liblttng-ust/lttng-ust-comm.c

   1 /*
   2  * lttng-ust-comm.c
   3  *
   4  * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
   5  * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   6  *
   7  * This library is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; only
  10  * version 2.1 of the License.
  11  *
  12  * This library is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with this library; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #define _LGPL_SOURCE
  23 #define _GNU_SOURCE
  24 #include <sys/types.h>
  25 #include <sys/socket.h>
  26 #include <sys/mman.h>
  27 #include <sys/stat.h>
  28 #include <sys/types.h>
  29 #include <sys/wait.h>
  30 #include <fcntl.h>
  31 #include <unistd.h>
  32 #include <errno.h>
  33 #include <pthread.h>
  34 #include <semaphore.h>
  35 #include <time.h>
  36 #include <assert.h>
  37 #include <signal.h>
  38 #include <urcu/uatomic.h>
  39 #include <urcu/futex.h>
  40 #include <urcu/compiler.h>
  41
  42 #include <lttng/ust-events.h>
  43 #include <lttng/ust-abi.h>
  44 #include <lttng/ust.h>
  45 #include <lttng/ust-error.h>
  46 #include <lttng/ust-ctl.h>
  47 #include <urcu/tls-compat.h>
  48 #include <ust-comm.h>
  49 #include <usterr-signal-safe.h>
  50 #include <helper.h>
  51 #include "tracepoint-internal.h"
  52 #include "lttng-tracer-core.h"
  53 #include "compat.h"
  54 #include "../libringbuffer/tlsfixup.h"
  55 #include "lttng-ust-statedump.h"
  56 #include "clock.h"
  57 #include "../libringbuffer/getcpu.h"
  58 #include "getenv.h"
  59
  60 /*
  61  * Has lttng ust comm constructor been called ?
  62  */
  63 static int initialized;
  64
  65 /*
  66  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  67  * Held when handling a command, also held by fork() to deal with
  68  * removal of threads, and by exit path.
  69  *
  70  * The UST lock is the centralized mutex across UST tracing control and
  71  * probe registration.
  72  *
  73  * ust_exit_mutex must never nest in ust_mutex.
  74  *
  75  * ust_fork_mutex must never nest in ust_mutex.
  76  *
  77  * ust_mutex_nest is a per-thread nesting counter, allowing the perf
  78  * counter lazy initialization called by events within the statedump,
  79  * which traces while the ust_mutex is held.
  80  *
  81  * ust_lock nests within the dynamic loader lock (within glibc) because
  82  * it is taken within the library constructor.
  83  */
  84 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
  85
  86 /* Allow nesting the ust_mutex within the same thread. */
  87 static DEFINE_URCU_TLS(int, ust_mutex_nest);
  88
  89 /*
  90  * ust_exit_mutex protects thread_active variable wrt thread exit. It
  91  * cannot be done by ust_mutex because pthread_cancel(), which takes an
  92  * internal libc lock, cannot nest within ust_mutex.
  93  *
  94  * It never nests within a ust_mutex.
  95  */
  96 static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
  97
  98 /*
  99  * ust_fork_mutex protects base address statedump tracing against forks. It
 100  * prevents the dynamic loader lock to be taken (by base address statedump
 101  * tracing) while a fork is happening, thus preventing deadlock issues with
 102  * the dynamic loader lock.
 103  */
 104 static pthread_mutex_t ust_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
 105
 106 /* Should the ust comm thread quit ? */
 107 static int lttng_ust_comm_should_quit;
 108
 109 /*
 110  * Return 0 on success, -1 if should quit.
 111  * The lock is taken in both cases.
 112  * Signal-safe.
 113  */
 114 int ust_lock(void)
 115 {
 116         sigset_t sig_all_blocked, orig_mask;
 117         int ret, oldstate;
 118
 119         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 120         if (ret) {
 121                 ERR("pthread_setcancelstate: %s", strerror(ret));
 122         }
 123         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 124                 ERR("pthread_setcancelstate: unexpected oldstate");
 125         }
 126         sigfillset(&sig_all_blocked);
 127         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 128         if (ret) {
 129                 ERR("pthread_sigmask: %s", strerror(ret));
 130         }
 131         if (!URCU_TLS(ust_mutex_nest)++)
 132                 pthread_mutex_lock(&ust_mutex);
 133         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 134         if (ret) {
 135                 ERR("pthread_sigmask: %s", strerror(ret));
 136         }
 137         if (lttng_ust_comm_should_quit) {
 138                 return -1;
 139         } else {
 140                 return 0;
 141         }
 142 }
 143
 144 /*
 145  * ust_lock_nocheck() can be used in constructors/destructors, because
 146  * they are already nested within the dynamic loader lock, and therefore
 147  * have exclusive access against execution of liblttng-ust destructor.
 148  * Signal-safe.
 149  */
 150 void ust_lock_nocheck(void)
 151 {
 152         sigset_t sig_all_blocked, orig_mask;
 153         int ret, oldstate;
 154
 155         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 156         if (ret) {
 157                 ERR("pthread_setcancelstate: %s", strerror(ret));
 158         }
 159         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 160                 ERR("pthread_setcancelstate: unexpected oldstate");
 161         }
 162         sigfillset(&sig_all_blocked);
 163         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 164         if (ret) {
 165                 ERR("pthread_sigmask: %s", strerror(ret));
 166         }
 167         if (!URCU_TLS(ust_mutex_nest)++)
 168                 pthread_mutex_lock(&ust_mutex);
 169         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 170         if (ret) {
 171                 ERR("pthread_sigmask: %s", strerror(ret));
 172         }
 173 }
 174
 175 /*
 176  * Signal-safe.
 177  */
 178 void ust_unlock(void)
 179 {
 180         sigset_t sig_all_blocked, orig_mask;
 181         int ret, oldstate;
 182
 183         sigfillset(&sig_all_blocked);
 184         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 185         if (ret) {
 186                 ERR("pthread_sigmask: %s", strerror(ret));
 187         }
 188         if (!--URCU_TLS(ust_mutex_nest))
 189                 pthread_mutex_unlock(&ust_mutex);
 190         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 191         if (ret) {
 192                 ERR("pthread_sigmask: %s", strerror(ret));
 193         }
 194         ret = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &oldstate);
 195         if (ret) {
 196                 ERR("pthread_setcancelstate: %s", strerror(ret));
 197         }
 198         if (oldstate != PTHREAD_CANCEL_DISABLE) {
 199                 ERR("pthread_setcancelstate: unexpected oldstate");
 200         }
 201 }
 202
 203 /*
 204  * Wait for either of these before continuing to the main
 205  * program:
 206  * - the register_done message from sessiond daemon
 207  *   (will let the sessiond daemon enable sessions before main
 208  *   starts.)
 209  * - sessiond daemon is not reachable.
 210  * - timeout (ensuring applications are resilient to session
 211  *   daemon problems).
 212  */
 213 static sem_t constructor_wait;
 214 /*
 215  * Doing this for both the global and local sessiond.
 216  */
 217 static int sem_count = { 2 };
 218
 219 /*
 220  * Counting nesting within lttng-ust. Used to ensure that calling fork()
 221  * from liblttng-ust does not execute the pre/post fork handlers.
 222  */
 223 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
 224
 225 /*
 226  * Info about socket and associated listener thread.
 227  */
 228 struct sock_info {
 229         const char *name;
 230         pthread_t ust_listener; /* listener thread */
 231         int root_handle;
 232         int constructor_sem_posted;
 233         int allowed;
 234         int global;
 235         int thread_active;
 236
 237         char sock_path[PATH_MAX];
 238         int socket;
 239         int notify_socket;
 240
 241         char wait_shm_path[PATH_MAX];
 242         char *wait_shm_mmap;
 243         /* Keep track of lazy state dump not performed yet. */
 244         int statedump_pending;
 245 };
 246
 247 /* Socket from app (connect) to session daemon (listen) for communication */
 248 struct sock_info global_apps = {
 249         .name = "global",
 250         .global = 1,
 251
 252         .root_handle = -1,
 253         .allowed = 1,
 254         .thread_active = 0,
 255
 256         .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
 257         .socket = -1,
 258         .notify_socket = -1,
 259
 260         .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 261
 262         .statedump_pending = 0,
 263 };
 264
 265 /* TODO: allow global_apps_sock_path override */
 266
 267 struct sock_info local_apps = {
 268         .name = "local",
 269         .global = 0,
 270         .root_handle = -1,
 271         .allowed = 0,   /* Check setuid bit first */
 272         .thread_active = 0,
 273
 274         .socket = -1,
 275         .notify_socket = -1,
 276
 277         .statedump_pending = 0,
 278 };
 279
 280 static int wait_poll_fallback;
 281
 282 static const char *cmd_name_mapping[] = {
 283         [ LTTNG_UST_RELEASE ] = "Release",
 284         [ LTTNG_UST_SESSION ] = "Create Session",
 285         [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
 286
 287         [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
 288         [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
 289         [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
 290         [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
 291
 292         /* Session FD commands */
 293         [ LTTNG_UST_CHANNEL ] = "Create Channel",
 294         [ LTTNG_UST_SESSION_START ] = "Start Session",
 295         [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
 296
 297         /* Channel FD commands */
 298         [ LTTNG_UST_STREAM ] = "Create Stream",
 299         [ LTTNG_UST_EVENT ] = "Create Event",
 300
 301         /* Event and Channel FD commands */
 302         [ LTTNG_UST_CONTEXT ] = "Create Context",
 303         [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
 304
 305         /* Event, Channel and Session commands */
 306         [ LTTNG_UST_ENABLE ] = "Enable",
 307         [ LTTNG_UST_DISABLE ] = "Disable",
 308
 309         /* Tracepoint list commands */
 310         [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
 311         [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
 312
 313         /* Event FD commands */
 314         [ LTTNG_UST_FILTER ] = "Create Filter",
 315         [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
 316 };
 317
 318 static const char *str_timeout;
 319 static int got_timeout_env;
 320
 321 extern void lttng_ring_buffer_client_overwrite_init(void);
 322 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 323 extern void lttng_ring_buffer_client_discard_init(void);
 324 extern void lttng_ring_buffer_client_discard_rt_init(void);
 325 extern void lttng_ring_buffer_metadata_client_init(void);
 326 extern void lttng_ring_buffer_client_overwrite_exit(void);
 327 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 328 extern void lttng_ring_buffer_client_discard_exit(void);
 329 extern void lttng_ring_buffer_client_discard_rt_exit(void);
 330 extern void lttng_ring_buffer_metadata_client_exit(void);
 331
 332 ssize_t lttng_ust_read(int fd, void *buf, size_t len)
 333 {
 334         ssize_t ret;
 335         size_t copied = 0, to_copy = len;
 336
 337         do {
 338                 ret = read(fd, buf + copied, to_copy);
 339                 if (ret > 0) {
 340                         copied += ret;
 341                         to_copy -= ret;
 342                 }
 343         } while ((ret > 0 && to_copy > 0)
 344                 || (ret < 0 && errno == EINTR));
 345         if (ret > 0) {
 346                 ret = copied;
 347         }
 348         return ret;
 349 }
 350 /*
 351  * Returns the HOME directory path. Caller MUST NOT free(3) the returned
 352  * pointer.
 353  */
 354 static
 355 const char *get_lttng_home_dir(void)
 356 {
 357        const char *val;
 358
 359        val = (const char *) lttng_secure_getenv("LTTNG_HOME");
 360        if (val != NULL) {
 361                return val;
 362        }
 363        return (const char *) lttng_secure_getenv("HOME");
 364 }
 365
 366 /*
 367  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 368  */
 369 static
 370 void lttng_fixup_nest_count_tls(void)
 371 {
 372         asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 373 }
 374
 375 static
 376 void lttng_fixup_ust_mutex_nest_tls(void)
 377 {
 378         asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest)));
 379 }
 380
 381 /*
 382  * Fixup urcu bp TLS.
 383  */
 384 static
 385 void lttng_fixup_urcu_bp_tls(void)
 386 {
 387         rcu_read_lock();
 388         rcu_read_unlock();
 389 }
 390
 391 int lttng_get_notify_socket(void *owner)
 392 {
 393         struct sock_info *info = owner;
 394
 395         return info->notify_socket;
 396 }
 397
 398 static
 399 void print_cmd(int cmd, int handle)
 400 {
 401         const char *cmd_name = "Unknown";
 402
 403         if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
 404                         && cmd_name_mapping[cmd]) {
 405                 cmd_name = cmd_name_mapping[cmd];
 406         }
 407         DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
 408                 cmd_name, cmd,
 409                 lttng_ust_obj_get_name(handle), handle);
 410 }
 411
 412 static
 413 int setup_local_apps(void)
 414 {
 415         const char *home_dir;
 416         uid_t uid;
 417
 418         uid = getuid();
 419         /*
 420          * Disallow per-user tracing for setuid binaries.
 421          */
 422         if (uid != geteuid()) {
 423                 assert(local_apps.allowed == 0);
 424                 return 0;
 425         }
 426         home_dir = get_lttng_home_dir();
 427         if (!home_dir) {
 428                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 429                 assert(local_apps.allowed == 0);
 430                 return -ENOENT;
 431         }
 432         local_apps.allowed = 1;
 433         snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
 434                 home_dir,
 435                 LTTNG_DEFAULT_HOME_RUNDIR,
 436                 LTTNG_UST_SOCK_FILENAME);
 437         snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
 438                 LTTNG_UST_WAIT_FILENAME,
 439                 uid);
 440         return 0;
 441 }
 442
 443 /*
 444  * Get notify_sock timeout, in ms.
 445  * -1: wait forever. 0: don't wait. >0: timeout, in ms.
 446  */
 447 static
 448 long get_timeout(void)
 449 {
 450         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 451
 452         if (!got_timeout_env) {
 453                 str_timeout = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 454                 got_timeout_env = 1;
 455         }
 456         if (str_timeout)
 457                 constructor_delay_ms = strtol(str_timeout, NULL, 10);
 458         return constructor_delay_ms;
 459 }
 460
 461 static
 462 long get_notify_sock_timeout(void)
 463 {
 464         return get_timeout();
 465 }
 466
 467 /*
 468  * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
 469  */
 470 static
 471 int get_constructor_timeout(struct timespec *constructor_timeout)
 472 {
 473         long constructor_delay_ms;
 474         int ret;
 475
 476         constructor_delay_ms = get_timeout();
 477
 478         switch (constructor_delay_ms) {
 479         case -1:/* fall-through */
 480         case 0:
 481                 return constructor_delay_ms;
 482         default:
 483                 break;
 484         }
 485
 486         /*
 487          * If we are unable to find the current time, don't wait.
 488          */
 489         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 490         if (ret) {
 491                 /* Don't wait. */
 492                 return 0;
 493         }
 494         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 495         constructor_timeout->tv_nsec +=
 496                 (constructor_delay_ms % 1000UL) * 1000000UL;
 497         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 498                 constructor_timeout->tv_sec++;
 499                 constructor_timeout->tv_nsec -= 1000000000UL;
 500         }
 501         /* Timeout wait (constructor_delay_ms). */
 502         return 1;
 503 }
 504
 505 static
 506 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 507 {
 508         return ustcomm_send_reg_msg(socket,
 509                 type,
 510                 CAA_BITS_PER_LONG,
 511                 lttng_alignof(uint8_t) * CHAR_BIT,
 512                 lttng_alignof(uint16_t) * CHAR_BIT,
 513                 lttng_alignof(uint32_t) * CHAR_BIT,
 514                 lttng_alignof(uint64_t) * CHAR_BIT,
 515                 lttng_alignof(unsigned long) * CHAR_BIT);
 516 }
 517
 518 static
 519 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 520 {
 521         ssize_t len;
 522
 523         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 524         switch (len) {
 525         case sizeof(*lur):
 526                 DBG("message successfully sent");
 527                 return 0;
 528         default:
 529                 if (len == -ECONNRESET) {
 530                         DBG("remote end closed connection");
 531                         return 0;
 532                 }
 533                 if (len < 0)
 534                         return len;
 535                 DBG("incorrect message size: %zd", len);
 536                 return -EINVAL;
 537         }
 538 }
 539
 540 static
 541 int handle_register_done(struct sock_info *sock_info)
 542 {
 543         int ret;
 544
 545         if (sock_info->constructor_sem_posted)
 546                 return 0;
 547         sock_info->constructor_sem_posted = 1;
 548         if (uatomic_read(&sem_count) <= 0) {
 549                 return 0;
 550         }
 551         ret = uatomic_add_return(&sem_count, -1);
 552         if (ret == 0) {
 553                 ret = sem_post(&constructor_wait);
 554                 assert(!ret);
 555         }
 556         return 0;
 557 }
 558
 559 /*
 560  * Only execute pending statedump after the constructor semaphore has
 561  * been posted by each listener thread. This means statedump will only
 562  * be performed after the "registration done" command is received from
 563  * each session daemon the application is connected to.
 564  *
 565  * This ensures we don't run into deadlock issues with the dynamic
 566  * loader mutex, which is held while the constructor is called and
 567  * waiting on the constructor semaphore. All operations requiring this
 568  * dynamic loader lock need to be postponed using this mechanism.
 569  */
 570 static
 571 void handle_pending_statedump(struct sock_info *sock_info)
 572 {
 573         int ctor_passed = sock_info->constructor_sem_posted;
 574
 575         if (ctor_passed && sock_info->statedump_pending) {
 576                 sock_info->statedump_pending = 0;
 577                 pthread_mutex_lock(&ust_fork_mutex);
 578                 lttng_handle_pending_statedump(sock_info);
 579                 pthread_mutex_unlock(&ust_fork_mutex);
 580         }
 581 }
 582
 583 static
 584 int handle_message(struct sock_info *sock_info,
 585                 int sock, struct ustcomm_ust_msg *lum)
 586 {
 587         int ret = 0;
 588         const struct lttng_ust_objd_ops *ops;
 589         struct ustcomm_ust_reply lur;
 590         union ust_args args;
 591         char ctxstr[LTTNG_UST_SYM_NAME_LEN];    /* App context string. */
 592         ssize_t len;
 593
 594         memset(&lur, 0, sizeof(lur));
 595
 596         if (ust_lock()) {
 597                 ret = -LTTNG_UST_ERR_EXITING;
 598                 goto error;
 599         }
 600
 601         ops = objd_ops(lum->handle);
 602         if (!ops) {
 603                 ret = -ENOENT;
 604                 goto error;
 605         }
 606
 607         switch (lum->cmd) {
 608         case LTTNG_UST_REGISTER_DONE:
 609                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 610                         ret = handle_register_done(sock_info);
 611                 else
 612                         ret = -EINVAL;
 613                 break;
 614         case LTTNG_UST_RELEASE:
 615                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 616                         ret = -EPERM;
 617                 else
 618                         ret = lttng_ust_objd_unref(lum->handle, 1);
 619                 break;
 620         case LTTNG_UST_FILTER:
 621         {
 622                 /* Receive filter data */
 623                 struct lttng_ust_filter_bytecode_node *bytecode;
 624
 625                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 626                         ERR("Filter data size is too large: %u bytes",
 627                                 lum->u.filter.data_size);
 628                         ret = -EINVAL;
 629                         goto error;
 630                 }
 631
 632                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 633                         ERR("Filter reloc offset %u is not within data",
 634                                 lum->u.filter.reloc_offset);
 635                         ret = -EINVAL;
 636                         goto error;
 637                 }
 638
 639                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 640                 if (!bytecode) {
 641                         ret = -ENOMEM;
 642                         goto error;
 643                 }
 644                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 645                                 lum->u.filter.data_size);
 646                 switch (len) {
 647                 case 0: /* orderly shutdown */
 648                         ret = 0;
 649                         free(bytecode);
 650                         goto error;
 651                 default:
 652                         if (len == lum->u.filter.data_size) {
 653                                 DBG("filter data received");
 654                                 break;
 655                         } else if (len < 0) {
 656                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 657                                 if (len == -ECONNRESET) {
 658                                         ERR("%s remote end closed connection", sock_info->name);
 659                                         ret = len;
 660                                         free(bytecode);
 661                                         goto error;
 662                                 }
 663                                 ret = len;
 664                                 free(bytecode);
 665                                 goto error;
 666                         } else {
 667                                 DBG("incorrect filter data message size: %zd", len);
 668                                 ret = -EINVAL;
 669                                 free(bytecode);
 670                                 goto error;
 671                         }
 672                 }
 673                 bytecode->bc.len = lum->u.filter.data_size;
 674                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 675                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 676                 if (ops->cmd) {
 677                         ret = ops->cmd(lum->handle, lum->cmd,
 678                                         (unsigned long) bytecode,
 679                                         &args, sock_info);
 680                         if (ret) {
 681                                 free(bytecode);
 682                         }
 683                         /* don't free bytecode if everything went fine. */
 684                 } else {
 685                         ret = -ENOSYS;
 686                         free(bytecode);
 687                 }
 688                 break;
 689         }
 690         case LTTNG_UST_EXCLUSION:
 691         {
 692                 /* Receive exclusion names */
 693                 struct lttng_ust_excluder_node *node;
 694                 unsigned int count;
 695
 696                 count = lum->u.exclusion.count;
 697                 if (count == 0) {
 698                         /* There are no names to read */
 699                         ret = 0;
 700                         goto error;
 701                 }
 702                 node = zmalloc(sizeof(*node) +
 703                                 count * LTTNG_UST_SYM_NAME_LEN);
 704                 if (!node) {
 705                         ret = -ENOMEM;
 706                         goto error;
 707                 }
 708                 node->excluder.count = count;
 709                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 710                                 count * LTTNG_UST_SYM_NAME_LEN);
 711                 switch (len) {
 712                 case 0: /* orderly shutdown */
 713                         ret = 0;
 714                         free(node);
 715                         goto error;
 716                 default:
 717                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 718                                 DBG("Exclusion data received");
 719                                 break;
 720                         } else if (len < 0) {
 721                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 722                                 if (len == -ECONNRESET) {
 723                                         ERR("%s remote end closed connection", sock_info->name);
 724                                         ret = len;
 725                                         free(node);
 726                                         goto error;
 727                                 }
 728                                 ret = len;
 729                                 free(node);
 730                                 goto error;
 731                         } else {
 732                                 DBG("Incorrect exclusion data message size: %zd", len);
 733                                 ret = -EINVAL;
 734                                 free(node);
 735                                 goto error;
 736                         }
 737                 }
 738                 if (ops->cmd) {
 739                         ret = ops->cmd(lum->handle, lum->cmd,
 740                                         (unsigned long) node,
 741                                         &args, sock_info);
 742                         if (ret) {
 743                                 free(node);
 744                         }
 745                         /* Don't free exclusion data if everything went fine. */
 746                 } else {
 747                         ret = -ENOSYS;
 748                         free(node);
 749                 }
 750                 break;
 751         }
 752         case LTTNG_UST_CHANNEL:
 753         {
 754                 void *chan_data;
 755                 int wakeup_fd;
 756
 757                 len = ustcomm_recv_channel_from_sessiond(sock,
 758                                 &chan_data, lum->u.channel.len,
 759                                 &wakeup_fd);
 760                 switch (len) {
 761                 case 0: /* orderly shutdown */
 762                         ret = 0;
 763                         goto error;
 764                 default:
 765                         if (len == lum->u.channel.len) {
 766                                 DBG("channel data received");
 767                                 break;
 768                         } else if (len < 0) {
 769                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 770                                 if (len == -ECONNRESET) {
 771                                         ERR("%s remote end closed connection", sock_info->name);
 772                                         ret = len;
 773                                         goto error;
 774                                 }
 775                                 ret = len;
 776                                 goto error;
 777                         } else {
 778                                 DBG("incorrect channel data message size: %zd", len);
 779                                 ret = -EINVAL;
 780                                 goto error;
 781                         }
 782                 }
 783                 args.channel.chan_data = chan_data;
 784                 args.channel.wakeup_fd = wakeup_fd;
 785                 if (ops->cmd)
 786                         ret = ops->cmd(lum->handle, lum->cmd,
 787                                         (unsigned long) &lum->u,
 788                                         &args, sock_info);
 789                 else
 790                         ret = -ENOSYS;
 791                 break;
 792         }
 793         case LTTNG_UST_STREAM:
 794         {
 795                 /* Receive shm_fd, wakeup_fd */
 796                 ret = ustcomm_recv_stream_from_sessiond(sock,
 797                         &lum->u.stream.len,
 798                         &args.stream.shm_fd,
 799                         &args.stream.wakeup_fd);
 800                 if (ret) {
 801                         goto error;
 802                 }
 803                 if (ops->cmd)
 804                         ret = ops->cmd(lum->handle, lum->cmd,
 805                                         (unsigned long) &lum->u,
 806                                         &args, sock_info);
 807                 else
 808                         ret = -ENOSYS;
 809                 break;
 810         }
 811         case LTTNG_UST_CONTEXT:
 812                 switch (lum->u.context.ctx) {
 813                 case LTTNG_UST_CONTEXT_APP_CONTEXT:
 814                 {
 815                         char *p;
 816                         size_t ctxlen, recvlen;
 817
 818                         ctxlen = strlen("$app.") + lum->u.context.u.app_ctx.provider_name_len - 1
 819                                         + strlen(":") + lum->u.context.u.app_ctx.ctx_name_len;
 820                         if (ctxlen >= LTTNG_UST_SYM_NAME_LEN) {
 821                                 ERR("Application context string length size is too large: %zu bytes",
 822                                         ctxlen);
 823                                 ret = -EINVAL;
 824                                 goto error;
 825                         }
 826                         strcpy(ctxstr, "$app.");
 827                         p = &ctxstr[strlen("$app.")];
 828                         recvlen = ctxlen - strlen("$app.");
 829                         len = ustcomm_recv_unix_sock(sock, p, recvlen);
 830                         switch (len) {
 831                         case 0: /* orderly shutdown */
 832                                 ret = 0;
 833                                 goto error;
 834                         default:
 835                                 if (len == recvlen) {
 836                                         DBG("app context data received");
 837                                         break;
 838                                 } else if (len < 0) {
 839                                         DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 840                                         if (len == -ECONNRESET) {
 841                                                 ERR("%s remote end closed connection", sock_info->name);
 842                                                 ret = len;
 843                                                 goto error;
 844                                         }
 845                                         ret = len;
 846                                         goto error;
 847                                 } else {
 848                                         DBG("incorrect app context data message size: %zd", len);
 849                                         ret = -EINVAL;
 850                                         goto error;
 851                                 }
 852                         }
 853                         /* Put : between provider and ctxname. */
 854                         p[lum->u.context.u.app_ctx.provider_name_len - 1] = ':';
 855                         args.app_context.ctxname = ctxstr;
 856                         break;
 857                 }
 858                 default:
 859                         break;
 860                 }
 861                 if (ops->cmd) {
 862                         ret = ops->cmd(lum->handle, lum->cmd,
 863                                         (unsigned long) &lum->u,
 864                                         &args, sock_info);
 865                 } else {
 866                         ret = -ENOSYS;
 867                 }
 868                 break;
 869         default:
 870                 if (ops->cmd)
 871                         ret = ops->cmd(lum->handle, lum->cmd,
 872                                         (unsigned long) &lum->u,
 873                                         &args, sock_info);
 874                 else
 875                         ret = -ENOSYS;
 876                 break;
 877         }
 878
 879         lur.handle = lum->handle;
 880         lur.cmd = lum->cmd;
 881         lur.ret_val = ret;
 882         if (ret >= 0) {
 883                 lur.ret_code = LTTNG_UST_OK;
 884         } else {
 885                 /*
 886                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 887                  * error that are not caused by the transport, except if
 888                  * we already have a more precise error message to
 889                  * report.
 890                  */
 891                 if (ret > -LTTNG_UST_ERR) {
 892                         /* Translate code to UST error. */
 893                         switch (ret) {
 894                         case -EEXIST:
 895                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 896                                 break;
 897                         case -EINVAL:
 898                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 899                                 break;
 900                         case -ENOENT:
 901                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 902                                 break;
 903                         case -EPERM:
 904                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 905                                 break;
 906                         case -ENOSYS:
 907                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 908                                 break;
 909                         default:
 910                                 lur.ret_code = -LTTNG_UST_ERR;
 911                                 break;
 912                         }
 913                 } else {
 914                         lur.ret_code = ret;
 915                 }
 916         }
 917         if (ret >= 0) {
 918                 switch (lum->cmd) {
 919                 case LTTNG_UST_TRACER_VERSION:
 920                         lur.u.version = lum->u.version;
 921                         break;
 922                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 923                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 924                         break;
 925                 }
 926         }
 927         DBG("Return value: %d", lur.ret_val);
 928         ret = send_reply(sock, &lur);
 929         if (ret < 0) {
 930                 DBG("error sending reply");
 931                 goto error;
 932         }
 933
 934         /*
 935          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 936          * after the reply.
 937          */
 938         if (lur.ret_code == LTTNG_UST_OK) {
 939                 switch (lum->cmd) {
 940                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 941                         len = ustcomm_send_unix_sock(sock,
 942                                 &args.field_list.entry,
 943                                 sizeof(args.field_list.entry));
 944                         if (len < 0) {
 945                                 ret = len;
 946                                 goto error;
 947                         }
 948                         if (len != sizeof(args.field_list.entry)) {
 949                                 ret = -EINVAL;
 950                                 goto error;
 951                         }
 952                 }
 953         }
 954
 955 error:
 956         ust_unlock();
 957
 958         /*
 959          * Performed delayed statedump operations outside of the UST
 960          * lock. We need to take the dynamic loader lock before we take
 961          * the UST lock internally within handle_pending_statedump().
 962           */
 963         handle_pending_statedump(sock_info);
 964
 965         return ret;
 966 }
 967
 968 static
 969 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 970 {
 971         int ret;
 972
 973         if (sock_info->root_handle != -1) {
 974                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 975                 if (ret) {
 976                         ERR("Error unref root handle");
 977                 }
 978                 sock_info->root_handle = -1;
 979         }
 980         sock_info->constructor_sem_posted = 0;
 981
 982         /*
 983          * wait_shm_mmap, socket and notify socket are used by listener
 984          * threads outside of the ust lock, so we cannot tear them down
 985          * ourselves, because we cannot join on these threads. Leave
 986          * responsibility of cleaning up these resources to the OS
 987          * process exit.
 988          */
 989         if (exiting)
 990                 return;
 991
 992         if (sock_info->socket != -1) {
 993                 ret = ustcomm_close_unix_sock(sock_info->socket);
 994                 if (ret) {
 995                         ERR("Error closing ust cmd socket");
 996                 }
 997                 sock_info->socket = -1;
 998         }
 999         if (sock_info->notify_socket != -1) {
1000                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1001                 if (ret) {
1002                         ERR("Error closing ust notify socket");
1003                 }
1004                 sock_info->notify_socket = -1;
1005         }
1006         if (sock_info->wait_shm_mmap) {
1007                 long page_size;
1008
1009                 page_size = sysconf(_SC_PAGE_SIZE);
1010                 if (page_size <= 0) {
1011                         if (!page_size) {
1012                                 errno = EINVAL;
1013                         }
1014                         PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1015                 } else {
1016                         ret = munmap(sock_info->wait_shm_mmap, page_size);
1017                         if (ret) {
1018                                 ERR("Error unmapping wait shm");
1019                         }
1020                 }
1021                 sock_info->wait_shm_mmap = NULL;
1022         }
1023 }
1024
1025 /*
1026  * Using fork to set umask in the child process (not multi-thread safe).
1027  * We deal with the shm_open vs ftruncate race (happening when the
1028  * sessiond owns the shm and does not let everybody modify it, to ensure
1029  * safety against shm_unlink) by simply letting the mmap fail and
1030  * retrying after a few seconds.
1031  * For global shm, everybody has rw access to it until the sessiond
1032  * starts.
1033  */
1034 static
1035 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
1036 {
1037         int wait_shm_fd, ret;
1038         pid_t pid;
1039
1040         /*
1041          * Try to open read-only.
1042          */
1043         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1044         if (wait_shm_fd >= 0) {
1045                 int32_t tmp_read;
1046                 ssize_t len;
1047                 size_t bytes_read = 0;
1048
1049                 /*
1050                  * Try to read the fd. If unable to do so, try opening
1051                  * it in write mode.
1052                  */
1053                 do {
1054                         len = read(wait_shm_fd,
1055                                 &((char *) &tmp_read)[bytes_read],
1056                                 sizeof(tmp_read) - bytes_read);
1057                         if (len > 0) {
1058                                 bytes_read += len;
1059                         }
1060                 } while ((len < 0 && errno == EINTR)
1061                         || (len > 0 && bytes_read < sizeof(tmp_read)));
1062                 if (bytes_read != sizeof(tmp_read)) {
1063                         ret = close(wait_shm_fd);
1064                         if (ret) {
1065                                 ERR("close wait_shm_fd");
1066                         }
1067                         goto open_write;
1068                 }
1069                 goto end;
1070         } else if (wait_shm_fd < 0 && errno != ENOENT) {
1071                 /*
1072                  * Real-only open did not work, and it's not because the
1073                  * entry was not present. It's a failure that prohibits
1074                  * using shm.
1075                  */
1076                 ERR("Error opening shm %s", sock_info->wait_shm_path);
1077                 goto end;
1078         }
1079
1080 open_write:
1081         /*
1082          * If the open failed because the file did not exist, or because
1083          * the file was not truncated yet, try creating it ourself.
1084          */
1085         URCU_TLS(lttng_ust_nest_count)++;
1086         pid = fork();
1087         URCU_TLS(lttng_ust_nest_count)--;
1088         if (pid > 0) {
1089                 int status;
1090
1091                 /*
1092                  * Parent: wait for child to return, in which case the
1093                  * shared memory map will have been created.
1094                  */
1095                 pid = wait(&status);
1096                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
1097                         wait_shm_fd = -1;
1098                         goto end;
1099                 }
1100                 /*
1101                  * Try to open read-only again after creation.
1102                  */
1103                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1104                 if (wait_shm_fd < 0) {
1105                         /*
1106                          * Real-only open did not work. It's a failure
1107                          * that prohibits using shm.
1108                          */
1109                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1110                         goto end;
1111                 }
1112                 goto end;
1113         } else if (pid == 0) {
1114                 int create_mode;
1115
1116                 /* Child */
1117                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1118                 if (sock_info->global)
1119                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1120                 /*
1121                  * We're alone in a child process, so we can modify the
1122                  * process-wide umask.
1123                  */
1124                 umask(~create_mode);
1125                 /*
1126                  * Try creating shm (or get rw access).
1127                  * We don't do an exclusive open, because we allow other
1128                  * processes to create+ftruncate it concurrently.
1129                  */
1130                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1131                                 O_RDWR | O_CREAT, create_mode);
1132                 if (wait_shm_fd >= 0) {
1133                         ret = ftruncate(wait_shm_fd, mmap_size);
1134                         if (ret) {
1135                                 PERROR("ftruncate");
1136                                 _exit(EXIT_FAILURE);
1137                         }
1138                         _exit(EXIT_SUCCESS);
1139                 }
1140                 /*
1141                  * For local shm, we need to have rw access to accept
1142                  * opening it: this means the local sessiond will be
1143                  * able to wake us up. For global shm, we open it even
1144                  * if rw access is not granted, because the root.root
1145                  * sessiond will be able to override all rights and wake
1146                  * us up.
1147                  */
1148                 if (!sock_info->global && errno != EACCES) {
1149                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1150                         _exit(EXIT_FAILURE);
1151                 }
1152                 /*
1153                  * The shm exists, but we cannot open it RW. Report
1154                  * success.
1155                  */
1156                 _exit(EXIT_SUCCESS);
1157         } else {
1158                 return -1;
1159         }
1160 end:
1161         if (wait_shm_fd >= 0 && !sock_info->global) {
1162                 struct stat statbuf;
1163
1164                 /*
1165                  * Ensure that our user is the owner of the shm file for
1166                  * local shm. If we do not own the file, it means our
1167                  * sessiond will not have access to wake us up (there is
1168                  * probably a rogue process trying to fake our
1169                  * sessiond). Fallback to polling method in this case.
1170                  */
1171                 ret = fstat(wait_shm_fd, &statbuf);
1172                 if (ret) {
1173                         PERROR("fstat");
1174                         goto error_close;
1175                 }
1176                 if (statbuf.st_uid != getuid())
1177                         goto error_close;
1178         }
1179         return wait_shm_fd;
1180
1181 error_close:
1182         ret = close(wait_shm_fd);
1183         if (ret) {
1184                 PERROR("Error closing fd");
1185         }
1186         return -1;
1187 }
1188
1189 static
1190 char *get_map_shm(struct sock_info *sock_info)
1191 {
1192         long page_size;
1193         int wait_shm_fd, ret;
1194         char *wait_shm_mmap;
1195
1196         page_size = sysconf(_SC_PAGE_SIZE);
1197         if (page_size <= 0) {
1198                 if (!page_size) {
1199                         errno = EINVAL;
1200                 }
1201                 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1202                 goto error;
1203         }
1204
1205         wait_shm_fd = get_wait_shm(sock_info, page_size);
1206         if (wait_shm_fd < 0) {
1207                 goto error;
1208         }
1209         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1210                   MAP_SHARED, wait_shm_fd, 0);
1211         /* close shm fd immediately after taking the mmap reference */
1212         ret = close(wait_shm_fd);
1213         if (ret) {
1214                 PERROR("Error closing fd");
1215         }
1216         if (wait_shm_mmap == MAP_FAILED) {
1217                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1218                 goto error;
1219         }
1220         return wait_shm_mmap;
1221
1222 error:
1223         return NULL;
1224 }
1225
1226 static
1227 void wait_for_sessiond(struct sock_info *sock_info)
1228 {
1229         if (ust_lock()) {
1230                 goto quit;
1231         }
1232         if (wait_poll_fallback) {
1233                 goto error;
1234         }
1235         if (!sock_info->wait_shm_mmap) {
1236                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1237                 if (!sock_info->wait_shm_mmap)
1238                         goto error;
1239         }
1240         ust_unlock();
1241
1242         DBG("Waiting for %s apps sessiond", sock_info->name);
1243         /* Wait for futex wakeup */
1244         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
1245                 goto end_wait;
1246
1247         while (futex_async((int32_t *) sock_info->wait_shm_mmap,
1248                         FUTEX_WAIT, 0, NULL, NULL, 0)) {
1249                 switch (errno) {
1250                 case EWOULDBLOCK:
1251                         /* Value already changed. */
1252                         goto end_wait;
1253                 case EINTR:
1254                         /* Retry if interrupted by signal. */
1255                         break;  /* Get out of switch. */
1256                 case EFAULT:
1257                         wait_poll_fallback = 1;
1258                         DBG(
1259 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1260 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1261 "Please upgrade your kernel "
1262 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1263 "mainline). LTTng-UST will use polling mode fallback.");
1264                         if (ust_debug())
1265                                 PERROR("futex");
1266                         goto end_wait;
1267                 }
1268         }
1269 end_wait:
1270         return;
1271
1272 quit:
1273         ust_unlock();
1274         return;
1275
1276 error:
1277         ust_unlock();
1278         return;
1279 }
1280
1281 /*
1282  * This thread does not allocate any resource, except within
1283  * handle_message, within mutex protection. This mutex protects against
1284  * fork and exit.
1285  * The other moment it allocates resources is at socket connection, which
1286  * is also protected by the mutex.
1287  */
1288 static
1289 void *ust_listener_thread(void *arg)
1290 {
1291         struct sock_info *sock_info = arg;
1292         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1293         long timeout;
1294
1295         /* Restart trying to connect to the session daemon */
1296 restart:
1297         if (prev_connect_failed) {
1298                 /* Wait for sessiond availability with pipe */
1299                 wait_for_sessiond(sock_info);
1300                 if (has_waited) {
1301                         has_waited = 0;
1302                         /*
1303                          * Sleep for 5 seconds before retrying after a
1304                          * sequence of failure / wait / failure. This
1305                          * deals with a killed or broken session daemon.
1306                          */
1307                         sleep(5);
1308                 } else {
1309                         has_waited = 1;
1310                 }
1311                 prev_connect_failed = 0;
1312         }
1313
1314         if (sock_info->socket != -1) {
1315                 ret = ustcomm_close_unix_sock(sock_info->socket);
1316                 if (ret) {
1317                         ERR("Error closing %s ust cmd socket",
1318                                 sock_info->name);
1319                 }
1320                 sock_info->socket = -1;
1321         }
1322         if (sock_info->notify_socket != -1) {
1323                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1324                 if (ret) {
1325                         ERR("Error closing %s ust notify socket",
1326                                 sock_info->name);
1327                 }
1328                 sock_info->notify_socket = -1;
1329         }
1330
1331         /*
1332          * Register. We need to perform both connect and sending
1333          * registration message before doing the next connect otherwise
1334          * we may reach unix socket connect queue max limits and block
1335          * on the 2nd connect while the session daemon is awaiting the
1336          * first connect registration message.
1337          */
1338         /* Connect cmd socket */
1339         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1340         if (ret < 0) {
1341                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1342                 prev_connect_failed = 1;
1343
1344                 if (ust_lock()) {
1345                         goto quit;
1346                 }
1347
1348                 /*
1349                  * If we cannot find the sessiond daemon, don't delay
1350                  * constructor execution.
1351                  */
1352                 ret = handle_register_done(sock_info);
1353                 assert(!ret);
1354                 ust_unlock();
1355                 goto restart;
1356         }
1357         sock_info->socket = ret;
1358
1359         if (ust_lock()) {
1360                 goto quit;
1361         }
1362
1363         /*
1364          * Create only one root handle per listener thread for the whole
1365          * process lifetime, so we ensure we get ID which is statically
1366          * assigned to the root handle.
1367          */
1368         if (sock_info->root_handle == -1) {
1369                 ret = lttng_abi_create_root_handle();
1370                 if (ret < 0) {
1371                         ERR("Error creating root handle");
1372                         goto quit;
1373                 }
1374                 sock_info->root_handle = ret;
1375         }
1376
1377         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1378         if (ret < 0) {
1379                 ERR("Error registering to %s ust cmd socket",
1380                         sock_info->name);
1381                 prev_connect_failed = 1;
1382                 /*
1383                  * If we cannot register to the sessiond daemon, don't
1384                  * delay constructor execution.
1385                  */
1386                 ret = handle_register_done(sock_info);
1387                 assert(!ret);
1388                 ust_unlock();
1389                 goto restart;
1390         }
1391
1392         ust_unlock();
1393
1394         /* Connect notify socket */
1395         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1396         if (ret < 0) {
1397                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1398                 prev_connect_failed = 1;
1399
1400                 if (ust_lock()) {
1401                         goto quit;
1402                 }
1403
1404                 /*
1405                  * If we cannot find the sessiond daemon, don't delay
1406                  * constructor execution.
1407                  */
1408                 ret = handle_register_done(sock_info);
1409                 assert(!ret);
1410                 ust_unlock();
1411                 goto restart;
1412         }
1413         sock_info->notify_socket = ret;
1414
1415         timeout = get_notify_sock_timeout();
1416         if (timeout >= 0) {
1417                 /*
1418                  * Give at least 10ms to sessiond to reply to
1419                  * notifications.
1420                  */
1421                 if (timeout < 10)
1422                         timeout = 10;
1423                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1424                                 timeout);
1425                 if (ret < 0) {
1426                         WARN("Error setting socket receive timeout");
1427                 }
1428                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1429                                 timeout);
1430                 if (ret < 0) {
1431                         WARN("Error setting socket send timeout");
1432                 }
1433         } else if (timeout < -1) {
1434                 WARN("Unsupported timeout value %ld", timeout);
1435         }
1436
1437         if (ust_lock()) {
1438                 goto quit;
1439         }
1440
1441         ret = register_to_sessiond(sock_info->notify_socket,
1442                         USTCTL_SOCKET_NOTIFY);
1443         if (ret < 0) {
1444                 ERR("Error registering to %s ust notify socket",
1445                         sock_info->name);
1446                 prev_connect_failed = 1;
1447                 /*
1448                  * If we cannot register to the sessiond daemon, don't
1449                  * delay constructor execution.
1450                  */
1451                 ret = handle_register_done(sock_info);
1452                 assert(!ret);
1453                 ust_unlock();
1454                 goto restart;
1455         }
1456         sock = sock_info->socket;
1457
1458         ust_unlock();
1459
1460         for (;;) {
1461                 ssize_t len;
1462                 struct ustcomm_ust_msg lum;
1463
1464                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1465                 switch (len) {
1466                 case 0: /* orderly shutdown */
1467                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1468                         if (ust_lock()) {
1469                                 goto quit;
1470                         }
1471                         /*
1472                          * Either sessiond has shutdown or refused us by closing the socket.
1473                          * In either case, we don't want to delay construction execution,
1474                          * and we need to wait before retry.
1475                          */
1476                         prev_connect_failed = 1;
1477                         /*
1478                          * If we cannot register to the sessiond daemon, don't
1479                          * delay constructor execution.
1480                          */
1481                         ret = handle_register_done(sock_info);
1482                         assert(!ret);
1483                         ust_unlock();
1484                         goto end;
1485                 case sizeof(lum):
1486                         print_cmd(lum.cmd, lum.handle);
1487                         ret = handle_message(sock_info, sock, &lum);
1488                         if (ret) {
1489                                 ERR("Error handling message for %s socket",
1490                                         sock_info->name);
1491                                 /*
1492                                  * Close socket if protocol error is
1493                                  * detected.
1494                                  */
1495                                 goto end;
1496                         }
1497                         continue;
1498                 default:
1499                         if (len < 0) {
1500                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1501                         } else {
1502                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1503                         }
1504                         if (len == -ECONNRESET) {
1505                                 DBG("%s remote end closed connection", sock_info->name);
1506                                 goto end;
1507                         }
1508                         goto end;
1509                 }
1510
1511         }
1512 end:
1513         if (ust_lock()) {
1514                 goto quit;
1515         }
1516         /* Cleanup socket handles before trying to reconnect */
1517         lttng_ust_objd_table_owner_cleanup(sock_info);
1518         ust_unlock();
1519         goto restart;   /* try to reconnect */
1520
1521 quit:
1522         ust_unlock();
1523
1524         pthread_mutex_lock(&ust_exit_mutex);
1525         sock_info->thread_active = 0;
1526         pthread_mutex_unlock(&ust_exit_mutex);
1527         return NULL;
1528 }
1529
1530 /*
1531  * Weak symbol to call when the ust malloc wrapper is not loaded.
1532  */
1533 __attribute__((weak))
1534 void lttng_ust_malloc_wrapper_init(void)
1535 {
1536 }
1537
1538 /*
1539  * sessiond monitoring thread: monitor presence of global and per-user
1540  * sessiond by polling the application common named pipe.
1541  */
1542 void __attribute__((constructor)) lttng_ust_init(void)
1543 {
1544         struct timespec constructor_timeout;
1545         sigset_t sig_all_blocked, orig_parent_mask;
1546         pthread_attr_t thread_attr;
1547         int timeout_mode;
1548         int ret;
1549
1550         if (uatomic_xchg(&initialized, 1) == 1)
1551                 return;
1552
1553         /*
1554          * Fixup interdependency between TLS fixup mutex (which happens
1555          * to be the dynamic linker mutex) and ust_lock, taken within
1556          * the ust lock.
1557          */
1558         lttng_fixup_urcu_bp_tls();
1559         lttng_fixup_ringbuffer_tls();
1560         lttng_fixup_vtid_tls();
1561         lttng_fixup_nest_count_tls();
1562         lttng_fixup_procname_tls();
1563         lttng_fixup_ust_mutex_nest_tls();
1564
1565         /*
1566          * We want precise control over the order in which we construct
1567          * our sub-libraries vs starting to receive commands from
1568          * sessiond (otherwise leading to errors when trying to create
1569          * sessiond before the init functions are completed).
1570          */
1571         init_usterr();
1572         init_tracepoint();
1573         lttng_ust_clock_init();
1574         lttng_ust_getcpu_init();
1575         lttng_ust_statedump_init();
1576         lttng_ring_buffer_metadata_client_init();
1577         lttng_ring_buffer_client_overwrite_init();
1578         lttng_ring_buffer_client_overwrite_rt_init();
1579         lttng_ring_buffer_client_discard_init();
1580         lttng_ring_buffer_client_discard_rt_init();
1581         lttng_perf_counter_init();
1582         /*
1583          * Invoke ust malloc wrapper init before starting other threads.
1584          */
1585         lttng_ust_malloc_wrapper_init();
1586
1587         timeout_mode = get_constructor_timeout(&constructor_timeout);
1588
1589         ret = sem_init(&constructor_wait, 0, 0);
1590         assert(!ret);
1591
1592         ret = setup_local_apps();
1593         if (ret) {
1594                 DBG("local apps setup returned %d", ret);
1595         }
1596
1597         /* A new thread created by pthread_create inherits the signal mask
1598          * from the parent. To avoid any signal being received by the
1599          * listener thread, we block all signals temporarily in the parent,
1600          * while we create the listener thread.
1601          */
1602         sigfillset(&sig_all_blocked);
1603         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1604         if (ret) {
1605                 ERR("pthread_sigmask: %s", strerror(ret));
1606         }
1607
1608         ret = pthread_attr_init(&thread_attr);
1609         if (ret) {
1610                 ERR("pthread_attr_init: %s", strerror(ret));
1611         }
1612         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1613         if (ret) {
1614                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1615         }
1616
1617         pthread_mutex_lock(&ust_exit_mutex);
1618         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1619                         ust_listener_thread, &global_apps);
1620         if (ret) {
1621                 ERR("pthread_create global: %s", strerror(ret));
1622         }
1623         global_apps.thread_active = 1;
1624         pthread_mutex_unlock(&ust_exit_mutex);
1625
1626         if (local_apps.allowed) {
1627                 pthread_mutex_lock(&ust_exit_mutex);
1628                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1629                                 ust_listener_thread, &local_apps);
1630                 if (ret) {
1631                         ERR("pthread_create local: %s", strerror(ret));
1632                 }
1633                 local_apps.thread_active = 1;
1634                 pthread_mutex_unlock(&ust_exit_mutex);
1635         } else {
1636                 handle_register_done(&local_apps);
1637         }
1638         ret = pthread_attr_destroy(&thread_attr);
1639         if (ret) {
1640                 ERR("pthread_attr_destroy: %s", strerror(ret));
1641         }
1642
1643         /* Restore original signal mask in parent */
1644         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1645         if (ret) {
1646                 ERR("pthread_sigmask: %s", strerror(ret));
1647         }
1648
1649         switch (timeout_mode) {
1650         case 1: /* timeout wait */
1651                 do {
1652                         ret = sem_timedwait(&constructor_wait,
1653                                         &constructor_timeout);
1654                 } while (ret < 0 && errno == EINTR);
1655                 if (ret < 0 && errno == ETIMEDOUT) {
1656                         ERR("Timed out waiting for lttng-sessiond");
1657                 } else {
1658                         assert(!ret);
1659                 }
1660                 break;
1661         case -1:/* wait forever */
1662                 do {
1663                         ret = sem_wait(&constructor_wait);
1664                 } while (ret < 0 && errno == EINTR);
1665                 assert(!ret);
1666                 break;
1667         case 0: /* no timeout */
1668                 break;
1669         }
1670 }
1671
1672 static
1673 void lttng_ust_cleanup(int exiting)
1674 {
1675         cleanup_sock_info(&global_apps, exiting);
1676         cleanup_sock_info(&local_apps, exiting);
1677         /*
1678          * The teardown in this function all affect data structures
1679          * accessed under the UST lock by the listener thread. This
1680          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1681          * that none of these threads are accessing this data at this
1682          * point.
1683          */
1684         lttng_ust_abi_exit();
1685         lttng_ust_events_exit();
1686         lttng_perf_counter_exit();
1687         lttng_ring_buffer_client_discard_rt_exit();
1688         lttng_ring_buffer_client_discard_exit();
1689         lttng_ring_buffer_client_overwrite_rt_exit();
1690         lttng_ring_buffer_client_overwrite_exit();
1691         lttng_ring_buffer_metadata_client_exit();
1692         lttng_ust_statedump_destroy();
1693         exit_tracepoint();
1694         if (!exiting) {
1695                 /* Reinitialize values for fork */
1696                 sem_count = 2;
1697                 lttng_ust_comm_should_quit = 0;
1698                 initialized = 0;
1699         }
1700 }
1701
1702 void __attribute__((destructor)) lttng_ust_exit(void)
1703 {
1704         int ret;
1705
1706         /*
1707          * Using pthread_cancel here because:
1708          * A) we don't want to hang application teardown.
1709          * B) the thread is not allocating any resource.
1710          */
1711
1712         /*
1713          * Require the communication thread to quit. Synchronize with
1714          * mutexes to ensure it is not in a mutex critical section when
1715          * pthread_cancel is later called.
1716          */
1717         ust_lock_nocheck();
1718         lttng_ust_comm_should_quit = 1;
1719         ust_unlock();
1720
1721         pthread_mutex_lock(&ust_exit_mutex);
1722         /* cancel threads */
1723         if (global_apps.thread_active) {
1724                 ret = pthread_cancel(global_apps.ust_listener);
1725                 if (ret) {
1726                         ERR("Error cancelling global ust listener thread: %s",
1727                                 strerror(ret));
1728                 } else {
1729                         global_apps.thread_active = 0;
1730                 }
1731         }
1732         if (local_apps.thread_active) {
1733                 ret = pthread_cancel(local_apps.ust_listener);
1734                 if (ret) {
1735                         ERR("Error cancelling local ust listener thread: %s",
1736                                 strerror(ret));
1737                 } else {
1738                         local_apps.thread_active = 0;
1739                 }
1740         }
1741         pthread_mutex_unlock(&ust_exit_mutex);
1742
1743         /*
1744          * Do NOT join threads: use of sys_futex makes it impossible to
1745          * join the threads without using async-cancel, but async-cancel
1746          * is delivered by a signal, which could hit the target thread
1747          * anywhere in its code path, including while the ust_lock() is
1748          * held, causing a deadlock for the other thread. Let the OS
1749          * cleanup the threads if there are stalled in a syscall.
1750          */
1751         lttng_ust_cleanup(1);
1752 }
1753
1754 /*
1755  * We exclude the worker threads across fork and clone (except
1756  * CLONE_VM), because these system calls only keep the forking thread
1757  * running in the child.  Therefore, we don't want to call fork or clone
1758  * in the middle of an tracepoint or ust tracing state modification.
1759  * Holding this mutex protects these structures across fork and clone.
1760  */
1761 void ust_before_fork(sigset_t *save_sigset)
1762 {
1763         /*
1764          * Disable signals. This is to avoid that the child intervenes
1765          * before it is properly setup for tracing. It is safer to
1766          * disable all signals, because then we know we are not breaking
1767          * anything by restoring the original mask.
1768          */
1769         sigset_t all_sigs;
1770         int ret;
1771
1772         if (URCU_TLS(lttng_ust_nest_count))
1773                 return;
1774         /* Disable signals */
1775         sigfillset(&all_sigs);
1776         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1777         if (ret == -1) {
1778                 PERROR("sigprocmask");
1779         }
1780
1781         pthread_mutex_lock(&ust_fork_mutex);
1782
1783         ust_lock_nocheck();
1784         rcu_bp_before_fork();
1785 }
1786
1787 static void ust_after_fork_common(sigset_t *restore_sigset)
1788 {
1789         int ret;
1790
1791         DBG("process %d", getpid());
1792         ust_unlock();
1793
1794         pthread_mutex_unlock(&ust_fork_mutex);
1795
1796         /* Restore signals */
1797         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1798         if (ret == -1) {
1799                 PERROR("sigprocmask");
1800         }
1801 }
1802
1803 void ust_after_fork_parent(sigset_t *restore_sigset)
1804 {
1805         if (URCU_TLS(lttng_ust_nest_count))
1806                 return;
1807         DBG("process %d", getpid());
1808         rcu_bp_after_fork_parent();
1809         /* Release mutexes and reenable signals */
1810         ust_after_fork_common(restore_sigset);
1811 }
1812
1813 /*
1814  * After fork, in the child, we need to cleanup all the leftover state,
1815  * except the worker thread which already magically disappeared thanks
1816  * to the weird Linux fork semantics. After tyding up, we call
1817  * lttng_ust_init() again to start over as a new PID.
1818  *
1819  * This is meant for forks() that have tracing in the child between the
1820  * fork and following exec call (if there is any).
1821  */
1822 void ust_after_fork_child(sigset_t *restore_sigset)
1823 {
1824         if (URCU_TLS(lttng_ust_nest_count))
1825                 return;
1826         DBG("process %d", getpid());
1827         /* Release urcu mutexes */
1828         rcu_bp_after_fork_child();
1829         lttng_ust_cleanup(0);
1830         lttng_context_vtid_reset();
1831         /* Release mutexes and reenable signals */
1832         ust_after_fork_common(restore_sigset);
1833         lttng_ust_init();
1834 }
1835
1836 void lttng_ust_sockinfo_session_enabled(void *owner)
1837 {
1838         struct sock_info *sock_info = owner;
1839         sock_info->statedump_pending = 1;
1840 }