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         ssize_t len;
 592
 593         memset(&lur, 0, sizeof(lur));
 594
 595         if (ust_lock()) {
 596                 ret = -LTTNG_UST_ERR_EXITING;
 597                 goto error;
 598         }
 599
 600         ops = objd_ops(lum->handle);
 601         if (!ops) {
 602                 ret = -ENOENT;
 603                 goto error;
 604         }
 605
 606         switch (lum->cmd) {
 607         case LTTNG_UST_REGISTER_DONE:
 608                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 609                         ret = handle_register_done(sock_info);
 610                 else
 611                         ret = -EINVAL;
 612                 break;
 613         case LTTNG_UST_RELEASE:
 614                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 615                         ret = -EPERM;
 616                 else
 617                         ret = lttng_ust_objd_unref(lum->handle, 1);
 618                 break;
 619         case LTTNG_UST_FILTER:
 620         {
 621                 /* Receive filter data */
 622                 struct lttng_ust_filter_bytecode_node *bytecode;
 623
 624                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 625                         ERR("Filter data size is too large: %u bytes",
 626                                 lum->u.filter.data_size);
 627                         ret = -EINVAL;
 628                         goto error;
 629                 }
 630
 631                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 632                         ERR("Filter reloc offset %u is not within data",
 633                                 lum->u.filter.reloc_offset);
 634                         ret = -EINVAL;
 635                         goto error;
 636                 }
 637
 638                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 639                 if (!bytecode) {
 640                         ret = -ENOMEM;
 641                         goto error;
 642                 }
 643                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 644                                 lum->u.filter.data_size);
 645                 switch (len) {
 646                 case 0: /* orderly shutdown */
 647                         ret = 0;
 648                         free(bytecode);
 649                         goto error;
 650                 default:
 651                         if (len == lum->u.filter.data_size) {
 652                                 DBG("filter data received");
 653                                 break;
 654                         } else if (len < 0) {
 655                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 656                                 if (len == -ECONNRESET) {
 657                                         ERR("%s remote end closed connection", sock_info->name);
 658                                         ret = len;
 659                                         free(bytecode);
 660                                         goto error;
 661                                 }
 662                                 ret = len;
 663                                 free(bytecode);
 664                                 goto error;
 665                         } else {
 666                                 DBG("incorrect filter data message size: %zd", len);
 667                                 ret = -EINVAL;
 668                                 free(bytecode);
 669                                 goto error;
 670                         }
 671                 }
 672                 bytecode->bc.len = lum->u.filter.data_size;
 673                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 674                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 675                 if (ops->cmd) {
 676                         ret = ops->cmd(lum->handle, lum->cmd,
 677                                         (unsigned long) bytecode,
 678                                         &args, sock_info);
 679                         if (ret) {
 680                                 free(bytecode);
 681                         }
 682                         /* don't free bytecode if everything went fine. */
 683                 } else {
 684                         ret = -ENOSYS;
 685                         free(bytecode);
 686                 }
 687                 break;
 688         }
 689         case LTTNG_UST_EXCLUSION:
 690         {
 691                 /* Receive exclusion names */
 692                 struct lttng_ust_excluder_node *node;
 693                 unsigned int count;
 694
 695                 count = lum->u.exclusion.count;
 696                 if (count == 0) {
 697                         /* There are no names to read */
 698                         ret = 0;
 699                         goto error;
 700                 }
 701                 node = zmalloc(sizeof(*node) +
 702                                 count * LTTNG_UST_SYM_NAME_LEN);
 703                 if (!node) {
 704                         ret = -ENOMEM;
 705                         goto error;
 706                 }
 707                 node->excluder.count = count;
 708                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 709                                 count * LTTNG_UST_SYM_NAME_LEN);
 710                 switch (len) {
 711                 case 0: /* orderly shutdown */
 712                         ret = 0;
 713                         free(node);
 714                         goto error;
 715                 default:
 716                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 717                                 DBG("Exclusion data received");
 718                                 break;
 719                         } else if (len < 0) {
 720                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 721                                 if (len == -ECONNRESET) {
 722                                         ERR("%s remote end closed connection", sock_info->name);
 723                                         ret = len;
 724                                         free(node);
 725                                         goto error;
 726                                 }
 727                                 ret = len;
 728                                 free(node);
 729                                 goto error;
 730                         } else {
 731                                 DBG("Incorrect exclusion data message size: %zd", len);
 732                                 ret = -EINVAL;
 733                                 free(node);
 734                                 goto error;
 735                         }
 736                 }
 737                 if (ops->cmd) {
 738                         ret = ops->cmd(lum->handle, lum->cmd,
 739                                         (unsigned long) node,
 740                                         &args, sock_info);
 741                         if (ret) {
 742                                 free(node);
 743                         }
 744                         /* Don't free exclusion data if everything went fine. */
 745                 } else {
 746                         ret = -ENOSYS;
 747                         free(node);
 748                 }
 749                 break;
 750         }
 751         case LTTNG_UST_CHANNEL:
 752         {
 753                 void *chan_data;
 754                 int wakeup_fd;
 755
 756                 len = ustcomm_recv_channel_from_sessiond(sock,
 757                                 &chan_data, lum->u.channel.len,
 758                                 &wakeup_fd);
 759                 switch (len) {
 760                 case 0: /* orderly shutdown */
 761                         ret = 0;
 762                         goto error;
 763                 default:
 764                         if (len == lum->u.channel.len) {
 765                                 DBG("channel data received");
 766                                 break;
 767                         } else if (len < 0) {
 768                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 769                                 if (len == -ECONNRESET) {
 770                                         ERR("%s remote end closed connection", sock_info->name);
 771                                         ret = len;
 772                                         goto error;
 773                                 }
 774                                 ret = len;
 775                                 goto error;
 776                         } else {
 777                                 DBG("incorrect channel data message size: %zd", len);
 778                                 ret = -EINVAL;
 779                                 goto error;
 780                         }
 781                 }
 782                 args.channel.chan_data = chan_data;
 783                 args.channel.wakeup_fd = wakeup_fd;
 784                 if (ops->cmd)
 785                         ret = ops->cmd(lum->handle, lum->cmd,
 786                                         (unsigned long) &lum->u,
 787                                         &args, sock_info);
 788                 else
 789                         ret = -ENOSYS;
 790                 break;
 791         }
 792         case LTTNG_UST_STREAM:
 793         {
 794                 /* Receive shm_fd, wakeup_fd */
 795                 ret = ustcomm_recv_stream_from_sessiond(sock,
 796                         &lum->u.stream.len,
 797                         &args.stream.shm_fd,
 798                         &args.stream.wakeup_fd);
 799                 if (ret) {
 800                         goto error;
 801                 }
 802                 if (ops->cmd)
 803                         ret = ops->cmd(lum->handle, lum->cmd,
 804                                         (unsigned long) &lum->u,
 805                                         &args, sock_info);
 806                 else
 807                         ret = -ENOSYS;
 808                 break;
 809         }
 810         default:
 811                 if (ops->cmd)
 812                         ret = ops->cmd(lum->handle, lum->cmd,
 813                                         (unsigned long) &lum->u,
 814                                         &args, sock_info);
 815                 else
 816                         ret = -ENOSYS;
 817                 break;
 818         }
 819
 820         lur.handle = lum->handle;
 821         lur.cmd = lum->cmd;
 822         lur.ret_val = ret;
 823         if (ret >= 0) {
 824                 lur.ret_code = LTTNG_UST_OK;
 825         } else {
 826                 /*
 827                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 828                  * error that are not caused by the transport, except if
 829                  * we already have a more precise error message to
 830                  * report.
 831                  */
 832                 if (ret > -LTTNG_UST_ERR) {
 833                         /* Translate code to UST error. */
 834                         switch (ret) {
 835                         case -EEXIST:
 836                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 837                                 break;
 838                         case -EINVAL:
 839                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 840                                 break;
 841                         case -ENOENT:
 842                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 843                                 break;
 844                         case -EPERM:
 845                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 846                                 break;
 847                         case -ENOSYS:
 848                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 849                                 break;
 850                         default:
 851                                 lur.ret_code = -LTTNG_UST_ERR;
 852                                 break;
 853                         }
 854                 } else {
 855                         lur.ret_code = ret;
 856                 }
 857         }
 858         if (ret >= 0) {
 859                 switch (lum->cmd) {
 860                 case LTTNG_UST_TRACER_VERSION:
 861                         lur.u.version = lum->u.version;
 862                         break;
 863                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 864                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 865                         break;
 866                 }
 867         }
 868         DBG("Return value: %d", lur.ret_val);
 869         ret = send_reply(sock, &lur);
 870         if (ret < 0) {
 871                 DBG("error sending reply");
 872                 goto error;
 873         }
 874
 875         /*
 876          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 877          * after the reply.
 878          */
 879         if (lur.ret_code == LTTNG_UST_OK) {
 880                 switch (lum->cmd) {
 881                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 882                         len = ustcomm_send_unix_sock(sock,
 883                                 &args.field_list.entry,
 884                                 sizeof(args.field_list.entry));
 885                         if (len < 0) {
 886                                 ret = len;
 887                                 goto error;
 888                         }
 889                         if (len != sizeof(args.field_list.entry)) {
 890                                 ret = -EINVAL;
 891                                 goto error;
 892                         }
 893                 }
 894         }
 895
 896 error:
 897         ust_unlock();
 898
 899         /*
 900          * Performed delayed statedump operations outside of the UST
 901          * lock. We need to take the dynamic loader lock before we take
 902          * the UST lock internally within handle_pending_statedump().
 903           */
 904         handle_pending_statedump(sock_info);
 905
 906         return ret;
 907 }
 908
 909 static
 910 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 911 {
 912         int ret;
 913
 914         if (sock_info->root_handle != -1) {
 915                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 916                 if (ret) {
 917                         ERR("Error unref root handle");
 918                 }
 919                 sock_info->root_handle = -1;
 920         }
 921         sock_info->constructor_sem_posted = 0;
 922
 923         /*
 924          * wait_shm_mmap, socket and notify socket are used by listener
 925          * threads outside of the ust lock, so we cannot tear them down
 926          * ourselves, because we cannot join on these threads. Leave
 927          * responsibility of cleaning up these resources to the OS
 928          * process exit.
 929          */
 930         if (exiting)
 931                 return;
 932
 933         if (sock_info->socket != -1) {
 934                 ret = ustcomm_close_unix_sock(sock_info->socket);
 935                 if (ret) {
 936                         ERR("Error closing ust cmd socket");
 937                 }
 938                 sock_info->socket = -1;
 939         }
 940         if (sock_info->notify_socket != -1) {
 941                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 942                 if (ret) {
 943                         ERR("Error closing ust notify socket");
 944                 }
 945                 sock_info->notify_socket = -1;
 946         }
 947         if (sock_info->wait_shm_mmap) {
 948                 long page_size;
 949
 950                 page_size = sysconf(_SC_PAGE_SIZE);
 951                 if (page_size <= 0) {
 952                         if (!page_size) {
 953                                 errno = EINVAL;
 954                         }
 955                         PERROR("Error in sysconf(_SC_PAGE_SIZE)");
 956                 } else {
 957                         ret = munmap(sock_info->wait_shm_mmap, page_size);
 958                         if (ret) {
 959                                 ERR("Error unmapping wait shm");
 960                         }
 961                 }
 962                 sock_info->wait_shm_mmap = NULL;
 963         }
 964 }
 965
 966 /*
 967  * Using fork to set umask in the child process (not multi-thread safe).
 968  * We deal with the shm_open vs ftruncate race (happening when the
 969  * sessiond owns the shm and does not let everybody modify it, to ensure
 970  * safety against shm_unlink) by simply letting the mmap fail and
 971  * retrying after a few seconds.
 972  * For global shm, everybody has rw access to it until the sessiond
 973  * starts.
 974  */
 975 static
 976 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 977 {
 978         int wait_shm_fd, ret;
 979         pid_t pid;
 980
 981         /*
 982          * Try to open read-only.
 983          */
 984         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 985         if (wait_shm_fd >= 0) {
 986                 int32_t tmp_read;
 987                 ssize_t len;
 988                 size_t bytes_read = 0;
 989
 990                 /*
 991                  * Try to read the fd. If unable to do so, try opening
 992                  * it in write mode.
 993                  */
 994                 do {
 995                         len = read(wait_shm_fd,
 996                                 &((char *) &tmp_read)[bytes_read],
 997                                 sizeof(tmp_read) - bytes_read);
 998                         if (len > 0) {
 999                                 bytes_read += len;
1000                         }
1001                 } while ((len < 0 && errno == EINTR)
1002                         || (len > 0 && bytes_read < sizeof(tmp_read)));
1003                 if (bytes_read != sizeof(tmp_read)) {
1004                         ret = close(wait_shm_fd);
1005                         if (ret) {
1006                                 ERR("close wait_shm_fd");
1007                         }
1008                         goto open_write;
1009                 }
1010                 goto end;
1011         } else if (wait_shm_fd < 0 && errno != ENOENT) {
1012                 /*
1013                  * Real-only open did not work, and it's not because the
1014                  * entry was not present. It's a failure that prohibits
1015                  * using shm.
1016                  */
1017                 ERR("Error opening shm %s", sock_info->wait_shm_path);
1018                 goto end;
1019         }
1020
1021 open_write:
1022         /*
1023          * If the open failed because the file did not exist, or because
1024          * the file was not truncated yet, try creating it ourself.
1025          */
1026         URCU_TLS(lttng_ust_nest_count)++;
1027         pid = fork();
1028         URCU_TLS(lttng_ust_nest_count)--;
1029         if (pid > 0) {
1030                 int status;
1031
1032                 /*
1033                  * Parent: wait for child to return, in which case the
1034                  * shared memory map will have been created.
1035                  */
1036                 pid = wait(&status);
1037                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
1038                         wait_shm_fd = -1;
1039                         goto end;
1040                 }
1041                 /*
1042                  * Try to open read-only again after creation.
1043                  */
1044                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1045                 if (wait_shm_fd < 0) {
1046                         /*
1047                          * Real-only open did not work. It's a failure
1048                          * that prohibits using shm.
1049                          */
1050                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1051                         goto end;
1052                 }
1053                 goto end;
1054         } else if (pid == 0) {
1055                 int create_mode;
1056
1057                 /* Child */
1058                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1059                 if (sock_info->global)
1060                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1061                 /*
1062                  * We're alone in a child process, so we can modify the
1063                  * process-wide umask.
1064                  */
1065                 umask(~create_mode);
1066                 /*
1067                  * Try creating shm (or get rw access).
1068                  * We don't do an exclusive open, because we allow other
1069                  * processes to create+ftruncate it concurrently.
1070                  */
1071                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1072                                 O_RDWR | O_CREAT, create_mode);
1073                 if (wait_shm_fd >= 0) {
1074                         ret = ftruncate(wait_shm_fd, mmap_size);
1075                         if (ret) {
1076                                 PERROR("ftruncate");
1077                                 _exit(EXIT_FAILURE);
1078                         }
1079                         _exit(EXIT_SUCCESS);
1080                 }
1081                 /*
1082                  * For local shm, we need to have rw access to accept
1083                  * opening it: this means the local sessiond will be
1084                  * able to wake us up. For global shm, we open it even
1085                  * if rw access is not granted, because the root.root
1086                  * sessiond will be able to override all rights and wake
1087                  * us up.
1088                  */
1089                 if (!sock_info->global && errno != EACCES) {
1090                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1091                         _exit(EXIT_FAILURE);
1092                 }
1093                 /*
1094                  * The shm exists, but we cannot open it RW. Report
1095                  * success.
1096                  */
1097                 _exit(EXIT_SUCCESS);
1098         } else {
1099                 return -1;
1100         }
1101 end:
1102         if (wait_shm_fd >= 0 && !sock_info->global) {
1103                 struct stat statbuf;
1104
1105                 /*
1106                  * Ensure that our user is the owner of the shm file for
1107                  * local shm. If we do not own the file, it means our
1108                  * sessiond will not have access to wake us up (there is
1109                  * probably a rogue process trying to fake our
1110                  * sessiond). Fallback to polling method in this case.
1111                  */
1112                 ret = fstat(wait_shm_fd, &statbuf);
1113                 if (ret) {
1114                         PERROR("fstat");
1115                         goto error_close;
1116                 }
1117                 if (statbuf.st_uid != getuid())
1118                         goto error_close;
1119         }
1120         return wait_shm_fd;
1121
1122 error_close:
1123         ret = close(wait_shm_fd);
1124         if (ret) {
1125                 PERROR("Error closing fd");
1126         }
1127         return -1;
1128 }
1129
1130 static
1131 char *get_map_shm(struct sock_info *sock_info)
1132 {
1133         long page_size;
1134         int wait_shm_fd, ret;
1135         char *wait_shm_mmap;
1136
1137         page_size = sysconf(_SC_PAGE_SIZE);
1138         if (page_size <= 0) {
1139                 if (!page_size) {
1140                         errno = EINVAL;
1141                 }
1142                 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1143                 goto error;
1144         }
1145
1146         wait_shm_fd = get_wait_shm(sock_info, page_size);
1147         if (wait_shm_fd < 0) {
1148                 goto error;
1149         }
1150         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1151                   MAP_SHARED, wait_shm_fd, 0);
1152         /* close shm fd immediately after taking the mmap reference */
1153         ret = close(wait_shm_fd);
1154         if (ret) {
1155                 PERROR("Error closing fd");
1156         }
1157         if (wait_shm_mmap == MAP_FAILED) {
1158                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1159                 goto error;
1160         }
1161         return wait_shm_mmap;
1162
1163 error:
1164         return NULL;
1165 }
1166
1167 static
1168 void wait_for_sessiond(struct sock_info *sock_info)
1169 {
1170         if (ust_lock()) {
1171                 goto quit;
1172         }
1173         if (wait_poll_fallback) {
1174                 goto error;
1175         }
1176         if (!sock_info->wait_shm_mmap) {
1177                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1178                 if (!sock_info->wait_shm_mmap)
1179                         goto error;
1180         }
1181         ust_unlock();
1182
1183         DBG("Waiting for %s apps sessiond", sock_info->name);
1184         /* Wait for futex wakeup */
1185         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
1186                 goto end_wait;
1187
1188         while (futex_async((int32_t *) sock_info->wait_shm_mmap,
1189                         FUTEX_WAIT, 0, NULL, NULL, 0)) {
1190                 switch (errno) {
1191                 case EWOULDBLOCK:
1192                         /* Value already changed. */
1193                         goto end_wait;
1194                 case EINTR:
1195                         /* Retry if interrupted by signal. */
1196                         break;  /* Get out of switch. */
1197                 case EFAULT:
1198                         wait_poll_fallback = 1;
1199                         DBG(
1200 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1201 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1202 "Please upgrade your kernel "
1203 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1204 "mainline). LTTng-UST will use polling mode fallback.");
1205                         if (ust_debug())
1206                                 PERROR("futex");
1207                         goto end_wait;
1208                 }
1209         }
1210 end_wait:
1211         return;
1212
1213 quit:
1214         ust_unlock();
1215         return;
1216
1217 error:
1218         ust_unlock();
1219         return;
1220 }
1221
1222 /*
1223  * This thread does not allocate any resource, except within
1224  * handle_message, within mutex protection. This mutex protects against
1225  * fork and exit.
1226  * The other moment it allocates resources is at socket connection, which
1227  * is also protected by the mutex.
1228  */
1229 static
1230 void *ust_listener_thread(void *arg)
1231 {
1232         struct sock_info *sock_info = arg;
1233         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1234         long timeout;
1235
1236         /* Restart trying to connect to the session daemon */
1237 restart:
1238         if (prev_connect_failed) {
1239                 /* Wait for sessiond availability with pipe */
1240                 wait_for_sessiond(sock_info);
1241                 if (has_waited) {
1242                         has_waited = 0;
1243                         /*
1244                          * Sleep for 5 seconds before retrying after a
1245                          * sequence of failure / wait / failure. This
1246                          * deals with a killed or broken session daemon.
1247                          */
1248                         sleep(5);
1249                 }
1250                 has_waited = 1;
1251                 prev_connect_failed = 0;
1252         }
1253
1254         if (sock_info->socket != -1) {
1255                 ret = ustcomm_close_unix_sock(sock_info->socket);
1256                 if (ret) {
1257                         ERR("Error closing %s ust cmd socket",
1258                                 sock_info->name);
1259                 }
1260                 sock_info->socket = -1;
1261         }
1262         if (sock_info->notify_socket != -1) {
1263                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1264                 if (ret) {
1265                         ERR("Error closing %s ust notify socket",
1266                                 sock_info->name);
1267                 }
1268                 sock_info->notify_socket = -1;
1269         }
1270
1271         /*
1272          * Register. We need to perform both connect and sending
1273          * registration message before doing the next connect otherwise
1274          * we may reach unix socket connect queue max limits and block
1275          * on the 2nd connect while the session daemon is awaiting the
1276          * first connect registration message.
1277          */
1278         /* Connect cmd socket */
1279         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1280         if (ret < 0) {
1281                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1282                 prev_connect_failed = 1;
1283
1284                 if (ust_lock()) {
1285                         goto quit;
1286                 }
1287
1288                 /*
1289                  * If we cannot find the sessiond daemon, don't delay
1290                  * constructor execution.
1291                  */
1292                 ret = handle_register_done(sock_info);
1293                 assert(!ret);
1294                 ust_unlock();
1295                 goto restart;
1296         }
1297         sock_info->socket = ret;
1298
1299         if (ust_lock()) {
1300                 goto quit;
1301         }
1302
1303         /*
1304          * Create only one root handle per listener thread for the whole
1305          * process lifetime, so we ensure we get ID which is statically
1306          * assigned to the root handle.
1307          */
1308         if (sock_info->root_handle == -1) {
1309                 ret = lttng_abi_create_root_handle();
1310                 if (ret < 0) {
1311                         ERR("Error creating root handle");
1312                         goto quit;
1313                 }
1314                 sock_info->root_handle = ret;
1315         }
1316
1317         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1318         if (ret < 0) {
1319                 ERR("Error registering to %s ust cmd socket",
1320                         sock_info->name);
1321                 prev_connect_failed = 1;
1322                 /*
1323                  * If we cannot register to the sessiond daemon, don't
1324                  * delay constructor execution.
1325                  */
1326                 ret = handle_register_done(sock_info);
1327                 assert(!ret);
1328                 ust_unlock();
1329                 goto restart;
1330         }
1331
1332         ust_unlock();
1333
1334         /* Connect notify socket */
1335         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1336         if (ret < 0) {
1337                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1338                 prev_connect_failed = 1;
1339
1340                 if (ust_lock()) {
1341                         goto quit;
1342                 }
1343
1344                 /*
1345                  * If we cannot find the sessiond daemon, don't delay
1346                  * constructor execution.
1347                  */
1348                 ret = handle_register_done(sock_info);
1349                 assert(!ret);
1350                 ust_unlock();
1351                 goto restart;
1352         }
1353         sock_info->notify_socket = ret;
1354
1355         timeout = get_notify_sock_timeout();
1356         if (timeout >= 0) {
1357                 /*
1358                  * Give at least 10ms to sessiond to reply to
1359                  * notifications.
1360                  */
1361                 if (timeout < 10)
1362                         timeout = 10;
1363                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1364                                 timeout);
1365                 if (ret < 0) {
1366                         WARN("Error setting socket receive timeout");
1367                 }
1368                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1369                                 timeout);
1370                 if (ret < 0) {
1371                         WARN("Error setting socket send timeout");
1372                 }
1373         } else if (timeout < -1) {
1374                 WARN("Unsupported timeout value %ld", timeout);
1375         }
1376
1377         if (ust_lock()) {
1378                 goto quit;
1379         }
1380
1381         ret = register_to_sessiond(sock_info->notify_socket,
1382                         USTCTL_SOCKET_NOTIFY);
1383         if (ret < 0) {
1384                 ERR("Error registering to %s ust notify socket",
1385                         sock_info->name);
1386                 prev_connect_failed = 1;
1387                 /*
1388                  * If we cannot register to the sessiond daemon, don't
1389                  * delay constructor execution.
1390                  */
1391                 ret = handle_register_done(sock_info);
1392                 assert(!ret);
1393                 ust_unlock();
1394                 goto restart;
1395         }
1396         sock = sock_info->socket;
1397
1398         ust_unlock();
1399
1400         for (;;) {
1401                 ssize_t len;
1402                 struct ustcomm_ust_msg lum;
1403
1404                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1405                 switch (len) {
1406                 case 0: /* orderly shutdown */
1407                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1408                         if (ust_lock()) {
1409                                 goto quit;
1410                         }
1411                         /*
1412                          * Either sessiond has shutdown or refused us by closing the socket.
1413                          * In either case, we don't want to delay construction execution,
1414                          * and we need to wait before retry.
1415                          */
1416                         prev_connect_failed = 1;
1417                         /*
1418                          * If we cannot register to the sessiond daemon, don't
1419                          * delay constructor execution.
1420                          */
1421                         ret = handle_register_done(sock_info);
1422                         assert(!ret);
1423                         ust_unlock();
1424                         goto end;
1425                 case sizeof(lum):
1426                         print_cmd(lum.cmd, lum.handle);
1427                         ret = handle_message(sock_info, sock, &lum);
1428                         if (ret) {
1429                                 ERR("Error handling message for %s socket",
1430                                         sock_info->name);
1431                                 /*
1432                                  * Close socket if protocol error is
1433                                  * detected.
1434                                  */
1435                                 goto end;
1436                         }
1437                         continue;
1438                 default:
1439                         if (len < 0) {
1440                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1441                         } else {
1442                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1443                         }
1444                         if (len == -ECONNRESET) {
1445                                 DBG("%s remote end closed connection", sock_info->name);
1446                                 goto end;
1447                         }
1448                         goto end;
1449                 }
1450
1451         }
1452 end:
1453         if (ust_lock()) {
1454                 goto quit;
1455         }
1456         /* Cleanup socket handles before trying to reconnect */
1457         lttng_ust_objd_table_owner_cleanup(sock_info);
1458         ust_unlock();
1459         goto restart;   /* try to reconnect */
1460
1461 quit:
1462         ust_unlock();
1463
1464         pthread_mutex_lock(&ust_exit_mutex);
1465         sock_info->thread_active = 0;
1466         pthread_mutex_unlock(&ust_exit_mutex);
1467         return NULL;
1468 }
1469
1470 /*
1471  * Weak symbol to call when the ust malloc wrapper is not loaded.
1472  */
1473 __attribute__((weak))
1474 void lttng_ust_malloc_wrapper_init(void)
1475 {
1476 }
1477
1478 /*
1479  * sessiond monitoring thread: monitor presence of global and per-user
1480  * sessiond by polling the application common named pipe.
1481  */
1482 void __attribute__((constructor)) lttng_ust_init(void)
1483 {
1484         struct timespec constructor_timeout;
1485         sigset_t sig_all_blocked, orig_parent_mask;
1486         pthread_attr_t thread_attr;
1487         int timeout_mode;
1488         int ret;
1489
1490         if (uatomic_xchg(&initialized, 1) == 1)
1491                 return;
1492
1493         /*
1494          * Fixup interdependency between TLS fixup mutex (which happens
1495          * to be the dynamic linker mutex) and ust_lock, taken within
1496          * the ust lock.
1497          */
1498         lttng_fixup_urcu_bp_tls();
1499         lttng_fixup_ringbuffer_tls();
1500         lttng_fixup_vtid_tls();
1501         lttng_fixup_nest_count_tls();
1502         lttng_fixup_procname_tls();
1503         lttng_fixup_ust_mutex_nest_tls();
1504
1505         /*
1506          * We want precise control over the order in which we construct
1507          * our sub-libraries vs starting to receive commands from
1508          * sessiond (otherwise leading to errors when trying to create
1509          * sessiond before the init functions are completed).
1510          */
1511         init_usterr();
1512         init_tracepoint();
1513         lttng_ust_clock_init();
1514         lttng_ust_getcpu_init();
1515         lttng_ust_statedump_init();
1516         lttng_ring_buffer_metadata_client_init();
1517         lttng_ring_buffer_client_overwrite_init();
1518         lttng_ring_buffer_client_overwrite_rt_init();
1519         lttng_ring_buffer_client_discard_init();
1520         lttng_ring_buffer_client_discard_rt_init();
1521         lttng_perf_counter_init();
1522         lttng_context_init();
1523         /*
1524          * Invoke ust malloc wrapper init before starting other threads.
1525          */
1526         lttng_ust_malloc_wrapper_init();
1527
1528         timeout_mode = get_constructor_timeout(&constructor_timeout);
1529
1530         ret = sem_init(&constructor_wait, 0, 0);
1531         assert(!ret);
1532
1533         ret = setup_local_apps();
1534         if (ret) {
1535                 DBG("local apps setup returned %d", ret);
1536         }
1537
1538         /* A new thread created by pthread_create inherits the signal mask
1539          * from the parent. To avoid any signal being received by the
1540          * listener thread, we block all signals temporarily in the parent,
1541          * while we create the listener thread.
1542          */
1543         sigfillset(&sig_all_blocked);
1544         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1545         if (ret) {
1546                 ERR("pthread_sigmask: %s", strerror(ret));
1547         }
1548
1549         ret = pthread_attr_init(&thread_attr);
1550         if (ret) {
1551                 ERR("pthread_attr_init: %s", strerror(ret));
1552         }
1553         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1554         if (ret) {
1555                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1556         }
1557
1558         pthread_mutex_lock(&ust_exit_mutex);
1559         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1560                         ust_listener_thread, &global_apps);
1561         if (ret) {
1562                 ERR("pthread_create global: %s", strerror(ret));
1563         }
1564         global_apps.thread_active = 1;
1565         pthread_mutex_unlock(&ust_exit_mutex);
1566
1567         if (local_apps.allowed) {
1568                 pthread_mutex_lock(&ust_exit_mutex);
1569                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1570                                 ust_listener_thread, &local_apps);
1571                 if (ret) {
1572                         ERR("pthread_create local: %s", strerror(ret));
1573                 }
1574                 local_apps.thread_active = 1;
1575                 pthread_mutex_unlock(&ust_exit_mutex);
1576         } else {
1577                 handle_register_done(&local_apps);
1578         }
1579         ret = pthread_attr_destroy(&thread_attr);
1580         if (ret) {
1581                 ERR("pthread_attr_destroy: %s", strerror(ret));
1582         }
1583
1584         /* Restore original signal mask in parent */
1585         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1586         if (ret) {
1587                 ERR("pthread_sigmask: %s", strerror(ret));
1588         }
1589
1590         switch (timeout_mode) {
1591         case 1: /* timeout wait */
1592                 do {
1593                         ret = sem_timedwait(&constructor_wait,
1594                                         &constructor_timeout);
1595                 } while (ret < 0 && errno == EINTR);
1596                 if (ret < 0 && errno == ETIMEDOUT) {
1597                         ERR("Timed out waiting for lttng-sessiond");
1598                 } else {
1599                         assert(!ret);
1600                 }
1601                 break;
1602         case -1:/* wait forever */
1603                 do {
1604                         ret = sem_wait(&constructor_wait);
1605                 } while (ret < 0 && errno == EINTR);
1606                 assert(!ret);
1607                 break;
1608         case 0: /* no timeout */
1609                 break;
1610         }
1611 }
1612
1613 static
1614 void lttng_ust_cleanup(int exiting)
1615 {
1616         cleanup_sock_info(&global_apps, exiting);
1617         cleanup_sock_info(&local_apps, exiting);
1618         /*
1619          * The teardown in this function all affect data structures
1620          * accessed under the UST lock by the listener thread. This
1621          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1622          * that none of these threads are accessing this data at this
1623          * point.
1624          */
1625         lttng_ust_abi_exit();
1626         lttng_ust_events_exit();
1627         lttng_context_exit();
1628         lttng_perf_counter_exit();
1629         lttng_ring_buffer_client_discard_rt_exit();
1630         lttng_ring_buffer_client_discard_exit();
1631         lttng_ring_buffer_client_overwrite_rt_exit();
1632         lttng_ring_buffer_client_overwrite_exit();
1633         lttng_ring_buffer_metadata_client_exit();
1634         lttng_ust_statedump_destroy();
1635         exit_tracepoint();
1636         if (!exiting) {
1637                 /* Reinitialize values for fork */
1638                 sem_count = 2;
1639                 lttng_ust_comm_should_quit = 0;
1640                 initialized = 0;
1641         }
1642 }
1643
1644 void __attribute__((destructor)) lttng_ust_exit(void)
1645 {
1646         int ret;
1647
1648         /*
1649          * Using pthread_cancel here because:
1650          * A) we don't want to hang application teardown.
1651          * B) the thread is not allocating any resource.
1652          */
1653
1654         /*
1655          * Require the communication thread to quit. Synchronize with
1656          * mutexes to ensure it is not in a mutex critical section when
1657          * pthread_cancel is later called.
1658          */
1659         ust_lock_nocheck();
1660         lttng_ust_comm_should_quit = 1;
1661         ust_unlock();
1662
1663         pthread_mutex_lock(&ust_exit_mutex);
1664         /* cancel threads */
1665         if (global_apps.thread_active) {
1666                 ret = pthread_cancel(global_apps.ust_listener);
1667                 if (ret) {
1668                         ERR("Error cancelling global ust listener thread: %s",
1669                                 strerror(ret));
1670                 } else {
1671                         global_apps.thread_active = 0;
1672                 }
1673         }
1674         if (local_apps.thread_active) {
1675                 ret = pthread_cancel(local_apps.ust_listener);
1676                 if (ret) {
1677                         ERR("Error cancelling local ust listener thread: %s",
1678                                 strerror(ret));
1679                 } else {
1680                         local_apps.thread_active = 0;
1681                 }
1682         }
1683         pthread_mutex_unlock(&ust_exit_mutex);
1684
1685         /*
1686          * Do NOT join threads: use of sys_futex makes it impossible to
1687          * join the threads without using async-cancel, but async-cancel
1688          * is delivered by a signal, which could hit the target thread
1689          * anywhere in its code path, including while the ust_lock() is
1690          * held, causing a deadlock for the other thread. Let the OS
1691          * cleanup the threads if there are stalled in a syscall.
1692          */
1693         lttng_ust_cleanup(1);
1694 }
1695
1696 /*
1697  * We exclude the worker threads across fork and clone (except
1698  * CLONE_VM), because these system calls only keep the forking thread
1699  * running in the child.  Therefore, we don't want to call fork or clone
1700  * in the middle of an tracepoint or ust tracing state modification.
1701  * Holding this mutex protects these structures across fork and clone.
1702  */
1703 void ust_before_fork(sigset_t *save_sigset)
1704 {
1705         /*
1706          * Disable signals. This is to avoid that the child intervenes
1707          * before it is properly setup for tracing. It is safer to
1708          * disable all signals, because then we know we are not breaking
1709          * anything by restoring the original mask.
1710          */
1711         sigset_t all_sigs;
1712         int ret;
1713
1714         if (URCU_TLS(lttng_ust_nest_count))
1715                 return;
1716         /* Disable signals */
1717         sigfillset(&all_sigs);
1718         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1719         if (ret == -1) {
1720                 PERROR("sigprocmask");
1721         }
1722
1723         pthread_mutex_lock(&ust_fork_mutex);
1724
1725         ust_lock_nocheck();
1726         rcu_bp_before_fork();
1727 }
1728
1729 static void ust_after_fork_common(sigset_t *restore_sigset)
1730 {
1731         int ret;
1732
1733         DBG("process %d", getpid());
1734         ust_unlock();
1735
1736         pthread_mutex_unlock(&ust_fork_mutex);
1737
1738         /* Restore signals */
1739         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1740         if (ret == -1) {
1741                 PERROR("sigprocmask");
1742         }
1743 }
1744
1745 void ust_after_fork_parent(sigset_t *restore_sigset)
1746 {
1747         if (URCU_TLS(lttng_ust_nest_count))
1748                 return;
1749         DBG("process %d", getpid());
1750         rcu_bp_after_fork_parent();
1751         /* Release mutexes and reenable signals */
1752         ust_after_fork_common(restore_sigset);
1753 }
1754
1755 /*
1756  * After fork, in the child, we need to cleanup all the leftover state,
1757  * except the worker thread which already magically disappeared thanks
1758  * to the weird Linux fork semantics. After tyding up, we call
1759  * lttng_ust_init() again to start over as a new PID.
1760  *
1761  * This is meant for forks() that have tracing in the child between the
1762  * fork and following exec call (if there is any).
1763  */
1764 void ust_after_fork_child(sigset_t *restore_sigset)
1765 {
1766         if (URCU_TLS(lttng_ust_nest_count))
1767                 return;
1768         DBG("process %d", getpid());
1769         /* Release urcu mutexes */
1770         rcu_bp_after_fork_child();
1771         lttng_ust_cleanup(0);
1772         lttng_context_vtid_reset();
1773         /* Release mutexes and reenable signals */
1774         ust_after_fork_common(restore_sigset);
1775         lttng_ust_init();
1776 }
1777
1778 void lttng_ust_sockinfo_session_enabled(void *owner)
1779 {
1780         struct sock_info *sock_info = owner;
1781         sock_info->statedump_pending = 1;
1782 }