2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 #include <sys/socket.h>
28 #include <sys/types.h>
33 #include <bin/lttng-consumerd/health-consumerd.h>
34 #include <common/common.h>
35 #include <common/utils.h>
36 #include <common/time.h>
37 #include <common/compat/poll.h>
38 #include <common/compat/endian.h>
39 #include <common/index/index.h>
40 #include <common/kernel-ctl/kernel-ctl.h>
41 #include <common/sessiond-comm/relayd.h>
42 #include <common/sessiond-comm/sessiond-comm.h>
43 #include <common/kernel-consumer/kernel-consumer.h>
44 #include <common/relayd/relayd.h>
45 #include <common/ust-consumer/ust-consumer.h>
46 #include <common/consumer/consumer-timer.h>
47 #include <common/consumer/consumer.h>
48 #include <common/consumer/consumer-stream.h>
49 #include <common/consumer/consumer-testpoint.h>
50 #include <common/align.h>
51 #include <common/consumer/consumer-metadata-cache.h>
52 #include <common/trace-chunk.h>
53 #include <common/trace-chunk-registry.h>
54 #include <common/string-utils/format.h>
55 #include <common/dynamic-array.h>
57 struct lttng_consumer_global_data consumer_data
= {
60 .type
= LTTNG_CONSUMER_UNKNOWN
,
63 enum consumer_channel_action
{
66 CONSUMER_CHANNEL_QUIT
,
69 struct consumer_channel_msg
{
70 enum consumer_channel_action action
;
71 struct lttng_consumer_channel
*chan
; /* add */
72 uint64_t key
; /* del */
75 /* Flag used to temporarily pause data consumption from testpoints. */
76 int data_consumption_paused
;
79 * Flag to inform the polling thread to quit when all fd hung up. Updated by
80 * the consumer_thread_receive_fds when it notices that all fds has hung up.
81 * Also updated by the signal handler (consumer_should_exit()). Read by the
87 * Global hash table containing respectively metadata and data streams. The
88 * stream element in this ht should only be updated by the metadata poll thread
89 * for the metadata and the data poll thread for the data.
91 static struct lttng_ht
*metadata_ht
;
92 static struct lttng_ht
*data_ht
;
95 * Notify a thread lttng pipe to poll back again. This usually means that some
96 * global state has changed so we just send back the thread in a poll wait
99 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
101 struct lttng_consumer_stream
*null_stream
= NULL
;
105 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
108 static void notify_health_quit_pipe(int *pipe
)
112 ret
= lttng_write(pipe
[1], "4", 1);
114 PERROR("write consumer health quit");
118 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
119 struct lttng_consumer_channel
*chan
,
121 enum consumer_channel_action action
)
123 struct consumer_channel_msg msg
;
126 memset(&msg
, 0, sizeof(msg
));
131 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
132 if (ret
< sizeof(msg
)) {
133 PERROR("notify_channel_pipe write error");
137 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
140 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
143 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
144 struct lttng_consumer_channel
**chan
,
146 enum consumer_channel_action
*action
)
148 struct consumer_channel_msg msg
;
151 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
152 if (ret
< sizeof(msg
)) {
156 *action
= msg
.action
;
164 * Cleanup the stream list of a channel. Those streams are not yet globally
167 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
169 struct lttng_consumer_stream
*stream
, *stmp
;
173 /* Delete streams that might have been left in the stream list. */
174 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
176 cds_list_del(&stream
->send_node
);
178 * Once a stream is added to this list, the buffers were created so we
179 * have a guarantee that this call will succeed. Setting the monitor
180 * mode to 0 so we don't lock nor try to delete the stream from the
184 consumer_stream_destroy(stream
, NULL
);
189 * Find a stream. The consumer_data.lock must be locked during this
192 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
195 struct lttng_ht_iter iter
;
196 struct lttng_ht_node_u64
*node
;
197 struct lttng_consumer_stream
*stream
= NULL
;
201 /* -1ULL keys are lookup failures */
202 if (key
== (uint64_t) -1ULL) {
208 lttng_ht_lookup(ht
, &key
, &iter
);
209 node
= lttng_ht_iter_get_node_u64(&iter
);
211 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
219 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
221 struct lttng_consumer_stream
*stream
;
224 stream
= find_stream(key
, ht
);
226 stream
->key
= (uint64_t) -1ULL;
228 * We don't want the lookup to match, but we still need
229 * to iterate on this stream when iterating over the hash table. Just
230 * change the node key.
232 stream
->node
.key
= (uint64_t) -1ULL;
238 * Return a channel object for the given key.
240 * RCU read side lock MUST be acquired before calling this function and
241 * protects the channel ptr.
243 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
245 struct lttng_ht_iter iter
;
246 struct lttng_ht_node_u64
*node
;
247 struct lttng_consumer_channel
*channel
= NULL
;
249 /* -1ULL keys are lookup failures */
250 if (key
== (uint64_t) -1ULL) {
254 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
255 node
= lttng_ht_iter_get_node_u64(&iter
);
257 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
264 * There is a possibility that the consumer does not have enough time between
265 * the close of the channel on the session daemon and the cleanup in here thus
266 * once we have a channel add with an existing key, we know for sure that this
267 * channel will eventually get cleaned up by all streams being closed.
269 * This function just nullifies the already existing channel key.
271 static void steal_channel_key(uint64_t key
)
273 struct lttng_consumer_channel
*channel
;
276 channel
= consumer_find_channel(key
);
278 channel
->key
= (uint64_t) -1ULL;
280 * We don't want the lookup to match, but we still need to iterate on
281 * this channel when iterating over the hash table. Just change the
284 channel
->node
.key
= (uint64_t) -1ULL;
289 static void free_channel_rcu(struct rcu_head
*head
)
291 struct lttng_ht_node_u64
*node
=
292 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
293 struct lttng_consumer_channel
*channel
=
294 caa_container_of(node
, struct lttng_consumer_channel
, node
);
296 switch (consumer_data
.type
) {
297 case LTTNG_CONSUMER_KERNEL
:
299 case LTTNG_CONSUMER32_UST
:
300 case LTTNG_CONSUMER64_UST
:
301 lttng_ustconsumer_free_channel(channel
);
304 ERR("Unknown consumer_data type");
311 * RCU protected relayd socket pair free.
313 static void free_relayd_rcu(struct rcu_head
*head
)
315 struct lttng_ht_node_u64
*node
=
316 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
317 struct consumer_relayd_sock_pair
*relayd
=
318 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
321 * Close all sockets. This is done in the call RCU since we don't want the
322 * socket fds to be reassigned thus potentially creating bad state of the
325 * We do not have to lock the control socket mutex here since at this stage
326 * there is no one referencing to this relayd object.
328 (void) relayd_close(&relayd
->control_sock
);
329 (void) relayd_close(&relayd
->data_sock
);
331 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
336 * Destroy and free relayd socket pair object.
338 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
341 struct lttng_ht_iter iter
;
343 if (relayd
== NULL
) {
347 DBG("Consumer destroy and close relayd socket pair");
349 iter
.iter
.node
= &relayd
->node
.node
;
350 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
352 /* We assume the relayd is being or is destroyed */
356 /* RCU free() call */
357 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
361 * Remove a channel from the global list protected by a mutex. This function is
362 * also responsible for freeing its data structures.
364 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
366 struct lttng_ht_iter iter
;
368 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
370 pthread_mutex_lock(&consumer_data
.lock
);
371 pthread_mutex_lock(&channel
->lock
);
373 /* Destroy streams that might have been left in the stream list. */
374 clean_channel_stream_list(channel
);
376 if (channel
->live_timer_enabled
== 1) {
377 consumer_timer_live_stop(channel
);
379 if (channel
->monitor_timer_enabled
== 1) {
380 consumer_timer_monitor_stop(channel
);
383 switch (consumer_data
.type
) {
384 case LTTNG_CONSUMER_KERNEL
:
386 case LTTNG_CONSUMER32_UST
:
387 case LTTNG_CONSUMER64_UST
:
388 lttng_ustconsumer_del_channel(channel
);
391 ERR("Unknown consumer_data type");
396 lttng_trace_chunk_put(channel
->trace_chunk
);
397 channel
->trace_chunk
= NULL
;
399 if (channel
->is_published
) {
403 iter
.iter
.node
= &channel
->node
.node
;
404 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
407 iter
.iter
.node
= &channel
->channels_by_session_id_ht_node
.node
;
408 ret
= lttng_ht_del(consumer_data
.channels_by_session_id_ht
,
414 channel
->is_deleted
= true;
415 call_rcu(&channel
->node
.head
, free_channel_rcu
);
417 pthread_mutex_unlock(&channel
->lock
);
418 pthread_mutex_unlock(&consumer_data
.lock
);
422 * Iterate over the relayd hash table and destroy each element. Finally,
423 * destroy the whole hash table.
425 static void cleanup_relayd_ht(void)
427 struct lttng_ht_iter iter
;
428 struct consumer_relayd_sock_pair
*relayd
;
432 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
434 consumer_destroy_relayd(relayd
);
439 lttng_ht_destroy(consumer_data
.relayd_ht
);
443 * Update the end point status of all streams having the given network sequence
444 * index (relayd index).
446 * It's atomically set without having the stream mutex locked which is fine
447 * because we handle the write/read race with a pipe wakeup for each thread.
449 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
450 enum consumer_endpoint_status status
)
452 struct lttng_ht_iter iter
;
453 struct lttng_consumer_stream
*stream
;
455 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
459 /* Let's begin with metadata */
460 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
461 if (stream
->net_seq_idx
== net_seq_idx
) {
462 uatomic_set(&stream
->endpoint_status
, status
);
463 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
467 /* Follow up by the data streams */
468 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
469 if (stream
->net_seq_idx
== net_seq_idx
) {
470 uatomic_set(&stream
->endpoint_status
, status
);
471 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
478 * Cleanup a relayd object by flagging every associated streams for deletion,
479 * destroying the object meaning removing it from the relayd hash table,
480 * closing the sockets and freeing the memory in a RCU call.
482 * If a local data context is available, notify the threads that the streams'
483 * state have changed.
485 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
491 DBG("Cleaning up relayd object ID %"PRIu64
, relayd
->net_seq_idx
);
493 /* Save the net sequence index before destroying the object */
494 netidx
= relayd
->net_seq_idx
;
497 * Delete the relayd from the relayd hash table, close the sockets and free
498 * the object in a RCU call.
500 consumer_destroy_relayd(relayd
);
502 /* Set inactive endpoint to all streams */
503 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
506 * With a local data context, notify the threads that the streams' state
507 * have changed. The write() action on the pipe acts as an "implicit"
508 * memory barrier ordering the updates of the end point status from the
509 * read of this status which happens AFTER receiving this notify.
511 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
512 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
516 * Flag a relayd socket pair for destruction. Destroy it if the refcount
519 * RCU read side lock MUST be aquired before calling this function.
521 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
525 /* Set destroy flag for this object */
526 uatomic_set(&relayd
->destroy_flag
, 1);
528 /* Destroy the relayd if refcount is 0 */
529 if (uatomic_read(&relayd
->refcount
) == 0) {
530 consumer_destroy_relayd(relayd
);
535 * Completly destroy stream from every visiable data structure and the given
538 * One this call returns, the stream object is not longer usable nor visible.
540 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
543 consumer_stream_destroy(stream
, ht
);
547 * XXX naming of del vs destroy is all mixed up.
549 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
551 consumer_stream_destroy(stream
, data_ht
);
554 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
556 consumer_stream_destroy(stream
, metadata_ht
);
559 void consumer_stream_update_channel_attributes(
560 struct lttng_consumer_stream
*stream
,
561 struct lttng_consumer_channel
*channel
)
563 stream
->channel_read_only_attributes
.tracefile_size
=
564 channel
->tracefile_size
;
567 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
569 const char *channel_name
,
572 struct lttng_trace_chunk
*trace_chunk
,
575 enum consumer_channel_type type
,
576 unsigned int monitor
)
579 struct lttng_consumer_stream
*stream
;
581 stream
= zmalloc(sizeof(*stream
));
582 if (stream
== NULL
) {
583 PERROR("malloc struct lttng_consumer_stream");
588 if (trace_chunk
&& !lttng_trace_chunk_get(trace_chunk
)) {
589 ERR("Failed to acquire trace chunk reference during the creation of a stream");
595 stream
->key
= stream_key
;
596 stream
->trace_chunk
= trace_chunk
;
598 stream
->out_fd_offset
= 0;
599 stream
->output_written
= 0;
600 stream
->net_seq_idx
= relayd_id
;
601 stream
->session_id
= session_id
;
602 stream
->monitor
= monitor
;
603 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
604 stream
->index_file
= NULL
;
605 stream
->last_sequence_number
= -1ULL;
606 pthread_mutex_init(&stream
->lock
, NULL
);
607 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
609 /* If channel is the metadata, flag this stream as metadata. */
610 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
611 stream
->metadata_flag
= 1;
612 /* Metadata is flat out. */
613 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
614 /* Live rendez-vous point. */
615 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
616 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
618 /* Format stream name to <channel_name>_<cpu_number> */
619 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
622 PERROR("snprintf stream name");
627 /* Key is always the wait_fd for streams. */
628 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
630 /* Init node per channel id key */
631 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
633 /* Init session id node with the stream session id */
634 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
636 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
637 " relayd_id %" PRIu64
", session_id %" PRIu64
,
638 stream
->name
, stream
->key
, channel_key
,
639 stream
->net_seq_idx
, stream
->session_id
);
646 lttng_trace_chunk_put(stream
->trace_chunk
);
656 * Add a stream to the global list protected by a mutex.
658 void consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
660 struct lttng_ht
*ht
= data_ht
;
665 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
667 pthread_mutex_lock(&consumer_data
.lock
);
668 pthread_mutex_lock(&stream
->chan
->lock
);
669 pthread_mutex_lock(&stream
->chan
->timer_lock
);
670 pthread_mutex_lock(&stream
->lock
);
673 /* Steal stream identifier to avoid having streams with the same key */
674 steal_stream_key(stream
->key
, ht
);
676 lttng_ht_add_unique_u64(ht
, &stream
->node
);
678 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
679 &stream
->node_channel_id
);
682 * Add stream to the stream_list_ht of the consumer data. No need to steal
683 * the key since the HT does not use it and we allow to add redundant keys
686 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
689 * When nb_init_stream_left reaches 0, we don't need to trigger any action
690 * in terms of destroying the associated channel, because the action that
691 * causes the count to become 0 also causes a stream to be added. The
692 * channel deletion will thus be triggered by the following removal of this
695 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
696 /* Increment refcount before decrementing nb_init_stream_left */
698 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
701 /* Update consumer data once the node is inserted. */
702 consumer_data
.stream_count
++;
703 consumer_data
.need_update
= 1;
706 pthread_mutex_unlock(&stream
->lock
);
707 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
708 pthread_mutex_unlock(&stream
->chan
->lock
);
709 pthread_mutex_unlock(&consumer_data
.lock
);
712 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
714 consumer_del_stream(stream
, data_ht
);
718 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
719 * be acquired before calling this.
721 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
724 struct lttng_ht_node_u64
*node
;
725 struct lttng_ht_iter iter
;
729 lttng_ht_lookup(consumer_data
.relayd_ht
,
730 &relayd
->net_seq_idx
, &iter
);
731 node
= lttng_ht_iter_get_node_u64(&iter
);
735 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
742 * Allocate and return a consumer relayd socket.
744 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
745 uint64_t net_seq_idx
)
747 struct consumer_relayd_sock_pair
*obj
= NULL
;
749 /* net sequence index of -1 is a failure */
750 if (net_seq_idx
== (uint64_t) -1ULL) {
754 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
756 PERROR("zmalloc relayd sock");
760 obj
->net_seq_idx
= net_seq_idx
;
762 obj
->destroy_flag
= 0;
763 obj
->control_sock
.sock
.fd
= -1;
764 obj
->data_sock
.sock
.fd
= -1;
765 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
766 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
773 * Find a relayd socket pair in the global consumer data.
775 * Return the object if found else NULL.
776 * RCU read-side lock must be held across this call and while using the
779 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
781 struct lttng_ht_iter iter
;
782 struct lttng_ht_node_u64
*node
;
783 struct consumer_relayd_sock_pair
*relayd
= NULL
;
785 /* Negative keys are lookup failures */
786 if (key
== (uint64_t) -1ULL) {
790 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
792 node
= lttng_ht_iter_get_node_u64(&iter
);
794 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
802 * Find a relayd and send the stream
804 * Returns 0 on success, < 0 on error
806 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
810 struct consumer_relayd_sock_pair
*relayd
;
813 assert(stream
->net_seq_idx
!= -1ULL);
816 /* The stream is not metadata. Get relayd reference if exists. */
818 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
819 if (relayd
!= NULL
) {
820 /* Add stream on the relayd */
821 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
822 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
823 path
, &stream
->relayd_stream_id
,
824 stream
->chan
->tracefile_size
,
825 stream
->chan
->tracefile_count
,
826 stream
->trace_chunk
);
827 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
829 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
830 lttng_consumer_cleanup_relayd(relayd
);
834 uatomic_inc(&relayd
->refcount
);
835 stream
->sent_to_relayd
= 1;
837 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
838 stream
->key
, stream
->net_seq_idx
);
843 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
844 stream
->name
, stream
->key
, stream
->net_seq_idx
);
852 * Find a relayd and send the streams sent message
854 * Returns 0 on success, < 0 on error
856 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
859 struct consumer_relayd_sock_pair
*relayd
;
861 assert(net_seq_idx
!= -1ULL);
863 /* The stream is not metadata. Get relayd reference if exists. */
865 relayd
= consumer_find_relayd(net_seq_idx
);
866 if (relayd
!= NULL
) {
867 /* Add stream on the relayd */
868 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
869 ret
= relayd_streams_sent(&relayd
->control_sock
);
870 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
872 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
873 lttng_consumer_cleanup_relayd(relayd
);
877 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
884 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
892 * Find a relayd and close the stream
894 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
896 struct consumer_relayd_sock_pair
*relayd
;
898 /* The stream is not metadata. Get relayd reference if exists. */
900 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
902 consumer_stream_relayd_close(stream
, relayd
);
908 * Handle stream for relayd transmission if the stream applies for network
909 * streaming where the net sequence index is set.
911 * Return destination file descriptor or negative value on error.
913 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
914 size_t data_size
, unsigned long padding
,
915 struct consumer_relayd_sock_pair
*relayd
)
918 struct lttcomm_relayd_data_hdr data_hdr
;
924 /* Reset data header */
925 memset(&data_hdr
, 0, sizeof(data_hdr
));
927 if (stream
->metadata_flag
) {
928 /* Caller MUST acquire the relayd control socket lock */
929 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
934 /* Metadata are always sent on the control socket. */
935 outfd
= relayd
->control_sock
.sock
.fd
;
937 /* Set header with stream information */
938 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
939 data_hdr
.data_size
= htobe32(data_size
);
940 data_hdr
.padding_size
= htobe32(padding
);
943 * Note that net_seq_num below is assigned with the *current* value of
944 * next_net_seq_num and only after that the next_net_seq_num will be
945 * increment. This is why when issuing a command on the relayd using
946 * this next value, 1 should always be substracted in order to compare
947 * the last seen sequence number on the relayd side to the last sent.
949 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
950 /* Other fields are zeroed previously */
952 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
958 ++stream
->next_net_seq_num
;
960 /* Set to go on data socket */
961 outfd
= relayd
->data_sock
.sock
.fd
;
969 * Trigger a dump of the metadata content. Following/during the succesful
970 * completion of this call, the metadata poll thread will start receiving
971 * metadata packets to consume.
973 * The caller must hold the channel and stream locks.
976 int consumer_metadata_stream_dump(struct lttng_consumer_stream
*stream
)
980 ASSERT_LOCKED(stream
->chan
->lock
);
981 ASSERT_LOCKED(stream
->lock
);
982 assert(stream
->metadata_flag
);
983 assert(stream
->chan
->trace_chunk
);
985 switch (consumer_data
.type
) {
986 case LTTNG_CONSUMER_KERNEL
:
988 * Reset the position of what has been read from the
989 * metadata cache to 0 so we can dump it again.
991 ret
= kernctl_metadata_cache_dump(stream
->wait_fd
);
993 case LTTNG_CONSUMER32_UST
:
994 case LTTNG_CONSUMER64_UST
:
996 * Reset the position pushed from the metadata cache so it
997 * will write from the beginning on the next push.
999 stream
->ust_metadata_pushed
= 0;
1000 ret
= consumer_metadata_wakeup_pipe(stream
->chan
);
1003 ERR("Unknown consumer_data type");
1007 ERR("Failed to dump the metadata cache");
1013 int lttng_consumer_channel_set_trace_chunk(
1014 struct lttng_consumer_channel
*channel
,
1015 struct lttng_trace_chunk
*new_trace_chunk
)
1017 pthread_mutex_lock(&channel
->lock
);
1018 if (channel
->is_deleted
) {
1020 * The channel has been logically deleted and should no longer
1021 * be used. It has released its reference to its current trace
1022 * chunk and should not acquire a new one.
1024 * Return success as there is nothing for the caller to do.
1030 * The acquisition of the reference cannot fail (barring
1031 * a severe internal error) since a reference to the published
1032 * chunk is already held by the caller.
1034 if (new_trace_chunk
) {
1035 const bool acquired_reference
= lttng_trace_chunk_get(
1038 assert(acquired_reference
);
1041 lttng_trace_chunk_put(channel
->trace_chunk
);
1042 channel
->trace_chunk
= new_trace_chunk
;
1044 pthread_mutex_unlock(&channel
->lock
);
1049 * Allocate and return a new lttng_consumer_channel object using the given key
1050 * to initialize the hash table node.
1052 * On error, return NULL.
1054 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
1055 uint64_t session_id
,
1056 const uint64_t *chunk_id
,
1057 const char *pathname
,
1060 enum lttng_event_output output
,
1061 uint64_t tracefile_size
,
1062 uint64_t tracefile_count
,
1063 uint64_t session_id_per_pid
,
1064 unsigned int monitor
,
1065 unsigned int live_timer_interval
,
1066 const char *root_shm_path
,
1067 const char *shm_path
)
1069 struct lttng_consumer_channel
*channel
= NULL
;
1070 struct lttng_trace_chunk
*trace_chunk
= NULL
;
1073 trace_chunk
= lttng_trace_chunk_registry_find_chunk(
1074 consumer_data
.chunk_registry
, session_id
,
1077 ERR("Failed to find trace chunk reference during creation of channel");
1082 channel
= zmalloc(sizeof(*channel
));
1083 if (channel
== NULL
) {
1084 PERROR("malloc struct lttng_consumer_channel");
1089 channel
->refcount
= 0;
1090 channel
->session_id
= session_id
;
1091 channel
->session_id_per_pid
= session_id_per_pid
;
1092 channel
->relayd_id
= relayd_id
;
1093 channel
->tracefile_size
= tracefile_size
;
1094 channel
->tracefile_count
= tracefile_count
;
1095 channel
->monitor
= monitor
;
1096 channel
->live_timer_interval
= live_timer_interval
;
1097 pthread_mutex_init(&channel
->lock
, NULL
);
1098 pthread_mutex_init(&channel
->timer_lock
, NULL
);
1101 case LTTNG_EVENT_SPLICE
:
1102 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
1104 case LTTNG_EVENT_MMAP
:
1105 channel
->output
= CONSUMER_CHANNEL_MMAP
;
1115 * In monitor mode, the streams associated with the channel will be put in
1116 * a special list ONLY owned by this channel. So, the refcount is set to 1
1117 * here meaning that the channel itself has streams that are referenced.
1119 * On a channel deletion, once the channel is no longer visible, the
1120 * refcount is decremented and checked for a zero value to delete it. With
1121 * streams in no monitor mode, it will now be safe to destroy the channel.
1123 if (!channel
->monitor
) {
1124 channel
->refcount
= 1;
1127 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1128 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1130 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1131 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1133 if (root_shm_path
) {
1134 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1135 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1138 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1139 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1142 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1143 lttng_ht_node_init_u64(&channel
->channels_by_session_id_ht_node
,
1144 channel
->session_id
);
1146 channel
->wait_fd
= -1;
1147 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1150 int ret
= lttng_consumer_channel_set_trace_chunk(channel
,
1157 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1160 lttng_trace_chunk_put(trace_chunk
);
1163 consumer_del_channel(channel
);
1169 * Add a channel to the global list protected by a mutex.
1171 * Always return 0 indicating success.
1173 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1174 struct lttng_consumer_local_data
*ctx
)
1176 pthread_mutex_lock(&consumer_data
.lock
);
1177 pthread_mutex_lock(&channel
->lock
);
1178 pthread_mutex_lock(&channel
->timer_lock
);
1181 * This gives us a guarantee that the channel we are about to add to the
1182 * channel hash table will be unique. See this function comment on the why
1183 * we need to steel the channel key at this stage.
1185 steal_channel_key(channel
->key
);
1188 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1189 lttng_ht_add_u64(consumer_data
.channels_by_session_id_ht
,
1190 &channel
->channels_by_session_id_ht_node
);
1192 channel
->is_published
= true;
1194 pthread_mutex_unlock(&channel
->timer_lock
);
1195 pthread_mutex_unlock(&channel
->lock
);
1196 pthread_mutex_unlock(&consumer_data
.lock
);
1198 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1199 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1206 * Allocate the pollfd structure and the local view of the out fds to avoid
1207 * doing a lookup in the linked list and concurrency issues when writing is
1208 * needed. Called with consumer_data.lock held.
1210 * Returns the number of fds in the structures.
1212 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1213 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1214 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1217 struct lttng_ht_iter iter
;
1218 struct lttng_consumer_stream
*stream
;
1223 assert(local_stream
);
1225 DBG("Updating poll fd array");
1226 *nb_inactive_fd
= 0;
1228 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1230 * Only active streams with an active end point can be added to the
1231 * poll set and local stream storage of the thread.
1233 * There is a potential race here for endpoint_status to be updated
1234 * just after the check. However, this is OK since the stream(s) will
1235 * be deleted once the thread is notified that the end point state has
1236 * changed where this function will be called back again.
1238 * We track the number of inactive FDs because they still need to be
1239 * closed by the polling thread after a wakeup on the data_pipe or
1242 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1243 (*nb_inactive_fd
)++;
1247 * This clobbers way too much the debug output. Uncomment that if you
1248 * need it for debugging purposes.
1250 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1251 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1252 local_stream
[i
] = stream
;
1258 * Insert the consumer_data_pipe at the end of the array and don't
1259 * increment i so nb_fd is the number of real FD.
1261 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1262 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1264 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1265 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1270 * Poll on the should_quit pipe and the command socket return -1 on
1271 * error, 1 if should exit, 0 if data is available on the command socket
1273 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1278 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1279 if (num_rdy
== -1) {
1281 * Restart interrupted system call.
1283 if (errno
== EINTR
) {
1286 PERROR("Poll error");
1289 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1290 DBG("consumer_should_quit wake up");
1297 * Set the error socket.
1299 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1302 ctx
->consumer_error_socket
= sock
;
1306 * Set the command socket path.
1308 void lttng_consumer_set_command_sock_path(
1309 struct lttng_consumer_local_data
*ctx
, char *sock
)
1311 ctx
->consumer_command_sock_path
= sock
;
1315 * Send return code to the session daemon.
1316 * If the socket is not defined, we return 0, it is not a fatal error
1318 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1320 if (ctx
->consumer_error_socket
> 0) {
1321 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1322 sizeof(enum lttcomm_sessiond_command
));
1329 * Close all the tracefiles and stream fds and MUST be called when all
1330 * instances are destroyed i.e. when all threads were joined and are ended.
1332 void lttng_consumer_cleanup(void)
1334 struct lttng_ht_iter iter
;
1335 struct lttng_consumer_channel
*channel
;
1336 unsigned int trace_chunks_left
;
1340 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1342 consumer_del_channel(channel
);
1347 lttng_ht_destroy(consumer_data
.channel_ht
);
1348 lttng_ht_destroy(consumer_data
.channels_by_session_id_ht
);
1350 cleanup_relayd_ht();
1352 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1355 * This HT contains streams that are freed by either the metadata thread or
1356 * the data thread so we do *nothing* on the hash table and simply destroy
1359 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1362 * Trace chunks in the registry may still exist if the session
1363 * daemon has encountered an internal error and could not
1364 * tear down its sessions and/or trace chunks properly.
1366 * Release the session daemon's implicit reference to any remaining
1367 * trace chunk and print an error if any trace chunk was found. Note
1368 * that there are _no_ legitimate cases for trace chunks to be left,
1369 * it is a leak. However, it can happen following a crash of the
1370 * session daemon and not emptying the registry would cause an assertion
1373 trace_chunks_left
= lttng_trace_chunk_registry_put_each_chunk(
1374 consumer_data
.chunk_registry
);
1375 if (trace_chunks_left
) {
1376 ERR("%u trace chunks are leaked by lttng-consumerd. "
1377 "This can be caused by an internal error of the session daemon.",
1380 /* Run all callbacks freeing each chunk. */
1382 lttng_trace_chunk_registry_destroy(consumer_data
.chunk_registry
);
1386 * Called from signal handler.
1388 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1392 CMM_STORE_SHARED(consumer_quit
, 1);
1393 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1395 PERROR("write consumer quit");
1398 DBG("Consumer flag that it should quit");
1403 * Flush pending writes to trace output disk file.
1406 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1410 int outfd
= stream
->out_fd
;
1413 * This does a blocking write-and-wait on any page that belongs to the
1414 * subbuffer prior to the one we just wrote.
1415 * Don't care about error values, as these are just hints and ways to
1416 * limit the amount of page cache used.
1418 if (orig_offset
< stream
->max_sb_size
) {
1421 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1422 stream
->max_sb_size
,
1423 SYNC_FILE_RANGE_WAIT_BEFORE
1424 | SYNC_FILE_RANGE_WRITE
1425 | SYNC_FILE_RANGE_WAIT_AFTER
);
1427 * Give hints to the kernel about how we access the file:
1428 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1431 * We need to call fadvise again after the file grows because the
1432 * kernel does not seem to apply fadvise to non-existing parts of the
1435 * Call fadvise _after_ having waited for the page writeback to
1436 * complete because the dirty page writeback semantic is not well
1437 * defined. So it can be expected to lead to lower throughput in
1440 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1441 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1442 if (ret
&& ret
!= -ENOSYS
) {
1444 PERROR("posix_fadvise on fd %i", outfd
);
1449 * Initialise the necessary environnement :
1450 * - create a new context
1451 * - create the poll_pipe
1452 * - create the should_quit pipe (for signal handler)
1453 * - create the thread pipe (for splice)
1455 * Takes a function pointer as argument, this function is called when data is
1456 * available on a buffer. This function is responsible to do the
1457 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1458 * buffer configuration and then kernctl_put_next_subbuf at the end.
1460 * Returns a pointer to the new context or NULL on error.
1462 struct lttng_consumer_local_data
*lttng_consumer_create(
1463 enum lttng_consumer_type type
,
1464 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1465 struct lttng_consumer_local_data
*ctx
),
1466 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1467 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1468 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1471 struct lttng_consumer_local_data
*ctx
;
1473 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1474 consumer_data
.type
== type
);
1475 consumer_data
.type
= type
;
1477 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1479 PERROR("allocating context");
1483 ctx
->consumer_error_socket
= -1;
1484 ctx
->consumer_metadata_socket
= -1;
1485 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1486 /* assign the callbacks */
1487 ctx
->on_buffer_ready
= buffer_ready
;
1488 ctx
->on_recv_channel
= recv_channel
;
1489 ctx
->on_recv_stream
= recv_stream
;
1490 ctx
->on_update_stream
= update_stream
;
1492 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1493 if (!ctx
->consumer_data_pipe
) {
1494 goto error_poll_pipe
;
1497 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1498 if (!ctx
->consumer_wakeup_pipe
) {
1499 goto error_wakeup_pipe
;
1502 ret
= pipe(ctx
->consumer_should_quit
);
1504 PERROR("Error creating recv pipe");
1505 goto error_quit_pipe
;
1508 ret
= pipe(ctx
->consumer_channel_pipe
);
1510 PERROR("Error creating channel pipe");
1511 goto error_channel_pipe
;
1514 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1515 if (!ctx
->consumer_metadata_pipe
) {
1516 goto error_metadata_pipe
;
1519 ctx
->channel_monitor_pipe
= -1;
1523 error_metadata_pipe
:
1524 utils_close_pipe(ctx
->consumer_channel_pipe
);
1526 utils_close_pipe(ctx
->consumer_should_quit
);
1528 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1530 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1538 * Iterate over all streams of the hashtable and free them properly.
1540 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1542 struct lttng_ht_iter iter
;
1543 struct lttng_consumer_stream
*stream
;
1550 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1552 * Ignore return value since we are currently cleaning up so any error
1555 (void) consumer_del_stream(stream
, ht
);
1559 lttng_ht_destroy(ht
);
1563 * Iterate over all streams of the metadata hashtable and free them
1566 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1568 struct lttng_ht_iter iter
;
1569 struct lttng_consumer_stream
*stream
;
1576 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1578 * Ignore return value since we are currently cleaning up so any error
1581 (void) consumer_del_metadata_stream(stream
, ht
);
1585 lttng_ht_destroy(ht
);
1589 * Close all fds associated with the instance and free the context.
1591 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1595 DBG("Consumer destroying it. Closing everything.");
1601 destroy_data_stream_ht(data_ht
);
1602 destroy_metadata_stream_ht(metadata_ht
);
1604 ret
= close(ctx
->consumer_error_socket
);
1608 ret
= close(ctx
->consumer_metadata_socket
);
1612 utils_close_pipe(ctx
->consumer_channel_pipe
);
1613 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1614 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1615 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1616 utils_close_pipe(ctx
->consumer_should_quit
);
1618 unlink(ctx
->consumer_command_sock_path
);
1623 * Write the metadata stream id on the specified file descriptor.
1625 static int write_relayd_metadata_id(int fd
,
1626 struct lttng_consumer_stream
*stream
,
1627 unsigned long padding
)
1630 struct lttcomm_relayd_metadata_payload hdr
;
1632 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1633 hdr
.padding_size
= htobe32(padding
);
1634 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1635 if (ret
< sizeof(hdr
)) {
1637 * This error means that the fd's end is closed so ignore the PERROR
1638 * not to clubber the error output since this can happen in a normal
1641 if (errno
!= EPIPE
) {
1642 PERROR("write metadata stream id");
1644 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1646 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1647 * handle writting the missing part so report that as an error and
1648 * don't lie to the caller.
1653 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1654 stream
->relayd_stream_id
, padding
);
1661 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1662 * core function for writing trace buffers to either the local filesystem or
1665 * It must be called with the stream and the channel lock held.
1667 * Careful review MUST be put if any changes occur!
1669 * Returns the number of bytes written
1671 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1672 struct lttng_consumer_local_data
*ctx
,
1673 struct lttng_consumer_stream
*stream
, unsigned long len
,
1674 unsigned long padding
,
1675 struct ctf_packet_index
*index
)
1677 unsigned long mmap_offset
;
1680 off_t orig_offset
= stream
->out_fd_offset
;
1681 /* Default is on the disk */
1682 int outfd
= stream
->out_fd
;
1683 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1684 unsigned int relayd_hang_up
= 0;
1686 /* RCU lock for the relayd pointer */
1688 assert(stream
->net_seq_idx
!= (uint64_t) -1ULL ||
1689 stream
->trace_chunk
);
1691 /* Flag that the current stream if set for network streaming. */
1692 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1693 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1694 if (relayd
== NULL
) {
1700 /* get the offset inside the fd to mmap */
1701 switch (consumer_data
.type
) {
1702 case LTTNG_CONSUMER_KERNEL
:
1703 mmap_base
= stream
->mmap_base
;
1704 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1706 PERROR("tracer ctl get_mmap_read_offset");
1710 case LTTNG_CONSUMER32_UST
:
1711 case LTTNG_CONSUMER64_UST
:
1712 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1714 ERR("read mmap get mmap base for stream %s", stream
->name
);
1718 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1720 PERROR("tracer ctl get_mmap_read_offset");
1726 ERR("Unknown consumer_data type");
1730 /* Handle stream on the relayd if the output is on the network */
1732 unsigned long netlen
= len
;
1735 * Lock the control socket for the complete duration of the function
1736 * since from this point on we will use the socket.
1738 if (stream
->metadata_flag
) {
1739 /* Metadata requires the control socket. */
1740 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1741 if (stream
->reset_metadata_flag
) {
1742 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1743 stream
->relayd_stream_id
,
1744 stream
->metadata_version
);
1749 stream
->reset_metadata_flag
= 0;
1751 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1754 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1759 /* Use the returned socket. */
1762 /* Write metadata stream id before payload */
1763 if (stream
->metadata_flag
) {
1764 ret
= write_relayd_metadata_id(outfd
, stream
, padding
);
1771 /* No streaming, we have to set the len with the full padding */
1774 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1775 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1777 ERR("Reset metadata file");
1780 stream
->reset_metadata_flag
= 0;
1784 * Check if we need to change the tracefile before writing the packet.
1786 if (stream
->chan
->tracefile_size
> 0 &&
1787 (stream
->tracefile_size_current
+ len
) >
1788 stream
->chan
->tracefile_size
) {
1789 ret
= consumer_stream_rotate_output_files(stream
);
1793 outfd
= stream
->out_fd
;
1796 stream
->tracefile_size_current
+= len
;
1798 index
->offset
= htobe64(stream
->out_fd_offset
);
1803 * This call guarantee that len or less is returned. It's impossible to
1804 * receive a ret value that is bigger than len.
1806 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1807 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1808 if (ret
< 0 || ((size_t) ret
!= len
)) {
1810 * Report error to caller if nothing was written else at least send the
1818 /* Socket operation failed. We consider the relayd dead */
1819 if (errno
== EPIPE
) {
1821 * This is possible if the fd is closed on the other side
1822 * (outfd) or any write problem. It can be verbose a bit for a
1823 * normal execution if for instance the relayd is stopped
1824 * abruptly. This can happen so set this to a DBG statement.
1826 DBG("Consumer mmap write detected relayd hang up");
1828 /* Unhandled error, print it and stop function right now. */
1829 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1833 stream
->output_written
+= ret
;
1835 /* This call is useless on a socket so better save a syscall. */
1837 /* This won't block, but will start writeout asynchronously */
1838 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1839 SYNC_FILE_RANGE_WRITE
);
1840 stream
->out_fd_offset
+= len
;
1841 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1846 * This is a special case that the relayd has closed its socket. Let's
1847 * cleanup the relayd object and all associated streams.
1849 if (relayd
&& relayd_hang_up
) {
1850 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1851 lttng_consumer_cleanup_relayd(relayd
);
1855 /* Unlock only if ctrl socket used */
1856 if (relayd
&& stream
->metadata_flag
) {
1857 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1865 * Splice the data from the ring buffer to the tracefile.
1867 * It must be called with the stream lock held.
1869 * Returns the number of bytes spliced.
1871 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1872 struct lttng_consumer_local_data
*ctx
,
1873 struct lttng_consumer_stream
*stream
, unsigned long len
,
1874 unsigned long padding
,
1875 struct ctf_packet_index
*index
)
1877 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1879 off_t orig_offset
= stream
->out_fd_offset
;
1880 int fd
= stream
->wait_fd
;
1881 /* Default is on the disk */
1882 int outfd
= stream
->out_fd
;
1883 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1885 unsigned int relayd_hang_up
= 0;
1887 switch (consumer_data
.type
) {
1888 case LTTNG_CONSUMER_KERNEL
:
1890 case LTTNG_CONSUMER32_UST
:
1891 case LTTNG_CONSUMER64_UST
:
1892 /* Not supported for user space tracing */
1895 ERR("Unknown consumer_data type");
1899 /* RCU lock for the relayd pointer */
1902 /* Flag that the current stream if set for network streaming. */
1903 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1904 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1905 if (relayd
== NULL
) {
1910 splice_pipe
= stream
->splice_pipe
;
1912 /* Write metadata stream id before payload */
1914 unsigned long total_len
= len
;
1916 if (stream
->metadata_flag
) {
1918 * Lock the control socket for the complete duration of the function
1919 * since from this point on we will use the socket.
1921 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1923 if (stream
->reset_metadata_flag
) {
1924 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1925 stream
->relayd_stream_id
,
1926 stream
->metadata_version
);
1931 stream
->reset_metadata_flag
= 0;
1933 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
,
1941 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1944 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1950 /* Use the returned socket. */
1953 /* No streaming, we have to set the len with the full padding */
1956 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1957 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1959 ERR("Reset metadata file");
1962 stream
->reset_metadata_flag
= 0;
1965 * Check if we need to change the tracefile before writing the packet.
1967 if (stream
->chan
->tracefile_size
> 0 &&
1968 (stream
->tracefile_size_current
+ len
) >
1969 stream
->chan
->tracefile_size
) {
1970 ret
= consumer_stream_rotate_output_files(stream
);
1975 outfd
= stream
->out_fd
;
1978 stream
->tracefile_size_current
+= len
;
1979 index
->offset
= htobe64(stream
->out_fd_offset
);
1983 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1984 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1985 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1986 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1987 DBG("splice chan to pipe, ret %zd", ret_splice
);
1988 if (ret_splice
< 0) {
1991 PERROR("Error in relay splice");
1995 /* Handle stream on the relayd if the output is on the network */
1996 if (relayd
&& stream
->metadata_flag
) {
1997 size_t metadata_payload_size
=
1998 sizeof(struct lttcomm_relayd_metadata_payload
);
2000 /* Update counter to fit the spliced data */
2001 ret_splice
+= metadata_payload_size
;
2002 len
+= metadata_payload_size
;
2004 * We do this so the return value can match the len passed as
2005 * argument to this function.
2007 written
-= metadata_payload_size
;
2010 /* Splice data out */
2011 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
2012 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
2013 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
2015 if (ret_splice
< 0) {
2020 } else if (ret_splice
> len
) {
2022 * We don't expect this code path to be executed but you never know
2023 * so this is an extra protection agains a buggy splice().
2026 written
+= ret_splice
;
2027 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
2031 /* All good, update current len and continue. */
2035 /* This call is useless on a socket so better save a syscall. */
2037 /* This won't block, but will start writeout asynchronously */
2038 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
2039 SYNC_FILE_RANGE_WRITE
);
2040 stream
->out_fd_offset
+= ret_splice
;
2042 stream
->output_written
+= ret_splice
;
2043 written
+= ret_splice
;
2046 lttng_consumer_sync_trace_file(stream
, orig_offset
);
2052 * This is a special case that the relayd has closed its socket. Let's
2053 * cleanup the relayd object and all associated streams.
2055 if (relayd
&& relayd_hang_up
) {
2056 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
2057 lttng_consumer_cleanup_relayd(relayd
);
2058 /* Skip splice error so the consumer does not fail */
2063 /* send the appropriate error description to sessiond */
2066 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
2069 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
2072 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
2077 if (relayd
&& stream
->metadata_flag
) {
2078 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2086 * Sample the snapshot positions for a specific fd
2088 * Returns 0 on success, < 0 on error
2090 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream
*stream
)
2092 switch (consumer_data
.type
) {
2093 case LTTNG_CONSUMER_KERNEL
:
2094 return lttng_kconsumer_sample_snapshot_positions(stream
);
2095 case LTTNG_CONSUMER32_UST
:
2096 case LTTNG_CONSUMER64_UST
:
2097 return lttng_ustconsumer_sample_snapshot_positions(stream
);
2099 ERR("Unknown consumer_data type");
2105 * Take a snapshot for a specific fd
2107 * Returns 0 on success, < 0 on error
2109 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
2111 switch (consumer_data
.type
) {
2112 case LTTNG_CONSUMER_KERNEL
:
2113 return lttng_kconsumer_take_snapshot(stream
);
2114 case LTTNG_CONSUMER32_UST
:
2115 case LTTNG_CONSUMER64_UST
:
2116 return lttng_ustconsumer_take_snapshot(stream
);
2118 ERR("Unknown consumer_data type");
2125 * Get the produced position
2127 * Returns 0 on success, < 0 on error
2129 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
2132 switch (consumer_data
.type
) {
2133 case LTTNG_CONSUMER_KERNEL
:
2134 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
2135 case LTTNG_CONSUMER32_UST
:
2136 case LTTNG_CONSUMER64_UST
:
2137 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2139 ERR("Unknown consumer_data type");
2146 * Get the consumed position (free-running counter position in bytes).
2148 * Returns 0 on success, < 0 on error
2150 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream
*stream
,
2153 switch (consumer_data
.type
) {
2154 case LTTNG_CONSUMER_KERNEL
:
2155 return lttng_kconsumer_get_consumed_snapshot(stream
, pos
);
2156 case LTTNG_CONSUMER32_UST
:
2157 case LTTNG_CONSUMER64_UST
:
2158 return lttng_ustconsumer_get_consumed_snapshot(stream
, pos
);
2160 ERR("Unknown consumer_data type");
2166 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2167 int sock
, struct pollfd
*consumer_sockpoll
)
2169 switch (consumer_data
.type
) {
2170 case LTTNG_CONSUMER_KERNEL
:
2171 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2172 case LTTNG_CONSUMER32_UST
:
2173 case LTTNG_CONSUMER64_UST
:
2174 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2176 ERR("Unknown consumer_data type");
2182 void lttng_consumer_close_all_metadata(void)
2184 switch (consumer_data
.type
) {
2185 case LTTNG_CONSUMER_KERNEL
:
2187 * The Kernel consumer has a different metadata scheme so we don't
2188 * close anything because the stream will be closed by the session
2192 case LTTNG_CONSUMER32_UST
:
2193 case LTTNG_CONSUMER64_UST
:
2195 * Close all metadata streams. The metadata hash table is passed and
2196 * this call iterates over it by closing all wakeup fd. This is safe
2197 * because at this point we are sure that the metadata producer is
2198 * either dead or blocked.
2200 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2203 ERR("Unknown consumer_data type");
2209 * Clean up a metadata stream and free its memory.
2211 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2212 struct lttng_ht
*ht
)
2214 struct lttng_consumer_channel
*channel
= NULL
;
2215 bool free_channel
= false;
2219 * This call should NEVER receive regular stream. It must always be
2220 * metadata stream and this is crucial for data structure synchronization.
2222 assert(stream
->metadata_flag
);
2224 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2226 pthread_mutex_lock(&consumer_data
.lock
);
2228 * Note that this assumes that a stream's channel is never changed and
2229 * that the stream's lock doesn't need to be taken to sample its
2232 channel
= stream
->chan
;
2233 pthread_mutex_lock(&channel
->lock
);
2234 pthread_mutex_lock(&stream
->lock
);
2235 if (channel
->metadata_cache
) {
2236 /* Only applicable to userspace consumers. */
2237 pthread_mutex_lock(&channel
->metadata_cache
->lock
);
2240 /* Remove any reference to that stream. */
2241 consumer_stream_delete(stream
, ht
);
2243 /* Close down everything including the relayd if one. */
2244 consumer_stream_close(stream
);
2245 /* Destroy tracer buffers of the stream. */
2246 consumer_stream_destroy_buffers(stream
);
2248 /* Atomically decrement channel refcount since other threads can use it. */
2249 if (!uatomic_sub_return(&channel
->refcount
, 1)
2250 && !uatomic_read(&channel
->nb_init_stream_left
)) {
2251 /* Go for channel deletion! */
2252 free_channel
= true;
2254 stream
->chan
= NULL
;
2257 * Nullify the stream reference so it is not used after deletion. The
2258 * channel lock MUST be acquired before being able to check for a NULL
2261 channel
->metadata_stream
= NULL
;
2263 if (channel
->metadata_cache
) {
2264 pthread_mutex_unlock(&channel
->metadata_cache
->lock
);
2266 pthread_mutex_unlock(&stream
->lock
);
2267 pthread_mutex_unlock(&channel
->lock
);
2268 pthread_mutex_unlock(&consumer_data
.lock
);
2271 consumer_del_channel(channel
);
2274 lttng_trace_chunk_put(stream
->trace_chunk
);
2275 stream
->trace_chunk
= NULL
;
2276 consumer_stream_free(stream
);
2280 * Action done with the metadata stream when adding it to the consumer internal
2281 * data structures to handle it.
2283 void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2285 struct lttng_ht
*ht
= metadata_ht
;
2286 struct lttng_ht_iter iter
;
2287 struct lttng_ht_node_u64
*node
;
2292 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2294 pthread_mutex_lock(&consumer_data
.lock
);
2295 pthread_mutex_lock(&stream
->chan
->lock
);
2296 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2297 pthread_mutex_lock(&stream
->lock
);
2300 * From here, refcounts are updated so be _careful_ when returning an error
2307 * Lookup the stream just to make sure it does not exist in our internal
2308 * state. This should NEVER happen.
2310 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2311 node
= lttng_ht_iter_get_node_u64(&iter
);
2315 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2316 * in terms of destroying the associated channel, because the action that
2317 * causes the count to become 0 also causes a stream to be added. The
2318 * channel deletion will thus be triggered by the following removal of this
2321 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2322 /* Increment refcount before decrementing nb_init_stream_left */
2324 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2327 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2329 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2330 &stream
->node_channel_id
);
2333 * Add stream to the stream_list_ht of the consumer data. No need to steal
2334 * the key since the HT does not use it and we allow to add redundant keys
2337 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2341 pthread_mutex_unlock(&stream
->lock
);
2342 pthread_mutex_unlock(&stream
->chan
->lock
);
2343 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2344 pthread_mutex_unlock(&consumer_data
.lock
);
2348 * Delete data stream that are flagged for deletion (endpoint_status).
2350 static void validate_endpoint_status_data_stream(void)
2352 struct lttng_ht_iter iter
;
2353 struct lttng_consumer_stream
*stream
;
2355 DBG("Consumer delete flagged data stream");
2358 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2359 /* Validate delete flag of the stream */
2360 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2363 /* Delete it right now */
2364 consumer_del_stream(stream
, data_ht
);
2370 * Delete metadata stream that are flagged for deletion (endpoint_status).
2372 static void validate_endpoint_status_metadata_stream(
2373 struct lttng_poll_event
*pollset
)
2375 struct lttng_ht_iter iter
;
2376 struct lttng_consumer_stream
*stream
;
2378 DBG("Consumer delete flagged metadata stream");
2383 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2384 /* Validate delete flag of the stream */
2385 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2389 * Remove from pollset so the metadata thread can continue without
2390 * blocking on a deleted stream.
2392 lttng_poll_del(pollset
, stream
->wait_fd
);
2394 /* Delete it right now */
2395 consumer_del_metadata_stream(stream
, metadata_ht
);
2401 * Thread polls on metadata file descriptor and write them on disk or on the
2404 void *consumer_thread_metadata_poll(void *data
)
2406 int ret
, i
, pollfd
, err
= -1;
2407 uint32_t revents
, nb_fd
;
2408 struct lttng_consumer_stream
*stream
= NULL
;
2409 struct lttng_ht_iter iter
;
2410 struct lttng_ht_node_u64
*node
;
2411 struct lttng_poll_event events
;
2412 struct lttng_consumer_local_data
*ctx
= data
;
2415 rcu_register_thread();
2417 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2419 if (testpoint(consumerd_thread_metadata
)) {
2420 goto error_testpoint
;
2423 health_code_update();
2425 DBG("Thread metadata poll started");
2427 /* Size is set to 1 for the consumer_metadata pipe */
2428 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2430 ERR("Poll set creation failed");
2434 ret
= lttng_poll_add(&events
,
2435 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2441 DBG("Metadata main loop started");
2445 health_code_update();
2446 health_poll_entry();
2447 DBG("Metadata poll wait");
2448 ret
= lttng_poll_wait(&events
, -1);
2449 DBG("Metadata poll return from wait with %d fd(s)",
2450 LTTNG_POLL_GETNB(&events
));
2452 DBG("Metadata event caught in thread");
2454 if (errno
== EINTR
) {
2455 ERR("Poll EINTR caught");
2458 if (LTTNG_POLL_GETNB(&events
) == 0) {
2459 err
= 0; /* All is OK */
2466 /* From here, the event is a metadata wait fd */
2467 for (i
= 0; i
< nb_fd
; i
++) {
2468 health_code_update();
2470 revents
= LTTNG_POLL_GETEV(&events
, i
);
2471 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2473 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2474 if (revents
& LPOLLIN
) {
2477 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2478 &stream
, sizeof(stream
));
2479 if (pipe_len
< sizeof(stream
)) {
2481 PERROR("read metadata stream");
2484 * Remove the pipe from the poll set and continue the loop
2485 * since their might be data to consume.
2487 lttng_poll_del(&events
,
2488 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2489 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2493 /* A NULL stream means that the state has changed. */
2494 if (stream
== NULL
) {
2495 /* Check for deleted streams. */
2496 validate_endpoint_status_metadata_stream(&events
);
2500 DBG("Adding metadata stream %d to poll set",
2503 /* Add metadata stream to the global poll events list */
2504 lttng_poll_add(&events
, stream
->wait_fd
,
2505 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2506 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2507 DBG("Metadata thread pipe hung up");
2509 * Remove the pipe from the poll set and continue the loop
2510 * since their might be data to consume.
2512 lttng_poll_del(&events
,
2513 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2514 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2517 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2521 /* Handle other stream */
2527 uint64_t tmp_id
= (uint64_t) pollfd
;
2529 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2531 node
= lttng_ht_iter_get_node_u64(&iter
);
2534 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2537 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2538 /* Get the data out of the metadata file descriptor */
2539 DBG("Metadata available on fd %d", pollfd
);
2540 assert(stream
->wait_fd
== pollfd
);
2543 health_code_update();
2545 len
= ctx
->on_buffer_ready(stream
, ctx
);
2547 * We don't check the return value here since if we get
2548 * a negative len, it means an error occurred thus we
2549 * simply remove it from the poll set and free the
2554 /* It's ok to have an unavailable sub-buffer */
2555 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2556 /* Clean up stream from consumer and free it. */
2557 lttng_poll_del(&events
, stream
->wait_fd
);
2558 consumer_del_metadata_stream(stream
, metadata_ht
);
2560 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2561 DBG("Metadata fd %d is hup|err.", pollfd
);
2562 if (!stream
->hangup_flush_done
2563 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2564 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2565 DBG("Attempting to flush and consume the UST buffers");
2566 lttng_ustconsumer_on_stream_hangup(stream
);
2568 /* We just flushed the stream now read it. */
2570 health_code_update();
2572 len
= ctx
->on_buffer_ready(stream
, ctx
);
2574 * We don't check the return value here since if we get
2575 * a negative len, it means an error occurred thus we
2576 * simply remove it from the poll set and free the
2582 lttng_poll_del(&events
, stream
->wait_fd
);
2584 * This call update the channel states, closes file descriptors
2585 * and securely free the stream.
2587 consumer_del_metadata_stream(stream
, metadata_ht
);
2589 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2593 /* Release RCU lock for the stream looked up */
2601 DBG("Metadata poll thread exiting");
2603 lttng_poll_clean(&events
);
2608 ERR("Health error occurred in %s", __func__
);
2610 health_unregister(health_consumerd
);
2611 rcu_unregister_thread();
2616 * This thread polls the fds in the set to consume the data and write
2617 * it to tracefile if necessary.
2619 void *consumer_thread_data_poll(void *data
)
2621 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2622 struct pollfd
*pollfd
= NULL
;
2623 /* local view of the streams */
2624 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2625 /* local view of consumer_data.fds_count */
2627 /* 2 for the consumer_data_pipe and wake up pipe */
2628 const int nb_pipes_fd
= 2;
2629 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2630 int nb_inactive_fd
= 0;
2631 struct lttng_consumer_local_data
*ctx
= data
;
2634 rcu_register_thread();
2636 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2638 if (testpoint(consumerd_thread_data
)) {
2639 goto error_testpoint
;
2642 health_code_update();
2644 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2645 if (local_stream
== NULL
) {
2646 PERROR("local_stream malloc");
2651 health_code_update();
2657 * the fds set has been updated, we need to update our
2658 * local array as well
2660 pthread_mutex_lock(&consumer_data
.lock
);
2661 if (consumer_data
.need_update
) {
2666 local_stream
= NULL
;
2668 /* Allocate for all fds */
2669 pollfd
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) * sizeof(struct pollfd
));
2670 if (pollfd
== NULL
) {
2671 PERROR("pollfd malloc");
2672 pthread_mutex_unlock(&consumer_data
.lock
);
2676 local_stream
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) *
2677 sizeof(struct lttng_consumer_stream
*));
2678 if (local_stream
== NULL
) {
2679 PERROR("local_stream malloc");
2680 pthread_mutex_unlock(&consumer_data
.lock
);
2683 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2684 data_ht
, &nb_inactive_fd
);
2686 ERR("Error in allocating pollfd or local_outfds");
2687 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2688 pthread_mutex_unlock(&consumer_data
.lock
);
2692 consumer_data
.need_update
= 0;
2694 pthread_mutex_unlock(&consumer_data
.lock
);
2696 /* No FDs and consumer_quit, consumer_cleanup the thread */
2697 if (nb_fd
== 0 && nb_inactive_fd
== 0 &&
2698 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2699 err
= 0; /* All is OK */
2702 /* poll on the array of fds */
2704 DBG("polling on %d fd", nb_fd
+ nb_pipes_fd
);
2705 if (testpoint(consumerd_thread_data_poll
)) {
2708 health_poll_entry();
2709 num_rdy
= poll(pollfd
, nb_fd
+ nb_pipes_fd
, -1);
2711 DBG("poll num_rdy : %d", num_rdy
);
2712 if (num_rdy
== -1) {
2714 * Restart interrupted system call.
2716 if (errno
== EINTR
) {
2719 PERROR("Poll error");
2720 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2722 } else if (num_rdy
== 0) {
2723 DBG("Polling thread timed out");
2727 if (caa_unlikely(data_consumption_paused
)) {
2728 DBG("Data consumption paused, sleeping...");
2734 * If the consumer_data_pipe triggered poll go directly to the
2735 * beginning of the loop to update the array. We want to prioritize
2736 * array update over low-priority reads.
2738 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2739 ssize_t pipe_readlen
;
2741 DBG("consumer_data_pipe wake up");
2742 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2743 &new_stream
, sizeof(new_stream
));
2744 if (pipe_readlen
< sizeof(new_stream
)) {
2745 PERROR("Consumer data pipe");
2746 /* Continue so we can at least handle the current stream(s). */
2751 * If the stream is NULL, just ignore it. It's also possible that
2752 * the sessiond poll thread changed the consumer_quit state and is
2753 * waking us up to test it.
2755 if (new_stream
== NULL
) {
2756 validate_endpoint_status_data_stream();
2760 /* Continue to update the local streams and handle prio ones */
2764 /* Handle wakeup pipe. */
2765 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2767 ssize_t pipe_readlen
;
2769 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2771 if (pipe_readlen
< 0) {
2772 PERROR("Consumer data wakeup pipe");
2774 /* We've been awakened to handle stream(s). */
2775 ctx
->has_wakeup
= 0;
2778 /* Take care of high priority channels first. */
2779 for (i
= 0; i
< nb_fd
; i
++) {
2780 health_code_update();
2782 if (local_stream
[i
] == NULL
) {
2785 if (pollfd
[i
].revents
& POLLPRI
) {
2786 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2788 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2789 /* it's ok to have an unavailable sub-buffer */
2790 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2791 /* Clean the stream and free it. */
2792 consumer_del_stream(local_stream
[i
], data_ht
);
2793 local_stream
[i
] = NULL
;
2794 } else if (len
> 0) {
2795 local_stream
[i
]->data_read
= 1;
2801 * If we read high prio channel in this loop, try again
2802 * for more high prio data.
2808 /* Take care of low priority channels. */
2809 for (i
= 0; i
< nb_fd
; i
++) {
2810 health_code_update();
2812 if (local_stream
[i
] == NULL
) {
2815 if ((pollfd
[i
].revents
& POLLIN
) ||
2816 local_stream
[i
]->hangup_flush_done
||
2817 local_stream
[i
]->has_data
) {
2818 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2819 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2820 /* it's ok to have an unavailable sub-buffer */
2821 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2822 /* Clean the stream and free it. */
2823 consumer_del_stream(local_stream
[i
], data_ht
);
2824 local_stream
[i
] = NULL
;
2825 } else if (len
> 0) {
2826 local_stream
[i
]->data_read
= 1;
2831 /* Handle hangup and errors */
2832 for (i
= 0; i
< nb_fd
; i
++) {
2833 health_code_update();
2835 if (local_stream
[i
] == NULL
) {
2838 if (!local_stream
[i
]->hangup_flush_done
2839 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2840 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2841 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2842 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2844 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2845 /* Attempt read again, for the data we just flushed. */
2846 local_stream
[i
]->data_read
= 1;
2849 * If the poll flag is HUP/ERR/NVAL and we have
2850 * read no data in this pass, we can remove the
2851 * stream from its hash table.
2853 if ((pollfd
[i
].revents
& POLLHUP
)) {
2854 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2855 if (!local_stream
[i
]->data_read
) {
2856 consumer_del_stream(local_stream
[i
], data_ht
);
2857 local_stream
[i
] = NULL
;
2860 } else if (pollfd
[i
].revents
& POLLERR
) {
2861 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2862 if (!local_stream
[i
]->data_read
) {
2863 consumer_del_stream(local_stream
[i
], data_ht
);
2864 local_stream
[i
] = NULL
;
2867 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2868 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2869 if (!local_stream
[i
]->data_read
) {
2870 consumer_del_stream(local_stream
[i
], data_ht
);
2871 local_stream
[i
] = NULL
;
2875 if (local_stream
[i
] != NULL
) {
2876 local_stream
[i
]->data_read
= 0;
2883 DBG("polling thread exiting");
2888 * Close the write side of the pipe so epoll_wait() in
2889 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2890 * read side of the pipe. If we close them both, epoll_wait strangely does
2891 * not return and could create a endless wait period if the pipe is the
2892 * only tracked fd in the poll set. The thread will take care of closing
2895 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2900 ERR("Health error occurred in %s", __func__
);
2902 health_unregister(health_consumerd
);
2904 rcu_unregister_thread();
2909 * Close wake-up end of each stream belonging to the channel. This will
2910 * allow the poll() on the stream read-side to detect when the
2911 * write-side (application) finally closes them.
2914 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2916 struct lttng_ht
*ht
;
2917 struct lttng_consumer_stream
*stream
;
2918 struct lttng_ht_iter iter
;
2920 ht
= consumer_data
.stream_per_chan_id_ht
;
2923 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2924 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2925 ht
->match_fct
, &channel
->key
,
2926 &iter
.iter
, stream
, node_channel_id
.node
) {
2928 * Protect against teardown with mutex.
2930 pthread_mutex_lock(&stream
->lock
);
2931 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2934 switch (consumer_data
.type
) {
2935 case LTTNG_CONSUMER_KERNEL
:
2937 case LTTNG_CONSUMER32_UST
:
2938 case LTTNG_CONSUMER64_UST
:
2939 if (stream
->metadata_flag
) {
2940 /* Safe and protected by the stream lock. */
2941 lttng_ustconsumer_close_metadata(stream
->chan
);
2944 * Note: a mutex is taken internally within
2945 * liblttng-ust-ctl to protect timer wakeup_fd
2946 * use from concurrent close.
2948 lttng_ustconsumer_close_stream_wakeup(stream
);
2952 ERR("Unknown consumer_data type");
2956 pthread_mutex_unlock(&stream
->lock
);
2961 static void destroy_channel_ht(struct lttng_ht
*ht
)
2963 struct lttng_ht_iter iter
;
2964 struct lttng_consumer_channel
*channel
;
2972 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2973 ret
= lttng_ht_del(ht
, &iter
);
2978 lttng_ht_destroy(ht
);
2982 * This thread polls the channel fds to detect when they are being
2983 * closed. It closes all related streams if the channel is detected as
2984 * closed. It is currently only used as a shim layer for UST because the
2985 * consumerd needs to keep the per-stream wakeup end of pipes open for
2988 void *consumer_thread_channel_poll(void *data
)
2990 int ret
, i
, pollfd
, err
= -1;
2991 uint32_t revents
, nb_fd
;
2992 struct lttng_consumer_channel
*chan
= NULL
;
2993 struct lttng_ht_iter iter
;
2994 struct lttng_ht_node_u64
*node
;
2995 struct lttng_poll_event events
;
2996 struct lttng_consumer_local_data
*ctx
= data
;
2997 struct lttng_ht
*channel_ht
;
2999 rcu_register_thread();
3001 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
3003 if (testpoint(consumerd_thread_channel
)) {
3004 goto error_testpoint
;
3007 health_code_update();
3009 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3011 /* ENOMEM at this point. Better to bail out. */
3015 DBG("Thread channel poll started");
3017 /* Size is set to 1 for the consumer_channel pipe */
3018 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
3020 ERR("Poll set creation failed");
3024 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
3030 DBG("Channel main loop started");
3034 health_code_update();
3035 DBG("Channel poll wait");
3036 health_poll_entry();
3037 ret
= lttng_poll_wait(&events
, -1);
3038 DBG("Channel poll return from wait with %d fd(s)",
3039 LTTNG_POLL_GETNB(&events
));
3041 DBG("Channel event caught in thread");
3043 if (errno
== EINTR
) {
3044 ERR("Poll EINTR caught");
3047 if (LTTNG_POLL_GETNB(&events
) == 0) {
3048 err
= 0; /* All is OK */
3055 /* From here, the event is a channel wait fd */
3056 for (i
= 0; i
< nb_fd
; i
++) {
3057 health_code_update();
3059 revents
= LTTNG_POLL_GETEV(&events
, i
);
3060 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
3062 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
3063 if (revents
& LPOLLIN
) {
3064 enum consumer_channel_action action
;
3067 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
3070 ERR("Error reading channel pipe");
3072 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3077 case CONSUMER_CHANNEL_ADD
:
3078 DBG("Adding channel %d to poll set",
3081 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
3084 lttng_ht_add_unique_u64(channel_ht
,
3085 &chan
->wait_fd_node
);
3087 /* Add channel to the global poll events list */
3088 lttng_poll_add(&events
, chan
->wait_fd
,
3089 LPOLLERR
| LPOLLHUP
);
3091 case CONSUMER_CHANNEL_DEL
:
3094 * This command should never be called if the channel
3095 * has streams monitored by either the data or metadata
3096 * thread. The consumer only notify this thread with a
3097 * channel del. command if it receives a destroy
3098 * channel command from the session daemon that send it
3099 * if a command prior to the GET_CHANNEL failed.
3103 chan
= consumer_find_channel(key
);
3106 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
3109 lttng_poll_del(&events
, chan
->wait_fd
);
3110 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
3111 ret
= lttng_ht_del(channel_ht
, &iter
);
3114 switch (consumer_data
.type
) {
3115 case LTTNG_CONSUMER_KERNEL
:
3117 case LTTNG_CONSUMER32_UST
:
3118 case LTTNG_CONSUMER64_UST
:
3119 health_code_update();
3120 /* Destroy streams that might have been left in the stream list. */
3121 clean_channel_stream_list(chan
);
3124 ERR("Unknown consumer_data type");
3129 * Release our own refcount. Force channel deletion even if
3130 * streams were not initialized.
3132 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
3133 consumer_del_channel(chan
);
3138 case CONSUMER_CHANNEL_QUIT
:
3140 * Remove the pipe from the poll set and continue the loop
3141 * since their might be data to consume.
3143 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3146 ERR("Unknown action");
3149 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3150 DBG("Channel thread pipe hung up");
3152 * Remove the pipe from the poll set and continue the loop
3153 * since their might be data to consume.
3155 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3158 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3162 /* Handle other stream */
3168 uint64_t tmp_id
= (uint64_t) pollfd
;
3170 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3172 node
= lttng_ht_iter_get_node_u64(&iter
);
3175 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3178 /* Check for error event */
3179 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3180 DBG("Channel fd %d is hup|err.", pollfd
);
3182 lttng_poll_del(&events
, chan
->wait_fd
);
3183 ret
= lttng_ht_del(channel_ht
, &iter
);
3187 * This will close the wait fd for each stream associated to
3188 * this channel AND monitored by the data/metadata thread thus
3189 * will be clean by the right thread.
3191 consumer_close_channel_streams(chan
);
3193 /* Release our own refcount */
3194 if (!uatomic_sub_return(&chan
->refcount
, 1)
3195 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3196 consumer_del_channel(chan
);
3199 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3204 /* Release RCU lock for the channel looked up */
3212 lttng_poll_clean(&events
);
3214 destroy_channel_ht(channel_ht
);
3217 DBG("Channel poll thread exiting");
3220 ERR("Health error occurred in %s", __func__
);
3222 health_unregister(health_consumerd
);
3223 rcu_unregister_thread();
3227 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3228 struct pollfd
*sockpoll
, int client_socket
)
3235 ret
= lttng_consumer_poll_socket(sockpoll
);
3239 DBG("Metadata connection on client_socket");
3241 /* Blocking call, waiting for transmission */
3242 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3243 if (ctx
->consumer_metadata_socket
< 0) {
3244 WARN("On accept metadata");
3255 * This thread listens on the consumerd socket and receives the file
3256 * descriptors from the session daemon.
3258 void *consumer_thread_sessiond_poll(void *data
)
3260 int sock
= -1, client_socket
, ret
, err
= -1;
3262 * structure to poll for incoming data on communication socket avoids
3263 * making blocking sockets.
3265 struct pollfd consumer_sockpoll
[2];
3266 struct lttng_consumer_local_data
*ctx
= data
;
3268 rcu_register_thread();
3270 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3272 if (testpoint(consumerd_thread_sessiond
)) {
3273 goto error_testpoint
;
3276 health_code_update();
3278 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3279 unlink(ctx
->consumer_command_sock_path
);
3280 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3281 if (client_socket
< 0) {
3282 ERR("Cannot create command socket");
3286 ret
= lttcomm_listen_unix_sock(client_socket
);
3291 DBG("Sending ready command to lttng-sessiond");
3292 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3293 /* return < 0 on error, but == 0 is not fatal */
3295 ERR("Error sending ready command to lttng-sessiond");
3299 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3300 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3301 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3302 consumer_sockpoll
[1].fd
= client_socket
;
3303 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3305 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3313 DBG("Connection on client_socket");
3315 /* Blocking call, waiting for transmission */
3316 sock
= lttcomm_accept_unix_sock(client_socket
);
3323 * Setup metadata socket which is the second socket connection on the
3324 * command unix socket.
3326 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3335 /* This socket is not useful anymore. */
3336 ret
= close(client_socket
);
3338 PERROR("close client_socket");
3342 /* update the polling structure to poll on the established socket */
3343 consumer_sockpoll
[1].fd
= sock
;
3344 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3347 health_code_update();
3349 health_poll_entry();
3350 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3359 DBG("Incoming command on sock");
3360 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3363 * This could simply be a session daemon quitting. Don't output
3366 DBG("Communication interrupted on command socket");
3370 if (CMM_LOAD_SHARED(consumer_quit
)) {
3371 DBG("consumer_thread_receive_fds received quit from signal");
3372 err
= 0; /* All is OK */
3375 DBG("received command on sock");
3381 DBG("Consumer thread sessiond poll exiting");
3384 * Close metadata streams since the producer is the session daemon which
3387 * NOTE: for now, this only applies to the UST tracer.
3389 lttng_consumer_close_all_metadata();
3392 * when all fds have hung up, the polling thread
3395 CMM_STORE_SHARED(consumer_quit
, 1);
3398 * Notify the data poll thread to poll back again and test the
3399 * consumer_quit state that we just set so to quit gracefully.
3401 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3403 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3405 notify_health_quit_pipe(health_quit_pipe
);
3407 /* Cleaning up possibly open sockets. */
3411 PERROR("close sock sessiond poll");
3414 if (client_socket
>= 0) {
3415 ret
= close(client_socket
);
3417 PERROR("close client_socket sessiond poll");
3424 ERR("Health error occurred in %s", __func__
);
3426 health_unregister(health_consumerd
);
3428 rcu_unregister_thread();
3432 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3433 struct lttng_consumer_local_data
*ctx
)
3437 pthread_mutex_lock(&stream
->chan
->lock
);
3438 pthread_mutex_lock(&stream
->lock
);
3439 if (stream
->metadata_flag
) {
3440 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3443 switch (consumer_data
.type
) {
3444 case LTTNG_CONSUMER_KERNEL
:
3445 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3447 case LTTNG_CONSUMER32_UST
:
3448 case LTTNG_CONSUMER64_UST
:
3449 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3452 ERR("Unknown consumer_data type");
3458 if (stream
->metadata_flag
) {
3459 pthread_cond_broadcast(&stream
->metadata_rdv
);
3460 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3462 pthread_mutex_unlock(&stream
->lock
);
3463 pthread_mutex_unlock(&stream
->chan
->lock
);
3468 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3470 switch (consumer_data
.type
) {
3471 case LTTNG_CONSUMER_KERNEL
:
3472 return lttng_kconsumer_on_recv_stream(stream
);
3473 case LTTNG_CONSUMER32_UST
:
3474 case LTTNG_CONSUMER64_UST
:
3475 return lttng_ustconsumer_on_recv_stream(stream
);
3477 ERR("Unknown consumer_data type");
3484 * Allocate and set consumer data hash tables.
3486 int lttng_consumer_init(void)
3488 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3489 if (!consumer_data
.channel_ht
) {
3493 consumer_data
.channels_by_session_id_ht
=
3494 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3495 if (!consumer_data
.channels_by_session_id_ht
) {
3499 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3500 if (!consumer_data
.relayd_ht
) {
3504 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3505 if (!consumer_data
.stream_list_ht
) {
3509 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3510 if (!consumer_data
.stream_per_chan_id_ht
) {
3514 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3519 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3524 consumer_data
.chunk_registry
= lttng_trace_chunk_registry_create();
3525 if (!consumer_data
.chunk_registry
) {
3536 * Process the ADD_RELAYD command receive by a consumer.
3538 * This will create a relayd socket pair and add it to the relayd hash table.
3539 * The caller MUST acquire a RCU read side lock before calling it.
3541 void consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3542 struct lttng_consumer_local_data
*ctx
, int sock
,
3543 struct pollfd
*consumer_sockpoll
,
3544 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3545 uint64_t relayd_session_id
)
3547 int fd
= -1, ret
= -1, relayd_created
= 0;
3548 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3549 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3552 assert(relayd_sock
);
3554 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3556 /* Get relayd reference if exists. */
3557 relayd
= consumer_find_relayd(net_seq_idx
);
3558 if (relayd
== NULL
) {
3559 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3560 /* Not found. Allocate one. */
3561 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3562 if (relayd
== NULL
) {
3563 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3566 relayd
->sessiond_session_id
= sessiond_id
;
3571 * This code path MUST continue to the consumer send status message to
3572 * we can notify the session daemon and continue our work without
3573 * killing everything.
3577 * relayd key should never be found for control socket.
3579 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3582 /* First send a status message before receiving the fds. */
3583 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3585 /* Somehow, the session daemon is not responding anymore. */
3586 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3587 goto error_nosignal
;
3590 /* Poll on consumer socket. */
3591 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3593 /* Needing to exit in the middle of a command: error. */
3594 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3595 goto error_nosignal
;
3598 /* Get relayd socket from session daemon */
3599 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3600 if (ret
!= sizeof(fd
)) {
3601 fd
= -1; /* Just in case it gets set with an invalid value. */
3604 * Failing to receive FDs might indicate a major problem such as
3605 * reaching a fd limit during the receive where the kernel returns a
3606 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3607 * don't take any chances and stop everything.
3609 * XXX: Feature request #558 will fix that and avoid this possible
3610 * issue when reaching the fd limit.
3612 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3613 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3617 /* Copy socket information and received FD */
3618 switch (sock_type
) {
3619 case LTTNG_STREAM_CONTROL
:
3620 /* Copy received lttcomm socket */
3621 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3622 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3623 /* Handle create_sock error. */
3625 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3629 * Close the socket created internally by
3630 * lttcomm_create_sock, so we can replace it by the one
3631 * received from sessiond.
3633 if (close(relayd
->control_sock
.sock
.fd
)) {
3637 /* Assign new file descriptor */
3638 relayd
->control_sock
.sock
.fd
= fd
;
3639 /* Assign version values. */
3640 relayd
->control_sock
.major
= relayd_sock
->major
;
3641 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3643 relayd
->relayd_session_id
= relayd_session_id
;
3646 case LTTNG_STREAM_DATA
:
3647 /* Copy received lttcomm socket */
3648 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3649 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3650 /* Handle create_sock error. */
3652 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3656 * Close the socket created internally by
3657 * lttcomm_create_sock, so we can replace it by the one
3658 * received from sessiond.
3660 if (close(relayd
->data_sock
.sock
.fd
)) {
3664 /* Assign new file descriptor */
3665 relayd
->data_sock
.sock
.fd
= fd
;
3666 /* Assign version values. */
3667 relayd
->data_sock
.major
= relayd_sock
->major
;
3668 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3671 ERR("Unknown relayd socket type (%d)", sock_type
);
3672 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3676 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3677 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3678 relayd
->net_seq_idx
, fd
);
3680 * We gave the ownership of the fd to the relayd structure. Set the
3681 * fd to -1 so we don't call close() on it in the error path below.
3685 /* We successfully added the socket. Send status back. */
3686 ret
= consumer_send_status_msg(sock
, ret_code
);
3688 /* Somehow, the session daemon is not responding anymore. */
3689 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3690 goto error_nosignal
;
3694 * Add relayd socket pair to consumer data hashtable. If object already
3695 * exists or on error, the function gracefully returns.
3704 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3705 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3709 /* Close received socket if valid. */
3712 PERROR("close received socket");
3716 if (relayd_created
) {
3722 * Search for a relayd associated to the session id and return the reference.
3724 * A rcu read side lock MUST be acquire before calling this function and locked
3725 * until the relayd object is no longer necessary.
3727 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3729 struct lttng_ht_iter iter
;
3730 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3732 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3733 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3736 * Check by sessiond id which is unique here where the relayd session
3737 * id might not be when having multiple relayd.
3739 if (relayd
->sessiond_session_id
== id
) {
3740 /* Found the relayd. There can be only one per id. */
3752 * Check if for a given session id there is still data needed to be extract
3755 * Return 1 if data is pending or else 0 meaning ready to be read.
3757 int consumer_data_pending(uint64_t id
)
3760 struct lttng_ht_iter iter
;
3761 struct lttng_ht
*ht
;
3762 struct lttng_consumer_stream
*stream
;
3763 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3764 int (*data_pending
)(struct lttng_consumer_stream
*);
3766 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3769 pthread_mutex_lock(&consumer_data
.lock
);
3771 switch (consumer_data
.type
) {
3772 case LTTNG_CONSUMER_KERNEL
:
3773 data_pending
= lttng_kconsumer_data_pending
;
3775 case LTTNG_CONSUMER32_UST
:
3776 case LTTNG_CONSUMER64_UST
:
3777 data_pending
= lttng_ustconsumer_data_pending
;
3780 ERR("Unknown consumer data type");
3784 /* Ease our life a bit */
3785 ht
= consumer_data
.stream_list_ht
;
3787 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3788 ht
->hash_fct(&id
, lttng_ht_seed
),
3790 &iter
.iter
, stream
, node_session_id
.node
) {
3791 pthread_mutex_lock(&stream
->lock
);
3794 * A removed node from the hash table indicates that the stream has
3795 * been deleted thus having a guarantee that the buffers are closed
3796 * on the consumer side. However, data can still be transmitted
3797 * over the network so don't skip the relayd check.
3799 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3801 /* Check the stream if there is data in the buffers. */
3802 ret
= data_pending(stream
);
3804 pthread_mutex_unlock(&stream
->lock
);
3809 pthread_mutex_unlock(&stream
->lock
);
3812 relayd
= find_relayd_by_session_id(id
);
3814 unsigned int is_data_inflight
= 0;
3816 /* Send init command for data pending. */
3817 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3818 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3819 relayd
->relayd_session_id
);
3821 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3822 /* Communication error thus the relayd so no data pending. */
3823 goto data_not_pending
;
3826 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3827 ht
->hash_fct(&id
, lttng_ht_seed
),
3829 &iter
.iter
, stream
, node_session_id
.node
) {
3830 if (stream
->metadata_flag
) {
3831 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3832 stream
->relayd_stream_id
);
3834 ret
= relayd_data_pending(&relayd
->control_sock
,
3835 stream
->relayd_stream_id
,
3836 stream
->next_net_seq_num
- 1);
3840 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3842 } else if (ret
< 0) {
3843 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3844 lttng_consumer_cleanup_relayd(relayd
);
3845 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3846 goto data_not_pending
;
3850 /* Send end command for data pending. */
3851 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3852 relayd
->relayd_session_id
, &is_data_inflight
);
3853 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3855 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3856 lttng_consumer_cleanup_relayd(relayd
);
3857 goto data_not_pending
;
3859 if (is_data_inflight
) {
3865 * Finding _no_ node in the hash table and no inflight data means that the
3866 * stream(s) have been removed thus data is guaranteed to be available for
3867 * analysis from the trace files.
3871 /* Data is available to be read by a viewer. */
3872 pthread_mutex_unlock(&consumer_data
.lock
);
3877 /* Data is still being extracted from buffers. */
3878 pthread_mutex_unlock(&consumer_data
.lock
);
3884 * Send a ret code status message to the sessiond daemon.
3886 * Return the sendmsg() return value.
3888 int consumer_send_status_msg(int sock
, int ret_code
)
3890 struct lttcomm_consumer_status_msg msg
;
3892 memset(&msg
, 0, sizeof(msg
));
3893 msg
.ret_code
= ret_code
;
3895 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3899 * Send a channel status message to the sessiond daemon.
3901 * Return the sendmsg() return value.
3903 int consumer_send_status_channel(int sock
,
3904 struct lttng_consumer_channel
*channel
)
3906 struct lttcomm_consumer_status_channel msg
;
3910 memset(&msg
, 0, sizeof(msg
));
3912 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3914 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3915 msg
.key
= channel
->key
;
3916 msg
.stream_count
= channel
->streams
.count
;
3919 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3922 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3923 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3924 uint64_t max_sb_size
)
3926 unsigned long start_pos
;
3928 if (!nb_packets_per_stream
) {
3929 return consumed_pos
; /* Grab everything */
3931 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3932 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3933 if ((long) (start_pos
- consumed_pos
) < 0) {
3934 return consumed_pos
; /* Grab everything */
3940 int consumer_flush_buffer(struct lttng_consumer_stream
*stream
, int producer_active
)
3944 switch (consumer_data
.type
) {
3945 case LTTNG_CONSUMER_KERNEL
:
3946 ret
= kernctl_buffer_flush(stream
->wait_fd
);
3948 ERR("Failed to flush kernel stream");
3952 case LTTNG_CONSUMER32_UST
:
3953 case LTTNG_CONSUMER64_UST
:
3954 lttng_ustctl_flush_buffer(stream
, producer_active
);
3957 ERR("Unknown consumer_data type");
3966 * Sample the rotate position for all the streams of a channel. If a stream
3967 * is already at the rotate position (produced == consumed), we flag it as
3968 * ready for rotation. The rotation of ready streams occurs after we have
3969 * replied to the session daemon that we have finished sampling the positions.
3970 * Must be called with RCU read-side lock held to ensure existence of channel.
3972 * Returns 0 on success, < 0 on error
3974 int lttng_consumer_rotate_channel(struct lttng_consumer_channel
*channel
,
3975 uint64_t key
, uint64_t relayd_id
, uint32_t metadata
,
3976 struct lttng_consumer_local_data
*ctx
)
3979 struct lttng_consumer_stream
*stream
;
3980 struct lttng_ht_iter iter
;
3981 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
3982 struct lttng_dynamic_array stream_rotation_positions
;
3983 uint64_t next_chunk_id
, stream_count
= 0;
3984 enum lttng_trace_chunk_status chunk_status
;
3985 const bool is_local_trace
= relayd_id
== -1ULL;
3986 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3987 bool rotating_to_new_chunk
= true;
3989 DBG("Consumer sample rotate position for channel %" PRIu64
, key
);
3991 lttng_dynamic_array_init(&stream_rotation_positions
,
3992 sizeof(struct relayd_stream_rotation_position
), NULL
);
3996 pthread_mutex_lock(&channel
->lock
);
3997 assert(channel
->trace_chunk
);
3998 chunk_status
= lttng_trace_chunk_get_id(channel
->trace_chunk
,
4000 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4002 goto end_unlock_channel
;
4005 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4006 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4007 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4008 stream
, node_channel_id
.node
) {
4009 unsigned long consumed_pos
;
4011 health_code_update();
4014 * Lock stream because we are about to change its state.
4016 pthread_mutex_lock(&stream
->lock
);
4018 if (stream
->trace_chunk
== stream
->chan
->trace_chunk
) {
4019 rotating_to_new_chunk
= false;
4022 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4024 ERR("Failed to sample snapshot position during channel rotation");
4025 goto end_unlock_stream
;
4028 ret
= lttng_consumer_get_produced_snapshot(stream
,
4029 &stream
->rotate_position
);
4031 ERR("Failed to sample produced position during channel rotation");
4032 goto end_unlock_stream
;
4035 lttng_consumer_get_consumed_snapshot(stream
,
4037 if (consumed_pos
== stream
->rotate_position
) {
4038 stream
->rotate_ready
= true;
4042 * Active flush; has no effect if the production position
4043 * is at a packet boundary.
4045 ret
= consumer_flush_buffer(stream
, 1);
4047 ERR("Failed to flush stream %" PRIu64
" during channel rotation",
4049 goto end_unlock_stream
;
4052 if (!is_local_trace
) {
4054 * The relay daemon control protocol expects a rotation
4055 * position as "the sequence number of the first packet
4056 * _after_ the current trace chunk.
4058 * At the moment when the positions of the buffers are
4059 * sampled, the production position does not necessarily
4060 * sit at a packet boundary. The 'active' flush
4061 * operation above will push the production position to
4062 * the next packet boundary _if_ it is not already
4063 * sitting at such a boundary.
4065 * Assuming a current production position that is not
4066 * on the bound of a packet, the 'target' sequence
4068 * (consumed_pos / subbuffer_size) + 1
4069 * Note the '+ 1' to ensure the current packet is
4070 * part of the current trace chunk.
4072 * However, if the production position is already at
4073 * a packet boundary, the '+ 1' is not necessary as the
4074 * last packet of the current chunk is already
4077 const struct relayd_stream_rotation_position position
= {
4078 .stream_id
= stream
->relayd_stream_id
,
4079 .rotate_at_seq_num
= (stream
->rotate_position
/ stream
->max_sb_size
) +
4080 !!(stream
->rotate_position
% stream
->max_sb_size
),
4083 ret
= lttng_dynamic_array_add_element(
4084 &stream_rotation_positions
,
4087 ERR("Failed to allocate stream rotation position");
4088 goto end_unlock_stream
;
4092 pthread_mutex_unlock(&stream
->lock
);
4095 pthread_mutex_unlock(&channel
->lock
);
4097 if (is_local_trace
) {
4102 relayd
= consumer_find_relayd(relayd_id
);
4104 ERR("Failed to find relayd %" PRIu64
, relayd_id
);
4109 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4110 ret
= relayd_rotate_streams(&relayd
->control_sock
, stream_count
,
4111 rotating_to_new_chunk
? &next_chunk_id
: NULL
,
4112 (const struct relayd_stream_rotation_position
*)
4113 stream_rotation_positions
.buffer
.data
);
4114 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4116 ERR("Relayd rotate stream failed. Cleaning up relayd %" PRIu64
,
4117 relayd
->net_seq_idx
);
4118 lttng_consumer_cleanup_relayd(relayd
);
4126 pthread_mutex_unlock(&stream
->lock
);
4128 pthread_mutex_unlock(&channel
->lock
);
4131 lttng_dynamic_array_reset(&stream_rotation_positions
);
4136 * Check if a stream is ready to be rotated after extracting it.
4138 * Return 1 if it is ready for rotation, 0 if it is not, a negative value on
4139 * error. Stream lock must be held.
4141 int lttng_consumer_stream_is_rotate_ready(struct lttng_consumer_stream
*stream
)
4144 unsigned long consumed_pos
;
4146 if (!stream
->rotate_position
&& !stream
->rotate_ready
) {
4151 if (stream
->rotate_ready
) {
4157 * If we don't have the rotate_ready flag, check the consumed position
4158 * to determine if we need to rotate.
4160 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4162 ERR("Taking snapshot positions");
4166 ret
= lttng_consumer_get_consumed_snapshot(stream
, &consumed_pos
);
4168 ERR("Consumed snapshot position");
4172 /* Rotate position not reached yet (with check for overflow). */
4173 if ((long) (consumed_pos
- stream
->rotate_position
) < 0) {
4184 * Reset the state for a stream after a rotation occurred.
4186 void lttng_consumer_reset_stream_rotate_state(struct lttng_consumer_stream
*stream
)
4188 stream
->rotate_position
= 0;
4189 stream
->rotate_ready
= false;
4193 * Perform the rotation a local stream file.
4196 int rotate_local_stream(struct lttng_consumer_local_data
*ctx
,
4197 struct lttng_consumer_stream
*stream
)
4201 DBG("Rotate local stream: stream key %" PRIu64
", channel key %" PRIu64
,
4204 stream
->tracefile_size_current
= 0;
4205 stream
->tracefile_count_current
= 0;
4207 if (stream
->out_fd
>= 0) {
4208 ret
= close(stream
->out_fd
);
4210 PERROR("Failed to close stream out_fd of channel \"%s\"",
4211 stream
->chan
->name
);
4213 stream
->out_fd
= -1;
4216 if (stream
->index_file
) {
4217 lttng_index_file_put(stream
->index_file
);
4218 stream
->index_file
= NULL
;
4221 if (!stream
->trace_chunk
) {
4225 ret
= consumer_stream_create_output_files(stream
, true);
4231 * Performs the stream rotation for the rotate session feature if needed.
4232 * It must be called with the channel and stream locks held.
4234 * Return 0 on success, a negative number of error.
4236 int lttng_consumer_rotate_stream(struct lttng_consumer_local_data
*ctx
,
4237 struct lttng_consumer_stream
*stream
)
4241 DBG("Consumer rotate stream %" PRIu64
, stream
->key
);
4244 * Update the stream's 'current' chunk to the session's (channel)
4245 * now-current chunk.
4247 lttng_trace_chunk_put(stream
->trace_chunk
);
4248 if (stream
->chan
->trace_chunk
== stream
->trace_chunk
) {
4250 * A channel can be rotated and not have a "next" chunk
4251 * to transition to. In that case, the channel's "current chunk"
4252 * has not been closed yet, but it has not been updated to
4253 * a "next" trace chunk either. Hence, the stream, like its
4254 * parent channel, becomes part of no chunk and can't output
4255 * anything until a new trace chunk is created.
4257 stream
->trace_chunk
= NULL
;
4258 } else if (stream
->chan
->trace_chunk
&&
4259 !lttng_trace_chunk_get(stream
->chan
->trace_chunk
)) {
4260 ERR("Failed to acquire a reference to channel's trace chunk during stream rotation");
4265 * Update the stream's trace chunk to its parent channel's
4266 * current trace chunk.
4268 stream
->trace_chunk
= stream
->chan
->trace_chunk
;
4271 if (stream
->net_seq_idx
== (uint64_t) -1ULL) {
4272 ret
= rotate_local_stream(ctx
, stream
);
4274 ERR("Failed to rotate stream, ret = %i", ret
);
4279 if (stream
->metadata_flag
&& stream
->trace_chunk
) {
4281 * If the stream has transitioned to a new trace
4282 * chunk, the metadata should be re-dumped to the
4285 * However, it is possible for a stream to transition to
4286 * a "no-chunk" state. This can happen if a rotation
4287 * occurs on an inactive session. In such cases, the metadata
4288 * regeneration will happen when the next trace chunk is
4291 ret
= consumer_metadata_stream_dump(stream
);
4296 lttng_consumer_reset_stream_rotate_state(stream
);
4305 * Rotate all the ready streams now.
4307 * This is especially important for low throughput streams that have already
4308 * been consumed, we cannot wait for their next packet to perform the
4310 * Need to be called with RCU read-side lock held to ensure existence of
4313 * Returns 0 on success, < 0 on error
4315 int lttng_consumer_rotate_ready_streams(struct lttng_consumer_channel
*channel
,
4316 uint64_t key
, struct lttng_consumer_local_data
*ctx
)
4319 struct lttng_consumer_stream
*stream
;
4320 struct lttng_ht_iter iter
;
4321 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
4325 DBG("Consumer rotate ready streams in channel %" PRIu64
, key
);
4327 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4328 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4329 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4330 stream
, node_channel_id
.node
) {
4331 health_code_update();
4333 pthread_mutex_lock(&stream
->chan
->lock
);
4334 pthread_mutex_lock(&stream
->lock
);
4336 if (!stream
->rotate_ready
) {
4337 pthread_mutex_unlock(&stream
->lock
);
4338 pthread_mutex_unlock(&stream
->chan
->lock
);
4341 DBG("Consumer rotate ready stream %" PRIu64
, stream
->key
);
4343 ret
= lttng_consumer_rotate_stream(ctx
, stream
);
4344 pthread_mutex_unlock(&stream
->lock
);
4345 pthread_mutex_unlock(&stream
->chan
->lock
);
4358 enum lttcomm_return_code
lttng_consumer_init_command(
4359 struct lttng_consumer_local_data
*ctx
,
4360 const lttng_uuid sessiond_uuid
)
4362 enum lttcomm_return_code ret
;
4363 char uuid_str
[UUID_STR_LEN
];
4365 if (ctx
->sessiond_uuid
.is_set
) {
4366 ret
= LTTCOMM_CONSUMERD_ALREADY_SET
;
4370 ctx
->sessiond_uuid
.is_set
= true;
4371 memcpy(ctx
->sessiond_uuid
.value
, sessiond_uuid
, sizeof(lttng_uuid
));
4372 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4373 lttng_uuid_to_str(sessiond_uuid
, uuid_str
);
4374 DBG("Received session daemon UUID: %s", uuid_str
);
4379 enum lttcomm_return_code
lttng_consumer_create_trace_chunk(
4380 const uint64_t *relayd_id
, uint64_t session_id
,
4382 time_t chunk_creation_timestamp
,
4383 const char *chunk_override_name
,
4384 const struct lttng_credentials
*credentials
,
4385 struct lttng_directory_handle
*chunk_directory_handle
)
4388 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4389 struct lttng_trace_chunk
*created_chunk
= NULL
, *published_chunk
= NULL
;
4390 enum lttng_trace_chunk_status chunk_status
;
4391 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4392 char creation_timestamp_buffer
[ISO8601_STR_LEN
];
4393 const char *relayd_id_str
= "(none)";
4394 const char *creation_timestamp_str
;
4395 struct lttng_ht_iter iter
;
4396 struct lttng_consumer_channel
*channel
;
4399 /* Only used for logging purposes. */
4400 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4401 "%" PRIu64
, *relayd_id
);
4402 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4403 relayd_id_str
= relayd_id_buffer
;
4405 relayd_id_str
= "(formatting error)";
4409 /* Local protocol error. */
4410 assert(chunk_creation_timestamp
);
4411 ret
= time_to_iso8601_str(chunk_creation_timestamp
,
4412 creation_timestamp_buffer
,
4413 sizeof(creation_timestamp_buffer
));
4414 creation_timestamp_str
= !ret
? creation_timestamp_buffer
:
4415 "(formatting error)";
4417 DBG("Consumer create trace chunk command: relay_id = %s"
4418 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4419 ", chunk_override_name = %s"
4420 ", chunk_creation_timestamp = %s",
4421 relayd_id_str
, session_id
, chunk_id
,
4422 chunk_override_name
? : "(none)",
4423 creation_timestamp_str
);
4426 * The trace chunk registry, as used by the consumer daemon, implicitly
4427 * owns the trace chunks. This is only needed in the consumer since
4428 * the consumer has no notion of a session beyond session IDs being
4429 * used to identify other objects.
4431 * The lttng_trace_chunk_registry_publish() call below provides a
4432 * reference which is not released; it implicitly becomes the session
4433 * daemon's reference to the chunk in the consumer daemon.
4435 * The lifetime of trace chunks in the consumer daemon is managed by
4436 * the session daemon through the LTTNG_CONSUMER_CREATE_TRACE_CHUNK
4437 * and LTTNG_CONSUMER_DESTROY_TRACE_CHUNK commands.
4439 created_chunk
= lttng_trace_chunk_create(chunk_id
,
4440 chunk_creation_timestamp
);
4441 if (!created_chunk
) {
4442 ERR("Failed to create trace chunk");
4443 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4447 if (chunk_override_name
) {
4448 chunk_status
= lttng_trace_chunk_override_name(created_chunk
,
4449 chunk_override_name
);
4450 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4451 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4456 if (chunk_directory_handle
) {
4457 chunk_status
= lttng_trace_chunk_set_credentials(created_chunk
,
4459 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4460 ERR("Failed to set trace chunk credentials");
4461 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4465 * The consumer daemon has no ownership of the chunk output
4468 chunk_status
= lttng_trace_chunk_set_as_user(created_chunk
,
4469 chunk_directory_handle
);
4470 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4471 ERR("Failed to set trace chunk's directory handle");
4472 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4477 published_chunk
= lttng_trace_chunk_registry_publish_chunk(
4478 consumer_data
.chunk_registry
, session_id
,
4480 lttng_trace_chunk_put(created_chunk
);
4481 created_chunk
= NULL
;
4482 if (!published_chunk
) {
4483 ERR("Failed to publish trace chunk");
4484 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4489 cds_lfht_for_each_entry_duplicate(consumer_data
.channels_by_session_id_ht
->ht
,
4490 consumer_data
.channels_by_session_id_ht
->hash_fct(
4491 &session_id
, lttng_ht_seed
),
4492 consumer_data
.channels_by_session_id_ht
->match_fct
,
4493 &session_id
, &iter
.iter
, channel
,
4494 channels_by_session_id_ht_node
.node
) {
4495 ret
= lttng_consumer_channel_set_trace_chunk(channel
,
4499 * Roll-back the creation of this chunk.
4501 * This is important since the session daemon will
4502 * assume that the creation of this chunk failed and
4503 * will never ask for it to be closed, resulting
4504 * in a leak and an inconsistent state for some
4507 enum lttcomm_return_code close_ret
;
4508 char path
[LTTNG_PATH_MAX
];
4510 DBG("Failed to set new trace chunk on existing channels, rolling back");
4511 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4512 session_id
, chunk_id
,
4513 chunk_creation_timestamp
, NULL
,
4515 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4516 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4517 session_id
, chunk_id
);
4520 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4526 struct consumer_relayd_sock_pair
*relayd
;
4528 relayd
= consumer_find_relayd(*relayd_id
);
4530 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4531 ret
= relayd_create_trace_chunk(
4532 &relayd
->control_sock
, published_chunk
);
4533 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4535 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
, *relayd_id
);
4538 if (!relayd
|| ret
) {
4539 enum lttcomm_return_code close_ret
;
4540 char path
[LTTNG_PATH_MAX
];
4542 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4545 chunk_creation_timestamp
,
4547 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4548 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4553 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4560 /* Release the reference returned by the "publish" operation. */
4561 lttng_trace_chunk_put(published_chunk
);
4562 lttng_trace_chunk_put(created_chunk
);
4566 enum lttcomm_return_code
lttng_consumer_close_trace_chunk(
4567 const uint64_t *relayd_id
, uint64_t session_id
,
4568 uint64_t chunk_id
, time_t chunk_close_timestamp
,
4569 const enum lttng_trace_chunk_command_type
*close_command
,
4572 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4573 struct lttng_trace_chunk
*chunk
;
4574 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4575 const char *relayd_id_str
= "(none)";
4576 const char *close_command_name
= "none";
4577 struct lttng_ht_iter iter
;
4578 struct lttng_consumer_channel
*channel
;
4579 enum lttng_trace_chunk_status chunk_status
;
4584 /* Only used for logging purposes. */
4585 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4586 "%" PRIu64
, *relayd_id
);
4587 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4588 relayd_id_str
= relayd_id_buffer
;
4590 relayd_id_str
= "(formatting error)";
4593 if (close_command
) {
4594 close_command_name
= lttng_trace_chunk_command_type_get_name(
4598 DBG("Consumer close trace chunk command: relayd_id = %s"
4599 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4600 ", close command = %s",
4601 relayd_id_str
, session_id
, chunk_id
,
4602 close_command_name
);
4604 chunk
= lttng_trace_chunk_registry_find_chunk(
4605 consumer_data
.chunk_registry
, session_id
, chunk_id
);
4607 ERR("Failed to find chunk: session_id = %" PRIu64
4608 ", chunk_id = %" PRIu64
,
4609 session_id
, chunk_id
);
4610 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4614 chunk_status
= lttng_trace_chunk_set_close_timestamp(chunk
,
4615 chunk_close_timestamp
);
4616 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4617 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4621 if (close_command
) {
4622 chunk_status
= lttng_trace_chunk_set_close_command(
4623 chunk
, *close_command
);
4624 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4625 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4631 * chunk is now invalid to access as we no longer hold a reference to
4632 * it; it is only kept around to compare it (by address) to the
4633 * current chunk found in the session's channels.
4636 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
,
4637 channel
, node
.node
) {
4641 * Only change the channel's chunk to NULL if it still
4642 * references the chunk being closed. The channel may
4643 * reference a newer channel in the case of a session
4644 * rotation. When a session rotation occurs, the "next"
4645 * chunk is created before the "current" chunk is closed.
4647 if (channel
->trace_chunk
!= chunk
) {
4650 ret
= lttng_consumer_channel_set_trace_chunk(channel
, NULL
);
4653 * Attempt to close the chunk on as many channels as
4656 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4662 struct consumer_relayd_sock_pair
*relayd
;
4664 relayd
= consumer_find_relayd(*relayd_id
);
4666 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4667 ret
= relayd_close_trace_chunk(
4668 &relayd
->control_sock
, chunk
,
4670 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4672 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
,
4676 if (!relayd
|| ret
) {
4677 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4685 * Release the reference returned by the "find" operation and
4686 * the session daemon's implicit reference to the chunk.
4688 lttng_trace_chunk_put(chunk
);
4689 lttng_trace_chunk_put(chunk
);
4694 enum lttcomm_return_code
lttng_consumer_trace_chunk_exists(
4695 const uint64_t *relayd_id
, uint64_t session_id
,
4699 enum lttcomm_return_code ret_code
;
4700 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4701 const char *relayd_id_str
= "(none)";
4702 const bool is_local_trace
= !relayd_id
;
4703 struct consumer_relayd_sock_pair
*relayd
= NULL
;
4704 bool chunk_exists_local
, chunk_exists_remote
;
4709 /* Only used for logging purposes. */
4710 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4711 "%" PRIu64
, *relayd_id
);
4712 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4713 relayd_id_str
= relayd_id_buffer
;
4715 relayd_id_str
= "(formatting error)";
4719 DBG("Consumer trace chunk exists command: relayd_id = %s"
4720 ", chunk_id = %" PRIu64
, relayd_id_str
,
4722 ret
= lttng_trace_chunk_registry_chunk_exists(
4723 consumer_data
.chunk_registry
, session_id
,
4724 chunk_id
, &chunk_exists_local
);
4726 /* Internal error. */
4727 ERR("Failed to query the existence of a trace chunk");
4728 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
4731 DBG("Trace chunk %s locally",
4732 chunk_exists_local
? "exists" : "does not exist");
4733 if (chunk_exists_local
) {
4734 ret_code
= LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_LOCAL
;
4736 } else if (is_local_trace
) {
4737 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4742 relayd
= consumer_find_relayd(*relayd_id
);
4744 ERR("Failed to find relayd %" PRIu64
, *relayd_id
);
4745 ret_code
= LTTCOMM_CONSUMERD_INVALID_PARAMETERS
;
4746 goto end_rcu_unlock
;
4748 DBG("Looking up existence of trace chunk on relay daemon");
4749 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4750 ret
= relayd_trace_chunk_exists(&relayd
->control_sock
, chunk_id
,
4751 &chunk_exists_remote
);
4752 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4754 ERR("Failed to look-up the existence of trace chunk on relay daemon");
4755 ret_code
= LTTCOMM_CONSUMERD_RELAYD_FAIL
;
4756 goto end_rcu_unlock
;
4759 ret_code
= chunk_exists_remote
?
4760 LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_REMOTE
:
4761 LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4762 DBG("Trace chunk %s on relay daemon",
4763 chunk_exists_remote
? "exists" : "does not exist");