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         /* All negative values are considered as "-1". */
 459         if (constructor_delay_ms < -1)
 460                 constructor_delay_ms = -1;
 461         return constructor_delay_ms;
 462 }
 463
 464 static
 465 long get_notify_sock_timeout(void)
 466 {
 467         return get_timeout();
 468 }
 469
 470 /*
 471  * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
 472  */
 473 static
 474 int get_constructor_timeout(struct timespec *constructor_timeout)
 475 {
 476         long constructor_delay_ms;
 477         int ret;
 478
 479         constructor_delay_ms = get_timeout();
 480
 481         switch (constructor_delay_ms) {
 482         case -1:/* fall-through */
 483         case 0:
 484                 return constructor_delay_ms;
 485         default:
 486                 break;
 487         }
 488
 489         /*
 490          * If we are unable to find the current time, don't wait.
 491          */
 492         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 493         if (ret) {
 494                 /* Don't wait. */
 495                 return 0;
 496         }
 497         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 498         constructor_timeout->tv_nsec +=
 499                 (constructor_delay_ms % 1000UL) * 1000000UL;
 500         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 501                 constructor_timeout->tv_sec++;
 502                 constructor_timeout->tv_nsec -= 1000000000UL;
 503         }
 504         /* Timeout wait (constructor_delay_ms). */
 505         return 1;
 506 }
 507
 508 static
 509 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 510 {
 511         return ustcomm_send_reg_msg(socket,
 512                 type,
 513                 CAA_BITS_PER_LONG,
 514                 lttng_alignof(uint8_t) * CHAR_BIT,
 515                 lttng_alignof(uint16_t) * CHAR_BIT,
 516                 lttng_alignof(uint32_t) * CHAR_BIT,
 517                 lttng_alignof(uint64_t) * CHAR_BIT,
 518                 lttng_alignof(unsigned long) * CHAR_BIT);
 519 }
 520
 521 static
 522 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 523 {
 524         ssize_t len;
 525
 526         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 527         switch (len) {
 528         case sizeof(*lur):
 529                 DBG("message successfully sent");
 530                 return 0;
 531         default:
 532                 if (len == -ECONNRESET) {
 533                         DBG("remote end closed connection");
 534                         return 0;
 535                 }
 536                 if (len < 0)
 537                         return len;
 538                 DBG("incorrect message size: %zd", len);
 539                 return -EINVAL;
 540         }
 541 }
 542
 543 static
 544 int handle_register_done(struct sock_info *sock_info)
 545 {
 546         int ret;
 547
 548         if (sock_info->constructor_sem_posted)
 549                 return 0;
 550         sock_info->constructor_sem_posted = 1;
 551         if (uatomic_read(&sem_count) <= 0) {
 552                 return 0;
 553         }
 554         ret = uatomic_add_return(&sem_count, -1);
 555         if (ret == 0) {
 556                 ret = sem_post(&constructor_wait);
 557                 assert(!ret);
 558         }
 559         return 0;
 560 }
 561
 562 /*
 563  * Only execute pending statedump after the constructor semaphore has
 564  * been posted by each listener thread. This means statedump will only
 565  * be performed after the "registration done" command is received from
 566  * each session daemon the application is connected to.
 567  *
 568  * This ensures we don't run into deadlock issues with the dynamic
 569  * loader mutex, which is held while the constructor is called and
 570  * waiting on the constructor semaphore. All operations requiring this
 571  * dynamic loader lock need to be postponed using this mechanism.
 572  */
 573 static
 574 void handle_pending_statedump(struct sock_info *sock_info)
 575 {
 576         int ctor_passed = sock_info->constructor_sem_posted;
 577
 578         if (ctor_passed && sock_info->statedump_pending) {
 579                 sock_info->statedump_pending = 0;
 580                 pthread_mutex_lock(&ust_fork_mutex);
 581                 lttng_handle_pending_statedump(sock_info);
 582                 pthread_mutex_unlock(&ust_fork_mutex);
 583         }
 584 }
 585
 586 static
 587 int handle_message(struct sock_info *sock_info,
 588                 int sock, struct ustcomm_ust_msg *lum)
 589 {
 590         int ret = 0;
 591         const struct lttng_ust_objd_ops *ops;
 592         struct ustcomm_ust_reply lur;
 593         union ust_args args;
 594         ssize_t len;
 595
 596         memset(&lur, 0, sizeof(lur));
 597
 598         if (ust_lock()) {
 599                 ret = -LTTNG_UST_ERR_EXITING;
 600                 goto error;
 601         }
 602
 603         ops = objd_ops(lum->handle);
 604         if (!ops) {
 605                 ret = -ENOENT;
 606                 goto error;
 607         }
 608
 609         switch (lum->cmd) {
 610         case LTTNG_UST_REGISTER_DONE:
 611                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 612                         ret = handle_register_done(sock_info);
 613                 else
 614                         ret = -EINVAL;
 615                 break;
 616         case LTTNG_UST_RELEASE:
 617                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 618                         ret = -EPERM;
 619                 else
 620                         ret = lttng_ust_objd_unref(lum->handle, 1);
 621                 break;
 622         case LTTNG_UST_FILTER:
 623         {
 624                 /* Receive filter data */
 625                 struct lttng_ust_filter_bytecode_node *bytecode;
 626
 627                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 628                         ERR("Filter data size is too large: %u bytes",
 629                                 lum->u.filter.data_size);
 630                         ret = -EINVAL;
 631                         goto error;
 632                 }
 633
 634                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 635                         ERR("Filter reloc offset %u is not within data",
 636                                 lum->u.filter.reloc_offset);
 637                         ret = -EINVAL;
 638                         goto error;
 639                 }
 640
 641                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 642                 if (!bytecode) {
 643                         ret = -ENOMEM;
 644                         goto error;
 645                 }
 646                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 647                                 lum->u.filter.data_size);
 648                 switch (len) {
 649                 case 0: /* orderly shutdown */
 650                         ret = 0;
 651                         free(bytecode);
 652                         goto error;
 653                 default:
 654                         if (len == lum->u.filter.data_size) {
 655                                 DBG("filter data received");
 656                                 break;
 657                         } else if (len < 0) {
 658                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 659                                 if (len == -ECONNRESET) {
 660                                         ERR("%s remote end closed connection", sock_info->name);
 661                                         ret = len;
 662                                         free(bytecode);
 663                                         goto error;
 664                                 }
 665                                 ret = len;
 666                                 free(bytecode);
 667                                 goto error;
 668                         } else {
 669                                 DBG("incorrect filter data message size: %zd", len);
 670                                 ret = -EINVAL;
 671                                 free(bytecode);
 672                                 goto error;
 673                         }
 674                 }
 675                 bytecode->bc.len = lum->u.filter.data_size;
 676                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 677                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 678                 if (ops->cmd) {
 679                         ret = ops->cmd(lum->handle, lum->cmd,
 680                                         (unsigned long) bytecode,
 681                                         &args, sock_info);
 682                         if (ret) {
 683                                 free(bytecode);
 684                         }
 685                         /* don't free bytecode if everything went fine. */
 686                 } else {
 687                         ret = -ENOSYS;
 688                         free(bytecode);
 689                 }
 690                 break;
 691         }
 692         case LTTNG_UST_EXCLUSION:
 693         {
 694                 /* Receive exclusion names */
 695                 struct lttng_ust_excluder_node *node;
 696                 unsigned int count;
 697
 698                 count = lum->u.exclusion.count;
 699                 if (count == 0) {
 700                         /* There are no names to read */
 701                         ret = 0;
 702                         goto error;
 703                 }
 704                 node = zmalloc(sizeof(*node) +
 705                                 count * LTTNG_UST_SYM_NAME_LEN);
 706                 if (!node) {
 707                         ret = -ENOMEM;
 708                         goto error;
 709                 }
 710                 node->excluder.count = count;
 711                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 712                                 count * LTTNG_UST_SYM_NAME_LEN);
 713                 switch (len) {
 714                 case 0: /* orderly shutdown */
 715                         ret = 0;
 716                         free(node);
 717                         goto error;
 718                 default:
 719                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 720                                 DBG("Exclusion data received");
 721                                 break;
 722                         } else if (len < 0) {
 723                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 724                                 if (len == -ECONNRESET) {
 725                                         ERR("%s remote end closed connection", sock_info->name);
 726                                         ret = len;
 727                                         free(node);
 728                                         goto error;
 729                                 }
 730                                 ret = len;
 731                                 free(node);
 732                                 goto error;
 733                         } else {
 734                                 DBG("Incorrect exclusion data message size: %zd", len);
 735                                 ret = -EINVAL;
 736                                 free(node);
 737                                 goto error;
 738                         }
 739                 }
 740                 if (ops->cmd) {
 741                         ret = ops->cmd(lum->handle, lum->cmd,
 742                                         (unsigned long) node,
 743                                         &args, sock_info);
 744                         if (ret) {
 745                                 free(node);
 746                         }
 747                         /* Don't free exclusion data if everything went fine. */
 748                 } else {
 749                         ret = -ENOSYS;
 750                         free(node);
 751                 }
 752                 break;
 753         }
 754         case LTTNG_UST_CHANNEL:
 755         {
 756                 void *chan_data;
 757                 int wakeup_fd;
 758
 759                 len = ustcomm_recv_channel_from_sessiond(sock,
 760                                 &chan_data, lum->u.channel.len,
 761                                 &wakeup_fd);
 762                 switch (len) {
 763                 case 0: /* orderly shutdown */
 764                         ret = 0;
 765                         goto error;
 766                 default:
 767                         if (len == lum->u.channel.len) {
 768                                 DBG("channel data received");
 769                                 break;
 770                         } else if (len < 0) {
 771                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 772                                 if (len == -ECONNRESET) {
 773                                         ERR("%s remote end closed connection", sock_info->name);
 774                                         ret = len;
 775                                         goto error;
 776                                 }
 777                                 ret = len;
 778                                 goto error;
 779                         } else {
 780                                 DBG("incorrect channel data message size: %zd", len);
 781                                 ret = -EINVAL;
 782                                 goto error;
 783                         }
 784                 }
 785                 args.channel.chan_data = chan_data;
 786                 args.channel.wakeup_fd = wakeup_fd;
 787                 if (ops->cmd)
 788                         ret = ops->cmd(lum->handle, lum->cmd,
 789                                         (unsigned long) &lum->u,
 790                                         &args, sock_info);
 791                 else
 792                         ret = -ENOSYS;
 793                 break;
 794         }
 795         case LTTNG_UST_STREAM:
 796         {
 797                 /* Receive shm_fd, wakeup_fd */
 798                 ret = ustcomm_recv_stream_from_sessiond(sock,
 799                         &lum->u.stream.len,
 800                         &args.stream.shm_fd,
 801                         &args.stream.wakeup_fd);
 802                 if (ret) {
 803                         goto error;
 804                 }
 805                 if (ops->cmd)
 806                         ret = ops->cmd(lum->handle, lum->cmd,
 807                                         (unsigned long) &lum->u,
 808                                         &args, sock_info);
 809                 else
 810                         ret = -ENOSYS;
 811                 break;
 812         }
 813         default:
 814                 if (ops->cmd)
 815                         ret = ops->cmd(lum->handle, lum->cmd,
 816                                         (unsigned long) &lum->u,
 817                                         &args, sock_info);
 818                 else
 819                         ret = -ENOSYS;
 820                 break;
 821         }
 822
 823         lur.handle = lum->handle;
 824         lur.cmd = lum->cmd;
 825         lur.ret_val = ret;
 826         if (ret >= 0) {
 827                 lur.ret_code = LTTNG_UST_OK;
 828         } else {
 829                 /*
 830                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 831                  * error that are not caused by the transport, except if
 832                  * we already have a more precise error message to
 833                  * report.
 834                  */
 835                 if (ret > -LTTNG_UST_ERR) {
 836                         /* Translate code to UST error. */
 837                         switch (ret) {
 838                         case -EEXIST:
 839                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 840                                 break;
 841                         case -EINVAL:
 842                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 843                                 break;
 844                         case -ENOENT:
 845                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 846                                 break;
 847                         case -EPERM:
 848                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 849                                 break;
 850                         case -ENOSYS:
 851                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 852                                 break;
 853                         default:
 854                                 lur.ret_code = -LTTNG_UST_ERR;
 855                                 break;
 856                         }
 857                 } else {
 858                         lur.ret_code = ret;
 859                 }
 860         }
 861         if (ret >= 0) {
 862                 switch (lum->cmd) {
 863                 case LTTNG_UST_TRACER_VERSION:
 864                         lur.u.version = lum->u.version;
 865                         break;
 866                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 867                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 868                         break;
 869                 }
 870         }
 871         DBG("Return value: %d", lur.ret_val);
 872         ret = send_reply(sock, &lur);
 873         if (ret < 0) {
 874                 DBG("error sending reply");
 875                 goto error;
 876         }
 877
 878         /*
 879          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 880          * after the reply.
 881          */
 882         if (lur.ret_code == LTTNG_UST_OK) {
 883                 switch (lum->cmd) {
 884                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 885                         len = ustcomm_send_unix_sock(sock,
 886                                 &args.field_list.entry,
 887                                 sizeof(args.field_list.entry));
 888                         if (len < 0) {
 889                                 ret = len;
 890                                 goto error;
 891                         }
 892                         if (len != sizeof(args.field_list.entry)) {
 893                                 ret = -EINVAL;
 894                                 goto error;
 895                         }
 896                 }
 897         }
 898
 899 error:
 900         ust_unlock();
 901
 902         /*
 903          * Performed delayed statedump operations outside of the UST
 904          * lock. We need to take the dynamic loader lock before we take
 905          * the UST lock internally within handle_pending_statedump().
 906           */
 907         handle_pending_statedump(sock_info);
 908
 909         return ret;
 910 }
 911
 912 static
 913 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 914 {
 915         int ret;
 916
 917         if (sock_info->root_handle != -1) {
 918                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
 919                 if (ret) {
 920                         ERR("Error unref root handle");
 921                 }
 922                 sock_info->root_handle = -1;
 923         }
 924         sock_info->constructor_sem_posted = 0;
 925
 926         /*
 927          * wait_shm_mmap, socket and notify socket are used by listener
 928          * threads outside of the ust lock, so we cannot tear them down
 929          * ourselves, because we cannot join on these threads. Leave
 930          * responsibility of cleaning up these resources to the OS
 931          * process exit.
 932          */
 933         if (exiting)
 934                 return;
 935
 936         if (sock_info->socket != -1) {
 937                 ret = ustcomm_close_unix_sock(sock_info->socket);
 938                 if (ret) {
 939                         ERR("Error closing ust cmd socket");
 940                 }
 941                 sock_info->socket = -1;
 942         }
 943         if (sock_info->notify_socket != -1) {
 944                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
 945                 if (ret) {
 946                         ERR("Error closing ust notify socket");
 947                 }
 948                 sock_info->notify_socket = -1;
 949         }
 950         if (sock_info->wait_shm_mmap) {
 951                 long page_size;
 952
 953                 page_size = sysconf(_SC_PAGE_SIZE);
 954                 if (page_size <= 0) {
 955                         if (!page_size) {
 956                                 errno = EINVAL;
 957                         }
 958                         PERROR("Error in sysconf(_SC_PAGE_SIZE)");
 959                 } else {
 960                         ret = munmap(sock_info->wait_shm_mmap, page_size);
 961                         if (ret) {
 962                                 ERR("Error unmapping wait shm");
 963                         }
 964                 }
 965                 sock_info->wait_shm_mmap = NULL;
 966         }
 967 }
 968
 969 /*
 970  * Using fork to set umask in the child process (not multi-thread safe).
 971  * We deal with the shm_open vs ftruncate race (happening when the
 972  * sessiond owns the shm and does not let everybody modify it, to ensure
 973  * safety against shm_unlink) by simply letting the mmap fail and
 974  * retrying after a few seconds.
 975  * For global shm, everybody has rw access to it until the sessiond
 976  * starts.
 977  */
 978 static
 979 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 980 {
 981         int wait_shm_fd, ret;
 982         pid_t pid;
 983
 984         /*
 985          * Try to open read-only.
 986          */
 987         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 988         if (wait_shm_fd >= 0) {
 989                 int32_t tmp_read;
 990                 ssize_t len;
 991                 size_t bytes_read = 0;
 992
 993                 /*
 994                  * Try to read the fd. If unable to do so, try opening
 995                  * it in write mode.
 996                  */
 997                 do {
 998                         len = read(wait_shm_fd,
 999                                 &((char *) &tmp_read)[bytes_read],
1000                                 sizeof(tmp_read) - bytes_read);
1001                         if (len > 0) {
1002                                 bytes_read += len;
1003                         }
1004                 } while ((len < 0 && errno == EINTR)
1005                         || (len > 0 && bytes_read < sizeof(tmp_read)));
1006                 if (bytes_read != sizeof(tmp_read)) {
1007                         ret = close(wait_shm_fd);
1008                         if (ret) {
1009                                 ERR("close wait_shm_fd");
1010                         }
1011                         goto open_write;
1012                 }
1013                 goto end;
1014         } else if (wait_shm_fd < 0 && errno != ENOENT) {
1015                 /*
1016                  * Real-only open did not work, and it's not because the
1017                  * entry was not present. It's a failure that prohibits
1018                  * using shm.
1019                  */
1020                 ERR("Error opening shm %s", sock_info->wait_shm_path);
1021                 goto end;
1022         }
1023
1024 open_write:
1025         /*
1026          * If the open failed because the file did not exist, or because
1027          * the file was not truncated yet, try creating it ourself.
1028          */
1029         URCU_TLS(lttng_ust_nest_count)++;
1030         pid = fork();
1031         URCU_TLS(lttng_ust_nest_count)--;
1032         if (pid > 0) {
1033                 int status;
1034
1035                 /*
1036                  * Parent: wait for child to return, in which case the
1037                  * shared memory map will have been created.
1038                  */
1039                 pid = wait(&status);
1040                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
1041                         wait_shm_fd = -1;
1042                         goto end;
1043                 }
1044                 /*
1045                  * Try to open read-only again after creation.
1046                  */
1047                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1048                 if (wait_shm_fd < 0) {
1049                         /*
1050                          * Real-only open did not work. It's a failure
1051                          * that prohibits using shm.
1052                          */
1053                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1054                         goto end;
1055                 }
1056                 goto end;
1057         } else if (pid == 0) {
1058                 int create_mode;
1059
1060                 /* Child */
1061                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1062                 if (sock_info->global)
1063                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1064                 /*
1065                  * We're alone in a child process, so we can modify the
1066                  * process-wide umask.
1067                  */
1068                 umask(~create_mode);
1069                 /*
1070                  * Try creating shm (or get rw access).
1071                  * We don't do an exclusive open, because we allow other
1072                  * processes to create+ftruncate it concurrently.
1073                  */
1074                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1075                                 O_RDWR | O_CREAT, create_mode);
1076                 if (wait_shm_fd >= 0) {
1077                         ret = ftruncate(wait_shm_fd, mmap_size);
1078                         if (ret) {
1079                                 PERROR("ftruncate");
1080                                 _exit(EXIT_FAILURE);
1081                         }
1082                         _exit(EXIT_SUCCESS);
1083                 }
1084                 /*
1085                  * For local shm, we need to have rw access to accept
1086                  * opening it: this means the local sessiond will be
1087                  * able to wake us up. For global shm, we open it even
1088                  * if rw access is not granted, because the root.root
1089                  * sessiond will be able to override all rights and wake
1090                  * us up.
1091                  */
1092                 if (!sock_info->global && errno != EACCES) {
1093                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1094                         _exit(EXIT_FAILURE);
1095                 }
1096                 /*
1097                  * The shm exists, but we cannot open it RW. Report
1098                  * success.
1099                  */
1100                 _exit(EXIT_SUCCESS);
1101         } else {
1102                 return -1;
1103         }
1104 end:
1105         if (wait_shm_fd >= 0 && !sock_info->global) {
1106                 struct stat statbuf;
1107
1108                 /*
1109                  * Ensure that our user is the owner of the shm file for
1110                  * local shm. If we do not own the file, it means our
1111                  * sessiond will not have access to wake us up (there is
1112                  * probably a rogue process trying to fake our
1113                  * sessiond). Fallback to polling method in this case.
1114                  */
1115                 ret = fstat(wait_shm_fd, &statbuf);
1116                 if (ret) {
1117                         PERROR("fstat");
1118                         goto error_close;
1119                 }
1120                 if (statbuf.st_uid != getuid())
1121                         goto error_close;
1122         }
1123         return wait_shm_fd;
1124
1125 error_close:
1126         ret = close(wait_shm_fd);
1127         if (ret) {
1128                 PERROR("Error closing fd");
1129         }
1130         return -1;
1131 }
1132
1133 static
1134 char *get_map_shm(struct sock_info *sock_info)
1135 {
1136         long page_size;
1137         int wait_shm_fd, ret;
1138         char *wait_shm_mmap;
1139
1140         page_size = sysconf(_SC_PAGE_SIZE);
1141         if (page_size <= 0) {
1142                 if (!page_size) {
1143                         errno = EINVAL;
1144                 }
1145                 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1146                 goto error;
1147         }
1148
1149         wait_shm_fd = get_wait_shm(sock_info, page_size);
1150         if (wait_shm_fd < 0) {
1151                 goto error;
1152         }
1153         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1154                   MAP_SHARED, wait_shm_fd, 0);
1155         /* close shm fd immediately after taking the mmap reference */
1156         ret = close(wait_shm_fd);
1157         if (ret) {
1158                 PERROR("Error closing fd");
1159         }
1160         if (wait_shm_mmap == MAP_FAILED) {
1161                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1162                 goto error;
1163         }
1164         return wait_shm_mmap;
1165
1166 error:
1167         return NULL;
1168 }
1169
1170 static
1171 void wait_for_sessiond(struct sock_info *sock_info)
1172 {
1173         if (ust_lock()) {
1174                 goto quit;
1175         }
1176         if (wait_poll_fallback) {
1177                 goto error;
1178         }
1179         if (!sock_info->wait_shm_mmap) {
1180                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1181                 if (!sock_info->wait_shm_mmap)
1182                         goto error;
1183         }
1184         ust_unlock();
1185
1186         DBG("Waiting for %s apps sessiond", sock_info->name);
1187         /* Wait for futex wakeup */
1188         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
1189                 goto end_wait;
1190
1191         while (futex_async((int32_t *) sock_info->wait_shm_mmap,
1192                         FUTEX_WAIT, 0, NULL, NULL, 0)) {
1193                 switch (errno) {
1194                 case EWOULDBLOCK:
1195                         /* Value already changed. */
1196                         goto end_wait;
1197                 case EINTR:
1198                         /* Retry if interrupted by signal. */
1199                         break;  /* Get out of switch. */
1200                 case EFAULT:
1201                         wait_poll_fallback = 1;
1202                         DBG(
1203 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1204 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1205 "Please upgrade your kernel "
1206 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1207 "mainline). LTTng-UST will use polling mode fallback.");
1208                         if (ust_debug())
1209                                 PERROR("futex");
1210                         goto end_wait;
1211                 }
1212         }
1213 end_wait:
1214         return;
1215
1216 quit:
1217         ust_unlock();
1218         return;
1219
1220 error:
1221         ust_unlock();
1222         return;
1223 }
1224
1225 /*
1226  * This thread does not allocate any resource, except within
1227  * handle_message, within mutex protection. This mutex protects against
1228  * fork and exit.
1229  * The other moment it allocates resources is at socket connection, which
1230  * is also protected by the mutex.
1231  */
1232 static
1233 void *ust_listener_thread(void *arg)
1234 {
1235         struct sock_info *sock_info = arg;
1236         int sock, ret, prev_connect_failed = 0, has_waited = 0;
1237         long timeout;
1238
1239         /* Restart trying to connect to the session daemon */
1240 restart:
1241         if (prev_connect_failed) {
1242                 /* Wait for sessiond availability with pipe */
1243                 wait_for_sessiond(sock_info);
1244                 if (has_waited) {
1245                         has_waited = 0;
1246                         /*
1247                          * Sleep for 5 seconds before retrying after a
1248                          * sequence of failure / wait / failure. This
1249                          * deals with a killed or broken session daemon.
1250                          */
1251                         sleep(5);
1252                 } else {
1253                         has_waited = 1;
1254                 }
1255                 prev_connect_failed = 0;
1256         }
1257
1258         if (sock_info->socket != -1) {
1259                 ret = ustcomm_close_unix_sock(sock_info->socket);
1260                 if (ret) {
1261                         ERR("Error closing %s ust cmd socket",
1262                                 sock_info->name);
1263                 }
1264                 sock_info->socket = -1;
1265         }
1266         if (sock_info->notify_socket != -1) {
1267                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1268                 if (ret) {
1269                         ERR("Error closing %s ust notify socket",
1270                                 sock_info->name);
1271                 }
1272                 sock_info->notify_socket = -1;
1273         }
1274
1275         /*
1276          * Register. We need to perform both connect and sending
1277          * registration message before doing the next connect otherwise
1278          * we may reach unix socket connect queue max limits and block
1279          * on the 2nd connect while the session daemon is awaiting the
1280          * first connect registration message.
1281          */
1282         /* Connect cmd socket */
1283         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
1284         if (ret < 0) {
1285                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1286                 prev_connect_failed = 1;
1287
1288                 if (ust_lock()) {
1289                         goto quit;
1290                 }
1291
1292                 /*
1293                  * If we cannot find the sessiond daemon, don't delay
1294                  * constructor execution.
1295                  */
1296                 ret = handle_register_done(sock_info);
1297                 assert(!ret);
1298                 ust_unlock();
1299                 goto restart;
1300         }
1301         sock_info->socket = ret;
1302
1303         if (ust_lock()) {
1304                 goto quit;
1305         }
1306
1307         /*
1308          * Create only one root handle per listener thread for the whole
1309          * process lifetime, so we ensure we get ID which is statically
1310          * assigned to the root handle.
1311          */
1312         if (sock_info->root_handle == -1) {
1313                 ret = lttng_abi_create_root_handle();
1314                 if (ret < 0) {
1315                         ERR("Error creating root handle");
1316                         goto quit;
1317                 }
1318                 sock_info->root_handle = ret;
1319         }
1320
1321         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1322         if (ret < 0) {
1323                 ERR("Error registering to %s ust cmd socket",
1324                         sock_info->name);
1325                 prev_connect_failed = 1;
1326                 /*
1327                  * If we cannot register to the sessiond daemon, don't
1328                  * delay constructor execution.
1329                  */
1330                 ret = handle_register_done(sock_info);
1331                 assert(!ret);
1332                 ust_unlock();
1333                 goto restart;
1334         }
1335
1336         ust_unlock();
1337
1338         /* Connect notify 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->notify_socket = ret;
1358
1359         timeout = get_notify_sock_timeout();
1360         if (timeout >= 0) {
1361                 /*
1362                  * Give at least 10ms to sessiond to reply to
1363                  * notifications.
1364                  */
1365                 if (timeout < 10)
1366                         timeout = 10;
1367                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1368                                 timeout);
1369                 if (ret < 0) {
1370                         WARN("Error setting socket receive timeout");
1371                 }
1372                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1373                                 timeout);
1374                 if (ret < 0) {
1375                         WARN("Error setting socket send timeout");
1376                 }
1377         } else if (timeout < -1) {
1378                 WARN("Unsupported timeout value %ld", timeout);
1379         }
1380
1381         if (ust_lock()) {
1382                 goto quit;
1383         }
1384
1385         ret = register_to_sessiond(sock_info->notify_socket,
1386                         USTCTL_SOCKET_NOTIFY);
1387         if (ret < 0) {
1388                 ERR("Error registering to %s ust notify socket",
1389                         sock_info->name);
1390                 prev_connect_failed = 1;
1391                 /*
1392                  * If we cannot register to the sessiond daemon, don't
1393                  * delay constructor execution.
1394                  */
1395                 ret = handle_register_done(sock_info);
1396                 assert(!ret);
1397                 ust_unlock();
1398                 goto restart;
1399         }
1400         sock = sock_info->socket;
1401
1402         ust_unlock();
1403
1404         for (;;) {
1405                 ssize_t len;
1406                 struct ustcomm_ust_msg lum;
1407
1408                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1409                 switch (len) {
1410                 case 0: /* orderly shutdown */
1411                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1412                         if (ust_lock()) {
1413                                 goto quit;
1414                         }
1415                         /*
1416                          * Either sessiond has shutdown or refused us by closing the socket.
1417                          * In either case, we don't want to delay construction execution,
1418                          * and we need to wait before retry.
1419                          */
1420                         prev_connect_failed = 1;
1421                         /*
1422                          * If we cannot register to the sessiond daemon, don't
1423                          * delay constructor execution.
1424                          */
1425                         ret = handle_register_done(sock_info);
1426                         assert(!ret);
1427                         ust_unlock();
1428                         goto end;
1429                 case sizeof(lum):
1430                         print_cmd(lum.cmd, lum.handle);
1431                         ret = handle_message(sock_info, sock, &lum);
1432                         if (ret) {
1433                                 ERR("Error handling message for %s socket",
1434                                         sock_info->name);
1435                                 /*
1436                                  * Close socket if protocol error is
1437                                  * detected.
1438                                  */
1439                                 goto end;
1440                         }
1441                         continue;
1442                 default:
1443                         if (len < 0) {
1444                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1445                         } else {
1446                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1447                         }
1448                         if (len == -ECONNRESET) {
1449                                 DBG("%s remote end closed connection", sock_info->name);
1450                                 goto end;
1451                         }
1452                         goto end;
1453                 }
1454
1455         }
1456 end:
1457         if (ust_lock()) {
1458                 goto quit;
1459         }
1460         /* Cleanup socket handles before trying to reconnect */
1461         lttng_ust_objd_table_owner_cleanup(sock_info);
1462         ust_unlock();
1463         goto restart;   /* try to reconnect */
1464
1465 quit:
1466         ust_unlock();
1467
1468         pthread_mutex_lock(&ust_exit_mutex);
1469         sock_info->thread_active = 0;
1470         pthread_mutex_unlock(&ust_exit_mutex);
1471         return NULL;
1472 }
1473
1474 /*
1475  * Weak symbol to call when the ust malloc wrapper is not loaded.
1476  */
1477 __attribute__((weak))
1478 void lttng_ust_malloc_wrapper_init(void)
1479 {
1480 }
1481
1482 /*
1483  * sessiond monitoring thread: monitor presence of global and per-user
1484  * sessiond by polling the application common named pipe.
1485  */
1486 void __attribute__((constructor)) lttng_ust_init(void)
1487 {
1488         struct timespec constructor_timeout;
1489         sigset_t sig_all_blocked, orig_parent_mask;
1490         pthread_attr_t thread_attr;
1491         int timeout_mode;
1492         int ret;
1493
1494         if (uatomic_xchg(&initialized, 1) == 1)
1495                 return;
1496
1497         /*
1498          * Fixup interdependency between TLS fixup mutex (which happens
1499          * to be the dynamic linker mutex) and ust_lock, taken within
1500          * the ust lock.
1501          */
1502         lttng_fixup_urcu_bp_tls();
1503         lttng_fixup_ringbuffer_tls();
1504         lttng_fixup_vtid_tls();
1505         lttng_fixup_nest_count_tls();
1506         lttng_fixup_procname_tls();
1507         lttng_fixup_ust_mutex_nest_tls();
1508
1509         /*
1510          * We want precise control over the order in which we construct
1511          * our sub-libraries vs starting to receive commands from
1512          * sessiond (otherwise leading to errors when trying to create
1513          * sessiond before the init functions are completed).
1514          */
1515         init_usterr();
1516         init_tracepoint();
1517         lttng_ust_clock_init();
1518         lttng_ust_getcpu_init();
1519         lttng_ust_statedump_init();
1520         lttng_ring_buffer_metadata_client_init();
1521         lttng_ring_buffer_client_overwrite_init();
1522         lttng_ring_buffer_client_overwrite_rt_init();
1523         lttng_ring_buffer_client_discard_init();
1524         lttng_ring_buffer_client_discard_rt_init();
1525         lttng_perf_counter_init();
1526         /*
1527          * Invoke ust malloc wrapper init before starting other threads.
1528          */
1529         lttng_ust_malloc_wrapper_init();
1530
1531         timeout_mode = get_constructor_timeout(&constructor_timeout);
1532
1533         ret = sem_init(&constructor_wait, 0, 0);
1534         assert(!ret);
1535
1536         ret = setup_local_apps();
1537         if (ret) {
1538                 DBG("local apps setup returned %d", ret);
1539         }
1540
1541         /* A new thread created by pthread_create inherits the signal mask
1542          * from the parent. To avoid any signal being received by the
1543          * listener thread, we block all signals temporarily in the parent,
1544          * while we create the listener thread.
1545          */
1546         sigfillset(&sig_all_blocked);
1547         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1548         if (ret) {
1549                 ERR("pthread_sigmask: %s", strerror(ret));
1550         }
1551
1552         ret = pthread_attr_init(&thread_attr);
1553         if (ret) {
1554                 ERR("pthread_attr_init: %s", strerror(ret));
1555         }
1556         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1557         if (ret) {
1558                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1559         }
1560
1561         pthread_mutex_lock(&ust_exit_mutex);
1562         ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1563                         ust_listener_thread, &global_apps);
1564         if (ret) {
1565                 ERR("pthread_create global: %s", strerror(ret));
1566         }
1567         global_apps.thread_active = 1;
1568         pthread_mutex_unlock(&ust_exit_mutex);
1569
1570         if (local_apps.allowed) {
1571                 pthread_mutex_lock(&ust_exit_mutex);
1572                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1573                                 ust_listener_thread, &local_apps);
1574                 if (ret) {
1575                         ERR("pthread_create local: %s", strerror(ret));
1576                 }
1577                 local_apps.thread_active = 1;
1578                 pthread_mutex_unlock(&ust_exit_mutex);
1579         } else {
1580                 handle_register_done(&local_apps);
1581         }
1582         ret = pthread_attr_destroy(&thread_attr);
1583         if (ret) {
1584                 ERR("pthread_attr_destroy: %s", strerror(ret));
1585         }
1586
1587         /* Restore original signal mask in parent */
1588         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1589         if (ret) {
1590                 ERR("pthread_sigmask: %s", strerror(ret));
1591         }
1592
1593         switch (timeout_mode) {
1594         case 1: /* timeout wait */
1595                 do {
1596                         ret = sem_timedwait(&constructor_wait,
1597                                         &constructor_timeout);
1598                 } while (ret < 0 && errno == EINTR);
1599                 if (ret < 0 && errno == ETIMEDOUT) {
1600                         ERR("Timed out waiting for lttng-sessiond");
1601                 } else {
1602                         assert(!ret);
1603                 }
1604                 break;
1605         case -1:/* wait forever */
1606                 do {
1607                         ret = sem_wait(&constructor_wait);
1608                 } while (ret < 0 && errno == EINTR);
1609                 assert(!ret);
1610                 break;
1611         case 0: /* no timeout */
1612                 break;
1613         }
1614 }
1615
1616 static
1617 void lttng_ust_cleanup(int exiting)
1618 {
1619         cleanup_sock_info(&global_apps, exiting);
1620         cleanup_sock_info(&local_apps, exiting);
1621         /*
1622          * The teardown in this function all affect data structures
1623          * accessed under the UST lock by the listener thread. This
1624          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1625          * that none of these threads are accessing this data at this
1626          * point.
1627          */
1628         lttng_ust_abi_exit();
1629         lttng_ust_events_exit();
1630         lttng_perf_counter_exit();
1631         lttng_ring_buffer_client_discard_rt_exit();
1632         lttng_ring_buffer_client_discard_exit();
1633         lttng_ring_buffer_client_overwrite_rt_exit();
1634         lttng_ring_buffer_client_overwrite_exit();
1635         lttng_ring_buffer_metadata_client_exit();
1636         lttng_ust_statedump_destroy();
1637         exit_tracepoint();
1638         if (!exiting) {
1639                 /* Reinitialize values for fork */
1640                 sem_count = 2;
1641                 lttng_ust_comm_should_quit = 0;
1642                 initialized = 0;
1643         }
1644 }
1645
1646 void __attribute__((destructor)) lttng_ust_exit(void)
1647 {
1648         int ret;
1649
1650         /*
1651          * Using pthread_cancel here because:
1652          * A) we don't want to hang application teardown.
1653          * B) the thread is not allocating any resource.
1654          */
1655
1656         /*
1657          * Require the communication thread to quit. Synchronize with
1658          * mutexes to ensure it is not in a mutex critical section when
1659          * pthread_cancel is later called.
1660          */
1661         ust_lock_nocheck();
1662         lttng_ust_comm_should_quit = 1;
1663         ust_unlock();
1664
1665         pthread_mutex_lock(&ust_exit_mutex);
1666         /* cancel threads */
1667         if (global_apps.thread_active) {
1668                 ret = pthread_cancel(global_apps.ust_listener);
1669                 if (ret) {
1670                         ERR("Error cancelling global ust listener thread: %s",
1671                                 strerror(ret));
1672                 } else {
1673                         global_apps.thread_active = 0;
1674                 }
1675         }
1676         if (local_apps.thread_active) {
1677                 ret = pthread_cancel(local_apps.ust_listener);
1678                 if (ret) {
1679                         ERR("Error cancelling local ust listener thread: %s",
1680                                 strerror(ret));
1681                 } else {
1682                         local_apps.thread_active = 0;
1683                 }
1684         }
1685         pthread_mutex_unlock(&ust_exit_mutex);
1686
1687         /*
1688          * Do NOT join threads: use of sys_futex makes it impossible to
1689          * join the threads without using async-cancel, but async-cancel
1690          * is delivered by a signal, which could hit the target thread
1691          * anywhere in its code path, including while the ust_lock() is
1692          * held, causing a deadlock for the other thread. Let the OS
1693          * cleanup the threads if there are stalled in a syscall.
1694          */
1695         lttng_ust_cleanup(1);
1696 }
1697
1698 /*
1699  * We exclude the worker threads across fork and clone (except
1700  * CLONE_VM), because these system calls only keep the forking thread
1701  * running in the child.  Therefore, we don't want to call fork or clone
1702  * in the middle of an tracepoint or ust tracing state modification.
1703  * Holding this mutex protects these structures across fork and clone.
1704  */
1705 void ust_before_fork(sigset_t *save_sigset)
1706 {
1707         /*
1708          * Disable signals. This is to avoid that the child intervenes
1709          * before it is properly setup for tracing. It is safer to
1710          * disable all signals, because then we know we are not breaking
1711          * anything by restoring the original mask.
1712          */
1713         sigset_t all_sigs;
1714         int ret;
1715
1716         if (URCU_TLS(lttng_ust_nest_count))
1717                 return;
1718         /* Disable signals */
1719         sigfillset(&all_sigs);
1720         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1721         if (ret == -1) {
1722                 PERROR("sigprocmask");
1723         }
1724
1725         pthread_mutex_lock(&ust_fork_mutex);
1726
1727         ust_lock_nocheck();
1728         rcu_bp_before_fork();
1729 }
1730
1731 static void ust_after_fork_common(sigset_t *restore_sigset)
1732 {
1733         int ret;
1734
1735         DBG("process %d", getpid());
1736         ust_unlock();
1737
1738         pthread_mutex_unlock(&ust_fork_mutex);
1739
1740         /* Restore signals */
1741         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1742         if (ret == -1) {
1743                 PERROR("sigprocmask");
1744         }
1745 }
1746
1747 void ust_after_fork_parent(sigset_t *restore_sigset)
1748 {
1749         if (URCU_TLS(lttng_ust_nest_count))
1750                 return;
1751         DBG("process %d", getpid());
1752         rcu_bp_after_fork_parent();
1753         /* Release mutexes and reenable signals */
1754         ust_after_fork_common(restore_sigset);
1755 }
1756
1757 /*
1758  * After fork, in the child, we need to cleanup all the leftover state,
1759  * except the worker thread which already magically disappeared thanks
1760  * to the weird Linux fork semantics. After tyding up, we call
1761  * lttng_ust_init() again to start over as a new PID.
1762  *
1763  * This is meant for forks() that have tracing in the child between the
1764  * fork and following exec call (if there is any).
1765  */
1766 void ust_after_fork_child(sigset_t *restore_sigset)
1767 {
1768         if (URCU_TLS(lttng_ust_nest_count))
1769                 return;
1770         DBG("process %d", getpid());
1771         /* Release urcu mutexes */
1772         rcu_bp_after_fork_child();
1773         lttng_ust_cleanup(0);
1774         lttng_context_vtid_reset();
1775         /* Release mutexes and reenable signals */
1776         ust_after_fork_common(restore_sigset);
1777         lttng_ust_init();
1778 }
1779
1780 void lttng_ust_sockinfo_session_enabled(void *owner)
1781 {
1782         struct sock_info *sock_info = owner;
1783         sock_info->statedump_pending = 1;
1784 }