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