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