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