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 <lttng/trigger/trigger.h>
  11 #include <common/error.hpp>
  12 #include <common/config/session-config.hpp>
  13 #include <common/defaults.hpp>
  14 #include <common/utils.hpp>
  15 #include <common/futex.hpp>
  16 #include <common/align.hpp>
  17 #include <common/time.hpp>
  18 #include <common/hashtable/utils.hpp>
  19 #include <sys/stat.h>
  20 #include <time.h>
  21 #include <signal.h>
  22 #include <inttypes.h>
  23
  24 #include <common/kernel-ctl/kernel-ctl.hpp>
  25 #include <lttng/notification/channel-internal.hpp>
  26 #include <lttng/rotate-internal.hpp>
  27 #include <lttng/location-internal.hpp>
  28 #include <lttng/condition/condition-internal.hpp>
  29 #include <lttng/notification/notification-internal.hpp>
  30
  31 #include "rotation-thread.hpp"
  32 #include "lttng-sessiond.hpp"
  33 #include "health-sessiond.hpp"
  34 #include "rotate.hpp"
  35 #include "cmd.hpp"
  36 #include "session.hpp"
  37 #include "timer.hpp"
  38 #include "notification-thread-commands.hpp"
  39 #include "utils.hpp"
  40 #include "thread.hpp"
  41
  42 #include <urcu.h>
  43 #include <urcu/list.h>
  44
  45 struct lttng_notification_channel *rotate_notification_channel = NULL;
  46
  47 struct rotation_thread {
  48         struct lttng_poll_event events;
  49 };
  50
  51 /*
  52  * The timer thread enqueues jobs and wakes up the rotation thread.
  53  * When the rotation thread wakes up, it empties the queue.
  54  */
  55 struct rotation_thread_timer_queue {
  56         struct lttng_pipe *event_pipe;
  57         struct cds_list_head list;
  58         pthread_mutex_t lock;
  59 };
  60
  61 struct rotation_thread_handle {
  62         struct rotation_thread_timer_queue *rotation_timer_queue;
  63         /* Access to the notification thread cmd_queue */
  64         struct notification_thread_handle *notification_thread_handle;
  65         /* Thread-specific quit pipe. */
  66         struct lttng_pipe *quit_pipe;
  67 };
  68
  69 namespace {
  70 struct rotation_thread_job {
  71         enum rotation_thread_job_type type;
  72         struct ltt_session *session;
  73         /* List member in struct rotation_thread_timer_queue. */
  74         struct cds_list_head head;
  75 };
  76 } /* namespace */
  77
  78 static
  79 const char *get_job_type_str(enum rotation_thread_job_type job_type)
  80 {
  81         switch (job_type) {
  82         case ROTATION_THREAD_JOB_TYPE_CHECK_PENDING_ROTATION:
  83                 return "CHECK_PENDING_ROTATION";
  84         case ROTATION_THREAD_JOB_TYPE_SCHEDULED_ROTATION:
  85                 return "SCHEDULED_ROTATION";
  86         default:
  87                 abort();
  88         }
  89 }
  90
  91 struct rotation_thread_timer_queue *rotation_thread_timer_queue_create(void)
  92 {
  93         struct rotation_thread_timer_queue *queue = NULL;
  94
  95         queue = zmalloc<rotation_thread_timer_queue>();
  96         if (!queue) {
  97                 PERROR("Failed to allocate timer rotate queue");
  98                 goto end;
  99         }
 100
 101         queue->event_pipe = lttng_pipe_open(FD_CLOEXEC | O_NONBLOCK);
 102         CDS_INIT_LIST_HEAD(&queue->list);
 103         pthread_mutex_init(&queue->lock, NULL);
 104 end:
 105         return queue;
 106 }
 107
 108 void rotation_thread_timer_queue_destroy(
 109                 struct rotation_thread_timer_queue *queue)
 110 {
 111         if (!queue) {
 112                 return;
 113         }
 114
 115         lttng_pipe_destroy(queue->event_pipe);
 116
 117         pthread_mutex_lock(&queue->lock);
 118         LTTNG_ASSERT(cds_list_empty(&queue->list));
 119         pthread_mutex_unlock(&queue->lock);
 120         pthread_mutex_destroy(&queue->lock);
 121         free(queue);
 122 }
 123
 124 /*
 125  * Destroy the thread data previously created by the init function.
 126  */
 127 void rotation_thread_handle_destroy(
 128                 struct rotation_thread_handle *handle)
 129 {
 130         lttng_pipe_destroy(handle->quit_pipe);
 131         free(handle);
 132 }
 133
 134 struct rotation_thread_handle *rotation_thread_handle_create(
 135                 struct rotation_thread_timer_queue *rotation_timer_queue,
 136                 struct notification_thread_handle *notification_thread_handle)
 137 {
 138         struct rotation_thread_handle *handle;
 139
 140         handle = zmalloc<rotation_thread_handle>();
 141         if (!handle) {
 142                 goto end;
 143         }
 144
 145         handle->rotation_timer_queue = rotation_timer_queue;
 146         handle->notification_thread_handle = notification_thread_handle;
 147         handle->quit_pipe = lttng_pipe_open(FD_CLOEXEC);
 148         if (!handle->quit_pipe) {
 149                 goto error;
 150         }
 151
 152 end:
 153         return handle;
 154 error:
 155         rotation_thread_handle_destroy(handle);
 156         return NULL;
 157 }
 158
 159 /*
 160  * Called with the rotation_thread_timer_queue lock held.
 161  * Return true if the same timer job already exists in the queue, false if not.
 162  */
 163 static
 164 bool timer_job_exists(const struct rotation_thread_timer_queue *queue,
 165                 enum rotation_thread_job_type job_type,
 166                 struct ltt_session *session)
 167 {
 168         bool exists = false;
 169         struct rotation_thread_job *job;
 170
 171         cds_list_for_each_entry(job, &queue->list, head) {
 172                 if (job->session == session && job->type == job_type) {
 173                         exists = true;
 174                         goto end;
 175                 }
 176         }
 177 end:
 178         return exists;
 179 }
 180
 181 void rotation_thread_enqueue_job(struct rotation_thread_timer_queue *queue,
 182                 enum rotation_thread_job_type job_type,
 183                 struct ltt_session *session)
 184 {
 185         int ret;
 186         const char dummy = '!';
 187         struct rotation_thread_job *job = NULL;
 188         const char *job_type_str = get_job_type_str(job_type);
 189
 190         pthread_mutex_lock(&queue->lock);
 191         if (timer_job_exists(queue, job_type, session)) {
 192                 /*
 193                  * This timer job is already pending, we don't need to add
 194                  * it.
 195                  */
 196                 goto end;
 197         }
 198
 199         job = zmalloc<rotation_thread_job>();
 200         if (!job) {
 201                 PERROR("Failed to allocate rotation thread job of type \"%s\" for session \"%s\"",
 202                                 job_type_str, session->name);
 203                 goto end;
 204         }
 205         /* No reason for this to fail as the caller must hold a reference. */
 206         (void) session_get(session);
 207
 208         job->session = session;
 209         job->type = job_type;
 210         cds_list_add_tail(&job->head, &queue->list);
 211
 212         ret = lttng_write(lttng_pipe_get_writefd(queue->event_pipe), &dummy,
 213                         sizeof(dummy));
 214         if (ret < 0) {
 215                 /*
 216                  * We do not want to block in the timer handler, the job has
 217                  * been enqueued in the list, the wakeup pipe is probably full,
 218                  * the job will be processed when the rotation_thread catches
 219                  * up.
 220                  */
 221                 DIAGNOSTIC_PUSH
 222                 DIAGNOSTIC_IGNORE_LOGICAL_OP
 223                 if (errno == EAGAIN || errno == EWOULDBLOCK) {
 224                 DIAGNOSTIC_POP
 225                         /*
 226                          * Not an error, but would be surprising and indicate
 227                          * that the rotation thread can't keep up with the
 228                          * current load.
 229                          */
 230                         DBG("Wake-up pipe of rotation thread job queue is full");
 231                         goto end;
 232                 }
 233                 PERROR("Failed to wake-up the rotation thread after pushing a job of type \"%s\" for session \"%s\"",
 234                                 job_type_str, session->name);
 235                 goto end;
 236         }
 237
 238 end:
 239         pthread_mutex_unlock(&queue->lock);
 240 }
 241
 242 static
 243 int init_poll_set(struct lttng_poll_event *poll_set,
 244                 struct rotation_thread_handle *handle)
 245 {
 246         int ret;
 247
 248         /*
 249          * Create pollset with size 3:
 250          *      - rotation thread quit pipe,
 251          *      - rotation thread timer queue pipe,
 252          *      - notification channel sock,
 253          */
 254         ret = lttng_poll_create(poll_set, 5, LTTNG_CLOEXEC);
 255         if (ret < 0) {
 256                 goto error;
 257         }
 258
 259         ret = lttng_poll_add(poll_set,
 260                         lttng_pipe_get_readfd(handle->quit_pipe),
 261                         LPOLLIN | LPOLLERR);
 262         if (ret < 0) {
 263                 ERR("Failed to add quit pipe read fd to poll set");
 264                 goto error;
 265         }
 266
 267         ret = lttng_poll_add(poll_set,
 268                         lttng_pipe_get_readfd(handle->rotation_timer_queue->event_pipe),
 269                         LPOLLIN | LPOLLERR);
 270         if (ret < 0) {
 271                 ERR("Failed to add rotate_pending fd to poll set");
 272                 goto error;
 273         }
 274
 275         return ret;
 276 error:
 277         lttng_poll_clean(poll_set);
 278         return ret;
 279 }
 280
 281 static
 282 void fini_thread_state(struct rotation_thread *state)
 283 {
 284         lttng_poll_clean(&state->events);
 285         if (rotate_notification_channel) {
 286                 lttng_notification_channel_destroy(rotate_notification_channel);
 287         }
 288 }
 289
 290 static
 291 int init_thread_state(struct rotation_thread_handle *handle,
 292                 struct rotation_thread *state)
 293 {
 294         int ret;
 295
 296         memset(state, 0, sizeof(*state));
 297         lttng_poll_init(&state->events);
 298
 299         ret = init_poll_set(&state->events, handle);
 300         if (ret) {
 301                 ERR("Failed to initialize rotation thread poll set");
 302                 goto end;
 303         }
 304
 305         rotate_notification_channel = lttng_notification_channel_create(
 306                         lttng_session_daemon_notification_endpoint);
 307         if (!rotate_notification_channel) {
 308                 ERR("Could not create notification channel");
 309                 ret = -1;
 310                 goto end;
 311         }
 312         ret = lttng_poll_add(&state->events, rotate_notification_channel->socket,
 313                         LPOLLIN | LPOLLERR);
 314         if (ret < 0) {
 315                 ERR("Failed to add notification fd to pollset");
 316                 goto end;
 317         }
 318
 319 end:
 320         return ret;
 321 }
 322
 323 static
 324 void check_session_rotation_pending_on_consumers(struct ltt_session *session,
 325                 bool *_rotation_completed)
 326 {
 327         int ret = 0;
 328         struct consumer_socket *socket;
 329         struct cds_lfht_iter iter;
 330         enum consumer_trace_chunk_exists_status exists_status;
 331         uint64_t relayd_id;
 332         bool chunk_exists_on_peer = false;
 333         enum lttng_trace_chunk_status chunk_status;
 334
 335         LTTNG_ASSERT(session->chunk_being_archived);
 336
 337         /*
 338          * Check for a local pending rotation on all consumers (32-bit
 339          * user space, 64-bit user space, and kernel).
 340          */
 341         rcu_read_lock();
 342         if (!session->ust_session) {
 343                 goto skip_ust;
 344         }
 345         cds_lfht_for_each_entry(session->ust_session->consumer->socks->ht,
 346                         &iter, socket, node.node) {
 347                 relayd_id = session->ust_session->consumer->type == CONSUMER_DST_LOCAL ?
 348                                 -1ULL :
 349                                 session->ust_session->consumer->net_seq_index;
 350
 351                 pthread_mutex_lock(socket->lock);
 352                 ret = consumer_trace_chunk_exists(socket,
 353                                 relayd_id,
 354                                 session->id, session->chunk_being_archived,
 355                                 &exists_status);
 356                 if (ret) {
 357                         pthread_mutex_unlock(socket->lock);
 358                         ERR("Error occurred while checking rotation status on consumer daemon");
 359                         goto end;
 360                 }
 361
 362                 if (exists_status != CONSUMER_TRACE_CHUNK_EXISTS_STATUS_UNKNOWN_CHUNK) {
 363                         pthread_mutex_unlock(socket->lock);
 364                         chunk_exists_on_peer = true;
 365                         goto end;
 366                 }
 367                 pthread_mutex_unlock(socket->lock);
 368         }
 369
 370 skip_ust:
 371         if (!session->kernel_session) {
 372                 goto skip_kernel;
 373         }
 374         cds_lfht_for_each_entry(session->kernel_session->consumer->socks->ht,
 375                                 &iter, socket, node.node) {
 376                 pthread_mutex_lock(socket->lock);
 377                 relayd_id = session->kernel_session->consumer->type == CONSUMER_DST_LOCAL ?
 378                                 -1ULL :
 379                                 session->kernel_session->consumer->net_seq_index;
 380
 381                 ret = consumer_trace_chunk_exists(socket,
 382                                 relayd_id,
 383                                 session->id, session->chunk_being_archived,
 384                                 &exists_status);
 385                 if (ret) {
 386                         pthread_mutex_unlock(socket->lock);
 387                         ERR("Error occurred while checking rotation status on consumer daemon");
 388                         goto end;
 389                 }
 390
 391                 if (exists_status != CONSUMER_TRACE_CHUNK_EXISTS_STATUS_UNKNOWN_CHUNK) {
 392                         pthread_mutex_unlock(socket->lock);
 393                         chunk_exists_on_peer = true;
 394                         goto end;
 395                 }
 396                 pthread_mutex_unlock(socket->lock);
 397         }
 398 skip_kernel:
 399 end:
 400         rcu_read_unlock();
 401
 402         if (!chunk_exists_on_peer) {
 403                 uint64_t chunk_being_archived_id;
 404
 405                 chunk_status = lttng_trace_chunk_get_id(
 406                                 session->chunk_being_archived,
 407                                 &chunk_being_archived_id);
 408                 LTTNG_ASSERT(chunk_status == LTTNG_TRACE_CHUNK_STATUS_OK);
 409                 DBG("Rotation of trace archive %" PRIu64 " of session \"%s\" is complete on all consumers",
 410                                 chunk_being_archived_id,
 411                                 session->name);
 412         }
 413         *_rotation_completed = !chunk_exists_on_peer;
 414         if (ret) {
 415                 ret = session_reset_rotation_state(session,
 416                                 LTTNG_ROTATION_STATE_ERROR);
 417                 if (ret) {
 418                         ERR("Failed to reset rotation state of session \"%s\"",
 419                                         session->name);
 420                 }
 421         }
 422 }
 423
 424 /*
 425  * Check if the last rotation was completed, called with session lock held.
 426  * Should only return non-zero in the event of a fatal error. Doing so will
 427  * shutdown the thread.
 428  */
 429 static
 430 int check_session_rotation_pending(struct ltt_session *session,
 431                 struct notification_thread_handle *notification_thread_handle)
 432 {
 433         int ret;
 434         struct lttng_trace_archive_location *location;
 435         enum lttng_trace_chunk_status chunk_status;
 436         bool rotation_completed = false;
 437         const char *archived_chunk_name;
 438         uint64_t chunk_being_archived_id;
 439
 440         if (!session->chunk_being_archived) {
 441                 ret = 0;
 442                 goto end;
 443         }
 444
 445         chunk_status = lttng_trace_chunk_get_id(session->chunk_being_archived,
 446                         &chunk_being_archived_id);
 447         LTTNG_ASSERT(chunk_status == LTTNG_TRACE_CHUNK_STATUS_OK);
 448
 449         DBG("Checking for pending rotation on session \"%s\", trace archive %" PRIu64,
 450                         session->name, chunk_being_archived_id);
 451
 452         /*
 453          * The rotation-pending check timer of a session is launched in
 454          * one-shot mode. If the rotation is incomplete, the rotation
 455          * thread will re-enable the pending-check timer.
 456          *
 457          * The timer thread can't stop the timer itself since it is involved
 458          * in the check for the timer's quiescence.
 459          */
 460         ret = timer_session_rotation_pending_check_stop(session);
 461         if (ret) {
 462                 goto check_ongoing_rotation;
 463         }
 464
 465         check_session_rotation_pending_on_consumers(session,
 466                         &rotation_completed);
 467         if (!rotation_completed ||
 468                         session->rotation_state == LTTNG_ROTATION_STATE_ERROR) {
 469                 goto check_ongoing_rotation;
 470         }
 471
 472         /*
 473          * Now we can clear the "ONGOING" state in the session. New
 474          * rotations can start now.
 475          */
 476         chunk_status = lttng_trace_chunk_get_name(session->chunk_being_archived,
 477                         &archived_chunk_name, NULL);
 478         LTTNG_ASSERT(chunk_status == LTTNG_TRACE_CHUNK_STATUS_OK);
 479         free(session->last_archived_chunk_name);
 480         session->last_archived_chunk_name = strdup(archived_chunk_name);
 481         if (!session->last_archived_chunk_name) {
 482                 PERROR("Failed to duplicate archived chunk name");
 483         }
 484         session_reset_rotation_state(session, LTTNG_ROTATION_STATE_COMPLETED);
 485
 486         if (!session->quiet_rotation) {
 487                 location = session_get_trace_archive_location(session);
 488                 ret = notification_thread_command_session_rotation_completed(
 489                                 notification_thread_handle,
 490                                 session->id,
 491                                 session->last_archived_chunk_id.value,
 492                                 location);
 493                 lttng_trace_archive_location_put(location);
 494                 if (ret != LTTNG_OK) {
 495                         ERR("Failed to notify notification thread of completed rotation for session %s",
 496                                         session->name);
 497                 }
 498         }
 499
 500         ret = 0;
 501 check_ongoing_rotation:
 502         if (session->rotation_state == LTTNG_ROTATION_STATE_ONGOING) {
 503                 chunk_status = lttng_trace_chunk_get_id(
 504                                 session->chunk_being_archived,
 505                                 &chunk_being_archived_id);
 506                 LTTNG_ASSERT(chunk_status == LTTNG_TRACE_CHUNK_STATUS_OK);
 507
 508                 DBG("Rotation of trace archive %" PRIu64 " is still pending for session %s",
 509                                 chunk_being_archived_id, session->name);
 510                 ret = timer_session_rotation_pending_check_start(session,
 511                                 DEFAULT_ROTATE_PENDING_TIMER);
 512                 if (ret) {
 513                         ERR("Failed to re-enable rotation pending timer");
 514                         ret = -1;
 515                         goto end;
 516                 }
 517         }
 518
 519 end:
 520         return ret;
 521 }
 522
 523 /* Call with the session and session_list locks held. */
 524 static
 525 int launch_session_rotation(struct ltt_session *session)
 526 {
 527         int ret;
 528         struct lttng_rotate_session_return rotation_return;
 529
 530         DBG("Launching scheduled time-based rotation on session \"%s\"",
 531                         session->name);
 532
 533         ret = cmd_rotate_session(session, &rotation_return, false,
 534                 LTTNG_TRACE_CHUNK_COMMAND_TYPE_MOVE_TO_COMPLETED);
 535         if (ret == LTTNG_OK) {
 536                 DBG("Scheduled time-based rotation successfully launched on session \"%s\"",
 537                                 session->name);
 538         } else {
 539                 /* Don't consider errors as fatal. */
 540                 DBG("Scheduled time-based rotation aborted for session %s: %s",
 541                                 session->name, lttng_strerror(ret));
 542         }
 543         return 0;
 544 }
 545
 546 static
 547 int run_job(struct rotation_thread_job *job, struct ltt_session *session,
 548                 struct notification_thread_handle *notification_thread_handle)
 549 {
 550         int ret;
 551
 552         switch (job->type) {
 553         case ROTATION_THREAD_JOB_TYPE_SCHEDULED_ROTATION:
 554                 ret = launch_session_rotation(session);
 555                 break;
 556         case ROTATION_THREAD_JOB_TYPE_CHECK_PENDING_ROTATION:
 557                 ret = check_session_rotation_pending(session,
 558                                 notification_thread_handle);
 559                 break;
 560         default:
 561                 abort();
 562         }
 563         return ret;
 564 }
 565
 566 static
 567 int handle_job_queue(struct rotation_thread_handle *handle,
 568                 struct rotation_thread *state __attribute__((unused)),
 569                 struct rotation_thread_timer_queue *queue)
 570 {
 571         int ret = 0;
 572
 573         for (;;) {
 574                 struct ltt_session *session;
 575                 struct rotation_thread_job *job;
 576
 577                 /* Take the queue lock only to pop an element from the list. */
 578                 pthread_mutex_lock(&queue->lock);
 579                 if (cds_list_empty(&queue->list)) {
 580                         pthread_mutex_unlock(&queue->lock);
 581                         break;
 582                 }
 583                 job = cds_list_first_entry(&queue->list,
 584                                 typeof(*job), head);
 585                 cds_list_del(&job->head);
 586                 pthread_mutex_unlock(&queue->lock);
 587
 588                 session_lock_list();
 589                 session = job->session;
 590                 if (!session) {
 591                         DBG("Session \"%s\" not found",
 592                                         session->name != NULL ? session->name : "");
 593                         /*
 594                          * This is a non-fatal error, and we cannot report it to
 595                          * the user (timer), so just print the error and
 596                          * continue the processing.
 597                          *
 598                          * While the timer thread will purge pending signals for
 599                          * a session on the session's destruction, it is
 600                          * possible for a job targeting that session to have
 601                          * already been queued before it was destroyed.
 602                          */
 603                         free(job);
 604                         session_put(session);
 605                         session_unlock_list();
 606                         continue;
 607                 }
 608
 609                 session_lock(session);
 610                 ret = run_job(job, session, handle->notification_thread_handle);
 611                 session_unlock(session);
 612                 /* Release reference held by the job. */
 613                 session_put(session);
 614                 session_unlock_list();
 615                 free(job);
 616                 if (ret) {
 617                         goto end;
 618                 }
 619         }
 620
 621         ret = 0;
 622
 623 end:
 624         return ret;
 625 }
 626
 627 static
 628 int handle_condition(const struct lttng_notification *notification,
 629                 struct notification_thread_handle *notification_thread_handle)
 630 {
 631         int ret = 0;
 632         const char *condition_session_name = NULL;
 633         enum lttng_condition_type condition_type;
 634         enum lttng_condition_status condition_status;
 635         enum lttng_evaluation_status evaluation_status;
 636         uint64_t consumed;
 637         struct ltt_session *session;
 638         const struct lttng_condition *condition =
 639                         lttng_notification_get_const_condition(notification);
 640         const struct lttng_evaluation *evaluation =
 641                         lttng_notification_get_const_evaluation(notification);
 642
 643         condition_type = lttng_condition_get_type(condition);
 644
 645         if (condition_type != LTTNG_CONDITION_TYPE_SESSION_CONSUMED_SIZE) {
 646                 ret = -1;
 647                 ERR("Condition type and session usage type are not the same");
 648                 goto end;
 649         }
 650
 651         /* Fetch info to test */
 652         condition_status = lttng_condition_session_consumed_size_get_session_name(
 653                         condition, &condition_session_name);
 654         if (condition_status != LTTNG_CONDITION_STATUS_OK) {
 655                 ERR("Session name could not be fetched");
 656                 ret = -1;
 657                 goto end;
 658         }
 659         evaluation_status = lttng_evaluation_session_consumed_size_get_consumed_size(evaluation,
 660                         &consumed);
 661         if (evaluation_status != LTTNG_EVALUATION_STATUS_OK) {
 662                 ERR("Failed to get evaluation");
 663                 ret = -1;
 664                 goto end;
 665         }
 666
 667         session_lock_list();
 668         session = session_find_by_name(condition_session_name);
 669         if (!session) {
 670                 DBG("Failed to find session while handling notification: notification type = %s, session name = `%s`",
 671                                 lttng_condition_type_str(condition_type),
 672                                 condition_session_name);
 673                 /*
 674                  * Not a fatal error: a session can be destroyed before we get
 675                  * the chance to handle the notification.
 676                  */
 677                 ret = 0;
 678                 session_unlock_list();
 679                 goto end;
 680         }
 681         session_lock(session);
 682
 683         if (!lttng_trigger_is_equal(session->rotate_trigger,
 684                         lttng_notification_get_const_trigger(notification))) {
 685                 /* Notification does not originate from our rotation trigger. */
 686                 ret = 0;
 687                 goto end_unlock;
 688         }
 689
 690         ret = unsubscribe_session_consumed_size_rotation(session,
 691                         notification_thread_handle);
 692         if (ret) {
 693                 goto end_unlock;
 694         }
 695
 696         ret = cmd_rotate_session(session, NULL, false,
 697                 LTTNG_TRACE_CHUNK_COMMAND_TYPE_MOVE_TO_COMPLETED);
 698         if (ret == -LTTNG_ERR_ROTATION_PENDING) {
 699                 DBG("Rotate already pending, subscribe to the next threshold value");
 700         } else if (ret != LTTNG_OK) {
 701                 ERR("Failed to rotate on size notification with error: %s",
 702                                 lttng_strerror(ret));
 703                 ret = -1;
 704                 goto end_unlock;
 705         }
 706         ret = subscribe_session_consumed_size_rotation(session,
 707                         consumed + session->rotate_size,
 708                         notification_thread_handle);
 709         if (ret) {
 710                 ERR("Failed to subscribe to session consumed size condition");
 711                 goto end_unlock;
 712         }
 713         ret = 0;
 714
 715 end_unlock:
 716         session_unlock(session);
 717         session_put(session);
 718         session_unlock_list();
 719 end:
 720         return ret;
 721 }
 722
 723 static
 724 int handle_notification_channel(int fd __attribute__((unused)),
 725                 struct rotation_thread_handle *handle,
 726                 struct rotation_thread *state __attribute__((unused)))
 727 {
 728         int ret;
 729         bool notification_pending;
 730         struct lttng_notification *notification = NULL;
 731         enum lttng_notification_channel_status status;
 732
 733         status = lttng_notification_channel_has_pending_notification(
 734                         rotate_notification_channel, &notification_pending);
 735         if (status != LTTNG_NOTIFICATION_CHANNEL_STATUS_OK) {
 736                 ERR("Error occurred while checking for pending notification");
 737                 ret = -1;
 738                 goto end;
 739         }
 740
 741         if (!notification_pending) {
 742                 ret = 0;
 743                 goto end;
 744         }
 745
 746         /* Receive the next notification. */
 747         status = lttng_notification_channel_get_next_notification(
 748                         rotate_notification_channel,
 749                         &notification);
 750
 751         switch (status) {
 752         case LTTNG_NOTIFICATION_CHANNEL_STATUS_OK:
 753                 break;
 754         case LTTNG_NOTIFICATION_CHANNEL_STATUS_NOTIFICATIONS_DROPPED:
 755                 /* Not an error, we will wait for the next one */
 756                 ret = 0;
 757                 goto end;;
 758         case LTTNG_NOTIFICATION_CHANNEL_STATUS_CLOSED:
 759                 ERR("Notification channel was closed");
 760                 ret = -1;
 761                 goto end;
 762         default:
 763                 /* Unhandled conditions / errors. */
 764                 ERR("Unknown notification channel status");
 765                 ret = -1;
 766                 goto end;
 767         }
 768
 769         ret = handle_condition(notification,
 770                         handle->notification_thread_handle);
 771
 772 end:
 773         lttng_notification_destroy(notification);
 774         return ret;
 775 }
 776
 777 static
 778 void *thread_rotation(void *data)
 779 {
 780         int ret;
 781         struct rotation_thread_handle *handle = (rotation_thread_handle *) data;
 782         struct rotation_thread thread;
 783         int queue_pipe_fd;
 784
 785         DBG("Started rotation thread");
 786         rcu_register_thread();
 787         rcu_thread_online();
 788         health_register(the_health_sessiond, HEALTH_SESSIOND_TYPE_ROTATION);
 789         health_code_update();
 790
 791         if (!handle) {
 792                 ERR("Invalid thread context provided");
 793                 goto end;
 794         }
 795
 796         queue_pipe_fd = lttng_pipe_get_readfd(
 797                         handle->rotation_timer_queue->event_pipe);
 798
 799
 800         ret = init_thread_state(handle, &thread);
 801         if (ret) {
 802                 goto error;
 803         }
 804
 805         while (true) {
 806                 int fd_count, i;
 807
 808                 health_poll_entry();
 809                 DBG("Entering poll wait");
 810                 ret = lttng_poll_wait(&thread.events, -1);
 811                 DBG("Poll wait returned (%i)", ret);
 812                 health_poll_exit();
 813                 if (ret < 0) {
 814                         /*
 815                          * Restart interrupted system call.
 816                          */
 817                         if (errno == EINTR) {
 818                                 continue;
 819                         }
 820                         ERR("Error encountered during lttng_poll_wait (%i)", ret);
 821                         goto error;
 822                 }
 823
 824                 fd_count = ret;
 825                 for (i = 0; i < fd_count; i++) {
 826                         int fd = LTTNG_POLL_GETFD(&thread.events, i);
 827                         uint32_t revents = LTTNG_POLL_GETEV(&thread.events, i);
 828
 829                         DBG("Handling fd (%i) activity (%u)",
 830                                         fd, revents);
 831
 832                         if (revents & LPOLLERR) {
 833                                 ERR("Polling returned an error on fd %i", fd);
 834                                 goto error;
 835                         }
 836
 837                         if (fd == rotate_notification_channel->socket) {
 838                                 ret = handle_notification_channel(fd, handle,
 839                                                 &thread);
 840                                 if (ret) {
 841                                         ERR("Error occurred while handling activity on notification channel socket");
 842                                         goto error;
 843                                 }
 844                         } else {
 845                                 /* Job queue or quit pipe activity. */
 846
 847                                 /*
 848                                  * The job queue is serviced if there is
 849                                  * activity on the quit pipe to ensure it is
 850                                  * flushed and all references held in the queue
 851                                  * are released.
 852                                  */
 853                                 ret = handle_job_queue(handle, &thread,
 854                                                 handle->rotation_timer_queue);
 855                                 if (ret) {
 856                                         ERR("Failed to handle rotation timer pipe event");
 857                                         goto error;
 858                                 }
 859
 860                                 if (fd == queue_pipe_fd) {
 861                                         char buf;
 862
 863                                         ret = lttng_read(fd, &buf, 1);
 864                                         if (ret != 1) {
 865                                                 ERR("Failed to read from wakeup pipe (fd = %i)", fd);
 866                                                 goto error;
 867                                         }
 868                                 } else {
 869                                         DBG("Quit pipe activity");
 870                                         goto exit;
 871                                 }
 872                         }
 873                 }
 874         }
 875 exit:
 876 error:
 877         DBG("Thread exit");
 878         fini_thread_state(&thread);
 879 end:
 880         health_unregister(the_health_sessiond);
 881         rcu_thread_offline();
 882         rcu_unregister_thread();
 883         return NULL;
 884 }
 885
 886 static
 887 bool shutdown_rotation_thread(void *thread_data)
 888 {
 889         struct rotation_thread_handle *handle = (rotation_thread_handle *) thread_data;
 890         const int write_fd = lttng_pipe_get_writefd(handle->quit_pipe);
 891
 892         return notify_thread_pipe(write_fd) == 1;
 893 }
 894
 895 bool launch_rotation_thread(struct rotation_thread_handle *handle)
 896 {
 897         struct lttng_thread *thread;
 898
 899         thread = lttng_thread_create("Rotation",
 900                         thread_rotation,
 901                         shutdown_rotation_thread,
 902                         NULL,
 903                         handle);
 904         if (!thread) {
 905                 goto error;
 906         }
 907         lttng_thread_put(thread);
 908         return true;
 909 error:
 910         return false;
 911 }