libust/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 #include <sys/types.h>
  23 #include <sys/socket.h>
  24 #include <sys/prctl.h>
  25 #include <unistd.h>
  26 #include <errno.h>
  27 #include <pthread.h>
  28 #include <semaphore.h>
  29 #include <time.h>
  30 #include <assert.h>
  31 #include <signal.h>
  32 #include <urcu/uatomic.h>
  33
  34 #include <lttng-ust-comm.h>
  35 #include <ust/usterr-signal-safe.h>
  36 #include <ust/lttng-ust-abi.h>
  37 #include <ust/tracepoint.h>
  38 #include <ust/tracepoint-internal.h>
  39 #include <ust/ust.h>
  40
  41 /*
  42  * Has lttng ust comm constructor been called ?
  43  */
  44 static int initialized;
  45
  46 /*
  47  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  48  * Held when handling a command, also held by fork() to deal with
  49  * removal of threads, and by exit path.
  50  */
  51
  52 /* Should the ust comm thread quit ? */
  53 static int lttng_ust_comm_should_quit;
  54
  55 /*
  56  * Wait for either of these before continuing to the main
  57  * program:
  58  * - the register_done message from sessiond daemon
  59  *   (will let the sessiond daemon enable sessions before main
  60  *   starts.)
  61  * - sessiond daemon is not reachable.
  62  * - timeout (ensuring applications are resilient to session
  63  *   daemon problems).
  64  */
  65 static sem_t constructor_wait;
  66 /*
  67  * Doing this for both the global and local sessiond.
  68  */
  69 static int sem_count = { 2 };
  70
  71 /*
  72  * Info about socket and associated listener thread.
  73  */
  74 struct sock_info {
  75         const char *name;
  76         char sock_path[PATH_MAX];
  77         int socket;
  78         pthread_t ust_listener; /* listener thread */
  79         int root_handle;
  80         int constructor_sem_posted;
  81         int allowed;
  82 };
  83
  84 /* Socket from app (connect) to session daemon (listen) for communication */
  85 struct sock_info global_apps = {
  86         .name = "global",
  87         .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
  88         .socket = -1,
  89         .root_handle = -1,
  90         .allowed = 1,
  91 };
  92
  93 /* TODO: allow global_apps_sock_path override */
  94
  95 struct sock_info local_apps = {
  96         .name = "local",
  97         .socket = -1,
  98         .root_handle = -1,
  99         .allowed = 0,   /* Check setuid bit first */
 100 };
 101
 102 extern void ltt_ring_buffer_client_overwrite_init(void);
 103 extern void ltt_ring_buffer_client_discard_init(void);
 104 extern void ltt_ring_buffer_metadata_client_init(void);
 105 extern void ltt_ring_buffer_client_overwrite_exit(void);
 106 extern void ltt_ring_buffer_client_discard_exit(void);
 107 extern void ltt_ring_buffer_metadata_client_exit(void);
 108
 109 static
 110 int setup_local_apps(void)
 111 {
 112         const char *home_dir;
 113
 114         /*
 115          * Disallow per-user tracing for setuid binaries.
 116          */
 117         if (getuid() != geteuid()) {
 118                 local_apps.allowed = 0;
 119                 return 0;
 120         } else {
 121                 local_apps.allowed = 1;
 122         }
 123         home_dir = (const char *) getenv("HOME");
 124         if (!home_dir)
 125                 return -ENOENT;
 126         snprintf(local_apps.sock_path, PATH_MAX,
 127                  DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
 128         return 0;
 129 }
 130
 131 static
 132 int register_app_to_sessiond(int socket)
 133 {
 134         ssize_t ret;
 135         int prctl_ret;
 136         struct {
 137                 uint32_t major;
 138                 uint32_t minor;
 139                 pid_t pid;
 140                 pid_t ppid;
 141                 uid_t uid;
 142                 gid_t gid;
 143                 char name[16];  /* process name */
 144         } reg_msg;
 145
 146         reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
 147         reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
 148         reg_msg.pid = getpid();
 149         reg_msg.ppid = getppid();
 150         reg_msg.uid = getuid();
 151         reg_msg.gid = getgid();
 152         prctl_ret = prctl(PR_GET_NAME, (unsigned long) reg_msg.name, 0, 0, 0);
 153         if (prctl_ret) {
 154                 ERR("Error executing prctl");
 155                 return -errno;
 156         }
 157
 158         ret = lttcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
 159         if (ret >= 0 && ret != sizeof(reg_msg))
 160                 return -EIO;
 161         return ret;
 162 }
 163
 164 static
 165 int send_reply(int sock, struct lttcomm_ust_reply *lur)
 166 {
 167         ssize_t len;
 168
 169         len = lttcomm_send_unix_sock(sock, lur, sizeof(*lur));
 170         switch (len) {
 171         case sizeof(*lur):
 172                 DBG("message successfully sent");
 173                 return 0;
 174         case -1:
 175                 if (errno == ECONNRESET) {
 176                         printf("remote end closed connection\n");
 177                         return 0;
 178                 }
 179                 return -1;
 180         default:
 181                 printf("incorrect message size: %zd\n", len);
 182                 return -1;
 183         }
 184 }
 185
 186 static
 187 int handle_register_done(struct sock_info *sock_info)
 188 {
 189         int ret;
 190
 191         if (sock_info->constructor_sem_posted)
 192                 return 0;
 193         sock_info->constructor_sem_posted = 1;
 194         ret = uatomic_add_return(&sem_count, -1);
 195         if (ret == 0) {
 196                 ret = sem_post(&constructor_wait);
 197                 assert(!ret);
 198         }
 199         return 0;
 200 }
 201
 202 static
 203 int handle_message(struct sock_info *sock_info,
 204                 int sock, struct lttcomm_ust_msg *lum)
 205 {
 206         int ret = 0;
 207         const struct objd_ops *ops;
 208         struct lttcomm_ust_reply lur;
 209
 210         ust_lock();
 211
 212         memset(&lur, 0, sizeof(lur));
 213
 214         if (lttng_ust_comm_should_quit) {
 215                 ret = -EPERM;
 216                 goto end;
 217         }
 218
 219         ops = objd_ops(lum->handle);
 220         if (!ops) {
 221                 ret = -ENOENT;
 222                 goto end;
 223         }
 224
 225         switch (lum->cmd) {
 226         case LTTNG_UST_REGISTER_DONE:
 227                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 228                         ret = handle_register_done(sock_info);
 229                 else
 230                         ret = -EINVAL;
 231                 break;
 232         case LTTNG_UST_RELEASE:
 233                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 234                         ret = -EPERM;
 235                 else
 236                         ret = objd_unref(lum->handle);
 237                 break;
 238         default:
 239                 if (ops->cmd)
 240                         ret = ops->cmd(lum->handle, lum->cmd,
 241                                         (unsigned long) &lum->u);
 242                 else
 243                         ret = -ENOSYS;
 244                 break;
 245         }
 246
 247 end:
 248         lur.handle = lum->handle;
 249         lur.cmd = lum->cmd;
 250         lur.ret_val = ret;
 251         if (ret >= 0) {
 252                 lur.ret_code = LTTCOMM_OK;
 253         } else {
 254                 lur.ret_code = LTTCOMM_SESSION_FAIL;
 255         }
 256         ret = send_reply(sock, &lur);
 257
 258         ust_unlock();
 259         return ret;
 260 }
 261
 262 static
 263 void cleanup_sock_info(struct sock_info *sock_info)
 264 {
 265         int ret;
 266
 267         if (sock_info->socket != -1) {
 268                 ret = close(sock_info->socket);
 269                 if (ret) {
 270                         ERR("Error closing local apps socket");
 271                 }
 272                 sock_info->socket = -1;
 273         }
 274         if (sock_info->root_handle != -1) {
 275                 ret = objd_unref(sock_info->root_handle);
 276                 if (ret) {
 277                         ERR("Error unref root handle");
 278                 }
 279                 sock_info->root_handle = -1;
 280         }
 281 }
 282
 283 /*
 284  * This thread does not allocate any resource, except within
 285  * handle_message, within mutex protection. This mutex protects against
 286  * fork and exit.
 287  * The other moment it allocates resources is at socket connexion, which
 288  * is also protected by the mutex.
 289  */
 290 static
 291 void *ust_listener_thread(void *arg)
 292 {
 293         struct sock_info *sock_info = arg;
 294         int sock, ret;
 295
 296         /* Restart trying to connect to the session daemon */
 297 restart:
 298         ust_lock();
 299
 300         if (lttng_ust_comm_should_quit) {
 301                 ust_unlock();
 302                 goto quit;
 303         }
 304
 305         if (sock_info->socket != -1) {
 306                 ret = close(sock_info->socket);
 307                 if (ret) {
 308                         ERR("Error closing %s apps socket", sock_info->name);
 309                 }
 310                 sock_info->socket = -1;
 311         }
 312
 313         /* Check for sessiond availability with pipe TODO */
 314
 315         /* Register */
 316         ret = lttcomm_connect_unix_sock(sock_info->sock_path);
 317         if (ret < 0) {
 318                 ERR("Error connecting to %s apps socket", sock_info->name);
 319                 /*
 320                  * If we cannot find the sessiond daemon, don't delay
 321                  * constructor execution.
 322                  */
 323                 ret = handle_register_done(sock_info);
 324                 assert(!ret);
 325                 ust_unlock();
 326                 sleep(5);
 327                 goto restart;
 328         }
 329
 330         sock_info->socket = sock = ret;
 331
 332         /*
 333          * Create only one root handle per listener thread for the whole
 334          * process lifetime.
 335          */
 336         if (sock_info->root_handle == -1) {
 337                 ret = lttng_abi_create_root_handle();
 338                 if (ret) {
 339                         ERR("Error creating root handle");
 340                         ust_unlock();
 341                         goto quit;
 342                 }
 343                 sock_info->root_handle = ret;
 344         }
 345
 346         ret = register_app_to_sessiond(sock);
 347         if (ret < 0) {
 348                 ERR("Error registering to %s apps socket", sock_info->name);
 349                 /*
 350                  * If we cannot register to the sessiond daemon, don't
 351                  * delay constructor execution.
 352                  */
 353                 ret = handle_register_done(sock_info);
 354                 assert(!ret);
 355                 ust_unlock();
 356                 sleep(5);
 357                 goto restart;
 358         }
 359         ust_unlock();
 360
 361         for (;;) {
 362                 ssize_t len;
 363                 struct lttcomm_ust_msg lum;
 364
 365                 len = lttcomm_recv_unix_sock(sock, &lum, sizeof(lum));
 366                 switch (len) {
 367                 case 0: /* orderly shutdown */
 368                         DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
 369                         goto end;
 370                 case sizeof(lum):
 371                         DBG("message received\n");
 372                         ret = handle_message(sock_info, sock, &lum);
 373                         if (ret < 0) {
 374                                 ERR("Error handling message for %s socket", sock_info->name);
 375                         }
 376                         continue;
 377                 case -1:
 378                         if (errno == ECONNRESET) {
 379                                 ERR("%s remote end closed connection\n", sock_info->name);
 380                                 goto end;
 381                         }
 382                         goto end;
 383                 default:
 384                         ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
 385                         continue;
 386                 }
 387
 388         }
 389 end:
 390         goto restart;   /* try to reconnect */
 391 quit:
 392         return NULL;
 393 }
 394
 395 /*
 396  * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
 397  */
 398 static
 399 int get_timeout(struct timespec *constructor_timeout)
 400 {
 401         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 402         char *str_delay;
 403         int ret;
 404
 405         str_delay = getenv("UST_REGISTER_TIMEOUT");
 406         if (str_delay) {
 407                 constructor_delay_ms = strtol(str_delay, NULL, 10);
 408         }
 409
 410         switch (constructor_delay_ms) {
 411         case -1:/* fall-through */
 412         case 0:
 413                 return constructor_delay_ms;
 414         default:
 415                 break;
 416         }
 417
 418         /*
 419          * If we are unable to find the current time, don't wait.
 420          */
 421         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 422         if (ret) {
 423                 return -1;
 424         }
 425         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 426         constructor_timeout->tv_nsec +=
 427                 (constructor_delay_ms % 1000UL) * 1000000UL;
 428         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 429                 constructor_timeout->tv_sec++;
 430                 constructor_timeout->tv_nsec -= 1000000000UL;
 431         }
 432         return 1;
 433 }
 434
 435 /*
 436  * sessiond monitoring thread: monitor presence of global and per-user
 437  * sessiond by polling the application common named pipe.
 438  */
 439 /* TODO */
 440
 441 void __attribute__((constructor)) lttng_ust_init(void)
 442 {
 443         struct timespec constructor_timeout;
 444         int timeout_mode;
 445         int ret;
 446
 447         if (uatomic_xchg(&initialized, 1) == 1)
 448                 return;
 449
 450         /*
 451          * We want precise control over the order in which we construct
 452          * our sub-libraries vs starting to receive commands from
 453          * sessiond (otherwise leading to errors when trying to create
 454          * sessiond before the init functions are completed).
 455          */
 456         init_usterr();
 457         init_tracepoint();
 458         ltt_ring_buffer_metadata_client_init();
 459         ltt_ring_buffer_client_overwrite_init();
 460         ltt_ring_buffer_client_discard_init();
 461
 462         timeout_mode = get_timeout(&constructor_timeout);
 463
 464         ret = sem_init(&constructor_wait, 0, 0);
 465         assert(!ret);
 466
 467         ret = setup_local_apps();
 468         if (ret) {
 469                 ERR("Error setting up to local apps");
 470         }
 471         ret = pthread_create(&local_apps.ust_listener, NULL,
 472                         ust_listener_thread, &local_apps);
 473
 474         if (local_apps.allowed) {
 475                 ret = pthread_create(&global_apps.ust_listener, NULL,
 476                                 ust_listener_thread, &global_apps);
 477         } else {
 478                 handle_register_done(&local_apps);
 479         }
 480
 481         switch (timeout_mode) {
 482         case 1: /* timeout wait */
 483                 do {
 484                         ret = sem_timedwait(&constructor_wait,
 485                                         &constructor_timeout);
 486                 } while (ret < 0 && errno == EINTR);
 487                 if (ret < 0 && errno == ETIMEDOUT) {
 488                         ERR("Timed out waiting for ltt-sessiond");
 489                 } else {
 490                         assert(!ret);
 491                 }
 492                 break;
 493         case -1:/* wait forever */
 494                 do {
 495                         ret = sem_wait(&constructor_wait);
 496                 } while (ret < 0 && errno == EINTR);
 497                 assert(!ret);
 498                 break;
 499         case 0: /* no timeout */
 500                 break;
 501         }
 502 }
 503
 504 static
 505 void lttng_ust_cleanup(int exiting)
 506 {
 507         cleanup_sock_info(&global_apps);
 508         if (local_apps.allowed) {
 509                 cleanup_sock_info(&local_apps);
 510         }
 511         lttng_ust_abi_exit();
 512         ltt_events_exit();
 513         ltt_ring_buffer_client_discard_exit();
 514         ltt_ring_buffer_client_overwrite_exit();
 515         ltt_ring_buffer_metadata_client_exit();
 516         exit_tracepoint();
 517         if (!exiting) {
 518                 /* Reinitialize values for fork */
 519                 sem_count = 2;
 520                 lttng_ust_comm_should_quit = 0;
 521                 initialized = 0;
 522         }
 523 }
 524
 525 void __attribute__((destructor)) lttng_ust_exit(void)
 526 {
 527         int ret;
 528
 529         /*
 530          * Using pthread_cancel here because:
 531          * A) we don't want to hang application teardown.
 532          * B) the thread is not allocating any resource.
 533          */
 534
 535         /*
 536          * Require the communication thread to quit. Synchronize with
 537          * mutexes to ensure it is not in a mutex critical section when
 538          * pthread_cancel is later called.
 539          */
 540         ust_lock();
 541         lttng_ust_comm_should_quit = 1;
 542         ust_unlock();
 543
 544         ret = pthread_cancel(global_apps.ust_listener);
 545         if (ret) {
 546                 ERR("Error cancelling global ust listener thread");
 547         }
 548         if (local_apps.allowed) {
 549                 ret = pthread_cancel(local_apps.ust_listener);
 550                 if (ret) {
 551                         ERR("Error cancelling local ust listener thread");
 552                 }
 553         }
 554         lttng_ust_cleanup(1);
 555 }
 556
 557 /*
 558  * We exclude the worker threads across fork and clone (except
 559  * CLONE_VM), because these system calls only keep the forking thread
 560  * running in the child.  Therefore, we don't want to call fork or clone
 561  * in the middle of an tracepoint or ust tracing state modification.
 562  * Holding this mutex protects these structures across fork and clone.
 563  */
 564 void ust_before_fork(ust_fork_info_t *fork_info)
 565 {
 566         /*
 567          * Disable signals. This is to avoid that the child intervenes
 568          * before it is properly setup for tracing. It is safer to
 569          * disable all signals, because then we know we are not breaking
 570          * anything by restoring the original mask.
 571          */
 572         sigset_t all_sigs;
 573         int ret;
 574
 575         /* Disable signals */
 576         sigfillset(&all_sigs);
 577         ret = sigprocmask(SIG_BLOCK, &all_sigs, &fork_info->orig_sigs);
 578         if (ret == -1) {
 579                 PERROR("sigprocmask");
 580         }
 581         ust_lock();
 582         rcu_bp_before_fork();
 583 }
 584
 585 static void ust_after_fork_common(ust_fork_info_t *fork_info)
 586 {
 587         int ret;
 588
 589         DBG("process %d", getpid());
 590         ust_unlock();
 591         /* Restore signals */
 592         ret = sigprocmask(SIG_SETMASK, &fork_info->orig_sigs, NULL);
 593         if (ret == -1) {
 594                 PERROR("sigprocmask");
 595         }
 596 }
 597
 598 void ust_after_fork_parent(ust_fork_info_t *fork_info)
 599 {
 600         DBG("process %d", getpid());
 601         rcu_bp_after_fork_parent();
 602         /* Release mutexes and reenable signals */
 603         ust_after_fork_common(fork_info);
 604 }
 605
 606 /*
 607  * After fork, in the child, we need to cleanup all the leftover state,
 608  * except the worker thread which already magically disappeared thanks
 609  * to the weird Linux fork semantics. After tyding up, we call
 610  * lttng_ust_init() again to start over as a new PID.
 611  *
 612  * This is meant for forks() that have tracing in the child between the
 613  * fork and following exec call (if there is any).
 614  */
 615 void ust_after_fork_child(ust_fork_info_t *fork_info)
 616 {
 617         DBG("process %d", getpid());
 618         /* Release urcu mutexes */
 619         rcu_bp_after_fork_child();
 620         lttng_ust_cleanup(0);
 621         lttng_ust_init();
 622         /* Release mutexes and reenable signals */
 623         ust_after_fork_common(fork_info);
 624 }