src/bin/lttng-relayd/live.c

   1 /*
   2  * Copyright (C) 2013 - Julien Desfossez <jdesfossez@efficios.com>
   3  *                      David Goulet <dgoulet@efficios.com>
   4  *               2015 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License, version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  13  * more details.
  14  *
  15  * You should have received a copy of the GNU General Public License along
  16  * with this program; if not, write to the Free Software Foundation, Inc.,
  17  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  18  */
  19
  20 #define _GNU_SOURCE
  21 #include <getopt.h>
  22 #include <grp.h>
  23 #include <limits.h>
  24 #include <pthread.h>
  25 #include <signal.h>
  26 #include <stdio.h>
  27 #include <stdlib.h>
  28 #include <string.h>
  29 #include <sys/mman.h>
  30 #include <sys/mount.h>
  31 #include <sys/resource.h>
  32 #include <sys/socket.h>
  33 #include <sys/stat.h>
  34 #include <sys/types.h>
  35 #include <sys/wait.h>
  36 #include <inttypes.h>
  37 #include <urcu/futex.h>
  38 #include <urcu/uatomic.h>
  39 #include <urcu/rculist.h>
  40 #include <unistd.h>
  41 #include <fcntl.h>
  42 #include <config.h>
  43
  44 #include <lttng/lttng.h>
  45 #include <common/common.h>
  46 #include <common/compat/poll.h>
  47 #include <common/compat/socket.h>
  48 #include <common/compat/endian.h>
  49 #include <common/defaults.h>
  50 #include <common/futex.h>
  51 #include <common/index/index.h>
  52 #include <common/sessiond-comm/sessiond-comm.h>
  53 #include <common/sessiond-comm/inet.h>
  54 #include <common/sessiond-comm/relayd.h>
  55 #include <common/uri.h>
  56 #include <common/utils.h>
  57
  58 #include "cmd.h"
  59 #include "live.h"
  60 #include "lttng-relayd.h"
  61 #include "utils.h"
  62 #include "health-relayd.h"
  63 #include "testpoint.h"
  64 #include "viewer-stream.h"
  65 #include "stream.h"
  66 #include "session.h"
  67 #include "ctf-trace.h"
  68 #include "connection.h"
  69 #include "viewer-session.h"
  70
  71 #define SESSION_BUF_DEFAULT_COUNT       16
  72
  73 static struct lttng_uri *live_uri;
  74
  75 /*
  76  * This pipe is used to inform the worker thread that a command is queued and
  77  * ready to be processed.
  78  */
  79 static int live_conn_pipe[2] = { -1, -1 };
  80
  81 /* Shared between threads */
  82 static int live_dispatch_thread_exit;
  83
  84 static pthread_t live_listener_thread;
  85 static pthread_t live_dispatcher_thread;
  86 static pthread_t live_worker_thread;
  87
  88 /*
  89  * Relay command queue.
  90  *
  91  * The live_thread_listener and live_thread_dispatcher communicate with this
  92  * queue.
  93  */
  94 static struct relay_conn_queue viewer_conn_queue;
  95
  96 static uint64_t last_relay_viewer_session_id;
  97 static pthread_mutex_t last_relay_viewer_session_id_lock =
  98                 PTHREAD_MUTEX_INITIALIZER;
  99
 100 /*
 101  * Cleanup the daemon
 102  */
 103 static
 104 void cleanup_relayd_live(void)
 105 {
 106         DBG("Cleaning up");
 107
 108         free(live_uri);
 109 }
 110
 111 /*
 112  * Receive a request buffer using a given socket, destination allocated buffer
 113  * of length size.
 114  *
 115  * Return the size of the received message or else a negative value on error
 116  * with errno being set by recvmsg() syscall.
 117  */
 118 static
 119 ssize_t recv_request(struct lttcomm_sock *sock, void *buf, size_t size)
 120 {
 121         ssize_t ret;
 122
 123         ret = sock->ops->recvmsg(sock, buf, size, 0);
 124         if (ret < 0 || ret != size) {
 125                 if (ret == 0) {
 126                         /* Orderly shutdown. Not necessary to print an error. */
 127                         DBG("Socket %d did an orderly shutdown", sock->fd);
 128                 } else {
 129                         ERR("Relay failed to receive request.");
 130                 }
 131                 ret = -1;
 132         }
 133
 134         return ret;
 135 }
 136
 137 /*
 138  * Send a response buffer using a given socket, source allocated buffer of
 139  * length size.
 140  *
 141  * Return the size of the sent message or else a negative value on error with
 142  * errno being set by sendmsg() syscall.
 143  */
 144 static
 145 ssize_t send_response(struct lttcomm_sock *sock, void *buf, size_t size)
 146 {
 147         ssize_t ret;
 148
 149         ret = sock->ops->sendmsg(sock, buf, size, 0);
 150         if (ret < 0) {
 151                 ERR("Relayd failed to send response.");
 152         }
 153
 154         return ret;
 155 }
 156
 157 /*
 158  * Atomically check if new streams got added in one of the sessions attached
 159  * and reset the flag to 0.
 160  *
 161  * Returns 1 if new streams got added, 0 if nothing changed, a negative value
 162  * on error.
 163  */
 164 static
 165 int check_new_streams(struct relay_connection *conn)
 166 {
 167         struct relay_session *session;
 168         unsigned long current_val;
 169         int ret = 0;
 170
 171         if (!conn->viewer_session) {
 172                 goto end;
 173         }
 174         rcu_read_lock();
 175         cds_list_for_each_entry_rcu(session,
 176                         &conn->viewer_session->session_list,
 177                         viewer_session_node) {
 178                 if (!session_get(session)) {
 179                         continue;
 180                 }
 181                 current_val = uatomic_cmpxchg(&session->new_streams, 1, 0);
 182                 ret = current_val;
 183                 session_put(session);
 184                 if (ret == 1) {
 185                         goto end;
 186                 }
 187         }
 188 end:
 189         rcu_read_unlock();
 190         return ret;
 191 }
 192
 193 /*
 194  * Send viewer streams to the given socket. The ignore_sent_flag indicates if
 195  * this function should ignore the sent flag or not.
 196  *
 197  * Return 0 on success or else a negative value.
 198  */
 199 static
 200 ssize_t send_viewer_streams(struct lttcomm_sock *sock,
 201                 struct relay_session *session, unsigned int ignore_sent_flag)
 202 {
 203         ssize_t ret;
 204         struct lttng_viewer_stream send_stream;
 205         struct lttng_ht_iter iter;
 206         struct relay_viewer_stream *vstream;
 207
 208         rcu_read_lock();
 209
 210         cds_lfht_for_each_entry(viewer_streams_ht->ht, &iter.iter, vstream,
 211                         stream_n.node) {
 212                 struct ctf_trace *ctf_trace;
 213
 214                 health_code_update();
 215
 216                 if (!viewer_stream_get(vstream)) {
 217                         continue;
 218                 }
 219
 220                 pthread_mutex_lock(&vstream->stream->lock);
 221                 /* Ignore if not the same session. */
 222                 if (vstream->stream->trace->session->id != session->id ||
 223                                 (!ignore_sent_flag && vstream->sent_flag)) {
 224                         pthread_mutex_unlock(&vstream->stream->lock);
 225                         viewer_stream_put(vstream);
 226                         continue;
 227                 }
 228
 229                 ctf_trace = vstream->stream->trace;
 230                 send_stream.id = htobe64(vstream->stream->stream_handle);
 231                 send_stream.ctf_trace_id = htobe64(ctf_trace->id);
 232                 send_stream.metadata_flag = htobe32(
 233                                 vstream->stream->is_metadata);
 234                 strncpy(send_stream.path_name, vstream->path_name,
 235                                 sizeof(send_stream.path_name));
 236                 strncpy(send_stream.channel_name, vstream->channel_name,
 237                                 sizeof(send_stream.channel_name));
 238
 239                 DBG("Sending stream %" PRIu64 " to viewer",
 240                                 vstream->stream->stream_handle);
 241                 vstream->sent_flag = 1;
 242                 pthread_mutex_unlock(&vstream->stream->lock);
 243
 244                 ret = send_response(sock, &send_stream, sizeof(send_stream));
 245                 viewer_stream_put(vstream);
 246                 if (ret < 0) {
 247                         goto end_unlock;
 248                 }
 249         }
 250
 251         ret = 0;
 252
 253 end_unlock:
 254         rcu_read_unlock();
 255         return ret;
 256 }
 257
 258 /*
 259  * Create every viewer stream possible for the given session with the seek
 260  * type. Three counters *can* be return which are in order the total amount of
 261  * viewer stream of the session, the number of unsent stream and the number of
 262  * stream created. Those counters can be NULL and thus will be ignored.
 263  *
 264  * Return 0 on success or else a negative value.
 265  */
 266 static
 267 int make_viewer_streams(struct relay_session *session,
 268                 enum lttng_viewer_seek seek_t, uint32_t *nb_total, uint32_t *nb_unsent,
 269                 uint32_t *nb_created, bool *closed)
 270 {
 271         int ret;
 272         struct lttng_ht_iter iter;
 273         struct ctf_trace *ctf_trace;
 274
 275         assert(session);
 276
 277         /*
 278          * Hold the session lock to ensure that we see either none or
 279          * all initial streams for a session, but no intermediate state.
 280          */
 281         pthread_mutex_lock(&session->lock);
 282
 283         if (session->connection_closed) {
 284                 *closed = true;
 285         }
 286
 287         /*
 288          * Create viewer streams for relay streams that are ready to be
 289          * used for a the given session id only.
 290          */
 291         rcu_read_lock();
 292         cds_lfht_for_each_entry(session->ctf_traces_ht->ht, &iter.iter, ctf_trace,
 293                         node.node) {
 294                 struct relay_stream *stream;
 295
 296                 health_code_update();
 297
 298                 if (!ctf_trace_get(ctf_trace)) {
 299                         continue;
 300                 }
 301
 302                 cds_list_for_each_entry_rcu(stream, &ctf_trace->stream_list, stream_node) {
 303                         struct relay_viewer_stream *vstream;
 304
 305                         if (!stream_get(stream)) {
 306                                 continue;
 307                         }
 308                         /*
 309                          * stream published is protected by the session
 310                          * lock.
 311                          */
 312                         if (!stream->published) {
 313                                 goto next;
 314                         }
 315                         vstream = viewer_stream_get_by_id(stream->stream_handle);
 316                         if (!vstream) {
 317                                 vstream = viewer_stream_create(stream, seek_t);
 318                                 if (!vstream) {
 319                                         ret = -1;
 320                                         ctf_trace_put(ctf_trace);
 321                                         stream_put(stream);
 322                                         goto error_unlock;
 323                                 }
 324
 325                                 if (nb_created) {
 326                                         /* Update number of created stream counter. */
 327                                         (*nb_created)++;
 328                                 }
 329                                 /*
 330                                  * Ensure a self-reference is preserved even
 331                                  * after we have put our local reference.
 332                                  */
 333                                 viewer_stream_get(vstream);
 334                         } else {
 335                                 if (!vstream->sent_flag && nb_unsent) {
 336                                         /* Update number of unsent stream counter. */
 337                                         (*nb_unsent)++;
 338                                 }
 339                         }
 340                         /* Update number of total stream counter. */
 341                         if (nb_total) {
 342                                 if (stream->is_metadata) {
 343                                         if (!stream->closed ||
 344                                                         stream->metadata_received > vstream->metadata_sent) {
 345                                                 (*nb_total)++;
 346                                         }
 347                                 } else {
 348                                         if (!stream->closed ||
 349                                                 !(((int64_t) (stream->prev_seq - stream->last_net_seq_num)) >= 0)) {
 350
 351                                                 (*nb_total)++;
 352                                         }
 353                                 }
 354                         }
 355                         /* Put local reference. */
 356                         viewer_stream_put(vstream);
 357                 next:
 358                         stream_put(stream);
 359                 }
 360                 ctf_trace_put(ctf_trace);
 361         }
 362
 363         ret = 0;
 364
 365 error_unlock:
 366         rcu_read_unlock();
 367         pthread_mutex_unlock(&session->lock);
 368         return ret;
 369 }
 370
 371 int relayd_live_stop(void)
 372 {
 373         /* Stop dispatch thread */
 374         CMM_STORE_SHARED(live_dispatch_thread_exit, 1);
 375         futex_nto1_wake(&viewer_conn_queue.futex);
 376         return 0;
 377 }
 378
 379 /*
 380  * Create a poll set with O_CLOEXEC and add the thread quit pipe to the set.
 381  */
 382 static
 383 int create_thread_poll_set(struct lttng_poll_event *events, int size)
 384 {
 385         int ret;
 386
 387         if (events == NULL || size == 0) {
 388                 ret = -1;
 389                 goto error;
 390         }
 391
 392         ret = lttng_poll_create(events, size, LTTNG_CLOEXEC);
 393         if (ret < 0) {
 394                 goto error;
 395         }
 396
 397         /* Add quit pipe */
 398         ret = lttng_poll_add(events, thread_quit_pipe[0], LPOLLIN | LPOLLERR);
 399         if (ret < 0) {
 400                 goto error;
 401         }
 402
 403         return 0;
 404
 405 error:
 406         return ret;
 407 }
 408
 409 /*
 410  * Check if the thread quit pipe was triggered.
 411  *
 412  * Return 1 if it was triggered else 0;
 413  */
 414 static
 415 int check_thread_quit_pipe(int fd, uint32_t events)
 416 {
 417         if (fd == thread_quit_pipe[0] && (events & LPOLLIN)) {
 418                 return 1;
 419         }
 420
 421         return 0;
 422 }
 423
 424 /*
 425  * Create and init socket from uri.
 426  */
 427 static
 428 struct lttcomm_sock *init_socket(struct lttng_uri *uri)
 429 {
 430         int ret;
 431         struct lttcomm_sock *sock = NULL;
 432
 433         sock = lttcomm_alloc_sock_from_uri(uri);
 434         if (sock == NULL) {
 435                 ERR("Allocating socket");
 436                 goto error;
 437         }
 438
 439         ret = lttcomm_create_sock(sock);
 440         if (ret < 0) {
 441                 goto error;
 442         }
 443         DBG("Listening on sock %d for live", sock->fd);
 444
 445         ret = sock->ops->bind(sock);
 446         if (ret < 0) {
 447                 goto error;
 448         }
 449
 450         ret = sock->ops->listen(sock, -1);
 451         if (ret < 0) {
 452                 goto error;
 453
 454         }
 455
 456         return sock;
 457
 458 error:
 459         if (sock) {
 460                 lttcomm_destroy_sock(sock);
 461         }
 462         return NULL;
 463 }
 464
 465 /*
 466  * This thread manages the listening for new connections on the network
 467  */
 468 static
 469 void *thread_listener(void *data)
 470 {
 471         int i, ret, pollfd, err = -1;
 472         uint32_t revents, nb_fd;
 473         struct lttng_poll_event events;
 474         struct lttcomm_sock *live_control_sock;
 475
 476         DBG("[thread] Relay live listener started");
 477
 478         health_register(health_relayd, HEALTH_RELAYD_TYPE_LIVE_LISTENER);
 479
 480         health_code_update();
 481
 482         live_control_sock = init_socket(live_uri);
 483         if (!live_control_sock) {
 484                 goto error_sock_control;
 485         }
 486
 487         /* Pass 2 as size here for the thread quit pipe and control sockets. */
 488         ret = create_thread_poll_set(&events, 2);
 489         if (ret < 0) {
 490                 goto error_create_poll;
 491         }
 492
 493         /* Add the control socket */
 494         ret = lttng_poll_add(&events, live_control_sock->fd, LPOLLIN | LPOLLRDHUP);
 495         if (ret < 0) {
 496                 goto error_poll_add;
 497         }
 498
 499         lttng_relay_notify_ready();
 500
 501         if (testpoint(relayd_thread_live_listener)) {
 502                 goto error_testpoint;
 503         }
 504
 505         while (1) {
 506                 health_code_update();
 507
 508                 DBG("Listener accepting live viewers connections");
 509
 510 restart:
 511                 health_poll_entry();
 512                 ret = lttng_poll_wait(&events, -1);
 513                 health_poll_exit();
 514                 if (ret < 0) {
 515                         /*
 516                          * Restart interrupted system call.
 517                          */
 518                         if (errno == EINTR) {
 519                                 goto restart;
 520                         }
 521                         goto error;
 522                 }
 523                 nb_fd = ret;
 524
 525                 DBG("Relay new viewer connection received");
 526                 for (i = 0; i < nb_fd; i++) {
 527                         health_code_update();
 528
 529                         /* Fetch once the poll data */
 530                         revents = LTTNG_POLL_GETEV(&events, i);
 531                         pollfd = LTTNG_POLL_GETFD(&events, i);
 532
 533                         if (!revents) {
 534                                 /* No activity for this FD (poll implementation). */
 535                                 continue;
 536                         }
 537
 538                         /* Thread quit pipe has been closed. Killing thread. */
 539                         ret = check_thread_quit_pipe(pollfd, revents);
 540                         if (ret) {
 541                                 err = 0;
 542                                 goto exit;
 543                         }
 544
 545                         if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
 546                                 ERR("socket poll error");
 547                                 goto error;
 548                         } else if (revents & LPOLLIN) {
 549                                 /*
 550                                  * A new connection is requested, therefore a
 551                                  * viewer connection is allocated in this
 552                                  * thread, enqueued to a global queue and
 553                                  * dequeued (and freed) in the worker thread.
 554                                  */
 555                                 int val = 1;
 556                                 struct relay_connection *new_conn;
 557                                 struct lttcomm_sock *newsock;
 558
 559                                 newsock = live_control_sock->ops->accept(live_control_sock);
 560                                 if (!newsock) {
 561                                         PERROR("accepting control sock");
 562                                         goto error;
 563                                 }
 564                                 DBG("Relay viewer connection accepted socket %d", newsock->fd);
 565
 566                                 ret = setsockopt(newsock->fd, SOL_SOCKET, SO_REUSEADDR, &val,
 567                                                 sizeof(val));
 568                                 if (ret < 0) {
 569                                         PERROR("setsockopt inet");
 570                                         lttcomm_destroy_sock(newsock);
 571                                         goto error;
 572                                 }
 573                                 new_conn = connection_create(newsock, RELAY_CONNECTION_UNKNOWN);
 574                                 if (!new_conn) {
 575                                         lttcomm_destroy_sock(newsock);
 576                                         goto error;
 577                                 }
 578                                 /* Ownership assumed by the connection. */
 579                                 newsock = NULL;
 580
 581                                 /* Enqueue request for the dispatcher thread. */
 582                                 cds_wfcq_enqueue(&viewer_conn_queue.head, &viewer_conn_queue.tail,
 583                                                  &new_conn->qnode);
 584
 585                                 /*
 586                                  * Wake the dispatch queue futex.
 587                                  * Implicit memory barrier with the
 588                                  * exchange in cds_wfcq_enqueue.
 589                                  */
 590                                 futex_nto1_wake(&viewer_conn_queue.futex);
 591                         }
 592                 }
 593         }
 594
 595 exit:
 596 error:
 597 error_poll_add:
 598 error_testpoint:
 599         lttng_poll_clean(&events);
 600 error_create_poll:
 601         if (live_control_sock->fd >= 0) {
 602                 ret = live_control_sock->ops->close(live_control_sock);
 603                 if (ret) {
 604                         PERROR("close");
 605                 }
 606         }
 607         lttcomm_destroy_sock(live_control_sock);
 608 error_sock_control:
 609         if (err) {
 610                 health_error();
 611                 DBG("Live viewer listener thread exited with error");
 612         }
 613         health_unregister(health_relayd);
 614         DBG("Live viewer listener thread cleanup complete");
 615         if (lttng_relay_stop_threads()) {
 616                 ERR("Error stopping threads");
 617         }
 618         return NULL;
 619 }
 620
 621 /*
 622  * This thread manages the dispatching of the requests to worker threads
 623  */
 624 static
 625 void *thread_dispatcher(void *data)
 626 {
 627         int err = -1;
 628         ssize_t ret;
 629         struct cds_wfcq_node *node;
 630         struct relay_connection *conn = NULL;
 631
 632         DBG("[thread] Live viewer relay dispatcher started");
 633
 634         health_register(health_relayd, HEALTH_RELAYD_TYPE_LIVE_DISPATCHER);
 635
 636         if (testpoint(relayd_thread_live_dispatcher)) {
 637                 goto error_testpoint;
 638         }
 639
 640         health_code_update();
 641
 642         while (!CMM_LOAD_SHARED(live_dispatch_thread_exit)) {
 643                 health_code_update();
 644
 645                 /* Atomically prepare the queue futex */
 646                 futex_nto1_prepare(&viewer_conn_queue.futex);
 647
 648                 do {
 649                         health_code_update();
 650
 651                         /* Dequeue commands */
 652                         node = cds_wfcq_dequeue_blocking(&viewer_conn_queue.head,
 653                                                          &viewer_conn_queue.tail);
 654                         if (node == NULL) {
 655                                 DBG("Woken up but nothing in the live-viewer "
 656                                                 "relay command queue");
 657                                 /* Continue thread execution */
 658                                 break;
 659                         }
 660                         conn = caa_container_of(node, struct relay_connection, qnode);
 661                         DBG("Dispatching viewer request waiting on sock %d",
 662                                         conn->sock->fd);
 663
 664                         /*
 665                          * Inform worker thread of the new request. This
 666                          * call is blocking so we can be assured that
 667                          * the data will be read at some point in time
 668                          * or wait to the end of the world :)
 669                          */
 670                         ret = lttng_write(live_conn_pipe[1], &conn, sizeof(conn));
 671                         if (ret < 0) {
 672                                 PERROR("write conn pipe");
 673                                 connection_put(conn);
 674                                 goto error;
 675                         }
 676                 } while (node != NULL);
 677
 678                 /* Futex wait on queue. Blocking call on futex() */
 679                 health_poll_entry();
 680                 futex_nto1_wait(&viewer_conn_queue.futex);
 681                 health_poll_exit();
 682         }
 683
 684         /* Normal exit, no error */
 685         err = 0;
 686
 687 error:
 688 error_testpoint:
 689         if (err) {
 690                 health_error();
 691                 ERR("Health error occurred in %s", __func__);
 692         }
 693         health_unregister(health_relayd);
 694         DBG("Live viewer dispatch thread dying");
 695         if (lttng_relay_stop_threads()) {
 696                 ERR("Error stopping threads");
 697         }
 698         return NULL;
 699 }
 700
 701 /*
 702  * Establish connection with the viewer and check the versions.
 703  *
 704  * Return 0 on success or else negative value.
 705  */
 706 static
 707 int viewer_connect(struct relay_connection *conn)
 708 {
 709         int ret;
 710         struct lttng_viewer_connect reply, msg;
 711
 712         conn->version_check_done = 1;
 713
 714         health_code_update();
 715
 716         DBG("Viewer is establishing a connection to the relayd.");
 717
 718         ret = recv_request(conn->sock, &msg, sizeof(msg));
 719         if (ret < 0) {
 720                 goto end;
 721         }
 722
 723         health_code_update();
 724
 725         memset(&reply, 0, sizeof(reply));
 726         reply.major = RELAYD_VERSION_COMM_MAJOR;
 727         reply.minor = RELAYD_VERSION_COMM_MINOR;
 728
 729         /* Major versions must be the same */
 730         if (reply.major != be32toh(msg.major)) {
 731                 DBG("Incompatible major versions ([relayd] %u vs [client] %u)",
 732                                 reply.major, be32toh(msg.major));
 733                 ret = -1;
 734                 goto end;
 735         }
 736
 737         conn->major = reply.major;
 738         /* We adapt to the lowest compatible version */
 739         if (reply.minor <= be32toh(msg.minor)) {
 740                 conn->minor = reply.minor;
 741         } else {
 742                 conn->minor = be32toh(msg.minor);
 743         }
 744
 745         if (be32toh(msg.type) == LTTNG_VIEWER_CLIENT_COMMAND) {
 746                 conn->type = RELAY_VIEWER_COMMAND;
 747         } else if (be32toh(msg.type) == LTTNG_VIEWER_CLIENT_NOTIFICATION) {
 748                 conn->type = RELAY_VIEWER_NOTIFICATION;
 749         } else {
 750                 ERR("Unknown connection type : %u", be32toh(msg.type));
 751                 ret = -1;
 752                 goto end;
 753         }
 754
 755         reply.major = htobe32(reply.major);
 756         reply.minor = htobe32(reply.minor);
 757         if (conn->type == RELAY_VIEWER_COMMAND) {
 758                 /*
 759                  * Increment outside of htobe64 macro, because the argument can
 760                  * be used more than once within the macro, and thus the
 761                  * operation may be undefined.
 762                  */
 763                 pthread_mutex_lock(&last_relay_viewer_session_id_lock);
 764                 last_relay_viewer_session_id++;
 765                 pthread_mutex_unlock(&last_relay_viewer_session_id_lock);
 766                 reply.viewer_session_id = htobe64(last_relay_viewer_session_id);
 767         }
 768
 769         health_code_update();
 770
 771         ret = send_response(conn->sock, &reply, sizeof(reply));
 772         if (ret < 0) {
 773                 goto end;
 774         }
 775
 776         health_code_update();
 777
 778         DBG("Version check done using protocol %u.%u", conn->major, conn->minor);
 779         ret = 0;
 780
 781 end:
 782         return ret;
 783 }
 784
 785 /*
 786  * Send the viewer the list of current sessions.
 787  * We need to create a copy of the hash table content because otherwise
 788  * we cannot assume the number of entries stays the same between getting
 789  * the number of HT elements and iteration over the HT.
 790  *
 791  * Return 0 on success or else a negative value.
 792  */
 793 static
 794 int viewer_list_sessions(struct relay_connection *conn)
 795 {
 796         int ret;
 797         struct lttng_viewer_list_sessions session_list;
 798         struct lttng_ht_iter iter;
 799         struct relay_session *session;
 800         struct lttng_viewer_session *send_session_buf = NULL;
 801         uint32_t buf_count = SESSION_BUF_DEFAULT_COUNT;
 802         uint32_t count = 0;
 803
 804         DBG("List sessions received");
 805
 806         send_session_buf = zmalloc(SESSION_BUF_DEFAULT_COUNT * sizeof(*send_session_buf));
 807         if (!send_session_buf) {
 808                 return -1;
 809         }
 810
 811         rcu_read_lock();
 812         cds_lfht_for_each_entry(sessions_ht->ht, &iter.iter, session,
 813                         session_n.node) {
 814                 struct lttng_viewer_session *send_session;
 815
 816                 health_code_update();
 817
 818                 if (count >= buf_count) {
 819                         struct lttng_viewer_session *newbuf;
 820                         uint32_t new_buf_count = buf_count << 1;
 821
 822                         newbuf = realloc(send_session_buf,
 823                                 new_buf_count * sizeof(*send_session_buf));
 824                         if (!newbuf) {
 825                                 ret = -1;
 826                                 rcu_read_unlock();
 827                                 goto end_free;
 828                         }
 829                         send_session_buf = newbuf;
 830                         buf_count = new_buf_count;
 831                 }
 832                 send_session = &send_session_buf[count];
 833                 strncpy(send_session->session_name, session->session_name,
 834                                 sizeof(send_session->session_name));
 835                 strncpy(send_session->hostname, session->hostname,
 836                                 sizeof(send_session->hostname));
 837                 send_session->id = htobe64(session->id);
 838                 send_session->live_timer = htobe32(session->live_timer);
 839                 if (session->viewer_attached) {
 840                         send_session->clients = htobe32(1);
 841                 } else {
 842                         send_session->clients = htobe32(0);
 843                 }
 844                 send_session->streams = htobe32(session->stream_count);
 845                 count++;
 846         }
 847         rcu_read_unlock();
 848
 849         session_list.sessions_count = htobe32(count);
 850
 851         health_code_update();
 852
 853         ret = send_response(conn->sock, &session_list, sizeof(session_list));
 854         if (ret < 0) {
 855                 goto end_free;
 856         }
 857
 858         health_code_update();
 859
 860         ret = send_response(conn->sock, send_session_buf,
 861                         count * sizeof(*send_session_buf));
 862         if (ret < 0) {
 863                 goto end_free;
 864         }
 865         health_code_update();
 866
 867         ret = 0;
 868 end_free:
 869         free(send_session_buf);
 870         return ret;
 871 }
 872
 873 /*
 874  * Send the viewer the list of current streams.
 875  */
 876 static
 877 int viewer_get_new_streams(struct relay_connection *conn)
 878 {
 879         int ret, send_streams = 0;
 880         uint32_t nb_created = 0, nb_unsent = 0, nb_streams = 0, nb_total = 0;
 881         struct lttng_viewer_new_streams_request request;
 882         struct lttng_viewer_new_streams_response response;
 883         struct relay_session *session;
 884         uint64_t session_id;
 885         bool closed = false;
 886
 887         assert(conn);
 888
 889         DBG("Get new streams received");
 890
 891         health_code_update();
 892
 893         /* Receive the request from the connected client. */
 894         ret = recv_request(conn->sock, &request, sizeof(request));
 895         if (ret < 0) {
 896                 goto error;
 897         }
 898         session_id = be64toh(request.session_id);
 899
 900         health_code_update();
 901
 902         memset(&response, 0, sizeof(response));
 903
 904         session = session_get_by_id(session_id);
 905         if (!session) {
 906                 DBG("Relay session %" PRIu64 " not found", session_id);
 907                 response.status = htobe32(LTTNG_VIEWER_NEW_STREAMS_ERR);
 908                 goto send_reply;
 909         }
 910
 911         if (!viewer_session_is_attached(conn->viewer_session, session)) {
 912                 send_streams = 0;
 913                 response.status = htobe32(LTTNG_VIEWER_NEW_STREAMS_ERR);
 914                 goto send_reply;
 915         }
 916
 917         send_streams = 1;
 918         response.status = htobe32(LTTNG_VIEWER_NEW_STREAMS_OK);
 919
 920         ret = make_viewer_streams(session, LTTNG_VIEWER_SEEK_LAST, &nb_total, &nb_unsent,
 921                         &nb_created, &closed);
 922         if (ret < 0) {
 923                 goto end_put_session;
 924         }
 925         /* Only send back the newly created streams with the unsent ones. */
 926         nb_streams = nb_created + nb_unsent;
 927         response.streams_count = htobe32(nb_streams);
 928
 929         /*
 930          * If the session is closed, HUP when there are no more streams
 931          * with data.
 932          */
 933         if (closed && nb_total == 0) {
 934                 send_streams = 0;
 935                 response.streams_count = 0;
 936                 response.status = htobe32(LTTNG_VIEWER_NEW_STREAMS_HUP);
 937                 goto send_reply;
 938         }
 939
 940 send_reply:
 941         health_code_update();
 942         ret = send_response(conn->sock, &response, sizeof(response));
 943         if (ret < 0) {
 944                 goto end_put_session;
 945         }
 946         health_code_update();
 947
 948         /*
 949          * Unknown or empty session, just return gracefully, the viewer
 950          * knows what is happening.
 951          */
 952         if (!send_streams || !nb_streams) {
 953                 ret = 0;
 954                 goto end_put_session;
 955         }
 956
 957         /*
 958          * Send stream and *DON'T* ignore the sent flag so every viewer
 959          * streams that were not sent from that point will be sent to
 960          * the viewer.
 961          */
 962         ret = send_viewer_streams(conn->sock, session, 0);
 963         if (ret < 0) {
 964                 goto end_put_session;
 965         }
 966
 967 end_put_session:
 968         if (session) {
 969                 session_put(session);
 970         }
 971 error:
 972         return ret;
 973 }
 974
 975 /*
 976  * Send the viewer the list of current sessions.
 977  */
 978 static
 979 int viewer_attach_session(struct relay_connection *conn)
 980 {
 981         int send_streams = 0;
 982         ssize_t ret;
 983         uint32_t nb_streams = 0;
 984         enum lttng_viewer_seek seek_type;
 985         struct lttng_viewer_attach_session_request request;
 986         struct lttng_viewer_attach_session_response response;
 987         struct relay_session *session = NULL;
 988         bool closed = false;
 989
 990         assert(conn);
 991
 992         health_code_update();
 993
 994         /* Receive the request from the connected client. */
 995         ret = recv_request(conn->sock, &request, sizeof(request));
 996         if (ret < 0) {
 997                 goto error;
 998         }
 999
1000         health_code_update();
1001
1002         memset(&response, 0, sizeof(response));
1003
1004         if (!conn->viewer_session) {
1005                 DBG("Client trying to attach before creating a live viewer session");
1006                 response.status = htobe32(LTTNG_VIEWER_ATTACH_NO_SESSION);
1007                 goto send_reply;
1008         }
1009
1010         session = session_get_by_id(be64toh(request.session_id));
1011         if (!session) {
1012                 DBG("Relay session %" PRIu64 " not found",
1013                                 be64toh(request.session_id));
1014                 response.status = htobe32(LTTNG_VIEWER_ATTACH_UNK);
1015                 goto send_reply;
1016         }
1017         DBG("Attach session ID %" PRIu64 " received",
1018                 be64toh(request.session_id));
1019
1020         if (session->live_timer == 0) {
1021                 DBG("Not live session");
1022                 response.status = htobe32(LTTNG_VIEWER_ATTACH_NOT_LIVE);
1023                 goto send_reply;
1024         }
1025
1026         send_streams = 1;
1027         ret = viewer_session_attach(conn->viewer_session, session);
1028         if (ret) {
1029                 DBG("Already a viewer attached");
1030                 response.status = htobe32(LTTNG_VIEWER_ATTACH_ALREADY);
1031                 goto send_reply;
1032         }
1033
1034         switch (be32toh(request.seek)) {
1035         case LTTNG_VIEWER_SEEK_BEGINNING:
1036         case LTTNG_VIEWER_SEEK_LAST:
1037                 response.status = htobe32(LTTNG_VIEWER_ATTACH_OK);
1038                 seek_type = be32toh(request.seek);
1039                 break;
1040         default:
1041                 ERR("Wrong seek parameter");
1042                 response.status = htobe32(LTTNG_VIEWER_ATTACH_SEEK_ERR);
1043                 send_streams = 0;
1044                 goto send_reply;
1045         }
1046
1047         ret = make_viewer_streams(session, seek_type, &nb_streams, NULL,
1048                         NULL, &closed);
1049         if (ret < 0) {
1050                 goto end_put_session;
1051         }
1052         response.streams_count = htobe32(nb_streams);
1053
1054         /*
1055          * If the session is closed when the viewer is attaching, it
1056          * means some of the streams may have been concurrently removed,
1057          * so we don't allow the viewer to attach, even if there are
1058          * streams available.
1059          */
1060         if (closed) {
1061                 send_streams = 0;
1062                 response.streams_count = 0;
1063                 response.status = htobe32(LTTNG_VIEWER_NEW_STREAMS_HUP);
1064                 goto send_reply;
1065         }
1066
1067 send_reply:
1068         health_code_update();
1069         ret = send_response(conn->sock, &response, sizeof(response));
1070         if (ret < 0) {
1071                 goto end_put_session;
1072         }
1073         health_code_update();
1074
1075         /*
1076          * Unknown or empty session, just return gracefully, the viewer
1077          * knows what is happening.
1078          */
1079         if (!send_streams || !nb_streams) {
1080                 ret = 0;
1081                 goto end_put_session;
1082         }
1083
1084         /* Send stream and ignore the sent flag. */
1085         ret = send_viewer_streams(conn->sock, session, 1);
1086         if (ret < 0) {
1087                 goto end_put_session;
1088         }
1089
1090 end_put_session:
1091         if (session) {
1092                 session_put(session);
1093         }
1094 error:
1095         return ret;
1096 }
1097
1098 /*
1099  * Open the index file if needed for the given vstream.
1100  *
1101  * If an index file is successfully opened, the vstream index_fd set with
1102  * it.
1103  *
1104  * Return 0 on success, a negative value on error (-ENOENT if not ready yet).
1105  *
1106  * Called with rstream lock held.
1107  */
1108 static int try_open_index(struct relay_viewer_stream *vstream,
1109                 struct relay_stream *rstream)
1110 {
1111         int ret = 0;
1112
1113         if (vstream->index_fd) {
1114                 goto end;
1115         }
1116
1117         /*
1118          * First time, we open the index file and at least one index is ready.
1119          */
1120         if (rstream->total_index_received == 0) {
1121                 ret = -ENOENT;
1122                 goto end;
1123         }
1124         ret = index_open(vstream->path_name, vstream->channel_name,
1125                         vstream->stream->tracefile_count,
1126                         vstream->current_tracefile_id);
1127         if (ret >= 0) {
1128                 vstream->index_fd = stream_fd_create(ret);
1129                 if (!vstream->index_fd) {
1130                         if (close(ret)) {
1131                                 PERROR("close");
1132                         }
1133                         ret = -1;
1134                 } else {
1135                         ret = 0;
1136                 }
1137                 goto end;
1138         }
1139
1140 end:
1141         return ret;
1142 }
1143
1144 /*
1145  * Check the status of the index for the given stream. This function
1146  * updates the index structure if needed and can put (close) the vstream
1147  * in the HUP situation.
1148  *
1149  * Return 0 means that we can proceed with the index. A value of 1 means
1150  * that the index has been updated and is ready to be sent to the
1151  * client. A negative value indicates an error that can't be handled.
1152  *
1153  * Called with rstream lock held.
1154  */
1155 static int check_index_status(struct relay_viewer_stream *vstream,
1156                 struct relay_stream *rstream, struct ctf_trace *trace,
1157                 struct lttng_viewer_index *index)
1158 {
1159         int ret;
1160
1161         if (trace->session->connection_closed
1162                         && rstream->total_index_received
1163                                 == vstream->last_sent_index) {
1164                 /* Last index sent and session connection is closed. */
1165                 index->status = htobe32(LTTNG_VIEWER_INDEX_HUP);
1166                 goto hup;
1167         } else if (rstream->beacon_ts_end != -1ULL &&
1168                         rstream->total_index_received
1169                                 == vstream->last_sent_index) {
1170                 /*
1171                  * We've received a synchronization beacon and the last index
1172                  * available has been sent, the index for now is inactive.
1173                  *
1174                  * In this case, we have received a beacon which allows us to
1175                  * inform the client of a time interval during which we can
1176                  * guarantee that there are no events to read (and never will
1177                  * be).
1178                  */
1179                 index->status = htobe32(LTTNG_VIEWER_INDEX_INACTIVE);
1180                 index->timestamp_end = htobe64(rstream->beacon_ts_end);
1181                 index->stream_id = htobe64(rstream->ctf_stream_id);
1182                 goto index_ready;
1183         } else if (rstream->total_index_received <= vstream->last_sent_index) {
1184                 /*
1185                  * This actually checks the case where recv == last_sent.
1186                  * In this case, we have not received a beacon. Therefore, we
1187                  * can only ask the client to retry later.
1188                  */
1189                 index->status = htobe32(LTTNG_VIEWER_INDEX_RETRY);
1190                 goto index_ready;
1191         } else if (!viewer_stream_is_tracefile_seq_readable(vstream,
1192                         vstream->current_tracefile_seq)) {
1193                 /*
1194                  * The producer has overwritten our current file. We
1195                  * need to rotate.
1196                  */
1197                 DBG("Viewer stream %" PRIu64 " rotation due to overwrite",
1198                                 vstream->stream->stream_handle);
1199                 ret = viewer_stream_rotate(vstream);
1200                 if (ret < 0) {
1201                         goto end;
1202                 } else if (ret == 1) {
1203                         /* EOF across entire stream. */
1204                         index->status = htobe32(LTTNG_VIEWER_INDEX_HUP);
1205                         goto hup;
1206                 }
1207                 assert(viewer_stream_is_tracefile_seq_readable(vstream,
1208                         vstream->current_tracefile_seq));
1209                 /* ret == 0 means successful so we continue. */
1210                 ret = 0;
1211         } else {
1212                 ssize_t read_ret;
1213                 char tmp[1];
1214
1215                 /*
1216                  * Use EOF on current index file to find out when we
1217                  * need to rotate.
1218                  */
1219                 read_ret = lttng_read(vstream->index_fd->fd, tmp, 1);
1220                 if (read_ret == 1) {
1221                         off_t seek_ret;
1222
1223                         /* There is still data to read. Rewind position. */
1224                         seek_ret = lseek(vstream->index_fd->fd, -1, SEEK_CUR);
1225                         if (seek_ret < 0) {
1226                                 ret = -1;
1227                                 goto end;
1228                         }
1229                         ret = 0;
1230                 } else if (read_ret == 0) {
1231                         /* EOF. We need to rotate. */
1232                         DBG("Viewer stream %" PRIu64 " rotation due to EOF",
1233                                         vstream->stream->stream_handle);
1234                         ret = viewer_stream_rotate(vstream);
1235                         if (ret < 0) {
1236                                 goto end;
1237                         } else if (ret == 1) {
1238                                 /* EOF across entire stream. */
1239                                 index->status = htobe32(LTTNG_VIEWER_INDEX_HUP);
1240                                 goto hup;
1241                         }
1242                         assert(viewer_stream_is_tracefile_seq_readable(vstream,
1243                                 vstream->current_tracefile_seq));
1244                         /* ret == 0 means successful so we continue. */
1245                         ret = 0;
1246                 } else {
1247                         /* Error reading index. */
1248                         ret = -1;
1249                 }
1250         }
1251 end:
1252         return ret;
1253
1254 hup:
1255         viewer_stream_put(vstream);
1256 index_ready:
1257         return 1;
1258 }
1259
1260 /*
1261  * Send the next index for a stream.
1262  *
1263  * Return 0 on success or else a negative value.
1264  */
1265 static
1266 int viewer_get_next_index(struct relay_connection *conn)
1267 {
1268         int ret;
1269         ssize_t read_ret;
1270         struct lttng_viewer_get_next_index request_index;
1271         struct lttng_viewer_index viewer_index;
1272         struct ctf_packet_index packet_index;
1273         struct relay_viewer_stream *vstream = NULL;
1274         struct relay_stream *rstream = NULL;
1275         struct ctf_trace *ctf_trace = NULL;
1276         struct relay_viewer_stream *metadata_viewer_stream = NULL;
1277
1278         assert(conn);
1279
1280         DBG("Viewer get next index");
1281
1282         memset(&viewer_index, 0, sizeof(viewer_index));
1283         health_code_update();
1284
1285         ret = recv_request(conn->sock, &request_index, sizeof(request_index));
1286         if (ret < 0) {
1287                 goto end;
1288         }
1289         health_code_update();
1290
1291         vstream = viewer_stream_get_by_id(be64toh(request_index.stream_id));
1292         if (!vstream) {
1293                 viewer_index.status = htobe32(LTTNG_VIEWER_INDEX_ERR);
1294                 goto send_reply;
1295         }
1296
1297         /* Use back. ref. Protected by refcounts. */
1298         rstream = vstream->stream;
1299         ctf_trace = rstream->trace;
1300
1301         /* metadata_viewer_stream may be NULL. */
1302         metadata_viewer_stream =
1303                         ctf_trace_get_viewer_metadata_stream(ctf_trace);
1304
1305         pthread_mutex_lock(&rstream->lock);
1306
1307         /*
1308          * The viewer should not ask for index on metadata stream.
1309          */
1310         if (rstream->is_metadata) {
1311                 viewer_index.status = htobe32(LTTNG_VIEWER_INDEX_HUP);
1312                 goto send_reply;
1313         }
1314
1315         /* Try to open an index if one is needed for that stream. */
1316         ret = try_open_index(vstream, rstream);
1317         if (ret < 0) {
1318                 if (ret == -ENOENT) {
1319                         /*
1320                          * The index is created only when the first data
1321                          * packet arrives, it might not be ready at the
1322                          * beginning of the session
1323                          */
1324                         viewer_index.status = htobe32(LTTNG_VIEWER_INDEX_RETRY);
1325                 } else {
1326                         /* Unhandled error. */
1327                         viewer_index.status = htobe32(LTTNG_VIEWER_INDEX_ERR);
1328                 }
1329                 goto send_reply;
1330         }
1331
1332         ret = check_index_status(vstream, rstream, ctf_trace, &viewer_index);
1333         if (ret < 0) {
1334                 goto error_put;
1335         } else if (ret == 1) {
1336                 /*
1337                  * We have no index to send and check_index_status has populated
1338                  * viewer_index's status.
1339                  */
1340                 goto send_reply;
1341         }
1342         /* At this point, ret is 0 thus we will be able to read the index. */
1343         assert(!ret);
1344
1345         /*
1346          * vstream->stream_fd may be NULL if it has been closed by
1347          * tracefile rotation, or if we are at the beginning of the
1348          * stream. We open the data stream file here to protect against
1349          * overwrite caused by tracefile rotation (in association with
1350          * unlink performed before overwrite).
1351          */
1352         if (!vstream->stream_fd) {
1353                 char fullpath[PATH_MAX];
1354
1355                 if (vstream->stream->tracefile_count > 0) {
1356                         ret = snprintf(fullpath, PATH_MAX, "%s/%s_%" PRIu64,
1357                                         vstream->path_name,
1358                                         vstream->channel_name,
1359                                         vstream->current_tracefile_id);
1360                 } else {
1361                         ret = snprintf(fullpath, PATH_MAX, "%s/%s",
1362                                         vstream->path_name,
1363                                         vstream->channel_name);
1364                 }
1365                 if (ret < 0) {
1366                         goto error_put;
1367                 }
1368                 ret = open(fullpath, O_RDONLY);
1369                 if (ret < 0) {
1370                         PERROR("Relay opening trace file");
1371                         goto error_put;
1372                 }
1373                 vstream->stream_fd = stream_fd_create(ret);
1374                 if (!vstream->stream_fd) {
1375                         if (close(ret)) {
1376                                 PERROR("close");
1377                         }
1378                         goto error_put;
1379                 }
1380         }
1381
1382         ret = check_new_streams(conn);
1383         if (ret < 0) {
1384                 viewer_index.status = htobe32(LTTNG_VIEWER_INDEX_ERR);
1385                 goto send_reply;
1386         } else if (ret == 1) {
1387                 viewer_index.flags |= LTTNG_VIEWER_FLAG_NEW_STREAM;
1388         }
1389
1390         read_ret = lttng_read(vstream->index_fd->fd, &packet_index,
1391                         sizeof(packet_index));
1392         if (read_ret < sizeof(packet_index)) {
1393                 ERR("Relay reading index file %d returned %zd",
1394                         vstream->index_fd->fd, read_ret);
1395                 viewer_index.status = htobe32(LTTNG_VIEWER_INDEX_ERR);
1396                 goto send_reply;
1397         } else {
1398                 viewer_index.status = htobe32(LTTNG_VIEWER_INDEX_OK);
1399                 vstream->last_sent_index++;
1400         }
1401
1402         /*
1403          * Indexes are stored in big endian, no need to switch before sending.
1404          */
1405         DBG("Sending viewer index for stream %" PRIu64 " offset %" PRIu64,
1406                 rstream->stream_handle,
1407                 be64toh(packet_index.offset));
1408         viewer_index.offset = packet_index.offset;
1409         viewer_index.packet_size = packet_index.packet_size;
1410         viewer_index.content_size = packet_index.content_size;
1411         viewer_index.timestamp_begin = packet_index.timestamp_begin;
1412         viewer_index.timestamp_end = packet_index.timestamp_end;
1413         viewer_index.events_discarded = packet_index.events_discarded;
1414         viewer_index.stream_id = packet_index.stream_id;
1415
1416 send_reply:
1417         if (rstream) {
1418                 pthread_mutex_unlock(&rstream->lock);
1419         }
1420
1421         if (metadata_viewer_stream) {
1422                 pthread_mutex_lock(&metadata_viewer_stream->stream->lock);
1423                 DBG("get next index metadata check: recv %" PRIu64
1424                                 " sent %" PRIu64,
1425                         metadata_viewer_stream->stream->metadata_received,
1426                         metadata_viewer_stream->metadata_sent);
1427                 if (!metadata_viewer_stream->stream->metadata_received ||
1428                                 metadata_viewer_stream->stream->metadata_received >
1429                                         metadata_viewer_stream->metadata_sent) {
1430                         viewer_index.flags |= LTTNG_VIEWER_FLAG_NEW_METADATA;
1431                 }
1432                 pthread_mutex_unlock(&metadata_viewer_stream->stream->lock);
1433         }
1434
1435         viewer_index.flags = htobe32(viewer_index.flags);
1436         health_code_update();
1437
1438         ret = send_response(conn->sock, &viewer_index, sizeof(viewer_index));
1439         if (ret < 0) {
1440                 goto end;
1441         }
1442         health_code_update();
1443
1444         DBG("Index %" PRIu64 " for stream %" PRIu64 " sent",
1445                         vstream->last_sent_index,
1446                         vstream->stream->stream_handle);
1447 end:
1448         if (metadata_viewer_stream) {
1449                 viewer_stream_put(metadata_viewer_stream);
1450         }
1451         if (vstream) {
1452                 viewer_stream_put(vstream);
1453         }
1454         return ret;
1455
1456 error_put:
1457         pthread_mutex_unlock(&rstream->lock);
1458         if (metadata_viewer_stream) {
1459                 viewer_stream_put(metadata_viewer_stream);
1460         }
1461         viewer_stream_put(vstream);
1462         return ret;
1463 }
1464
1465 /*
1466  * Send the next index for a stream
1467  *
1468  * Return 0 on success or else a negative value.
1469  */
1470 static
1471 int viewer_get_packet(struct relay_connection *conn)
1472 {
1473         int ret, send_data = 0;
1474         char *data = NULL;
1475         uint32_t len = 0;
1476         ssize_t read_len;
1477         struct lttng_viewer_get_packet get_packet_info;
1478         struct lttng_viewer_trace_packet reply;
1479         struct relay_viewer_stream *vstream = NULL;
1480         struct ctf_trace *ctf_trace;
1481         struct relay_viewer_stream *metadata_viewer_stream = NULL;
1482
1483         DBG2("Relay get data packet");
1484
1485         health_code_update();
1486
1487         ret = recv_request(conn->sock, &get_packet_info,
1488                         sizeof(get_packet_info));
1489         if (ret < 0) {
1490                 goto end;
1491         }
1492         health_code_update();
1493
1494         /* From this point on, the error label can be reached. */
1495         memset(&reply, 0, sizeof(reply));
1496
1497         vstream = viewer_stream_get_by_id(be64toh(get_packet_info.stream_id));
1498         if (!vstream) {
1499                 reply.status = htobe32(LTTNG_VIEWER_GET_PACKET_ERR);
1500                 goto send_reply_nolock;
1501         }
1502
1503         ctf_trace = vstream->stream->trace;
1504
1505         /* metadata_viewer_stream may be NULL. */
1506         metadata_viewer_stream =
1507                         ctf_trace_get_viewer_metadata_stream(ctf_trace);
1508
1509         if (metadata_viewer_stream) {
1510                 bool get_packet_err = false;
1511
1512                 pthread_mutex_lock(&metadata_viewer_stream->stream->lock);
1513                 DBG("get packet metadata check: recv %" PRIu64 " sent %" PRIu64,
1514                         metadata_viewer_stream->stream->metadata_received,
1515                         metadata_viewer_stream->metadata_sent);
1516                 if (!metadata_viewer_stream->stream->metadata_received ||
1517                                 metadata_viewer_stream->stream->metadata_received >
1518                                         metadata_viewer_stream->metadata_sent) {
1519                         /*
1520                          * We prevent the client from reading a data stream as
1521                          * long as there is metadata left to consume. This
1522                          * ensures that the client won't receive data of which
1523                          * it can't make sense.
1524                          */
1525                         get_packet_err = true;
1526                 }
1527                 pthread_mutex_unlock(&metadata_viewer_stream->stream->lock);
1528                 viewer_stream_put(metadata_viewer_stream);
1529                 if (get_packet_err) {
1530                         reply.status = htobe32(LTTNG_VIEWER_GET_PACKET_ERR);
1531                         reply.flags |= LTTNG_VIEWER_FLAG_NEW_METADATA;
1532                         goto send_reply_nolock;
1533                 }
1534         }
1535
1536         pthread_mutex_lock(&vstream->stream->lock);
1537
1538         len = be32toh(get_packet_info.len);
1539         data = zmalloc(len);
1540         if (!data) {
1541                 PERROR("relay data zmalloc");
1542                 goto error;
1543         }
1544
1545         ret = lseek(vstream->stream_fd->fd, be64toh(get_packet_info.offset),
1546                         SEEK_SET);
1547         if (ret < 0) {
1548                 PERROR("lseek fd %d to offset %" PRIu64, vstream->stream_fd->fd,
1549                         be64toh(get_packet_info.offset));
1550                 goto error;
1551         }
1552         read_len = lttng_read(vstream->stream_fd->fd, data, len);
1553         if (read_len < len) {
1554                 PERROR("Relay reading trace file, fd: %d, offset: %" PRIu64,
1555                                 vstream->stream_fd->fd,
1556                                 be64toh(get_packet_info.offset));
1557                 goto error;
1558         }
1559         reply.status = htobe32(LTTNG_VIEWER_GET_PACKET_OK);
1560         reply.len = htobe32(len);
1561         send_data = 1;
1562         goto send_reply;
1563
1564 error:
1565         reply.status = htobe32(LTTNG_VIEWER_GET_PACKET_ERR);
1566
1567 send_reply:
1568         if (vstream) {
1569                 pthread_mutex_unlock(&vstream->stream->lock);
1570         }
1571 send_reply_nolock:
1572         reply.flags = htobe32(reply.flags);
1573
1574         health_code_update();
1575
1576         ret = send_response(conn->sock, &reply, sizeof(reply));
1577         if (ret < 0) {
1578                 goto end_free;
1579         }
1580         health_code_update();
1581
1582         if (send_data) {
1583                 health_code_update();
1584                 ret = send_response(conn->sock, data, len);
1585                 if (ret < 0) {
1586                         goto end_free;
1587                 }
1588                 health_code_update();
1589         }
1590
1591         DBG("Sent %u bytes for stream %" PRIu64, len,
1592                         be64toh(get_packet_info.stream_id));
1593
1594 end_free:
1595         free(data);
1596 end:
1597         if (vstream) {
1598                 viewer_stream_put(vstream);
1599         }
1600         return ret;
1601 }
1602
1603 /*
1604  * Send the session's metadata
1605  *
1606  * Return 0 on success else a negative value.
1607  */
1608 static
1609 int viewer_get_metadata(struct relay_connection *conn)
1610 {
1611         int ret = 0;
1612         ssize_t read_len;
1613         uint64_t len = 0;
1614         char *data = NULL;
1615         struct lttng_viewer_get_metadata request;
1616         struct lttng_viewer_metadata_packet reply;
1617         struct relay_viewer_stream *vstream = NULL;
1618
1619         assert(conn);
1620
1621         DBG("Relay get metadata");
1622
1623         health_code_update();
1624
1625         ret = recv_request(conn->sock, &request, sizeof(request));
1626         if (ret < 0) {
1627                 goto end;
1628         }
1629         health_code_update();
1630
1631         memset(&reply, 0, sizeof(reply));
1632
1633         vstream = viewer_stream_get_by_id(be64toh(request.stream_id));
1634         if (!vstream) {
1635                 /*
1636                  * The metadata stream can be closed by a CLOSE command
1637                  * just before we attach. It can also be closed by
1638                  * per-pid tracing during tracing. Therefore, it is
1639                  * possible that we cannot find this viewer stream.
1640                  * Reply back to the client with an error if we cannot
1641                  * find it.
1642                  */
1643                 reply.status = htobe32(LTTNG_VIEWER_METADATA_ERR);
1644                 goto send_reply;
1645         }
1646         pthread_mutex_lock(&vstream->stream->lock);
1647         if (!vstream->stream->is_metadata) {
1648                 ERR("Invalid metadata stream");
1649                 goto error;
1650         }
1651
1652         assert(vstream->metadata_sent <= vstream->stream->metadata_received);
1653
1654         len = vstream->stream->metadata_received - vstream->metadata_sent;
1655         if (len == 0) {
1656                 reply.status = htobe32(LTTNG_VIEWER_NO_NEW_METADATA);
1657                 goto send_reply;
1658         }
1659
1660         /* first time, we open the metadata file */
1661         if (!vstream->stream_fd) {
1662                 char fullpath[PATH_MAX];
1663
1664                 ret = snprintf(fullpath, PATH_MAX, "%s/%s", vstream->path_name,
1665                                 vstream->channel_name);
1666                 if (ret < 0) {
1667                         goto error;
1668                 }
1669                 ret = open(fullpath, O_RDONLY);
1670                 if (ret < 0) {
1671                         PERROR("Relay opening metadata file");
1672                         goto error;
1673                 }
1674                 vstream->stream_fd = stream_fd_create(ret);
1675                 if (!vstream->stream_fd) {
1676                         if (close(ret)) {
1677                                 PERROR("close");
1678                         }
1679                         goto error;
1680                 }
1681         }
1682
1683         reply.len = htobe64(len);
1684         data = zmalloc(len);
1685         if (!data) {
1686                 PERROR("viewer metadata zmalloc");
1687                 goto error;
1688         }
1689
1690         read_len = lttng_read(vstream->stream_fd->fd, data, len);
1691         if (read_len < len) {
1692                 PERROR("Relay reading metadata file");
1693                 goto error;
1694         }
1695         vstream->metadata_sent += read_len;
1696         if (vstream->metadata_sent == vstream->stream->metadata_received
1697                         && vstream->stream->closed) {
1698                 /* Release ownership for the viewer metadata stream. */
1699                 viewer_stream_put(vstream);
1700         }
1701
1702         reply.status = htobe32(LTTNG_VIEWER_METADATA_OK);
1703
1704         goto send_reply;
1705
1706 error:
1707         reply.status = htobe32(LTTNG_VIEWER_METADATA_ERR);
1708
1709 send_reply:
1710         health_code_update();
1711         if (vstream) {
1712                 pthread_mutex_unlock(&vstream->stream->lock);
1713         }
1714         ret = send_response(conn->sock, &reply, sizeof(reply));
1715         if (ret < 0) {
1716                 goto end_free;
1717         }
1718         health_code_update();
1719
1720         if (len > 0) {
1721                 ret = send_response(conn->sock, data, len);
1722                 if (ret < 0) {
1723                         goto end_free;
1724                 }
1725         }
1726
1727         DBG("Sent %" PRIu64 " bytes of metadata for stream %" PRIu64, len,
1728                         be64toh(request.stream_id));
1729
1730         DBG("Metadata sent");
1731
1732 end_free:
1733         free(data);
1734 end:
1735         if (vstream) {
1736                 viewer_stream_put(vstream);
1737         }
1738         return ret;
1739 }
1740
1741 /*
1742  * Create a viewer session.
1743  *
1744  * Return 0 on success or else a negative value.
1745  */
1746 static
1747 int viewer_create_session(struct relay_connection *conn)
1748 {
1749         int ret;
1750         struct lttng_viewer_create_session_response resp;
1751
1752         DBG("Viewer create session received");
1753
1754         memset(&resp, 0, sizeof(resp));
1755         resp.status = htobe32(LTTNG_VIEWER_CREATE_SESSION_OK);
1756         conn->viewer_session = viewer_session_create();
1757         if (!conn->viewer_session) {
1758                 ERR("Allocation viewer session");
1759                 resp.status = htobe32(LTTNG_VIEWER_CREATE_SESSION_ERR);
1760                 goto send_reply;
1761         }
1762
1763 send_reply:
1764         health_code_update();
1765         ret = send_response(conn->sock, &resp, sizeof(resp));
1766         if (ret < 0) {
1767                 goto end;
1768         }
1769         health_code_update();
1770         ret = 0;
1771
1772 end:
1773         return ret;
1774 }
1775
1776
1777 /*
1778  * live_relay_unknown_command: send -1 if received unknown command
1779  */
1780 static
1781 void live_relay_unknown_command(struct relay_connection *conn)
1782 {
1783         struct lttcomm_relayd_generic_reply reply;
1784
1785         memset(&reply, 0, sizeof(reply));
1786         reply.ret_code = htobe32(LTTNG_ERR_UNK);
1787         (void) send_response(conn->sock, &reply, sizeof(reply));
1788 }
1789
1790 /*
1791  * Process the commands received on the control socket
1792  */
1793 static
1794 int process_control(struct lttng_viewer_cmd *recv_hdr,
1795                 struct relay_connection *conn)
1796 {
1797         int ret = 0;
1798         uint32_t msg_value;
1799
1800         msg_value = be32toh(recv_hdr->cmd);
1801
1802         /*
1803          * Make sure we've done the version check before any command other then a
1804          * new client connection.
1805          */
1806         if (msg_value != LTTNG_VIEWER_CONNECT && !conn->version_check_done) {
1807                 ERR("Viewer conn value %" PRIu32 " before version check", msg_value);
1808                 ret = -1;
1809                 goto end;
1810         }
1811
1812         switch (msg_value) {
1813         case LTTNG_VIEWER_CONNECT:
1814                 ret = viewer_connect(conn);
1815                 break;
1816         case LTTNG_VIEWER_LIST_SESSIONS:
1817                 ret = viewer_list_sessions(conn);
1818                 break;
1819         case LTTNG_VIEWER_ATTACH_SESSION:
1820                 ret = viewer_attach_session(conn);
1821                 break;
1822         case LTTNG_VIEWER_GET_NEXT_INDEX:
1823                 ret = viewer_get_next_index(conn);
1824                 break;
1825         case LTTNG_VIEWER_GET_PACKET:
1826                 ret = viewer_get_packet(conn);
1827                 break;
1828         case LTTNG_VIEWER_GET_METADATA:
1829                 ret = viewer_get_metadata(conn);
1830                 break;
1831         case LTTNG_VIEWER_GET_NEW_STREAMS:
1832                 ret = viewer_get_new_streams(conn);
1833                 break;
1834         case LTTNG_VIEWER_CREATE_SESSION:
1835                 ret = viewer_create_session(conn);
1836                 break;
1837         default:
1838                 ERR("Received unknown viewer command (%u)",
1839                                 be32toh(recv_hdr->cmd));
1840                 live_relay_unknown_command(conn);
1841                 ret = -1;
1842                 goto end;
1843         }
1844
1845 end:
1846         return ret;
1847 }
1848
1849 static
1850 void cleanup_connection_pollfd(struct lttng_poll_event *events, int pollfd)
1851 {
1852         int ret;
1853
1854         (void) lttng_poll_del(events, pollfd);
1855
1856         ret = close(pollfd);
1857         if (ret < 0) {
1858                 ERR("Closing pollfd %d", pollfd);
1859         }
1860 }
1861
1862 /*
1863  * This thread does the actual work
1864  */
1865 static
1866 void *thread_worker(void *data)
1867 {
1868         int ret, err = -1;
1869         uint32_t nb_fd;
1870         struct lttng_poll_event events;
1871         struct lttng_ht *viewer_connections_ht;
1872         struct lttng_ht_iter iter;
1873         struct lttng_viewer_cmd recv_hdr;
1874         struct relay_connection *destroy_conn;
1875
1876         DBG("[thread] Live viewer relay worker started");
1877
1878         rcu_register_thread();
1879
1880         health_register(health_relayd, HEALTH_RELAYD_TYPE_LIVE_WORKER);
1881
1882         if (testpoint(relayd_thread_live_worker)) {
1883                 goto error_testpoint;
1884         }
1885
1886         /* table of connections indexed on socket */
1887         viewer_connections_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
1888         if (!viewer_connections_ht) {
1889                 goto viewer_connections_ht_error;
1890         }
1891
1892         ret = create_thread_poll_set(&events, 2);
1893         if (ret < 0) {
1894                 goto error_poll_create;
1895         }
1896
1897         ret = lttng_poll_add(&events, live_conn_pipe[0], LPOLLIN | LPOLLRDHUP);
1898         if (ret < 0) {
1899                 goto error;
1900         }
1901
1902 restart:
1903         while (1) {
1904                 int i;
1905
1906                 health_code_update();
1907
1908                 /* Infinite blocking call, waiting for transmission */
1909                 DBG3("Relayd live viewer worker thread polling...");
1910                 health_poll_entry();
1911                 ret = lttng_poll_wait(&events, -1);
1912                 health_poll_exit();
1913                 if (ret < 0) {
1914                         /*
1915                          * Restart interrupted system call.
1916                          */
1917                         if (errno == EINTR) {
1918                                 goto restart;
1919                         }
1920                         goto error;
1921                 }
1922
1923                 nb_fd = ret;
1924
1925                 /*
1926                  * Process control. The control connection is prioritised so we don't
1927                  * starve it with high throughput tracing data on the data
1928                  * connection.
1929                  */
1930                 for (i = 0; i < nb_fd; i++) {
1931                         /* Fetch once the poll data */
1932                         uint32_t revents = LTTNG_POLL_GETEV(&events, i);
1933                         int pollfd = LTTNG_POLL_GETFD(&events, i);
1934
1935                         health_code_update();
1936
1937                         if (!revents) {
1938                                 /* No activity for this FD (poll implementation). */
1939                                 continue;
1940                         }
1941
1942                         /* Thread quit pipe has been closed. Killing thread. */
1943                         ret = check_thread_quit_pipe(pollfd, revents);
1944                         if (ret) {
1945                                 err = 0;
1946                                 goto exit;
1947                         }
1948
1949                         /* Inspect the relay conn pipe for new connection. */
1950                         if (pollfd == live_conn_pipe[0]) {
1951                                 if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
1952                                         ERR("Relay live pipe error");
1953                                         goto error;
1954                                 } else if (revents & LPOLLIN) {
1955                                         struct relay_connection *conn;
1956
1957                                         ret = lttng_read(live_conn_pipe[0],
1958                                                         &conn, sizeof(conn));
1959                                         if (ret < 0) {
1960                                                 goto error;
1961                                         }
1962                                         lttng_poll_add(&events, conn->sock->fd,
1963                                                         LPOLLIN | LPOLLRDHUP);
1964                                         connection_ht_add(viewer_connections_ht, conn);
1965                                         DBG("Connection socket %d added to poll", conn->sock->fd);
1966                                 }
1967                         } else {
1968                                 /* Connection activity. */
1969                                 struct relay_connection *conn;
1970
1971                                 conn = connection_get_by_sock(viewer_connections_ht, pollfd);
1972                                 if (!conn) {
1973                                         continue;
1974                                 }
1975
1976                                 if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
1977                                         cleanup_connection_pollfd(&events, pollfd);
1978                                         /* Put "create" ownership reference. */
1979                                         connection_put(conn);
1980                                 } else if (revents & LPOLLIN) {
1981                                         ret = conn->sock->ops->recvmsg(conn->sock, &recv_hdr,
1982                                                         sizeof(recv_hdr), 0);
1983                                         if (ret <= 0) {
1984                                                 /* Connection closed. */
1985                                                 cleanup_connection_pollfd(&events, pollfd);
1986                                                 /* Put "create" ownership reference. */
1987                                                 connection_put(conn);
1988                                                 DBG("Viewer control conn closed with %d", pollfd);
1989                                         } else {
1990                                                 ret = process_control(&recv_hdr, conn);
1991                                                 if (ret < 0) {
1992                                                         /* Clear the session on error. */
1993                                                         cleanup_connection_pollfd(&events, pollfd);
1994                                                         /* Put "create" ownership reference. */
1995                                                         connection_put(conn);
1996                                                         DBG("Viewer connection closed with %d", pollfd);
1997                                                 }
1998                                         }
1999                                 }
2000                                 /* Put local "get_by_sock" reference. */
2001                                 connection_put(conn);
2002                         }
2003                 }
2004         }
2005
2006 exit:
2007 error:
2008         lttng_poll_clean(&events);
2009
2010         /* Cleanup reamaining connection object. */
2011         rcu_read_lock();
2012         cds_lfht_for_each_entry(viewer_connections_ht->ht, &iter.iter,
2013                         destroy_conn,
2014                         sock_n.node) {
2015                 health_code_update();
2016                 connection_put(destroy_conn);
2017         }
2018         rcu_read_unlock();
2019 error_poll_create:
2020         lttng_ht_destroy(viewer_connections_ht);
2021 viewer_connections_ht_error:
2022         /* Close relay conn pipes */
2023         utils_close_pipe(live_conn_pipe);
2024         if (err) {
2025                 DBG("Viewer worker thread exited with error");
2026         }
2027         DBG("Viewer worker thread cleanup complete");
2028 error_testpoint:
2029         if (err) {
2030                 health_error();
2031                 ERR("Health error occurred in %s", __func__);
2032         }
2033         health_unregister(health_relayd);
2034         if (lttng_relay_stop_threads()) {
2035                 ERR("Error stopping threads");
2036         }
2037         rcu_unregister_thread();
2038         return NULL;
2039 }
2040
2041 /*
2042  * Create the relay command pipe to wake thread_manage_apps.
2043  * Closed in cleanup().
2044  */
2045 static int create_conn_pipe(void)
2046 {
2047         return utils_create_pipe_cloexec(live_conn_pipe);
2048 }
2049
2050 int relayd_live_join(void)
2051 {
2052         int ret, retval = 0;
2053         void *status;
2054
2055         ret = pthread_join(live_listener_thread, &status);
2056         if (ret) {
2057                 errno = ret;
2058                 PERROR("pthread_join live listener");
2059                 retval = -1;
2060         }
2061
2062         ret = pthread_join(live_worker_thread, &status);
2063         if (ret) {
2064                 errno = ret;
2065                 PERROR("pthread_join live worker");
2066                 retval = -1;
2067         }
2068
2069         ret = pthread_join(live_dispatcher_thread, &status);
2070         if (ret) {
2071                 errno = ret;
2072                 PERROR("pthread_join live dispatcher");
2073                 retval = -1;
2074         }
2075
2076         cleanup_relayd_live();
2077
2078         return retval;
2079 }
2080
2081 /*
2082  * main
2083  */
2084 int relayd_live_create(struct lttng_uri *uri)
2085 {
2086         int ret = 0, retval = 0;
2087         void *status;
2088         int is_root;
2089
2090         if (!uri) {
2091                 retval = -1;
2092                 goto exit_init_data;
2093         }
2094         live_uri = uri;
2095
2096         /* Check if daemon is UID = 0 */
2097         is_root = !getuid();
2098
2099         if (!is_root) {
2100                 if (live_uri->port < 1024) {
2101                         ERR("Need to be root to use ports < 1024");
2102                         retval = -1;
2103                         goto exit_init_data;
2104                 }
2105         }
2106
2107         /* Setup the thread apps communication pipe. */
2108         if (create_conn_pipe()) {
2109                 retval = -1;
2110                 goto exit_init_data;
2111         }
2112
2113         /* Init relay command queue. */
2114         cds_wfcq_init(&viewer_conn_queue.head, &viewer_conn_queue.tail);
2115
2116         /* Set up max poll set size */
2117         lttng_poll_set_max_size();
2118
2119         /* Setup the dispatcher thread */
2120         ret = pthread_create(&live_dispatcher_thread, NULL,
2121                         thread_dispatcher, (void *) NULL);
2122         if (ret) {
2123                 errno = ret;
2124                 PERROR("pthread_create viewer dispatcher");
2125                 retval = -1;
2126                 goto exit_dispatcher_thread;
2127         }
2128
2129         /* Setup the worker thread */
2130         ret = pthread_create(&live_worker_thread, NULL,
2131                         thread_worker, NULL);
2132         if (ret) {
2133                 errno = ret;
2134                 PERROR("pthread_create viewer worker");
2135                 retval = -1;
2136                 goto exit_worker_thread;
2137         }
2138
2139         /* Setup the listener thread */
2140         ret = pthread_create(&live_listener_thread, NULL,
2141                         thread_listener, (void *) NULL);
2142         if (ret) {
2143                 errno = ret;
2144                 PERROR("pthread_create viewer listener");
2145                 retval = -1;
2146                 goto exit_listener_thread;
2147         }
2148
2149         /*
2150          * All OK, started all threads.
2151          */
2152         return retval;
2153
2154         /*
2155          * Join on the live_listener_thread should anything be added after
2156          * the live_listener thread's creation.
2157          */
2158
2159 exit_listener_thread:
2160
2161         ret = pthread_join(live_worker_thread, &status);
2162         if (ret) {
2163                 errno = ret;
2164                 PERROR("pthread_join live worker");
2165                 retval = -1;
2166         }
2167 exit_worker_thread:
2168
2169         ret = pthread_join(live_dispatcher_thread, &status);
2170         if (ret) {
2171                 errno = ret;
2172                 PERROR("pthread_join live dispatcher");
2173                 retval = -1;
2174         }
2175 exit_dispatcher_thread:
2176
2177 exit_init_data:
2178         cleanup_relayd_live();
2179
2180         return retval;
2181 }