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