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