liblttng-ust/lttng-ust-comm.c

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