src/bin/lttng-relayd/live.c

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