liblttng-ust/lttng-ust-comm.c

   1 /*
   2  * lttng-ust-comm.c
   3  *
   4  * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
   5  * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   6  *
   7  * This library is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; only
  10  * version 2.1 of the License.
  11  *
  12  * This library is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with this library; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #define _LGPL_SOURCE
  23 #define _GNU_SOURCE
  24 #include <sys/types.h>
  25 #include <sys/socket.h>
  26 #include <sys/mman.h>
  27 #include <sys/stat.h>
  28 #include <sys/types.h>
  29 #include <sys/wait.h>
  30 #include <fcntl.h>
  31 #include <unistd.h>
  32 #include <errno.h>
  33 #include <pthread.h>
  34 #include <semaphore.h>
  35 #include <time.h>
  36 #include <assert.h>
  37 #include <signal.h>
  38 #include <urcu/uatomic.h>
  39 #include <urcu/futex.h>
  40 #include <urcu/compiler.h>
  41
  42 #include <lttng/ust-events.h>
  43 #include <lttng/ust-abi.h>
  44 #include <lttng/ust.h>
  45 #include <lttng/ust-error.h>
  46 #include <lttng/ust-ctl.h>
  47 #include <urcu/tls-compat.h>
  48 #include <ust-comm.h>
  49 #include <usterr-signal-safe.h>
  50 #include <helper.h>
  51 #include "tracepoint-internal.h"
  52 #include "lttng-tracer-core.h"
  53 #include "compat.h"
  54 #include "../libringbuffer/tlsfixup.h"
  55 #include "lttng-ust-statedump.h"
  56 #include "clock.h"
  57 #include "../libringbuffer/getcpu.h"
  58 #include "getenv.h"
  59
  60 /*
  61  * Has lttng ust comm constructor been called ?
  62  */
  63 static int initialized;
  64
  65 /*
  66  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  67  * Held when handling a command, also held by fork() to deal with
  68  * removal of threads, and by exit path.
  69  *
  70  * The UST lock is the centralized mutex across UST tracing control and
  71  * probe registration.
  72  *
  73  * ust_exit_mutex must never nest in ust_mutex.
  74  *
  75  * ust_fork_mutex must never nest in ust_mutex.
  76  *
  77  * ust_mutex_nest is a per-thread nesting counter, allowing the perf
  78  * counter lazy initialization called by events within the statedump,
  79  * which traces while the ust_mutex is held.
  80  *
  81  * ust_lock nests within the dynamic loader lock (within glibc) because
  82  * it is taken within the library constructor.
  83  */
  84 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
  85
  86 /* Allow nesting the ust_mutex within the same thread. */
  87 static DEFINE_URCU_TLS(int, ust_mutex_nest);
  88
  89 /*
  90  * ust_exit_mutex protects thread_active variable wrt thread exit. It
  91  * cannot be done by ust_mutex because pthread_cancel(), which takes an
  92  * internal libc lock, cannot nest within ust_mutex.
  93  *
  94  * It never nests within a ust_mutex.
  95  */
  96 static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
  97
  98 /*
  99  * ust_fork_mutex protects base address statedump tracing against forks. It
 100  * prevents the dynamic loader lock to be taken (by base address statedump
 101  * tracing) while a fork is happening, thus preventing deadlock issues with
 102  * the dynamic loader lock.
 103  */
 104 static pthread_mutex_t ust_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
 105
 106 /* Should the ust comm thread quit ? */
 107 static int lttng_ust_comm_should_quit;
 108
 109 /*
 110  * Return 0 on success, -1 if should quit.
 111  * The lock is taken in both cases.
 112  * Signal-safe.
 113  */
 114 int ust_lock(void)
 115 {
 116         sigset_t sig_all_blocked, orig_mask;
 117         int ret, oldstate;
 118
 119         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 120         if (ret) {
 121                 ERR("pthread_setcancelstate: %s", strerror(ret));
 122         }
 123         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 124                 ERR("pthread_setcancelstate: unexpected oldstate");
 125         }
 126         sigfillset(&sig_all_blocked);
 127         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 128         if (ret) {
 129                 ERR("pthread_sigmask: %s", strerror(ret));
 130         }
 131         if (!URCU_TLS(ust_mutex_nest)++)
 132                 pthread_mutex_lock(&ust_mutex);
 133         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 134         if (ret) {
 135                 ERR("pthread_sigmask: %s", strerror(ret));
 136         }
 137         if (lttng_ust_comm_should_quit) {
 138                 return -1;
 139         } else {
 140                 return 0;
 141         }
 142 }
 143
 144 /*
 145  * ust_lock_nocheck() can be used in constructors/destructors, because
 146  * they are already nested within the dynamic loader lock, and therefore
 147  * have exclusive access against execution of liblttng-ust destructor.
 148  * Signal-safe.
 149  */
 150 void ust_lock_nocheck(void)
 151 {
 152         sigset_t sig_all_blocked, orig_mask;
 153         int ret, oldstate;
 154
 155         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 156         if (ret) {
 157                 ERR("pthread_setcancelstate: %s", strerror(ret));
 158         }
 159         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 160                 ERR("pthread_setcancelstate: unexpected oldstate");
 161         }
 162         sigfillset(&sig_all_blocked);
 163         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 164         if (ret) {
 165                 ERR("pthread_sigmask: %s", strerror(ret));
 166         }
 167         if (!URCU_TLS(ust_mutex_nest)++)
 168                 pthread_mutex_lock(&ust_mutex);
 169         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 170         if (ret) {
 171                 ERR("pthread_sigmask: %s", strerror(ret));
 172         }
 173 }
 174
 175 /*
 176  * Signal-safe.
 177  */
 178 void ust_unlock(void)
 179 {
 180         sigset_t sig_all_blocked, orig_mask;
 181         int ret, oldstate;
 182
 183         sigfillset(&sig_all_blocked);
 184         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 185         if (ret) {
 186                 ERR("pthread_sigmask: %s", strerror(ret));
 187         }
 188         if (!--URCU_TLS(ust_mutex_nest))
 189                 pthread_mutex_unlock(&ust_mutex);
 190         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 191         if (ret) {
 192                 ERR("pthread_sigmask: %s", strerror(ret));
 193         }
 194         ret = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &oldstate);
 195         if (ret) {
 196                 ERR("pthread_setcancelstate: %s", strerror(ret));
 197         }
 198         if (oldstate != PTHREAD_CANCEL_DISABLE) {
 199                 ERR("pthread_setcancelstate: unexpected oldstate");
 200         }
 201 }
 202
 203 /*
 204  * Wait for either of these before continuing to the main
 205  * program:
 206  * - the register_done message from sessiond daemon
 207  *   (will let the sessiond daemon enable sessions before main
 208  *   starts.)
 209  * - sessiond daemon is not reachable.
 210  * - timeout (ensuring applications are resilient to session
 211  *   daemon problems).
 212  */
 213 static sem_t constructor_wait;
 214 /*
 215  * Doing this for both the global and local sessiond.
 216  */
 217 static int sem_count = { 2 };
 218
 219 /*
 220  * Counting nesting within lttng-ust. Used to ensure that calling fork()
 221  * from liblttng-ust does not execute the pre/post fork handlers.
 222  */
 223 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
 224
 225 /*
 226  * Info about socket and associated listener thread.
 227  */
 228 struct sock_info {
 229         const char *name;
 230         pthread_t ust_listener; /* listener thread */
 231         int root_handle;
 232         int constructor_sem_posted;
 233         int allowed;
 234         int global;
 235         int thread_active;
 236
 237         char sock_path[PATH_MAX];
 238         int socket;
 239         int notify_socket;
 240
 241         char wait_shm_path[PATH_MAX];
 242         char *wait_shm_mmap;
 243         /* Keep track of lazy state dump not performed yet. */
 244         int statedump_pending;
 245 };
 246
 247 /* Socket from app (connect) to session daemon (listen) for communication */
 248 struct sock_info global_apps = {
 249         .name = "global",
 250         .global = 1,
 251
 252         .root_handle = -1,
 253         .allowed = 1,
 254         .thread_active = 0,
 255
 256         .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
 257         .socket = -1,
 258         .notify_socket = -1,
 259
 260         .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 261
 262         .statedump_pending = 0,
 263 };
 264
 265 /* TODO: allow global_apps_sock_path override */
 266
 267 struct sock_info local_apps = {
 268         .name = "local",
 269         .global = 0,
 270         .root_handle = -1,
 271         .allowed = 0,   /* Check setuid bit first */
 272         .thread_active = 0,
 273
 274         .socket = -1,
 275         .notify_socket = -1,
 276
 277         .statedump_pending = 0,
 278 };
 279
 280 static int wait_poll_fallback;
 281
 282 static const char *cmd_name_mapping[] = {
 283         [ LTTNG_UST_RELEASE ] = "Release",
 284         [ LTTNG_UST_SESSION ] = "Create Session",
 285         [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
 286
 287         [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
 288         [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
 289         [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
 290         [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
 291
 292         /* Session FD commands */
 293         [ LTTNG_UST_CHANNEL ] = "Create Channel",
 294         [ LTTNG_UST_SESSION_START ] = "Start Session",
 295         [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
 296
 297         /* Channel FD commands */
 298         [ LTTNG_UST_STREAM ] = "Create Stream",
 299         [ LTTNG_UST_EVENT ] = "Create Event",
 300
 301         /* Event and Channel FD commands */
 302         [ LTTNG_UST_CONTEXT ] = "Create Context",
 303         [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
 304
 305         /* Event, Channel and Session commands */
 306         [ LTTNG_UST_ENABLE ] = "Enable",
 307         [ LTTNG_UST_DISABLE ] = "Disable",
 308
 309         /* Tracepoint list commands */
 310         [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
 311         [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
 312
 313         /* Event FD commands */
 314         [ LTTNG_UST_FILTER ] = "Create Filter",
 315         [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
 316 };
 317
 318 static const char *str_timeout;
 319 static int got_timeout_env;
 320
 321 extern void lttng_ring_buffer_client_overwrite_init(void);
 322 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 323 extern void lttng_ring_buffer_client_discard_init(void);
 324 extern void lttng_ring_buffer_client_discard_rt_init(void);
 325 extern void lttng_ring_buffer_metadata_client_init(void);
 326 extern void lttng_ring_buffer_client_overwrite_exit(void);
 327 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 328 extern void lttng_ring_buffer_client_discard_exit(void);
 329 extern void lttng_ring_buffer_client_discard_rt_exit(void);
 330 extern void lttng_ring_buffer_metadata_client_exit(void);
 331
 332 /*
 333  * Returns the HOME directory path. Caller MUST NOT free(3) the returned
 334  * pointer.
 335  */
 336 static
 337 const char *get_lttng_home_dir(void)
 338 {
 339        const char *val;
 340
 341        val = (const char *) lttng_secure_getenv("LTTNG_HOME");
 342        if (val != NULL) {
 343                return val;
 344        }
 345        return (const char *) lttng_secure_getenv("HOME");
 346 }
 347
 348 /*
 349  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 350  */
 351 static
 352 void lttng_fixup_nest_count_tls(void)
 353 {
 354         asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 355 }
 356
 357 static
 358 void lttng_fixup_ust_mutex_nest_tls(void)
 359 {
 360         asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest)));
 361 }
 362
 363 /*
 364  * Fixup urcu bp TLS.
 365  */
 366 static
 367 void lttng_fixup_urcu_bp_tls(void)
 368 {
 369         rcu_read_lock();
 370         rcu_read_unlock();
 371 }
 372
 373 int lttng_get_notify_socket(void *owner)
 374 {
 375         struct sock_info *info = owner;
 376
 377         return info->notify_socket;
 378 }
 379
 380 static
 381 void print_cmd(int cmd, int handle)
 382 {
 383         const char *cmd_name = "Unknown";
 384
 385         if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
 386                         && cmd_name_mapping[cmd]) {
 387                 cmd_name = cmd_name_mapping[cmd];
 388         }
 389         DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
 390                 cmd_name, cmd,
 391                 lttng_ust_obj_get_name(handle), handle);
 392 }
 393
 394 static
 395 int setup_local_apps(void)
 396 {
 397         const char *home_dir;
 398         uid_t uid;
 399
 400         uid = getuid();
 401         /*
 402          * Disallow per-user tracing for setuid binaries.
 403          */
 404         if (uid != geteuid()) {
 405                 assert(local_apps.allowed == 0);
 406                 return 0;
 407         }
 408         home_dir = get_lttng_home_dir();
 409         if (!home_dir) {
 410                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 411                 assert(local_apps.allowed == 0);
 412                 return -ENOENT;
 413         }
 414         local_apps.allowed = 1;
 415         snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
 416                 home_dir,
 417                 LTTNG_DEFAULT_HOME_RUNDIR,
 418                 LTTNG_UST_SOCK_FILENAME);
 419         snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
 420                 LTTNG_UST_WAIT_FILENAME,
 421                 uid);
 422         return 0;
 423 }
 424
 425 /*
 426  * Get notify_sock timeout, in ms.
 427  * -1: wait forever. 0: don't wait. >0: timeout, in ms.
 428  */
 429 static
 430 long get_timeout(void)
 431 {
 432         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 433
 434         if (!got_timeout_env) {
 435                 str_timeout = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 436                 got_timeout_env = 1;
 437         }
 438         if (str_timeout)
 439                 constructor_delay_ms = strtol(str_timeout, NULL, 10);
 440         return constructor_delay_ms;
 441 }
 442
 443 static
 444 long get_notify_sock_timeout(void)
 445 {
 446         return get_timeout();
 447 }
 448
 449 /*
 450  * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
 451  */
 452 static
 453 int get_constructor_timeout(struct timespec *constructor_timeout)
 454 {
 455         long constructor_delay_ms;
 456         int ret;
 457
 458         constructor_delay_ms = get_timeout();
 459
 460         switch (constructor_delay_ms) {
 461         case -1:/* fall-through */
 462         case 0:
 463                 return constructor_delay_ms;
 464         default:
 465                 break;
 466         }
 467
 468         /*
 469          * If we are unable to find the current time, don't wait.
 470          */
 471         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 472         if (ret) {
 473                 /* Don't wait. */
 474                 return 0;
 475         }
 476         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 477         constructor_timeout->tv_nsec +=
 478                 (constructor_delay_ms % 1000UL) * 1000000UL;
 479         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 480                 constructor_timeout->tv_sec++;
 481                 constructor_timeout->tv_nsec -= 1000000000UL;
 482         }
 483         /* Timeout wait (constructor_delay_ms). */
 484         return 1;
 485 }
 486
 487 static
 488 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 489 {
 490         return ustcomm_send_reg_msg(socket,
 491                 type,
 492                 CAA_BITS_PER_LONG,
 493                 lttng_alignof(uint8_t) * CHAR_BIT,
 494                 lttng_alignof(uint16_t) * CHAR_BIT,
 495                 lttng_alignof(uint32_t) * CHAR_BIT,
 496                 lttng_alignof(uint64_t) * CHAR_BIT,
 497                 lttng_alignof(unsigned long) * CHAR_BIT);
 498 }
 499
 500 static
 501 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 502 {
 503         ssize_t len;
 504
 505         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 506         switch (len) {
 507         case sizeof(*lur):
 508                 DBG("message successfully sent");
 509                 return 0;
 510         default:
 511                 if (len == -ECONNRESET) {
 512                         DBG("remote end closed connection");
 513                         return 0;
 514                 }
 515                 if (len < 0)
 516                         return len;
 517                 DBG("incorrect message size: %zd", len);
 518                 return -EINVAL;
 519         }
 520 }
 521
 522 static
 523 int handle_register_done(struct sock_info *sock_info)
 524 {
 525         int ret;
 526
 527         if (sock_info->constructor_sem_posted)
 528                 return 0;
 529         sock_info->constructor_sem_posted = 1;
 530         if (uatomic_read(&sem_count) <= 0) {
 531                 return 0;
 532         }
 533         ret = uatomic_add_return(&sem_count, -1);
 534         if (ret == 0) {
 535                 ret = sem_post(&constructor_wait);
 536                 assert(!ret);
 537         }
 538         return 0;
 539 }
 540
 541 /*
 542  * Only execute pending statedump after the constructor semaphore has
 543  * been posted by each listener thread. This means statedump will only
 544  * be performed after the "registration done" command is received from
 545  * each session daemon the application is connected to.
 546  *
 547  * This ensures we don't run into deadlock issues with the dynamic
 548  * loader mutex, which is held while the constructor is called and
 549  * waiting on the constructor semaphore. All operations requiring this
 550  * dynamic loader lock need to be postponed using this mechanism.
 551  */
 552 static
 553 void handle_pending_statedump(struct sock_info *sock_info)
 554 {
 555         int ctor_passed = sock_info->constructor_sem_posted;
 556
 557         if (ctor_passed && sock_info->statedump_pending) {
 558                 sock_info->statedump_pending = 0;
 559                 pthread_mutex_lock(&ust_fork_mutex);
 560                 lttng_handle_pending_statedump(sock_info);
 561                 pthread_mutex_unlock(&ust_fork_mutex);
 562         }
 563 }
 564
 565 static
 566 int handle_message(struct sock_info *sock_info,
 567                 int sock, struct ustcomm_ust_msg *lum)
 568 {
 569         int ret = 0;
 570         const struct lttng_ust_objd_ops *ops;
 571         struct ustcomm_ust_reply lur;
 572         union ust_args args;
 573         ssize_t len;
 574
 575         memset(&lur, 0, sizeof(lur));
 576
 577         if (ust_lock()) {
 578                 ret = -LTTNG_UST_ERR_EXITING;
 579                 goto error;
 580         }
 581
 582         ops = objd_ops(lum->handle);
 583         if (!ops) {
 584                 ret = -ENOENT;
 585                 goto error;
 586         }
 587
 588         switch (lum->cmd) {
 589         case LTTNG_UST_REGISTER_DONE:
 590                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 591                         ret = handle_register_done(sock_info);
 592                 else
 593                         ret = -EINVAL;
 594                 break;
 595         case LTTNG_UST_RELEASE:
 596                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 597                         ret = -EPERM;
 598                 else
 599                         ret = lttng_ust_objd_unref(lum->handle, 1);
 600                 break;
 601         case LTTNG_UST_FILTER:
 602         {
 603                 /* Receive filter data */
 604                 struct lttng_ust_filter_bytecode_node *bytecode;
 605
 606                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 607                         ERR("Filter data size is too large: %u bytes",
 608                                 lum->u.filter.data_size);
 609                         ret = -EINVAL;
 610                         goto error;
 611                 }
 612
 613                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 614                         ERR("Filter reloc offset %u is not within data",
 615                                 lum->u.filter.reloc_offset);
 616                         ret = -EINVAL;
 617                         goto error;
 618                 }
 619
 620                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 621                 if (!bytecode) {
 622                         ret = -ENOMEM;
 623                         goto error;
 624                 }
 625                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 626                                 lum->u.filter.data_size);
 627                 switch (len) {
 628                 case 0: /* orderly shutdown */
 629                         ret = 0;
 630                         free(bytecode);
 631                         goto error;
 632                 default:
 633                         if (len == lum->u.filter.data_size) {
 634                                 DBG("filter data received");
 635                                 break;
 636                         } else if (len < 0) {
 637                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 638                                 if (len == -ECONNRESET) {
 639                                         ERR("%s remote end closed connection", sock_info->name);
 640                                         ret = len;
 641                                         free(bytecode);
 642                                         goto error;
 643                                 }
 644                                 ret = len;
 645                                 free(bytecode);
 646                                 goto error;
 647                         } else {
 648                                 DBG("incorrect filter data message size: %zd", len);
 649                                 ret = -EINVAL;
 650                                 free(bytecode);
 651                                 goto error;
 652                         }
 653                 }
 654                 bytecode->bc.len = lum->u.filter.data_size;
 655                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 656                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 657                 if (ops->cmd) {
 658                         ret = ops->cmd(lum->handle, lum->cmd,
 659                                         (unsigned long) bytecode,
 660                                         &args, sock_info);
 661                         if (ret) {
 662                                 free(bytecode);
 663                         }
 664                         /* don't free bytecode if everything went fine. */
 665                 } else {
 666                         ret = -ENOSYS;
 667                         free(bytecode);
 668                 }
 669                 break;
 670         }
 671         case LTTNG_UST_EXCLUSION:
 672         {
 673                 /* Receive exclusion names */
 674                 struct lttng_ust_excluder_node *node;
 675                 unsigned int count;
 676
 677                 count = lum->u.exclusion.count;
 678                 if (count == 0) {
 679                         /* There are no names to read */
 680                         ret = 0;
 681                         goto error;
 682                 }
 683                 node = zmalloc(sizeof(*node) +
 684                                 count * LTTNG_UST_SYM_NAME_LEN);
 685                 if (!node) {
 686                         ret = -ENOMEM;
 687                         goto error;
 688                 }
 689                 node->excluder.count = count;
 690                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 691                                 count * LTTNG_UST_SYM_NAME_LEN);
 692                 switch (len) {
 693                 case 0: /* orderly shutdown */
 694                         ret = 0;
 695                         free(node);
 696                         goto error;
 697                 default:
 698                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 699                                 DBG("Exclusion data received");
 700                                 break;
 701                         } else if (len < 0) {
 702                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 703                                 if (len == -ECONNRESET) {
 704                                         ERR("%s remote end closed connection", sock_info->name);
 705                                         ret = len;
 706                                         free(node);
 707                                         goto error;
 708                                 }
 709                                 ret = len;
 710                                 free(node);
 711                                 goto error;
 712                         } else {
 713                                 DBG("Incorrect exclusion data message size: %zd", len);
 714                                 ret = -EINVAL;
 715                                 free(node);
 716                                 goto error;
 717                         }
 718                 }
 719                 if (ops->cmd) {
 720                         ret = ops->cmd(lum->handle, lum->cmd,
 721                                         (unsigned long) node,
 722                                         &args, sock_info);
 723                         if (ret) {
 724                                 free(node);
 725                         }
 726                         /* Don't free exclusion data if everything went fine. */
 727                 } else {
 728                         ret = -ENOSYS;
 729                         free(node);
 730                 }
 731                 break;
 732         }
 733         case LTTNG_UST_CHANNEL:
 734         {
 735                 void *chan_data;
 736                 int wakeup_fd;
 737
 738                 len = ustcomm_recv_channel_from_sessiond(sock,
 739                                 &chan_data, lum->u.channel.len,
 740                                 &wakeup_fd);
 741                 switch (len) {
 742                 case 0: /* orderly shutdown */
 743                         ret = 0;
 744                         goto error;
 745                 default:
 746                         if (len == lum->u.channel.len) {
 747                                 DBG("channel data received");
 748                                 break;
 749                         } else if (len < 0) {
 750                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 751                                 if (len == -ECONNRESET) {
 752                                         ERR("%s remote end closed connection", sock_info->name);
 753                                         ret = len;
 754                                         goto error;
 755                                 }
 756                                 ret = len;
 757                                 goto error;
 758                         } else {
 759                                 DBG("incorrect channel data message size: %zd", len);
 760                                 ret = -EINVAL;
 761                                 goto error;
 762                         }
 763                 }
 764                 args.channel.chan_data = chan_data;
 765                 args.channel.wakeup_fd = wakeup_fd;
 766                 if (ops->cmd)
 767                         ret = ops->cmd(lum->handle, lum->cmd,
 768                                         (unsigned long) &lum->u,
 769                                         &args, sock_info);
 770                 else
 771                         ret = -ENOSYS;
 772                 break;
 773         }
 774         case LTTNG_UST_STREAM:
 775         {
 776                 /* Receive shm_fd, wakeup_fd */
 777                 ret = ustcomm_recv_stream_from_sessiond(sock,
 778                         &lum->u.stream.len,
 779                         &args.stream.shm_fd,
 780                         &args.stream.wakeup_fd);
 781                 if (ret) {
 782                         goto error;
 783                 }
 784                 if (ops->cmd)
 785                         ret = ops->cmd(lum->handle, lum->cmd,
 786                                         (unsigned long) &lum->u,
 787                                         &args, sock_info);
 788                 else
 789                         ret = -ENOSYS;
 790                 break;
 791         }
 792         default:
 793                 if (ops->cmd)
 794                         ret = ops->cmd(lum->handle, lum->cmd,
 795                                         (unsigned long) &lum->u,
 796                                         &args, sock_info);
 797                 else
 798                         ret = -ENOSYS;
 799                 break;
 800         }
 801
 802         lur.handle = lum->handle;
 803         lur.cmd = lum->cmd;
 804         lur.ret_val = ret;
 805         if (ret >= 0) {
 806                 lur.ret_code = LTTNG_UST_OK;
 807         } else {
 808                 /*
 809                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 810                  * error that are not caused by the transport, except if
 811                  * we already have a more precise error message to
 812                  * report.
 813                  */
 814                 if (ret > -LTTNG_UST_ERR) {
 815                         /* Translate code to UST error. */
 816                         switch (ret) {
 817                         case -EEXIST:
 818                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 819                                 break;
 820                         case -EINVAL:
 821                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 822                                 break;
 823                         case -ENOENT:
 824                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 825                                 break;
 826                         case -EPERM:
 827                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 828                                 break;
 829                         case -ENOSYS:
 830                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 831                                 break;
 832                         default:
 833                                 lur.ret_code = -LTTNG_UST_ERR;
 834                                 break;
 835                         }
 836                 } else {
 837                         lur.ret_code = ret;
 838                 }
 839         }
 840         if (ret >= 0) {
 841                 switch (lum->cmd) {
 842                 case LTTNG_UST_TRACER_VERSION:
 843                         lur.u.version = lum->u.version;
 844                         break;
 845                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 846                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 847                         break;
 848                 }
 849         }
 850         DBG("Return value: %d", lur.ret_val);
 851         ret = send_reply(sock, &lur);
 852         if (ret < 0) {
 853                 DBG("error sending reply");
 854                 goto error;
 855         }
 856
 857         /*
 858          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 859          * after the reply.
 860          */
 861         if (lur.ret_code == LTTNG_UST_OK) {
 862                 switch (lum->cmd) {
 863                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 864                         len = ustcomm_send_unix_sock(sock,
 865                                 &args.field_list.entry,
 866                                 sizeof(args.field_list.entry));
 867                         if (len < 0) {
 868                                 ret = len;
 869                                 goto error;
 870                         }
 871                         if (len != sizeof(args.field_list.entry)) {
 872                                 ret = -EINVAL;
 873                                 goto error;
 874                         }
 875                 }
 876         }
 877
 878 error:
 879         ust_unlock();
 880
 881         /*
 882          * Performed delayed statedump operations outside of the UST
 883          * lock. We need to take the dynamic loader lock before we take
 884          * the UST lock internally within handle_pending_statedump().
 885           */
 886         handle_pending_statedump(sock_info);
 887
 888         return ret;
 889 }
 890
 891 static
 892 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 893 {
 894         int ret;
 895
 896         if (sock_info->root_handle != -1) {
 897                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 898                 if (ret) {
 899                         ERR("Error unref root handle");
 900                 }
 901                 sock_info->root_handle = -1;
 902         }
 903         sock_info->constructor_sem_posted = 0;
 904
 905         /*
 906          * wait_shm_mmap, socket and notify socket are used by listener
 907          * threads outside of the ust lock, so we cannot tear them down
 908          * ourselves, because we cannot join on these threads. Leave
 909          * responsibility of cleaning up these resources to the OS
 910          * process exit.
 911          */
 912         if (exiting)
 913                 return;
 914
 915         if (sock_info->socket != -1) {
 916                 ret = ustcomm_close_unix_sock(sock_info->socket);
 917                 if (ret) {
 918                         ERR("Error closing ust cmd socket");
 919                 }
 920                 sock_info->socket = -1;
 921         }
 922         if (sock_info->notify_socket != -1) {
 923                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 924                 if (ret) {
 925                         ERR("Error closing ust notify socket");
 926                 }
 927                 sock_info->notify_socket = -1;
 928         }
 929         if (sock_info->wait_shm_mmap) {
 930                 long page_size;
 931
 932                 page_size = sysconf(_SC_PAGE_SIZE);
 933                 if (page_size <= 0) {
 934                         if (!page_size) {
 935                                 errno = EINVAL;
 936                         }
 937                         PERROR("Error in sysconf(_SC_PAGE_SIZE)");
 938                 } else {
 939                         ret = munmap(sock_info->wait_shm_mmap, page_size);
 940                         if (ret) {
 941                                 ERR("Error unmapping wait shm");
 942                         }
 943                 }
 944                 sock_info->wait_shm_mmap = NULL;
 945         }
 946 }
 947
 948 /*
 949  * Using fork to set umask in the child process (not multi-thread safe).
 950  * We deal with the shm_open vs ftruncate race (happening when the
 951  * sessiond owns the shm and does not let everybody modify it, to ensure
 952  * safety against shm_unlink) by simply letting the mmap fail and
 953  * retrying after a few seconds.
 954  * For global shm, everybody has rw access to it until the sessiond
 955  * starts.
 956  */
 957 static
 958 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 959 {
 960         int wait_shm_fd, ret;
 961         pid_t pid;
 962
 963         /*
 964          * Try to open read-only.
 965          */
 966         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 967         if (wait_shm_fd >= 0) {
 968                 int32_t tmp_read;
 969                 ssize_t len;
 970                 size_t bytes_read = 0;
 971
 972                 /*
 973                  * Try to read the fd. If unable to do so, try opening
 974                  * it in write mode.
 975                  */
 976                 do {
 977                         len = read(wait_shm_fd,
 978                                 &((char *) &tmp_read)[bytes_read],
 979                                 sizeof(tmp_read) - bytes_read);
 980                         if (len > 0) {
 981                                 bytes_read += len;
 982                         }
 983                 } while ((len < 0 && errno == EINTR)
 984                         || (len > 0 && bytes_read < sizeof(tmp_read)));
 985                 if (bytes_read != sizeof(tmp_read)) {
 986                         ret = close(wait_shm_fd);
 987                         if (ret) {
 988                                 ERR("close wait_shm_fd");
 989                         }
 990                         goto open_write;
 991                 }
 992                 goto end;
 993         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 994                 /*
 995                  * Real-only open did not work, and it's not because the
 996                  * entry was not present. It's a failure that prohibits
 997                  * using shm.
 998                  */
 999                 ERR("Error opening shm %s", sock_info->wait_shm_path);
1000                 goto end;
1001         }
1002
1003 open_write:
1004         /*
1005          * If the open failed because the file did not exist, or because
1006          * the file was not truncated yet, try creating it ourself.
1007          */
1008         URCU_TLS(lttng_ust_nest_count)++;
1009         pid = fork();
1010         URCU_TLS(lttng_ust_nest_count)--;
1011         if (pid > 0) {
1012                 int status;
1013
1014                 /*
1015                  * Parent: wait for child to return, in which case the
1016                  * shared memory map will have been created.
1017                  */
1018                 pid = wait(&status);
1019                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
1020                         wait_shm_fd = -1;
1021                         goto end;
1022                 }
1023                 /*
1024                  * Try to open read-only again after creation.
1025                  */
1026                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1027                 if (wait_shm_fd < 0) {
1028                         /*
1029                          * Real-only open did not work. It's a failure
1030                          * that prohibits using shm.
1031                          */
1032                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1033                         goto end;
1034                 }
1035                 goto end;
1036         } else if (pid == 0) {
1037                 int create_mode;
1038
1039                 /* Child */
1040                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1041                 if (sock_info->global)
1042                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1043                 /*
1044                  * We're alone in a child process, so we can modify the
1045                  * process-wide umask.
1046                  */
1047                 umask(~create_mode);
1048                 /*
1049                  * Try creating shm (or get rw access).
1050                  * We don't do an exclusive open, because we allow other
1051                  * processes to create+ftruncate it concurrently.
1052                  */
1053                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1054                                 O_RDWR | O_CREAT, create_mode);
1055                 if (wait_shm_fd >= 0) {
1056                         ret = ftruncate(wait_shm_fd, mmap_size);
1057                         if (ret) {
1058                                 PERROR("ftruncate");
1059                                 _exit(EXIT_FAILURE);
1060                         }
1061                         _exit(EXIT_SUCCESS);
1062                 }
1063                 /*
1064                  * For local shm, we need to have rw access to accept
1065                  * opening it: this means the local sessiond will be
1066                  * able to wake us up. For global shm, we open it even
1067                  * if rw access is not granted, because the root.root
1068                  * sessiond will be able to override all rights and wake
1069                  * us up.
1070                  */
1071                 if (!sock_info->global && errno != EACCES) {
1072                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1073                         _exit(EXIT_FAILURE);
1074                 }
1075                 /*
1076                  * The shm exists, but we cannot open it RW. Report
1077                  * success.
1078                  */
1079                 _exit(EXIT_SUCCESS);
1080         } else {
1081                 return -1;
1082         }
1083 end:
1084         if (wait_shm_fd >= 0 && !sock_info->global) {
1085                 struct stat statbuf;
1086
1087                 /*
1088                  * Ensure that our user is the owner of the shm file for
1089                  * local shm. If we do not own the file, it means our
1090                  * sessiond will not have access to wake us up (there is
1091                  * probably a rogue process trying to fake our
1092                  * sessiond). Fallback to polling method in this case.
1093                  */
1094                 ret = fstat(wait_shm_fd, &statbuf);
1095                 if (ret) {
1096                         PERROR("fstat");
1097                         goto error_close;
1098                 }
1099                 if (statbuf.st_uid != getuid())
1100                         goto error_close;
1101         }
1102         return wait_shm_fd;
1103
1104 error_close:
1105         ret = close(wait_shm_fd);
1106         if (ret) {
1107                 PERROR("Error closing fd");
1108         }
1109         return -1;
1110 }
1111
1112 static
1113 char *get_map_shm(struct sock_info *sock_info)
1114 {
1115         long page_size;
1116         int wait_shm_fd, ret;
1117         char *wait_shm_mmap;
1118
1119         page_size = sysconf(_SC_PAGE_SIZE);
1120         if (page_size <= 0) {
1121                 if (!page_size) {
1122                         errno = EINVAL;
1123                 }
1124                 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1125                 goto error;
1126         }
1127
1128         wait_shm_fd = get_wait_shm(sock_info, page_size);
1129         if (wait_shm_fd < 0) {
1130                 goto error;
1131         }
1132         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1133                   MAP_SHARED, wait_shm_fd, 0);
1134         /* close shm fd immediately after taking the mmap reference */
1135         ret = close(wait_shm_fd);
1136         if (ret) {
1137                 PERROR("Error closing fd");
1138         }
1139         if (wait_shm_mmap == MAP_FAILED) {
1140                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1141                 goto error;
1142         }
1143         return wait_shm_mmap;
1144
1145 error:
1146         return NULL;
1147 }
1148
1149 static
1150 void wait_for_sessiond(struct sock_info *sock_info)
1151 {
1152         if (ust_lock()) {
1153                 goto quit;
1154         }
1155         if (wait_poll_fallback) {
1156                 goto error;
1157         }
1158         if (!sock_info->wait_shm_mmap) {
1159                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1160                 if (!sock_info->wait_shm_mmap)
1161                         goto error;
1162         }
1163         ust_unlock();
1164
1165         DBG("Waiting for %s apps sessiond", sock_info->name);
1166         /* Wait for futex wakeup */
1167         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
1168                 goto end_wait;
1169
1170         while (futex_async((int32_t *) sock_info->wait_shm_mmap,
1171                         FUTEX_WAIT, 0, NULL, NULL, 0)) {
1172                 switch (errno) {
1173                 case EWOULDBLOCK:
1174                         /* Value already changed. */
1175                         goto end_wait;
1176                 case EINTR:
1177                         /* Retry if interrupted by signal. */
1178                         break;  /* Get out of switch. */
1179                 case EFAULT:
1180                         wait_poll_fallback = 1;
1181                         DBG(
1182 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1183 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1184 "Please upgrade your kernel "
1185 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1186 "mainline). LTTng-UST will use polling mode fallback.");
1187                         if (ust_debug())
1188                                 PERROR("futex");
1189                         goto end_wait;
1190                 }
1191         }
1192 end_wait:
1193         return;
1194
1195 quit:
1196         ust_unlock();
1197         return;
1198
1199 error:
1200         ust_unlock();
1201         return;
1202 }
1203
1204 /*
1205  * This thread does not allocate any resource, except within
1206  * handle_message, within mutex protection. This mutex protects against
1207  * fork and exit.
1208  * The other moment it allocates resources is at socket connection, which
1209  * is also protected by the mutex.
1210  */
1211 static
1212 void *ust_listener_thread(void *arg)
1213 {
1214         struct sock_info *sock_info = arg;
1215         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1216         long timeout;
1217
1218         /* Restart trying to connect to the session daemon */
1219 restart:
1220         if (prev_connect_failed) {
1221                 /* Wait for sessiond availability with pipe */
1222                 wait_for_sessiond(sock_info);
1223                 if (has_waited) {
1224                         has_waited = 0;
1225                         /*
1226                          * Sleep for 5 seconds before retrying after a
1227                          * sequence of failure / wait / failure. This
1228                          * deals with a killed or broken session daemon.
1229                          */
1230                         sleep(5);
1231                 } else {
1232                         has_waited = 1;
1233                 }
1234                 prev_connect_failed = 0;
1235         }
1236
1237         if (sock_info->socket != -1) {
1238                 ret = ustcomm_close_unix_sock(sock_info->socket);
1239                 if (ret) {
1240                         ERR("Error closing %s ust cmd socket",
1241                                 sock_info->name);
1242                 }
1243                 sock_info->socket = -1;
1244         }
1245         if (sock_info->notify_socket != -1) {
1246                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1247                 if (ret) {
1248                         ERR("Error closing %s ust notify socket",
1249                                 sock_info->name);
1250                 }
1251                 sock_info->notify_socket = -1;
1252         }
1253
1254         /*
1255          * Register. We need to perform both connect and sending
1256          * registration message before doing the next connect otherwise
1257          * we may reach unix socket connect queue max limits and block
1258          * on the 2nd connect while the session daemon is awaiting the
1259          * first connect registration message.
1260          */
1261         /* Connect cmd socket */
1262         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1263         if (ret < 0) {
1264                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1265                 prev_connect_failed = 1;
1266
1267                 if (ust_lock()) {
1268                         goto quit;
1269                 }
1270
1271                 /*
1272                  * If we cannot find the sessiond daemon, don't delay
1273                  * constructor execution.
1274                  */
1275                 ret = handle_register_done(sock_info);
1276                 assert(!ret);
1277                 ust_unlock();
1278                 goto restart;
1279         }
1280         sock_info->socket = ret;
1281
1282         if (ust_lock()) {
1283                 goto quit;
1284         }
1285
1286         /*
1287          * Create only one root handle per listener thread for the whole
1288          * process lifetime, so we ensure we get ID which is statically
1289          * assigned to the root handle.
1290          */
1291         if (sock_info->root_handle == -1) {
1292                 ret = lttng_abi_create_root_handle();
1293                 if (ret < 0) {
1294                         ERR("Error creating root handle");
1295                         goto quit;
1296                 }
1297                 sock_info->root_handle = ret;
1298         }
1299
1300         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1301         if (ret < 0) {
1302                 ERR("Error registering to %s ust cmd socket",
1303                         sock_info->name);
1304                 prev_connect_failed = 1;
1305                 /*
1306                  * If we cannot register to the sessiond daemon, don't
1307                  * delay constructor execution.
1308                  */
1309                 ret = handle_register_done(sock_info);
1310                 assert(!ret);
1311                 ust_unlock();
1312                 goto restart;
1313         }
1314
1315         ust_unlock();
1316
1317         /* Connect notify socket */
1318         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1319         if (ret < 0) {
1320                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1321                 prev_connect_failed = 1;
1322
1323                 if (ust_lock()) {
1324                         goto quit;
1325                 }
1326
1327                 /*
1328                  * If we cannot find the sessiond daemon, don't delay
1329                  * constructor execution.
1330                  */
1331                 ret = handle_register_done(sock_info);
1332                 assert(!ret);
1333                 ust_unlock();
1334                 goto restart;
1335         }
1336         sock_info->notify_socket = ret;
1337
1338         timeout = get_notify_sock_timeout();
1339         if (timeout >= 0) {
1340                 /*
1341                  * Give at least 10ms to sessiond to reply to
1342                  * notifications.
1343                  */
1344                 if (timeout < 10)
1345                         timeout = 10;
1346                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1347                                 timeout);
1348                 if (ret < 0) {
1349                         WARN("Error setting socket receive timeout");
1350                 }
1351                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1352                                 timeout);
1353                 if (ret < 0) {
1354                         WARN("Error setting socket send timeout");
1355                 }
1356         } else if (timeout < -1) {
1357                 WARN("Unsupported timeout value %ld", timeout);
1358         }
1359
1360         if (ust_lock()) {
1361                 goto quit;
1362         }
1363
1364         ret = register_to_sessiond(sock_info->notify_socket,
1365                         USTCTL_SOCKET_NOTIFY);
1366         if (ret < 0) {
1367                 ERR("Error registering to %s ust notify socket",
1368                         sock_info->name);
1369                 prev_connect_failed = 1;
1370                 /*
1371                  * If we cannot register to the sessiond daemon, don't
1372                  * delay constructor execution.
1373                  */
1374                 ret = handle_register_done(sock_info);
1375                 assert(!ret);
1376                 ust_unlock();
1377                 goto restart;
1378         }
1379         sock = sock_info->socket;
1380
1381         ust_unlock();
1382
1383         for (;;) {
1384                 ssize_t len;
1385                 struct ustcomm_ust_msg lum;
1386
1387                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1388                 switch (len) {
1389                 case 0: /* orderly shutdown */
1390                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1391                         if (ust_lock()) {
1392                                 goto quit;
1393                         }
1394                         /*
1395                          * Either sessiond has shutdown or refused us by closing the socket.
1396                          * In either case, we don't want to delay construction execution,
1397                          * and we need to wait before retry.
1398                          */
1399                         prev_connect_failed = 1;
1400                         /*
1401                          * If we cannot register to the sessiond daemon, don't
1402                          * delay constructor execution.
1403                          */
1404                         ret = handle_register_done(sock_info);
1405                         assert(!ret);
1406                         ust_unlock();
1407                         goto end;
1408                 case sizeof(lum):
1409                         print_cmd(lum.cmd, lum.handle);
1410                         ret = handle_message(sock_info, sock, &lum);
1411                         if (ret) {
1412                                 ERR("Error handling message for %s socket",
1413                                         sock_info->name);
1414                                 /*
1415                                  * Close socket if protocol error is
1416                                  * detected.
1417                                  */
1418                                 goto end;
1419                         }
1420                         continue;
1421                 default:
1422                         if (len < 0) {
1423                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1424                         } else {
1425                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1426                         }
1427                         if (len == -ECONNRESET) {
1428                                 DBG("%s remote end closed connection", sock_info->name);
1429                                 goto end;
1430                         }
1431                         goto end;
1432                 }
1433
1434         }
1435 end:
1436         if (ust_lock()) {
1437                 goto quit;
1438         }
1439         /* Cleanup socket handles before trying to reconnect */
1440         lttng_ust_objd_table_owner_cleanup(sock_info);
1441         ust_unlock();
1442         goto restart;   /* try to reconnect */
1443
1444 quit:
1445         ust_unlock();
1446
1447         pthread_mutex_lock(&ust_exit_mutex);
1448         sock_info->thread_active = 0;
1449         pthread_mutex_unlock(&ust_exit_mutex);
1450         return NULL;
1451 }
1452
1453 /*
1454  * Weak symbol to call when the ust malloc wrapper is not loaded.
1455  */
1456 __attribute__((weak))
1457 void lttng_ust_malloc_wrapper_init(void)
1458 {
1459 }
1460
1461 /*
1462  * sessiond monitoring thread: monitor presence of global and per-user
1463  * sessiond by polling the application common named pipe.
1464  */
1465 void __attribute__((constructor)) lttng_ust_init(void)
1466 {
1467         struct timespec constructor_timeout;
1468         sigset_t sig_all_blocked, orig_parent_mask;
1469         pthread_attr_t thread_attr;
1470         int timeout_mode;
1471         int ret;
1472
1473         if (uatomic_xchg(&initialized, 1) == 1)
1474                 return;
1475
1476         /*
1477          * Fixup interdependency between TLS fixup mutex (which happens
1478          * to be the dynamic linker mutex) and ust_lock, taken within
1479          * the ust lock.
1480          */
1481         lttng_fixup_urcu_bp_tls();
1482         lttng_fixup_ringbuffer_tls();
1483         lttng_fixup_vtid_tls();
1484         lttng_fixup_nest_count_tls();
1485         lttng_fixup_procname_tls();
1486         lttng_fixup_ust_mutex_nest_tls();
1487
1488         /*
1489          * We want precise control over the order in which we construct
1490          * our sub-libraries vs starting to receive commands from
1491          * sessiond (otherwise leading to errors when trying to create
1492          * sessiond before the init functions are completed).
1493          */
1494         init_usterr();
1495         init_tracepoint();
1496         lttng_ust_clock_init();
1497         lttng_ust_getcpu_init();
1498         lttng_ust_statedump_init();
1499         lttng_ring_buffer_metadata_client_init();
1500         lttng_ring_buffer_client_overwrite_init();
1501         lttng_ring_buffer_client_overwrite_rt_init();
1502         lttng_ring_buffer_client_discard_init();
1503         lttng_ring_buffer_client_discard_rt_init();
1504         lttng_perf_counter_init();
1505         lttng_context_init();
1506         /*
1507          * Invoke ust malloc wrapper init before starting other threads.
1508          */
1509         lttng_ust_malloc_wrapper_init();
1510
1511         timeout_mode = get_constructor_timeout(&constructor_timeout);
1512
1513         ret = sem_init(&constructor_wait, 0, 0);
1514         assert(!ret);
1515
1516         ret = setup_local_apps();
1517         if (ret) {
1518                 DBG("local apps setup returned %d", ret);
1519         }
1520
1521         /* A new thread created by pthread_create inherits the signal mask
1522          * from the parent. To avoid any signal being received by the
1523          * listener thread, we block all signals temporarily in the parent,
1524          * while we create the listener thread.
1525          */
1526         sigfillset(&sig_all_blocked);
1527         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1528         if (ret) {
1529                 ERR("pthread_sigmask: %s", strerror(ret));
1530         }
1531
1532         ret = pthread_attr_init(&thread_attr);
1533         if (ret) {
1534                 ERR("pthread_attr_init: %s", strerror(ret));
1535         }
1536         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1537         if (ret) {
1538                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1539         }
1540
1541         pthread_mutex_lock(&ust_exit_mutex);
1542         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1543                         ust_listener_thread, &global_apps);
1544         if (ret) {
1545                 ERR("pthread_create global: %s", strerror(ret));
1546         }
1547         global_apps.thread_active = 1;
1548         pthread_mutex_unlock(&ust_exit_mutex);
1549
1550         if (local_apps.allowed) {
1551                 pthread_mutex_lock(&ust_exit_mutex);
1552                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1553                                 ust_listener_thread, &local_apps);
1554                 if (ret) {
1555                         ERR("pthread_create local: %s", strerror(ret));
1556                 }
1557                 local_apps.thread_active = 1;
1558                 pthread_mutex_unlock(&ust_exit_mutex);
1559         } else {
1560                 handle_register_done(&local_apps);
1561         }
1562         ret = pthread_attr_destroy(&thread_attr);
1563         if (ret) {
1564                 ERR("pthread_attr_destroy: %s", strerror(ret));
1565         }
1566
1567         /* Restore original signal mask in parent */
1568         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1569         if (ret) {
1570                 ERR("pthread_sigmask: %s", strerror(ret));
1571         }
1572
1573         switch (timeout_mode) {
1574         case 1: /* timeout wait */
1575                 do {
1576                         ret = sem_timedwait(&constructor_wait,
1577                                         &constructor_timeout);
1578                 } while (ret < 0 && errno == EINTR);
1579                 if (ret < 0 && errno == ETIMEDOUT) {
1580                         ERR("Timed out waiting for lttng-sessiond");
1581                 } else {
1582                         assert(!ret);
1583                 }
1584                 break;
1585         case -1:/* wait forever */
1586                 do {
1587                         ret = sem_wait(&constructor_wait);
1588                 } while (ret < 0 && errno == EINTR);
1589                 assert(!ret);
1590                 break;
1591         case 0: /* no timeout */
1592                 break;
1593         }
1594 }
1595
1596 static
1597 void lttng_ust_cleanup(int exiting)
1598 {
1599         cleanup_sock_info(&global_apps, exiting);
1600         cleanup_sock_info(&local_apps, exiting);
1601         /*
1602          * The teardown in this function all affect data structures
1603          * accessed under the UST lock by the listener thread. This
1604          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1605          * that none of these threads are accessing this data at this
1606          * point.
1607          */
1608         lttng_ust_abi_exit();
1609         lttng_ust_events_exit();
1610         lttng_context_exit();
1611         lttng_perf_counter_exit();
1612         lttng_ring_buffer_client_discard_rt_exit();
1613         lttng_ring_buffer_client_discard_exit();
1614         lttng_ring_buffer_client_overwrite_rt_exit();
1615         lttng_ring_buffer_client_overwrite_exit();
1616         lttng_ring_buffer_metadata_client_exit();
1617         lttng_ust_statedump_destroy();
1618         exit_tracepoint();
1619         if (!exiting) {
1620                 /* Reinitialize values for fork */
1621                 sem_count = 2;
1622                 lttng_ust_comm_should_quit = 0;
1623                 initialized = 0;
1624         }
1625 }
1626
1627 void __attribute__((destructor)) lttng_ust_exit(void)
1628 {
1629         int ret;
1630
1631         /*
1632          * Using pthread_cancel here because:
1633          * A) we don't want to hang application teardown.
1634          * B) the thread is not allocating any resource.
1635          */
1636
1637         /*
1638          * Require the communication thread to quit. Synchronize with
1639          * mutexes to ensure it is not in a mutex critical section when
1640          * pthread_cancel is later called.
1641          */
1642         ust_lock_nocheck();
1643         lttng_ust_comm_should_quit = 1;
1644         ust_unlock();
1645
1646         pthread_mutex_lock(&ust_exit_mutex);
1647         /* cancel threads */
1648         if (global_apps.thread_active) {
1649                 ret = pthread_cancel(global_apps.ust_listener);
1650                 if (ret) {
1651                         ERR("Error cancelling global ust listener thread: %s",
1652                                 strerror(ret));
1653                 } else {
1654                         global_apps.thread_active = 0;
1655                 }
1656         }
1657         if (local_apps.thread_active) {
1658                 ret = pthread_cancel(local_apps.ust_listener);
1659                 if (ret) {
1660                         ERR("Error cancelling local ust listener thread: %s",
1661                                 strerror(ret));
1662                 } else {
1663                         local_apps.thread_active = 0;
1664                 }
1665         }
1666         pthread_mutex_unlock(&ust_exit_mutex);
1667
1668         /*
1669          * Do NOT join threads: use of sys_futex makes it impossible to
1670          * join the threads without using async-cancel, but async-cancel
1671          * is delivered by a signal, which could hit the target thread
1672          * anywhere in its code path, including while the ust_lock() is
1673          * held, causing a deadlock for the other thread. Let the OS
1674          * cleanup the threads if there are stalled in a syscall.
1675          */
1676         lttng_ust_cleanup(1);
1677 }
1678
1679 /*
1680  * We exclude the worker threads across fork and clone (except
1681  * CLONE_VM), because these system calls only keep the forking thread
1682  * running in the child.  Therefore, we don't want to call fork or clone
1683  * in the middle of an tracepoint or ust tracing state modification.
1684  * Holding this mutex protects these structures across fork and clone.
1685  */
1686 void ust_before_fork(sigset_t *save_sigset)
1687 {
1688         /*
1689          * Disable signals. This is to avoid that the child intervenes
1690          * before it is properly setup for tracing. It is safer to
1691          * disable all signals, because then we know we are not breaking
1692          * anything by restoring the original mask.
1693          */
1694         sigset_t all_sigs;
1695         int ret;
1696
1697         if (URCU_TLS(lttng_ust_nest_count))
1698                 return;
1699         /* Disable signals */
1700         sigfillset(&all_sigs);
1701         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1702         if (ret == -1) {
1703                 PERROR("sigprocmask");
1704         }
1705
1706         pthread_mutex_lock(&ust_fork_mutex);
1707
1708         ust_lock_nocheck();
1709         rcu_bp_before_fork();
1710 }
1711
1712 static void ust_after_fork_common(sigset_t *restore_sigset)
1713 {
1714         int ret;
1715
1716         DBG("process %d", getpid());
1717         ust_unlock();
1718
1719         pthread_mutex_unlock(&ust_fork_mutex);
1720
1721         /* Restore signals */
1722         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1723         if (ret == -1) {
1724                 PERROR("sigprocmask");
1725         }
1726 }
1727
1728 void ust_after_fork_parent(sigset_t *restore_sigset)
1729 {
1730         if (URCU_TLS(lttng_ust_nest_count))
1731                 return;
1732         DBG("process %d", getpid());
1733         rcu_bp_after_fork_parent();
1734         /* Release mutexes and reenable signals */
1735         ust_after_fork_common(restore_sigset);
1736 }
1737
1738 /*
1739  * After fork, in the child, we need to cleanup all the leftover state,
1740  * except the worker thread which already magically disappeared thanks
1741  * to the weird Linux fork semantics. After tyding up, we call
1742  * lttng_ust_init() again to start over as a new PID.
1743  *
1744  * This is meant for forks() that have tracing in the child between the
1745  * fork and following exec call (if there is any).
1746  */
1747 void ust_after_fork_child(sigset_t *restore_sigset)
1748 {
1749         if (URCU_TLS(lttng_ust_nest_count))
1750                 return;
1751         DBG("process %d", getpid());
1752         /* Release urcu mutexes */
1753         rcu_bp_after_fork_child();
1754         lttng_ust_cleanup(0);
1755         lttng_context_vtid_reset();
1756         /* Release mutexes and reenable signals */
1757         ust_after_fork_common(restore_sigset);
1758         lttng_ust_init();
1759 }
1760
1761 void lttng_ust_sockinfo_session_enabled(void *owner)
1762 {
1763         struct sock_info *sock_info = owner;
1764         sock_info->statedump_pending = 1;
1765 }