2 * Copyright (C) 2017 Julien Desfossez <jdesfossez@efficios.com>
3 * Copyright (C) 2018 Jérémie Galarneau <jeremie.galarneau@efficios.com>
5 * SPDX-License-Identifier: GPL-2.0-only
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>
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>
30 #include "rotation-thread.hpp"
31 #include "lttng-sessiond.hpp"
32 #include "health-sessiond.hpp"
35 #include "session.hpp"
37 #include "notification-thread-commands.hpp"
42 #include <urcu/list.h>
44 struct lttng_notification_channel
*rotate_notification_channel
= NULL
;
46 struct rotation_thread
{
47 struct lttng_poll_event events
;
51 * The timer thread enqueues jobs and wakes up the rotation thread.
52 * When the rotation thread wakes up, it empties the queue.
54 struct rotation_thread_timer_queue
{
55 struct lttng_pipe
*event_pipe
;
56 struct cds_list_head list
;
60 struct rotation_thread_handle
{
61 struct rotation_thread_timer_queue
*rotation_timer_queue
;
62 /* Access to the notification thread cmd_queue */
63 struct notification_thread_handle
*notification_thread_handle
;
64 /* Thread-specific quit pipe. */
65 struct lttng_pipe
*quit_pipe
;
69 struct rotation_thread_job
{
70 enum rotation_thread_job_type type
;
71 struct ltt_session
*session
;
72 /* List member in struct rotation_thread_timer_queue. */
73 struct cds_list_head head
;
78 const char *get_job_type_str(enum rotation_thread_job_type job_type
)
81 case ROTATION_THREAD_JOB_TYPE_CHECK_PENDING_ROTATION
:
82 return "CHECK_PENDING_ROTATION";
83 case ROTATION_THREAD_JOB_TYPE_SCHEDULED_ROTATION
:
84 return "SCHEDULED_ROTATION";
90 struct rotation_thread_timer_queue
*rotation_thread_timer_queue_create(void)
92 struct rotation_thread_timer_queue
*queue
= NULL
;
94 queue
= zmalloc
<rotation_thread_timer_queue
>();
96 PERROR("Failed to allocate timer rotate queue");
100 queue
->event_pipe
= lttng_pipe_open(FD_CLOEXEC
| O_NONBLOCK
);
101 CDS_INIT_LIST_HEAD(&queue
->list
);
102 pthread_mutex_init(&queue
->lock
, NULL
);
107 void rotation_thread_timer_queue_destroy(
108 struct rotation_thread_timer_queue
*queue
)
114 lttng_pipe_destroy(queue
->event_pipe
);
116 pthread_mutex_lock(&queue
->lock
);
117 LTTNG_ASSERT(cds_list_empty(&queue
->list
));
118 pthread_mutex_unlock(&queue
->lock
);
119 pthread_mutex_destroy(&queue
->lock
);
124 * Destroy the thread data previously created by the init function.
126 void rotation_thread_handle_destroy(
127 struct rotation_thread_handle
*handle
)
129 lttng_pipe_destroy(handle
->quit_pipe
);
133 struct rotation_thread_handle
*rotation_thread_handle_create(
134 struct rotation_thread_timer_queue
*rotation_timer_queue
,
135 struct notification_thread_handle
*notification_thread_handle
)
137 struct rotation_thread_handle
*handle
;
139 handle
= zmalloc
<rotation_thread_handle
>();
144 handle
->rotation_timer_queue
= rotation_timer_queue
;
145 handle
->notification_thread_handle
= notification_thread_handle
;
146 handle
->quit_pipe
= lttng_pipe_open(FD_CLOEXEC
);
147 if (!handle
->quit_pipe
) {
154 rotation_thread_handle_destroy(handle
);
159 * Called with the rotation_thread_timer_queue lock held.
160 * Return true if the same timer job already exists in the queue, false if not.
163 bool timer_job_exists(const struct rotation_thread_timer_queue
*queue
,
164 enum rotation_thread_job_type job_type
,
165 struct ltt_session
*session
)
168 struct rotation_thread_job
*job
;
170 cds_list_for_each_entry(job
, &queue
->list
, head
) {
171 if (job
->session
== session
&& job
->type
== job_type
) {
180 void rotation_thread_enqueue_job(struct rotation_thread_timer_queue
*queue
,
181 enum rotation_thread_job_type job_type
,
182 struct ltt_session
*session
)
185 const char dummy
= '!';
186 struct rotation_thread_job
*job
= NULL
;
187 const char *job_type_str
= get_job_type_str(job_type
);
189 pthread_mutex_lock(&queue
->lock
);
190 if (timer_job_exists(queue
, job_type
, session
)) {
192 * This timer job is already pending, we don't need to add
198 job
= zmalloc
<rotation_thread_job
>();
200 PERROR("Failed to allocate rotation thread job of type \"%s\" for session \"%s\"",
201 job_type_str
, session
->name
);
204 /* No reason for this to fail as the caller must hold a reference. */
205 (void) session_get(session
);
207 job
->session
= session
;
208 job
->type
= job_type
;
209 cds_list_add_tail(&job
->head
, &queue
->list
);
211 ret
= lttng_write(lttng_pipe_get_writefd(queue
->event_pipe
), &dummy
,
215 * We do not want to block in the timer handler, the job has
216 * been enqueued in the list, the wakeup pipe is probably full,
217 * the job will be processed when the rotation_thread catches
221 DIAGNOSTIC_IGNORE_LOGICAL_OP
222 if (errno
== EAGAIN
|| errno
== EWOULDBLOCK
) {
225 * Not an error, but would be surprising and indicate
226 * that the rotation thread can't keep up with the
229 DBG("Wake-up pipe of rotation thread job queue is full");
232 PERROR("Failed to wake-up the rotation thread after pushing a job of type \"%s\" for session \"%s\"",
233 job_type_str
, session
->name
);
238 pthread_mutex_unlock(&queue
->lock
);
242 int init_poll_set(struct lttng_poll_event
*poll_set
,
243 struct rotation_thread_handle
*handle
)
248 * Create pollset with size 3:
249 * - rotation thread quit pipe,
250 * - rotation thread timer queue pipe,
251 * - notification channel sock,
253 ret
= lttng_poll_create(poll_set
, 5, LTTNG_CLOEXEC
);
258 ret
= lttng_poll_add(poll_set
,
259 lttng_pipe_get_readfd(handle
->quit_pipe
),
262 ERR("Failed to add quit pipe read fd to poll set");
266 ret
= lttng_poll_add(poll_set
,
267 lttng_pipe_get_readfd(handle
->rotation_timer_queue
->event_pipe
),
270 ERR("Failed to add rotate_pending fd to poll set");
276 lttng_poll_clean(poll_set
);
281 void fini_thread_state(struct rotation_thread
*state
)
283 lttng_poll_clean(&state
->events
);
284 if (rotate_notification_channel
) {
285 lttng_notification_channel_destroy(rotate_notification_channel
);
290 int init_thread_state(struct rotation_thread_handle
*handle
,
291 struct rotation_thread
*state
)
295 memset(state
, 0, sizeof(*state
));
296 lttng_poll_init(&state
->events
);
298 ret
= init_poll_set(&state
->events
, handle
);
300 ERR("Failed to initialize rotation thread poll set");
304 rotate_notification_channel
= lttng_notification_channel_create(
305 lttng_session_daemon_notification_endpoint
);
306 if (!rotate_notification_channel
) {
307 ERR("Could not create notification channel");
311 ret
= lttng_poll_add(&state
->events
, rotate_notification_channel
->socket
,
314 ERR("Failed to add notification fd to pollset");
323 void check_session_rotation_pending_on_consumers(struct ltt_session
*session
,
324 bool *_rotation_completed
)
327 struct consumer_socket
*socket
;
328 struct cds_lfht_iter iter
;
329 enum consumer_trace_chunk_exists_status exists_status
;
331 bool chunk_exists_on_peer
= false;
332 enum lttng_trace_chunk_status chunk_status
;
334 LTTNG_ASSERT(session
->chunk_being_archived
);
337 * Check for a local pending rotation on all consumers (32-bit
338 * user space, 64-bit user space, and kernel).
341 if (!session
->ust_session
) {
344 cds_lfht_for_each_entry(session
->ust_session
->consumer
->socks
->ht
,
345 &iter
, socket
, node
.node
) {
346 relayd_id
= session
->ust_session
->consumer
->type
== CONSUMER_DST_LOCAL
?
348 session
->ust_session
->consumer
->net_seq_index
;
350 pthread_mutex_lock(socket
->lock
);
351 ret
= consumer_trace_chunk_exists(socket
,
353 session
->id
, session
->chunk_being_archived
,
356 pthread_mutex_unlock(socket
->lock
);
357 ERR("Error occurred while checking rotation status on consumer daemon");
361 if (exists_status
!= CONSUMER_TRACE_CHUNK_EXISTS_STATUS_UNKNOWN_CHUNK
) {
362 pthread_mutex_unlock(socket
->lock
);
363 chunk_exists_on_peer
= true;
366 pthread_mutex_unlock(socket
->lock
);
370 if (!session
->kernel_session
) {
373 cds_lfht_for_each_entry(session
->kernel_session
->consumer
->socks
->ht
,
374 &iter
, socket
, node
.node
) {
375 pthread_mutex_lock(socket
->lock
);
376 relayd_id
= session
->kernel_session
->consumer
->type
== CONSUMER_DST_LOCAL
?
378 session
->kernel_session
->consumer
->net_seq_index
;
380 ret
= consumer_trace_chunk_exists(socket
,
382 session
->id
, session
->chunk_being_archived
,
385 pthread_mutex_unlock(socket
->lock
);
386 ERR("Error occurred while checking rotation status on consumer daemon");
390 if (exists_status
!= CONSUMER_TRACE_CHUNK_EXISTS_STATUS_UNKNOWN_CHUNK
) {
391 pthread_mutex_unlock(socket
->lock
);
392 chunk_exists_on_peer
= true;
395 pthread_mutex_unlock(socket
->lock
);
401 if (!chunk_exists_on_peer
) {
402 uint64_t chunk_being_archived_id
;
404 chunk_status
= lttng_trace_chunk_get_id(
405 session
->chunk_being_archived
,
406 &chunk_being_archived_id
);
407 LTTNG_ASSERT(chunk_status
== LTTNG_TRACE_CHUNK_STATUS_OK
);
408 DBG("Rotation of trace archive %" PRIu64
" of session \"%s\" is complete on all consumers",
409 chunk_being_archived_id
,
412 *_rotation_completed
= !chunk_exists_on_peer
;
414 ret
= session_reset_rotation_state(session
,
415 LTTNG_ROTATION_STATE_ERROR
);
417 ERR("Failed to reset rotation state of session \"%s\"",
424 * Check if the last rotation was completed, called with session lock held.
425 * Should only return non-zero in the event of a fatal error. Doing so will
426 * shutdown the thread.
429 int check_session_rotation_pending(struct ltt_session
*session
,
430 struct notification_thread_handle
*notification_thread_handle
)
433 struct lttng_trace_archive_location
*location
;
434 enum lttng_trace_chunk_status chunk_status
;
435 bool rotation_completed
= false;
436 const char *archived_chunk_name
;
437 uint64_t chunk_being_archived_id
;
439 if (!session
->chunk_being_archived
) {
444 chunk_status
= lttng_trace_chunk_get_id(session
->chunk_being_archived
,
445 &chunk_being_archived_id
);
446 LTTNG_ASSERT(chunk_status
== LTTNG_TRACE_CHUNK_STATUS_OK
);
448 DBG("Checking for pending rotation on session \"%s\", trace archive %" PRIu64
,
449 session
->name
, chunk_being_archived_id
);
452 * The rotation-pending check timer of a session is launched in
453 * one-shot mode. If the rotation is incomplete, the rotation
454 * thread will re-enable the pending-check timer.
456 * The timer thread can't stop the timer itself since it is involved
457 * in the check for the timer's quiescence.
459 ret
= timer_session_rotation_pending_check_stop(session
);
461 goto check_ongoing_rotation
;
464 check_session_rotation_pending_on_consumers(session
,
465 &rotation_completed
);
466 if (!rotation_completed
||
467 session
->rotation_state
== LTTNG_ROTATION_STATE_ERROR
) {
468 goto check_ongoing_rotation
;
472 * Now we can clear the "ONGOING" state in the session. New
473 * rotations can start now.
475 chunk_status
= lttng_trace_chunk_get_name(session
->chunk_being_archived
,
476 &archived_chunk_name
, NULL
);
477 LTTNG_ASSERT(chunk_status
== LTTNG_TRACE_CHUNK_STATUS_OK
);
478 free(session
->last_archived_chunk_name
);
479 session
->last_archived_chunk_name
= strdup(archived_chunk_name
);
480 if (!session
->last_archived_chunk_name
) {
481 PERROR("Failed to duplicate archived chunk name");
483 session_reset_rotation_state(session
, LTTNG_ROTATION_STATE_COMPLETED
);
485 if (!session
->quiet_rotation
) {
486 location
= session_get_trace_archive_location(session
);
487 ret
= notification_thread_command_session_rotation_completed(
488 notification_thread_handle
,
492 session
->last_archived_chunk_id
.value
,
494 lttng_trace_archive_location_put(location
);
495 if (ret
!= LTTNG_OK
) {
496 ERR("Failed to notify notification thread of completed rotation for session %s",
502 check_ongoing_rotation
:
503 if (session
->rotation_state
== LTTNG_ROTATION_STATE_ONGOING
) {
504 chunk_status
= lttng_trace_chunk_get_id(
505 session
->chunk_being_archived
,
506 &chunk_being_archived_id
);
507 LTTNG_ASSERT(chunk_status
== LTTNG_TRACE_CHUNK_STATUS_OK
);
509 DBG("Rotation of trace archive %" PRIu64
" is still pending for session %s",
510 chunk_being_archived_id
, session
->name
);
511 ret
= timer_session_rotation_pending_check_start(session
,
512 DEFAULT_ROTATE_PENDING_TIMER
);
514 ERR("Failed to re-enable rotation pending timer");
524 /* Call with the session and session_list locks held. */
526 int launch_session_rotation(struct ltt_session
*session
)
529 struct lttng_rotate_session_return rotation_return
;
531 DBG("Launching scheduled time-based rotation on session \"%s\"",
534 ret
= cmd_rotate_session(session
, &rotation_return
, false,
535 LTTNG_TRACE_CHUNK_COMMAND_TYPE_MOVE_TO_COMPLETED
);
536 if (ret
== LTTNG_OK
) {
537 DBG("Scheduled time-based rotation successfully launched on session \"%s\"",
540 /* Don't consider errors as fatal. */
541 DBG("Scheduled time-based rotation aborted for session %s: %s",
542 session
->name
, lttng_strerror(ret
));
548 int run_job(struct rotation_thread_job
*job
, struct ltt_session
*session
,
549 struct notification_thread_handle
*notification_thread_handle
)
554 case ROTATION_THREAD_JOB_TYPE_SCHEDULED_ROTATION
:
555 ret
= launch_session_rotation(session
);
557 case ROTATION_THREAD_JOB_TYPE_CHECK_PENDING_ROTATION
:
558 ret
= check_session_rotation_pending(session
,
559 notification_thread_handle
);
568 int handle_job_queue(struct rotation_thread_handle
*handle
,
569 struct rotation_thread
*state
__attribute__((unused
)),
570 struct rotation_thread_timer_queue
*queue
)
575 struct ltt_session
*session
;
576 struct rotation_thread_job
*job
;
578 /* Take the queue lock only to pop an element from the list. */
579 pthread_mutex_lock(&queue
->lock
);
580 if (cds_list_empty(&queue
->list
)) {
581 pthread_mutex_unlock(&queue
->lock
);
584 job
= cds_list_first_entry(&queue
->list
,
586 cds_list_del(&job
->head
);
587 pthread_mutex_unlock(&queue
->lock
);
590 session
= job
->session
;
592 DBG("Session \"%s\" not found",
593 session
->name
!= NULL
? session
->name
: "");
595 * This is a non-fatal error, and we cannot report it to
596 * the user (timer), so just print the error and
597 * continue the processing.
599 * While the timer thread will purge pending signals for
600 * a session on the session's destruction, it is
601 * possible for a job targeting that session to have
602 * already been queued before it was destroyed.
605 session_put(session
);
606 session_unlock_list();
610 session_lock(session
);
611 ret
= run_job(job
, session
, handle
->notification_thread_handle
);
612 session_unlock(session
);
613 /* Release reference held by the job. */
614 session_put(session
);
615 session_unlock_list();
629 int handle_condition(const struct lttng_condition
*condition
,
630 const struct lttng_evaluation
*evaluation
,
631 struct notification_thread_handle
*notification_thread_handle
)
634 const char *condition_session_name
= NULL
;
635 enum lttng_condition_type condition_type
;
636 enum lttng_condition_status condition_status
;
637 enum lttng_evaluation_status evaluation_status
;
639 struct ltt_session
*session
;
641 condition_type
= lttng_condition_get_type(condition
);
643 if (condition_type
!= LTTNG_CONDITION_TYPE_SESSION_CONSUMED_SIZE
) {
645 ERR("Condition type and session usage type are not the same");
649 /* Fetch info to test */
650 condition_status
= lttng_condition_session_consumed_size_get_session_name(
651 condition
, &condition_session_name
);
652 if (condition_status
!= LTTNG_CONDITION_STATUS_OK
) {
653 ERR("Session name could not be fetched");
657 evaluation_status
= lttng_evaluation_session_consumed_size_get_consumed_size(evaluation
,
659 if (evaluation_status
!= LTTNG_EVALUATION_STATUS_OK
) {
660 ERR("Failed to get evaluation");
666 session
= session_find_by_name(condition_session_name
);
668 DBG("Failed to find session while handling notification: notification type = %s, session name = `%s`",
669 lttng_condition_type_str(condition_type
),
670 condition_session_name
);
672 * Not a fatal error: a session can be destroyed before we get
673 * the chance to handle the notification.
676 session_unlock_list();
679 session_lock(session
);
681 ret
= unsubscribe_session_consumed_size_rotation(session
,
682 notification_thread_handle
);
687 ret
= cmd_rotate_session(session
, NULL
, false,
688 LTTNG_TRACE_CHUNK_COMMAND_TYPE_MOVE_TO_COMPLETED
);
689 if (ret
== -LTTNG_ERR_ROTATION_PENDING
) {
690 DBG("Rotate already pending, subscribe to the next threshold value");
691 } else if (ret
!= LTTNG_OK
) {
692 ERR("Failed to rotate on size notification with error: %s",
693 lttng_strerror(ret
));
697 ret
= subscribe_session_consumed_size_rotation(session
,
698 consumed
+ session
->rotate_size
,
699 notification_thread_handle
);
701 ERR("Failed to subscribe to session consumed size condition");
707 session_unlock(session
);
708 session_put(session
);
709 session_unlock_list();
715 int handle_notification_channel(int fd
__attribute__((unused
)),
716 struct rotation_thread_handle
*handle
,
717 struct rotation_thread
*state
__attribute__((unused
)))
720 bool notification_pending
;
721 struct lttng_notification
*notification
= NULL
;
722 enum lttng_notification_channel_status status
;
723 const struct lttng_evaluation
*notification_evaluation
;
724 const struct lttng_condition
*notification_condition
;
726 status
= lttng_notification_channel_has_pending_notification(
727 rotate_notification_channel
, ¬ification_pending
);
728 if (status
!= LTTNG_NOTIFICATION_CHANNEL_STATUS_OK
) {
729 ERR("Error occurred while checking for pending notification");
734 if (!notification_pending
) {
739 /* Receive the next notification. */
740 status
= lttng_notification_channel_get_next_notification(
741 rotate_notification_channel
,
745 case LTTNG_NOTIFICATION_CHANNEL_STATUS_OK
:
747 case LTTNG_NOTIFICATION_CHANNEL_STATUS_NOTIFICATIONS_DROPPED
:
748 /* Not an error, we will wait for the next one */
751 case LTTNG_NOTIFICATION_CHANNEL_STATUS_CLOSED
:
752 ERR("Notification channel was closed");
756 /* Unhandled conditions / errors. */
757 ERR("Unknown notification channel status");
762 notification_condition
= lttng_notification_get_condition(notification
);
763 notification_evaluation
= lttng_notification_get_evaluation(notification
);
765 ret
= handle_condition(notification_condition
, notification_evaluation
,
766 handle
->notification_thread_handle
);
769 lttng_notification_destroy(notification
);
774 void *thread_rotation(void *data
)
777 struct rotation_thread_handle
*handle
= (rotation_thread_handle
*) data
;
778 struct rotation_thread thread
;
781 DBG("Started rotation thread");
782 rcu_register_thread();
784 health_register(the_health_sessiond
, HEALTH_SESSIOND_TYPE_ROTATION
);
785 health_code_update();
788 ERR("Invalid thread context provided");
792 queue_pipe_fd
= lttng_pipe_get_readfd(
793 handle
->rotation_timer_queue
->event_pipe
);
796 ret
= init_thread_state(handle
, &thread
);
805 DBG("Entering poll wait");
806 ret
= lttng_poll_wait(&thread
.events
, -1);
807 DBG("Poll wait returned (%i)", ret
);
811 * Restart interrupted system call.
813 if (errno
== EINTR
) {
816 ERR("Error encountered during lttng_poll_wait (%i)", ret
);
821 for (i
= 0; i
< fd_count
; i
++) {
822 int fd
= LTTNG_POLL_GETFD(&thread
.events
, i
);
823 uint32_t revents
= LTTNG_POLL_GETEV(&thread
.events
, i
);
825 DBG("Handling fd (%i) activity (%u)",
828 if (revents
& LPOLLERR
) {
829 ERR("Polling returned an error on fd %i", fd
);
833 if (fd
== rotate_notification_channel
->socket
) {
834 ret
= handle_notification_channel(fd
, handle
,
837 ERR("Error occurred while handling activity on notification channel socket");
841 /* Job queue or quit pipe activity. */
844 * The job queue is serviced if there is
845 * activity on the quit pipe to ensure it is
846 * flushed and all references held in the queue
849 ret
= handle_job_queue(handle
, &thread
,
850 handle
->rotation_timer_queue
);
852 ERR("Failed to handle rotation timer pipe event");
856 if (fd
== queue_pipe_fd
) {
859 ret
= lttng_read(fd
, &buf
, 1);
861 ERR("Failed to read from wakeup pipe (fd = %i)", fd
);
865 DBG("Quit pipe activity");
874 fini_thread_state(&thread
);
876 health_unregister(the_health_sessiond
);
877 rcu_thread_offline();
878 rcu_unregister_thread();
883 bool shutdown_rotation_thread(void *thread_data
)
885 struct rotation_thread_handle
*handle
= (rotation_thread_handle
*) thread_data
;
886 const int write_fd
= lttng_pipe_get_writefd(handle
->quit_pipe
);
888 return notify_thread_pipe(write_fd
) == 1;
891 bool launch_rotation_thread(struct rotation_thread_handle
*handle
)
893 struct lttng_thread
*thread
;
895 thread
= lttng_thread_create("Rotation",
897 shutdown_rotation_thread
,
903 lttng_thread_put(thread
);