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