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 <ust-comm.h>
  45 #include <usterr-signal-safe.h>
  46 #include "tracepoint-internal.h"
  47 #include "ltt-tracer-core.h"
  48 #include "compat.h"
  49 #include "../libringbuffer/tlsfixup.h"
  50
  51 /*
  52  * Has lttng ust comm constructor been called ?
  53  */
  54 static int initialized;
  55
  56 /*
  57  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  58  * Held when handling a command, also held by fork() to deal with
  59  * removal of threads, and by exit path.
  60  */
  61
  62 /* Should the ust comm thread quit ? */
  63 static int lttng_ust_comm_should_quit;
  64
  65 /*
  66  * Wait for either of these before continuing to the main
  67  * program:
  68  * - the register_done message from sessiond daemon
  69  *   (will let the sessiond daemon enable sessions before main
  70  *   starts.)
  71  * - sessiond daemon is not reachable.
  72  * - timeout (ensuring applications are resilient to session
  73  *   daemon problems).
  74  */
  75 static sem_t constructor_wait;
  76 /*
  77  * Doing this for both the global and local sessiond.
  78  */
  79 static int sem_count = { 2 };
  80
  81 /*
  82  * Counting nesting within lttng-ust. Used to ensure that calling fork()
  83  * from liblttng-ust does not execute the pre/post fork handlers.
  84  */
  85 static int __thread lttng_ust_nest_count;
  86
  87 /*
  88  * Info about socket and associated listener thread.
  89  */
  90 struct sock_info {
  91         const char *name;
  92         pthread_t ust_listener; /* listener thread */
  93         int root_handle;
  94         int constructor_sem_posted;
  95         int allowed;
  96         int global;
  97
  98         char sock_path[PATH_MAX];
  99         int socket;
 100
 101         char wait_shm_path[PATH_MAX];
 102         char *wait_shm_mmap;
 103 };
 104
 105 /* Socket from app (connect) to session daemon (listen) for communication */
 106 struct sock_info global_apps = {
 107         .name = "global",
 108         .global = 1,
 109
 110         .root_handle = -1,
 111         .allowed = 1,
 112
 113         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
 114         .socket = -1,
 115
 116         .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
 117 };
 118
 119 /* TODO: allow global_apps_sock_path override */
 120
 121 struct sock_info local_apps = {
 122         .name = "local",
 123         .global = 0,
 124         .root_handle = -1,
 125         .allowed = 0,   /* Check setuid bit first */
 126
 127         .socket = -1,
 128 };
 129
 130 static int wait_poll_fallback;
 131
 132 extern void ltt_ring_buffer_client_overwrite_init(void);
 133 extern void ltt_ring_buffer_client_discard_init(void);
 134 extern void ltt_ring_buffer_metadata_client_init(void);
 135 extern void ltt_ring_buffer_client_overwrite_exit(void);
 136 extern void ltt_ring_buffer_client_discard_exit(void);
 137 extern void ltt_ring_buffer_metadata_client_exit(void);
 138
 139 /*
 140  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 141  */
 142 static
 143 void lttng_fixup_nest_count_tls(void)
 144 {
 145         asm volatile ("" : : "m" (lttng_ust_nest_count));
 146 }
 147
 148 static
 149 int setup_local_apps(void)
 150 {
 151         const char *home_dir;
 152         uid_t uid;
 153
 154         uid = getuid();
 155         /*
 156          * Disallow per-user tracing for setuid binaries.
 157          */
 158         if (uid != geteuid()) {
 159                 local_apps.allowed = 0;
 160                 return 0;
 161         } else {
 162                 local_apps.allowed = 1;
 163         }
 164         home_dir = (const char *) getenv("HOME");
 165         if (!home_dir)
 166                 return -ENOENT;
 167         snprintf(local_apps.sock_path, PATH_MAX,
 168                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 169         snprintf(local_apps.wait_shm_path, PATH_MAX,
 170                  DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
 171         return 0;
 172 }
 173
 174 static
 175 int register_app_to_sessiond(int socket)
 176 {
 177         ssize_t ret;
 178         struct {
 179                 uint32_t major;
 180                 uint32_t minor;
 181                 pid_t pid;
 182                 pid_t ppid;
 183                 uid_t uid;
 184                 gid_t gid;
 185                 uint32_t bits_per_long;
 186                 char name[16];  /* process name */
 187         } reg_msg;
 188
 189         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 190         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 191         reg_msg.pid = getpid();
 192         reg_msg.ppid = getppid();
 193         reg_msg.uid = getuid();
 194         reg_msg.gid = getgid();
 195         reg_msg.bits_per_long = CAA_BITS_PER_LONG;
 196         lttng_ust_getprocname(reg_msg.name);
 197
 198         ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 199         if (ret >= 0 && ret != sizeof(reg_msg))
 200                 return -EIO;
 201         return ret;
 202 }
 203
 204 static
 205 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 206 {
 207         ssize_t len;
 208
 209         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 210         switch (len) {
 211         case sizeof(*lur):
 212                 DBG("message successfully sent");
 213                 return 0;
 214         case -1:
 215                 if (errno == ECONNRESET) {
 216                         printf("remote end closed connection\n");
 217                         return 0;
 218                 }
 219                 return -1;
 220         default:
 221                 printf("incorrect message size: %zd\n", len);
 222                 return -1;
 223         }
 224 }
 225
 226 static
 227 int handle_register_done(struct sock_info *sock_info)
 228 {
 229         int ret;
 230
 231         if (sock_info->constructor_sem_posted)
 232                 return 0;
 233         sock_info->constructor_sem_posted = 1;
 234         if (uatomic_read(&sem_count) <= 0) {
 235                 return 0;
 236         }
 237         ret = uatomic_add_return(&sem_count, -1);
 238         if (ret == 0) {
 239                 ret = sem_post(&constructor_wait);
 240                 assert(!ret);
 241         }
 242         return 0;
 243 }
 244
 245 static
 246 int handle_message(struct sock_info *sock_info,
 247                 int sock, struct ustcomm_ust_msg *lum)
 248 {
 249         int ret = 0;
 250         const struct lttng_ust_objd_ops *ops;
 251         struct ustcomm_ust_reply lur;
 252         int shm_fd, wait_fd;
 253         union ust_args args;
 254         ssize_t len;
 255
 256         ust_lock();
 257
 258         memset(&lur, 0, sizeof(lur));
 259
 260         if (lttng_ust_comm_should_quit) {
 261                 ret = -EPERM;
 262                 goto end;
 263         }
 264
 265         ops = objd_ops(lum->handle);
 266         if (!ops) {
 267                 ret = -ENOENT;
 268                 goto end;
 269         }
 270
 271         switch (lum->cmd) {
 272         case LTTNG_UST_REGISTER_DONE:
 273                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 274                         ret = handle_register_done(sock_info);
 275                 else
 276                         ret = -EINVAL;
 277                 break;
 278         case LTTNG_UST_RELEASE:
 279                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 280                         ret = -EPERM;
 281                 else
 282                         ret = lttng_ust_objd_unref(lum->handle);
 283                 break;
 284         case LTTNG_UST_FILTER:
 285         {
 286                 /* Receive filter data */
 287                 struct {
 288                         uint16_t len;
 289                         uint16_t reloc_offset;
 290                         char data[FILTER_BYTECODE_MAX_LEN];
 291                 } filter_data;
 292
 293                 if (lum->u.filter.data_size > 65536) {
 294                         ERR("Filter data size is too large: %u bytes\n",
 295                                 lum->u.filter.data_size);
 296                         ret = -EINVAL;
 297                         goto error;
 298                 }
 299                 len = ustcomm_recv_unix_sock(sock, filter_data.data,
 300                                 lum->u.filter.data_size);
 301                 switch (len) {
 302                 case 0: /* orderly shutdown */
 303                         ret = 0;
 304                         goto error;
 305                 case -1:
 306                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 307                         if (errno == ECONNRESET) {
 308                                 ERR("%s remote end closed connection\n", sock_info->name);
 309                                 ret = -EINVAL;
 310                                 goto error;
 311                         }
 312                         ret = -EINVAL;
 313                         goto end;
 314                 default:
 315                         if (len == lum->u.filter.data_size) {
 316                                 DBG("filter data received\n");
 317                                 break;
 318                         } else {
 319                                 ERR("incorrect filter data message size: %zd\n", len);
 320                                 ret = -EINVAL;
 321                                 goto end;
 322                         }
 323                 }
 324                 filter_data.len = lum->u.filter.data_size;
 325                 filter_data.reloc_offset = lum->u.filter.reloc_offset;
 326                 if (ops->cmd)
 327                         ret = ops->cmd(lum->handle, lum->cmd,
 328                                         (unsigned long) &filter_data,
 329                                         &args);
 330                 else
 331                         ret = -ENOSYS;
 332                 break;
 333         }
 334         default:
 335                 if (ops->cmd)
 336                         ret = ops->cmd(lum->handle, lum->cmd,
 337                                         (unsigned long) &lum->u,
 338                                         &args);
 339                 else
 340                         ret = -ENOSYS;
 341                 break;
 342         }
 343
 344 end:
 345         lur.handle = lum->handle;
 346         lur.cmd = lum->cmd;
 347         lur.ret_val = ret;
 348         if (ret >= 0) {
 349                 lur.ret_code = USTCOMM_OK;
 350         } else {
 351                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 352                 lur.ret_code = ret;
 353         }
 354         if (ret >= 0) {
 355                 switch (lum->cmd) {
 356                 case LTTNG_UST_STREAM:
 357                         /*
 358                          * Special-case reply to send stream info.
 359                          * Use lum.u output.
 360                          */
 361                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 362                         shm_fd = *args.stream.shm_fd;
 363                         wait_fd = *args.stream.wait_fd;
 364                         break;
 365                 case LTTNG_UST_METADATA:
 366                 case LTTNG_UST_CHANNEL:
 367                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 368                         shm_fd = *args.channel.shm_fd;
 369                         wait_fd = *args.channel.wait_fd;
 370                         break;
 371                 case LTTNG_UST_TRACER_VERSION:
 372                         lur.u.version = lum->u.version;
 373                         break;
 374                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 375                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 376                         break;
 377                 }
 378         }
 379         ret = send_reply(sock, &lur);
 380         if (ret < 0) {
 381                 perror("error sending reply");
 382                 goto error;
 383         }
 384
 385         if ((lum->cmd == LTTNG_UST_STREAM
 386              || lum->cmd == LTTNG_UST_CHANNEL
 387              || lum->cmd == LTTNG_UST_METADATA)
 388                         && lur.ret_code == USTCOMM_OK) {
 389                 int sendret = 0;
 390
 391                 /* we also need to send the file descriptors. */
 392                 ret = ustcomm_send_fds_unix_sock(sock,
 393                         &shm_fd, &shm_fd,
 394                         1, sizeof(int));
 395                 if (ret < 0) {
 396                         perror("send shm_fd");
 397                         sendret = ret;
 398                 }
 399                 /*
 400                  * The sessiond expects 2 file descriptors, even upon
 401                  * error.
 402                  */
 403                 ret = ustcomm_send_fds_unix_sock(sock,
 404                         &wait_fd, &wait_fd,
 405                         1, sizeof(int));
 406                 if (ret < 0) {
 407                         perror("send wait_fd");
 408                         goto error;
 409                 }
 410                 if (sendret) {
 411                         ret = sendret;
 412                         goto error;
 413                 }
 414         }
 415         /*
 416          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 417          * after the reply.
 418          */
 419         if (lur.ret_code == USTCOMM_OK) {
 420                 switch (lum->cmd) {
 421                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 422                         len = ustcomm_send_unix_sock(sock,
 423                                 &args.field_list.entry,
 424                                 sizeof(args.field_list.entry));
 425                         if (len != sizeof(args.field_list.entry)) {
 426                                 ret = -1;
 427                                 goto error;
 428                         }
 429                 }
 430         }
 431         /*
 432          * We still have the memory map reference, and the fds have been
 433          * sent to the sessiond. We can therefore close those fds. Note
 434          * that we keep the write side of the wait_fd open, but close
 435          * the read side.
 436          */
 437         if (lur.ret_code == USTCOMM_OK) {
 438                 switch (lum->cmd) {
 439                 case LTTNG_UST_STREAM:
 440                         if (shm_fd >= 0) {
 441                                 ret = close(shm_fd);
 442                                 if (ret) {
 443                                         PERROR("Error closing stream shm_fd");
 444                                 }
 445                                 *args.stream.shm_fd = -1;
 446                         }
 447                         if (wait_fd >= 0) {
 448                                 ret = close(wait_fd);
 449                                 if (ret) {
 450                                         PERROR("Error closing stream wait_fd");
 451                                 }
 452                                 *args.stream.wait_fd = -1;
 453                         }
 454                         break;
 455                 case LTTNG_UST_METADATA:
 456                 case LTTNG_UST_CHANNEL:
 457                         if (shm_fd >= 0) {
 458                                 ret = close(shm_fd);
 459                                 if (ret) {
 460                                         PERROR("Error closing channel shm_fd");
 461                                 }
 462                                 *args.channel.shm_fd = -1;
 463                         }
 464                         if (wait_fd >= 0) {
 465                                 ret = close(wait_fd);
 466                                 if (ret) {
 467                                         PERROR("Error closing channel wait_fd");
 468                                 }
 469                                 *args.channel.wait_fd = -1;
 470                         }
 471                         break;
 472                 }
 473         }
 474
 475 error:
 476         ust_unlock();
 477         return ret;
 478 }
 479
 480 static
 481 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 482 {
 483         int ret;
 484
 485         if (sock_info->socket != -1) {
 486                 ret = ustcomm_close_unix_sock(sock_info->socket);
 487                 if (ret) {
 488                         ERR("Error closing apps socket");
 489                 }
 490                 sock_info->socket = -1;
 491         }
 492         if (sock_info->root_handle != -1) {
 493                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 494                 if (ret) {
 495                         ERR("Error unref root handle");
 496                 }
 497                 sock_info->root_handle = -1;
 498         }
 499         sock_info->constructor_sem_posted = 0;
 500         /*
 501          * wait_shm_mmap is used by listener threads outside of the
 502          * ust lock, so we cannot tear it down ourselves, because we
 503          * cannot join on these threads. Leave this task to the OS
 504          * process exit.
 505          */
 506         if (!exiting && sock_info->wait_shm_mmap) {
 507                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 508                 if (ret) {
 509                         ERR("Error unmapping wait shm");
 510                 }
 511                 sock_info->wait_shm_mmap = NULL;
 512         }
 513 }
 514
 515 /*
 516  * Using fork to set umask in the child process (not multi-thread safe).
 517  * We deal with the shm_open vs ftruncate race (happening when the
 518  * sessiond owns the shm and does not let everybody modify it, to ensure
 519  * safety against shm_unlink) by simply letting the mmap fail and
 520  * retrying after a few seconds.
 521  * For global shm, everybody has rw access to it until the sessiond
 522  * starts.
 523  */
 524 static
 525 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 526 {
 527         int wait_shm_fd, ret;
 528         pid_t pid;
 529
 530         /*
 531          * Try to open read-only.
 532          */
 533         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 534         if (wait_shm_fd >= 0) {
 535                 goto end;
 536         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 537                 /*
 538                  * Real-only open did not work, and it's not because the
 539                  * entry was not present. It's a failure that prohibits
 540                  * using shm.
 541                  */
 542                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 543                 goto end;
 544         }
 545         /*
 546          * If the open failed because the file did not exist, try
 547          * creating it ourself.
 548          */
 549         lttng_ust_nest_count++;
 550         pid = fork();
 551         lttng_ust_nest_count--;
 552         if (pid > 0) {
 553                 int status;
 554
 555                 /*
 556                  * Parent: wait for child to return, in which case the
 557                  * shared memory map will have been created.
 558                  */
 559                 pid = wait(&status);
 560                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 561                         wait_shm_fd = -1;
 562                         goto end;
 563                 }
 564                 /*
 565                  * Try to open read-only again after creation.
 566                  */
 567                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 568                 if (wait_shm_fd < 0) {
 569                         /*
 570                          * Real-only open did not work. It's a failure
 571                          * that prohibits using shm.
 572                          */
 573                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 574                         goto end;
 575                 }
 576                 goto end;
 577         } else if (pid == 0) {
 578                 int create_mode;
 579
 580                 /* Child */
 581                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 582                 if (sock_info->global)
 583                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 584                 /*
 585                  * We're alone in a child process, so we can modify the
 586                  * process-wide umask.
 587                  */
 588                 umask(~create_mode);
 589                 /*
 590                  * Try creating shm (or get rw access).
 591                  * We don't do an exclusive open, because we allow other
 592                  * processes to create+ftruncate it concurrently.
 593                  */
 594                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 595                                 O_RDWR | O_CREAT, create_mode);
 596                 if (wait_shm_fd >= 0) {
 597                         ret = ftruncate(wait_shm_fd, mmap_size);
 598                         if (ret) {
 599                                 PERROR("ftruncate");
 600                                 exit(EXIT_FAILURE);
 601                         }
 602                         exit(EXIT_SUCCESS);
 603                 }
 604                 /*
 605                  * For local shm, we need to have rw access to accept
 606                  * opening it: this means the local sessiond will be
 607                  * able to wake us up. For global shm, we open it even
 608                  * if rw access is not granted, because the root.root
 609                  * sessiond will be able to override all rights and wake
 610                  * us up.
 611                  */
 612                 if (!sock_info->global && errno != EACCES) {
 613                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 614                         exit(EXIT_FAILURE);
 615                 }
 616                 /*
 617                  * The shm exists, but we cannot open it RW. Report
 618                  * success.
 619                  */
 620                 exit(EXIT_SUCCESS);
 621         } else {
 622                 return -1;
 623         }
 624 end:
 625         if (wait_shm_fd >= 0 && !sock_info->global) {
 626                 struct stat statbuf;
 627
 628                 /*
 629                  * Ensure that our user is the owner of the shm file for
 630                  * local shm. If we do not own the file, it means our
 631                  * sessiond will not have access to wake us up (there is
 632                  * probably a rogue process trying to fake our
 633                  * sessiond). Fallback to polling method in this case.
 634                  */
 635                 ret = fstat(wait_shm_fd, &statbuf);
 636                 if (ret) {
 637                         PERROR("fstat");
 638                         goto error_close;
 639                 }
 640                 if (statbuf.st_uid != getuid())
 641                         goto error_close;
 642         }
 643         return wait_shm_fd;
 644
 645 error_close:
 646         ret = close(wait_shm_fd);
 647         if (ret) {
 648                 PERROR("Error closing fd");
 649         }
 650         return -1;
 651 }
 652
 653 static
 654 char *get_map_shm(struct sock_info *sock_info)
 655 {
 656         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 657         int wait_shm_fd, ret;
 658         char *wait_shm_mmap;
 659
 660         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 661         if (wait_shm_fd < 0) {
 662                 goto error;
 663         }
 664         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 665                   MAP_SHARED, wait_shm_fd, 0);
 666         /* close shm fd immediately after taking the mmap reference */
 667         ret = close(wait_shm_fd);
 668         if (ret) {
 669                 PERROR("Error closing fd");
 670         }
 671         if (wait_shm_mmap == MAP_FAILED) {
 672                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 673                 goto error;
 674         }
 675         return wait_shm_mmap;
 676
 677 error:
 678         return NULL;
 679 }
 680
 681 static
 682 void wait_for_sessiond(struct sock_info *sock_info)
 683 {
 684         int ret;
 685
 686         ust_lock();
 687         if (lttng_ust_comm_should_quit) {
 688                 goto quit;
 689         }
 690         if (wait_poll_fallback) {
 691                 goto error;
 692         }
 693         if (!sock_info->wait_shm_mmap) {
 694                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 695                 if (!sock_info->wait_shm_mmap)
 696                         goto error;
 697         }
 698         ust_unlock();
 699
 700         DBG("Waiting for %s apps sessiond", sock_info->name);
 701         /* Wait for futex wakeup */
 702         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 703                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 704                         FUTEX_WAIT, 0, NULL, NULL, 0);
 705                 if (ret < 0) {
 706                         if (errno == EFAULT) {
 707                                 wait_poll_fallback = 1;
 708                                 DBG(
 709 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 710 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 711 "Please upgrade your kernel "
 712 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 713 "mainline). LTTng-UST will use polling mode fallback.");
 714                                 if (ust_debug())
 715                                         PERROR("futex");
 716                         }
 717                 }
 718         }
 719         return;
 720
 721 quit:
 722         ust_unlock();
 723         return;
 724
 725 error:
 726         ust_unlock();
 727         return;
 728 }
 729
 730 /*
 731  * This thread does not allocate any resource, except within
 732  * handle_message, within mutex protection. This mutex protects against
 733  * fork and exit.
 734  * The other moment it allocates resources is at socket connection, which
 735  * is also protected by the mutex.
 736  */
 737 static
 738 void *ust_listener_thread(void *arg)
 739 {
 740         struct sock_info *sock_info = arg;
 741         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 742
 743         /* Restart trying to connect to the session daemon */
 744 restart:
 745         if (prev_connect_failed) {
 746                 /* Wait for sessiond availability with pipe */
 747                 wait_for_sessiond(sock_info);
 748                 if (has_waited) {
 749                         has_waited = 0;
 750                         /*
 751                          * Sleep for 5 seconds before retrying after a
 752                          * sequence of failure / wait / failure. This
 753                          * deals with a killed or broken session daemon.
 754                          */
 755                         sleep(5);
 756                 }
 757                 has_waited = 1;
 758                 prev_connect_failed = 0;
 759         }
 760         ust_lock();
 761
 762         if (lttng_ust_comm_should_quit) {
 763                 ust_unlock();
 764                 goto quit;
 765         }
 766
 767         if (sock_info->socket != -1) {
 768                 ret = ustcomm_close_unix_sock(sock_info->socket);
 769                 if (ret) {
 770                         ERR("Error closing %s apps socket", sock_info->name);
 771                 }
 772                 sock_info->socket = -1;
 773         }
 774
 775         /* Register */
 776         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 777         if (ret < 0) {
 778                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 779                 prev_connect_failed = 1;
 780                 /*
 781                  * If we cannot find the sessiond daemon, don't delay
 782                  * constructor execution.
 783                  */
 784                 ret = handle_register_done(sock_info);
 785                 assert(!ret);
 786                 ust_unlock();
 787                 goto restart;
 788         }
 789
 790         sock_info->socket = sock = ret;
 791
 792         /*
 793          * Create only one root handle per listener thread for the whole
 794          * process lifetime.
 795          */
 796         if (sock_info->root_handle == -1) {
 797                 ret = lttng_abi_create_root_handle();
 798                 if (ret < 0) {
 799                         ERR("Error creating root handle");
 800                         ust_unlock();
 801                         goto quit;
 802                 }
 803                 sock_info->root_handle = ret;
 804         }
 805
 806         ret = register_app_to_sessiond(sock);
 807         if (ret < 0) {
 808                 ERR("Error registering to %s apps socket", sock_info->name);
 809                 prev_connect_failed = 1;
 810                 /*
 811                  * If we cannot register to the sessiond daemon, don't
 812                  * delay constructor execution.
 813                  */
 814                 ret = handle_register_done(sock_info);
 815                 assert(!ret);
 816                 ust_unlock();
 817                 goto restart;
 818         }
 819         ust_unlock();
 820
 821         for (;;) {
 822                 ssize_t len;
 823                 struct ustcomm_ust_msg lum;
 824
 825                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 826                 switch (len) {
 827                 case 0: /* orderly shutdown */
 828                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 829                         ust_lock();
 830                         /*
 831                          * Either sessiond has shutdown or refused us by closing the socket.
 832                          * In either case, we don't want to delay construction execution,
 833                          * and we need to wait before retry.
 834                          */
 835                         prev_connect_failed = 1;
 836                         /*
 837                          * If we cannot register to the sessiond daemon, don't
 838                          * delay constructor execution.
 839                          */
 840                         ret = handle_register_done(sock_info);
 841                         assert(!ret);
 842                         ust_unlock();
 843                         goto end;
 844                 case sizeof(lum):
 845                         DBG("message received\n");
 846                         ret = handle_message(sock_info, sock, &lum);
 847                         if (ret < 0) {
 848                                 ERR("Error handling message for %s socket", sock_info->name);
 849                         }
 850                         continue;
 851                 case -1:
 852                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 853                         if (errno == ECONNRESET) {
 854                                 ERR("%s remote end closed connection\n", sock_info->name);
 855                                 goto end;
 856                         }
 857                         goto end;
 858                 default:
 859                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 860                         continue;
 861                 }
 862
 863         }
 864 end:
 865         goto restart;   /* try to reconnect */
 866 quit:
 867         return NULL;
 868 }
 869
 870 /*
 871  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 872  */
 873 static
 874 int get_timeout(struct timespec *constructor_timeout)
 875 {
 876         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 877         char *str_delay;
 878         int ret;
 879
 880         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 881         if (str_delay) {
 882                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 883         }
 884
 885         switch (constructor_delay_ms) {
 886         case -1:/* fall-through */
 887         case 0:
 888                 return constructor_delay_ms;
 889         default:
 890                 break;
 891         }
 892
 893         /*
 894          * If we are unable to find the current time, don't wait.
 895          */
 896         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 897         if (ret) {
 898                 return -1;
 899         }
 900         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 901         constructor_timeout->tv_nsec +=
 902                 (constructor_delay_ms % 1000UL) * 1000000UL;
 903         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 904                 constructor_timeout->tv_sec++;
 905                 constructor_timeout->tv_nsec -= 1000000000UL;
 906         }
 907         return 1;
 908 }
 909
 910 /*
 911  * sessiond monitoring thread: monitor presence of global and per-user
 912  * sessiond by polling the application common named pipe.
 913  */
 914 /* TODO */
 915
 916 void __attribute__((constructor)) lttng_ust_init(void)
 917 {
 918         struct timespec constructor_timeout;
 919         sigset_t sig_all_blocked, orig_parent_mask;
 920         int timeout_mode;
 921         int ret;
 922
 923         if (uatomic_xchg(&initialized, 1) == 1)
 924                 return;
 925
 926         /*
 927          * Fixup interdependency between TLS fixup mutex (which happens
 928          * to be the dynamic linker mutex) and ust_lock, taken within
 929          * the ust lock.
 930          */
 931         lttng_fixup_event_tls();
 932         lttng_fixup_ringbuffer_tls();
 933         lttng_fixup_vtid_tls();
 934         lttng_fixup_nest_count_tls();
 935
 936         /*
 937          * We want precise control over the order in which we construct
 938          * our sub-libraries vs starting to receive commands from
 939          * sessiond (otherwise leading to errors when trying to create
 940          * sessiond before the init functions are completed).
 941          */
 942         init_usterr();
 943         init_tracepoint();
 944         ltt_ring_buffer_metadata_client_init();
 945         ltt_ring_buffer_client_overwrite_init();
 946         ltt_ring_buffer_client_discard_init();
 947
 948         timeout_mode = get_timeout(&constructor_timeout);
 949
 950         ret = sem_init(&constructor_wait, 0, 0);
 951         assert(!ret);
 952
 953         ret = setup_local_apps();
 954         if (ret) {
 955                 ERR("Error setting up to local apps");
 956         }
 957
 958         /* A new thread created by pthread_create inherits the signal mask
 959          * from the parent. To avoid any signal being received by the
 960          * listener thread, we block all signals temporarily in the parent,
 961          * while we create the listener thread.
 962          */
 963         sigfillset(&sig_all_blocked);
 964         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 965         if (ret) {
 966                 ERR("pthread_sigmask: %s", strerror(ret));
 967         }
 968
 969         ret = pthread_create(&global_apps.ust_listener, NULL,
 970                         ust_listener_thread, &global_apps);
 971         if (ret) {
 972                 ERR("pthread_create global: %s", strerror(ret));
 973         }
 974         if (local_apps.allowed) {
 975                 ret = pthread_create(&local_apps.ust_listener, NULL,
 976                                 ust_listener_thread, &local_apps);
 977                 if (ret) {
 978                         ERR("pthread_create local: %s", strerror(ret));
 979                 }
 980         } else {
 981                 handle_register_done(&local_apps);
 982         }
 983
 984         /* Restore original signal mask in parent */
 985         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
 986         if (ret) {
 987                 ERR("pthread_sigmask: %s", strerror(ret));
 988         }
 989
 990         switch (timeout_mode) {
 991         case 1: /* timeout wait */
 992                 do {
 993                         ret = sem_timedwait(&constructor_wait,
 994                                         &constructor_timeout);
 995                 } while (ret < 0 && errno == EINTR);
 996                 if (ret < 0 && errno == ETIMEDOUT) {
 997                         ERR("Timed out waiting for ltt-sessiond");
 998                 } else {
 999                         assert(!ret);
1000                 }
1001                 break;
1002         case -1:/* wait forever */
1003                 do {
1004                         ret = sem_wait(&constructor_wait);
1005                 } while (ret < 0 && errno == EINTR);
1006                 assert(!ret);
1007                 break;
1008         case 0: /* no timeout */
1009                 break;
1010         }
1011 }
1012
1013 static
1014 void lttng_ust_cleanup(int exiting)
1015 {
1016         cleanup_sock_info(&global_apps, exiting);
1017         if (local_apps.allowed) {
1018                 cleanup_sock_info(&local_apps, exiting);
1019         }
1020         /*
1021          * The teardown in this function all affect data structures
1022          * accessed under the UST lock by the listener thread. This
1023          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1024          * that none of these threads are accessing this data at this
1025          * point.
1026          */
1027         lttng_ust_abi_exit();
1028         lttng_ust_events_exit();
1029         ltt_ring_buffer_client_discard_exit();
1030         ltt_ring_buffer_client_overwrite_exit();
1031         ltt_ring_buffer_metadata_client_exit();
1032         exit_tracepoint();
1033         if (!exiting) {
1034                 /* Reinitialize values for fork */
1035                 sem_count = 2;
1036                 lttng_ust_comm_should_quit = 0;
1037                 initialized = 0;
1038         }
1039 }
1040
1041 void __attribute__((destructor)) lttng_ust_exit(void)
1042 {
1043         int ret;
1044
1045         /*
1046          * Using pthread_cancel here because:
1047          * A) we don't want to hang application teardown.
1048          * B) the thread is not allocating any resource.
1049          */
1050
1051         /*
1052          * Require the communication thread to quit. Synchronize with
1053          * mutexes to ensure it is not in a mutex critical section when
1054          * pthread_cancel is later called.
1055          */
1056         ust_lock();
1057         lttng_ust_comm_should_quit = 1;
1058         ust_unlock();
1059
1060         /* cancel threads */
1061         ret = pthread_cancel(global_apps.ust_listener);
1062         if (ret) {
1063                 ERR("Error cancelling global ust listener thread: %s",
1064                         strerror(ret));
1065         }
1066         if (local_apps.allowed) {
1067                 ret = pthread_cancel(local_apps.ust_listener);
1068                 if (ret) {
1069                         ERR("Error cancelling local ust listener thread: %s",
1070                                 strerror(ret));
1071                 }
1072         }
1073         /*
1074          * Do NOT join threads: use of sys_futex makes it impossible to
1075          * join the threads without using async-cancel, but async-cancel
1076          * is delivered by a signal, which could hit the target thread
1077          * anywhere in its code path, including while the ust_lock() is
1078          * held, causing a deadlock for the other thread. Let the OS
1079          * cleanup the threads if there are stalled in a syscall.
1080          */
1081         lttng_ust_cleanup(1);
1082 }
1083
1084 /*
1085  * We exclude the worker threads across fork and clone (except
1086  * CLONE_VM), because these system calls only keep the forking thread
1087  * running in the child.  Therefore, we don't want to call fork or clone
1088  * in the middle of an tracepoint or ust tracing state modification.
1089  * Holding this mutex protects these structures across fork and clone.
1090  */
1091 void ust_before_fork(sigset_t *save_sigset)
1092 {
1093         /*
1094          * Disable signals. This is to avoid that the child intervenes
1095          * before it is properly setup for tracing. It is safer to
1096          * disable all signals, because then we know we are not breaking
1097          * anything by restoring the original mask.
1098          */
1099         sigset_t all_sigs;
1100         int ret;
1101
1102         if (lttng_ust_nest_count)
1103                 return;
1104         /* Disable signals */
1105         sigfillset(&all_sigs);
1106         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1107         if (ret == -1) {
1108                 PERROR("sigprocmask");
1109         }
1110         ust_lock();
1111         rcu_bp_before_fork();
1112 }
1113
1114 static void ust_after_fork_common(sigset_t *restore_sigset)
1115 {
1116         int ret;
1117
1118         DBG("process %d", getpid());
1119         ust_unlock();
1120         /* Restore signals */
1121         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1122         if (ret == -1) {
1123                 PERROR("sigprocmask");
1124         }
1125 }
1126
1127 void ust_after_fork_parent(sigset_t *restore_sigset)
1128 {
1129         if (lttng_ust_nest_count)
1130                 return;
1131         DBG("process %d", getpid());
1132         rcu_bp_after_fork_parent();
1133         /* Release mutexes and reenable signals */
1134         ust_after_fork_common(restore_sigset);
1135 }
1136
1137 /*
1138  * After fork, in the child, we need to cleanup all the leftover state,
1139  * except the worker thread which already magically disappeared thanks
1140  * to the weird Linux fork semantics. After tyding up, we call
1141  * lttng_ust_init() again to start over as a new PID.
1142  *
1143  * This is meant for forks() that have tracing in the child between the
1144  * fork and following exec call (if there is any).
1145  */
1146 void ust_after_fork_child(sigset_t *restore_sigset)
1147 {
1148         if (lttng_ust_nest_count)
1149                 return;
1150         DBG("process %d", getpid());
1151         /* Release urcu mutexes */
1152         rcu_bp_after_fork_child();
1153         lttng_ust_cleanup(0);
1154         lttng_context_vtid_reset();
1155         /* Release mutexes and reenable signals */
1156         ust_after_fork_common(restore_sigset);
1157         lttng_ust_init();
1158 }