liblttng-ust/lttng-ust-comm.c

   1 /*
   2  * lttng-ust-comm.c
   3  *
   4  * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
   5  * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   6  *
   7  * This library is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; only
  10  * version 2.1 of the License.
  11  *
  12  * This library is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with this library; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #define _LGPL_SOURCE
  23 #include <sys/types.h>
  24 #include <sys/socket.h>
  25 #include <sys/mman.h>
  26 #include <sys/stat.h>
  27 #include <sys/types.h>
  28 #include <sys/wait.h>
  29 #include <fcntl.h>
  30 #include <unistd.h>
  31 #include <errno.h>
  32 #include <pthread.h>
  33 #include <semaphore.h>
  34 #include <time.h>
  35 #include <assert.h>
  36 #include <signal.h>
  37 #include <urcu/uatomic.h>
  38 #include <urcu/futex.h>
  39 #include <urcu/compiler.h>
  40
  41 #include <lttng/ust-events.h>
  42 #include <lttng/ust-abi.h>
  43 #include <lttng/ust.h>
  44 #include <lttng/ust-error.h>
  45 #include <lttng/ust-ctl.h>
  46 #include <urcu/tls-compat.h>
  47 #include <ust-comm.h>
  48 #include <usterr-signal-safe.h>
  49 #include <helper.h>
  50 #include "tracepoint-internal.h"
  51 #include "lttng-tracer-core.h"
  52 #include "compat.h"
  53 #include "../libringbuffer/tlsfixup.h"
  54 #include "lttng-ust-baddr.h"
  55 #include "getenv.h"
  56
  57 /*
  58  * Has lttng ust comm constructor been called ?
  59  */
  60 static int initialized;
  61
  62 /*
  63  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  64  * Held when handling a command, also held by fork() to deal with
  65  * removal of threads, and by exit path.
  66  *
  67  * The UST lock is the centralized mutex across UST tracing control and
  68  * probe registration.
  69  *
  70  * ust_exit_mutex must never nest in ust_mutex.
  71  *
  72  * ust_fork_mutex must never nest in ust_mutex.
  73  *
  74  * ust_mutex_nest is a per-thread nesting counter, allowing the perf
  75  * counter lazy initialization called by events within the statedump,
  76  * which traces while the ust_mutex is held.
  77  *
  78  * ust_lock nests within the dynamic loader lock (within glibc) because
  79  * it is taken within the library constructor.
  80  */
  81 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
  82
  83 /* Allow nesting the ust_mutex within the same thread. */
  84 static DEFINE_URCU_TLS(int, ust_mutex_nest);
  85
  86 /*
  87  * ust_exit_mutex protects thread_active variable wrt thread exit. It
  88  * cannot be done by ust_mutex because pthread_cancel(), which takes an
  89  * internal libc lock, cannot nest within ust_mutex.
  90  *
  91  * It never nests within a ust_mutex.
  92  */
  93 static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
  94
  95 /*
  96  * ust_fork_mutex protects base address statedump tracing against forks. It
  97  * prevents the dynamic loader lock to be taken (by base address statedump
  98  * tracing) while a fork is happening, thus preventing deadlock issues with
  99  * the dynamic loader lock.
 100  */
 101 static pthread_mutex_t ust_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
 102
 103 /* Should the ust comm thread quit ? */
 104 static int lttng_ust_comm_should_quit;
 105
 106 /*
 107  * Return 0 on success, -1 if should quit.
 108  * The lock is taken in both cases.
 109  * Signal-safe.
 110  */
 111 int ust_lock(void)
 112 {
 113         sigset_t sig_all_blocked, orig_mask;
 114         int ret, oldstate;
 115
 116         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 117         if (ret) {
 118                 ERR("pthread_setcancelstate: %s", strerror(ret));
 119         }
 120         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 121                 ERR("pthread_setcancelstate: unexpected oldstate");
 122         }
 123         sigfillset(&sig_all_blocked);
 124         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 125         if (ret) {
 126                 ERR("pthread_sigmask: %s", strerror(ret));
 127         }
 128         if (!URCU_TLS(ust_mutex_nest)++)
 129                 pthread_mutex_lock(&ust_mutex);
 130         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 131         if (ret) {
 132                 ERR("pthread_sigmask: %s", strerror(ret));
 133         }
 134         if (lttng_ust_comm_should_quit) {
 135                 return -1;
 136         } else {
 137                 return 0;
 138         }
 139 }
 140
 141 /*
 142  * ust_lock_nocheck() can be used in constructors/destructors, because
 143  * they are already nested within the dynamic loader lock, and therefore
 144  * have exclusive access against execution of liblttng-ust destructor.
 145  * Signal-safe.
 146  */
 147 void ust_lock_nocheck(void)
 148 {
 149         sigset_t sig_all_blocked, orig_mask;
 150         int ret, oldstate;
 151
 152         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 153         if (ret) {
 154                 ERR("pthread_setcancelstate: %s", strerror(ret));
 155         }
 156         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 157                 ERR("pthread_setcancelstate: unexpected oldstate");
 158         }
 159         sigfillset(&sig_all_blocked);
 160         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 161         if (ret) {
 162                 ERR("pthread_sigmask: %s", strerror(ret));
 163         }
 164         if (!URCU_TLS(ust_mutex_nest)++)
 165                 pthread_mutex_lock(&ust_mutex);
 166         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 167         if (ret) {
 168                 ERR("pthread_sigmask: %s", strerror(ret));
 169         }
 170 }
 171
 172 /*
 173  * Signal-safe.
 174  */
 175 void ust_unlock(void)
 176 {
 177         sigset_t sig_all_blocked, orig_mask;
 178         int ret, oldstate;
 179
 180         sigfillset(&sig_all_blocked);
 181         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 182         if (ret) {
 183                 ERR("pthread_sigmask: %s", strerror(ret));
 184         }
 185         if (!--URCU_TLS(ust_mutex_nest))
 186                 pthread_mutex_unlock(&ust_mutex);
 187         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 188         if (ret) {
 189                 ERR("pthread_sigmask: %s", strerror(ret));
 190         }
 191         ret = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &oldstate);
 192         if (ret) {
 193                 ERR("pthread_setcancelstate: %s", strerror(ret));
 194         }
 195         if (oldstate != PTHREAD_CANCEL_DISABLE) {
 196                 ERR("pthread_setcancelstate: unexpected oldstate");
 197         }
 198 }
 199
 200 /*
 201  * Wait for either of these before continuing to the main
 202  * program:
 203  * - the register_done message from sessiond daemon
 204  *   (will let the sessiond daemon enable sessions before main
 205  *   starts.)
 206  * - sessiond daemon is not reachable.
 207  * - timeout (ensuring applications are resilient to session
 208  *   daemon problems).
 209  */
 210 static sem_t constructor_wait;
 211 /*
 212  * Doing this for both the global and local sessiond.
 213  */
 214 static int sem_count = { 2 };
 215
 216 /*
 217  * Counting nesting within lttng-ust. Used to ensure that calling fork()
 218  * from liblttng-ust does not execute the pre/post fork handlers.
 219  */
 220 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
 221
 222 /*
 223  * Info about socket and associated listener thread.
 224  */
 225 struct sock_info {
 226         const char *name;
 227         pthread_t ust_listener; /* listener thread */
 228         int root_handle;
 229         int constructor_sem_posted;
 230         int allowed;
 231         int global;
 232         int thread_active;
 233
 234         char sock_path[PATH_MAX];
 235         int socket;
 236         int notify_socket;
 237
 238         char wait_shm_path[PATH_MAX];
 239         char *wait_shm_mmap;
 240         /* Keep track of lazy state dump not performed yet. */
 241         int statedump_pending;
 242 };
 243
 244 /* Socket from app (connect) to session daemon (listen) for communication */
 245 struct sock_info global_apps = {
 246         .name = "global",
 247         .global = 1,
 248
 249         .root_handle = -1,
 250         .allowed = 1,
 251         .thread_active = 0,
 252
 253         .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
 254         .socket = -1,
 255         .notify_socket = -1,
 256
 257         .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 258
 259         .statedump_pending = 0,
 260 };
 261
 262 /* TODO: allow global_apps_sock_path override */
 263
 264 struct sock_info local_apps = {
 265         .name = "local",
 266         .global = 0,
 267         .root_handle = -1,
 268         .allowed = 0,   /* Check setuid bit first */
 269         .thread_active = 0,
 270
 271         .socket = -1,
 272         .notify_socket = -1,
 273
 274         .statedump_pending = 0,
 275 };
 276
 277 static int wait_poll_fallback;
 278
 279 static const char *cmd_name_mapping[] = {
 280         [ LTTNG_UST_RELEASE ] = "Release",
 281         [ LTTNG_UST_SESSION ] = "Create Session",
 282         [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
 283
 284         [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
 285         [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
 286         [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
 287         [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
 288
 289         /* Session FD commands */
 290         [ LTTNG_UST_CHANNEL ] = "Create Channel",
 291         [ LTTNG_UST_SESSION_START ] = "Start Session",
 292         [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
 293
 294         /* Channel FD commands */
 295         [ LTTNG_UST_STREAM ] = "Create Stream",
 296         [ LTTNG_UST_EVENT ] = "Create Event",
 297
 298         /* Event and Channel FD commands */
 299         [ LTTNG_UST_CONTEXT ] = "Create Context",
 300         [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
 301
 302         /* Event, Channel and Session commands */
 303         [ LTTNG_UST_ENABLE ] = "Enable",
 304         [ LTTNG_UST_DISABLE ] = "Disable",
 305
 306         /* Tracepoint list commands */
 307         [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
 308         [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
 309
 310         /* Event FD commands */
 311         [ LTTNG_UST_FILTER ] = "Create Filter",
 312         [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
 313 };
 314
 315 static const char *str_timeout;
 316 static int got_timeout_env;
 317
 318 extern void lttng_ring_buffer_client_overwrite_init(void);
 319 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 320 extern void lttng_ring_buffer_client_discard_init(void);
 321 extern void lttng_ring_buffer_client_discard_rt_init(void);
 322 extern void lttng_ring_buffer_metadata_client_init(void);
 323 extern void lttng_ring_buffer_client_overwrite_exit(void);
 324 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 325 extern void lttng_ring_buffer_client_discard_exit(void);
 326 extern void lttng_ring_buffer_client_discard_rt_exit(void);
 327 extern void lttng_ring_buffer_metadata_client_exit(void);
 328
 329 /*
 330  * Returns the HOME directory path. Caller MUST NOT free(3) the returned
 331  * pointer.
 332  */
 333 static
 334 const char *get_lttng_home_dir(void)
 335 {
 336        const char *val;
 337
 338        val = (const char *) lttng_secure_getenv("LTTNG_HOME");
 339        if (val != NULL) {
 340                return val;
 341        }
 342        return (const char *) lttng_secure_getenv("HOME");
 343 }
 344
 345 /*
 346  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 347  */
 348 static
 349 void lttng_fixup_nest_count_tls(void)
 350 {
 351         asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 352 }
 353
 354 static
 355 void lttng_fixup_ust_mutex_nest_tls(void)
 356 {
 357         asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest)));
 358 }
 359
 360 /*
 361  * Fixup urcu bp TLS.
 362  */
 363 static
 364 void lttng_fixup_urcu_bp_tls(void)
 365 {
 366         rcu_read_lock();
 367         rcu_read_unlock();
 368 }
 369
 370 int lttng_get_notify_socket(void *owner)
 371 {
 372         struct sock_info *info = owner;
 373
 374         return info->notify_socket;
 375 }
 376
 377 static
 378 void print_cmd(int cmd, int handle)
 379 {
 380         const char *cmd_name = "Unknown";
 381
 382         if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
 383                         && cmd_name_mapping[cmd]) {
 384                 cmd_name = cmd_name_mapping[cmd];
 385         }
 386         DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
 387                 cmd_name, cmd,
 388                 lttng_ust_obj_get_name(handle), handle);
 389 }
 390
 391 static
 392 int setup_local_apps(void)
 393 {
 394         const char *home_dir;
 395         uid_t uid;
 396
 397         uid = getuid();
 398         /*
 399          * Disallow per-user tracing for setuid binaries.
 400          */
 401         if (uid != geteuid()) {
 402                 assert(local_apps.allowed == 0);
 403                 return 0;
 404         }
 405         home_dir = get_lttng_home_dir();
 406         if (!home_dir) {
 407                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 408                 assert(local_apps.allowed == 0);
 409                 return -ENOENT;
 410         }
 411         local_apps.allowed = 1;
 412         snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
 413                 home_dir,
 414                 LTTNG_DEFAULT_HOME_RUNDIR,
 415                 LTTNG_UST_SOCK_FILENAME);
 416         snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
 417                 LTTNG_UST_WAIT_FILENAME,
 418                 uid);
 419         return 0;
 420 }
 421
 422 /*
 423  * Get notify_sock timeout, in ms.
 424  * -1: wait forever. 0: don't wait. >0: timeout, in ms.
 425  */
 426 static
 427 long get_timeout(void)
 428 {
 429         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 430
 431         if (!got_timeout_env) {
 432                 str_timeout = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 433                 got_timeout_env = 1;
 434         }
 435         if (str_timeout)
 436                 constructor_delay_ms = strtol(str_timeout, NULL, 10);
 437         return constructor_delay_ms;
 438 }
 439
 440 static
 441 long get_notify_sock_timeout(void)
 442 {
 443         return get_timeout();
 444 }
 445
 446 /*
 447  * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
 448  */
 449 static
 450 int get_constructor_timeout(struct timespec *constructor_timeout)
 451 {
 452         long constructor_delay_ms;
 453         int ret;
 454
 455         constructor_delay_ms = get_timeout();
 456
 457         switch (constructor_delay_ms) {
 458         case -1:/* fall-through */
 459         case 0:
 460                 return constructor_delay_ms;
 461         default:
 462                 break;
 463         }
 464
 465         /*
 466          * If we are unable to find the current time, don't wait.
 467          */
 468         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 469         if (ret) {
 470                 /* Don't wait. */
 471                 return 0;
 472         }
 473         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 474         constructor_timeout->tv_nsec +=
 475                 (constructor_delay_ms % 1000UL) * 1000000UL;
 476         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 477                 constructor_timeout->tv_sec++;
 478                 constructor_timeout->tv_nsec -= 1000000000UL;
 479         }
 480         /* Timeout wait (constructor_delay_ms). */
 481         return 1;
 482 }
 483
 484 static
 485 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 486 {
 487         return ustcomm_send_reg_msg(socket,
 488                 type,
 489                 CAA_BITS_PER_LONG,
 490                 lttng_alignof(uint8_t) * CHAR_BIT,
 491                 lttng_alignof(uint16_t) * CHAR_BIT,
 492                 lttng_alignof(uint32_t) * CHAR_BIT,
 493                 lttng_alignof(uint64_t) * CHAR_BIT,
 494                 lttng_alignof(unsigned long) * CHAR_BIT);
 495 }
 496
 497 static
 498 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 499 {
 500         ssize_t len;
 501
 502         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 503         switch (len) {
 504         case sizeof(*lur):
 505                 DBG("message successfully sent");
 506                 return 0;
 507         default:
 508                 if (len == -ECONNRESET) {
 509                         DBG("remote end closed connection");
 510                         return 0;
 511                 }
 512                 if (len < 0)
 513                         return len;
 514                 DBG("incorrect message size: %zd", len);
 515                 return -EINVAL;
 516         }
 517 }
 518
 519 static
 520 int handle_register_done(struct sock_info *sock_info)
 521 {
 522         int ret;
 523
 524         if (sock_info->constructor_sem_posted)
 525                 return 0;
 526         sock_info->constructor_sem_posted = 1;
 527         if (uatomic_read(&sem_count) <= 0) {
 528                 return 0;
 529         }
 530         ret = uatomic_add_return(&sem_count, -1);
 531         if (ret == 0) {
 532                 ret = sem_post(&constructor_wait);
 533                 assert(!ret);
 534         }
 535         return 0;
 536 }
 537
 538 /*
 539  * Only execute pending statedump after the constructor semaphore has
 540  * been posted by each listener thread. This means statedump will only
 541  * be performed after the "registration done" command is received from
 542  * each session daemon the application is connected to.
 543  *
 544  * This ensures we don't run into deadlock issues with the dynamic
 545  * loader mutex, which is held while the constructor is called and
 546  * waiting on the constructor semaphore. All operations requiring this
 547  * dynamic loader lock need to be postponed using this mechanism.
 548  */
 549 static
 550 void handle_pending_statedump(struct sock_info *sock_info)
 551 {
 552         int ctor_passed = sock_info->constructor_sem_posted;
 553
 554         if (ctor_passed && sock_info->statedump_pending) {
 555                 sock_info->statedump_pending = 0;
 556                 pthread_mutex_lock(&ust_fork_mutex);
 557                 lttng_handle_pending_statedump(sock_info);
 558                 pthread_mutex_unlock(&ust_fork_mutex);
 559         }
 560 }
 561
 562 static
 563 int handle_message(struct sock_info *sock_info,
 564                 int sock, struct ustcomm_ust_msg *lum)
 565 {
 566         int ret = 0;
 567         const struct lttng_ust_objd_ops *ops;
 568         struct ustcomm_ust_reply lur;
 569         union ust_args args;
 570         ssize_t len;
 571
 572         memset(&lur, 0, sizeof(lur));
 573
 574         if (ust_lock()) {
 575                 ret = -LTTNG_UST_ERR_EXITING;
 576                 goto error;
 577         }
 578
 579         ops = objd_ops(lum->handle);
 580         if (!ops) {
 581                 ret = -ENOENT;
 582                 goto error;
 583         }
 584
 585         switch (lum->cmd) {
 586         case LTTNG_UST_REGISTER_DONE:
 587                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 588                         ret = handle_register_done(sock_info);
 589                 else
 590                         ret = -EINVAL;
 591                 break;
 592         case LTTNG_UST_RELEASE:
 593                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 594                         ret = -EPERM;
 595                 else
 596                         ret = lttng_ust_objd_unref(lum->handle, 1);
 597                 break;
 598         case LTTNG_UST_FILTER:
 599         {
 600                 /* Receive filter data */
 601                 struct lttng_ust_filter_bytecode_node *bytecode;
 602
 603                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 604                         ERR("Filter data size is too large: %u bytes",
 605                                 lum->u.filter.data_size);
 606                         ret = -EINVAL;
 607                         goto error;
 608                 }
 609
 610                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 611                         ERR("Filter reloc offset %u is not within data",
 612                                 lum->u.filter.reloc_offset);
 613                         ret = -EINVAL;
 614                         goto error;
 615                 }
 616
 617                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 618                 if (!bytecode) {
 619                         ret = -ENOMEM;
 620                         goto error;
 621                 }
 622                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 623                                 lum->u.filter.data_size);
 624                 switch (len) {
 625                 case 0: /* orderly shutdown */
 626                         ret = 0;
 627                         free(bytecode);
 628                         goto error;
 629                 default:
 630                         if (len == lum->u.filter.data_size) {
 631                                 DBG("filter data received");
 632                                 break;
 633                         } else if (len < 0) {
 634                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 635                                 if (len == -ECONNRESET) {
 636                                         ERR("%s remote end closed connection", sock_info->name);
 637                                         ret = len;
 638                                         free(bytecode);
 639                                         goto error;
 640                                 }
 641                                 ret = len;
 642                                 free(bytecode);
 643                                 goto error;
 644                         } else {
 645                                 DBG("incorrect filter data message size: %zd", len);
 646                                 ret = -EINVAL;
 647                                 free(bytecode);
 648                                 goto error;
 649                         }
 650                 }
 651                 bytecode->bc.len = lum->u.filter.data_size;
 652                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 653                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 654                 if (ops->cmd) {
 655                         ret = ops->cmd(lum->handle, lum->cmd,
 656                                         (unsigned long) bytecode,
 657                                         &args, sock_info);
 658                         if (ret) {
 659                                 free(bytecode);
 660                         }
 661                         /* don't free bytecode if everything went fine. */
 662                 } else {
 663                         ret = -ENOSYS;
 664                         free(bytecode);
 665                 }
 666                 break;
 667         }
 668         case LTTNG_UST_EXCLUSION:
 669         {
 670                 /* Receive exclusion names */
 671                 struct lttng_ust_excluder_node *node;
 672                 unsigned int count;
 673
 674                 count = lum->u.exclusion.count;
 675                 if (count == 0) {
 676                         /* There are no names to read */
 677                         ret = 0;
 678                         goto error;
 679                 }
 680                 node = zmalloc(sizeof(*node) +
 681                                 count * LTTNG_UST_SYM_NAME_LEN);
 682                 if (!node) {
 683                         ret = -ENOMEM;
 684                         goto error;
 685                 }
 686                 node->excluder.count = count;
 687                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 688                                 count * LTTNG_UST_SYM_NAME_LEN);
 689                 switch (len) {
 690                 case 0: /* orderly shutdown */
 691                         ret = 0;
 692                         free(node);
 693                         goto error;
 694                 default:
 695                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 696                                 DBG("Exclusion data received");
 697                                 break;
 698                         } else if (len < 0) {
 699                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 700                                 if (len == -ECONNRESET) {
 701                                         ERR("%s remote end closed connection", sock_info->name);
 702                                         ret = len;
 703                                         free(node);
 704                                         goto error;
 705                                 }
 706                                 ret = len;
 707                                 free(node);
 708                                 goto error;
 709                         } else {
 710                                 DBG("Incorrect exclusion data message size: %zd", len);
 711                                 ret = -EINVAL;
 712                                 free(node);
 713                                 goto error;
 714                         }
 715                 }
 716                 if (ops->cmd) {
 717                         ret = ops->cmd(lum->handle, lum->cmd,
 718                                         (unsigned long) node,
 719                                         &args, sock_info);
 720                         if (ret) {
 721                                 free(node);
 722                         }
 723                         /* Don't free exclusion data if everything went fine. */
 724                 } else {
 725                         ret = -ENOSYS;
 726                         free(node);
 727                 }
 728                 break;
 729         }
 730         case LTTNG_UST_CHANNEL:
 731         {
 732                 void *chan_data;
 733                 int wakeup_fd;
 734
 735                 len = ustcomm_recv_channel_from_sessiond(sock,
 736                                 &chan_data, lum->u.channel.len,
 737                                 &wakeup_fd);
 738                 switch (len) {
 739                 case 0: /* orderly shutdown */
 740                         ret = 0;
 741                         goto error;
 742                 default:
 743                         if (len == lum->u.channel.len) {
 744                                 DBG("channel data received");
 745                                 break;
 746                         } else if (len < 0) {
 747                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 748                                 if (len == -ECONNRESET) {
 749                                         ERR("%s remote end closed connection", sock_info->name);
 750                                         ret = len;
 751                                         goto error;
 752                                 }
 753                                 ret = len;
 754                                 goto error;
 755                         } else {
 756                                 DBG("incorrect channel data message size: %zd", len);
 757                                 ret = -EINVAL;
 758                                 goto error;
 759                         }
 760                 }
 761                 args.channel.chan_data = chan_data;
 762                 args.channel.wakeup_fd = wakeup_fd;
 763                 if (ops->cmd)
 764                         ret = ops->cmd(lum->handle, lum->cmd,
 765                                         (unsigned long) &lum->u,
 766                                         &args, sock_info);
 767                 else
 768                         ret = -ENOSYS;
 769                 break;
 770         }
 771         case LTTNG_UST_STREAM:
 772         {
 773                 /* Receive shm_fd, wakeup_fd */
 774                 ret = ustcomm_recv_stream_from_sessiond(sock,
 775                         &lum->u.stream.len,
 776                         &args.stream.shm_fd,
 777                         &args.stream.wakeup_fd);
 778                 if (ret) {
 779                         goto error;
 780                 }
 781                 if (ops->cmd)
 782                         ret = ops->cmd(lum->handle, lum->cmd,
 783                                         (unsigned long) &lum->u,
 784                                         &args, sock_info);
 785                 else
 786                         ret = -ENOSYS;
 787                 break;
 788         }
 789         default:
 790                 if (ops->cmd)
 791                         ret = ops->cmd(lum->handle, lum->cmd,
 792                                         (unsigned long) &lum->u,
 793                                         &args, sock_info);
 794                 else
 795                         ret = -ENOSYS;
 796                 break;
 797         }
 798
 799         lur.handle = lum->handle;
 800         lur.cmd = lum->cmd;
 801         lur.ret_val = ret;
 802         if (ret >= 0) {
 803                 lur.ret_code = LTTNG_UST_OK;
 804         } else {
 805                 /*
 806                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 807                  * error that are not caused by the transport, except if
 808                  * we already have a more precise error message to
 809                  * report.
 810                  */
 811                 if (ret > -LTTNG_UST_ERR) {
 812                         /* Translate code to UST error. */
 813                         switch (ret) {
 814                         case -EEXIST:
 815                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 816                                 break;
 817                         case -EINVAL:
 818                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 819                                 break;
 820                         case -ENOENT:
 821                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 822                                 break;
 823                         case -EPERM:
 824                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 825                                 break;
 826                         case -ENOSYS:
 827                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 828                                 break;
 829                         default:
 830                                 lur.ret_code = -LTTNG_UST_ERR;
 831                                 break;
 832                         }
 833                 } else {
 834                         lur.ret_code = ret;
 835                 }
 836         }
 837         if (ret >= 0) {
 838                 switch (lum->cmd) {
 839                 case LTTNG_UST_TRACER_VERSION:
 840                         lur.u.version = lum->u.version;
 841                         break;
 842                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 843                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 844                         break;
 845                 }
 846         }
 847         DBG("Return value: %d", lur.ret_val);
 848         ret = send_reply(sock, &lur);
 849         if (ret < 0) {
 850                 DBG("error sending reply");
 851                 goto error;
 852         }
 853
 854         /*
 855          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 856          * after the reply.
 857          */
 858         if (lur.ret_code == LTTNG_UST_OK) {
 859                 switch (lum->cmd) {
 860                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 861                         len = ustcomm_send_unix_sock(sock,
 862                                 &args.field_list.entry,
 863                                 sizeof(args.field_list.entry));
 864                         if (len < 0) {
 865                                 ret = len;
 866                                 goto error;
 867                         }
 868                         if (len != sizeof(args.field_list.entry)) {
 869                                 ret = -EINVAL;
 870                                 goto error;
 871                         }
 872                 }
 873         }
 874
 875 error:
 876         ust_unlock();
 877
 878         /*
 879          * Performed delayed statedump operations outside of the UST
 880          * lock. We need to take the dynamic loader lock before we take
 881          * the UST lock internally within handle_pending_statedump().
 882           */
 883         handle_pending_statedump(sock_info);
 884
 885         return ret;
 886 }
 887
 888 static
 889 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 890 {
 891         int ret;
 892
 893         if (sock_info->root_handle != -1) {
 894                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 895                 if (ret) {
 896                         ERR("Error unref root handle");
 897                 }
 898                 sock_info->root_handle = -1;
 899         }
 900         sock_info->constructor_sem_posted = 0;
 901
 902         /*
 903          * wait_shm_mmap, socket and notify socket are used by listener
 904          * threads outside of the ust lock, so we cannot tear them down
 905          * ourselves, because we cannot join on these threads. Leave
 906          * responsibility of cleaning up these resources to the OS
 907          * process exit.
 908          */
 909         if (exiting)
 910                 return;
 911
 912         if (sock_info->socket != -1) {
 913                 ret = ustcomm_close_unix_sock(sock_info->socket);
 914                 if (ret) {
 915                         ERR("Error closing ust cmd socket");
 916                 }
 917                 sock_info->socket = -1;
 918         }
 919         if (sock_info->notify_socket != -1) {
 920                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 921                 if (ret) {
 922                         ERR("Error closing ust notify socket");
 923                 }
 924                 sock_info->notify_socket = -1;
 925         }
 926         if (sock_info->wait_shm_mmap) {
 927                 long page_size;
 928
 929                 page_size = sysconf(_SC_PAGE_SIZE);
 930                 if (page_size <= 0) {
 931                         if (!page_size) {
 932                                 errno = EINVAL;
 933                         }
 934                         PERROR("Error in sysconf(_SC_PAGE_SIZE)");
 935                 } else {
 936                         ret = munmap(sock_info->wait_shm_mmap, page_size);
 937                         if (ret) {
 938                                 ERR("Error unmapping wait shm");
 939                         }
 940                 }
 941                 sock_info->wait_shm_mmap = NULL;
 942         }
 943 }
 944
 945 /*
 946  * Using fork to set umask in the child process (not multi-thread safe).
 947  * We deal with the shm_open vs ftruncate race (happening when the
 948  * sessiond owns the shm and does not let everybody modify it, to ensure
 949  * safety against shm_unlink) by simply letting the mmap fail and
 950  * retrying after a few seconds.
 951  * For global shm, everybody has rw access to it until the sessiond
 952  * starts.
 953  */
 954 static
 955 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 956 {
 957         int wait_shm_fd, ret;
 958         pid_t pid;
 959
 960         /*
 961          * Try to open read-only.
 962          */
 963         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 964         if (wait_shm_fd >= 0) {
 965                 int32_t tmp_read;
 966                 ssize_t len;
 967                 size_t bytes_read = 0;
 968
 969                 /*
 970                  * Try to read the fd. If unable to do so, try opening
 971                  * it in write mode.
 972                  */
 973                 do {
 974                         len = read(wait_shm_fd,
 975                                 &((char *) &tmp_read)[bytes_read],
 976                                 sizeof(tmp_read) - bytes_read);
 977                         if (len > 0) {
 978                                 bytes_read += len;
 979                         }
 980                 } while ((len < 0 && errno == EINTR)
 981                         || (len > 0 && bytes_read < sizeof(tmp_read)));
 982                 if (bytes_read != sizeof(tmp_read)) {
 983                         ret = close(wait_shm_fd);
 984                         if (ret) {
 985                                 ERR("close wait_shm_fd");
 986                         }
 987                         goto open_write;
 988                 }
 989                 goto end;
 990         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 991                 /*
 992                  * Real-only open did not work, and it's not because the
 993                  * entry was not present. It's a failure that prohibits
 994                  * using shm.
 995                  */
 996                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 997                 goto end;
 998         }
 999
1000 open_write:
1001         /*
1002          * If the open failed because the file did not exist, or because
1003          * the file was not truncated yet, try creating it ourself.
1004          */
1005         URCU_TLS(lttng_ust_nest_count)++;
1006         pid = fork();
1007         URCU_TLS(lttng_ust_nest_count)--;
1008         if (pid > 0) {
1009                 int status;
1010
1011                 /*
1012                  * Parent: wait for child to return, in which case the
1013                  * shared memory map will have been created.
1014                  */
1015                 pid = wait(&status);
1016                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
1017                         wait_shm_fd = -1;
1018                         goto end;
1019                 }
1020                 /*
1021                  * Try to open read-only again after creation.
1022                  */
1023                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1024                 if (wait_shm_fd < 0) {
1025                         /*
1026                          * Real-only open did not work. It's a failure
1027                          * that prohibits using shm.
1028                          */
1029                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1030                         goto end;
1031                 }
1032                 goto end;
1033         } else if (pid == 0) {
1034                 int create_mode;
1035
1036                 /* Child */
1037                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1038                 if (sock_info->global)
1039                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1040                 /*
1041                  * We're alone in a child process, so we can modify the
1042                  * process-wide umask.
1043                  */
1044                 umask(~create_mode);
1045                 /*
1046                  * Try creating shm (or get rw access).
1047                  * We don't do an exclusive open, because we allow other
1048                  * processes to create+ftruncate it concurrently.
1049                  */
1050                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1051                                 O_RDWR | O_CREAT, create_mode);
1052                 if (wait_shm_fd >= 0) {
1053                         ret = ftruncate(wait_shm_fd, mmap_size);
1054                         if (ret) {
1055                                 PERROR("ftruncate");
1056                                 _exit(EXIT_FAILURE);
1057                         }
1058                         _exit(EXIT_SUCCESS);
1059                 }
1060                 /*
1061                  * For local shm, we need to have rw access to accept
1062                  * opening it: this means the local sessiond will be
1063                  * able to wake us up. For global shm, we open it even
1064                  * if rw access is not granted, because the root.root
1065                  * sessiond will be able to override all rights and wake
1066                  * us up.
1067                  */
1068                 if (!sock_info->global && errno != EACCES) {
1069                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1070                         _exit(EXIT_FAILURE);
1071                 }
1072                 /*
1073                  * The shm exists, but we cannot open it RW. Report
1074                  * success.
1075                  */
1076                 _exit(EXIT_SUCCESS);
1077         } else {
1078                 return -1;
1079         }
1080 end:
1081         if (wait_shm_fd >= 0 && !sock_info->global) {
1082                 struct stat statbuf;
1083
1084                 /*
1085                  * Ensure that our user is the owner of the shm file for
1086                  * local shm. If we do not own the file, it means our
1087                  * sessiond will not have access to wake us up (there is
1088                  * probably a rogue process trying to fake our
1089                  * sessiond). Fallback to polling method in this case.
1090                  */
1091                 ret = fstat(wait_shm_fd, &statbuf);
1092                 if (ret) {
1093                         PERROR("fstat");
1094                         goto error_close;
1095                 }
1096                 if (statbuf.st_uid != getuid())
1097                         goto error_close;
1098         }
1099         return wait_shm_fd;
1100
1101 error_close:
1102         ret = close(wait_shm_fd);
1103         if (ret) {
1104                 PERROR("Error closing fd");
1105         }
1106         return -1;
1107 }
1108
1109 static
1110 char *get_map_shm(struct sock_info *sock_info)
1111 {
1112         long page_size;
1113         int wait_shm_fd, ret;
1114         char *wait_shm_mmap;
1115
1116         page_size = sysconf(_SC_PAGE_SIZE);
1117         if (page_size <= 0) {
1118                 if (!page_size) {
1119                         errno = EINVAL;
1120                 }
1121                 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1122                 goto error;
1123         }
1124
1125         wait_shm_fd = get_wait_shm(sock_info, page_size);
1126         if (wait_shm_fd < 0) {
1127                 goto error;
1128         }
1129         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1130                   MAP_SHARED, wait_shm_fd, 0);
1131         /* close shm fd immediately after taking the mmap reference */
1132         ret = close(wait_shm_fd);
1133         if (ret) {
1134                 PERROR("Error closing fd");
1135         }
1136         if (wait_shm_mmap == MAP_FAILED) {
1137                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1138                 goto error;
1139         }
1140         return wait_shm_mmap;
1141
1142 error:
1143         return NULL;
1144 }
1145
1146 static
1147 void wait_for_sessiond(struct sock_info *sock_info)
1148 {
1149         if (ust_lock()) {
1150                 goto quit;
1151         }
1152         if (wait_poll_fallback) {
1153                 goto error;
1154         }
1155         if (!sock_info->wait_shm_mmap) {
1156                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1157                 if (!sock_info->wait_shm_mmap)
1158                         goto error;
1159         }
1160         ust_unlock();
1161
1162         DBG("Waiting for %s apps sessiond", sock_info->name);
1163         /* Wait for futex wakeup */
1164         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
1165                 goto end_wait;
1166
1167         while (futex_async((int32_t *) sock_info->wait_shm_mmap,
1168                         FUTEX_WAIT, 0, NULL, NULL, 0)) {
1169                 switch (errno) {
1170                 case EWOULDBLOCK:
1171                         /* Value already changed. */
1172                         goto end_wait;
1173                 case EINTR:
1174                         /* Retry if interrupted by signal. */
1175                         break;  /* Get out of switch. */
1176                 case EFAULT:
1177                         wait_poll_fallback = 1;
1178                         DBG(
1179 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1180 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1181 "Please upgrade your kernel "
1182 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1183 "mainline). LTTng-UST will use polling mode fallback.");
1184                         if (ust_debug())
1185                                 PERROR("futex");
1186                         goto end_wait;
1187                 }
1188         }
1189 end_wait:
1190         return;
1191
1192 quit:
1193         ust_unlock();
1194         return;
1195
1196 error:
1197         ust_unlock();
1198         return;
1199 }
1200
1201 /*
1202  * This thread does not allocate any resource, except within
1203  * handle_message, within mutex protection. This mutex protects against
1204  * fork and exit.
1205  * The other moment it allocates resources is at socket connection, which
1206  * is also protected by the mutex.
1207  */
1208 static
1209 void *ust_listener_thread(void *arg)
1210 {
1211         struct sock_info *sock_info = arg;
1212         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1213         long timeout;
1214
1215         /* Restart trying to connect to the session daemon */
1216 restart:
1217         if (prev_connect_failed) {
1218                 /* Wait for sessiond availability with pipe */
1219                 wait_for_sessiond(sock_info);
1220                 if (has_waited) {
1221                         has_waited = 0;
1222                         /*
1223                          * Sleep for 5 seconds before retrying after a
1224                          * sequence of failure / wait / failure. This
1225                          * deals with a killed or broken session daemon.
1226                          */
1227                         sleep(5);
1228                 } else {
1229                         has_waited = 1;
1230                 }
1231                 prev_connect_failed = 0;
1232         }
1233
1234         if (sock_info->socket != -1) {
1235                 ret = ustcomm_close_unix_sock(sock_info->socket);
1236                 if (ret) {
1237                         ERR("Error closing %s ust cmd socket",
1238                                 sock_info->name);
1239                 }
1240                 sock_info->socket = -1;
1241         }
1242         if (sock_info->notify_socket != -1) {
1243                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1244                 if (ret) {
1245                         ERR("Error closing %s ust notify socket",
1246                                 sock_info->name);
1247                 }
1248                 sock_info->notify_socket = -1;
1249         }
1250
1251         /*
1252          * Register. We need to perform both connect and sending
1253          * registration message before doing the next connect otherwise
1254          * we may reach unix socket connect queue max limits and block
1255          * on the 2nd connect while the session daemon is awaiting the
1256          * first connect registration message.
1257          */
1258         /* Connect cmd socket */
1259         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1260         if (ret < 0) {
1261                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1262                 prev_connect_failed = 1;
1263
1264                 if (ust_lock()) {
1265                         goto quit;
1266                 }
1267
1268                 /*
1269                  * If we cannot find the sessiond daemon, don't delay
1270                  * constructor execution.
1271                  */
1272                 ret = handle_register_done(sock_info);
1273                 assert(!ret);
1274                 ust_unlock();
1275                 goto restart;
1276         }
1277         sock_info->socket = ret;
1278
1279         if (ust_lock()) {
1280                 goto quit;
1281         }
1282
1283         /*
1284          * Create only one root handle per listener thread for the whole
1285          * process lifetime, so we ensure we get ID which is statically
1286          * assigned to the root handle.
1287          */
1288         if (sock_info->root_handle == -1) {
1289                 ret = lttng_abi_create_root_handle();
1290                 if (ret < 0) {
1291                         ERR("Error creating root handle");
1292                         goto quit;
1293                 }
1294                 sock_info->root_handle = ret;
1295         }
1296
1297         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1298         if (ret < 0) {
1299                 ERR("Error registering to %s ust cmd socket",
1300                         sock_info->name);
1301                 prev_connect_failed = 1;
1302                 /*
1303                  * If we cannot register to the sessiond daemon, don't
1304                  * delay constructor execution.
1305                  */
1306                 ret = handle_register_done(sock_info);
1307                 assert(!ret);
1308                 ust_unlock();
1309                 goto restart;
1310         }
1311
1312         ust_unlock();
1313
1314         /* Connect notify socket */
1315         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1316         if (ret < 0) {
1317                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1318                 prev_connect_failed = 1;
1319
1320                 if (ust_lock()) {
1321                         goto quit;
1322                 }
1323
1324                 /*
1325                  * If we cannot find the sessiond daemon, don't delay
1326                  * constructor execution.
1327                  */
1328                 ret = handle_register_done(sock_info);
1329                 assert(!ret);
1330                 ust_unlock();
1331                 goto restart;
1332         }
1333         sock_info->notify_socket = ret;
1334
1335         timeout = get_notify_sock_timeout();
1336         if (timeout >= 0) {
1337                 /*
1338                  * Give at least 10ms to sessiond to reply to
1339                  * notifications.
1340                  */
1341                 if (timeout < 10)
1342                         timeout = 10;
1343                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1344                                 timeout);
1345                 if (ret < 0) {
1346                         WARN("Error setting socket receive timeout");
1347                 }
1348                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1349                                 timeout);
1350                 if (ret < 0) {
1351                         WARN("Error setting socket send timeout");
1352                 }
1353         } else if (timeout < -1) {
1354                 WARN("Unsupported timeout value %ld", timeout);
1355         }
1356
1357         if (ust_lock()) {
1358                 goto quit;
1359         }
1360
1361         ret = register_to_sessiond(sock_info->notify_socket,
1362                         USTCTL_SOCKET_NOTIFY);
1363         if (ret < 0) {
1364                 ERR("Error registering to %s ust notify socket",
1365                         sock_info->name);
1366                 prev_connect_failed = 1;
1367                 /*
1368                  * If we cannot register to the sessiond daemon, don't
1369                  * delay constructor execution.
1370                  */
1371                 ret = handle_register_done(sock_info);
1372                 assert(!ret);
1373                 ust_unlock();
1374                 goto restart;
1375         }
1376         sock = sock_info->socket;
1377
1378         ust_unlock();
1379
1380         for (;;) {
1381                 ssize_t len;
1382                 struct ustcomm_ust_msg lum;
1383
1384                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1385                 switch (len) {
1386                 case 0: /* orderly shutdown */
1387                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1388                         if (ust_lock()) {
1389                                 goto quit;
1390                         }
1391                         /*
1392                          * Either sessiond has shutdown or refused us by closing the socket.
1393                          * In either case, we don't want to delay construction execution,
1394                          * and we need to wait before retry.
1395                          */
1396                         prev_connect_failed = 1;
1397                         /*
1398                          * If we cannot register to the sessiond daemon, don't
1399                          * delay constructor execution.
1400                          */
1401                         ret = handle_register_done(sock_info);
1402                         assert(!ret);
1403                         ust_unlock();
1404                         goto end;
1405                 case sizeof(lum):
1406                         print_cmd(lum.cmd, lum.handle);
1407                         ret = handle_message(sock_info, sock, &lum);
1408                         if (ret) {
1409                                 ERR("Error handling message for %s socket",
1410                                         sock_info->name);
1411                                 /*
1412                                  * Close socket if protocol error is
1413                                  * detected.
1414                                  */
1415                                 goto end;
1416                         }
1417                         continue;
1418                 default:
1419                         if (len < 0) {
1420                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1421                         } else {
1422                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1423                         }
1424                         if (len == -ECONNRESET) {
1425                                 DBG("%s remote end closed connection", sock_info->name);
1426                                 goto end;
1427                         }
1428                         goto end;
1429                 }
1430
1431         }
1432 end:
1433         if (ust_lock()) {
1434                 goto quit;
1435         }
1436         /* Cleanup socket handles before trying to reconnect */
1437         lttng_ust_objd_table_owner_cleanup(sock_info);
1438         ust_unlock();
1439         goto restart;   /* try to reconnect */
1440
1441 quit:
1442         ust_unlock();
1443
1444         pthread_mutex_lock(&ust_exit_mutex);
1445         sock_info->thread_active = 0;
1446         pthread_mutex_unlock(&ust_exit_mutex);
1447         return NULL;
1448 }
1449
1450 /*
1451  * Weak symbol to call when the ust malloc wrapper is not loaded.
1452  */
1453 __attribute__((weak))
1454 void lttng_ust_malloc_wrapper_init(void)
1455 {
1456 }
1457
1458 /*
1459  * sessiond monitoring thread: monitor presence of global and per-user
1460  * sessiond by polling the application common named pipe.
1461  */
1462 void __attribute__((constructor)) lttng_ust_init(void)
1463 {
1464         struct timespec constructor_timeout;
1465         sigset_t sig_all_blocked, orig_parent_mask;
1466         pthread_attr_t thread_attr;
1467         int timeout_mode;
1468         int ret;
1469
1470         if (uatomic_xchg(&initialized, 1) == 1)
1471                 return;
1472
1473         /*
1474          * Fixup interdependency between TLS fixup mutex (which happens
1475          * to be the dynamic linker mutex) and ust_lock, taken within
1476          * the ust lock.
1477          */
1478         lttng_fixup_urcu_bp_tls();
1479         lttng_fixup_ringbuffer_tls();
1480         lttng_fixup_vtid_tls();
1481         lttng_fixup_nest_count_tls();
1482         lttng_fixup_procname_tls();
1483         lttng_fixup_ust_mutex_nest_tls();
1484
1485         /*
1486          * We want precise control over the order in which we construct
1487          * our sub-libraries vs starting to receive commands from
1488          * sessiond (otherwise leading to errors when trying to create
1489          * sessiond before the init functions are completed).
1490          */
1491         init_usterr();
1492         init_tracepoint();
1493         lttng_ust_baddr_statedump_init();
1494         lttng_ring_buffer_metadata_client_init();
1495         lttng_ring_buffer_client_overwrite_init();
1496         lttng_ring_buffer_client_overwrite_rt_init();
1497         lttng_ring_buffer_client_discard_init();
1498         lttng_ring_buffer_client_discard_rt_init();
1499         lttng_perf_counter_init();
1500         lttng_context_init();
1501         /*
1502          * Invoke ust malloc wrapper init before starting other threads.
1503          */
1504         lttng_ust_malloc_wrapper_init();
1505
1506         timeout_mode = get_constructor_timeout(&constructor_timeout);
1507
1508         ret = sem_init(&constructor_wait, 0, 0);
1509         assert(!ret);
1510
1511         ret = setup_local_apps();
1512         if (ret) {
1513                 DBG("local apps setup returned %d", ret);
1514         }
1515
1516         /* A new thread created by pthread_create inherits the signal mask
1517          * from the parent. To avoid any signal being received by the
1518          * listener thread, we block all signals temporarily in the parent,
1519          * while we create the listener thread.
1520          */
1521         sigfillset(&sig_all_blocked);
1522         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1523         if (ret) {
1524                 ERR("pthread_sigmask: %s", strerror(ret));
1525         }
1526
1527         ret = pthread_attr_init(&thread_attr);
1528         if (ret) {
1529                 ERR("pthread_attr_init: %s", strerror(ret));
1530         }
1531         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1532         if (ret) {
1533                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1534         }
1535
1536         pthread_mutex_lock(&ust_exit_mutex);
1537         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1538                         ust_listener_thread, &global_apps);
1539         if (ret) {
1540                 ERR("pthread_create global: %s", strerror(ret));
1541         }
1542         global_apps.thread_active = 1;
1543         pthread_mutex_unlock(&ust_exit_mutex);
1544
1545         if (local_apps.allowed) {
1546                 pthread_mutex_lock(&ust_exit_mutex);
1547                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1548                                 ust_listener_thread, &local_apps);
1549                 if (ret) {
1550                         ERR("pthread_create local: %s", strerror(ret));
1551                 }
1552                 local_apps.thread_active = 1;
1553                 pthread_mutex_unlock(&ust_exit_mutex);
1554         } else {
1555                 handle_register_done(&local_apps);
1556         }
1557         ret = pthread_attr_destroy(&thread_attr);
1558         if (ret) {
1559                 ERR("pthread_attr_destroy: %s", strerror(ret));
1560         }
1561
1562         /* Restore original signal mask in parent */
1563         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1564         if (ret) {
1565                 ERR("pthread_sigmask: %s", strerror(ret));
1566         }
1567
1568         switch (timeout_mode) {
1569         case 1: /* timeout wait */
1570                 do {
1571                         ret = sem_timedwait(&constructor_wait,
1572                                         &constructor_timeout);
1573                 } while (ret < 0 && errno == EINTR);
1574                 if (ret < 0 && errno == ETIMEDOUT) {
1575                         ERR("Timed out waiting for lttng-sessiond");
1576                 } else {
1577                         assert(!ret);
1578                 }
1579                 break;
1580         case -1:/* wait forever */
1581                 do {
1582                         ret = sem_wait(&constructor_wait);
1583                 } while (ret < 0 && errno == EINTR);
1584                 assert(!ret);
1585                 break;
1586         case 0: /* no timeout */
1587                 break;
1588         }
1589 }
1590
1591 static
1592 void lttng_ust_cleanup(int exiting)
1593 {
1594         cleanup_sock_info(&global_apps, exiting);
1595         cleanup_sock_info(&local_apps, exiting);
1596         /*
1597          * The teardown in this function all affect data structures
1598          * accessed under the UST lock by the listener thread. This
1599          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1600          * that none of these threads are accessing this data at this
1601          * point.
1602          */
1603         lttng_ust_abi_exit();
1604         lttng_ust_events_exit();
1605         lttng_context_exit();
1606         lttng_perf_counter_exit();
1607         lttng_ring_buffer_client_discard_rt_exit();
1608         lttng_ring_buffer_client_discard_exit();
1609         lttng_ring_buffer_client_overwrite_rt_exit();
1610         lttng_ring_buffer_client_overwrite_exit();
1611         lttng_ring_buffer_metadata_client_exit();
1612         lttng_ust_baddr_statedump_destroy();
1613         exit_tracepoint();
1614         if (!exiting) {
1615                 /* Reinitialize values for fork */
1616                 sem_count = 2;
1617                 lttng_ust_comm_should_quit = 0;
1618                 initialized = 0;
1619         }
1620 }
1621
1622 void __attribute__((destructor)) lttng_ust_exit(void)
1623 {
1624         int ret;
1625
1626         /*
1627          * Using pthread_cancel here because:
1628          * A) we don't want to hang application teardown.
1629          * B) the thread is not allocating any resource.
1630          */
1631
1632         /*
1633          * Require the communication thread to quit. Synchronize with
1634          * mutexes to ensure it is not in a mutex critical section when
1635          * pthread_cancel is later called.
1636          */
1637         ust_lock_nocheck();
1638         lttng_ust_comm_should_quit = 1;
1639         ust_unlock();
1640
1641         pthread_mutex_lock(&ust_exit_mutex);
1642         /* cancel threads */
1643         if (global_apps.thread_active) {
1644                 ret = pthread_cancel(global_apps.ust_listener);
1645                 if (ret) {
1646                         ERR("Error cancelling global ust listener thread: %s",
1647                                 strerror(ret));
1648                 } else {
1649                         global_apps.thread_active = 0;
1650                 }
1651         }
1652         if (local_apps.thread_active) {
1653                 ret = pthread_cancel(local_apps.ust_listener);
1654                 if (ret) {
1655                         ERR("Error cancelling local ust listener thread: %s",
1656                                 strerror(ret));
1657                 } else {
1658                         local_apps.thread_active = 0;
1659                 }
1660         }
1661         pthread_mutex_unlock(&ust_exit_mutex);
1662
1663         /*
1664          * Do NOT join threads: use of sys_futex makes it impossible to
1665          * join the threads without using async-cancel, but async-cancel
1666          * is delivered by a signal, which could hit the target thread
1667          * anywhere in its code path, including while the ust_lock() is
1668          * held, causing a deadlock for the other thread. Let the OS
1669          * cleanup the threads if there are stalled in a syscall.
1670          */
1671         lttng_ust_cleanup(1);
1672 }
1673
1674 /*
1675  * We exclude the worker threads across fork and clone (except
1676  * CLONE_VM), because these system calls only keep the forking thread
1677  * running in the child.  Therefore, we don't want to call fork or clone
1678  * in the middle of an tracepoint or ust tracing state modification.
1679  * Holding this mutex protects these structures across fork and clone.
1680  */
1681 void ust_before_fork(sigset_t *save_sigset)
1682 {
1683         /*
1684          * Disable signals. This is to avoid that the child intervenes
1685          * before it is properly setup for tracing. It is safer to
1686          * disable all signals, because then we know we are not breaking
1687          * anything by restoring the original mask.
1688          */
1689         sigset_t all_sigs;
1690         int ret;
1691
1692         if (URCU_TLS(lttng_ust_nest_count))
1693                 return;
1694         /* Disable signals */
1695         sigfillset(&all_sigs);
1696         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1697         if (ret == -1) {
1698                 PERROR("sigprocmask");
1699         }
1700
1701         pthread_mutex_lock(&ust_fork_mutex);
1702
1703         ust_lock_nocheck();
1704         rcu_bp_before_fork();
1705 }
1706
1707 static void ust_after_fork_common(sigset_t *restore_sigset)
1708 {
1709         int ret;
1710
1711         DBG("process %d", getpid());
1712         ust_unlock();
1713
1714         pthread_mutex_unlock(&ust_fork_mutex);
1715
1716         /* Restore signals */
1717         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1718         if (ret == -1) {
1719                 PERROR("sigprocmask");
1720         }
1721 }
1722
1723 void ust_after_fork_parent(sigset_t *restore_sigset)
1724 {
1725         if (URCU_TLS(lttng_ust_nest_count))
1726                 return;
1727         DBG("process %d", getpid());
1728         rcu_bp_after_fork_parent();
1729         /* Release mutexes and reenable signals */
1730         ust_after_fork_common(restore_sigset);
1731 }
1732
1733 /*
1734  * After fork, in the child, we need to cleanup all the leftover state,
1735  * except the worker thread which already magically disappeared thanks
1736  * to the weird Linux fork semantics. After tyding up, we call
1737  * lttng_ust_init() again to start over as a new PID.
1738  *
1739  * This is meant for forks() that have tracing in the child between the
1740  * fork and following exec call (if there is any).
1741  */
1742 void ust_after_fork_child(sigset_t *restore_sigset)
1743 {
1744         if (URCU_TLS(lttng_ust_nest_count))
1745                 return;
1746         DBG("process %d", getpid());
1747         /* Release urcu mutexes */
1748         rcu_bp_after_fork_child();
1749         lttng_ust_cleanup(0);
1750         lttng_context_vtid_reset();
1751         /* Release mutexes and reenable signals */
1752         ust_after_fork_common(restore_sigset);
1753         lttng_ust_init();
1754 }
1755
1756 void lttng_ust_sockinfo_session_enabled(void *owner)
1757 {
1758         struct sock_info *sock_info = owner;
1759         sock_info->statedump_pending = 1;
1760 }