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