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

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