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