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-statedump.h"
  56 #include "clock.h"
  57 #include "../libringbuffer/getcpu.h"
  58 #include "getenv.h"
  59
  60 /*
  61  * Has lttng ust comm constructor been called ?
  62  */
  63 static int initialized;
  64
  65 /*
  66  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  67  * Held when handling a command, also held by fork() to deal with
  68  * removal of threads, and by exit path.
  69  *
  70  * The UST lock is the centralized mutex across UST tracing control and
  71  * probe registration.
  72  *
  73  * ust_exit_mutex must never nest in ust_mutex.
  74  *
  75  * ust_fork_mutex must never nest in ust_mutex.
  76  *
  77  * ust_mutex_nest is a per-thread nesting counter, allowing the perf
  78  * counter lazy initialization called by events within the statedump,
  79  * which traces while the ust_mutex is held.
  80  *
  81  * ust_lock nests within the dynamic loader lock (within glibc) because
  82  * it is taken within the library constructor.
  83  */
  84 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
  85
  86 /* Allow nesting the ust_mutex within the same thread. */
  87 static DEFINE_URCU_TLS(int, ust_mutex_nest);
  88
  89 /*
  90  * ust_exit_mutex protects thread_active variable wrt thread exit. It
  91  * cannot be done by ust_mutex because pthread_cancel(), which takes an
  92  * internal libc lock, cannot nest within ust_mutex.
  93  *
  94  * It never nests within a ust_mutex.
  95  */
  96 static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
  97
  98 /*
  99  * ust_fork_mutex protects base address statedump tracing against forks. It
 100  * prevents the dynamic loader lock to be taken (by base address statedump
 101  * tracing) while a fork is happening, thus preventing deadlock issues with
 102  * the dynamic loader lock.
 103  */
 104 static pthread_mutex_t ust_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
 105
 106 /* Should the ust comm thread quit ? */
 107 static int lttng_ust_comm_should_quit;
 108
 109 /*
 110  * Return 0 on success, -1 if should quit.
 111  * The lock is taken in both cases.
 112  * Signal-safe.
 113  */
 114 int ust_lock(void)
 115 {
 116         sigset_t sig_all_blocked, orig_mask;
 117         int ret, oldstate;
 118
 119         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 120         if (ret) {
 121                 ERR("pthread_setcancelstate: %s", strerror(ret));
 122         }
 123         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 124                 ERR("pthread_setcancelstate: unexpected oldstate");
 125         }
 126         sigfillset(&sig_all_blocked);
 127         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 128         if (ret) {
 129                 ERR("pthread_sigmask: %s", strerror(ret));
 130         }
 131         if (!URCU_TLS(ust_mutex_nest)++)
 132                 pthread_mutex_lock(&ust_mutex);
 133         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 134         if (ret) {
 135                 ERR("pthread_sigmask: %s", strerror(ret));
 136         }
 137         if (lttng_ust_comm_should_quit) {
 138                 return -1;
 139         } else {
 140                 return 0;
 141         }
 142 }
 143
 144 /*
 145  * ust_lock_nocheck() can be used in constructors/destructors, because
 146  * they are already nested within the dynamic loader lock, and therefore
 147  * have exclusive access against execution of liblttng-ust destructor.
 148  * Signal-safe.
 149  */
 150 void ust_lock_nocheck(void)
 151 {
 152         sigset_t sig_all_blocked, orig_mask;
 153         int ret, oldstate;
 154
 155         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 156         if (ret) {
 157                 ERR("pthread_setcancelstate: %s", strerror(ret));
 158         }
 159         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 160                 ERR("pthread_setcancelstate: unexpected oldstate");
 161         }
 162         sigfillset(&sig_all_blocked);
 163         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 164         if (ret) {
 165                 ERR("pthread_sigmask: %s", strerror(ret));
 166         }
 167         if (!URCU_TLS(ust_mutex_nest)++)
 168                 pthread_mutex_lock(&ust_mutex);
 169         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 170         if (ret) {
 171                 ERR("pthread_sigmask: %s", strerror(ret));
 172         }
 173 }
 174
 175 /*
 176  * Signal-safe.
 177  */
 178 void ust_unlock(void)
 179 {
 180         sigset_t sig_all_blocked, orig_mask;
 181         int ret, oldstate;
 182
 183         sigfillset(&sig_all_blocked);
 184         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 185         if (ret) {
 186                 ERR("pthread_sigmask: %s", strerror(ret));
 187         }
 188         if (!--URCU_TLS(ust_mutex_nest))
 189                 pthread_mutex_unlock(&ust_mutex);
 190         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 191         if (ret) {
 192                 ERR("pthread_sigmask: %s", strerror(ret));
 193         }
 194         ret = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &oldstate);
 195         if (ret) {
 196                 ERR("pthread_setcancelstate: %s", strerror(ret));
 197         }
 198         if (oldstate != PTHREAD_CANCEL_DISABLE) {
 199                 ERR("pthread_setcancelstate: unexpected oldstate");
 200         }
 201 }
 202
 203 /*
 204  * Wait for either of these before continuing to the main
 205  * program:
 206  * - the register_done message from sessiond daemon
 207  *   (will let the sessiond daemon enable sessions before main
 208  *   starts.)
 209  * - sessiond daemon is not reachable.
 210  * - timeout (ensuring applications are resilient to session
 211  *   daemon problems).
 212  */
 213 static sem_t constructor_wait;
 214 /*
 215  * Doing this for both the global and local sessiond.
 216  */
 217 static int sem_count = { 2 };
 218
 219 /*
 220  * Counting nesting within lttng-ust. Used to ensure that calling fork()
 221  * from liblttng-ust does not execute the pre/post fork handlers.
 222  */
 223 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
 224
 225 /*
 226  * Info about socket and associated listener thread.
 227  */
 228 struct sock_info {
 229         const char *name;
 230         pthread_t ust_listener; /* listener thread */
 231         int root_handle;
 232         int constructor_sem_posted;
 233         int allowed;
 234         int global;
 235         int thread_active;
 236
 237         char sock_path[PATH_MAX];
 238         int socket;
 239         int notify_socket;
 240
 241         char wait_shm_path[PATH_MAX];
 242         char *wait_shm_mmap;
 243         /* Keep track of lazy state dump not performed yet. */
 244         int statedump_pending;
 245 };
 246
 247 /* Socket from app (connect) to session daemon (listen) for communication */
 248 struct sock_info global_apps = {
 249         .name = "global",
 250         .global = 1,
 251
 252         .root_handle = -1,
 253         .allowed = 1,
 254         .thread_active = 0,
 255
 256         .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
 257         .socket = -1,
 258         .notify_socket = -1,
 259
 260         .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 261
 262         .statedump_pending = 0,
 263 };
 264
 265 /* TODO: allow global_apps_sock_path override */
 266
 267 struct sock_info local_apps = {
 268         .name = "local",
 269         .global = 0,
 270         .root_handle = -1,
 271         .allowed = 0,   /* Check setuid bit first */
 272         .thread_active = 0,
 273
 274         .socket = -1,
 275         .notify_socket = -1,
 276
 277         .statedump_pending = 0,
 278 };
 279
 280 static int wait_poll_fallback;
 281
 282 static const char *cmd_name_mapping[] = {
 283         [ LTTNG_UST_RELEASE ] = "Release",
 284         [ LTTNG_UST_SESSION ] = "Create Session",
 285         [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
 286
 287         [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
 288         [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
 289         [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
 290         [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
 291
 292         /* Session FD commands */
 293         [ LTTNG_UST_CHANNEL ] = "Create Channel",
 294         [ LTTNG_UST_SESSION_START ] = "Start Session",
 295         [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
 296
 297         /* Channel FD commands */
 298         [ LTTNG_UST_STREAM ] = "Create Stream",
 299         [ LTTNG_UST_EVENT ] = "Create Event",
 300
 301         /* Event and Channel FD commands */
 302         [ LTTNG_UST_CONTEXT ] = "Create Context",
 303         [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
 304
 305         /* Event, Channel and Session commands */
 306         [ LTTNG_UST_ENABLE ] = "Enable",
 307         [ LTTNG_UST_DISABLE ] = "Disable",
 308
 309         /* Tracepoint list commands */
 310         [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
 311         [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
 312
 313         /* Event FD commands */
 314         [ LTTNG_UST_FILTER ] = "Create Filter",
 315         [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
 316 };
 317
 318 static const char *str_timeout;
 319 static int got_timeout_env;
 320
 321 extern void lttng_ring_buffer_client_overwrite_init(void);
 322 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 323 extern void lttng_ring_buffer_client_discard_init(void);
 324 extern void lttng_ring_buffer_client_discard_rt_init(void);
 325 extern void lttng_ring_buffer_metadata_client_init(void);
 326 extern void lttng_ring_buffer_client_overwrite_exit(void);
 327 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 328 extern void lttng_ring_buffer_client_discard_exit(void);
 329 extern void lttng_ring_buffer_client_discard_rt_exit(void);
 330 extern void lttng_ring_buffer_metadata_client_exit(void);
 331
 332 /*
 333  * Returns the HOME directory path. Caller MUST NOT free(3) the returned
 334  * pointer.
 335  */
 336 static
 337 const char *get_lttng_home_dir(void)
 338 {
 339        const char *val;
 340
 341        val = (const char *) lttng_secure_getenv("LTTNG_HOME");
 342        if (val != NULL) {
 343                return val;
 344        }
 345        return (const char *) lttng_secure_getenv("HOME");
 346 }
 347
 348 /*
 349  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 350  */
 351 static
 352 void lttng_fixup_nest_count_tls(void)
 353 {
 354         asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 355 }
 356
 357 static
 358 void lttng_fixup_ust_mutex_nest_tls(void)
 359 {
 360         asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest)));
 361 }
 362
 363 /*
 364  * Fixup urcu bp TLS.
 365  */
 366 static
 367 void lttng_fixup_urcu_bp_tls(void)
 368 {
 369         rcu_read_lock();
 370         rcu_read_unlock();
 371 }
 372
 373 int lttng_get_notify_socket(void *owner)
 374 {
 375         struct sock_info *info = owner;
 376
 377         return info->notify_socket;
 378 }
 379
 380 static
 381 void print_cmd(int cmd, int handle)
 382 {
 383         const char *cmd_name = "Unknown";
 384
 385         if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
 386                         && cmd_name_mapping[cmd]) {
 387                 cmd_name = cmd_name_mapping[cmd];
 388         }
 389         DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
 390                 cmd_name, cmd,
 391                 lttng_ust_obj_get_name(handle), handle);
 392 }
 393
 394 static
 395 int setup_local_apps(void)
 396 {
 397         const char *home_dir;
 398         uid_t uid;
 399
 400         uid = getuid();
 401         /*
 402          * Disallow per-user tracing for setuid binaries.
 403          */
 404         if (uid != geteuid()) {
 405                 assert(local_apps.allowed == 0);
 406                 return 0;
 407         }
 408         home_dir = get_lttng_home_dir();
 409         if (!home_dir) {
 410                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 411                 assert(local_apps.allowed == 0);
 412                 return -ENOENT;
 413         }
 414         local_apps.allowed = 1;
 415         snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
 416                 home_dir,
 417                 LTTNG_DEFAULT_HOME_RUNDIR,
 418                 LTTNG_UST_SOCK_FILENAME);
 419         snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
 420                 LTTNG_UST_WAIT_FILENAME,
 421                 uid);
 422         return 0;
 423 }
 424
 425 /*
 426  * Get notify_sock timeout, in ms.
 427  * -1: wait forever. 0: don't wait. >0: timeout, in ms.
 428  */
 429 static
 430 long get_timeout(void)
 431 {
 432         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 433
 434         if (!got_timeout_env) {
 435                 str_timeout = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 436                 got_timeout_env = 1;
 437         }
 438         if (str_timeout)
 439                 constructor_delay_ms = strtol(str_timeout, NULL, 10);
 440         /* All negative values are considered as "-1". */
 441         if (constructor_delay_ms < -1)
 442                 constructor_delay_ms = -1;
 443         return constructor_delay_ms;
 444 }
 445
 446 static
 447 long get_notify_sock_timeout(void)
 448 {
 449         return get_timeout();
 450 }
 451
 452 /*
 453  * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
 454  */
 455 static
 456 int get_constructor_timeout(struct timespec *constructor_timeout)
 457 {
 458         long constructor_delay_ms;
 459         int ret;
 460
 461         constructor_delay_ms = get_timeout();
 462
 463         switch (constructor_delay_ms) {
 464         case -1:/* fall-through */
 465         case 0:
 466                 return constructor_delay_ms;
 467         default:
 468                 break;
 469         }
 470
 471         /*
 472          * If we are unable to find the current time, don't wait.
 473          */
 474         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 475         if (ret) {
 476                 /* Don't wait. */
 477                 return 0;
 478         }
 479         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 480         constructor_timeout->tv_nsec +=
 481                 (constructor_delay_ms % 1000UL) * 1000000UL;
 482         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 483                 constructor_timeout->tv_sec++;
 484                 constructor_timeout->tv_nsec -= 1000000000UL;
 485         }
 486         /* Timeout wait (constructor_delay_ms). */
 487         return 1;
 488 }
 489
 490 static
 491 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 492 {
 493         return ustcomm_send_reg_msg(socket,
 494                 type,
 495                 CAA_BITS_PER_LONG,
 496                 lttng_alignof(uint8_t) * CHAR_BIT,
 497                 lttng_alignof(uint16_t) * CHAR_BIT,
 498                 lttng_alignof(uint32_t) * CHAR_BIT,
 499                 lttng_alignof(uint64_t) * CHAR_BIT,
 500                 lttng_alignof(unsigned long) * CHAR_BIT);
 501 }
 502
 503 static
 504 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 505 {
 506         ssize_t len;
 507
 508         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 509         switch (len) {
 510         case sizeof(*lur):
 511                 DBG("message successfully sent");
 512                 return 0;
 513         default:
 514                 if (len == -ECONNRESET) {
 515                         DBG("remote end closed connection");
 516                         return 0;
 517                 }
 518                 if (len < 0)
 519                         return len;
 520                 DBG("incorrect message size: %zd", len);
 521                 return -EINVAL;
 522         }
 523 }
 524
 525 static
 526 int handle_register_done(struct sock_info *sock_info)
 527 {
 528         int ret;
 529
 530         if (sock_info->constructor_sem_posted)
 531                 return 0;
 532         sock_info->constructor_sem_posted = 1;
 533         if (uatomic_read(&sem_count) <= 0) {
 534                 return 0;
 535         }
 536         ret = uatomic_add_return(&sem_count, -1);
 537         if (ret == 0) {
 538                 ret = sem_post(&constructor_wait);
 539                 assert(!ret);
 540         }
 541         return 0;
 542 }
 543
 544 /*
 545  * Only execute pending statedump after the constructor semaphore has
 546  * been posted by each listener thread. This means statedump will only
 547  * be performed after the "registration done" command is received from
 548  * each session daemon the application is connected to.
 549  *
 550  * This ensures we don't run into deadlock issues with the dynamic
 551  * loader mutex, which is held while the constructor is called and
 552  * waiting on the constructor semaphore. All operations requiring this
 553  * dynamic loader lock need to be postponed using this mechanism.
 554  */
 555 static
 556 void handle_pending_statedump(struct sock_info *sock_info)
 557 {
 558         int ctor_passed = sock_info->constructor_sem_posted;
 559
 560         if (ctor_passed && sock_info->statedump_pending) {
 561                 sock_info->statedump_pending = 0;
 562                 pthread_mutex_lock(&ust_fork_mutex);
 563                 lttng_handle_pending_statedump(sock_info);
 564                 pthread_mutex_unlock(&ust_fork_mutex);
 565         }
 566 }
 567
 568 static
 569 int handle_message(struct sock_info *sock_info,
 570                 int sock, struct ustcomm_ust_msg *lum)
 571 {
 572         int ret = 0;
 573         const struct lttng_ust_objd_ops *ops;
 574         struct ustcomm_ust_reply lur;
 575         union ust_args args;
 576         ssize_t len;
 577
 578         memset(&lur, 0, sizeof(lur));
 579
 580         if (ust_lock()) {
 581                 ret = -LTTNG_UST_ERR_EXITING;
 582                 goto error;
 583         }
 584
 585         ops = objd_ops(lum->handle);
 586         if (!ops) {
 587                 ret = -ENOENT;
 588                 goto error;
 589         }
 590
 591         switch (lum->cmd) {
 592         case LTTNG_UST_REGISTER_DONE:
 593                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 594                         ret = handle_register_done(sock_info);
 595                 else
 596                         ret = -EINVAL;
 597                 break;
 598         case LTTNG_UST_RELEASE:
 599                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 600                         ret = -EPERM;
 601                 else
 602                         ret = lttng_ust_objd_unref(lum->handle, 1);
 603                 break;
 604         case LTTNG_UST_FILTER:
 605         {
 606                 /* Receive filter data */
 607                 struct lttng_ust_filter_bytecode_node *bytecode;
 608
 609                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 610                         ERR("Filter data size is too large: %u bytes",
 611                                 lum->u.filter.data_size);
 612                         ret = -EINVAL;
 613                         goto error;
 614                 }
 615
 616                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 617                         ERR("Filter reloc offset %u is not within data",
 618                                 lum->u.filter.reloc_offset);
 619                         ret = -EINVAL;
 620                         goto error;
 621                 }
 622
 623                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 624                 if (!bytecode) {
 625                         ret = -ENOMEM;
 626                         goto error;
 627                 }
 628                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 629                                 lum->u.filter.data_size);
 630                 switch (len) {
 631                 case 0: /* orderly shutdown */
 632                         ret = 0;
 633                         free(bytecode);
 634                         goto error;
 635                 default:
 636                         if (len == lum->u.filter.data_size) {
 637                                 DBG("filter data received");
 638                                 break;
 639                         } else if (len < 0) {
 640                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 641                                 if (len == -ECONNRESET) {
 642                                         ERR("%s remote end closed connection", sock_info->name);
 643                                         ret = len;
 644                                         free(bytecode);
 645                                         goto error;
 646                                 }
 647                                 ret = len;
 648                                 free(bytecode);
 649                                 goto error;
 650                         } else {
 651                                 DBG("incorrect filter data message size: %zd", len);
 652                                 ret = -EINVAL;
 653                                 free(bytecode);
 654                                 goto error;
 655                         }
 656                 }
 657                 bytecode->bc.len = lum->u.filter.data_size;
 658                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 659                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 660                 if (ops->cmd) {
 661                         ret = ops->cmd(lum->handle, lum->cmd,
 662                                         (unsigned long) bytecode,
 663                                         &args, sock_info);
 664                         if (ret) {
 665                                 free(bytecode);
 666                         }
 667                         /* don't free bytecode if everything went fine. */
 668                 } else {
 669                         ret = -ENOSYS;
 670                         free(bytecode);
 671                 }
 672                 break;
 673         }
 674         case LTTNG_UST_EXCLUSION:
 675         {
 676                 /* Receive exclusion names */
 677                 struct lttng_ust_excluder_node *node;
 678                 unsigned int count;
 679
 680                 count = lum->u.exclusion.count;
 681                 if (count == 0) {
 682                         /* There are no names to read */
 683                         ret = 0;
 684                         goto error;
 685                 }
 686                 node = zmalloc(sizeof(*node) +
 687                                 count * LTTNG_UST_SYM_NAME_LEN);
 688                 if (!node) {
 689                         ret = -ENOMEM;
 690                         goto error;
 691                 }
 692                 node->excluder.count = count;
 693                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 694                                 count * LTTNG_UST_SYM_NAME_LEN);
 695                 switch (len) {
 696                 case 0: /* orderly shutdown */
 697                         ret = 0;
 698                         free(node);
 699                         goto error;
 700                 default:
 701                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 702                                 DBG("Exclusion data received");
 703                                 break;
 704                         } else if (len < 0) {
 705                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 706                                 if (len == -ECONNRESET) {
 707                                         ERR("%s remote end closed connection", sock_info->name);
 708                                         ret = len;
 709                                         free(node);
 710                                         goto error;
 711                                 }
 712                                 ret = len;
 713                                 free(node);
 714                                 goto error;
 715                         } else {
 716                                 DBG("Incorrect exclusion data message size: %zd", len);
 717                                 ret = -EINVAL;
 718                                 free(node);
 719                                 goto error;
 720                         }
 721                 }
 722                 if (ops->cmd) {
 723                         ret = ops->cmd(lum->handle, lum->cmd,
 724                                         (unsigned long) node,
 725                                         &args, sock_info);
 726                         if (ret) {
 727                                 free(node);
 728                         }
 729                         /* Don't free exclusion data if everything went fine. */
 730                 } else {
 731                         ret = -ENOSYS;
 732                         free(node);
 733                 }
 734                 break;
 735         }
 736         case LTTNG_UST_CHANNEL:
 737         {
 738                 void *chan_data;
 739                 int wakeup_fd;
 740
 741                 len = ustcomm_recv_channel_from_sessiond(sock,
 742                                 &chan_data, lum->u.channel.len,
 743                                 &wakeup_fd);
 744                 switch (len) {
 745                 case 0: /* orderly shutdown */
 746                         ret = 0;
 747                         goto error;
 748                 default:
 749                         if (len == lum->u.channel.len) {
 750                                 DBG("channel data received");
 751                                 break;
 752                         } else if (len < 0) {
 753                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 754                                 if (len == -ECONNRESET) {
 755                                         ERR("%s remote end closed connection", sock_info->name);
 756                                         ret = len;
 757                                         goto error;
 758                                 }
 759                                 ret = len;
 760                                 goto error;
 761                         } else {
 762                                 DBG("incorrect channel data message size: %zd", len);
 763                                 ret = -EINVAL;
 764                                 goto error;
 765                         }
 766                 }
 767                 args.channel.chan_data = chan_data;
 768                 args.channel.wakeup_fd = wakeup_fd;
 769                 if (ops->cmd)
 770                         ret = ops->cmd(lum->handle, lum->cmd,
 771                                         (unsigned long) &lum->u,
 772                                         &args, sock_info);
 773                 else
 774                         ret = -ENOSYS;
 775                 break;
 776         }
 777         case LTTNG_UST_STREAM:
 778         {
 779                 /* Receive shm_fd, wakeup_fd */
 780                 ret = ustcomm_recv_stream_from_sessiond(sock,
 781                         &lum->u.stream.len,
 782                         &args.stream.shm_fd,
 783                         &args.stream.wakeup_fd);
 784                 if (ret) {
 785                         goto error;
 786                 }
 787                 if (ops->cmd)
 788                         ret = ops->cmd(lum->handle, lum->cmd,
 789                                         (unsigned long) &lum->u,
 790                                         &args, sock_info);
 791                 else
 792                         ret = -ENOSYS;
 793                 break;
 794         }
 795         default:
 796                 if (ops->cmd)
 797                         ret = ops->cmd(lum->handle, lum->cmd,
 798                                         (unsigned long) &lum->u,
 799                                         &args, sock_info);
 800                 else
 801                         ret = -ENOSYS;
 802                 break;
 803         }
 804
 805         lur.handle = lum->handle;
 806         lur.cmd = lum->cmd;
 807         lur.ret_val = ret;
 808         if (ret >= 0) {
 809                 lur.ret_code = LTTNG_UST_OK;
 810         } else {
 811                 /*
 812                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 813                  * error that are not caused by the transport, except if
 814                  * we already have a more precise error message to
 815                  * report.
 816                  */
 817                 if (ret > -LTTNG_UST_ERR) {
 818                         /* Translate code to UST error. */
 819                         switch (ret) {
 820                         case -EEXIST:
 821                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 822                                 break;
 823                         case -EINVAL:
 824                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 825                                 break;
 826                         case -ENOENT:
 827                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 828                                 break;
 829                         case -EPERM:
 830                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 831                                 break;
 832                         case -ENOSYS:
 833                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 834                                 break;
 835                         default:
 836                                 lur.ret_code = -LTTNG_UST_ERR;
 837                                 break;
 838                         }
 839                 } else {
 840                         lur.ret_code = ret;
 841                 }
 842         }
 843         if (ret >= 0) {
 844                 switch (lum->cmd) {
 845                 case LTTNG_UST_TRACER_VERSION:
 846                         lur.u.version = lum->u.version;
 847                         break;
 848                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 849                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 850                         break;
 851                 }
 852         }
 853         DBG("Return value: %d", lur.ret_val);
 854         ret = send_reply(sock, &lur);
 855         if (ret < 0) {
 856                 DBG("error sending reply");
 857                 goto error;
 858         }
 859
 860         /*
 861          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 862          * after the reply.
 863          */
 864         if (lur.ret_code == LTTNG_UST_OK) {
 865                 switch (lum->cmd) {
 866                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 867                         len = ustcomm_send_unix_sock(sock,
 868                                 &args.field_list.entry,
 869                                 sizeof(args.field_list.entry));
 870                         if (len < 0) {
 871                                 ret = len;
 872                                 goto error;
 873                         }
 874                         if (len != sizeof(args.field_list.entry)) {
 875                                 ret = -EINVAL;
 876                                 goto error;
 877                         }
 878                 }
 879         }
 880
 881 error:
 882         ust_unlock();
 883
 884         /*
 885          * Performed delayed statedump operations outside of the UST
 886          * lock. We need to take the dynamic loader lock before we take
 887          * the UST lock internally within handle_pending_statedump().
 888           */
 889         handle_pending_statedump(sock_info);
 890
 891         return ret;
 892 }
 893
 894 static
 895 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 896 {
 897         int ret;
 898
 899         if (sock_info->root_handle != -1) {
 900                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 901                 if (ret) {
 902                         ERR("Error unref root handle");
 903                 }
 904                 sock_info->root_handle = -1;
 905         }
 906         sock_info->constructor_sem_posted = 0;
 907
 908         /*
 909          * wait_shm_mmap, socket and notify socket are used by listener
 910          * threads outside of the ust lock, so we cannot tear them down
 911          * ourselves, because we cannot join on these threads. Leave
 912          * responsibility of cleaning up these resources to the OS
 913          * process exit.
 914          */
 915         if (exiting)
 916                 return;
 917
 918         if (sock_info->socket != -1) {
 919                 ret = ustcomm_close_unix_sock(sock_info->socket);
 920                 if (ret) {
 921                         ERR("Error closing ust cmd socket");
 922                 }
 923                 sock_info->socket = -1;
 924         }
 925         if (sock_info->notify_socket != -1) {
 926                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 927                 if (ret) {
 928                         ERR("Error closing ust notify socket");
 929                 }
 930                 sock_info->notify_socket = -1;
 931         }
 932         if (sock_info->wait_shm_mmap) {
 933                 long page_size;
 934
 935                 page_size = sysconf(_SC_PAGE_SIZE);
 936                 if (page_size <= 0) {
 937                         if (!page_size) {
 938                                 errno = EINVAL;
 939                         }
 940                         PERROR("Error in sysconf(_SC_PAGE_SIZE)");
 941                 } else {
 942                         ret = munmap(sock_info->wait_shm_mmap, page_size);
 943                         if (ret) {
 944                                 ERR("Error unmapping wait shm");
 945                         }
 946                 }
 947                 sock_info->wait_shm_mmap = NULL;
 948         }
 949 }
 950
 951 /*
 952  * Using fork to set umask in the child process (not multi-thread safe).
 953  * We deal with the shm_open vs ftruncate race (happening when the
 954  * sessiond owns the shm and does not let everybody modify it, to ensure
 955  * safety against shm_unlink) by simply letting the mmap fail and
 956  * retrying after a few seconds.
 957  * For global shm, everybody has rw access to it until the sessiond
 958  * starts.
 959  */
 960 static
 961 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 962 {
 963         int wait_shm_fd, ret;
 964         pid_t pid;
 965
 966         /*
 967          * Try to open read-only.
 968          */
 969         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 970         if (wait_shm_fd >= 0) {
 971                 int32_t tmp_read;
 972                 ssize_t len;
 973                 size_t bytes_read = 0;
 974
 975                 /*
 976                  * Try to read the fd. If unable to do so, try opening
 977                  * it in write mode.
 978                  */
 979                 do {
 980                         len = read(wait_shm_fd,
 981                                 &((char *) &tmp_read)[bytes_read],
 982                                 sizeof(tmp_read) - bytes_read);
 983                         if (len > 0) {
 984                                 bytes_read += len;
 985                         }
 986                 } while ((len < 0 && errno == EINTR)
 987                         || (len > 0 && bytes_read < sizeof(tmp_read)));
 988                 if (bytes_read != sizeof(tmp_read)) {
 989                         ret = close(wait_shm_fd);
 990                         if (ret) {
 991                                 ERR("close wait_shm_fd");
 992                         }
 993                         goto open_write;
 994                 }
 995                 goto end;
 996         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 997                 /*
 998                  * Real-only open did not work, and it's not because the
 999                  * entry was not present. It's a failure that prohibits
1000                  * using shm.
1001                  */
1002                 ERR("Error opening shm %s", sock_info->wait_shm_path);
1003                 goto end;
1004         }
1005
1006 open_write:
1007         /*
1008          * If the open failed because the file did not exist, or because
1009          * the file was not truncated yet, try creating it ourself.
1010          */
1011         URCU_TLS(lttng_ust_nest_count)++;
1012         pid = fork();
1013         URCU_TLS(lttng_ust_nest_count)--;
1014         if (pid > 0) {
1015                 int status;
1016
1017                 /*
1018                  * Parent: wait for child to return, in which case the
1019                  * shared memory map will have been created.
1020                  */
1021                 pid = wait(&status);
1022                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
1023                         wait_shm_fd = -1;
1024                         goto end;
1025                 }
1026                 /*
1027                  * Try to open read-only again after creation.
1028                  */
1029                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1030                 if (wait_shm_fd < 0) {
1031                         /*
1032                          * Real-only open did not work. It's a failure
1033                          * that prohibits using shm.
1034                          */
1035                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1036                         goto end;
1037                 }
1038                 goto end;
1039         } else if (pid == 0) {
1040                 int create_mode;
1041
1042                 /* Child */
1043                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1044                 if (sock_info->global)
1045                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1046                 /*
1047                  * We're alone in a child process, so we can modify the
1048                  * process-wide umask.
1049                  */
1050                 umask(~create_mode);
1051                 /*
1052                  * Try creating shm (or get rw access).
1053                  * We don't do an exclusive open, because we allow other
1054                  * processes to create+ftruncate it concurrently.
1055                  */
1056                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1057                                 O_RDWR | O_CREAT, create_mode);
1058                 if (wait_shm_fd >= 0) {
1059                         ret = ftruncate(wait_shm_fd, mmap_size);
1060                         if (ret) {
1061                                 PERROR("ftruncate");
1062                                 _exit(EXIT_FAILURE);
1063                         }
1064                         _exit(EXIT_SUCCESS);
1065                 }
1066                 /*
1067                  * For local shm, we need to have rw access to accept
1068                  * opening it: this means the local sessiond will be
1069                  * able to wake us up. For global shm, we open it even
1070                  * if rw access is not granted, because the root.root
1071                  * sessiond will be able to override all rights and wake
1072                  * us up.
1073                  */
1074                 if (!sock_info->global && errno != EACCES) {
1075                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1076                         _exit(EXIT_FAILURE);
1077                 }
1078                 /*
1079                  * The shm exists, but we cannot open it RW. Report
1080                  * success.
1081                  */
1082                 _exit(EXIT_SUCCESS);
1083         } else {
1084                 return -1;
1085         }
1086 end:
1087         if (wait_shm_fd >= 0 && !sock_info->global) {
1088                 struct stat statbuf;
1089
1090                 /*
1091                  * Ensure that our user is the owner of the shm file for
1092                  * local shm. If we do not own the file, it means our
1093                  * sessiond will not have access to wake us up (there is
1094                  * probably a rogue process trying to fake our
1095                  * sessiond). Fallback to polling method in this case.
1096                  */
1097                 ret = fstat(wait_shm_fd, &statbuf);
1098                 if (ret) {
1099                         PERROR("fstat");
1100                         goto error_close;
1101                 }
1102                 if (statbuf.st_uid != getuid())
1103                         goto error_close;
1104         }
1105         return wait_shm_fd;
1106
1107 error_close:
1108         ret = close(wait_shm_fd);
1109         if (ret) {
1110                 PERROR("Error closing fd");
1111         }
1112         return -1;
1113 }
1114
1115 static
1116 char *get_map_shm(struct sock_info *sock_info)
1117 {
1118         long page_size;
1119         int wait_shm_fd, ret;
1120         char *wait_shm_mmap;
1121
1122         page_size = sysconf(_SC_PAGE_SIZE);
1123         if (page_size <= 0) {
1124                 if (!page_size) {
1125                         errno = EINVAL;
1126                 }
1127                 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1128                 goto error;
1129         }
1130
1131         wait_shm_fd = get_wait_shm(sock_info, page_size);
1132         if (wait_shm_fd < 0) {
1133                 goto error;
1134         }
1135         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1136                   MAP_SHARED, wait_shm_fd, 0);
1137         /* close shm fd immediately after taking the mmap reference */
1138         ret = close(wait_shm_fd);
1139         if (ret) {
1140                 PERROR("Error closing fd");
1141         }
1142         if (wait_shm_mmap == MAP_FAILED) {
1143                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1144                 goto error;
1145         }
1146         return wait_shm_mmap;
1147
1148 error:
1149         return NULL;
1150 }
1151
1152 static
1153 void wait_for_sessiond(struct sock_info *sock_info)
1154 {
1155         if (ust_lock()) {
1156                 goto quit;
1157         }
1158         if (wait_poll_fallback) {
1159                 goto error;
1160         }
1161         if (!sock_info->wait_shm_mmap) {
1162                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1163                 if (!sock_info->wait_shm_mmap)
1164                         goto error;
1165         }
1166         ust_unlock();
1167
1168         DBG("Waiting for %s apps sessiond", sock_info->name);
1169         /* Wait for futex wakeup */
1170         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
1171                 goto end_wait;
1172
1173         while (futex_async((int32_t *) sock_info->wait_shm_mmap,
1174                         FUTEX_WAIT, 0, NULL, NULL, 0)) {
1175                 switch (errno) {
1176                 case EWOULDBLOCK:
1177                         /* Value already changed. */
1178                         goto end_wait;
1179                 case EINTR:
1180                         /* Retry if interrupted by signal. */
1181                         break;  /* Get out of switch. */
1182                 case EFAULT:
1183                         wait_poll_fallback = 1;
1184                         DBG(
1185 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1186 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1187 "Please upgrade your kernel "
1188 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1189 "mainline). LTTng-UST will use polling mode fallback.");
1190                         if (ust_debug())
1191                                 PERROR("futex");
1192                         goto end_wait;
1193                 }
1194         }
1195 end_wait:
1196         return;
1197
1198 quit:
1199         ust_unlock();
1200         return;
1201
1202 error:
1203         ust_unlock();
1204         return;
1205 }
1206
1207 /*
1208  * This thread does not allocate any resource, except within
1209  * handle_message, within mutex protection. This mutex protects against
1210  * fork and exit.
1211  * The other moment it allocates resources is at socket connection, which
1212  * is also protected by the mutex.
1213  */
1214 static
1215 void *ust_listener_thread(void *arg)
1216 {
1217         struct sock_info *sock_info = arg;
1218         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1219         long timeout;
1220
1221         /* Restart trying to connect to the session daemon */
1222 restart:
1223         if (prev_connect_failed) {
1224                 /* Wait for sessiond availability with pipe */
1225                 wait_for_sessiond(sock_info);
1226                 if (has_waited) {
1227                         has_waited = 0;
1228                         /*
1229                          * Sleep for 5 seconds before retrying after a
1230                          * sequence of failure / wait / failure. This
1231                          * deals with a killed or broken session daemon.
1232                          */
1233                         sleep(5);
1234                 } else {
1235                         has_waited = 1;
1236                 }
1237                 prev_connect_failed = 0;
1238         }
1239
1240         if (sock_info->socket != -1) {
1241                 ret = ustcomm_close_unix_sock(sock_info->socket);
1242                 if (ret) {
1243                         ERR("Error closing %s ust cmd socket",
1244                                 sock_info->name);
1245                 }
1246                 sock_info->socket = -1;
1247         }
1248         if (sock_info->notify_socket != -1) {
1249                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1250                 if (ret) {
1251                         ERR("Error closing %s ust notify socket",
1252                                 sock_info->name);
1253                 }
1254                 sock_info->notify_socket = -1;
1255         }
1256
1257         /*
1258          * Register. We need to perform both connect and sending
1259          * registration message before doing the next connect otherwise
1260          * we may reach unix socket connect queue max limits and block
1261          * on the 2nd connect while the session daemon is awaiting the
1262          * first connect registration message.
1263          */
1264         /* Connect cmd socket */
1265         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1266         if (ret < 0) {
1267                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1268                 prev_connect_failed = 1;
1269
1270                 if (ust_lock()) {
1271                         goto quit;
1272                 }
1273
1274                 /*
1275                  * If we cannot find the sessiond daemon, don't delay
1276                  * constructor execution.
1277                  */
1278                 ret = handle_register_done(sock_info);
1279                 assert(!ret);
1280                 ust_unlock();
1281                 goto restart;
1282         }
1283         sock_info->socket = ret;
1284
1285         if (ust_lock()) {
1286                 goto quit;
1287         }
1288
1289         /*
1290          * Create only one root handle per listener thread for the whole
1291          * process lifetime, so we ensure we get ID which is statically
1292          * assigned to the root handle.
1293          */
1294         if (sock_info->root_handle == -1) {
1295                 ret = lttng_abi_create_root_handle();
1296                 if (ret < 0) {
1297                         ERR("Error creating root handle");
1298                         goto quit;
1299                 }
1300                 sock_info->root_handle = ret;
1301         }
1302
1303         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1304         if (ret < 0) {
1305                 ERR("Error registering to %s ust cmd socket",
1306                         sock_info->name);
1307                 prev_connect_failed = 1;
1308                 /*
1309                  * If we cannot register to the sessiond daemon, don't
1310                  * delay constructor execution.
1311                  */
1312                 ret = handle_register_done(sock_info);
1313                 assert(!ret);
1314                 ust_unlock();
1315                 goto restart;
1316         }
1317
1318         ust_unlock();
1319
1320         /* Connect notify socket */
1321         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1322         if (ret < 0) {
1323                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1324                 prev_connect_failed = 1;
1325
1326                 if (ust_lock()) {
1327                         goto quit;
1328                 }
1329
1330                 /*
1331                  * If we cannot find the sessiond daemon, don't delay
1332                  * constructor execution.
1333                  */
1334                 ret = handle_register_done(sock_info);
1335                 assert(!ret);
1336                 ust_unlock();
1337                 goto restart;
1338         }
1339         sock_info->notify_socket = ret;
1340
1341         timeout = get_notify_sock_timeout();
1342         if (timeout >= 0) {
1343                 /*
1344                  * Give at least 10ms to sessiond to reply to
1345                  * notifications.
1346                  */
1347                 if (timeout < 10)
1348                         timeout = 10;
1349                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1350                                 timeout);
1351                 if (ret < 0) {
1352                         WARN("Error setting socket receive timeout");
1353                 }
1354                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1355                                 timeout);
1356                 if (ret < 0) {
1357                         WARN("Error setting socket send timeout");
1358                 }
1359         } else if (timeout < -1) {
1360                 WARN("Unsupported timeout value %ld", timeout);
1361         }
1362
1363         if (ust_lock()) {
1364                 goto quit;
1365         }
1366
1367         ret = register_to_sessiond(sock_info->notify_socket,
1368                         USTCTL_SOCKET_NOTIFY);
1369         if (ret < 0) {
1370                 ERR("Error registering to %s ust notify socket",
1371                         sock_info->name);
1372                 prev_connect_failed = 1;
1373                 /*
1374                  * If we cannot register to the sessiond daemon, don't
1375                  * delay constructor execution.
1376                  */
1377                 ret = handle_register_done(sock_info);
1378                 assert(!ret);
1379                 ust_unlock();
1380                 goto restart;
1381         }
1382         sock = sock_info->socket;
1383
1384         ust_unlock();
1385
1386         for (;;) {
1387                 ssize_t len;
1388                 struct ustcomm_ust_msg lum;
1389
1390                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1391                 switch (len) {
1392                 case 0: /* orderly shutdown */
1393                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1394                         if (ust_lock()) {
1395                                 goto quit;
1396                         }
1397                         /*
1398                          * Either sessiond has shutdown or refused us by closing the socket.
1399                          * In either case, we don't want to delay construction execution,
1400                          * and we need to wait before retry.
1401                          */
1402                         prev_connect_failed = 1;
1403                         /*
1404                          * If we cannot register to the sessiond daemon, don't
1405                          * delay constructor execution.
1406                          */
1407                         ret = handle_register_done(sock_info);
1408                         assert(!ret);
1409                         ust_unlock();
1410                         goto end;
1411                 case sizeof(lum):
1412                         print_cmd(lum.cmd, lum.handle);
1413                         ret = handle_message(sock_info, sock, &lum);
1414                         if (ret) {
1415                                 ERR("Error handling message for %s socket",
1416                                         sock_info->name);
1417                                 /*
1418                                  * Close socket if protocol error is
1419                                  * detected.
1420                                  */
1421                                 goto end;
1422                         }
1423                         continue;
1424                 default:
1425                         if (len < 0) {
1426                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1427                         } else {
1428                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1429                         }
1430                         if (len == -ECONNRESET) {
1431                                 DBG("%s remote end closed connection", sock_info->name);
1432                                 goto end;
1433                         }
1434                         goto end;
1435                 }
1436
1437         }
1438 end:
1439         if (ust_lock()) {
1440                 goto quit;
1441         }
1442         /* Cleanup socket handles before trying to reconnect */
1443         lttng_ust_objd_table_owner_cleanup(sock_info);
1444         ust_unlock();
1445         goto restart;   /* try to reconnect */
1446
1447 quit:
1448         ust_unlock();
1449
1450         pthread_mutex_lock(&ust_exit_mutex);
1451         sock_info->thread_active = 0;
1452         pthread_mutex_unlock(&ust_exit_mutex);
1453         return NULL;
1454 }
1455
1456 /*
1457  * Weak symbol to call when the ust malloc wrapper is not loaded.
1458  */
1459 __attribute__((weak))
1460 void lttng_ust_malloc_wrapper_init(void)
1461 {
1462 }
1463
1464 /*
1465  * sessiond monitoring thread: monitor presence of global and per-user
1466  * sessiond by polling the application common named pipe.
1467  */
1468 void __attribute__((constructor)) lttng_ust_init(void)
1469 {
1470         struct timespec constructor_timeout;
1471         sigset_t sig_all_blocked, orig_parent_mask;
1472         pthread_attr_t thread_attr;
1473         int timeout_mode;
1474         int ret;
1475
1476         if (uatomic_xchg(&initialized, 1) == 1)
1477                 return;
1478
1479         /*
1480          * Fixup interdependency between TLS fixup mutex (which happens
1481          * to be the dynamic linker mutex) and ust_lock, taken within
1482          * the ust lock.
1483          */
1484         lttng_fixup_urcu_bp_tls();
1485         lttng_fixup_ringbuffer_tls();
1486         lttng_fixup_vtid_tls();
1487         lttng_fixup_nest_count_tls();
1488         lttng_fixup_procname_tls();
1489         lttng_fixup_ust_mutex_nest_tls();
1490
1491         /*
1492          * We want precise control over the order in which we construct
1493          * our sub-libraries vs starting to receive commands from
1494          * sessiond (otherwise leading to errors when trying to create
1495          * sessiond before the init functions are completed).
1496          */
1497         init_usterr();
1498         init_tracepoint();
1499         lttng_ust_clock_init();
1500         lttng_ust_getcpu_init();
1501         lttng_ust_statedump_init();
1502         lttng_ring_buffer_metadata_client_init();
1503         lttng_ring_buffer_client_overwrite_init();
1504         lttng_ring_buffer_client_overwrite_rt_init();
1505         lttng_ring_buffer_client_discard_init();
1506         lttng_ring_buffer_client_discard_rt_init();
1507         lttng_perf_counter_init();
1508         lttng_context_init();
1509         /*
1510          * Invoke ust malloc wrapper init before starting other threads.
1511          */
1512         lttng_ust_malloc_wrapper_init();
1513
1514         timeout_mode = get_constructor_timeout(&constructor_timeout);
1515
1516         ret = sem_init(&constructor_wait, 0, 0);
1517         if (ret) {
1518                 PERROR("sem_init");
1519         }
1520
1521         ret = setup_local_apps();
1522         if (ret) {
1523                 DBG("local apps setup returned %d", ret);
1524         }
1525
1526         /* A new thread created by pthread_create inherits the signal mask
1527          * from the parent. To avoid any signal being received by the
1528          * listener thread, we block all signals temporarily in the parent,
1529          * while we create the listener thread.
1530          */
1531         sigfillset(&sig_all_blocked);
1532         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1533         if (ret) {
1534                 ERR("pthread_sigmask: %s", strerror(ret));
1535         }
1536
1537         ret = pthread_attr_init(&thread_attr);
1538         if (ret) {
1539                 ERR("pthread_attr_init: %s", strerror(ret));
1540         }
1541         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1542         if (ret) {
1543                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1544         }
1545
1546         pthread_mutex_lock(&ust_exit_mutex);
1547         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1548                         ust_listener_thread, &global_apps);
1549         if (ret) {
1550                 ERR("pthread_create global: %s", strerror(ret));
1551         }
1552         global_apps.thread_active = 1;
1553         pthread_mutex_unlock(&ust_exit_mutex);
1554
1555         if (local_apps.allowed) {
1556                 pthread_mutex_lock(&ust_exit_mutex);
1557                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1558                                 ust_listener_thread, &local_apps);
1559                 if (ret) {
1560                         ERR("pthread_create local: %s", strerror(ret));
1561                 }
1562                 local_apps.thread_active = 1;
1563                 pthread_mutex_unlock(&ust_exit_mutex);
1564         } else {
1565                 handle_register_done(&local_apps);
1566         }
1567         ret = pthread_attr_destroy(&thread_attr);
1568         if (ret) {
1569                 ERR("pthread_attr_destroy: %s", strerror(ret));
1570         }
1571
1572         /* Restore original signal mask in parent */
1573         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1574         if (ret) {
1575                 ERR("pthread_sigmask: %s", strerror(ret));
1576         }
1577
1578         switch (timeout_mode) {
1579         case 1: /* timeout wait */
1580                 do {
1581                         ret = sem_timedwait(&constructor_wait,
1582                                         &constructor_timeout);
1583                 } while (ret < 0 && errno == EINTR);
1584                 if (ret < 0) {
1585                         switch (errno) {
1586                         case ETIMEDOUT:
1587                                 ERR("Timed out waiting for lttng-sessiond");
1588                                 break;
1589                         case EINVAL:
1590                                 PERROR("sem_timedwait");
1591                                 break;
1592                         default:
1593                                 ERR("Unexpected error \"%s\" returned by sem_timedwait",
1594                                         strerror(errno));
1595                         }
1596                 }
1597                 break;
1598         case -1:/* wait forever */
1599                 do {
1600                         ret = sem_wait(&constructor_wait);
1601                 } while (ret < 0 && errno == EINTR);
1602                 if (ret < 0) {
1603                         switch (errno) {
1604                         case EINVAL:
1605                                 PERROR("sem_wait");
1606                                 break;
1607                         default:
1608                                 ERR("Unexpected error \"%s\" returned by sem_wait",
1609                                         strerror(errno));
1610                         }
1611                 }
1612                 break;
1613         case 0: /* no timeout */
1614                 break;
1615         }
1616 }
1617
1618 static
1619 void lttng_ust_cleanup(int exiting)
1620 {
1621         cleanup_sock_info(&global_apps, exiting);
1622         cleanup_sock_info(&local_apps, exiting);
1623         local_apps.allowed = 0;
1624         /*
1625          * The teardown in this function all affect data structures
1626          * accessed under the UST lock by the listener thread. This
1627          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1628          * that none of these threads are accessing this data at this
1629          * point.
1630          */
1631         lttng_ust_abi_exit();
1632         lttng_ust_events_exit();
1633         lttng_context_exit();
1634         lttng_perf_counter_exit();
1635         lttng_ring_buffer_client_discard_rt_exit();
1636         lttng_ring_buffer_client_discard_exit();
1637         lttng_ring_buffer_client_overwrite_rt_exit();
1638         lttng_ring_buffer_client_overwrite_exit();
1639         lttng_ring_buffer_metadata_client_exit();
1640         lttng_ust_statedump_destroy();
1641         exit_tracepoint();
1642         if (!exiting) {
1643                 /* Reinitialize values for fork */
1644                 sem_count = 2;
1645                 lttng_ust_comm_should_quit = 0;
1646                 initialized = 0;
1647         }
1648 }
1649
1650 void __attribute__((destructor)) lttng_ust_exit(void)
1651 {
1652         int ret;
1653
1654         /*
1655          * Using pthread_cancel here because:
1656          * A) we don't want to hang application teardown.
1657          * B) the thread is not allocating any resource.
1658          */
1659
1660         /*
1661          * Require the communication thread to quit. Synchronize with
1662          * mutexes to ensure it is not in a mutex critical section when
1663          * pthread_cancel is later called.
1664          */
1665         ust_lock_nocheck();
1666         lttng_ust_comm_should_quit = 1;
1667         ust_unlock();
1668
1669         pthread_mutex_lock(&ust_exit_mutex);
1670         /* cancel threads */
1671         if (global_apps.thread_active) {
1672                 ret = pthread_cancel(global_apps.ust_listener);
1673                 if (ret) {
1674                         ERR("Error cancelling global ust listener thread: %s",
1675                                 strerror(ret));
1676                 } else {
1677                         global_apps.thread_active = 0;
1678                 }
1679         }
1680         if (local_apps.thread_active) {
1681                 ret = pthread_cancel(local_apps.ust_listener);
1682                 if (ret) {
1683                         ERR("Error cancelling local ust listener thread: %s",
1684                                 strerror(ret));
1685                 } else {
1686                         local_apps.thread_active = 0;
1687                 }
1688         }
1689         pthread_mutex_unlock(&ust_exit_mutex);
1690
1691         /*
1692          * Do NOT join threads: use of sys_futex makes it impossible to
1693          * join the threads without using async-cancel, but async-cancel
1694          * is delivered by a signal, which could hit the target thread
1695          * anywhere in its code path, including while the ust_lock() is
1696          * held, causing a deadlock for the other thread. Let the OS
1697          * cleanup the threads if there are stalled in a syscall.
1698          */
1699         lttng_ust_cleanup(1);
1700 }
1701
1702 /*
1703  * We exclude the worker threads across fork and clone (except
1704  * CLONE_VM), because these system calls only keep the forking thread
1705  * running in the child.  Therefore, we don't want to call fork or clone
1706  * in the middle of an tracepoint or ust tracing state modification.
1707  * Holding this mutex protects these structures across fork and clone.
1708  */
1709 void ust_before_fork(sigset_t *save_sigset)
1710 {
1711         /*
1712          * Disable signals. This is to avoid that the child intervenes
1713          * before it is properly setup for tracing. It is safer to
1714          * disable all signals, because then we know we are not breaking
1715          * anything by restoring the original mask.
1716          */
1717         sigset_t all_sigs;
1718         int ret;
1719
1720         if (URCU_TLS(lttng_ust_nest_count))
1721                 return;
1722         /* Disable signals */
1723         sigfillset(&all_sigs);
1724         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1725         if (ret == -1) {
1726                 PERROR("sigprocmask");
1727         }
1728
1729         pthread_mutex_lock(&ust_fork_mutex);
1730
1731         ust_lock_nocheck();
1732         rcu_bp_before_fork();
1733 }
1734
1735 static void ust_after_fork_common(sigset_t *restore_sigset)
1736 {
1737         int ret;
1738
1739         DBG("process %d", getpid());
1740         ust_unlock();
1741
1742         pthread_mutex_unlock(&ust_fork_mutex);
1743
1744         /* Restore signals */
1745         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1746         if (ret == -1) {
1747                 PERROR("sigprocmask");
1748         }
1749 }
1750
1751 void ust_after_fork_parent(sigset_t *restore_sigset)
1752 {
1753         if (URCU_TLS(lttng_ust_nest_count))
1754                 return;
1755         DBG("process %d", getpid());
1756         rcu_bp_after_fork_parent();
1757         /* Release mutexes and reenable signals */
1758         ust_after_fork_common(restore_sigset);
1759 }
1760
1761 /*
1762  * After fork, in the child, we need to cleanup all the leftover state,
1763  * except the worker thread which already magically disappeared thanks
1764  * to the weird Linux fork semantics. After tyding up, we call
1765  * lttng_ust_init() again to start over as a new PID.
1766  *
1767  * This is meant for forks() that have tracing in the child between the
1768  * fork and following exec call (if there is any).
1769  */
1770 void ust_after_fork_child(sigset_t *restore_sigset)
1771 {
1772         if (URCU_TLS(lttng_ust_nest_count))
1773                 return;
1774         DBG("process %d", getpid());
1775         /* Release urcu mutexes */
1776         rcu_bp_after_fork_child();
1777         lttng_ust_cleanup(0);
1778         lttng_context_vtid_reset();
1779         /* Release mutexes and reenable signals */
1780         ust_after_fork_common(restore_sigset);
1781         lttng_ust_init();
1782 }
1783
1784 void lttng_ust_sockinfo_session_enabled(void *owner)
1785 {
1786         struct sock_info *sock_info = owner;
1787         sock_info->statedump_pending = 1;
1788 }