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 end;
 580         }
 581
 582         ops = objd_ops(lum->handle);
 583         if (!ops) {
 584                 ret = -ENOENT;
 585                 goto end;
 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 end;
 647                         } else {
 648                                 DBG("incorrect filter data message size: %zd", len);
 649                                 ret = -EINVAL;
 650                                 free(bytecode);
 651                                 goto end;
 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 end;
 712                         } else {
 713                                 DBG("Incorrect exclusion data message size: %zd", len);
 714                                 ret = -EINVAL;
 715                                 free(node);
 716                                 goto end;
 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 end;
 758                         } else {
 759                                 DBG("incorrect channel data message size: %zd", len);
 760                                 ret = -EINVAL;
 761                                 goto end;
 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 end;
 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 end:
 803         lur.handle = lum->handle;
 804         lur.cmd = lum->cmd;
 805         lur.ret_val = ret;
 806         if (ret >= 0) {
 807                 lur.ret_code = LTTNG_UST_OK;
 808         } else {
 809                 /*
 810                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 811                  * error that are not caused by the transport, except if
 812                  * we already have a more precise error message to
 813                  * report.
 814                  */
 815                 if (ret > -LTTNG_UST_ERR) {
 816                         /* Translate code to UST error. */
 817                         switch (ret) {
 818                         case -EEXIST:
 819                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 820                                 break;
 821                         case -EINVAL:
 822                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 823                                 break;
 824                         case -ENOENT:
 825                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 826                                 break;
 827                         case -EPERM:
 828                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 829                                 break;
 830                         case -ENOSYS:
 831                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 832                                 break;
 833                         default:
 834                                 lur.ret_code = -LTTNG_UST_ERR;
 835                                 break;
 836                         }
 837                 } else {
 838                         lur.ret_code = ret;
 839                 }
 840         }
 841         if (ret >= 0) {
 842                 switch (lum->cmd) {
 843                 case LTTNG_UST_TRACER_VERSION:
 844                         lur.u.version = lum->u.version;
 845                         break;
 846                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 847                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 848                         break;
 849                 }
 850         }
 851         DBG("Return value: %d", lur.ret_val);
 852         ret = send_reply(sock, &lur);
 853         if (ret < 0) {
 854                 DBG("error sending reply");
 855                 goto error;
 856         }
 857
 858         /*
 859          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 860          * after the reply.
 861          */
 862         if (lur.ret_code == LTTNG_UST_OK) {
 863                 switch (lum->cmd) {
 864                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 865                         len = ustcomm_send_unix_sock(sock,
 866                                 &args.field_list.entry,
 867                                 sizeof(args.field_list.entry));
 868                         if (len < 0) {
 869                                 ret = len;
 870                                 goto error;
 871                         }
 872                         if (len != sizeof(args.field_list.entry)) {
 873                                 ret = -EINVAL;
 874                                 goto error;
 875                         }
 876                 }
 877         }
 878
 879 error:
 880         ust_unlock();
 881
 882         /*
 883          * Performed delayed statedump operations outside of the UST
 884          * lock. We need to take the dynamic loader lock before we take
 885          * the UST lock internally within handle_pending_statedump().
 886           */
 887         handle_pending_statedump(sock_info);
 888
 889         return ret;
 890 }
 891
 892 static
 893 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 894 {
 895         int ret;
 896
 897         if (sock_info->root_handle != -1) {
 898                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 899                 if (ret) {
 900                         ERR("Error unref root handle");
 901                 }
 902                 sock_info->root_handle = -1;
 903         }
 904         sock_info->constructor_sem_posted = 0;
 905
 906         /*
 907          * wait_shm_mmap, socket and notify socket are used by listener
 908          * threads outside of the ust lock, so we cannot tear them down
 909          * ourselves, because we cannot join on these threads. Leave
 910          * responsibility of cleaning up these resources to the OS
 911          * process exit.
 912          */
 913         if (exiting)
 914                 return;
 915
 916         if (sock_info->socket != -1) {
 917                 ret = ustcomm_close_unix_sock(sock_info->socket);
 918                 if (ret) {
 919                         ERR("Error closing ust cmd socket");
 920                 }
 921                 sock_info->socket = -1;
 922         }
 923         if (sock_info->notify_socket != -1) {
 924                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 925                 if (ret) {
 926                         ERR("Error closing ust notify socket");
 927                 }
 928                 sock_info->notify_socket = -1;
 929         }
 930         if (sock_info->wait_shm_mmap) {
 931                 long page_size;
 932
 933                 page_size = sysconf(_SC_PAGE_SIZE);
 934                 if (page_size > 0) {
 935                         ret = munmap(sock_info->wait_shm_mmap, page_size);
 936                         if (ret) {
 937                                 ERR("Error unmapping wait shm");
 938                         }
 939                 }
 940                 sock_info->wait_shm_mmap = NULL;
 941         }
 942 }
 943
 944 /*
 945  * Using fork to set umask in the child process (not multi-thread safe).
 946  * We deal with the shm_open vs ftruncate race (happening when the
 947  * sessiond owns the shm and does not let everybody modify it, to ensure
 948  * safety against shm_unlink) by simply letting the mmap fail and
 949  * retrying after a few seconds.
 950  * For global shm, everybody has rw access to it until the sessiond
 951  * starts.
 952  */
 953 static
 954 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 955 {
 956         int wait_shm_fd, ret;
 957         pid_t pid;
 958
 959         /*
 960          * Try to open read-only.
 961          */
 962         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 963         if (wait_shm_fd >= 0) {
 964                 int32_t tmp_read;
 965                 ssize_t len;
 966                 size_t bytes_read = 0;
 967
 968                 /*
 969                  * Try to read the fd. If unable to do so, try opening
 970                  * it in write mode.
 971                  */
 972                 do {
 973                         len = read(wait_shm_fd,
 974                                 &((char *) &tmp_read)[bytes_read],
 975                                 sizeof(tmp_read) - bytes_read);
 976                         if (len > 0) {
 977                                 bytes_read += len;
 978                         }
 979                 } while ((len < 0 && errno == EINTR)
 980                         || (len > 0 && bytes_read < sizeof(tmp_read)));
 981                 if (bytes_read != sizeof(tmp_read)) {
 982                         ret = close(wait_shm_fd);
 983                         if (ret) {
 984                                 ERR("close wait_shm_fd");
 985                         }
 986                         goto open_write;
 987                 }
 988                 goto end;
 989         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 990                 /*
 991                  * Real-only open did not work, and it's not because the
 992                  * entry was not present. It's a failure that prohibits
 993                  * using shm.
 994                  */
 995                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 996                 goto end;
 997         }
 998
 999 open_write:
1000         /*
1001          * If the open failed because the file did not exist, or because
1002          * the file was not truncated yet, try creating it ourself.
1003          */
1004         URCU_TLS(lttng_ust_nest_count)++;
1005         pid = fork();
1006         URCU_TLS(lttng_ust_nest_count)--;
1007         if (pid > 0) {
1008                 int status;
1009
1010                 /*
1011                  * Parent: wait for child to return, in which case the
1012                  * shared memory map will have been created.
1013                  */
1014                 pid = wait(&status);
1015                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
1016                         wait_shm_fd = -1;
1017                         goto end;
1018                 }
1019                 /*
1020                  * Try to open read-only again after creation.
1021                  */
1022                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1023                 if (wait_shm_fd < 0) {
1024                         /*
1025                          * Real-only open did not work. It's a failure
1026                          * that prohibits using shm.
1027                          */
1028                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1029                         goto end;
1030                 }
1031                 goto end;
1032         } else if (pid == 0) {
1033                 int create_mode;
1034
1035                 /* Child */
1036                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1037                 if (sock_info->global)
1038                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1039                 /*
1040                  * We're alone in a child process, so we can modify the
1041                  * process-wide umask.
1042                  */
1043                 umask(~create_mode);
1044                 /*
1045                  * Try creating shm (or get rw access).
1046                  * We don't do an exclusive open, because we allow other
1047                  * processes to create+ftruncate it concurrently.
1048                  */
1049                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1050                                 O_RDWR | O_CREAT, create_mode);
1051                 if (wait_shm_fd >= 0) {
1052                         ret = ftruncate(wait_shm_fd, mmap_size);
1053                         if (ret) {
1054                                 PERROR("ftruncate");
1055                                 _exit(EXIT_FAILURE);
1056                         }
1057                         _exit(EXIT_SUCCESS);
1058                 }
1059                 /*
1060                  * For local shm, we need to have rw access to accept
1061                  * opening it: this means the local sessiond will be
1062                  * able to wake us up. For global shm, we open it even
1063                  * if rw access is not granted, because the root.root
1064                  * sessiond will be able to override all rights and wake
1065                  * us up.
1066                  */
1067                 if (!sock_info->global && errno != EACCES) {
1068                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1069                         _exit(EXIT_FAILURE);
1070                 }
1071                 /*
1072                  * The shm exists, but we cannot open it RW. Report
1073                  * success.
1074                  */
1075                 _exit(EXIT_SUCCESS);
1076         } else {
1077                 return -1;
1078         }
1079 end:
1080         if (wait_shm_fd >= 0 && !sock_info->global) {
1081                 struct stat statbuf;
1082
1083                 /*
1084                  * Ensure that our user is the owner of the shm file for
1085                  * local shm. If we do not own the file, it means our
1086                  * sessiond will not have access to wake us up (there is
1087                  * probably a rogue process trying to fake our
1088                  * sessiond). Fallback to polling method in this case.
1089                  */
1090                 ret = fstat(wait_shm_fd, &statbuf);
1091                 if (ret) {
1092                         PERROR("fstat");
1093                         goto error_close;
1094                 }
1095                 if (statbuf.st_uid != getuid())
1096                         goto error_close;
1097         }
1098         return wait_shm_fd;
1099
1100 error_close:
1101         ret = close(wait_shm_fd);
1102         if (ret) {
1103                 PERROR("Error closing fd");
1104         }
1105         return -1;
1106 }
1107
1108 static
1109 char *get_map_shm(struct sock_info *sock_info)
1110 {
1111         long page_size;
1112         int wait_shm_fd, ret;
1113         char *wait_shm_mmap;
1114
1115         page_size = sysconf(_SC_PAGE_SIZE);
1116         if (page_size < 0) {
1117                 goto error;
1118         }
1119
1120         wait_shm_fd = get_wait_shm(sock_info, page_size);
1121         if (wait_shm_fd < 0) {
1122                 goto error;
1123         }
1124         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1125                   MAP_SHARED, wait_shm_fd, 0);
1126         /* close shm fd immediately after taking the mmap reference */
1127         ret = close(wait_shm_fd);
1128         if (ret) {
1129                 PERROR("Error closing fd");
1130         }
1131         if (wait_shm_mmap == MAP_FAILED) {
1132                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1133                 goto error;
1134         }
1135         return wait_shm_mmap;
1136
1137 error:
1138         return NULL;
1139 }
1140
1141 static
1142 void wait_for_sessiond(struct sock_info *sock_info)
1143 {
1144         if (ust_lock()) {
1145                 goto quit;
1146         }
1147         if (wait_poll_fallback) {
1148                 goto error;
1149         }
1150         if (!sock_info->wait_shm_mmap) {
1151                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1152                 if (!sock_info->wait_shm_mmap)
1153                         goto error;
1154         }
1155         ust_unlock();
1156
1157         DBG("Waiting for %s apps sessiond", sock_info->name);
1158         /* Wait for futex wakeup */
1159         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
1160                 goto end_wait;
1161
1162         while (futex_async((int32_t *) sock_info->wait_shm_mmap,
1163                         FUTEX_WAIT, 0, NULL, NULL, 0)) {
1164                 switch (errno) {
1165                 case EWOULDBLOCK:
1166                         /* Value already changed. */
1167                         goto end_wait;
1168                 case EINTR:
1169                         /* Retry if interrupted by signal. */
1170                         break;  /* Get out of switch. */
1171                 case EFAULT:
1172                         wait_poll_fallback = 1;
1173                         DBG(
1174 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1175 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1176 "Please upgrade your kernel "
1177 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1178 "mainline). LTTng-UST will use polling mode fallback.");
1179                         if (ust_debug())
1180                                 PERROR("futex");
1181                         goto end_wait;
1182                 }
1183         }
1184 end_wait:
1185         return;
1186
1187 quit:
1188         ust_unlock();
1189         return;
1190
1191 error:
1192         ust_unlock();
1193         return;
1194 }
1195
1196 /*
1197  * This thread does not allocate any resource, except within
1198  * handle_message, within mutex protection. This mutex protects against
1199  * fork and exit.
1200  * The other moment it allocates resources is at socket connection, which
1201  * is also protected by the mutex.
1202  */
1203 static
1204 void *ust_listener_thread(void *arg)
1205 {
1206         struct sock_info *sock_info = arg;
1207         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1208         long timeout;
1209
1210         /* Restart trying to connect to the session daemon */
1211 restart:
1212         if (prev_connect_failed) {
1213                 /* Wait for sessiond availability with pipe */
1214                 wait_for_sessiond(sock_info);
1215                 if (has_waited) {
1216                         has_waited = 0;
1217                         /*
1218                          * Sleep for 5 seconds before retrying after a
1219                          * sequence of failure / wait / failure. This
1220                          * deals with a killed or broken session daemon.
1221                          */
1222                         sleep(5);
1223                 }
1224                 has_waited = 1;
1225                 prev_connect_failed = 0;
1226         }
1227
1228         if (sock_info->socket != -1) {
1229                 ret = ustcomm_close_unix_sock(sock_info->socket);
1230                 if (ret) {
1231                         ERR("Error closing %s ust cmd socket",
1232                                 sock_info->name);
1233                 }
1234                 sock_info->socket = -1;
1235         }
1236         if (sock_info->notify_socket != -1) {
1237                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1238                 if (ret) {
1239                         ERR("Error closing %s ust notify socket",
1240                                 sock_info->name);
1241                 }
1242                 sock_info->notify_socket = -1;
1243         }
1244
1245         /*
1246          * Register. We need to perform both connect and sending
1247          * registration message before doing the next connect otherwise
1248          * we may reach unix socket connect queue max limits and block
1249          * on the 2nd connect while the session daemon is awaiting the
1250          * first connect registration message.
1251          */
1252         /* Connect cmd socket */
1253         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1254         if (ret < 0) {
1255                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1256                 prev_connect_failed = 1;
1257
1258                 if (ust_lock()) {
1259                         goto quit;
1260                 }
1261
1262                 /*
1263                  * If we cannot find the sessiond daemon, don't delay
1264                  * constructor execution.
1265                  */
1266                 ret = handle_register_done(sock_info);
1267                 assert(!ret);
1268                 ust_unlock();
1269                 goto restart;
1270         }
1271         sock_info->socket = ret;
1272
1273         if (ust_lock()) {
1274                 goto quit;
1275         }
1276
1277         /*
1278          * Create only one root handle per listener thread for the whole
1279          * process lifetime, so we ensure we get ID which is statically
1280          * assigned to the root handle.
1281          */
1282         if (sock_info->root_handle == -1) {
1283                 ret = lttng_abi_create_root_handle();
1284                 if (ret < 0) {
1285                         ERR("Error creating root handle");
1286                         goto quit;
1287                 }
1288                 sock_info->root_handle = ret;
1289         }
1290
1291         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1292         if (ret < 0) {
1293                 ERR("Error registering to %s ust cmd socket",
1294                         sock_info->name);
1295                 prev_connect_failed = 1;
1296                 /*
1297                  * If we cannot register to the sessiond daemon, don't
1298                  * delay constructor execution.
1299                  */
1300                 ret = handle_register_done(sock_info);
1301                 assert(!ret);
1302                 ust_unlock();
1303                 goto restart;
1304         }
1305
1306         ust_unlock();
1307
1308         /* Connect notify socket */
1309         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1310         if (ret < 0) {
1311                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1312                 prev_connect_failed = 1;
1313
1314                 if (ust_lock()) {
1315                         goto quit;
1316                 }
1317
1318                 /*
1319                  * If we cannot find the sessiond daemon, don't delay
1320                  * constructor execution.
1321                  */
1322                 ret = handle_register_done(sock_info);
1323                 assert(!ret);
1324                 ust_unlock();
1325                 goto restart;
1326         }
1327         sock_info->notify_socket = ret;
1328
1329         timeout = get_notify_sock_timeout();
1330         if (timeout >= 0) {
1331                 /*
1332                  * Give at least 10ms to sessiond to reply to
1333                  * notifications.
1334                  */
1335                 if (timeout < 10)
1336                         timeout = 10;
1337                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1338                                 timeout);
1339                 if (ret < 0) {
1340                         WARN("Error setting socket receive timeout");
1341                 }
1342                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1343                                 timeout);
1344                 if (ret < 0) {
1345                         WARN("Error setting socket send timeout");
1346                 }
1347         } else if (timeout < -1) {
1348                 WARN("Unsupported timeout value %ld", timeout);
1349         }
1350
1351         if (ust_lock()) {
1352                 goto quit;
1353         }
1354
1355         ret = register_to_sessiond(sock_info->notify_socket,
1356                         USTCTL_SOCKET_NOTIFY);
1357         if (ret < 0) {
1358                 ERR("Error registering to %s ust notify socket",
1359                         sock_info->name);
1360                 prev_connect_failed = 1;
1361                 /*
1362                  * If we cannot register to the sessiond daemon, don't
1363                  * delay constructor execution.
1364                  */
1365                 ret = handle_register_done(sock_info);
1366                 assert(!ret);
1367                 ust_unlock();
1368                 goto restart;
1369         }
1370         sock = sock_info->socket;
1371
1372         ust_unlock();
1373
1374         for (;;) {
1375                 ssize_t len;
1376                 struct ustcomm_ust_msg lum;
1377
1378                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1379                 switch (len) {
1380                 case 0: /* orderly shutdown */
1381                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1382                         if (ust_lock()) {
1383                                 goto quit;
1384                         }
1385                         /*
1386                          * Either sessiond has shutdown or refused us by closing the socket.
1387                          * In either case, we don't want to delay construction execution,
1388                          * and we need to wait before retry.
1389                          */
1390                         prev_connect_failed = 1;
1391                         /*
1392                          * If we cannot register to the sessiond daemon, don't
1393                          * delay constructor execution.
1394                          */
1395                         ret = handle_register_done(sock_info);
1396                         assert(!ret);
1397                         ust_unlock();
1398                         goto end;
1399                 case sizeof(lum):
1400                         print_cmd(lum.cmd, lum.handle);
1401                         ret = handle_message(sock_info, sock, &lum);
1402                         if (ret) {
1403                                 ERR("Error handling message for %s socket", sock_info->name);
1404                         }
1405                         continue;
1406                 default:
1407                         if (len < 0) {
1408                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1409                         } else {
1410                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1411                         }
1412                         if (len == -ECONNRESET) {
1413                                 DBG("%s remote end closed connection", sock_info->name);
1414                                 goto end;
1415                         }
1416                         goto end;
1417                 }
1418
1419         }
1420 end:
1421         if (ust_lock()) {
1422                 goto quit;
1423         }
1424         /* Cleanup socket handles before trying to reconnect */
1425         lttng_ust_objd_table_owner_cleanup(sock_info);
1426         ust_unlock();
1427         goto restart;   /* try to reconnect */
1428
1429 quit:
1430         ust_unlock();
1431
1432         pthread_mutex_lock(&ust_exit_mutex);
1433         sock_info->thread_active = 0;
1434         pthread_mutex_unlock(&ust_exit_mutex);
1435         return NULL;
1436 }
1437
1438 /*
1439  * Weak symbol to call when the ust malloc wrapper is not loaded.
1440  */
1441 __attribute__((weak))
1442 void lttng_ust_malloc_wrapper_init(void)
1443 {
1444 }
1445
1446 /*
1447  * sessiond monitoring thread: monitor presence of global and per-user
1448  * sessiond by polling the application common named pipe.
1449  */
1450 void __attribute__((constructor)) lttng_ust_init(void)
1451 {
1452         struct timespec constructor_timeout;
1453         sigset_t sig_all_blocked, orig_parent_mask;
1454         pthread_attr_t thread_attr;
1455         int timeout_mode;
1456         int ret;
1457
1458         if (uatomic_xchg(&initialized, 1) == 1)
1459                 return;
1460
1461         /*
1462          * Fixup interdependency between TLS fixup mutex (which happens
1463          * to be the dynamic linker mutex) and ust_lock, taken within
1464          * the ust lock.
1465          */
1466         lttng_fixup_urcu_bp_tls();
1467         lttng_fixup_ringbuffer_tls();
1468         lttng_fixup_vtid_tls();
1469         lttng_fixup_nest_count_tls();
1470         lttng_fixup_procname_tls();
1471         lttng_fixup_ust_mutex_nest_tls();
1472
1473         /*
1474          * We want precise control over the order in which we construct
1475          * our sub-libraries vs starting to receive commands from
1476          * sessiond (otherwise leading to errors when trying to create
1477          * sessiond before the init functions are completed).
1478          */
1479         init_usterr();
1480         init_tracepoint();
1481         lttng_ust_clock_init();
1482         lttng_ust_getcpu_init();
1483         lttng_ust_statedump_init();
1484         lttng_ring_buffer_metadata_client_init();
1485         lttng_ring_buffer_client_overwrite_init();
1486         lttng_ring_buffer_client_overwrite_rt_init();
1487         lttng_ring_buffer_client_discard_init();
1488         lttng_ring_buffer_client_discard_rt_init();
1489         lttng_perf_counter_init();
1490         lttng_context_init();
1491         /*
1492          * Invoke ust malloc wrapper init before starting other threads.
1493          */
1494         lttng_ust_malloc_wrapper_init();
1495
1496         timeout_mode = get_constructor_timeout(&constructor_timeout);
1497
1498         ret = sem_init(&constructor_wait, 0, 0);
1499         assert(!ret);
1500
1501         ret = setup_local_apps();
1502         if (ret) {
1503                 DBG("local apps setup returned %d", ret);
1504         }
1505
1506         /* A new thread created by pthread_create inherits the signal mask
1507          * from the parent. To avoid any signal being received by the
1508          * listener thread, we block all signals temporarily in the parent,
1509          * while we create the listener thread.
1510          */
1511         sigfillset(&sig_all_blocked);
1512         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1513         if (ret) {
1514                 ERR("pthread_sigmask: %s", strerror(ret));
1515         }
1516
1517         ret = pthread_attr_init(&thread_attr);
1518         if (ret) {
1519                 ERR("pthread_attr_init: %s", strerror(ret));
1520         }
1521         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1522         if (ret) {
1523                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1524         }
1525
1526         pthread_mutex_lock(&ust_exit_mutex);
1527         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1528                         ust_listener_thread, &global_apps);
1529         if (ret) {
1530                 ERR("pthread_create global: %s", strerror(ret));
1531         }
1532         global_apps.thread_active = 1;
1533         pthread_mutex_unlock(&ust_exit_mutex);
1534
1535         if (local_apps.allowed) {
1536                 pthread_mutex_lock(&ust_exit_mutex);
1537                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1538                                 ust_listener_thread, &local_apps);
1539                 if (ret) {
1540                         ERR("pthread_create local: %s", strerror(ret));
1541                 }
1542                 local_apps.thread_active = 1;
1543                 pthread_mutex_unlock(&ust_exit_mutex);
1544         } else {
1545                 handle_register_done(&local_apps);
1546         }
1547         ret = pthread_attr_destroy(&thread_attr);
1548         if (ret) {
1549                 ERR("pthread_attr_destroy: %s", strerror(ret));
1550         }
1551
1552         /* Restore original signal mask in parent */
1553         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1554         if (ret) {
1555                 ERR("pthread_sigmask: %s", strerror(ret));
1556         }
1557
1558         switch (timeout_mode) {
1559         case 1: /* timeout wait */
1560                 do {
1561                         ret = sem_timedwait(&constructor_wait,
1562                                         &constructor_timeout);
1563                 } while (ret < 0 && errno == EINTR);
1564                 if (ret < 0 && errno == ETIMEDOUT) {
1565                         ERR("Timed out waiting for lttng-sessiond");
1566                 } else {
1567                         assert(!ret);
1568                 }
1569                 break;
1570         case -1:/* wait forever */
1571                 do {
1572                         ret = sem_wait(&constructor_wait);
1573                 } while (ret < 0 && errno == EINTR);
1574                 assert(!ret);
1575                 break;
1576         case 0: /* no timeout */
1577                 break;
1578         }
1579 }
1580
1581 static
1582 void lttng_ust_cleanup(int exiting)
1583 {
1584         cleanup_sock_info(&global_apps, exiting);
1585         cleanup_sock_info(&local_apps, exiting);
1586         /*
1587          * The teardown in this function all affect data structures
1588          * accessed under the UST lock by the listener thread. This
1589          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1590          * that none of these threads are accessing this data at this
1591          * point.
1592          */
1593         lttng_ust_abi_exit();
1594         lttng_ust_events_exit();
1595         lttng_context_exit();
1596         lttng_perf_counter_exit();
1597         lttng_ring_buffer_client_discard_rt_exit();
1598         lttng_ring_buffer_client_discard_exit();
1599         lttng_ring_buffer_client_overwrite_rt_exit();
1600         lttng_ring_buffer_client_overwrite_exit();
1601         lttng_ring_buffer_metadata_client_exit();
1602         lttng_ust_statedump_destroy();
1603         exit_tracepoint();
1604         if (!exiting) {
1605                 /* Reinitialize values for fork */
1606                 sem_count = 2;
1607                 lttng_ust_comm_should_quit = 0;
1608                 initialized = 0;
1609         }
1610 }
1611
1612 void __attribute__((destructor)) lttng_ust_exit(void)
1613 {
1614         int ret;
1615
1616         /*
1617          * Using pthread_cancel here because:
1618          * A) we don't want to hang application teardown.
1619          * B) the thread is not allocating any resource.
1620          */
1621
1622         /*
1623          * Require the communication thread to quit. Synchronize with
1624          * mutexes to ensure it is not in a mutex critical section when
1625          * pthread_cancel is later called.
1626          */
1627         ust_lock_nocheck();
1628         lttng_ust_comm_should_quit = 1;
1629         ust_unlock();
1630
1631         pthread_mutex_lock(&ust_exit_mutex);
1632         /* cancel threads */
1633         if (global_apps.thread_active) {
1634                 ret = pthread_cancel(global_apps.ust_listener);
1635                 if (ret) {
1636                         ERR("Error cancelling global ust listener thread: %s",
1637                                 strerror(ret));
1638                 } else {
1639                         global_apps.thread_active = 0;
1640                 }
1641         }
1642         if (local_apps.thread_active) {
1643                 ret = pthread_cancel(local_apps.ust_listener);
1644                 if (ret) {
1645                         ERR("Error cancelling local ust listener thread: %s",
1646                                 strerror(ret));
1647                 } else {
1648                         local_apps.thread_active = 0;
1649                 }
1650         }
1651         pthread_mutex_unlock(&ust_exit_mutex);
1652
1653         /*
1654          * Do NOT join threads: use of sys_futex makes it impossible to
1655          * join the threads without using async-cancel, but async-cancel
1656          * is delivered by a signal, which could hit the target thread
1657          * anywhere in its code path, including while the ust_lock() is
1658          * held, causing a deadlock for the other thread. Let the OS
1659          * cleanup the threads if there are stalled in a syscall.
1660          */
1661         lttng_ust_cleanup(1);
1662 }
1663
1664 /*
1665  * We exclude the worker threads across fork and clone (except
1666  * CLONE_VM), because these system calls only keep the forking thread
1667  * running in the child.  Therefore, we don't want to call fork or clone
1668  * in the middle of an tracepoint or ust tracing state modification.
1669  * Holding this mutex protects these structures across fork and clone.
1670  */
1671 void ust_before_fork(sigset_t *save_sigset)
1672 {
1673         /*
1674          * Disable signals. This is to avoid that the child intervenes
1675          * before it is properly setup for tracing. It is safer to
1676          * disable all signals, because then we know we are not breaking
1677          * anything by restoring the original mask.
1678          */
1679         sigset_t all_sigs;
1680         int ret;
1681
1682         if (URCU_TLS(lttng_ust_nest_count))
1683                 return;
1684         /* Disable signals */
1685         sigfillset(&all_sigs);
1686         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1687         if (ret == -1) {
1688                 PERROR("sigprocmask");
1689         }
1690
1691         pthread_mutex_lock(&ust_fork_mutex);
1692
1693         ust_lock_nocheck();
1694         rcu_bp_before_fork();
1695 }
1696
1697 static void ust_after_fork_common(sigset_t *restore_sigset)
1698 {
1699         int ret;
1700
1701         DBG("process %d", getpid());
1702         ust_unlock();
1703
1704         pthread_mutex_unlock(&ust_fork_mutex);
1705
1706         /* Restore signals */
1707         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1708         if (ret == -1) {
1709                 PERROR("sigprocmask");
1710         }
1711 }
1712
1713 void ust_after_fork_parent(sigset_t *restore_sigset)
1714 {
1715         if (URCU_TLS(lttng_ust_nest_count))
1716                 return;
1717         DBG("process %d", getpid());
1718         rcu_bp_after_fork_parent();
1719         /* Release mutexes and reenable signals */
1720         ust_after_fork_common(restore_sigset);
1721 }
1722
1723 /*
1724  * After fork, in the child, we need to cleanup all the leftover state,
1725  * except the worker thread which already magically disappeared thanks
1726  * to the weird Linux fork semantics. After tyding up, we call
1727  * lttng_ust_init() again to start over as a new PID.
1728  *
1729  * This is meant for forks() that have tracing in the child between the
1730  * fork and following exec call (if there is any).
1731  */
1732 void ust_after_fork_child(sigset_t *restore_sigset)
1733 {
1734         if (URCU_TLS(lttng_ust_nest_count))
1735                 return;
1736         DBG("process %d", getpid());
1737         /* Release urcu mutexes */
1738         rcu_bp_after_fork_child();
1739         lttng_ust_cleanup(0);
1740         lttng_context_vtid_reset();
1741         /* Release mutexes and reenable signals */
1742         ust_after_fork_common(restore_sigset);
1743         lttng_ust_init();
1744 }
1745
1746 void lttng_ust_sockinfo_session_enabled(void *owner)
1747 {
1748         struct sock_info *sock_info = owner;
1749         sock_info->statedump_pending = 1;
1750 }