src/bin/lttng-sessiond/rotation-thread.cpp

   1 /*
   2  * Copyright (C) 2017 Julien Desfossez <jdesfossez@efficios.com>
   3  * Copyright (C) 2018 Jérémie Galarneau <jeremie.galarneau@efficios.com>
   4  *
   5  * SPDX-License-Identifier: GPL-2.0-only
   6  *
   7  */
   8
   9 #define _LGPL_SOURCE
  10 #include "cmd.hpp"
  11 #include "health-sessiond.hpp"
  12 #include "lttng-sessiond.hpp"
  13 #include "notification-thread-commands.hpp"
  14 #include "rotation-thread.hpp"
  15 #include "session.hpp"
  16 #include "thread.hpp"
  17 #include "timer.hpp"
  18 #include "utils.hpp"
  19
  20 #include <common/align.hpp>
  21 #include <common/config/session-config.hpp>
  22 #include <common/defaults.hpp>
  23 #include <common/error.hpp>
  24 #include <common/eventfd.hpp>
  25 #include <common/exception.hpp>
  26 #include <common/file-descriptor.hpp>
  27 #include <common/format.hpp>
  28 #include <common/futex.hpp>
  29 #include <common/hashtable/utils.hpp>
  30 #include <common/kernel-ctl/kernel-ctl.hpp>
  31 #include <common/locked-reference.hpp>
  32 #include <common/make-unique-wrapper.hpp>
  33 #include <common/pthread-lock.hpp>
  34 #include <common/scope-exit.hpp>
  35 #include <common/time.hpp>
  36 #include <common/urcu.hpp>
  37 #include <common/utils.hpp>
  38
  39 #include <lttng/action/action-internal.hpp>
  40 #include <lttng/condition/condition-internal.hpp>
  41 #include <lttng/location-internal.hpp>
  42 #include <lttng/notification/channel-internal.hpp>
  43 #include <lttng/notification/notification-internal.hpp>
  44 #include <lttng/rotate-internal.hpp>
  45 #include <lttng/trigger/trigger.h>
  46
  47 #include <inttypes.h>
  48 #include <memory>
  49 #include <signal.h>
  50 #include <sys/eventfd.h>
  51 #include <sys/stat.h>
  52 #include <time.h>
  53 #include <urcu.h>
  54 #include <urcu/list.h>
  55
  56 namespace ls = lttng::sessiond;
  57
  58 /*
  59  * The timer thread enqueues jobs and wakes up the rotation thread.
  60  * When the rotation thread wakes up, it empties the queue.
  61  */
  62 struct ls::rotation_thread_timer_queue {
  63         struct lttng_pipe *event_pipe;
  64         struct cds_list_head list;
  65         pthread_mutex_t lock;
  66 };
  67
  68 namespace {
  69 struct rotation_thread_job {
  70         using uptr = std::unique_ptr<
  71                 rotation_thread_job,
  72                 lttng::details::create_unique_class<rotation_thread_job, lttng::free>>;
  73
  74         enum ls::rotation_thread_job_type type;
  75         struct ltt_session *session;
  76         /* List member in struct rotation_thread_timer_queue. */
  77         struct cds_list_head head;
  78 };
  79
  80 const char *get_job_type_str(enum ls::rotation_thread_job_type job_type)
  81 {
  82         switch (job_type) {
  83         case ls::rotation_thread_job_type::CHECK_PENDING_ROTATION:
  84                 return "CHECK_PENDING_ROTATION";
  85         case ls::rotation_thread_job_type::SCHEDULED_ROTATION:
  86                 return "SCHEDULED_ROTATION";
  87         default:
  88                 abort();
  89         }
  90 }
  91
  92 /*
  93  * Called with the rotation_thread_timer_queue lock held.
  94  * Return true if the same timer job already exists in the queue, false if not.
  95  */
  96 bool timer_job_exists(const ls::rotation_thread_timer_queue *queue,
  97                       ls::rotation_thread_job_type job_type,
  98                       ltt_session *session)
  99 {
 100         bool exists = false;
 101         struct rotation_thread_job *job;
 102
 103         cds_list_for_each_entry (job, &queue->list, head) {
 104                 if (job->session == session && job->type == job_type) {
 105                         exists = true;
 106                         goto end;
 107                 }
 108         }
 109 end:
 110         return exists;
 111 }
 112
 113 void check_session_rotation_pending_on_consumers(ltt_session& session, bool& _rotation_completed)
 114 {
 115         int ret = 0;
 116         struct consumer_socket *socket;
 117         struct cds_lfht_iter iter;
 118         enum consumer_trace_chunk_exists_status exists_status;
 119         uint64_t relayd_id;
 120         bool chunk_exists_on_peer = false;
 121         enum lttng_trace_chunk_status chunk_status;
 122         lttng::urcu::read_lock_guard read_lock;
 123
 124         LTTNG_ASSERT(session.chunk_being_archived);
 125
 126         /*
 127          * Check for a local pending rotation on all consumers (32-bit
 128          * user space, 64-bit user space, and kernel).
 129          */
 130         if (!session.ust_session) {
 131                 goto skip_ust;
 132         }
 133
 134         cds_lfht_for_each_entry (
 135                 session.ust_session->consumer->socks->ht, &iter, socket, node.node) {
 136                 relayd_id = session.ust_session->consumer->type == CONSUMER_DST_LOCAL ?
 137                         -1ULL :
 138                         session.ust_session->consumer->net_seq_index;
 139
 140                 lttng::pthread::lock_guard socket_lock(*socket->lock);
 141                 ret = consumer_trace_chunk_exists(socket,
 142                                                   relayd_id,
 143                                                   session.id,
 144                                                   session.chunk_being_archived,
 145                                                   &exists_status);
 146                 if (ret) {
 147                         ERR("Error occurred while checking rotation status on consumer daemon");
 148                         goto end;
 149                 }
 150
 151                 if (exists_status != CONSUMER_TRACE_CHUNK_EXISTS_STATUS_UNKNOWN_CHUNK) {
 152                         chunk_exists_on_peer = true;
 153                         goto end;
 154                 }
 155         }
 156
 157 skip_ust:
 158         if (!session.kernel_session) {
 159                 goto skip_kernel;
 160         }
 161
 162         cds_lfht_for_each_entry (
 163                 session.kernel_session->consumer->socks->ht, &iter, socket, node.node) {
 164                 lttng::pthread::lock_guard socket_lock(*socket->lock);
 165
 166                 relayd_id = session.kernel_session->consumer->type == CONSUMER_DST_LOCAL ?
 167                         -1ULL :
 168                         session.kernel_session->consumer->net_seq_index;
 169
 170                 ret = consumer_trace_chunk_exists(socket,
 171                                                   relayd_id,
 172                                                   session.id,
 173                                                   session.chunk_being_archived,
 174                                                   &exists_status);
 175                 if (ret) {
 176                         ERR("Error occurred while checking rotation status on consumer daemon");
 177                         goto end;
 178                 }
 179
 180                 if (exists_status != CONSUMER_TRACE_CHUNK_EXISTS_STATUS_UNKNOWN_CHUNK) {
 181                         chunk_exists_on_peer = true;
 182                         goto end;
 183                 }
 184         }
 185 skip_kernel:
 186 end:
 187
 188         if (!chunk_exists_on_peer) {
 189                 uint64_t chunk_being_archived_id;
 190
 191                 chunk_status = lttng_trace_chunk_get_id(session.chunk_being_archived,
 192                                                         &chunk_being_archived_id);
 193                 LTTNG_ASSERT(chunk_status == LTTNG_TRACE_CHUNK_STATUS_OK);
 194                 DBG("Rotation of trace archive %" PRIu64
 195                     " of session \"%s\" is complete on all consumers",
 196                     chunk_being_archived_id,
 197                     session.name);
 198         }
 199
 200         _rotation_completed = !chunk_exists_on_peer;
 201         if (ret) {
 202                 ret = session_reset_rotation_state(session, LTTNG_ROTATION_STATE_ERROR);
 203                 if (ret) {
 204                         ERR("Failed to reset rotation state of session \"%s\"", session.name);
 205                 }
 206         }
 207 }
 208
 209 /*
 210  * Check if the last rotation was completed, called with session lock held.
 211  * Should only return non-zero in the event of a fatal error. Doing so will
 212  * shutdown the thread.
 213  */
 214 int check_session_rotation_pending(ltt_session& session,
 215                                    notification_thread_handle& notification_thread_handle)
 216 {
 217         int ret;
 218         struct lttng_trace_archive_location *location;
 219         enum lttng_trace_chunk_status chunk_status;
 220         bool rotation_completed = false;
 221         const char *archived_chunk_name;
 222         uint64_t chunk_being_archived_id;
 223
 224         if (!session.chunk_being_archived) {
 225                 ret = 0;
 226                 goto end;
 227         }
 228
 229         chunk_status =
 230                 lttng_trace_chunk_get_id(session.chunk_being_archived, &chunk_being_archived_id);
 231         LTTNG_ASSERT(chunk_status == LTTNG_TRACE_CHUNK_STATUS_OK);
 232
 233         DBG("Checking for pending rotation on session \"%s\", trace archive %" PRIu64,
 234             session.name,
 235             chunk_being_archived_id);
 236
 237         /*
 238          * The rotation-pending check timer of a session is launched in
 239          * one-shot mode. If the rotation is incomplete, the rotation
 240          * thread will re-enable the pending-check timer.
 241          *
 242          * The timer thread can't stop the timer itself since it is involved
 243          * in the check for the timer's quiescence.
 244          */
 245         ret = timer_session_rotation_pending_check_stop(session);
 246         if (ret) {
 247                 goto check_ongoing_rotation;
 248         }
 249
 250         check_session_rotation_pending_on_consumers(session, rotation_completed);
 251         if (!rotation_completed || session.rotation_state == LTTNG_ROTATION_STATE_ERROR) {
 252                 goto check_ongoing_rotation;
 253         }
 254
 255         /*
 256          * Now we can clear the "ONGOING" state in the session. New
 257          * rotations can start now.
 258          */
 259         chunk_status = lttng_trace_chunk_get_name(
 260                 session.chunk_being_archived, &archived_chunk_name, nullptr);
 261         LTTNG_ASSERT(chunk_status == LTTNG_TRACE_CHUNK_STATUS_OK);
 262         free(session.last_archived_chunk_name);
 263         session.last_archived_chunk_name = strdup(archived_chunk_name);
 264         if (!session.last_archived_chunk_name) {
 265                 PERROR("Failed to duplicate archived chunk name");
 266         }
 267
 268         session_reset_rotation_state(session, LTTNG_ROTATION_STATE_COMPLETED);
 269
 270         if (!session.quiet_rotation) {
 271                 location = session_get_trace_archive_location(&session);
 272                 ret = notification_thread_command_session_rotation_completed(
 273                         &notification_thread_handle,
 274                         session.id,
 275                         session.last_archived_chunk_id.value,
 276                         location);
 277                 lttng_trace_archive_location_put(location);
 278                 if (ret != LTTNG_OK) {
 279                         ERR("Failed to notify notification thread of completed rotation for session %s",
 280                             session.name);
 281                 }
 282         }
 283
 284         ret = 0;
 285 check_ongoing_rotation:
 286         if (session.rotation_state == LTTNG_ROTATION_STATE_ONGOING) {
 287                 chunk_status = lttng_trace_chunk_get_id(session.chunk_being_archived,
 288                                                         &chunk_being_archived_id);
 289                 LTTNG_ASSERT(chunk_status == LTTNG_TRACE_CHUNK_STATUS_OK);
 290
 291                 DBG("Rotation of trace archive %" PRIu64 " is still pending for session %s",
 292                     chunk_being_archived_id,
 293                     session.name);
 294                 ret = timer_session_rotation_pending_check_start(&session,
 295                                                                  DEFAULT_ROTATE_PENDING_TIMER);
 296                 if (ret) {
 297                         ERR("Failed to re-enable rotation pending timer");
 298                         ret = -1;
 299                         goto end;
 300                 }
 301         }
 302
 303 end:
 304         return ret;
 305 }
 306
 307 /* Call with the session and session_list locks held. */
 308 int launch_session_rotation(ltt_session& session)
 309 {
 310         int ret;
 311         struct lttng_rotate_session_return rotation_return;
 312
 313         DBG("Launching scheduled time-based rotation on session \"%s\"", session.name);
 314
 315         ASSERT_SESSION_LIST_LOCKED();
 316         ASSERT_LOCKED(session.lock);
 317
 318         ret = cmd_rotate_session(&session,
 319                                  &rotation_return,
 320                                  false,
 321                                  LTTNG_TRACE_CHUNK_COMMAND_TYPE_MOVE_TO_COMPLETED);
 322         if (ret == LTTNG_OK) {
 323                 DBG("Scheduled time-based rotation successfully launched on session \"%s\"",
 324                     session.name);
 325         } else {
 326                 /* Don't consider errors as fatal. */
 327                 DBG("Scheduled time-based rotation aborted for session %s: %s",
 328                     session.name,
 329                     lttng_strerror(ret));
 330         }
 331
 332         return 0;
 333 }
 334
 335 int run_job(const rotation_thread_job& job,
 336             ltt_session& session,
 337             notification_thread_handle& notification_thread_handle)
 338 {
 339         int ret;
 340
 341         switch (job.type) {
 342         case ls::rotation_thread_job_type::SCHEDULED_ROTATION:
 343                 ret = launch_session_rotation(session);
 344                 break;
 345         case ls::rotation_thread_job_type::CHECK_PENDING_ROTATION:
 346                 ret = check_session_rotation_pending(session, notification_thread_handle);
 347                 break;
 348         default:
 349                 abort();
 350         }
 351
 352         return ret;
 353 }
 354
 355 bool shutdown_rotation_thread(void *thread_data)
 356 {
 357         auto *handle = reinterpret_cast<const ls::rotation_thread *>(thread_data);
 358
 359         return handle->shutdown();
 360 }
 361 } /* namespace */
 362
 363 ls::rotation_thread_timer_queue *ls::rotation_thread_timer_queue_create()
 364 {
 365         auto queue = zmalloc<ls::rotation_thread_timer_queue>();
 366         if (!queue) {
 367                 PERROR("Failed to allocate timer rotate queue");
 368                 goto end;
 369         }
 370
 371         queue->event_pipe = lttng_pipe_open(FD_CLOEXEC | O_NONBLOCK);
 372         CDS_INIT_LIST_HEAD(&queue->list);
 373         pthread_mutex_init(&queue->lock, nullptr);
 374 end:
 375         return queue;
 376 }
 377
 378 void ls::rotation_thread_timer_queue_destroy(struct rotation_thread_timer_queue *queue)
 379 {
 380         if (!queue) {
 381                 return;
 382         }
 383
 384         lttng_pipe_destroy(queue->event_pipe);
 385
 386         {
 387                 lttng::pthread::lock_guard queue_lock(queue->lock);
 388
 389                 LTTNG_ASSERT(cds_list_empty(&queue->list));
 390         }
 391
 392         pthread_mutex_destroy(&queue->lock);
 393         free(queue);
 394 }
 395
 396 ls::rotation_thread::rotation_thread(
 397         rotation_thread_timer_queue& rotation_timer_queue,
 398         notification_thread_handle& notification_thread_handle) :
 399         _rotation_timer_queue{ rotation_timer_queue },
 400         _notification_thread_handle{ notification_thread_handle }
 401 {
 402         _quit_pipe.reset([]() {
 403                 auto raw_pipe = lttng_pipe_open(FD_CLOEXEC);
 404                 if (!raw_pipe) {
 405                         LTTNG_THROW_POSIX("Failed to rotation thread's quit pipe", errno);
 406                 }
 407
 408                 return raw_pipe;
 409         }());
 410
 411         _notification_channel.reset([]() {
 412                 auto channel = lttng_notification_channel_create(
 413                         lttng_session_daemon_notification_endpoint);
 414                 if (!channel) {
 415                         LTTNG_THROW_ERROR(
 416                                 "Failed to create notification channel of rotation thread");
 417                 }
 418
 419                 return channel;
 420         }());
 421
 422         lttng_poll_init(&_events);
 423
 424         /*
 425          * Create pollset with size 4:
 426          *      - rotation thread quit pipe,
 427          *      - rotation thread timer queue pipe,
 428          *      - notification channel sock,
 429          *      - subscribtion change event fd
 430          */
 431         if (lttng_poll_create(&_events, 4, LTTNG_CLOEXEC) < 0) {
 432                 LTTNG_THROW_ERROR("Failed to create poll object for rotation thread");
 433         }
 434
 435         if (lttng_poll_add(&_events, lttng_pipe_get_readfd(_quit_pipe.get()), LPOLLIN) < 0) {
 436                 LTTNG_THROW_ERROR("Failed to add quit pipe read fd to poll set");
 437         }
 438
 439         if (lttng_poll_add(&_events,
 440                            lttng_pipe_get_readfd(_rotation_timer_queue.event_pipe),
 441                            LPOLLIN) < 0) {
 442                 LTTNG_THROW_ERROR("Failed to add rotation timer queue event pipe fd to poll set");
 443         }
 444
 445         if (lttng_poll_add(&_events,
 446                            _notification_channel_subscribtion_change_eventfd.fd(),
 447                            LPOLLIN) < 0) {
 448                 LTTNG_THROW_ERROR(
 449                         "Failed to add rotation thread notification channel subscription change eventfd to poll set");
 450         }
 451
 452         if (lttng_poll_add(&_events, _notification_channel->socket, LPOLLIN) < 0) {
 453                 LTTNG_THROW_ERROR("Failed to add notification channel socket fd to pollset");
 454         }
 455 }
 456
 457 ls::rotation_thread::~rotation_thread()
 458 {
 459         lttng_poll_clean(&_events);
 460 }
 461
 462 void ls::rotation_thread_enqueue_job(ls::rotation_thread_timer_queue *queue,
 463                                  ls::rotation_thread_job_type job_type,
 464                                  ltt_session *session)
 465 {
 466         const char dummy = '!';
 467         struct rotation_thread_job *job = nullptr;
 468         const char *job_type_str = get_job_type_str(job_type);
 469         lttng::pthread::lock_guard queue_lock(queue->lock);
 470
 471         if (timer_job_exists(queue, job_type, session)) {
 472                 /*
 473                  * This timer job is already pending, we don't need to add
 474                  * it.
 475                  */
 476                 return;
 477         }
 478
 479         job = zmalloc<rotation_thread_job>();
 480         if (!job) {
 481                 PERROR("Failed to allocate rotation thread job of type \"%s\" for session \"%s\"",
 482                        job_type_str,
 483                        session->name);
 484                 return;
 485         }
 486
 487         /* No reason for this to fail as the caller must hold a reference. */
 488         (void) session_get(session);
 489
 490         job->session = session;
 491         job->type = job_type;
 492         cds_list_add_tail(&job->head, &queue->list);
 493
 494         const int write_ret =
 495                 lttng_write(lttng_pipe_get_writefd(queue->event_pipe), &dummy, sizeof(dummy));
 496         if (write_ret < 0) {
 497                 /*
 498                  * We do not want to block in the timer handler, the job has
 499                  * been enqueued in the list, the wakeup pipe is probably full,
 500                  * the job will be processed when the rotation_thread catches
 501                  * up.
 502                  */
 503                 DIAGNOSTIC_PUSH
 504                 DIAGNOSTIC_IGNORE_LOGICAL_OP
 505                 if (errno == EAGAIN || errno == EWOULDBLOCK) {
 506                         DIAGNOSTIC_POP
 507                         /*
 508                          * Not an error, but would be surprising and indicate
 509                          * that the rotation thread can't keep up with the
 510                          * current load.
 511                          */
 512                         DBG("Wake-up pipe of rotation thread job queue is full");
 513                         return;
 514                 }
 515
 516                 PERROR("Failed to wake-up the rotation thread after pushing a job of type \"%s\" for session \"%s\"",
 517                        job_type_str,
 518                        session->name);
 519                 return;
 520         }
 521 }
 522
 523 void ls::rotation_thread::_handle_job_queue()
 524 {
 525         for (;;) {
 526                 rotation_thread_job::uptr job;
 527
 528                 {
 529                         /* Take the queue lock only to pop an element from the list. */
 530                         lttng::pthread::lock_guard rotation_timer_queue_lock(
 531                                 _rotation_timer_queue.lock);
 532                         if (cds_list_empty(&_rotation_timer_queue.list)) {
 533                                 break;
 534                         }
 535
 536                         job.reset(cds_list_first_entry(
 537                                 &_rotation_timer_queue.list, typeof(rotation_thread_job), head));
 538                         cds_list_del(&job->head);
 539                 }
 540
 541                 session_lock_list();
 542                 const auto unlock_list = lttng::make_scope_exit([]() noexcept { session_unlock_list(); });
 543
 544                 /* locked_ptr will unlock the session and release the ref held by the job. */
 545                 session_lock(job->session);
 546                 auto session = ltt_session::locked_ptr(job->session);
 547
 548                 if (run_job(*job, *session, _notification_thread_handle)) {
 549                         return;
 550                 }
 551         }
 552 }
 553
 554 void ls::rotation_thread::_handle_notification(const lttng_notification &notification)
 555 {
 556         int ret = 0;
 557         const char *condition_session_name = nullptr;
 558         enum lttng_condition_status condition_status;
 559         enum lttng_evaluation_status evaluation_status;
 560         uint64_t consumed;
 561         auto *condition = lttng_notification_get_const_condition(&notification);
 562         auto *evaluation = lttng_notification_get_const_evaluation(&notification);
 563         const auto condition_type = lttng_condition_get_type(condition);
 564
 565         if (condition_type != LTTNG_CONDITION_TYPE_SESSION_CONSUMED_SIZE) {
 566                 LTTNG_THROW_ERROR("Unexpected condition type");
 567         }
 568
 569         /* Fetch info to test. */
 570         condition_status = lttng_condition_session_consumed_size_get_session_name(
 571                 condition, &condition_session_name);
 572         if (condition_status != LTTNG_CONDITION_STATUS_OK) {
 573                 LTTNG_THROW_ERROR("Session name could not be fetched from notification");
 574         }
 575
 576         evaluation_status =
 577                 lttng_evaluation_session_consumed_size_get_consumed_size(evaluation, &consumed);
 578         if (evaluation_status != LTTNG_EVALUATION_STATUS_OK) {
 579                 LTTNG_THROW_ERROR("Failed to get consumed size from evaluation");
 580         }
 581
 582         DBG_FMT("Handling session consumed size condition: session_name=`{}`, consumed_size={}",
 583                 condition_session_name,
 584                 consumed);
 585
 586         session_lock_list();
 587         const auto unlock_list = lttng::make_scope_exit([]() noexcept { session_unlock_list(); });
 588
 589         ltt_session::locked_ptr session{ [&condition_session_name]() {
 590                 auto raw_session_ptr = session_find_by_name(condition_session_name);
 591
 592                 if (raw_session_ptr) {
 593                         session_lock(raw_session_ptr);
 594                 }
 595
 596                 return raw_session_ptr;
 597         }() };
 598         if (!session) {
 599                 DBG_FMT("Failed to find session while handling notification: notification_type={}, session name=`{}`",
 600                         lttng_condition_type_str(condition_type),
 601                         condition_session_name);
 602                 /*
 603                  * Not a fatal error: a session can be destroyed before we get
 604                  * the chance to handle the notification.
 605                  */
 606                 return;
 607         }
 608
 609         if (!lttng_trigger_is_equal(session->rotate_trigger,
 610                                     lttng_notification_get_const_trigger(&notification))) {
 611                 DBG("Notification does not originate from the internal size-based scheduled rotation trigger, skipping");
 612                 return;
 613         }
 614
 615         unsubscribe_session_consumed_size_rotation(*session);
 616
 617         ret = cmd_rotate_session(
 618                 session.get(), nullptr, false, LTTNG_TRACE_CHUNK_COMMAND_TYPE_MOVE_TO_COMPLETED);
 619         switch (ret) {
 620         case LTTNG_OK:
 621                 break;
 622         case -LTTNG_ERR_ROTATION_PENDING:
 623                 DBG("Rotate already pending, subscribe to the next threshold value");
 624                 break;
 625         case -LTTNG_ERR_ROTATION_MULTIPLE_AFTER_STOP:
 626                 DBG("Rotation already happened since last stop, subscribe to the next threshold value");
 627                 break;
 628         case -LTTNG_ERR_ROTATION_AFTER_STOP_CLEAR:
 629                 DBG("Rotation already happened since last stop and clear, subscribe to the next threshold value");
 630                 break;
 631         default:
 632                 LTTNG_THROW_CTL("Failed to rotate on consumed size notification",
 633                                 static_cast<lttng_error_code>(-ret));
 634         }
 635
 636         subscribe_session_consumed_size_rotation(*session, consumed + session->rotate_size);
 637 }
 638
 639 void ls::rotation_thread::_handle_notification_channel_activity()
 640 {
 641         bool notification_pending = true;
 642
 643         /*
 644          * A notification channel may have multiple notifications queued-up internally in
 645          * its buffers. This is because a notification channel multiplexes command replies
 646          * and notifications. The current protocol specifies that multiple notifications can be
 647          * received before the reply to a command.
 648          *
 649          * In such cases, the notification channel client implementation internally queues them and
 650          * provides them on the next calls to lttng_notification_channel_get_next_notification().
 651          * This is correct with respect to the public API, which is intended to be used in "blocking
 652          * mode".
 653          *
 654          * However, this internal user relies on poll/epoll to wake-up when data is available
 655          * on the notification channel's socket. As such, it can't assume that a wake-up means only
 656          * one notification is available for consumption since many of them may have been queued in
 657          * the channel's internal buffers.
 658          */
 659         while (notification_pending) {
 660                 const auto pending_status = lttng_notification_channel_has_pending_notification(
 661                         _notification_channel.get(), &notification_pending);
 662                 if (pending_status != LTTNG_NOTIFICATION_CHANNEL_STATUS_OK) {
 663                         LTTNG_THROW_ERROR("Error occurred while checking for pending notification");
 664                 }
 665
 666                 if (!notification_pending) {
 667                         return;
 668                 }
 669
 670                 /* Receive the next notification. */
 671                 lttng_notification::uptr notification;
 672                 enum lttng_notification_channel_status next_notification_status;
 673
 674                 {
 675                         struct lttng_notification *raw_notification_ptr;
 676
 677                         next_notification_status = lttng_notification_channel_get_next_notification(
 678                                 _notification_channel.get(), &raw_notification_ptr);
 679                         notification.reset(raw_notification_ptr);
 680                 }
 681
 682                 switch (next_notification_status) {
 683                 case LTTNG_NOTIFICATION_CHANNEL_STATUS_OK:
 684                         break;
 685                 case LTTNG_NOTIFICATION_CHANNEL_STATUS_NOTIFICATIONS_DROPPED:
 686                         WARN("Dropped notification detected on notification channel used by the rotation management thread.");
 687                         return;
 688                 case LTTNG_NOTIFICATION_CHANNEL_STATUS_CLOSED:
 689                         LTTNG_THROW_ERROR("Notification channel was closed");
 690                 default:
 691                         /* Unhandled conditions / errors. */
 692                         LTTNG_THROW_ERROR("Unknown notification channel status");
 693                 }
 694
 695                 _handle_notification(*notification);
 696         }
 697 }
 698
 699 void ls::rotation_thread::_thread_function() noexcept
 700 {
 701         DBG("Started rotation thread");
 702
 703         try {
 704                 _run();
 705         } catch (const std::exception& e) {
 706                 ERR_FMT("Fatal rotation thread error: {}", e.what());
 707         }
 708
 709         DBG("Thread exit");
 710 }
 711
 712 void ls::rotation_thread::_run()
 713 {
 714         rcu_register_thread();
 715         const auto unregister_rcu_thread =
 716                 lttng::make_scope_exit([]() noexcept { rcu_unregister_thread(); });
 717
 718         rcu_thread_online();
 719         const auto offline_rcu_thread =
 720                 lttng::make_scope_exit([]() noexcept { rcu_thread_offline(); });
 721
 722         health_register(the_health_sessiond, HEALTH_SESSIOND_TYPE_ROTATION);
 723         health_code_update();
 724         const auto unregister_health =
 725                 lttng::make_scope_exit([]() noexcept { health_unregister(the_health_sessiond); });
 726
 727         const auto queue_pipe_fd = lttng_pipe_get_readfd(_rotation_timer_queue.event_pipe);
 728
 729         while (true) {
 730                 health_poll_entry();
 731                 DBG("Entering poll wait");
 732                 auto poll_wait_ret = lttng_poll_wait(&_events, -1);
 733                 DBG_FMT("Poll wait returned: ret={}", poll_wait_ret);
 734                 health_poll_exit();
 735                 if (poll_wait_ret < 0) {
 736                         /*
 737                          * Restart interrupted system call.
 738                          */
 739                         if (errno == EINTR) {
 740                                 continue;
 741                         }
 742
 743                         LTTNG_THROW_POSIX("Error encountered during lttng_poll_wait", errno);
 744                 }
 745
 746                 const auto fd_count = poll_wait_ret;
 747                 for (int i = 0; i < fd_count; i++) {
 748                         const auto fd = LTTNG_POLL_GETFD(&_events, i);
 749                         const auto revents = LTTNG_POLL_GETEV(&_events, i);
 750
 751                         DBG_FMT("Handling descriptor activity: fd={}, events={:b}", fd, revents);
 752
 753                         if (revents & LPOLLERR) {
 754                                 LTTNG_THROW_ERROR(
 755                                         fmt::format("Polling returned an error on fd: fd={}", fd));
 756                         }
 757
 758                         if (fd == _notification_channel->socket ||
 759                             fd == _notification_channel_subscribtion_change_eventfd.fd()) {
 760                                 try {
 761                                         _handle_notification_channel_activity();
 762                                 } catch (const lttng::ctl::error& e) {
 763                                         /*
 764                                          * The only non-fatal error (rotation failed), others
 765                                          * are caught at the top-level.
 766                                          */
 767                                         DBG_FMT("Control error occurred while handling activity on notification channel socket: {}",
 768                                                 e.what());
 769                                         continue;
 770                                 }
 771
 772                                 if (fd == _notification_channel_subscribtion_change_eventfd.fd()) {
 773                                         _notification_channel_subscribtion_change_eventfd.decrement();
 774                                 }
 775                         } else {
 776                                 /* Job queue or quit pipe activity. */
 777
 778                                 /*
 779                                  * The job queue is serviced if there is
 780                                  * activity on the quit pipe to ensure it is
 781                                  * flushed and all references held in the queue
 782                                  * are released.
 783                                  */
 784                                 _handle_job_queue();
 785                                 if (fd == queue_pipe_fd) {
 786                                         char buf;
 787
 788                                         if (lttng_read(fd, &buf, 1) != 1) {
 789                                                 LTTNG_THROW_POSIX(
 790                                                         fmt::format(
 791                                                                 "Failed to read from wakeup pipe: fd={}",
 792                                                                 fd),
 793                                                         errno);
 794                                         }
 795                                 } else {
 796                                         DBG("Quit pipe activity");
 797                                         return;
 798                                 }
 799                         }
 800                 }
 801         }
 802 }
 803
 804 bool ls::rotation_thread::shutdown() const noexcept
 805 {
 806         const int write_fd = lttng_pipe_get_writefd(_quit_pipe.get());
 807
 808         return notify_thread_pipe(write_fd) == 1;
 809 }
 810
 811 void ls::rotation_thread::launch_thread()
 812 {
 813         auto thread = lttng_thread_create(
 814                 "Rotation",
 815                 [](void *ptr) {
 816                         auto handle = reinterpret_cast<rotation_thread *>(ptr);
 817
 818                         handle->_thread_function();
 819                         return static_cast<void *>(nullptr);
 820                 },
 821                 shutdown_rotation_thread,
 822                 nullptr,
 823                 this);
 824         if (!thread) {
 825                 LTTNG_THROW_ERROR("Failed to launch rotation thread");
 826         }
 827
 828         lttng_thread_put(thread);
 829 }
 830
 831 void ls::rotation_thread::subscribe_session_consumed_size_rotation(ltt_session& session,
 832                                                                           std::uint64_t size)
 833 {
 834         const struct lttng_credentials session_creds = {
 835                 .uid = LTTNG_OPTIONAL_INIT_VALUE(session.uid),
 836                 .gid = LTTNG_OPTIONAL_INIT_VALUE(session.gid),
 837         };
 838
 839         ASSERT_LOCKED(session.lock);
 840
 841         auto rotate_condition = lttng::make_unique_wrapper<lttng_condition, lttng_condition_put>(
 842                 lttng_condition_session_consumed_size_create());
 843         if (!rotate_condition) {
 844                 LTTNG_THROW_POSIX("Failed to create session consumed size condition object", errno);
 845         }
 846
 847         auto condition_status =
 848                 lttng_condition_session_consumed_size_set_threshold(rotate_condition.get(), size);
 849         if (condition_status != LTTNG_CONDITION_STATUS_OK) {
 850                 LTTNG_THROW_ERROR(fmt::format(
 851                         "Could not set session consumed size condition threshold: size={}", size));
 852         }
 853
 854         condition_status = lttng_condition_session_consumed_size_set_session_name(rotate_condition.get(),
 855                                                                                   session.name);
 856         if (condition_status != LTTNG_CONDITION_STATUS_OK) {
 857                 LTTNG_THROW_ERROR(fmt::format(
 858                         "Could not set session consumed size condition session name: name=`{}`",
 859                         session.name));
 860         }
 861
 862         auto notify_action = lttng::make_unique_wrapper<lttng_action, lttng_action_put>(
 863                 lttng_action_notify_create());
 864         if (!notify_action) {
 865                 LTTNG_THROW_POSIX("Could not create notify action", errno);
 866         }
 867
 868         LTTNG_ASSERT(!session.rotate_trigger);
 869         /* trigger acquires its own reference to condition and action on success. */
 870         auto trigger = lttng::make_unique_wrapper<lttng_trigger, lttng_trigger_put>(
 871                 lttng_trigger_create(rotate_condition.get(), notify_action.get()));
 872         if (!trigger)
 873         {
 874                 LTTNG_THROW_POSIX("Could not create size-based rotation trigger", errno);
 875         }
 876
 877         /* Ensure this trigger is not visible to external users. */
 878         lttng_trigger_set_hidden(trigger.get());
 879         lttng_trigger_set_credentials(trigger.get(), &session_creds);
 880
 881         auto nc_status =
 882                 lttng_notification_channel_subscribe(_notification_channel.get(), rotate_condition.get());
 883         if (nc_status != LTTNG_NOTIFICATION_CHANNEL_STATUS_OK) {
 884                 LTTNG_THROW_ERROR("Could not subscribe to session consumed size notification");
 885         }
 886
 887         /*
 888          * Ensure any notification queued during the subscription are consumed by queueing an
 889          * event.
 890          */
 891         _notification_channel_subscribtion_change_eventfd.increment();
 892
 893         const auto register_ret = notification_thread_command_register_trigger(
 894                 &_notification_thread_handle, trigger.get(), true);
 895         if (register_ret != LTTNG_OK) {
 896                 LTTNG_THROW_CTL(
 897                         fmt::format(
 898                                 "Failed to register trigger for automatic size-based rotation: session_name{}, size={}",
 899                                 session.name,
 900                                 size),
 901                         register_ret);
 902         }
 903
 904         /* Ownership transferred to the session. */
 905         session.rotate_trigger = trigger.release();
 906 }
 907
 908 void ls::rotation_thread::unsubscribe_session_consumed_size_rotation(ltt_session& session)
 909 {
 910         LTTNG_ASSERT(session.rotate_trigger);
 911
 912         const auto remove_session_trigger = lttng::make_scope_exit([&session]() noexcept {
 913                 lttng_trigger_put(session.rotate_trigger);
 914                 session.rotate_trigger = nullptr;
 915         });
 916
 917         const auto unsubscribe_status = lttng_notification_channel_unsubscribe(
 918                 _notification_channel.get(),
 919                 lttng_trigger_get_const_condition(session.rotate_trigger));
 920         if (unsubscribe_status != LTTNG_NOTIFICATION_CHANNEL_STATUS_OK) {
 921                 LTTNG_THROW_ERROR(fmt::format(
 922                         "Failed to unsubscribe from consumed size condition used to control automatic size-based rotations: session_name=`{}` return_code={}",
 923                         session.name,
 924                         static_cast<int>(unsubscribe_status)));
 925         }
 926
 927         /*
 928          * Ensure any notification queued during the un-subscription are consumed by queueing an
 929          * event.
 930          */
 931         _notification_channel_subscribtion_change_eventfd.increment();
 932
 933         const auto unregister_status = notification_thread_command_unregister_trigger(
 934                 &_notification_thread_handle, session.rotate_trigger);
 935         if (unregister_status != LTTNG_OK) {
 936                 LTTNG_THROW_CTL(
 937                         fmt::format(
 938                                 "Failed to unregister trigger for automatic size-based rotation: session_name{}",
 939                                 session.name),
 940                         unregister_status);
 941         }
 942 }