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 <dlfcn.h>
  31 #include <fcntl.h>
  32 #include <unistd.h>
  33 #include <errno.h>
  34 #include <pthread.h>
  35 #include <semaphore.h>
  36 #include <time.h>
  37 #include <assert.h>
  38 #include <signal.h>
  39 #include <limits.h>
  40 #include <urcu/uatomic.h>
  41 #include <urcu/futex.h>
  42 #include <urcu/compiler.h>
  43
  44 #include <lttng/ust-events.h>
  45 #include <lttng/ust-abi.h>
  46 #include <lttng/ust.h>
  47 #include <lttng/ust-error.h>
  48 #include <lttng/ust-ctl.h>
  49 #include <urcu/tls-compat.h>
  50 #include <ust-comm.h>
  51 #include <ust-fd.h>
  52 #include <usterr-signal-safe.h>
  53 #include <helper.h>
  54 #include "tracepoint-internal.h"
  55 #include "lttng-tracer-core.h"
  56 #include "compat.h"
  57 #include "../libringbuffer/rb-init.h"
  58 #include "lttng-ust-statedump.h"
  59 #include "clock.h"
  60 #include "../libringbuffer/getcpu.h"
  61 #include "getenv.h"
  62
  63 /* Concatenate lttng ust shared library name with its major version number. */
  64 #define LTTNG_UST_LIB_SO_NAME "liblttng-ust.so." LTTNG_UST_LIBRARY_VERSION_MAJOR
  65
  66 /*
  67  * Has lttng ust comm constructor been called ?
  68  */
  69 static int initialized;
  70
  71 /*
  72  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  73  * Held when handling a command, also held by fork() to deal with
  74  * removal of threads, and by exit path.
  75  *
  76  * The UST lock is the centralized mutex across UST tracing control and
  77  * probe registration.
  78  *
  79  * ust_exit_mutex must never nest in ust_mutex.
  80  *
  81  * ust_fork_mutex must never nest in ust_mutex.
  82  *
  83  * ust_mutex_nest is a per-thread nesting counter, allowing the perf
  84  * counter lazy initialization called by events within the statedump,
  85  * which traces while the ust_mutex is held.
  86  *
  87  * ust_lock nests within the dynamic loader lock (within glibc) because
  88  * it is taken within the library constructor.
  89  */
  90 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
  91
  92 /* Allow nesting the ust_mutex within the same thread. */
  93 static DEFINE_URCU_TLS(int, ust_mutex_nest);
  94
  95 /*
  96  * ust_exit_mutex protects thread_active variable wrt thread exit. It
  97  * cannot be done by ust_mutex because pthread_cancel(), which takes an
  98  * internal libc lock, cannot nest within ust_mutex.
  99  *
 100  * It never nests within a ust_mutex.
 101  */
 102 static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
 103
 104 /*
 105  * ust_fork_mutex protects base address statedump tracing against forks. It
 106  * prevents the dynamic loader lock to be taken (by base address statedump
 107  * tracing) while a fork is happening, thus preventing deadlock issues with
 108  * the dynamic loader lock.
 109  */
 110 static pthread_mutex_t ust_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
 111
 112 /* Should the ust comm thread quit ? */
 113 static int lttng_ust_comm_should_quit;
 114
 115 /*
 116  * This variable can be tested by applications to check whether
 117  * lttng-ust is loaded. They simply have to define their own
 118  * "lttng_ust_loaded" weak symbol, and test it. It is set to 1 by the
 119  * library constructor.
 120  */
 121 int lttng_ust_loaded __attribute__((weak));
 122
 123 /*
 124  * Return 0 on success, -1 if should quit.
 125  * The lock is taken in both cases.
 126  * Signal-safe.
 127  */
 128 int ust_lock(void)
 129 {
 130         sigset_t sig_all_blocked, orig_mask;
 131         int ret, oldstate;
 132
 133         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 134         if (ret) {
 135                 ERR("pthread_setcancelstate: %s", strerror(ret));
 136         }
 137         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 138                 ERR("pthread_setcancelstate: unexpected oldstate");
 139         }
 140         sigfillset(&sig_all_blocked);
 141         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 142         if (ret) {
 143                 ERR("pthread_sigmask: %s", strerror(ret));
 144         }
 145         if (!URCU_TLS(ust_mutex_nest)++)
 146                 pthread_mutex_lock(&ust_mutex);
 147         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 148         if (ret) {
 149                 ERR("pthread_sigmask: %s", strerror(ret));
 150         }
 151         if (lttng_ust_comm_should_quit) {
 152                 return -1;
 153         } else {
 154                 return 0;
 155         }
 156 }
 157
 158 /*
 159  * ust_lock_nocheck() can be used in constructors/destructors, because
 160  * they are already nested within the dynamic loader lock, and therefore
 161  * have exclusive access against execution of liblttng-ust destructor.
 162  * Signal-safe.
 163  */
 164 void ust_lock_nocheck(void)
 165 {
 166         sigset_t sig_all_blocked, orig_mask;
 167         int ret, oldstate;
 168
 169         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 170         if (ret) {
 171                 ERR("pthread_setcancelstate: %s", strerror(ret));
 172         }
 173         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 174                 ERR("pthread_setcancelstate: unexpected oldstate");
 175         }
 176         sigfillset(&sig_all_blocked);
 177         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 178         if (ret) {
 179                 ERR("pthread_sigmask: %s", strerror(ret));
 180         }
 181         if (!URCU_TLS(ust_mutex_nest)++)
 182                 pthread_mutex_lock(&ust_mutex);
 183         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 184         if (ret) {
 185                 ERR("pthread_sigmask: %s", strerror(ret));
 186         }
 187 }
 188
 189 /*
 190  * Signal-safe.
 191  */
 192 void ust_unlock(void)
 193 {
 194         sigset_t sig_all_blocked, orig_mask;
 195         int ret, oldstate;
 196
 197         sigfillset(&sig_all_blocked);
 198         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 199         if (ret) {
 200                 ERR("pthread_sigmask: %s", strerror(ret));
 201         }
 202         if (!--URCU_TLS(ust_mutex_nest))
 203                 pthread_mutex_unlock(&ust_mutex);
 204         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 205         if (ret) {
 206                 ERR("pthread_sigmask: %s", strerror(ret));
 207         }
 208         ret = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &oldstate);
 209         if (ret) {
 210                 ERR("pthread_setcancelstate: %s", strerror(ret));
 211         }
 212         if (oldstate != PTHREAD_CANCEL_DISABLE) {
 213                 ERR("pthread_setcancelstate: unexpected oldstate");
 214         }
 215 }
 216
 217 /*
 218  * Wait for either of these before continuing to the main
 219  * program:
 220  * - the register_done message from sessiond daemon
 221  *   (will let the sessiond daemon enable sessions before main
 222  *   starts.)
 223  * - sessiond daemon is not reachable.
 224  * - timeout (ensuring applications are resilient to session
 225  *   daemon problems).
 226  */
 227 static sem_t constructor_wait;
 228 /*
 229  * Doing this for both the global and local sessiond.
 230  */
 231 static int sem_count = { 2 };
 232
 233 /*
 234  * Counting nesting within lttng-ust. Used to ensure that calling fork()
 235  * from liblttng-ust does not execute the pre/post fork handlers.
 236  */
 237 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
 238
 239 /*
 240  * Info about socket and associated listener thread.
 241  */
 242 struct sock_info {
 243         const char *name;
 244         pthread_t ust_listener; /* listener thread */
 245         int root_handle;
 246         int constructor_sem_posted;
 247         int allowed;
 248         int global;
 249         int thread_active;
 250
 251         char sock_path[PATH_MAX];
 252         int socket;
 253         int notify_socket;
 254
 255         char wait_shm_path[PATH_MAX];
 256         char *wait_shm_mmap;
 257         /* Keep track of lazy state dump not performed yet. */
 258         int statedump_pending;
 259 };
 260
 261 /* Socket from app (connect) to session daemon (listen) for communication */
 262 struct sock_info global_apps = {
 263         .name = "global",
 264         .global = 1,
 265
 266         .root_handle = -1,
 267         .allowed = 1,
 268         .thread_active = 0,
 269
 270         .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
 271         .socket = -1,
 272         .notify_socket = -1,
 273
 274         .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 275
 276         .statedump_pending = 0,
 277 };
 278
 279 /* TODO: allow global_apps_sock_path override */
 280
 281 struct sock_info local_apps = {
 282         .name = "local",
 283         .global = 0,
 284         .root_handle = -1,
 285         .allowed = 0,   /* Check setuid bit first */
 286         .thread_active = 0,
 287
 288         .socket = -1,
 289         .notify_socket = -1,
 290
 291         .statedump_pending = 0,
 292 };
 293
 294 static int wait_poll_fallback;
 295
 296 static const char *cmd_name_mapping[] = {
 297         [ LTTNG_UST_RELEASE ] = "Release",
 298         [ LTTNG_UST_SESSION ] = "Create Session",
 299         [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
 300
 301         [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
 302         [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
 303         [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
 304         [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
 305
 306         /* Session FD commands */
 307         [ LTTNG_UST_CHANNEL ] = "Create Channel",
 308         [ LTTNG_UST_SESSION_START ] = "Start Session",
 309         [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
 310
 311         /* Channel FD commands */
 312         [ LTTNG_UST_STREAM ] = "Create Stream",
 313         [ LTTNG_UST_EVENT ] = "Create Event",
 314
 315         /* Event and Channel FD commands */
 316         [ LTTNG_UST_CONTEXT ] = "Create Context",
 317         [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
 318
 319         /* Event, Channel and Session commands */
 320         [ LTTNG_UST_ENABLE ] = "Enable",
 321         [ LTTNG_UST_DISABLE ] = "Disable",
 322
 323         /* Tracepoint list commands */
 324         [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
 325         [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
 326
 327         /* Event FD commands */
 328         [ LTTNG_UST_FILTER ] = "Create Filter",
 329         [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
 330 };
 331
 332 static const char *str_timeout;
 333 static int got_timeout_env;
 334
 335 extern void lttng_ring_buffer_client_overwrite_init(void);
 336 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 337 extern void lttng_ring_buffer_client_discard_init(void);
 338 extern void lttng_ring_buffer_client_discard_rt_init(void);
 339 extern void lttng_ring_buffer_metadata_client_init(void);
 340 extern void lttng_ring_buffer_client_overwrite_exit(void);
 341 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 342 extern void lttng_ring_buffer_client_discard_exit(void);
 343 extern void lttng_ring_buffer_client_discard_rt_exit(void);
 344 extern void lttng_ring_buffer_metadata_client_exit(void);
 345
 346 ssize_t lttng_ust_read(int fd, void *buf, size_t len)
 347 {
 348         ssize_t ret;
 349         size_t copied = 0, to_copy = len;
 350
 351         do {
 352                 ret = read(fd, buf + copied, to_copy);
 353                 if (ret > 0) {
 354                         copied += ret;
 355                         to_copy -= ret;
 356                 }
 357         } while ((ret > 0 && to_copy > 0)
 358                 || (ret < 0 && errno == EINTR));
 359         if (ret > 0) {
 360                 ret = copied;
 361         }
 362         return ret;
 363 }
 364 /*
 365  * Returns the HOME directory path. Caller MUST NOT free(3) the returned
 366  * pointer.
 367  */
 368 static
 369 const char *get_lttng_home_dir(void)
 370 {
 371        const char *val;
 372
 373        val = (const char *) lttng_getenv("LTTNG_HOME");
 374        if (val != NULL) {
 375                return val;
 376        }
 377        return (const char *) lttng_getenv("HOME");
 378 }
 379
 380 /*
 381  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 382  */
 383 static
 384 void lttng_fixup_nest_count_tls(void)
 385 {
 386         asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 387 }
 388
 389 static
 390 void lttng_fixup_ust_mutex_nest_tls(void)
 391 {
 392         asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest)));
 393 }
 394
 395 /*
 396  * Fixup urcu bp TLS.
 397  */
 398 static
 399 void lttng_fixup_urcu_bp_tls(void)
 400 {
 401         rcu_read_lock();
 402         rcu_read_unlock();
 403 }
 404
 405 void lttng_ust_fixup_tls(void)
 406 {
 407         lttng_fixup_urcu_bp_tls();
 408         lttng_fixup_ringbuffer_tls();
 409         lttng_fixup_vtid_tls();
 410         lttng_fixup_nest_count_tls();
 411         lttng_fixup_procname_tls();
 412         lttng_fixup_ust_mutex_nest_tls();
 413         lttng_ust_fixup_fd_tracker_tls();
 414 }
 415
 416 int lttng_get_notify_socket(void *owner)
 417 {
 418         struct sock_info *info = owner;
 419
 420         return info->notify_socket;
 421 }
 422
 423 static
 424 void print_cmd(int cmd, int handle)
 425 {
 426         const char *cmd_name = "Unknown";
 427
 428         if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
 429                         && cmd_name_mapping[cmd]) {
 430                 cmd_name = cmd_name_mapping[cmd];
 431         }
 432         DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
 433                 cmd_name, cmd,
 434                 lttng_ust_obj_get_name(handle), handle);
 435 }
 436
 437 static
 438 int setup_local_apps(void)
 439 {
 440         const char *home_dir;
 441         uid_t uid;
 442
 443         uid = getuid();
 444         /*
 445          * Disallow per-user tracing for setuid binaries.
 446          */
 447         if (uid != geteuid()) {
 448                 assert(local_apps.allowed == 0);
 449                 return 0;
 450         }
 451         home_dir = get_lttng_home_dir();
 452         if (!home_dir) {
 453                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 454                 assert(local_apps.allowed == 0);
 455                 return -ENOENT;
 456         }
 457         local_apps.allowed = 1;
 458         snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
 459                 home_dir,
 460                 LTTNG_DEFAULT_HOME_RUNDIR,
 461                 LTTNG_UST_SOCK_FILENAME);
 462         snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
 463                 LTTNG_UST_WAIT_FILENAME,
 464                 uid);
 465         return 0;
 466 }
 467
 468 /*
 469  * Get socket timeout, in ms.
 470  * -1: wait forever. 0: don't wait. >0: timeout, in ms.
 471  */
 472 static
 473 long get_timeout(void)
 474 {
 475         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 476
 477         if (!got_timeout_env) {
 478                 str_timeout = lttng_getenv("LTTNG_UST_REGISTER_TIMEOUT");
 479                 got_timeout_env = 1;
 480         }
 481         if (str_timeout)
 482                 constructor_delay_ms = strtol(str_timeout, NULL, 10);
 483         /* All negative values are considered as "-1". */
 484         if (constructor_delay_ms < -1)
 485                 constructor_delay_ms = -1;
 486         return constructor_delay_ms;
 487 }
 488
 489 /* Timeout for notify socket send and recv. */
 490 static
 491 long get_notify_sock_timeout(void)
 492 {
 493         return get_timeout();
 494 }
 495
 496 /* Timeout for connecting to cmd and notify sockets. */
 497 static
 498 long get_connect_sock_timeout(void)
 499 {
 500         return get_timeout();
 501 }
 502
 503 /*
 504  * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
 505  */
 506 static
 507 int get_constructor_timeout(struct timespec *constructor_timeout)
 508 {
 509         long constructor_delay_ms;
 510         int ret;
 511
 512         constructor_delay_ms = get_timeout();
 513
 514         switch (constructor_delay_ms) {
 515         case -1:/* fall-through */
 516         case 0:
 517                 return constructor_delay_ms;
 518         default:
 519                 break;
 520         }
 521
 522         /*
 523          * If we are unable to find the current time, don't wait.
 524          */
 525         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 526         if (ret) {
 527                 /* Don't wait. */
 528                 return 0;
 529         }
 530         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 531         constructor_timeout->tv_nsec +=
 532                 (constructor_delay_ms % 1000UL) * 1000000UL;
 533         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 534                 constructor_timeout->tv_sec++;
 535                 constructor_timeout->tv_nsec -= 1000000000UL;
 536         }
 537         /* Timeout wait (constructor_delay_ms). */
 538         return 1;
 539 }
 540
 541 static
 542 void get_allow_blocking(void)
 543 {
 544         const char *str_allow_blocking =
 545                 lttng_getenv("LTTNG_UST_ALLOW_BLOCKING");
 546
 547         if (str_allow_blocking) {
 548                 DBG("%s environment variable is set",
 549                         "LTTNG_UST_ALLOW_BLOCKING");
 550                 lttng_ust_ringbuffer_set_allow_blocking();
 551         }
 552 }
 553
 554 static
 555 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 556 {
 557         return ustcomm_send_reg_msg(socket,
 558                 type,
 559                 CAA_BITS_PER_LONG,
 560                 lttng_alignof(uint8_t) * CHAR_BIT,
 561                 lttng_alignof(uint16_t) * CHAR_BIT,
 562                 lttng_alignof(uint32_t) * CHAR_BIT,
 563                 lttng_alignof(uint64_t) * CHAR_BIT,
 564                 lttng_alignof(unsigned long) * CHAR_BIT);
 565 }
 566
 567 static
 568 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 569 {
 570         ssize_t len;
 571
 572         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 573         switch (len) {
 574         case sizeof(*lur):
 575                 DBG("message successfully sent");
 576                 return 0;
 577         default:
 578                 if (len == -ECONNRESET) {
 579                         DBG("remote end closed connection");
 580                         return 0;
 581                 }
 582                 if (len < 0)
 583                         return len;
 584                 DBG("incorrect message size: %zd", len);
 585                 return -EINVAL;
 586         }
 587 }
 588
 589 static
 590 int handle_register_done(struct sock_info *sock_info)
 591 {
 592         int ret;
 593
 594         if (sock_info->constructor_sem_posted)
 595                 return 0;
 596         sock_info->constructor_sem_posted = 1;
 597         if (uatomic_read(&sem_count) <= 0) {
 598                 return 0;
 599         }
 600         ret = uatomic_add_return(&sem_count, -1);
 601         if (ret == 0) {
 602                 ret = sem_post(&constructor_wait);
 603                 assert(!ret);
 604         }
 605         return 0;
 606 }
 607
 608 /*
 609  * Only execute pending statedump after the constructor semaphore has
 610  * been posted by each listener thread. This means statedump will only
 611  * be performed after the "registration done" command is received from
 612  * each session daemon the application is connected to.
 613  *
 614  * This ensures we don't run into deadlock issues with the dynamic
 615  * loader mutex, which is held while the constructor is called and
 616  * waiting on the constructor semaphore. All operations requiring this
 617  * dynamic loader lock need to be postponed using this mechanism.
 618  */
 619 static
 620 void handle_pending_statedump(struct sock_info *sock_info)
 621 {
 622         int ctor_passed = sock_info->constructor_sem_posted;
 623
 624         if (ctor_passed && sock_info->statedump_pending) {
 625                 sock_info->statedump_pending = 0;
 626                 pthread_mutex_lock(&ust_fork_mutex);
 627                 lttng_handle_pending_statedump(sock_info);
 628                 pthread_mutex_unlock(&ust_fork_mutex);
 629         }
 630 }
 631
 632 static
 633 int handle_message(struct sock_info *sock_info,
 634                 int sock, struct ustcomm_ust_msg *lum)
 635 {
 636         int ret = 0;
 637         const struct lttng_ust_objd_ops *ops;
 638         struct ustcomm_ust_reply lur;
 639         union ust_args args;
 640         char ctxstr[LTTNG_UST_SYM_NAME_LEN];    /* App context string. */
 641         ssize_t len;
 642
 643         memset(&lur, 0, sizeof(lur));
 644
 645         if (ust_lock()) {
 646                 ret = -LTTNG_UST_ERR_EXITING;
 647                 goto error;
 648         }
 649
 650         ops = objd_ops(lum->handle);
 651         if (!ops) {
 652                 ret = -ENOENT;
 653                 goto error;
 654         }
 655
 656         switch (lum->cmd) {
 657         case LTTNG_UST_REGISTER_DONE:
 658                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 659                         ret = handle_register_done(sock_info);
 660                 else
 661                         ret = -EINVAL;
 662                 break;
 663         case LTTNG_UST_RELEASE:
 664                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 665                         ret = -EPERM;
 666                 else
 667                         ret = lttng_ust_objd_unref(lum->handle, 1);
 668                 break;
 669         case LTTNG_UST_FILTER:
 670         {
 671                 /* Receive filter data */
 672                 struct lttng_ust_filter_bytecode_node *bytecode;
 673
 674                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 675                         ERR("Filter data size is too large: %u bytes",
 676                                 lum->u.filter.data_size);
 677                         ret = -EINVAL;
 678                         goto error;
 679                 }
 680
 681                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 682                         ERR("Filter reloc offset %u is not within data",
 683                                 lum->u.filter.reloc_offset);
 684                         ret = -EINVAL;
 685                         goto error;
 686                 }
 687
 688                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 689                 if (!bytecode) {
 690                         ret = -ENOMEM;
 691                         goto error;
 692                 }
 693                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 694                                 lum->u.filter.data_size);
 695                 switch (len) {
 696                 case 0: /* orderly shutdown */
 697                         ret = 0;
 698                         free(bytecode);
 699                         goto error;
 700                 default:
 701                         if (len == lum->u.filter.data_size) {
 702                                 DBG("filter data received");
 703                                 break;
 704                         } else if (len < 0) {
 705                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 706                                 if (len == -ECONNRESET) {
 707                                         ERR("%s remote end closed connection", sock_info->name);
 708                                         ret = len;
 709                                         free(bytecode);
 710                                         goto error;
 711                                 }
 712                                 ret = len;
 713                                 free(bytecode);
 714                                 goto error;
 715                         } else {
 716                                 DBG("incorrect filter data message size: %zd", len);
 717                                 ret = -EINVAL;
 718                                 free(bytecode);
 719                                 goto error;
 720                         }
 721                 }
 722                 bytecode->bc.len = lum->u.filter.data_size;
 723                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 724                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 725                 if (ops->cmd) {
 726                         ret = ops->cmd(lum->handle, lum->cmd,
 727                                         (unsigned long) bytecode,
 728                                         &args, sock_info);
 729                         if (ret) {
 730                                 free(bytecode);
 731                         }
 732                         /* don't free bytecode if everything went fine. */
 733                 } else {
 734                         ret = -ENOSYS;
 735                         free(bytecode);
 736                 }
 737                 break;
 738         }
 739         case LTTNG_UST_EXCLUSION:
 740         {
 741                 /* Receive exclusion names */
 742                 struct lttng_ust_excluder_node *node;
 743                 unsigned int count;
 744
 745                 count = lum->u.exclusion.count;
 746                 if (count == 0) {
 747                         /* There are no names to read */
 748                         ret = 0;
 749                         goto error;
 750                 }
 751                 node = zmalloc(sizeof(*node) +
 752                                 count * LTTNG_UST_SYM_NAME_LEN);
 753                 if (!node) {
 754                         ret = -ENOMEM;
 755                         goto error;
 756                 }
 757                 node->excluder.count = count;
 758                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 759                                 count * LTTNG_UST_SYM_NAME_LEN);
 760                 switch (len) {
 761                 case 0: /* orderly shutdown */
 762                         ret = 0;
 763                         free(node);
 764                         goto error;
 765                 default:
 766                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 767                                 DBG("Exclusion data received");
 768                                 break;
 769                         } else if (len < 0) {
 770                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 771                                 if (len == -ECONNRESET) {
 772                                         ERR("%s remote end closed connection", sock_info->name);
 773                                         ret = len;
 774                                         free(node);
 775                                         goto error;
 776                                 }
 777                                 ret = len;
 778                                 free(node);
 779                                 goto error;
 780                         } else {
 781                                 DBG("Incorrect exclusion data message size: %zd", len);
 782                                 ret = -EINVAL;
 783                                 free(node);
 784                                 goto error;
 785                         }
 786                 }
 787                 if (ops->cmd) {
 788                         ret = ops->cmd(lum->handle, lum->cmd,
 789                                         (unsigned long) node,
 790                                         &args, sock_info);
 791                         if (ret) {
 792                                 free(node);
 793                         }
 794                         /* Don't free exclusion data if everything went fine. */
 795                 } else {
 796                         ret = -ENOSYS;
 797                         free(node);
 798                 }
 799                 break;
 800         }
 801         case LTTNG_UST_CHANNEL:
 802         {
 803                 void *chan_data;
 804                 int wakeup_fd;
 805
 806                 len = ustcomm_recv_channel_from_sessiond(sock,
 807                                 &chan_data, lum->u.channel.len,
 808                                 &wakeup_fd);
 809                 switch (len) {
 810                 case 0: /* orderly shutdown */
 811                         ret = 0;
 812                         goto error;
 813                 default:
 814                         if (len == lum->u.channel.len) {
 815                                 DBG("channel data received");
 816                                 break;
 817                         } else if (len < 0) {
 818                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 819                                 if (len == -ECONNRESET) {
 820                                         ERR("%s remote end closed connection", sock_info->name);
 821                                         ret = len;
 822                                         goto error;
 823                                 }
 824                                 ret = len;
 825                                 goto error;
 826                         } else {
 827                                 DBG("incorrect channel data message size: %zd", len);
 828                                 ret = -EINVAL;
 829                                 goto error;
 830                         }
 831                 }
 832                 args.channel.chan_data = chan_data;
 833                 args.channel.wakeup_fd = wakeup_fd;
 834                 if (ops->cmd)
 835                         ret = ops->cmd(lum->handle, lum->cmd,
 836                                         (unsigned long) &lum->u,
 837                                         &args, sock_info);
 838                 else
 839                         ret = -ENOSYS;
 840                 break;
 841         }
 842         case LTTNG_UST_STREAM:
 843         {
 844                 /* Receive shm_fd, wakeup_fd */
 845                 ret = ustcomm_recv_stream_from_sessiond(sock,
 846                         &lum->u.stream.len,
 847                         &args.stream.shm_fd,
 848                         &args.stream.wakeup_fd);
 849                 if (ret) {
 850                         goto error;
 851                 }
 852                 if (ops->cmd)
 853                         ret = ops->cmd(lum->handle, lum->cmd,
 854                                         (unsigned long) &lum->u,
 855                                         &args, sock_info);
 856                 else
 857                         ret = -ENOSYS;
 858                 break;
 859         }
 860         case LTTNG_UST_CONTEXT:
 861                 switch (lum->u.context.ctx) {
 862                 case LTTNG_UST_CONTEXT_APP_CONTEXT:
 863                 {
 864                         char *p;
 865                         size_t ctxlen, recvlen;
 866
 867                         ctxlen = strlen("$app.") + lum->u.context.u.app_ctx.provider_name_len - 1
 868                                         + strlen(":") + lum->u.context.u.app_ctx.ctx_name_len;
 869                         if (ctxlen >= LTTNG_UST_SYM_NAME_LEN) {
 870                                 ERR("Application context string length size is too large: %zu bytes",
 871                                         ctxlen);
 872                                 ret = -EINVAL;
 873                                 goto error;
 874                         }
 875                         strcpy(ctxstr, "$app.");
 876                         p = &ctxstr[strlen("$app.")];
 877                         recvlen = ctxlen - strlen("$app.");
 878                         len = ustcomm_recv_unix_sock(sock, p, recvlen);
 879                         switch (len) {
 880                         case 0: /* orderly shutdown */
 881                                 ret = 0;
 882                                 goto error;
 883                         default:
 884                                 if (len == recvlen) {
 885                                         DBG("app context data received");
 886                                         break;
 887                                 } else if (len < 0) {
 888                                         DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 889                                         if (len == -ECONNRESET) {
 890                                                 ERR("%s remote end closed connection", sock_info->name);
 891                                                 ret = len;
 892                                                 goto error;
 893                                         }
 894                                         ret = len;
 895                                         goto error;
 896                                 } else {
 897                                         DBG("incorrect app context data message size: %zd", len);
 898                                         ret = -EINVAL;
 899                                         goto error;
 900                                 }
 901                         }
 902                         /* Put : between provider and ctxname. */
 903                         p[lum->u.context.u.app_ctx.provider_name_len - 1] = ':';
 904                         args.app_context.ctxname = ctxstr;
 905                         break;
 906                 }
 907                 default:
 908                         break;
 909                 }
 910                 if (ops->cmd) {
 911                         ret = ops->cmd(lum->handle, lum->cmd,
 912                                         (unsigned long) &lum->u,
 913                                         &args, sock_info);
 914                 } else {
 915                         ret = -ENOSYS;
 916                 }
 917                 break;
 918         default:
 919                 if (ops->cmd)
 920                         ret = ops->cmd(lum->handle, lum->cmd,
 921                                         (unsigned long) &lum->u,
 922                                         &args, sock_info);
 923                 else
 924                         ret = -ENOSYS;
 925                 break;
 926         }
 927
 928         lur.handle = lum->handle;
 929         lur.cmd = lum->cmd;
 930         lur.ret_val = ret;
 931         if (ret >= 0) {
 932                 lur.ret_code = LTTNG_UST_OK;
 933         } else {
 934                 /*
 935                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 936                  * error that are not caused by the transport, except if
 937                  * we already have a more precise error message to
 938                  * report.
 939                  */
 940                 if (ret > -LTTNG_UST_ERR) {
 941                         /* Translate code to UST error. */
 942                         switch (ret) {
 943                         case -EEXIST:
 944                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 945                                 break;
 946                         case -EINVAL:
 947                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 948                                 break;
 949                         case -ENOENT:
 950                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 951                                 break;
 952                         case -EPERM:
 953                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 954                                 break;
 955                         case -ENOSYS:
 956                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 957                                 break;
 958                         default:
 959                                 lur.ret_code = -LTTNG_UST_ERR;
 960                                 break;
 961                         }
 962                 } else {
 963                         lur.ret_code = ret;
 964                 }
 965         }
 966         if (ret >= 0) {
 967                 switch (lum->cmd) {
 968                 case LTTNG_UST_TRACER_VERSION:
 969                         lur.u.version = lum->u.version;
 970                         break;
 971                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 972                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 973                         break;
 974                 }
 975         }
 976         DBG("Return value: %d", lur.ret_val);
 977
 978         ust_unlock();
 979
 980         /*
 981          * Performed delayed statedump operations outside of the UST
 982          * lock. We need to take the dynamic loader lock before we take
 983          * the UST lock internally within handle_pending_statedump().
 984           */
 985         handle_pending_statedump(sock_info);
 986
 987         if (ust_lock()) {
 988                 ret = -LTTNG_UST_ERR_EXITING;
 989                 goto error;
 990         }
 991
 992         ret = send_reply(sock, &lur);
 993         if (ret < 0) {
 994                 DBG("error sending reply");
 995                 goto error;
 996         }
 997
 998         /*
 999          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
1000          * after the reply.
1001          */
1002         if (lur.ret_code == LTTNG_UST_OK) {
1003                 switch (lum->cmd) {
1004                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
1005                         len = ustcomm_send_unix_sock(sock,
1006                                 &args.field_list.entry,
1007                                 sizeof(args.field_list.entry));
1008                         if (len < 0) {
1009                                 ret = len;
1010                                 goto error;
1011                         }
1012                         if (len != sizeof(args.field_list.entry)) {
1013                                 ret = -EINVAL;
1014                                 goto error;
1015                         }
1016                 }
1017         }
1018
1019 error:
1020         ust_unlock();
1021
1022         return ret;
1023 }
1024
1025 static
1026 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
1027 {
1028         int ret;
1029
1030         if (sock_info->root_handle != -1) {
1031                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
1032                 if (ret) {
1033                         ERR("Error unref root handle");
1034                 }
1035                 sock_info->root_handle = -1;
1036         }
1037         sock_info->constructor_sem_posted = 0;
1038
1039         /*
1040          * wait_shm_mmap, socket and notify socket are used by listener
1041          * threads outside of the ust lock, so we cannot tear them down
1042          * ourselves, because we cannot join on these threads. Leave
1043          * responsibility of cleaning up these resources to the OS
1044          * process exit.
1045          */
1046         if (exiting)
1047                 return;
1048
1049         if (sock_info->socket != -1) {
1050                 ret = ustcomm_close_unix_sock(sock_info->socket);
1051                 if (ret) {
1052                         ERR("Error closing ust cmd socket");
1053                 }
1054                 sock_info->socket = -1;
1055         }
1056         if (sock_info->notify_socket != -1) {
1057                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1058                 if (ret) {
1059                         ERR("Error closing ust notify socket");
1060                 }
1061                 sock_info->notify_socket = -1;
1062         }
1063         if (sock_info->wait_shm_mmap) {
1064                 long page_size;
1065
1066                 page_size = sysconf(_SC_PAGE_SIZE);
1067                 if (page_size <= 0) {
1068                         if (!page_size) {
1069                                 errno = EINVAL;
1070                         }
1071                         PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1072                 } else {
1073                         ret = munmap(sock_info->wait_shm_mmap, page_size);
1074                         if (ret) {
1075                                 ERR("Error unmapping wait shm");
1076                         }
1077                 }
1078                 sock_info->wait_shm_mmap = NULL;
1079         }
1080 }
1081
1082 /*
1083  * Using fork to set umask in the child process (not multi-thread safe).
1084  * We deal with the shm_open vs ftruncate race (happening when the
1085  * sessiond owns the shm and does not let everybody modify it, to ensure
1086  * safety against shm_unlink) by simply letting the mmap fail and
1087  * retrying after a few seconds.
1088  * For global shm, everybody has rw access to it until the sessiond
1089  * starts.
1090  */
1091 static
1092 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
1093 {
1094         int wait_shm_fd, ret;
1095         pid_t pid;
1096
1097         /*
1098          * Try to open read-only.
1099          */
1100         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1101         if (wait_shm_fd >= 0) {
1102                 int32_t tmp_read;
1103                 ssize_t len;
1104                 size_t bytes_read = 0;
1105
1106                 /*
1107                  * Try to read the fd. If unable to do so, try opening
1108                  * it in write mode.
1109                  */
1110                 do {
1111                         len = read(wait_shm_fd,
1112                                 &((char *) &tmp_read)[bytes_read],
1113                                 sizeof(tmp_read) - bytes_read);
1114                         if (len > 0) {
1115                                 bytes_read += len;
1116                         }
1117                 } while ((len < 0 && errno == EINTR)
1118                         || (len > 0 && bytes_read < sizeof(tmp_read)));
1119                 if (bytes_read != sizeof(tmp_read)) {
1120                         ret = close(wait_shm_fd);
1121                         if (ret) {
1122                                 ERR("close wait_shm_fd");
1123                         }
1124                         goto open_write;
1125                 }
1126                 goto end;
1127         } else if (wait_shm_fd < 0 && errno != ENOENT) {
1128                 /*
1129                  * Real-only open did not work, and it's not because the
1130                  * entry was not present. It's a failure that prohibits
1131                  * using shm.
1132                  */
1133                 ERR("Error opening shm %s", sock_info->wait_shm_path);
1134                 goto end;
1135         }
1136
1137 open_write:
1138         /*
1139          * If the open failed because the file did not exist, or because
1140          * the file was not truncated yet, try creating it ourself.
1141          */
1142         URCU_TLS(lttng_ust_nest_count)++;
1143         pid = fork();
1144         URCU_TLS(lttng_ust_nest_count)--;
1145         if (pid > 0) {
1146                 int status;
1147
1148                 /*
1149                  * Parent: wait for child to return, in which case the
1150                  * shared memory map will have been created.
1151                  */
1152                 pid = wait(&status);
1153                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
1154                         wait_shm_fd = -1;
1155                         goto end;
1156                 }
1157                 /*
1158                  * Try to open read-only again after creation.
1159                  */
1160                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1161                 if (wait_shm_fd < 0) {
1162                         /*
1163                          * Real-only open did not work. It's a failure
1164                          * that prohibits using shm.
1165                          */
1166                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1167                         goto end;
1168                 }
1169                 goto end;
1170         } else if (pid == 0) {
1171                 int create_mode;
1172
1173                 /* Child */
1174                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1175                 if (sock_info->global)
1176                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1177                 /*
1178                  * We're alone in a child process, so we can modify the
1179                  * process-wide umask.
1180                  */
1181                 umask(~create_mode);
1182                 /*
1183                  * Try creating shm (or get rw access).
1184                  * We don't do an exclusive open, because we allow other
1185                  * processes to create+ftruncate it concurrently.
1186                  */
1187                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1188                                 O_RDWR | O_CREAT, create_mode);
1189                 if (wait_shm_fd >= 0) {
1190                         ret = ftruncate(wait_shm_fd, mmap_size);
1191                         if (ret) {
1192                                 PERROR("ftruncate");
1193                                 _exit(EXIT_FAILURE);
1194                         }
1195                         _exit(EXIT_SUCCESS);
1196                 }
1197                 /*
1198                  * For local shm, we need to have rw access to accept
1199                  * opening it: this means the local sessiond will be
1200                  * able to wake us up. For global shm, we open it even
1201                  * if rw access is not granted, because the root.root
1202                  * sessiond will be able to override all rights and wake
1203                  * us up.
1204                  */
1205                 if (!sock_info->global && errno != EACCES) {
1206                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1207                         _exit(EXIT_FAILURE);
1208                 }
1209                 /*
1210                  * The shm exists, but we cannot open it RW. Report
1211                  * success.
1212                  */
1213                 _exit(EXIT_SUCCESS);
1214         } else {
1215                 return -1;
1216         }
1217 end:
1218         if (wait_shm_fd >= 0 && !sock_info->global) {
1219                 struct stat statbuf;
1220
1221                 /*
1222                  * Ensure that our user is the owner of the shm file for
1223                  * local shm. If we do not own the file, it means our
1224                  * sessiond will not have access to wake us up (there is
1225                  * probably a rogue process trying to fake our
1226                  * sessiond). Fallback to polling method in this case.
1227                  */
1228                 ret = fstat(wait_shm_fd, &statbuf);
1229                 if (ret) {
1230                         PERROR("fstat");
1231                         goto error_close;
1232                 }
1233                 if (statbuf.st_uid != getuid())
1234                         goto error_close;
1235         }
1236         return wait_shm_fd;
1237
1238 error_close:
1239         ret = close(wait_shm_fd);
1240         if (ret) {
1241                 PERROR("Error closing fd");
1242         }
1243         return -1;
1244 }
1245
1246 static
1247 char *get_map_shm(struct sock_info *sock_info)
1248 {
1249         long page_size;
1250         int wait_shm_fd, ret;
1251         char *wait_shm_mmap;
1252
1253         page_size = sysconf(_SC_PAGE_SIZE);
1254         if (page_size <= 0) {
1255                 if (!page_size) {
1256                         errno = EINVAL;
1257                 }
1258                 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1259                 goto error;
1260         }
1261
1262         lttng_ust_lock_fd_tracker();
1263         wait_shm_fd = get_wait_shm(sock_info, page_size);
1264         if (wait_shm_fd < 0) {
1265                 lttng_ust_unlock_fd_tracker();
1266                 goto error;
1267         }
1268
1269         ret = lttng_ust_add_fd_to_tracker(wait_shm_fd);
1270         if (ret < 0) {
1271                 ret = close(wait_shm_fd);
1272                 if (!ret) {
1273                         PERROR("Error closing fd");
1274                 }
1275                 lttng_ust_unlock_fd_tracker();
1276                 goto error;
1277         }
1278
1279         wait_shm_fd = ret;
1280         lttng_ust_unlock_fd_tracker();
1281
1282         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1283                   MAP_SHARED, wait_shm_fd, 0);
1284
1285         /* close shm fd immediately after taking the mmap reference */
1286         lttng_ust_lock_fd_tracker();
1287         ret = close(wait_shm_fd);
1288         if (!ret) {
1289                 lttng_ust_delete_fd_from_tracker(wait_shm_fd);
1290         } else {
1291                 PERROR("Error closing fd");
1292         }
1293         lttng_ust_unlock_fd_tracker();
1294
1295         if (wait_shm_mmap == MAP_FAILED) {
1296                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1297                 goto error;
1298         }
1299         return wait_shm_mmap;
1300
1301 error:
1302         return NULL;
1303 }
1304
1305 static
1306 void wait_for_sessiond(struct sock_info *sock_info)
1307 {
1308         if (ust_lock()) {
1309                 goto quit;
1310         }
1311         if (wait_poll_fallback) {
1312                 goto error;
1313         }
1314         if (!sock_info->wait_shm_mmap) {
1315                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1316                 if (!sock_info->wait_shm_mmap)
1317                         goto error;
1318         }
1319         ust_unlock();
1320
1321         DBG("Waiting for %s apps sessiond", sock_info->name);
1322         /* Wait for futex wakeup */
1323         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
1324                 goto end_wait;
1325
1326         while (futex_async((int32_t *) sock_info->wait_shm_mmap,
1327                         FUTEX_WAIT, 0, NULL, NULL, 0)) {
1328                 switch (errno) {
1329                 case EWOULDBLOCK:
1330                         /* Value already changed. */
1331                         goto end_wait;
1332                 case EINTR:
1333                         /* Retry if interrupted by signal. */
1334                         break;  /* Get out of switch. */
1335                 case EFAULT:
1336                         wait_poll_fallback = 1;
1337                         DBG(
1338 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1339 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1340 "Please upgrade your kernel "
1341 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1342 "mainline). LTTng-UST will use polling mode fallback.");
1343                         if (ust_debug())
1344                                 PERROR("futex");
1345                         goto end_wait;
1346                 }
1347         }
1348 end_wait:
1349         return;
1350
1351 quit:
1352         ust_unlock();
1353         return;
1354
1355 error:
1356         ust_unlock();
1357         return;
1358 }
1359
1360 /*
1361  * This thread does not allocate any resource, except within
1362  * handle_message, within mutex protection. This mutex protects against
1363  * fork and exit.
1364  * The other moment it allocates resources is at socket connection, which
1365  * is also protected by the mutex.
1366  */
1367 static
1368 void *ust_listener_thread(void *arg)
1369 {
1370         struct sock_info *sock_info = arg;
1371         int sock, ret, prev_connect_failed = 0, has_waited = 0, fd;
1372         long timeout;
1373
1374         lttng_ust_fixup_tls();
1375         /*
1376          * If available, add '-ust' to the end of this thread's
1377          * process name
1378          */
1379         ret = lttng_ust_setustprocname();
1380         if (ret) {
1381                 ERR("Unable to set UST process name");
1382         }
1383
1384         /* Restart trying to connect to the session daemon */
1385 restart:
1386         if (prev_connect_failed) {
1387                 /* Wait for sessiond availability with pipe */
1388                 wait_for_sessiond(sock_info);
1389                 if (has_waited) {
1390                         has_waited = 0;
1391                         /*
1392                          * Sleep for 5 seconds before retrying after a
1393                          * sequence of failure / wait / failure. This
1394                          * deals with a killed or broken session daemon.
1395                          */
1396                         sleep(5);
1397                 } else {
1398                         has_waited = 1;
1399                 }
1400                 prev_connect_failed = 0;
1401         }
1402
1403         if (ust_lock()) {
1404                 goto quit;
1405         }
1406
1407         if (sock_info->socket != -1) {
1408                 /* FD tracker is updated by ustcomm_close_unix_sock() */
1409                 ret = ustcomm_close_unix_sock(sock_info->socket);
1410                 if (ret) {
1411                         ERR("Error closing %s ust cmd socket",
1412                                 sock_info->name);
1413                 }
1414                 sock_info->socket = -1;
1415         }
1416         if (sock_info->notify_socket != -1) {
1417                 /* FD tracker is updated by ustcomm_close_unix_sock() */
1418                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1419                 if (ret) {
1420                         ERR("Error closing %s ust notify socket",
1421                                 sock_info->name);
1422                 }
1423                 sock_info->notify_socket = -1;
1424         }
1425
1426
1427         /*
1428          * Register. We need to perform both connect and sending
1429          * registration message before doing the next connect otherwise
1430          * we may reach unix socket connect queue max limits and block
1431          * on the 2nd connect while the session daemon is awaiting the
1432          * first connect registration message.
1433          */
1434         /* Connect cmd socket */
1435         lttng_ust_lock_fd_tracker();
1436         ret = ustcomm_connect_unix_sock(sock_info->sock_path,
1437                 get_connect_sock_timeout());
1438         if (ret < 0) {
1439                 lttng_ust_unlock_fd_tracker();
1440                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1441                 prev_connect_failed = 1;
1442
1443                 /*
1444                  * If we cannot find the sessiond daemon, don't delay
1445                  * constructor execution.
1446                  */
1447                 ret = handle_register_done(sock_info);
1448                 assert(!ret);
1449                 ust_unlock();
1450                 goto restart;
1451         }
1452         fd = ret;
1453         ret = lttng_ust_add_fd_to_tracker(fd);
1454         if (ret < 0) {
1455                 ret = close(fd);
1456                 if (ret) {
1457                         PERROR("close on sock_info->socket");
1458                 }
1459                 ret = -1;
1460                 lttng_ust_unlock_fd_tracker();
1461                 ust_unlock();
1462                 goto quit;
1463         }
1464
1465         sock_info->socket = ret;
1466         lttng_ust_unlock_fd_tracker();
1467
1468         ust_unlock();
1469         /*
1470          * Unlock/relock ust lock because connect is blocking (with
1471          * timeout). Don't delay constructors on the ust lock for too
1472          * long.
1473          */
1474         if (ust_lock()) {
1475                 goto quit;
1476         }
1477
1478         /*
1479          * Create only one root handle per listener thread for the whole
1480          * process lifetime, so we ensure we get ID which is statically
1481          * assigned to the root handle.
1482          */
1483         if (sock_info->root_handle == -1) {
1484                 ret = lttng_abi_create_root_handle();
1485                 if (ret < 0) {
1486                         ERR("Error creating root handle");
1487                         goto quit;
1488                 }
1489                 sock_info->root_handle = ret;
1490         }
1491
1492         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1493         if (ret < 0) {
1494                 ERR("Error registering to %s ust cmd socket",
1495                         sock_info->name);
1496                 prev_connect_failed = 1;
1497                 /*
1498                  * If we cannot register to the sessiond daemon, don't
1499                  * delay constructor execution.
1500                  */
1501                 ret = handle_register_done(sock_info);
1502                 assert(!ret);
1503                 ust_unlock();
1504                 goto restart;
1505         }
1506
1507         ust_unlock();
1508         /*
1509          * Unlock/relock ust lock because connect is blocking (with
1510          * timeout). Don't delay constructors on the ust lock for too
1511          * long.
1512          */
1513         if (ust_lock()) {
1514                 goto quit;
1515         }
1516
1517         /* Connect notify socket */
1518         lttng_ust_lock_fd_tracker();
1519         ret = ustcomm_connect_unix_sock(sock_info->sock_path,
1520                 get_connect_sock_timeout());
1521         if (ret < 0) {
1522                 lttng_ust_unlock_fd_tracker();
1523                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1524                 prev_connect_failed = 1;
1525
1526                 /*
1527                  * If we cannot find the sessiond daemon, don't delay
1528                  * constructor execution.
1529                  */
1530                 ret = handle_register_done(sock_info);
1531                 assert(!ret);
1532                 ust_unlock();
1533                 goto restart;
1534         }
1535
1536         fd = ret;
1537         ret = lttng_ust_add_fd_to_tracker(fd);
1538         if (ret < 0) {
1539                 ret = close(fd);
1540                 if (ret) {
1541                         PERROR("close on sock_info->notify_socket");
1542                 }
1543                 ret = -1;
1544                 lttng_ust_unlock_fd_tracker();
1545                 ust_unlock();
1546                 goto quit;
1547         }
1548
1549         sock_info->notify_socket = ret;
1550         lttng_ust_unlock_fd_tracker();
1551
1552         ust_unlock();
1553         /*
1554          * Unlock/relock ust lock because connect is blocking (with
1555          * timeout). Don't delay constructors on the ust lock for too
1556          * long.
1557          */
1558         if (ust_lock()) {
1559                 goto quit;
1560         }
1561
1562         timeout = get_notify_sock_timeout();
1563         if (timeout >= 0) {
1564                 /*
1565                  * Give at least 10ms to sessiond to reply to
1566                  * notifications.
1567                  */
1568                 if (timeout < 10)
1569                         timeout = 10;
1570                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1571                                 timeout);
1572                 if (ret < 0) {
1573                         WARN("Error setting socket receive timeout");
1574                 }
1575                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1576                                 timeout);
1577                 if (ret < 0) {
1578                         WARN("Error setting socket send timeout");
1579                 }
1580         } else if (timeout < -1) {
1581                 WARN("Unsupported timeout value %ld", timeout);
1582         }
1583
1584         ret = register_to_sessiond(sock_info->notify_socket,
1585                         USTCTL_SOCKET_NOTIFY);
1586         if (ret < 0) {
1587                 ERR("Error registering to %s ust notify socket",
1588                         sock_info->name);
1589                 prev_connect_failed = 1;
1590                 /*
1591                  * If we cannot register to the sessiond daemon, don't
1592                  * delay constructor execution.
1593                  */
1594                 ret = handle_register_done(sock_info);
1595                 assert(!ret);
1596                 ust_unlock();
1597                 goto restart;
1598         }
1599         sock = sock_info->socket;
1600
1601         ust_unlock();
1602
1603         for (;;) {
1604                 ssize_t len;
1605                 struct ustcomm_ust_msg lum;
1606
1607                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1608                 switch (len) {
1609                 case 0: /* orderly shutdown */
1610                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1611                         if (ust_lock()) {
1612                                 goto quit;
1613                         }
1614                         /*
1615                          * Either sessiond has shutdown or refused us by closing the socket.
1616                          * In either case, we don't want to delay construction execution,
1617                          * and we need to wait before retry.
1618                          */
1619                         prev_connect_failed = 1;
1620                         /*
1621                          * If we cannot register to the sessiond daemon, don't
1622                          * delay constructor execution.
1623                          */
1624                         ret = handle_register_done(sock_info);
1625                         assert(!ret);
1626                         ust_unlock();
1627                         goto end;
1628                 case sizeof(lum):
1629                         print_cmd(lum.cmd, lum.handle);
1630                         ret = handle_message(sock_info, sock, &lum);
1631                         if (ret) {
1632                                 ERR("Error handling message for %s socket",
1633                                         sock_info->name);
1634                                 /*
1635                                  * Close socket if protocol error is
1636                                  * detected.
1637                                  */
1638                                 goto end;
1639                         }
1640                         continue;
1641                 default:
1642                         if (len < 0) {
1643                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1644                         } else {
1645                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1646                         }
1647                         if (len == -ECONNRESET) {
1648                                 DBG("%s remote end closed connection", sock_info->name);
1649                                 goto end;
1650                         }
1651                         goto end;
1652                 }
1653
1654         }
1655 end:
1656         if (ust_lock()) {
1657                 goto quit;
1658         }
1659         /* Cleanup socket handles before trying to reconnect */
1660         lttng_ust_objd_table_owner_cleanup(sock_info);
1661         ust_unlock();
1662         goto restart;   /* try to reconnect */
1663
1664 quit:
1665         ust_unlock();
1666
1667         pthread_mutex_lock(&ust_exit_mutex);
1668         sock_info->thread_active = 0;
1669         pthread_mutex_unlock(&ust_exit_mutex);
1670         return NULL;
1671 }
1672
1673 /*
1674  * Weak symbol to call when the ust malloc wrapper is not loaded.
1675  */
1676 __attribute__((weak))
1677 void lttng_ust_malloc_wrapper_init(void)
1678 {
1679 }
1680
1681 /*
1682  * sessiond monitoring thread: monitor presence of global and per-user
1683  * sessiond by polling the application common named pipe.
1684  */
1685 void __attribute__((constructor)) lttng_ust_init(void)
1686 {
1687         struct timespec constructor_timeout;
1688         sigset_t sig_all_blocked, orig_parent_mask;
1689         pthread_attr_t thread_attr;
1690         int timeout_mode;
1691         int ret;
1692         void *handle;
1693
1694         if (uatomic_xchg(&initialized, 1) == 1)
1695                 return;
1696
1697         /*
1698          * Fixup interdependency between TLS fixup mutex (which happens
1699          * to be the dynamic linker mutex) and ust_lock, taken within
1700          * the ust lock.
1701          */
1702         lttng_ust_fixup_tls();
1703
1704         lttng_ust_loaded = 1;
1705
1706         /*
1707          * We need to ensure that the liblttng-ust library is not unloaded to avoid
1708          * the unloading of code used by the ust_listener_threads as we can not
1709          * reliably know when they exited. To do that, manually load
1710          * liblttng-ust.so to increment the dynamic loader's internal refcount for
1711          * this library so it never becomes zero, thus never gets unloaded from the
1712          * address space of the process. Since we are already running in the
1713          * constructor of the LTTNG_UST_LIB_SO_NAME library, calling dlopen will
1714          * simply increment the refcount and no additionnal work is needed by the
1715          * dynamic loader as the shared library is already loaded in the address
1716          * space. As a safe guard, we use the RTLD_NODELETE flag to prevent
1717          * unloading of the UST library if its refcount becomes zero (which should
1718          * never happen). Do the return value check but discard the handle at the
1719          * end of the function as it's not needed.
1720          */
1721         handle = dlopen(LTTNG_UST_LIB_SO_NAME, RTLD_LAZY | RTLD_NODELETE);
1722         if (!handle) {
1723                 ERR("dlopen of liblttng-ust shared library (%s).", LTTNG_UST_LIB_SO_NAME);
1724         }
1725
1726         /*
1727          * We want precise control over the order in which we construct
1728          * our sub-libraries vs starting to receive commands from
1729          * sessiond (otherwise leading to errors when trying to create
1730          * sessiond before the init functions are completed).
1731          */
1732         init_usterr();
1733         lttng_ust_getenv_init();        /* Needs init_usterr() to be completed. */
1734         init_tracepoint();
1735         lttng_ust_init_fd_tracker();
1736         lttng_ust_clock_init();
1737         lttng_ust_getcpu_init();
1738         lttng_ust_statedump_init();
1739         lttng_ring_buffer_metadata_client_init();
1740         lttng_ring_buffer_client_overwrite_init();
1741         lttng_ring_buffer_client_overwrite_rt_init();
1742         lttng_ring_buffer_client_discard_init();
1743         lttng_ring_buffer_client_discard_rt_init();
1744         lttng_perf_counter_init();
1745         /*
1746          * Invoke ust malloc wrapper init before starting other threads.
1747          */
1748         lttng_ust_malloc_wrapper_init();
1749
1750         timeout_mode = get_constructor_timeout(&constructor_timeout);
1751
1752         get_allow_blocking();
1753
1754         ret = sem_init(&constructor_wait, 0, 0);
1755         if (ret) {
1756                 PERROR("sem_init");
1757         }
1758
1759         ret = setup_local_apps();
1760         if (ret) {
1761                 DBG("local apps setup returned %d", ret);
1762         }
1763
1764         /* A new thread created by pthread_create inherits the signal mask
1765          * from the parent. To avoid any signal being received by the
1766          * listener thread, we block all signals temporarily in the parent,
1767          * while we create the listener thread.
1768          */
1769         sigfillset(&sig_all_blocked);
1770         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1771         if (ret) {
1772                 ERR("pthread_sigmask: %s", strerror(ret));
1773         }
1774
1775         ret = pthread_attr_init(&thread_attr);
1776         if (ret) {
1777                 ERR("pthread_attr_init: %s", strerror(ret));
1778         }
1779         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1780         if (ret) {
1781                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1782         }
1783
1784         pthread_mutex_lock(&ust_exit_mutex);
1785         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1786                         ust_listener_thread, &global_apps);
1787         if (ret) {
1788                 ERR("pthread_create global: %s", strerror(ret));
1789         }
1790         global_apps.thread_active = 1;
1791         pthread_mutex_unlock(&ust_exit_mutex);
1792
1793         if (local_apps.allowed) {
1794                 pthread_mutex_lock(&ust_exit_mutex);
1795                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1796                                 ust_listener_thread, &local_apps);
1797                 if (ret) {
1798                         ERR("pthread_create local: %s", strerror(ret));
1799                 }
1800                 local_apps.thread_active = 1;
1801                 pthread_mutex_unlock(&ust_exit_mutex);
1802         } else {
1803                 handle_register_done(&local_apps);
1804         }
1805         ret = pthread_attr_destroy(&thread_attr);
1806         if (ret) {
1807                 ERR("pthread_attr_destroy: %s", strerror(ret));
1808         }
1809
1810         /* Restore original signal mask in parent */
1811         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1812         if (ret) {
1813                 ERR("pthread_sigmask: %s", strerror(ret));
1814         }
1815
1816         switch (timeout_mode) {
1817         case 1: /* timeout wait */
1818                 do {
1819                         ret = sem_timedwait(&constructor_wait,
1820                                         &constructor_timeout);
1821                 } while (ret < 0 && errno == EINTR);
1822                 if (ret < 0) {
1823                         switch (errno) {
1824                         case ETIMEDOUT:
1825                                 ERR("Timed out waiting for lttng-sessiond");
1826                                 break;
1827                         case EINVAL:
1828                                 PERROR("sem_timedwait");
1829                                 break;
1830                         default:
1831                                 ERR("Unexpected error \"%s\" returned by sem_timedwait",
1832                                         strerror(errno));
1833                         }
1834                 }
1835                 break;
1836         case -1:/* wait forever */
1837                 do {
1838                         ret = sem_wait(&constructor_wait);
1839                 } while (ret < 0 && errno == EINTR);
1840                 if (ret < 0) {
1841                         switch (errno) {
1842                         case EINVAL:
1843                                 PERROR("sem_wait");
1844                                 break;
1845                         default:
1846                                 ERR("Unexpected error \"%s\" returned by sem_wait",
1847                                         strerror(errno));
1848                         }
1849                 }
1850                 break;
1851         case 0: /* no timeout */
1852                 break;
1853         }
1854 }
1855
1856 static
1857 void lttng_ust_cleanup(int exiting)
1858 {
1859         cleanup_sock_info(&global_apps, exiting);
1860         cleanup_sock_info(&local_apps, exiting);
1861         local_apps.allowed = 0;
1862         /*
1863          * The teardown in this function all affect data structures
1864          * accessed under the UST lock by the listener thread. This
1865          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1866          * that none of these threads are accessing this data at this
1867          * point.
1868          */
1869         lttng_ust_abi_exit();
1870         lttng_ust_events_exit();
1871         lttng_perf_counter_exit();
1872         lttng_ring_buffer_client_discard_rt_exit();
1873         lttng_ring_buffer_client_discard_exit();
1874         lttng_ring_buffer_client_overwrite_rt_exit();
1875         lttng_ring_buffer_client_overwrite_exit();
1876         lttng_ring_buffer_metadata_client_exit();
1877         lttng_ust_statedump_destroy();
1878         exit_tracepoint();
1879         if (!exiting) {
1880                 /* Reinitialize values for fork */
1881                 sem_count = 2;
1882                 lttng_ust_comm_should_quit = 0;
1883                 initialized = 0;
1884         }
1885 }
1886
1887 void __attribute__((destructor)) lttng_ust_exit(void)
1888 {
1889         int ret;
1890
1891         /*
1892          * Using pthread_cancel here because:
1893          * A) we don't want to hang application teardown.
1894          * B) the thread is not allocating any resource.
1895          */
1896
1897         /*
1898          * Require the communication thread to quit. Synchronize with
1899          * mutexes to ensure it is not in a mutex critical section when
1900          * pthread_cancel is later called.
1901          */
1902         ust_lock_nocheck();
1903         lttng_ust_comm_should_quit = 1;
1904         ust_unlock();
1905
1906         pthread_mutex_lock(&ust_exit_mutex);
1907         /* cancel threads */
1908         if (global_apps.thread_active) {
1909                 ret = pthread_cancel(global_apps.ust_listener);
1910                 if (ret) {
1911                         ERR("Error cancelling global ust listener thread: %s",
1912                                 strerror(ret));
1913                 } else {
1914                         global_apps.thread_active = 0;
1915                 }
1916         }
1917         if (local_apps.thread_active) {
1918                 ret = pthread_cancel(local_apps.ust_listener);
1919                 if (ret) {
1920                         ERR("Error cancelling local ust listener thread: %s",
1921                                 strerror(ret));
1922                 } else {
1923                         local_apps.thread_active = 0;
1924                 }
1925         }
1926         pthread_mutex_unlock(&ust_exit_mutex);
1927
1928         /*
1929          * Do NOT join threads: use of sys_futex makes it impossible to
1930          * join the threads without using async-cancel, but async-cancel
1931          * is delivered by a signal, which could hit the target thread
1932          * anywhere in its code path, including while the ust_lock() is
1933          * held, causing a deadlock for the other thread. Let the OS
1934          * cleanup the threads if there are stalled in a syscall.
1935          */
1936         lttng_ust_cleanup(1);
1937 }
1938
1939 /*
1940  * We exclude the worker threads across fork and clone (except
1941  * CLONE_VM), because these system calls only keep the forking thread
1942  * running in the child.  Therefore, we don't want to call fork or clone
1943  * in the middle of an tracepoint or ust tracing state modification.
1944  * Holding this mutex protects these structures across fork and clone.
1945  */
1946 void ust_before_fork(sigset_t *save_sigset)
1947 {
1948         /*
1949          * Disable signals. This is to avoid that the child intervenes
1950          * before it is properly setup for tracing. It is safer to
1951          * disable all signals, because then we know we are not breaking
1952          * anything by restoring the original mask.
1953          */
1954         sigset_t all_sigs;
1955         int ret;
1956
1957         /* Fixup lttng-ust TLS. */
1958         lttng_ust_fixup_tls();
1959
1960         if (URCU_TLS(lttng_ust_nest_count))
1961                 return;
1962         /* Disable signals */
1963         sigfillset(&all_sigs);
1964         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1965         if (ret == -1) {
1966                 PERROR("sigprocmask");
1967         }
1968
1969         pthread_mutex_lock(&ust_fork_mutex);
1970
1971         ust_lock_nocheck();
1972         rcu_bp_before_fork();
1973 }
1974
1975 static void ust_after_fork_common(sigset_t *restore_sigset)
1976 {
1977         int ret;
1978
1979         DBG("process %d", getpid());
1980         ust_unlock();
1981
1982         pthread_mutex_unlock(&ust_fork_mutex);
1983
1984         /* Restore signals */
1985         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1986         if (ret == -1) {
1987                 PERROR("sigprocmask");
1988         }
1989 }
1990
1991 void ust_after_fork_parent(sigset_t *restore_sigset)
1992 {
1993         if (URCU_TLS(lttng_ust_nest_count))
1994                 return;
1995         DBG("process %d", getpid());
1996         rcu_bp_after_fork_parent();
1997         /* Release mutexes and reenable signals */
1998         ust_after_fork_common(restore_sigset);
1999 }
2000
2001 /*
2002  * After fork, in the child, we need to cleanup all the leftover state,
2003  * except the worker thread which already magically disappeared thanks
2004  * to the weird Linux fork semantics. After tyding up, we call
2005  * lttng_ust_init() again to start over as a new PID.
2006  *
2007  * This is meant for forks() that have tracing in the child between the
2008  * fork and following exec call (if there is any).
2009  */
2010 void ust_after_fork_child(sigset_t *restore_sigset)
2011 {
2012         if (URCU_TLS(lttng_ust_nest_count))
2013                 return;
2014         lttng_context_vpid_reset();
2015         lttng_context_vtid_reset();
2016         lttng_context_procname_reset();
2017         DBG("process %d", getpid());
2018         /* Release urcu mutexes */
2019         rcu_bp_after_fork_child();
2020         lttng_ust_cleanup(0);
2021         /* Release mutexes and reenable signals */
2022         ust_after_fork_common(restore_sigset);
2023         lttng_ust_init();
2024 }
2025
2026 void lttng_ust_sockinfo_session_enabled(void *owner)
2027 {
2028         struct sock_info *sock_info = owner;
2029         sock_info->statedump_pending = 1;
2030 }