         /* Release urcu mutexes */
1711         rcu_bp_after_fork_child();
1712         lttng_ust_cleanup(0);
1713         lttng_context_vtid_reset();
1714         /* Release mutexes and reenable signals */
1715         ust_after_fork_common(restore_sigset);
1716         lttng_ust_init();
1717 }
1718
1719 void lttng_ust_sockinfo_session_enabled(void *owner)
1720 {
1721         struct sock_info *sock_info = owner;
1722         sock_info->statedump_pending = 1;
1723 }