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         default:
 285                 if (ops->cmd)
 286                         ret = ops->cmd(lum->handle, lum->cmd,
 287                                         (unsigned long) &lum->u,
 288                                         &args);
 289                 else
 290                         ret = -ENOSYS;
 291                 break;
 292         }
 293
 294 end:
 295         lur.handle = lum->handle;
 296         lur.cmd = lum->cmd;
 297         lur.ret_val = ret;
 298         if (ret >= 0) {
 299                 lur.ret_code = USTCOMM_OK;
 300         } else {
 301                 //lur.ret_code = USTCOMM_SESSION_FAIL;
 302                 lur.ret_code = ret;
 303         }
 304         if (ret >= 0) {
 305                 switch (lum->cmd) {
 306                 case LTTNG_UST_STREAM:
 307                         /*
 308                          * Special-case reply to send stream info.
 309                          * Use lum.u output.
 310                          */
 311                         lur.u.stream.memory_map_size = *args.stream.memory_map_size;
 312                         shm_fd = *args.stream.shm_fd;
 313                         wait_fd = *args.stream.wait_fd;
 314                         break;
 315                 case LTTNG_UST_METADATA:
 316                 case LTTNG_UST_CHANNEL:
 317                         lur.u.channel.memory_map_size = *args.channel.memory_map_size;
 318                         shm_fd = *args.channel.shm_fd;
 319                         wait_fd = *args.channel.wait_fd;
 320                         break;
 321                 case LTTNG_UST_TRACER_VERSION:
 322                         lur.u.version = lum->u.version;
 323                         break;
 324                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 325                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 326                         break;
 327                 }
 328         }
 329         ret = send_reply(sock, &lur);
 330         if (ret < 0) {
 331                 perror("error sending reply");
 332                 goto error;
 333         }
 334
 335         if ((lum->cmd == LTTNG_UST_STREAM
 336              || lum->cmd == LTTNG_UST_CHANNEL
 337              || lum->cmd == LTTNG_UST_METADATA)
 338                         && lur.ret_code == USTCOMM_OK) {
 339                 int sendret = 0;
 340
 341                 /* we also need to send the file descriptors. */
 342                 ret = ustcomm_send_fds_unix_sock(sock,
 343                         &shm_fd, &shm_fd,
 344                         1, sizeof(int));
 345                 if (ret < 0) {
 346                         perror("send shm_fd");
 347                         sendret = ret;
 348                 }
 349                 /*
 350                  * The sessiond expects 2 file descriptors, even upon
 351                  * error.
 352                  */
 353                 ret = ustcomm_send_fds_unix_sock(sock,
 354                         &wait_fd, &wait_fd,
 355                         1, sizeof(int));
 356                 if (ret < 0) {
 357                         perror("send wait_fd");
 358                         goto error;
 359                 }
 360                 if (sendret) {
 361                         ret = sendret;
 362                         goto error;
 363                 }
 364         }
 365         /*
 366          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
 367          * after the reply.
 368          */
 369         if (lur.ret_code == USTCOMM_OK) {
 370                 switch (lum->cmd) {
 371                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
 372                         len = ustcomm_send_unix_sock(sock,
 373                                 &args.field_list.entry,
 374                                 sizeof(args.field_list.entry));
 375                         if (len != sizeof(args.field_list.entry)) {
 376                                 ret = -1;
 377                                 goto error;
 378                         }
 379                 }
 380         }
 381         /*
 382          * We still have the memory map reference, and the fds have been
 383          * sent to the sessiond. We can therefore close those fds. Note
 384          * that we keep the write side of the wait_fd open, but close
 385          * the read side.
 386          */
 387         if (lur.ret_code == USTCOMM_OK) {
 388                 switch (lum->cmd) {
 389                 case LTTNG_UST_STREAM:
 390                         if (shm_fd >= 0) {
 391                                 ret = close(shm_fd);
 392                                 if (ret) {
 393                                         PERROR("Error closing stream shm_fd");
 394                                 }
 395                                 *args.stream.shm_fd = -1;
 396                         }
 397                         if (wait_fd >= 0) {
 398                                 ret = close(wait_fd);
 399                                 if (ret) {
 400                                         PERROR("Error closing stream wait_fd");
 401                                 }
 402                                 *args.stream.wait_fd = -1;
 403                         }
 404                         break;
 405                 case LTTNG_UST_METADATA:
 406                 case LTTNG_UST_CHANNEL:
 407                         if (shm_fd >= 0) {
 408                                 ret = close(shm_fd);
 409                                 if (ret) {
 410                                         PERROR("Error closing channel shm_fd");
 411                                 }
 412                                 *args.channel.shm_fd = -1;
 413                         }
 414                         if (wait_fd >= 0) {
 415                                 ret = close(wait_fd);
 416                                 if (ret) {
 417                                         PERROR("Error closing channel wait_fd");
 418                                 }
 419                                 *args.channel.wait_fd = -1;
 420                         }
 421                         break;
 422                 }
 423         }
 424
 425 error:
 426         ust_unlock();
 427         return ret;
 428 }
 429
 430 static
 431 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 432 {
 433         int ret;
 434
 435         if (sock_info->socket != -1) {
 436                 ret = ustcomm_close_unix_sock(sock_info->socket);
 437                 if (ret) {
 438                         ERR("Error closing apps socket");
 439                 }
 440                 sock_info->socket = -1;
 441         }
 442         if (sock_info->root_handle != -1) {
 443                 ret = lttng_ust_objd_unref(sock_info->root_handle);
 444                 if (ret) {
 445                         ERR("Error unref root handle");
 446                 }
 447                 sock_info->root_handle = -1;
 448         }
 449         sock_info->constructor_sem_posted = 0;
 450         /*
 451          * wait_shm_mmap is used by listener threads outside of the
 452          * ust lock, so we cannot tear it down ourselves, because we
 453          * cannot join on these threads. Leave this task to the OS
 454          * process exit.
 455          */
 456         if (!exiting && sock_info->wait_shm_mmap) {
 457                 ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
 458                 if (ret) {
 459                         ERR("Error unmapping wait shm");
 460                 }
 461                 sock_info->wait_shm_mmap = NULL;
 462         }
 463 }
 464
 465 /*
 466  * Using fork to set umask in the child process (not multi-thread safe).
 467  * We deal with the shm_open vs ftruncate race (happening when the
 468  * sessiond owns the shm and does not let everybody modify it, to ensure
 469  * safety against shm_unlink) by simply letting the mmap fail and
 470  * retrying after a few seconds.
 471  * For global shm, everybody has rw access to it until the sessiond
 472  * starts.
 473  */
 474 static
 475 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
 476 {
 477         int wait_shm_fd, ret;
 478         pid_t pid;
 479
 480         /*
 481          * Try to open read-only.
 482          */
 483         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 484         if (wait_shm_fd >= 0) {
 485                 goto end;
 486         } else if (wait_shm_fd < 0 && errno != ENOENT) {
 487                 /*
 488                  * Real-only open did not work, and it's not because the
 489                  * entry was not present. It's a failure that prohibits
 490                  * using shm.
 491                  */
 492                 ERR("Error opening shm %s", sock_info->wait_shm_path);
 493                 goto end;
 494         }
 495         /*
 496          * If the open failed because the file did not exist, try
 497          * creating it ourself.
 498          */
 499         lttng_ust_nest_count++;
 500         pid = fork();
 501         lttng_ust_nest_count--;
 502         if (pid > 0) {
 503                 int status;
 504
 505                 /*
 506                  * Parent: wait for child to return, in which case the
 507                  * shared memory map will have been created.
 508                  */
 509                 pid = wait(&status);
 510                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
 511                         wait_shm_fd = -1;
 512                         goto end;
 513                 }
 514                 /*
 515                  * Try to open read-only again after creation.
 516                  */
 517                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
 518                 if (wait_shm_fd < 0) {
 519                         /*
 520                          * Real-only open did not work. It's a failure
 521                          * that prohibits using shm.
 522                          */
 523                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 524                         goto end;
 525                 }
 526                 goto end;
 527         } else if (pid == 0) {
 528                 int create_mode;
 529
 530                 /* Child */
 531                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
 532                 if (sock_info->global)
 533                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
 534                 /*
 535                  * We're alone in a child process, so we can modify the
 536                  * process-wide umask.
 537                  */
 538                 umask(~create_mode);
 539                 /*
 540                  * Try creating shm (or get rw access).
 541                  * We don't do an exclusive open, because we allow other
 542                  * processes to create+ftruncate it concurrently.
 543                  */
 544                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
 545                                 O_RDWR | O_CREAT, create_mode);
 546                 if (wait_shm_fd >= 0) {
 547                         ret = ftruncate(wait_shm_fd, mmap_size);
 548                         if (ret) {
 549                                 PERROR("ftruncate");
 550                                 exit(EXIT_FAILURE);
 551                         }
 552                         exit(EXIT_SUCCESS);
 553                 }
 554                 /*
 555                  * For local shm, we need to have rw access to accept
 556                  * opening it: this means the local sessiond will be
 557                  * able to wake us up. For global shm, we open it even
 558                  * if rw access is not granted, because the root.root
 559                  * sessiond will be able to override all rights and wake
 560                  * us up.
 561                  */
 562                 if (!sock_info->global && errno != EACCES) {
 563                         ERR("Error opening shm %s", sock_info->wait_shm_path);
 564                         exit(EXIT_FAILURE);
 565                 }
 566                 /*
 567                  * The shm exists, but we cannot open it RW. Report
 568                  * success.
 569                  */
 570                 exit(EXIT_SUCCESS);
 571         } else {
 572                 return -1;
 573         }
 574 end:
 575         if (wait_shm_fd >= 0 && !sock_info->global) {
 576                 struct stat statbuf;
 577
 578                 /*
 579                  * Ensure that our user is the owner of the shm file for
 580                  * local shm. If we do not own the file, it means our
 581                  * sessiond will not have access to wake us up (there is
 582                  * probably a rogue process trying to fake our
 583                  * sessiond). Fallback to polling method in this case.
 584                  */
 585                 ret = fstat(wait_shm_fd, &statbuf);
 586                 if (ret) {
 587                         PERROR("fstat");
 588                         goto error_close;
 589                 }
 590                 if (statbuf.st_uid != getuid())
 591                         goto error_close;
 592         }
 593         return wait_shm_fd;
 594
 595 error_close:
 596         ret = close(wait_shm_fd);
 597         if (ret) {
 598                 PERROR("Error closing fd");
 599         }
 600         return -1;
 601 }
 602
 603 static
 604 char *get_map_shm(struct sock_info *sock_info)
 605 {
 606         size_t mmap_size = sysconf(_SC_PAGE_SIZE);
 607         int wait_shm_fd, ret;
 608         char *wait_shm_mmap;
 609
 610         wait_shm_fd = get_wait_shm(sock_info, mmap_size);
 611         if (wait_shm_fd < 0) {
 612                 goto error;
 613         }
 614         wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
 615                   MAP_SHARED, wait_shm_fd, 0);
 616         /* close shm fd immediately after taking the mmap reference */
 617         ret = close(wait_shm_fd);
 618         if (ret) {
 619                 PERROR("Error closing fd");
 620         }
 621         if (wait_shm_mmap == MAP_FAILED) {
 622                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
 623                 goto error;
 624         }
 625         return wait_shm_mmap;
 626
 627 error:
 628         return NULL;
 629 }
 630
 631 static
 632 void wait_for_sessiond(struct sock_info *sock_info)
 633 {
 634         int ret;
 635
 636         ust_lock();
 637         if (lttng_ust_comm_should_quit) {
 638                 goto quit;
 639         }
 640         if (wait_poll_fallback) {
 641                 goto error;
 642         }
 643         if (!sock_info->wait_shm_mmap) {
 644                 sock_info->wait_shm_mmap = get_map_shm(sock_info);
 645                 if (!sock_info->wait_shm_mmap)
 646                         goto error;
 647         }
 648         ust_unlock();
 649
 650         DBG("Waiting for %s apps sessiond", sock_info->name);
 651         /* Wait for futex wakeup */
 652         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
 653                 ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
 654                         FUTEX_WAIT, 0, NULL, NULL, 0);
 655                 if (ret < 0) {
 656                         if (errno == EFAULT) {
 657                                 wait_poll_fallback = 1;
 658                                 DBG(
 659 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 660 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 661 "Please upgrade your kernel "
 662 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 663 "mainline). LTTng-UST will use polling mode fallback.");
 664                                 if (ust_debug())
 665                                         PERROR("futex");
 666                         }
 667                 }
 668         }
 669         return;
 670
 671 quit:
 672         ust_unlock();
 673         return;
 674
 675 error:
 676         ust_unlock();
 677         return;
 678 }
 679
 680 /*
 681  * This thread does not allocate any resource, except within
 682  * handle_message, within mutex protection. This mutex protects against
 683  * fork and exit.
 684  * The other moment it allocates resources is at socket connexion, which
 685  * is also protected by the mutex.
 686  */
 687 static
 688 void *ust_listener_thread(void *arg)
 689 {
 690         struct sock_info *sock_info = arg;
 691         int sock, ret, prev_connect_failed = 0, has_waited = 0;
 692
 693         /* Restart trying to connect to the session daemon */
 694 restart:
 695         if (prev_connect_failed) {
 696                 /* Wait for sessiond availability with pipe */
 697                 wait_for_sessiond(sock_info);
 698                 if (has_waited) {
 699                         has_waited = 0;
 700                         /*
 701                          * Sleep for 5 seconds before retrying after a
 702                          * sequence of failure / wait / failure. This
 703                          * deals with a killed or broken session daemon.
 704                          */
 705                         sleep(5);
 706                 }
 707                 has_waited = 1;
 708                 prev_connect_failed = 0;
 709         }
 710         ust_lock();
 711
 712         if (lttng_ust_comm_should_quit) {
 713                 ust_unlock();
 714                 goto quit;
 715         }
 716
 717         if (sock_info->socket != -1) {
 718                 ret = ustcomm_close_unix_sock(sock_info->socket);
 719                 if (ret) {
 720                         ERR("Error closing %s apps socket", sock_info->name);
 721                 }
 722                 sock_info->socket = -1;
 723         }
 724
 725         /* Register */
 726         ret = ustcomm_connect_unix_sock(sock_info->sock_path);
 727         if (ret < 0) {
 728                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
 729                 prev_connect_failed = 1;
 730                 /*
 731                  * If we cannot find the sessiond daemon, don't delay
 732                  * constructor execution.
 733                  */
 734                 ret = handle_register_done(sock_info);
 735                 assert(!ret);
 736                 ust_unlock();
 737                 goto restart;
 738         }
 739
 740         sock_info->socket = sock = ret;
 741
 742         /*
 743          * Create only one root handle per listener thread for the whole
 744          * process lifetime.
 745          */
 746         if (sock_info->root_handle == -1) {
 747                 ret = lttng_abi_create_root_handle();
 748                 if (ret < 0) {
 749                         ERR("Error creating root handle");
 750                         ust_unlock();
 751                         goto quit;
 752                 }
 753                 sock_info->root_handle = ret;
 754         }
 755
 756         ret = register_app_to_sessiond(sock);
 757         if (ret < 0) {
 758                 ERR("Error registering to %s apps socket", sock_info->name);
 759                 prev_connect_failed = 1;
 760                 /*
 761                  * If we cannot register to the sessiond daemon, don't
 762                  * delay constructor execution.
 763                  */
 764                 ret = handle_register_done(sock_info);
 765                 assert(!ret);
 766                 ust_unlock();
 767                 goto restart;
 768         }
 769         ust_unlock();
 770
 771         for (;;) {
 772                 ssize_t len;
 773                 struct ustcomm_ust_msg lum;
 774
 775                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 776                 switch (len) {
 777                 case 0: /* orderly shutdown */
 778                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 779                         ust_lock();
 780                         /*
 781                          * Either sessiond has shutdown or refused us by closing the socket.
 782                          * In either case, we don't want to delay construction execution,
 783                          * and we need to wait before retry.
 784                          */
 785                         prev_connect_failed = 1;
 786                         /*
 787                          * If we cannot register to the sessiond daemon, don't
 788                          * delay constructor execution.
 789                          */
 790                         ret = handle_register_done(sock_info);
 791                         assert(!ret);
 792                         ust_unlock();
 793                         goto end;
 794                 case sizeof(lum):
 795                         DBG("message received\n");
 796                         ret = handle_message(sock_info, sock, &lum);
 797                         if (ret < 0) {
 798                                 ERR("Error handling message for %s socket", sock_info->name);
 799                         }
 800                         continue;
 801                 case -1:
 802                         DBG("Receive failed from lttng-sessiond with errno %d", errno);
 803                         if (errno == ECONNRESET) {
 804                                 ERR("%s remote end closed connection\n", sock_info->name);
 805                                 goto end;
 806                         }
 807                         goto end;
 808                 default:
 809                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 810                         continue;
 811                 }
 812
 813         }
 814 end:
 815         goto restart;   /* try to reconnect */
 816 quit:
 817         return NULL;
 818 }
 819
 820 /*
 821  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 822  */
 823 static
 824 int get_timeout(struct timespec *constructor_timeout)
 825 {
 826         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 827         char *str_delay;
 828         int ret;
 829
 830         str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
 831         if (str_delay) {
 832                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 833         }
 834
 835         switch (constructor_delay_ms) {
 836         case -1:/* fall-through */
 837         case 0:
 838                 return constructor_delay_ms;
 839         default:
 840                 break;
 841         }
 842
 843         /*
 844          * If we are unable to find the current time, don't wait.
 845          */
 846         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 847         if (ret) {
 848                 return -1;
 849         }
 850         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 851         constructor_timeout->tv_nsec +=
 852                 (constructor_delay_ms % 1000UL) * 1000000UL;
 853         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 854                 constructor_timeout->tv_sec++;
 855                 constructor_timeout->tv_nsec -= 1000000000UL;
 856         }
 857         return 1;
 858 }
 859
 860 /*
 861  * sessiond monitoring thread: monitor presence of global and per-user
 862  * sessiond by polling the application common named pipe.
 863  */
 864 /* TODO */
 865
 866 void __attribute__((constructor)) lttng_ust_init(void)
 867 {
 868         struct timespec constructor_timeout;
 869         sigset_t sig_all_blocked, orig_parent_mask;
 870         int timeout_mode;
 871         int ret;
 872
 873         if (uatomic_xchg(&initialized, 1) == 1)
 874                 return;
 875
 876         /*
 877          * Fixup interdependency between TLS fixup mutex (which happens
 878          * to be the dynamic linker mutex) and ust_lock, taken within
 879          * the ust lock.
 880          */
 881         lttng_fixup_event_tls();
 882         lttng_fixup_ringbuffer_tls();
 883         lttng_fixup_vtid_tls();
 884         lttng_fixup_nest_count_tls();
 885
 886         /*
 887          * We want precise control over the order in which we construct
 888          * our sub-libraries vs starting to receive commands from
 889          * sessiond (otherwise leading to errors when trying to create
 890          * sessiond before the init functions are completed).
 891          */
 892         init_usterr();
 893         init_tracepoint();
 894         ltt_ring_buffer_metadata_client_init();
 895         ltt_ring_buffer_client_overwrite_init();
 896         ltt_ring_buffer_client_discard_init();
 897
 898         timeout_mode = get_timeout(&constructor_timeout);
 899
 900         ret = sem_init(&constructor_wait, 0, 0);
 901         assert(!ret);
 902
 903         ret = setup_local_apps();
 904         if (ret) {
 905                 ERR("Error setting up to local apps");
 906         }
 907
 908         /* A new thread created by pthread_create inherits the signal mask
 909          * from the parent. To avoid any signal being received by the
 910          * listener thread, we block all signals temporarily in the parent,
 911          * while we create the listener thread.
 912          */
 913         sigfillset(&sig_all_blocked);
 914         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
 915         if (ret) {
 916                 ERR("pthread_sigmask: %s", strerror(ret));
 917         }
 918
 919         ret = pthread_create(&global_apps.ust_listener, NULL,
 920                         ust_listener_thread, &global_apps);
 921         if (ret) {
 922                 ERR("pthread_create global: %s", strerror(ret));
 923         }
 924         if (local_apps.allowed) {
 925                 ret = pthread_create(&local_apps.ust_listener, NULL,
 926                                 ust_listener_thread, &local_apps);
 927                 if (ret) {
 928                         ERR("pthread_create local: %s", strerror(ret));
 929                 }
 930         } else {
 931                 handle_register_done(&local_apps);
 932         }
 933
 934         /* Restore original signal mask in parent */
 935         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
 936         if (ret) {
 937                 ERR("pthread_sigmask: %s", strerror(ret));
 938         }
 939
 940         switch (timeout_mode) {
 941         case 1: /* timeout wait */
 942                 do {
 943                         ret = sem_timedwait(&constructor_wait,
 944                                         &constructor_timeout);
 945                 } while (ret < 0 && errno == EINTR);
 946                 if (ret < 0 && errno == ETIMEDOUT) {
 947                         ERR("Timed out waiting for ltt-sessiond");
 948                 } else {
 949                         assert(!ret);
 950                 }
 951                 break;
 952         case -1:/* wait forever */
 953                 do {
 954                         ret = sem_wait(&constructor_wait);
 955                 } while (ret < 0 && errno == EINTR);
 956                 assert(!ret);
 957                 break;
 958         case 0: /* no timeout */
 959                 break;
 960         }
 961 }
 962
 963 static
 964 void lttng_ust_cleanup(int exiting)
 965 {
 966         cleanup_sock_info(&global_apps, exiting);
 967         if (local_apps.allowed) {
 968                 cleanup_sock_info(&local_apps, exiting);
 969         }
 970         /*
 971          * The teardown in this function all affect data structures
 972          * accessed under the UST lock by the listener thread. This
 973          * lock, along with the lttng_ust_comm_should_quit flag, ensure
 974          * that none of these threads are accessing this data at this
 975          * point.
 976          */
 977         lttng_ust_abi_exit();
 978         lttng_ust_events_exit();
 979         ltt_ring_buffer_client_discard_exit();
 980         ltt_ring_buffer_client_overwrite_exit();
 981         ltt_ring_buffer_metadata_client_exit();
 982         exit_tracepoint();
 983         if (!exiting) {
 984                 /* Reinitialize values for fork */
 985                 sem_count = 2;
 986                 lttng_ust_comm_should_quit = 0;
 987                 initialized = 0;
 988         }
 989 }
 990
 991 void __attribute__((destructor)) lttng_ust_exit(void)
 992 {
 993         int ret;
 994
 995         /*
 996          * Using pthread_cancel here because:
 997          * A) we don't want to hang application teardown.
 998          * B) the thread is not allocating any resource.
 999          */
1000
1001         /*
1002          * Require the communication thread to quit. Synchronize with
1003          * mutexes to ensure it is not in a mutex critical section when
1004          * pthread_cancel is later called.
1005          */
1006         ust_lock();
1007         lttng_ust_comm_should_quit = 1;
1008         ust_unlock();
1009
1010         /* cancel threads */
1011         ret = pthread_cancel(global_apps.ust_listener);
1012         if (ret) {
1013                 ERR("Error cancelling global ust listener thread: %s",
1014                         strerror(ret));
1015         }
1016         if (local_apps.allowed) {
1017                 ret = pthread_cancel(local_apps.ust_listener);
1018                 if (ret) {
1019                         ERR("Error cancelling local ust listener thread: %s",
1020                                 strerror(ret));
1021                 }
1022         }
1023         /*
1024          * Do NOT join threads: use of sys_futex makes it impossible to
1025          * join the threads without using async-cancel, but async-cancel
1026          * is delivered by a signal, which could hit the target thread
1027          * anywhere in its code path, including while the ust_lock() is
1028          * held, causing a deadlock for the other thread. Let the OS
1029          * cleanup the threads if there are stalled in a syscall.
1030          */
1031         lttng_ust_cleanup(1);
1032 }
1033
1034 /*
1035  * We exclude the worker threads across fork and clone (except
1036  * CLONE_VM), because these system calls only keep the forking thread
1037  * running in the child.  Therefore, we don't want to call fork or clone
1038  * in the middle of an tracepoint or ust tracing state modification.
1039  * Holding this mutex protects these structures across fork and clone.
1040  */
1041 void ust_before_fork(sigset_t *save_sigset)
1042 {
1043         /*
1044          * Disable signals. This is to avoid that the child intervenes
1045          * before it is properly setup for tracing. It is safer to
1046          * disable all signals, because then we know we are not breaking
1047          * anything by restoring the original mask.
1048          */
1049         sigset_t all_sigs;
1050         int ret;
1051
1052         if (lttng_ust_nest_count)
1053                 return;
1054         /* Disable signals */
1055         sigfillset(&all_sigs);
1056         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1057         if (ret == -1) {
1058                 PERROR("sigprocmask");
1059         }
1060         ust_lock();
1061         rcu_bp_before_fork();
1062 }
1063
1064 static void ust_after_fork_common(sigset_t *restore_sigset)
1065 {
1066         int ret;
1067
1068         DBG("process %d", getpid());
1069         ust_unlock();
1070         /* Restore signals */
1071         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1072         if (ret == -1) {
1073                 PERROR("sigprocmask");
1074         }
1075 }
1076
1077 void ust_after_fork_parent(sigset_t *restore_sigset)
1078 {
1079         if (lttng_ust_nest_count)
1080                 return;
1081         DBG("process %d", getpid());
1082         rcu_bp_after_fork_parent();
1083         /* Release mutexes and reenable signals */
1084         ust_after_fork_common(restore_sigset);
1085 }
1086
1087 /*
1088  * After fork, in the child, we need to cleanup all the leftover state,
1089  * except the worker thread which already magically disappeared thanks
1090  * to the weird Linux fork semantics. After tyding up, we call
1091  * lttng_ust_init() again to start over as a new PID.
1092  *
1093  * This is meant for forks() that have tracing in the child between the
1094  * fork and following exec call (if there is any).
1095  */
1096 void ust_after_fork_child(sigset_t *restore_sigset)
1097 {
1098         if (lttng_ust_nest_count)
1099                 return;
1100         DBG("process %d", getpid());
1101         /* Release urcu mutexes */
1102         rcu_bp_after_fork_child();
1103         lttng_ust_cleanup(0);
1104         lttng_context_vtid_reset();
1105         /* Release mutexes and reenable signals */
1106         ust_after_fork_common(restore_sigset);
1107         lttng_ust_init();
1108 }