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>
56 struct lttng_consumer_global_data consumer_data
= {
59 .type
= LTTNG_CONSUMER_UNKNOWN
,
62 enum consumer_channel_action
{
65 CONSUMER_CHANNEL_QUIT
,
68 struct consumer_channel_msg
{
69 enum consumer_channel_action action
;
70 struct lttng_consumer_channel
*chan
; /* add */
71 uint64_t key
; /* del */
74 /* Flag used to temporarily pause data consumption from testpoints. */
75 int data_consumption_paused
;
78 * Flag to inform the polling thread to quit when all fd hung up. Updated by
79 * the consumer_thread_receive_fds when it notices that all fds has hung up.
80 * Also updated by the signal handler (consumer_should_exit()). Read by the
86 * Global hash table containing respectively metadata and data streams. The
87 * stream element in this ht should only be updated by the metadata poll thread
88 * for the metadata and the data poll thread for the data.
90 static struct lttng_ht
*metadata_ht
;
91 static struct lttng_ht
*data_ht
;
94 * Notify a thread lttng pipe to poll back again. This usually means that some
95 * global state has changed so we just send back the thread in a poll wait
98 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
100 struct lttng_consumer_stream
*null_stream
= NULL
;
104 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
107 static void notify_health_quit_pipe(int *pipe
)
111 ret
= lttng_write(pipe
[1], "4", 1);
113 PERROR("write consumer health quit");
117 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
118 struct lttng_consumer_channel
*chan
,
120 enum consumer_channel_action action
)
122 struct consumer_channel_msg msg
;
125 memset(&msg
, 0, sizeof(msg
));
130 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
131 if (ret
< sizeof(msg
)) {
132 PERROR("notify_channel_pipe write error");
136 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
139 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
142 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
143 struct lttng_consumer_channel
**chan
,
145 enum consumer_channel_action
*action
)
147 struct consumer_channel_msg msg
;
150 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
151 if (ret
< sizeof(msg
)) {
155 *action
= msg
.action
;
163 * Cleanup the stream list of a channel. Those streams are not yet globally
166 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
168 struct lttng_consumer_stream
*stream
, *stmp
;
172 /* Delete streams that might have been left in the stream list. */
173 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
175 cds_list_del(&stream
->send_node
);
177 * Once a stream is added to this list, the buffers were created so we
178 * have a guarantee that this call will succeed. Setting the monitor
179 * mode to 0 so we don't lock nor try to delete the stream from the
183 consumer_stream_destroy(stream
, NULL
);
188 * Find a stream. The consumer_data.lock must be locked during this
191 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
194 struct lttng_ht_iter iter
;
195 struct lttng_ht_node_u64
*node
;
196 struct lttng_consumer_stream
*stream
= NULL
;
200 /* -1ULL keys are lookup failures */
201 if (key
== (uint64_t) -1ULL) {
207 lttng_ht_lookup(ht
, &key
, &iter
);
208 node
= lttng_ht_iter_get_node_u64(&iter
);
210 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
218 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
220 struct lttng_consumer_stream
*stream
;
223 stream
= find_stream(key
, ht
);
225 stream
->key
= (uint64_t) -1ULL;
227 * We don't want the lookup to match, but we still need
228 * to iterate on this stream when iterating over the hash table. Just
229 * change the node key.
231 stream
->node
.key
= (uint64_t) -1ULL;
237 * Return a channel object for the given key.
239 * RCU read side lock MUST be acquired before calling this function and
240 * protects the channel ptr.
242 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
244 struct lttng_ht_iter iter
;
245 struct lttng_ht_node_u64
*node
;
246 struct lttng_consumer_channel
*channel
= NULL
;
248 /* -1ULL keys are lookup failures */
249 if (key
== (uint64_t) -1ULL) {
253 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
254 node
= lttng_ht_iter_get_node_u64(&iter
);
256 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
263 * There is a possibility that the consumer does not have enough time between
264 * the close of the channel on the session daemon and the cleanup in here thus
265 * once we have a channel add with an existing key, we know for sure that this
266 * channel will eventually get cleaned up by all streams being closed.
268 * This function just nullifies the already existing channel key.
270 static void steal_channel_key(uint64_t key
)
272 struct lttng_consumer_channel
*channel
;
275 channel
= consumer_find_channel(key
);
277 channel
->key
= (uint64_t) -1ULL;
279 * We don't want the lookup to match, but we still need to iterate on
280 * this channel when iterating over the hash table. Just change the
283 channel
->node
.key
= (uint64_t) -1ULL;
288 static void free_channel_rcu(struct rcu_head
*head
)
290 struct lttng_ht_node_u64
*node
=
291 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
292 struct lttng_consumer_channel
*channel
=
293 caa_container_of(node
, struct lttng_consumer_channel
, node
);
295 switch (consumer_data
.type
) {
296 case LTTNG_CONSUMER_KERNEL
:
298 case LTTNG_CONSUMER32_UST
:
299 case LTTNG_CONSUMER64_UST
:
300 lttng_ustconsumer_free_channel(channel
);
303 ERR("Unknown consumer_data type");
310 * RCU protected relayd socket pair free.
312 static void free_relayd_rcu(struct rcu_head
*head
)
314 struct lttng_ht_node_u64
*node
=
315 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
316 struct consumer_relayd_sock_pair
*relayd
=
317 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
320 * Close all sockets. This is done in the call RCU since we don't want the
321 * socket fds to be reassigned thus potentially creating bad state of the
324 * We do not have to lock the control socket mutex here since at this stage
325 * there is no one referencing to this relayd object.
327 (void) relayd_close(&relayd
->control_sock
);
328 (void) relayd_close(&relayd
->data_sock
);
330 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
335 * Destroy and free relayd socket pair object.
337 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
340 struct lttng_ht_iter iter
;
342 if (relayd
== NULL
) {
346 DBG("Consumer destroy and close relayd socket pair");
348 iter
.iter
.node
= &relayd
->node
.node
;
349 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
351 /* We assume the relayd is being or is destroyed */
355 /* RCU free() call */
356 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
360 * Remove a channel from the global list protected by a mutex. This function is
361 * also responsible for freeing its data structures.
363 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
365 struct lttng_ht_iter iter
;
367 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
369 pthread_mutex_lock(&consumer_data
.lock
);
370 pthread_mutex_lock(&channel
->lock
);
372 /* Destroy streams that might have been left in the stream list. */
373 clean_channel_stream_list(channel
);
375 if (channel
->live_timer_enabled
== 1) {
376 consumer_timer_live_stop(channel
);
378 if (channel
->monitor_timer_enabled
== 1) {
379 consumer_timer_monitor_stop(channel
);
382 switch (consumer_data
.type
) {
383 case LTTNG_CONSUMER_KERNEL
:
385 case LTTNG_CONSUMER32_UST
:
386 case LTTNG_CONSUMER64_UST
:
387 lttng_ustconsumer_del_channel(channel
);
390 ERR("Unknown consumer_data type");
395 lttng_trace_chunk_put(channel
->trace_chunk
);
396 channel
->trace_chunk
= NULL
;
398 if (channel
->is_published
) {
402 iter
.iter
.node
= &channel
->node
.node
;
403 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
406 iter
.iter
.node
= &channel
->channels_by_session_id_ht_node
.node
;
407 ret
= lttng_ht_del(consumer_data
.channels_by_session_id_ht
,
413 call_rcu(&channel
->node
.head
, free_channel_rcu
);
415 pthread_mutex_unlock(&channel
->lock
);
416 pthread_mutex_unlock(&consumer_data
.lock
);
420 * Iterate over the relayd hash table and destroy each element. Finally,
421 * destroy the whole hash table.
423 static void cleanup_relayd_ht(void)
425 struct lttng_ht_iter iter
;
426 struct consumer_relayd_sock_pair
*relayd
;
430 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
432 consumer_destroy_relayd(relayd
);
437 lttng_ht_destroy(consumer_data
.relayd_ht
);
441 * Update the end point status of all streams having the given network sequence
442 * index (relayd index).
444 * It's atomically set without having the stream mutex locked which is fine
445 * because we handle the write/read race with a pipe wakeup for each thread.
447 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
448 enum consumer_endpoint_status status
)
450 struct lttng_ht_iter iter
;
451 struct lttng_consumer_stream
*stream
;
453 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
457 /* Let's begin with metadata */
458 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
459 if (stream
->net_seq_idx
== net_seq_idx
) {
460 uatomic_set(&stream
->endpoint_status
, status
);
461 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
465 /* Follow up by the data streams */
466 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
467 if (stream
->net_seq_idx
== net_seq_idx
) {
468 uatomic_set(&stream
->endpoint_status
, status
);
469 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
476 * Cleanup a relayd object by flagging every associated streams for deletion,
477 * destroying the object meaning removing it from the relayd hash table,
478 * closing the sockets and freeing the memory in a RCU call.
480 * If a local data context is available, notify the threads that the streams'
481 * state have changed.
483 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
489 DBG("Cleaning up relayd object ID %"PRIu64
, relayd
->net_seq_idx
);
491 /* Save the net sequence index before destroying the object */
492 netidx
= relayd
->net_seq_idx
;
495 * Delete the relayd from the relayd hash table, close the sockets and free
496 * the object in a RCU call.
498 consumer_destroy_relayd(relayd
);
500 /* Set inactive endpoint to all streams */
501 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
504 * With a local data context, notify the threads that the streams' state
505 * have changed. The write() action on the pipe acts as an "implicit"
506 * memory barrier ordering the updates of the end point status from the
507 * read of this status which happens AFTER receiving this notify.
509 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
510 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
514 * Flag a relayd socket pair for destruction. Destroy it if the refcount
517 * RCU read side lock MUST be aquired before calling this function.
519 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
523 /* Set destroy flag for this object */
524 uatomic_set(&relayd
->destroy_flag
, 1);
526 /* Destroy the relayd if refcount is 0 */
527 if (uatomic_read(&relayd
->refcount
) == 0) {
528 consumer_destroy_relayd(relayd
);
533 * Completly destroy stream from every visiable data structure and the given
536 * One this call returns, the stream object is not longer usable nor visible.
538 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
541 consumer_stream_destroy(stream
, ht
);
545 * XXX naming of del vs destroy is all mixed up.
547 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
549 consumer_stream_destroy(stream
, data_ht
);
552 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
554 consumer_stream_destroy(stream
, metadata_ht
);
557 void consumer_stream_update_channel_attributes(
558 struct lttng_consumer_stream
*stream
,
559 struct lttng_consumer_channel
*channel
)
561 stream
->channel_read_only_attributes
.tracefile_size
=
562 channel
->tracefile_size
;
565 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
567 const char *channel_name
,
570 struct lttng_trace_chunk
*trace_chunk
,
573 enum consumer_channel_type type
,
574 unsigned int monitor
)
577 struct lttng_consumer_stream
*stream
;
579 stream
= zmalloc(sizeof(*stream
));
580 if (stream
== NULL
) {
581 PERROR("malloc struct lttng_consumer_stream");
586 if (trace_chunk
&& !lttng_trace_chunk_get(trace_chunk
)) {
587 ERR("Failed to acquire trace chunk reference during the creation of a stream");
593 stream
->key
= stream_key
;
594 stream
->trace_chunk
= trace_chunk
;
596 stream
->out_fd_offset
= 0;
597 stream
->output_written
= 0;
598 stream
->net_seq_idx
= relayd_id
;
599 stream
->session_id
= session_id
;
600 stream
->monitor
= monitor
;
601 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
602 stream
->index_file
= NULL
;
603 stream
->last_sequence_number
= -1ULL;
604 pthread_mutex_init(&stream
->lock
, NULL
);
605 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
607 /* If channel is the metadata, flag this stream as metadata. */
608 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
609 stream
->metadata_flag
= 1;
610 /* Metadata is flat out. */
611 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
612 /* Live rendez-vous point. */
613 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
614 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
616 /* Format stream name to <channel_name>_<cpu_number> */
617 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
620 PERROR("snprintf stream name");
625 /* Key is always the wait_fd for streams. */
626 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
628 /* Init node per channel id key */
629 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
631 /* Init session id node with the stream session id */
632 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
634 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
635 " relayd_id %" PRIu64
", session_id %" PRIu64
,
636 stream
->name
, stream
->key
, channel_key
,
637 stream
->net_seq_idx
, stream
->session_id
);
644 lttng_trace_chunk_put(stream
->trace_chunk
);
654 * Add a stream to the global list protected by a mutex.
656 void consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
658 struct lttng_ht
*ht
= data_ht
;
663 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
665 pthread_mutex_lock(&consumer_data
.lock
);
666 pthread_mutex_lock(&stream
->chan
->lock
);
667 pthread_mutex_lock(&stream
->chan
->timer_lock
);
668 pthread_mutex_lock(&stream
->lock
);
671 /* Steal stream identifier to avoid having streams with the same key */
672 steal_stream_key(stream
->key
, ht
);
674 lttng_ht_add_unique_u64(ht
, &stream
->node
);
676 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
677 &stream
->node_channel_id
);
680 * Add stream to the stream_list_ht of the consumer data. No need to steal
681 * the key since the HT does not use it and we allow to add redundant keys
684 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
687 * When nb_init_stream_left reaches 0, we don't need to trigger any action
688 * in terms of destroying the associated channel, because the action that
689 * causes the count to become 0 also causes a stream to be added. The
690 * channel deletion will thus be triggered by the following removal of this
693 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
694 /* Increment refcount before decrementing nb_init_stream_left */
696 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
699 /* Update consumer data once the node is inserted. */
700 consumer_data
.stream_count
++;
701 consumer_data
.need_update
= 1;
704 pthread_mutex_unlock(&stream
->lock
);
705 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
706 pthread_mutex_unlock(&stream
->chan
->lock
);
707 pthread_mutex_unlock(&consumer_data
.lock
);
710 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
712 consumer_del_stream(stream
, data_ht
);
716 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
717 * be acquired before calling this.
719 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
722 struct lttng_ht_node_u64
*node
;
723 struct lttng_ht_iter iter
;
727 lttng_ht_lookup(consumer_data
.relayd_ht
,
728 &relayd
->net_seq_idx
, &iter
);
729 node
= lttng_ht_iter_get_node_u64(&iter
);
733 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
740 * Allocate and return a consumer relayd socket.
742 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
743 uint64_t net_seq_idx
)
745 struct consumer_relayd_sock_pair
*obj
= NULL
;
747 /* net sequence index of -1 is a failure */
748 if (net_seq_idx
== (uint64_t) -1ULL) {
752 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
754 PERROR("zmalloc relayd sock");
758 obj
->net_seq_idx
= net_seq_idx
;
760 obj
->destroy_flag
= 0;
761 obj
->control_sock
.sock
.fd
= -1;
762 obj
->data_sock
.sock
.fd
= -1;
763 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
764 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
771 * Find a relayd socket pair in the global consumer data.
773 * Return the object if found else NULL.
774 * RCU read-side lock must be held across this call and while using the
777 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
779 struct lttng_ht_iter iter
;
780 struct lttng_ht_node_u64
*node
;
781 struct consumer_relayd_sock_pair
*relayd
= NULL
;
783 /* Negative keys are lookup failures */
784 if (key
== (uint64_t) -1ULL) {
788 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
790 node
= lttng_ht_iter_get_node_u64(&iter
);
792 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
800 * Find a relayd and send the stream
802 * Returns 0 on success, < 0 on error
804 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
808 struct consumer_relayd_sock_pair
*relayd
;
811 assert(stream
->net_seq_idx
!= -1ULL);
814 /* The stream is not metadata. Get relayd reference if exists. */
816 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
817 if (relayd
!= NULL
) {
818 /* Add stream on the relayd */
819 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
820 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
821 path
, &stream
->relayd_stream_id
,
822 stream
->chan
->tracefile_size
,
823 stream
->chan
->tracefile_count
,
824 stream
->trace_chunk
);
825 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
827 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
828 lttng_consumer_cleanup_relayd(relayd
);
832 uatomic_inc(&relayd
->refcount
);
833 stream
->sent_to_relayd
= 1;
835 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
836 stream
->key
, stream
->net_seq_idx
);
841 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
842 stream
->name
, stream
->key
, stream
->net_seq_idx
);
850 * Find a relayd and send the streams sent message
852 * Returns 0 on success, < 0 on error
854 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
857 struct consumer_relayd_sock_pair
*relayd
;
859 assert(net_seq_idx
!= -1ULL);
861 /* The stream is not metadata. Get relayd reference if exists. */
863 relayd
= consumer_find_relayd(net_seq_idx
);
864 if (relayd
!= NULL
) {
865 /* Add stream on the relayd */
866 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
867 ret
= relayd_streams_sent(&relayd
->control_sock
);
868 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
870 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
871 lttng_consumer_cleanup_relayd(relayd
);
875 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
882 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
890 * Find a relayd and close the stream
892 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
894 struct consumer_relayd_sock_pair
*relayd
;
896 /* The stream is not metadata. Get relayd reference if exists. */
898 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
900 consumer_stream_relayd_close(stream
, relayd
);
906 * Handle stream for relayd transmission if the stream applies for network
907 * streaming where the net sequence index is set.
909 * Return destination file descriptor or negative value on error.
911 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
912 size_t data_size
, unsigned long padding
,
913 struct consumer_relayd_sock_pair
*relayd
)
916 struct lttcomm_relayd_data_hdr data_hdr
;
922 /* Reset data header */
923 memset(&data_hdr
, 0, sizeof(data_hdr
));
925 if (stream
->metadata_flag
) {
926 /* Caller MUST acquire the relayd control socket lock */
927 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
932 /* Metadata are always sent on the control socket. */
933 outfd
= relayd
->control_sock
.sock
.fd
;
935 /* Set header with stream information */
936 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
937 data_hdr
.data_size
= htobe32(data_size
);
938 data_hdr
.padding_size
= htobe32(padding
);
940 * Note that net_seq_num below is assigned with the *current* value of
941 * next_net_seq_num and only after that the next_net_seq_num will be
942 * increment. This is why when issuing a command on the relayd using
943 * this next value, 1 should always be substracted in order to compare
944 * the last seen sequence number on the relayd side to the last sent.
946 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
947 /* Other fields are zeroed previously */
949 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
955 ++stream
->next_net_seq_num
;
957 /* Set to go on data socket */
958 outfd
= relayd
->data_sock
.sock
.fd
;
966 * Trigger a dump of the metadata content. Following/during the succesful
967 * completion of this call, the metadata poll thread will start receiving
968 * metadata packets to consume.
970 * The caller must hold the channel and stream locks.
973 int consumer_metadata_stream_dump(struct lttng_consumer_stream
*stream
)
977 ASSERT_LOCKED(stream
->chan
->lock
);
978 ASSERT_LOCKED(stream
->lock
);
979 assert(stream
->metadata_flag
);
980 assert(stream
->chan
->trace_chunk
);
982 switch (consumer_data
.type
) {
983 case LTTNG_CONSUMER_KERNEL
:
985 * Reset the position of what has been read from the
986 * metadata cache to 0 so we can dump it again.
988 ret
= kernctl_metadata_cache_dump(stream
->wait_fd
);
990 case LTTNG_CONSUMER32_UST
:
991 case LTTNG_CONSUMER64_UST
:
993 * Reset the position pushed from the metadata cache so it
994 * will write from the beginning on the next push.
996 stream
->ust_metadata_pushed
= 0;
997 ret
= consumer_metadata_wakeup_pipe(stream
->chan
);
1000 ERR("Unknown consumer_data type");
1004 ERR("Failed to dump the metadata cache");
1010 int lttng_consumer_channel_set_trace_chunk(
1011 struct lttng_consumer_channel
*channel
,
1012 struct lttng_trace_chunk
*new_trace_chunk
)
1015 const bool is_local_trace
= channel
->relayd_id
== -1ULL;
1016 bool update_stream_trace_chunk
;
1017 struct cds_lfht_iter iter
;
1018 struct lttng_consumer_stream
*stream
;
1019 unsigned long channel_hash
;
1021 pthread_mutex_lock(&channel
->lock
);
1023 * A stream can transition to a state where it and its channel
1024 * no longer belong to a trace chunk. For instance, this happens when
1025 * a session is rotated while it is inactive. After the rotation
1026 * of an inactive session completes, the channel and its streams no
1027 * longer belong to a trace chunk.
1029 * However, if a session is stopped, rotated, and started again,
1030 * the session daemon will create a new chunk and send it to its peers.
1031 * In that case, the streams' transition to a new chunk can be performed
1034 * This trace chunk transition could also be performed lazily when
1035 * a buffer is consumed. However, creating the files here allows the
1036 * consumer daemon to report any creation error to the session daemon
1037 * and cause the start of the tracing session to fail.
1039 update_stream_trace_chunk
= !channel
->trace_chunk
&& new_trace_chunk
;
1042 * The acquisition of the reference cannot fail (barring
1043 * a severe internal error) since a reference to the published
1044 * chunk is already held by the caller.
1046 if (new_trace_chunk
) {
1047 const bool acquired_reference
= lttng_trace_chunk_get(
1050 assert(acquired_reference
);
1053 lttng_trace_chunk_put(channel
->trace_chunk
);
1054 channel
->trace_chunk
= new_trace_chunk
;
1055 if (!is_local_trace
|| !new_trace_chunk
) {
1060 if (!update_stream_trace_chunk
) {
1064 channel_hash
= consumer_data
.stream_per_chan_id_ht
->hash_fct(
1065 &channel
->key
, lttng_ht_seed
);
1067 cds_lfht_for_each_entry_duplicate(consumer_data
.stream_per_chan_id_ht
->ht
,
1069 consumer_data
.stream_per_chan_id_ht
->match_fct
,
1070 &channel
->key
, &iter
, stream
, node_channel_id
.node
) {
1071 bool acquired_reference
, should_regenerate_metadata
= false;
1073 acquired_reference
= lttng_trace_chunk_get(channel
->trace_chunk
);
1074 assert(acquired_reference
);
1076 pthread_mutex_lock(&stream
->lock
);
1079 * On a transition from "no-chunk" to a new chunk, a metadata
1080 * stream's content must be entirely dumped. This must occcur
1081 * _after_ the creation of the metadata stream's output files
1082 * as the consumption thread (not necessarily the one executing
1083 * this) may start to consume during the call to
1084 * consumer_metadata_stream_dump().
1086 should_regenerate_metadata
=
1087 stream
->metadata_flag
&&
1088 !stream
->trace_chunk
&& channel
->trace_chunk
;
1089 stream
->trace_chunk
= channel
->trace_chunk
;
1090 ret
= consumer_stream_create_output_files(stream
, true);
1092 pthread_mutex_unlock(&stream
->lock
);
1093 goto end_rcu_unlock
;
1095 if (should_regenerate_metadata
) {
1096 ret
= consumer_metadata_stream_dump(stream
);
1098 pthread_mutex_unlock(&stream
->lock
);
1100 goto end_rcu_unlock
;
1106 pthread_mutex_unlock(&channel
->lock
);
1111 * Allocate and return a new lttng_consumer_channel object using the given key
1112 * to initialize the hash table node.
1114 * On error, return NULL.
1116 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
1117 uint64_t session_id
,
1118 const uint64_t *chunk_id
,
1119 const char *pathname
,
1122 enum lttng_event_output output
,
1123 uint64_t tracefile_size
,
1124 uint64_t tracefile_count
,
1125 uint64_t session_id_per_pid
,
1126 unsigned int monitor
,
1127 unsigned int live_timer_interval
,
1128 const char *root_shm_path
,
1129 const char *shm_path
)
1131 struct lttng_consumer_channel
*channel
= NULL
;
1132 struct lttng_trace_chunk
*trace_chunk
= NULL
;
1135 trace_chunk
= lttng_trace_chunk_registry_find_chunk(
1136 consumer_data
.chunk_registry
, session_id
,
1139 ERR("Failed to find trace chunk reference during creation of channel");
1144 channel
= zmalloc(sizeof(*channel
));
1145 if (channel
== NULL
) {
1146 PERROR("malloc struct lttng_consumer_channel");
1151 channel
->refcount
= 0;
1152 channel
->session_id
= session_id
;
1153 channel
->session_id_per_pid
= session_id_per_pid
;
1154 channel
->relayd_id
= relayd_id
;
1155 channel
->tracefile_size
= tracefile_size
;
1156 channel
->tracefile_count
= tracefile_count
;
1157 channel
->monitor
= monitor
;
1158 channel
->live_timer_interval
= live_timer_interval
;
1159 pthread_mutex_init(&channel
->lock
, NULL
);
1160 pthread_mutex_init(&channel
->timer_lock
, NULL
);
1163 case LTTNG_EVENT_SPLICE
:
1164 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
1166 case LTTNG_EVENT_MMAP
:
1167 channel
->output
= CONSUMER_CHANNEL_MMAP
;
1177 * In monitor mode, the streams associated with the channel will be put in
1178 * a special list ONLY owned by this channel. So, the refcount is set to 1
1179 * here meaning that the channel itself has streams that are referenced.
1181 * On a channel deletion, once the channel is no longer visible, the
1182 * refcount is decremented and checked for a zero value to delete it. With
1183 * streams in no monitor mode, it will now be safe to destroy the channel.
1185 if (!channel
->monitor
) {
1186 channel
->refcount
= 1;
1189 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1190 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1192 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1193 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1195 if (root_shm_path
) {
1196 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1197 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1200 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1201 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1204 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1205 lttng_ht_node_init_u64(&channel
->channels_by_session_id_ht_node
,
1206 channel
->session_id
);
1208 channel
->wait_fd
= -1;
1209 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1212 int ret
= lttng_consumer_channel_set_trace_chunk(channel
,
1219 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1222 lttng_trace_chunk_put(trace_chunk
);
1225 consumer_del_channel(channel
);
1231 * Add a channel to the global list protected by a mutex.
1233 * Always return 0 indicating success.
1235 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1236 struct lttng_consumer_local_data
*ctx
)
1238 pthread_mutex_lock(&consumer_data
.lock
);
1239 pthread_mutex_lock(&channel
->lock
);
1240 pthread_mutex_lock(&channel
->timer_lock
);
1243 * This gives us a guarantee that the channel we are about to add to the
1244 * channel hash table will be unique. See this function comment on the why
1245 * we need to steel the channel key at this stage.
1247 steal_channel_key(channel
->key
);
1250 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1251 lttng_ht_add_u64(consumer_data
.channels_by_session_id_ht
,
1252 &channel
->channels_by_session_id_ht_node
);
1254 channel
->is_published
= true;
1256 pthread_mutex_unlock(&channel
->timer_lock
);
1257 pthread_mutex_unlock(&channel
->lock
);
1258 pthread_mutex_unlock(&consumer_data
.lock
);
1260 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1261 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1268 * Allocate the pollfd structure and the local view of the out fds to avoid
1269 * doing a lookup in the linked list and concurrency issues when writing is
1270 * needed. Called with consumer_data.lock held.
1272 * Returns the number of fds in the structures.
1274 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1275 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1276 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1279 struct lttng_ht_iter iter
;
1280 struct lttng_consumer_stream
*stream
;
1285 assert(local_stream
);
1287 DBG("Updating poll fd array");
1288 *nb_inactive_fd
= 0;
1290 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1292 * Only active streams with an active end point can be added to the
1293 * poll set and local stream storage of the thread.
1295 * There is a potential race here for endpoint_status to be updated
1296 * just after the check. However, this is OK since the stream(s) will
1297 * be deleted once the thread is notified that the end point state has
1298 * changed where this function will be called back again.
1300 * We track the number of inactive FDs because they still need to be
1301 * closed by the polling thread after a wakeup on the data_pipe or
1304 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1305 (*nb_inactive_fd
)++;
1309 * This clobbers way too much the debug output. Uncomment that if you
1310 * need it for debugging purposes.
1312 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1313 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1314 local_stream
[i
] = stream
;
1320 * Insert the consumer_data_pipe at the end of the array and don't
1321 * increment i so nb_fd is the number of real FD.
1323 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1324 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1326 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1327 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1332 * Poll on the should_quit pipe and the command socket return -1 on
1333 * error, 1 if should exit, 0 if data is available on the command socket
1335 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1340 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1341 if (num_rdy
== -1) {
1343 * Restart interrupted system call.
1345 if (errno
== EINTR
) {
1348 PERROR("Poll error");
1351 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1352 DBG("consumer_should_quit wake up");
1359 * Set the error socket.
1361 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1364 ctx
->consumer_error_socket
= sock
;
1368 * Set the command socket path.
1370 void lttng_consumer_set_command_sock_path(
1371 struct lttng_consumer_local_data
*ctx
, char *sock
)
1373 ctx
->consumer_command_sock_path
= sock
;
1377 * Send return code to the session daemon.
1378 * If the socket is not defined, we return 0, it is not a fatal error
1380 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1382 if (ctx
->consumer_error_socket
> 0) {
1383 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1384 sizeof(enum lttcomm_sessiond_command
));
1391 * Close all the tracefiles and stream fds and MUST be called when all
1392 * instances are destroyed i.e. when all threads were joined and are ended.
1394 void lttng_consumer_cleanup(void)
1396 struct lttng_ht_iter iter
;
1397 struct lttng_consumer_channel
*channel
;
1401 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1403 consumer_del_channel(channel
);
1408 lttng_ht_destroy(consumer_data
.channel_ht
);
1409 lttng_ht_destroy(consumer_data
.channels_by_session_id_ht
);
1411 cleanup_relayd_ht();
1413 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1416 * This HT contains streams that are freed by either the metadata thread or
1417 * the data thread so we do *nothing* on the hash table and simply destroy
1420 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1422 lttng_trace_chunk_registry_destroy(consumer_data
.chunk_registry
);
1426 * Called from signal handler.
1428 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1432 CMM_STORE_SHARED(consumer_quit
, 1);
1433 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1435 PERROR("write consumer quit");
1438 DBG("Consumer flag that it should quit");
1443 * Flush pending writes to trace output disk file.
1446 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1450 int outfd
= stream
->out_fd
;
1453 * This does a blocking write-and-wait on any page that belongs to the
1454 * subbuffer prior to the one we just wrote.
1455 * Don't care about error values, as these are just hints and ways to
1456 * limit the amount of page cache used.
1458 if (orig_offset
< stream
->max_sb_size
) {
1461 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1462 stream
->max_sb_size
,
1463 SYNC_FILE_RANGE_WAIT_BEFORE
1464 | SYNC_FILE_RANGE_WRITE
1465 | SYNC_FILE_RANGE_WAIT_AFTER
);
1467 * Give hints to the kernel about how we access the file:
1468 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1471 * We need to call fadvise again after the file grows because the
1472 * kernel does not seem to apply fadvise to non-existing parts of the
1475 * Call fadvise _after_ having waited for the page writeback to
1476 * complete because the dirty page writeback semantic is not well
1477 * defined. So it can be expected to lead to lower throughput in
1480 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1481 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1482 if (ret
&& ret
!= -ENOSYS
) {
1484 PERROR("posix_fadvise on fd %i", outfd
);
1489 * Initialise the necessary environnement :
1490 * - create a new context
1491 * - create the poll_pipe
1492 * - create the should_quit pipe (for signal handler)
1493 * - create the thread pipe (for splice)
1495 * Takes a function pointer as argument, this function is called when data is
1496 * available on a buffer. This function is responsible to do the
1497 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1498 * buffer configuration and then kernctl_put_next_subbuf at the end.
1500 * Returns a pointer to the new context or NULL on error.
1502 struct lttng_consumer_local_data
*lttng_consumer_create(
1503 enum lttng_consumer_type type
,
1504 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1505 struct lttng_consumer_local_data
*ctx
),
1506 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1507 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1508 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1511 struct lttng_consumer_local_data
*ctx
;
1513 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1514 consumer_data
.type
== type
);
1515 consumer_data
.type
= type
;
1517 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1519 PERROR("allocating context");
1523 ctx
->consumer_error_socket
= -1;
1524 ctx
->consumer_metadata_socket
= -1;
1525 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1526 /* assign the callbacks */
1527 ctx
->on_buffer_ready
= buffer_ready
;
1528 ctx
->on_recv_channel
= recv_channel
;
1529 ctx
->on_recv_stream
= recv_stream
;
1530 ctx
->on_update_stream
= update_stream
;
1532 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1533 if (!ctx
->consumer_data_pipe
) {
1534 goto error_poll_pipe
;
1537 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1538 if (!ctx
->consumer_wakeup_pipe
) {
1539 goto error_wakeup_pipe
;
1542 ret
= pipe(ctx
->consumer_should_quit
);
1544 PERROR("Error creating recv pipe");
1545 goto error_quit_pipe
;
1548 ret
= pipe(ctx
->consumer_channel_pipe
);
1550 PERROR("Error creating channel pipe");
1551 goto error_channel_pipe
;
1554 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1555 if (!ctx
->consumer_metadata_pipe
) {
1556 goto error_metadata_pipe
;
1559 ctx
->channel_monitor_pipe
= -1;
1563 error_metadata_pipe
:
1564 utils_close_pipe(ctx
->consumer_channel_pipe
);
1566 utils_close_pipe(ctx
->consumer_should_quit
);
1568 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1570 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1578 * Iterate over all streams of the hashtable and free them properly.
1580 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1582 struct lttng_ht_iter iter
;
1583 struct lttng_consumer_stream
*stream
;
1590 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1592 * Ignore return value since we are currently cleaning up so any error
1595 (void) consumer_del_stream(stream
, ht
);
1599 lttng_ht_destroy(ht
);
1603 * Iterate over all streams of the metadata hashtable and free them
1606 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1608 struct lttng_ht_iter iter
;
1609 struct lttng_consumer_stream
*stream
;
1616 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1618 * Ignore return value since we are currently cleaning up so any error
1621 (void) consumer_del_metadata_stream(stream
, ht
);
1625 lttng_ht_destroy(ht
);
1629 * Close all fds associated with the instance and free the context.
1631 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1635 DBG("Consumer destroying it. Closing everything.");
1641 destroy_data_stream_ht(data_ht
);
1642 destroy_metadata_stream_ht(metadata_ht
);
1644 ret
= close(ctx
->consumer_error_socket
);
1648 ret
= close(ctx
->consumer_metadata_socket
);
1652 utils_close_pipe(ctx
->consumer_channel_pipe
);
1653 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1654 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1655 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1656 utils_close_pipe(ctx
->consumer_should_quit
);
1658 unlink(ctx
->consumer_command_sock_path
);
1663 * Write the metadata stream id on the specified file descriptor.
1665 static int write_relayd_metadata_id(int fd
,
1666 struct lttng_consumer_stream
*stream
,
1667 unsigned long padding
)
1670 struct lttcomm_relayd_metadata_payload hdr
;
1672 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1673 hdr
.padding_size
= htobe32(padding
);
1674 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1675 if (ret
< sizeof(hdr
)) {
1677 * This error means that the fd's end is closed so ignore the PERROR
1678 * not to clubber the error output since this can happen in a normal
1681 if (errno
!= EPIPE
) {
1682 PERROR("write metadata stream id");
1684 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1686 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1687 * handle writting the missing part so report that as an error and
1688 * don't lie to the caller.
1693 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1694 stream
->relayd_stream_id
, padding
);
1701 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1702 * core function for writing trace buffers to either the local filesystem or
1705 * It must be called with the stream and the channel lock held.
1707 * Careful review MUST be put if any changes occur!
1709 * Returns the number of bytes written
1711 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1712 struct lttng_consumer_local_data
*ctx
,
1713 struct lttng_consumer_stream
*stream
, unsigned long len
,
1714 unsigned long padding
,
1715 struct ctf_packet_index
*index
)
1717 unsigned long mmap_offset
;
1720 off_t orig_offset
= stream
->out_fd_offset
;
1721 /* Default is on the disk */
1722 int outfd
= stream
->out_fd
;
1723 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1724 unsigned int relayd_hang_up
= 0;
1726 /* RCU lock for the relayd pointer */
1729 assert(stream
->net_seq_idx
!= (uint64_t) -1ULL ||
1730 stream
->chan
->trace_chunk
);
1732 /* Flag that the current stream if set for network streaming. */
1733 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1734 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1735 if (relayd
== NULL
) {
1741 /* get the offset inside the fd to mmap */
1742 switch (consumer_data
.type
) {
1743 case LTTNG_CONSUMER_KERNEL
:
1744 mmap_base
= stream
->mmap_base
;
1745 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1747 PERROR("tracer ctl get_mmap_read_offset");
1751 case LTTNG_CONSUMER32_UST
:
1752 case LTTNG_CONSUMER64_UST
:
1753 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1755 ERR("read mmap get mmap base for stream %s", stream
->name
);
1759 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1761 PERROR("tracer ctl get_mmap_read_offset");
1767 ERR("Unknown consumer_data type");
1771 /* Handle stream on the relayd if the output is on the network */
1773 unsigned long netlen
= len
;
1776 * Lock the control socket for the complete duration of the function
1777 * since from this point on we will use the socket.
1779 if (stream
->metadata_flag
) {
1780 /* Metadata requires the control socket. */
1781 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1782 if (stream
->reset_metadata_flag
) {
1783 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1784 stream
->relayd_stream_id
,
1785 stream
->metadata_version
);
1790 stream
->reset_metadata_flag
= 0;
1792 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1795 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1800 /* Use the returned socket. */
1803 /* Write metadata stream id before payload */
1804 if (stream
->metadata_flag
) {
1805 ret
= write_relayd_metadata_id(outfd
, stream
, padding
);
1812 /* No streaming, we have to set the len with the full padding */
1815 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1816 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1818 ERR("Reset metadata file");
1821 stream
->reset_metadata_flag
= 0;
1825 * Check if we need to change the tracefile before writing the packet.
1827 if (stream
->chan
->tracefile_size
> 0 &&
1828 (stream
->tracefile_size_current
+ len
) >
1829 stream
->chan
->tracefile_size
) {
1830 ret
= consumer_stream_rotate_output_files(stream
);
1834 outfd
= stream
->out_fd
;
1837 stream
->tracefile_size_current
+= len
;
1839 index
->offset
= htobe64(stream
->out_fd_offset
);
1844 * This call guarantee that len or less is returned. It's impossible to
1845 * receive a ret value that is bigger than len.
1847 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1848 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1849 if (ret
< 0 || ((size_t) ret
!= len
)) {
1851 * Report error to caller if nothing was written else at least send the
1859 /* Socket operation failed. We consider the relayd dead */
1860 if (errno
== EPIPE
) {
1862 * This is possible if the fd is closed on the other side
1863 * (outfd) or any write problem. It can be verbose a bit for a
1864 * normal execution if for instance the relayd is stopped
1865 * abruptly. This can happen so set this to a DBG statement.
1867 DBG("Consumer mmap write detected relayd hang up");
1869 /* Unhandled error, print it and stop function right now. */
1870 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1874 stream
->output_written
+= ret
;
1876 /* This call is useless on a socket so better save a syscall. */
1878 /* This won't block, but will start writeout asynchronously */
1879 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1880 SYNC_FILE_RANGE_WRITE
);
1881 stream
->out_fd_offset
+= len
;
1882 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1887 * This is a special case that the relayd has closed its socket. Let's
1888 * cleanup the relayd object and all associated streams.
1890 if (relayd
&& relayd_hang_up
) {
1891 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1892 lttng_consumer_cleanup_relayd(relayd
);
1896 /* Unlock only if ctrl socket used */
1897 if (relayd
&& stream
->metadata_flag
) {
1898 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1906 * Splice the data from the ring buffer to the tracefile.
1908 * It must be called with the stream lock held.
1910 * Returns the number of bytes spliced.
1912 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1913 struct lttng_consumer_local_data
*ctx
,
1914 struct lttng_consumer_stream
*stream
, unsigned long len
,
1915 unsigned long padding
,
1916 struct ctf_packet_index
*index
)
1918 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1920 off_t orig_offset
= stream
->out_fd_offset
;
1921 int fd
= stream
->wait_fd
;
1922 /* Default is on the disk */
1923 int outfd
= stream
->out_fd
;
1924 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1926 unsigned int relayd_hang_up
= 0;
1928 switch (consumer_data
.type
) {
1929 case LTTNG_CONSUMER_KERNEL
:
1931 case LTTNG_CONSUMER32_UST
:
1932 case LTTNG_CONSUMER64_UST
:
1933 /* Not supported for user space tracing */
1936 ERR("Unknown consumer_data type");
1940 /* RCU lock for the relayd pointer */
1943 /* Flag that the current stream if set for network streaming. */
1944 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1945 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1946 if (relayd
== NULL
) {
1951 splice_pipe
= stream
->splice_pipe
;
1953 /* Write metadata stream id before payload */
1955 unsigned long total_len
= len
;
1957 if (stream
->metadata_flag
) {
1959 * Lock the control socket for the complete duration of the function
1960 * since from this point on we will use the socket.
1962 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1964 if (stream
->reset_metadata_flag
) {
1965 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1966 stream
->relayd_stream_id
,
1967 stream
->metadata_version
);
1972 stream
->reset_metadata_flag
= 0;
1974 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
,
1982 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1985 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1991 /* Use the returned socket. */
1994 /* No streaming, we have to set the len with the full padding */
1997 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1998 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
2000 ERR("Reset metadata file");
2003 stream
->reset_metadata_flag
= 0;
2006 * Check if we need to change the tracefile before writing the packet.
2008 if (stream
->chan
->tracefile_size
> 0 &&
2009 (stream
->tracefile_size_current
+ len
) >
2010 stream
->chan
->tracefile_size
) {
2011 ret
= consumer_stream_rotate_output_files(stream
);
2016 outfd
= stream
->out_fd
;
2019 stream
->tracefile_size_current
+= len
;
2020 index
->offset
= htobe64(stream
->out_fd_offset
);
2024 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
2025 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
2026 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
2027 SPLICE_F_MOVE
| SPLICE_F_MORE
);
2028 DBG("splice chan to pipe, ret %zd", ret_splice
);
2029 if (ret_splice
< 0) {
2032 PERROR("Error in relay splice");
2036 /* Handle stream on the relayd if the output is on the network */
2037 if (relayd
&& stream
->metadata_flag
) {
2038 size_t metadata_payload_size
=
2039 sizeof(struct lttcomm_relayd_metadata_payload
);
2041 /* Update counter to fit the spliced data */
2042 ret_splice
+= metadata_payload_size
;
2043 len
+= metadata_payload_size
;
2045 * We do this so the return value can match the len passed as
2046 * argument to this function.
2048 written
-= metadata_payload_size
;
2051 /* Splice data out */
2052 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
2053 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
2054 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
2056 if (ret_splice
< 0) {
2061 } else if (ret_splice
> len
) {
2063 * We don't expect this code path to be executed but you never know
2064 * so this is an extra protection agains a buggy splice().
2067 written
+= ret_splice
;
2068 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
2072 /* All good, update current len and continue. */
2076 /* This call is useless on a socket so better save a syscall. */
2078 /* This won't block, but will start writeout asynchronously */
2079 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
2080 SYNC_FILE_RANGE_WRITE
);
2081 stream
->out_fd_offset
+= ret_splice
;
2083 stream
->output_written
+= ret_splice
;
2084 written
+= ret_splice
;
2087 lttng_consumer_sync_trace_file(stream
, orig_offset
);
2093 * This is a special case that the relayd has closed its socket. Let's
2094 * cleanup the relayd object and all associated streams.
2096 if (relayd
&& relayd_hang_up
) {
2097 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
2098 lttng_consumer_cleanup_relayd(relayd
);
2099 /* Skip splice error so the consumer does not fail */
2104 /* send the appropriate error description to sessiond */
2107 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
2110 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
2113 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
2118 if (relayd
&& stream
->metadata_flag
) {
2119 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2127 * Sample the snapshot positions for a specific fd
2129 * Returns 0 on success, < 0 on error
2131 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream
*stream
)
2133 switch (consumer_data
.type
) {
2134 case LTTNG_CONSUMER_KERNEL
:
2135 return lttng_kconsumer_sample_snapshot_positions(stream
);
2136 case LTTNG_CONSUMER32_UST
:
2137 case LTTNG_CONSUMER64_UST
:
2138 return lttng_ustconsumer_sample_snapshot_positions(stream
);
2140 ERR("Unknown consumer_data type");
2146 * Take a snapshot for a specific fd
2148 * Returns 0 on success, < 0 on error
2150 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
2152 switch (consumer_data
.type
) {
2153 case LTTNG_CONSUMER_KERNEL
:
2154 return lttng_kconsumer_take_snapshot(stream
);
2155 case LTTNG_CONSUMER32_UST
:
2156 case LTTNG_CONSUMER64_UST
:
2157 return lttng_ustconsumer_take_snapshot(stream
);
2159 ERR("Unknown consumer_data type");
2166 * Get the produced position
2168 * Returns 0 on success, < 0 on error
2170 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
2173 switch (consumer_data
.type
) {
2174 case LTTNG_CONSUMER_KERNEL
:
2175 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
2176 case LTTNG_CONSUMER32_UST
:
2177 case LTTNG_CONSUMER64_UST
:
2178 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2180 ERR("Unknown consumer_data type");
2187 * Get the consumed position (free-running counter position in bytes).
2189 * Returns 0 on success, < 0 on error
2191 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream
*stream
,
2194 switch (consumer_data
.type
) {
2195 case LTTNG_CONSUMER_KERNEL
:
2196 return lttng_kconsumer_get_consumed_snapshot(stream
, pos
);
2197 case LTTNG_CONSUMER32_UST
:
2198 case LTTNG_CONSUMER64_UST
:
2199 return lttng_ustconsumer_get_consumed_snapshot(stream
, pos
);
2201 ERR("Unknown consumer_data type");
2207 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2208 int sock
, struct pollfd
*consumer_sockpoll
)
2210 switch (consumer_data
.type
) {
2211 case LTTNG_CONSUMER_KERNEL
:
2212 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2213 case LTTNG_CONSUMER32_UST
:
2214 case LTTNG_CONSUMER64_UST
:
2215 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2217 ERR("Unknown consumer_data type");
2223 void lttng_consumer_close_all_metadata(void)
2225 switch (consumer_data
.type
) {
2226 case LTTNG_CONSUMER_KERNEL
:
2228 * The Kernel consumer has a different metadata scheme so we don't
2229 * close anything because the stream will be closed by the session
2233 case LTTNG_CONSUMER32_UST
:
2234 case LTTNG_CONSUMER64_UST
:
2236 * Close all metadata streams. The metadata hash table is passed and
2237 * this call iterates over it by closing all wakeup fd. This is safe
2238 * because at this point we are sure that the metadata producer is
2239 * either dead or blocked.
2241 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2244 ERR("Unknown consumer_data type");
2250 * Clean up a metadata stream and free its memory.
2252 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2253 struct lttng_ht
*ht
)
2255 struct lttng_consumer_channel
*free_chan
= NULL
;
2259 * This call should NEVER receive regular stream. It must always be
2260 * metadata stream and this is crucial for data structure synchronization.
2262 assert(stream
->metadata_flag
);
2264 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2266 pthread_mutex_lock(&consumer_data
.lock
);
2267 pthread_mutex_lock(&stream
->chan
->lock
);
2268 pthread_mutex_lock(&stream
->lock
);
2269 if (stream
->chan
->metadata_cache
) {
2270 /* Only applicable to userspace consumers. */
2271 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
2274 /* Remove any reference to that stream. */
2275 consumer_stream_delete(stream
, ht
);
2277 /* Close down everything including the relayd if one. */
2278 consumer_stream_close(stream
);
2279 /* Destroy tracer buffers of the stream. */
2280 consumer_stream_destroy_buffers(stream
);
2282 /* Atomically decrement channel refcount since other threads can use it. */
2283 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2284 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2285 /* Go for channel deletion! */
2286 free_chan
= stream
->chan
;
2290 * Nullify the stream reference so it is not used after deletion. The
2291 * channel lock MUST be acquired before being able to check for a NULL
2294 stream
->chan
->metadata_stream
= NULL
;
2296 if (stream
->chan
->metadata_cache
) {
2297 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
2299 pthread_mutex_unlock(&stream
->lock
);
2300 pthread_mutex_unlock(&stream
->chan
->lock
);
2301 pthread_mutex_unlock(&consumer_data
.lock
);
2304 consumer_del_channel(free_chan
);
2307 lttng_trace_chunk_put(stream
->trace_chunk
);
2308 stream
->trace_chunk
= NULL
;
2309 consumer_stream_free(stream
);
2313 * Action done with the metadata stream when adding it to the consumer internal
2314 * data structures to handle it.
2316 void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2318 struct lttng_ht
*ht
= metadata_ht
;
2319 struct lttng_ht_iter iter
;
2320 struct lttng_ht_node_u64
*node
;
2325 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2327 pthread_mutex_lock(&consumer_data
.lock
);
2328 pthread_mutex_lock(&stream
->chan
->lock
);
2329 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2330 pthread_mutex_lock(&stream
->lock
);
2333 * From here, refcounts are updated so be _careful_ when returning an error
2340 * Lookup the stream just to make sure it does not exist in our internal
2341 * state. This should NEVER happen.
2343 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2344 node
= lttng_ht_iter_get_node_u64(&iter
);
2348 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2349 * in terms of destroying the associated channel, because the action that
2350 * causes the count to become 0 also causes a stream to be added. The
2351 * channel deletion will thus be triggered by the following removal of this
2354 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2355 /* Increment refcount before decrementing nb_init_stream_left */
2357 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2360 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2362 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2363 &stream
->node_channel_id
);
2366 * Add stream to the stream_list_ht of the consumer data. No need to steal
2367 * the key since the HT does not use it and we allow to add redundant keys
2370 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2374 pthread_mutex_unlock(&stream
->lock
);
2375 pthread_mutex_unlock(&stream
->chan
->lock
);
2376 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2377 pthread_mutex_unlock(&consumer_data
.lock
);
2381 * Delete data stream that are flagged for deletion (endpoint_status).
2383 static void validate_endpoint_status_data_stream(void)
2385 struct lttng_ht_iter iter
;
2386 struct lttng_consumer_stream
*stream
;
2388 DBG("Consumer delete flagged data stream");
2391 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2392 /* Validate delete flag of the stream */
2393 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2396 /* Delete it right now */
2397 consumer_del_stream(stream
, data_ht
);
2403 * Delete metadata stream that are flagged for deletion (endpoint_status).
2405 static void validate_endpoint_status_metadata_stream(
2406 struct lttng_poll_event
*pollset
)
2408 struct lttng_ht_iter iter
;
2409 struct lttng_consumer_stream
*stream
;
2411 DBG("Consumer delete flagged metadata stream");
2416 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2417 /* Validate delete flag of the stream */
2418 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2422 * Remove from pollset so the metadata thread can continue without
2423 * blocking on a deleted stream.
2425 lttng_poll_del(pollset
, stream
->wait_fd
);
2427 /* Delete it right now */
2428 consumer_del_metadata_stream(stream
, metadata_ht
);
2434 * Thread polls on metadata file descriptor and write them on disk or on the
2437 void *consumer_thread_metadata_poll(void *data
)
2439 int ret
, i
, pollfd
, err
= -1;
2440 uint32_t revents
, nb_fd
;
2441 struct lttng_consumer_stream
*stream
= NULL
;
2442 struct lttng_ht_iter iter
;
2443 struct lttng_ht_node_u64
*node
;
2444 struct lttng_poll_event events
;
2445 struct lttng_consumer_local_data
*ctx
= data
;
2448 rcu_register_thread();
2450 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2452 if (testpoint(consumerd_thread_metadata
)) {
2453 goto error_testpoint
;
2456 health_code_update();
2458 DBG("Thread metadata poll started");
2460 /* Size is set to 1 for the consumer_metadata pipe */
2461 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2463 ERR("Poll set creation failed");
2467 ret
= lttng_poll_add(&events
,
2468 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2474 DBG("Metadata main loop started");
2478 health_code_update();
2479 health_poll_entry();
2480 DBG("Metadata poll wait");
2481 ret
= lttng_poll_wait(&events
, -1);
2482 DBG("Metadata poll return from wait with %d fd(s)",
2483 LTTNG_POLL_GETNB(&events
));
2485 DBG("Metadata event caught in thread");
2487 if (errno
== EINTR
) {
2488 ERR("Poll EINTR caught");
2491 if (LTTNG_POLL_GETNB(&events
) == 0) {
2492 err
= 0; /* All is OK */
2499 /* From here, the event is a metadata wait fd */
2500 for (i
= 0; i
< nb_fd
; i
++) {
2501 health_code_update();
2503 revents
= LTTNG_POLL_GETEV(&events
, i
);
2504 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2506 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2507 if (revents
& LPOLLIN
) {
2510 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2511 &stream
, sizeof(stream
));
2512 if (pipe_len
< sizeof(stream
)) {
2514 PERROR("read metadata stream");
2517 * Remove the pipe from the poll set and continue the loop
2518 * since their might be data to consume.
2520 lttng_poll_del(&events
,
2521 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2522 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2526 /* A NULL stream means that the state has changed. */
2527 if (stream
== NULL
) {
2528 /* Check for deleted streams. */
2529 validate_endpoint_status_metadata_stream(&events
);
2533 DBG("Adding metadata stream %d to poll set",
2536 /* Add metadata stream to the global poll events list */
2537 lttng_poll_add(&events
, stream
->wait_fd
,
2538 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2539 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2540 DBG("Metadata thread pipe hung up");
2542 * Remove the pipe from the poll set and continue the loop
2543 * since their might be data to consume.
2545 lttng_poll_del(&events
,
2546 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2547 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2550 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2554 /* Handle other stream */
2560 uint64_t tmp_id
= (uint64_t) pollfd
;
2562 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2564 node
= lttng_ht_iter_get_node_u64(&iter
);
2567 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2570 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2571 /* Get the data out of the metadata file descriptor */
2572 DBG("Metadata available on fd %d", pollfd
);
2573 assert(stream
->wait_fd
== pollfd
);
2576 health_code_update();
2578 len
= ctx
->on_buffer_ready(stream
, ctx
);
2580 * We don't check the return value here since if we get
2581 * a negative len, it means an error occurred thus we
2582 * simply remove it from the poll set and free the
2587 /* It's ok to have an unavailable sub-buffer */
2588 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2589 /* Clean up stream from consumer and free it. */
2590 lttng_poll_del(&events
, stream
->wait_fd
);
2591 consumer_del_metadata_stream(stream
, metadata_ht
);
2593 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2594 DBG("Metadata fd %d is hup|err.", pollfd
);
2595 if (!stream
->hangup_flush_done
2596 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2597 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2598 DBG("Attempting to flush and consume the UST buffers");
2599 lttng_ustconsumer_on_stream_hangup(stream
);
2601 /* We just flushed the stream now read it. */
2603 health_code_update();
2605 len
= ctx
->on_buffer_ready(stream
, ctx
);
2607 * We don't check the return value here since if we get
2608 * a negative len, it means an error occurred thus we
2609 * simply remove it from the poll set and free the
2615 lttng_poll_del(&events
, stream
->wait_fd
);
2617 * This call update the channel states, closes file descriptors
2618 * and securely free the stream.
2620 consumer_del_metadata_stream(stream
, metadata_ht
);
2622 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2626 /* Release RCU lock for the stream looked up */
2634 DBG("Metadata poll thread exiting");
2636 lttng_poll_clean(&events
);
2641 ERR("Health error occurred in %s", __func__
);
2643 health_unregister(health_consumerd
);
2644 rcu_unregister_thread();
2649 * This thread polls the fds in the set to consume the data and write
2650 * it to tracefile if necessary.
2652 void *consumer_thread_data_poll(void *data
)
2654 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2655 struct pollfd
*pollfd
= NULL
;
2656 /* local view of the streams */
2657 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2658 /* local view of consumer_data.fds_count */
2660 /* 2 for the consumer_data_pipe and wake up pipe */
2661 const int nb_pipes_fd
= 2;
2662 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2663 int nb_inactive_fd
= 0;
2664 struct lttng_consumer_local_data
*ctx
= data
;
2667 rcu_register_thread();
2669 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2671 if (testpoint(consumerd_thread_data
)) {
2672 goto error_testpoint
;
2675 health_code_update();
2677 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2678 if (local_stream
== NULL
) {
2679 PERROR("local_stream malloc");
2684 health_code_update();
2690 * the fds set has been updated, we need to update our
2691 * local array as well
2693 pthread_mutex_lock(&consumer_data
.lock
);
2694 if (consumer_data
.need_update
) {
2699 local_stream
= NULL
;
2701 /* Allocate for all fds */
2702 pollfd
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) * sizeof(struct pollfd
));
2703 if (pollfd
== NULL
) {
2704 PERROR("pollfd malloc");
2705 pthread_mutex_unlock(&consumer_data
.lock
);
2709 local_stream
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) *
2710 sizeof(struct lttng_consumer_stream
*));
2711 if (local_stream
== NULL
) {
2712 PERROR("local_stream malloc");
2713 pthread_mutex_unlock(&consumer_data
.lock
);
2716 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2717 data_ht
, &nb_inactive_fd
);
2719 ERR("Error in allocating pollfd or local_outfds");
2720 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2721 pthread_mutex_unlock(&consumer_data
.lock
);
2725 consumer_data
.need_update
= 0;
2727 pthread_mutex_unlock(&consumer_data
.lock
);
2729 /* No FDs and consumer_quit, consumer_cleanup the thread */
2730 if (nb_fd
== 0 && nb_inactive_fd
== 0 &&
2731 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2732 err
= 0; /* All is OK */
2735 /* poll on the array of fds */
2737 DBG("polling on %d fd", nb_fd
+ nb_pipes_fd
);
2738 if (testpoint(consumerd_thread_data_poll
)) {
2741 health_poll_entry();
2742 num_rdy
= poll(pollfd
, nb_fd
+ nb_pipes_fd
, -1);
2744 DBG("poll num_rdy : %d", num_rdy
);
2745 if (num_rdy
== -1) {
2747 * Restart interrupted system call.
2749 if (errno
== EINTR
) {
2752 PERROR("Poll error");
2753 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2755 } else if (num_rdy
== 0) {
2756 DBG("Polling thread timed out");
2760 if (caa_unlikely(data_consumption_paused
)) {
2761 DBG("Data consumption paused, sleeping...");
2767 * If the consumer_data_pipe triggered poll go directly to the
2768 * beginning of the loop to update the array. We want to prioritize
2769 * array update over low-priority reads.
2771 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2772 ssize_t pipe_readlen
;
2774 DBG("consumer_data_pipe wake up");
2775 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2776 &new_stream
, sizeof(new_stream
));
2777 if (pipe_readlen
< sizeof(new_stream
)) {
2778 PERROR("Consumer data pipe");
2779 /* Continue so we can at least handle the current stream(s). */
2784 * If the stream is NULL, just ignore it. It's also possible that
2785 * the sessiond poll thread changed the consumer_quit state and is
2786 * waking us up to test it.
2788 if (new_stream
== NULL
) {
2789 validate_endpoint_status_data_stream();
2793 /* Continue to update the local streams and handle prio ones */
2797 /* Handle wakeup pipe. */
2798 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2800 ssize_t pipe_readlen
;
2802 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2804 if (pipe_readlen
< 0) {
2805 PERROR("Consumer data wakeup pipe");
2807 /* We've been awakened to handle stream(s). */
2808 ctx
->has_wakeup
= 0;
2811 /* Take care of high priority channels first. */
2812 for (i
= 0; i
< nb_fd
; i
++) {
2813 health_code_update();
2815 if (local_stream
[i
] == NULL
) {
2818 if (pollfd
[i
].revents
& POLLPRI
) {
2819 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2821 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2822 /* it's ok to have an unavailable sub-buffer */
2823 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2824 /* Clean the stream and free it. */
2825 consumer_del_stream(local_stream
[i
], data_ht
);
2826 local_stream
[i
] = NULL
;
2827 } else if (len
> 0) {
2828 local_stream
[i
]->data_read
= 1;
2834 * If we read high prio channel in this loop, try again
2835 * for more high prio data.
2841 /* Take care of low priority channels. */
2842 for (i
= 0; i
< nb_fd
; i
++) {
2843 health_code_update();
2845 if (local_stream
[i
] == NULL
) {
2848 if ((pollfd
[i
].revents
& POLLIN
) ||
2849 local_stream
[i
]->hangup_flush_done
||
2850 local_stream
[i
]->has_data
) {
2851 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2852 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2853 /* it's ok to have an unavailable sub-buffer */
2854 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2855 /* Clean the stream and free it. */
2856 consumer_del_stream(local_stream
[i
], data_ht
);
2857 local_stream
[i
] = NULL
;
2858 } else if (len
> 0) {
2859 local_stream
[i
]->data_read
= 1;
2864 /* Handle hangup and errors */
2865 for (i
= 0; i
< nb_fd
; i
++) {
2866 health_code_update();
2868 if (local_stream
[i
] == NULL
) {
2871 if (!local_stream
[i
]->hangup_flush_done
2872 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2873 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2874 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2875 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2877 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2878 /* Attempt read again, for the data we just flushed. */
2879 local_stream
[i
]->data_read
= 1;
2882 * If the poll flag is HUP/ERR/NVAL and we have
2883 * read no data in this pass, we can remove the
2884 * stream from its hash table.
2886 if ((pollfd
[i
].revents
& POLLHUP
)) {
2887 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2888 if (!local_stream
[i
]->data_read
) {
2889 consumer_del_stream(local_stream
[i
], data_ht
);
2890 local_stream
[i
] = NULL
;
2893 } else if (pollfd
[i
].revents
& POLLERR
) {
2894 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2895 if (!local_stream
[i
]->data_read
) {
2896 consumer_del_stream(local_stream
[i
], data_ht
);
2897 local_stream
[i
] = NULL
;
2900 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2901 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2902 if (!local_stream
[i
]->data_read
) {
2903 consumer_del_stream(local_stream
[i
], data_ht
);
2904 local_stream
[i
] = NULL
;
2908 if (local_stream
[i
] != NULL
) {
2909 local_stream
[i
]->data_read
= 0;
2916 DBG("polling thread exiting");
2921 * Close the write side of the pipe so epoll_wait() in
2922 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2923 * read side of the pipe. If we close them both, epoll_wait strangely does
2924 * not return and could create a endless wait period if the pipe is the
2925 * only tracked fd in the poll set. The thread will take care of closing
2928 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2933 ERR("Health error occurred in %s", __func__
);
2935 health_unregister(health_consumerd
);
2937 rcu_unregister_thread();
2942 * Close wake-up end of each stream belonging to the channel. This will
2943 * allow the poll() on the stream read-side to detect when the
2944 * write-side (application) finally closes them.
2947 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2949 struct lttng_ht
*ht
;
2950 struct lttng_consumer_stream
*stream
;
2951 struct lttng_ht_iter iter
;
2953 ht
= consumer_data
.stream_per_chan_id_ht
;
2956 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2957 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2958 ht
->match_fct
, &channel
->key
,
2959 &iter
.iter
, stream
, node_channel_id
.node
) {
2961 * Protect against teardown with mutex.
2963 pthread_mutex_lock(&stream
->lock
);
2964 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2967 switch (consumer_data
.type
) {
2968 case LTTNG_CONSUMER_KERNEL
:
2970 case LTTNG_CONSUMER32_UST
:
2971 case LTTNG_CONSUMER64_UST
:
2972 if (stream
->metadata_flag
) {
2973 /* Safe and protected by the stream lock. */
2974 lttng_ustconsumer_close_metadata(stream
->chan
);
2977 * Note: a mutex is taken internally within
2978 * liblttng-ust-ctl to protect timer wakeup_fd
2979 * use from concurrent close.
2981 lttng_ustconsumer_close_stream_wakeup(stream
);
2985 ERR("Unknown consumer_data type");
2989 pthread_mutex_unlock(&stream
->lock
);
2994 static void destroy_channel_ht(struct lttng_ht
*ht
)
2996 struct lttng_ht_iter iter
;
2997 struct lttng_consumer_channel
*channel
;
3005 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
3006 ret
= lttng_ht_del(ht
, &iter
);
3011 lttng_ht_destroy(ht
);
3015 * This thread polls the channel fds to detect when they are being
3016 * closed. It closes all related streams if the channel is detected as
3017 * closed. It is currently only used as a shim layer for UST because the
3018 * consumerd needs to keep the per-stream wakeup end of pipes open for
3021 void *consumer_thread_channel_poll(void *data
)
3023 int ret
, i
, pollfd
, err
= -1;
3024 uint32_t revents
, nb_fd
;
3025 struct lttng_consumer_channel
*chan
= NULL
;
3026 struct lttng_ht_iter iter
;
3027 struct lttng_ht_node_u64
*node
;
3028 struct lttng_poll_event events
;
3029 struct lttng_consumer_local_data
*ctx
= data
;
3030 struct lttng_ht
*channel_ht
;
3032 rcu_register_thread();
3034 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
3036 if (testpoint(consumerd_thread_channel
)) {
3037 goto error_testpoint
;
3040 health_code_update();
3042 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3044 /* ENOMEM at this point. Better to bail out. */
3048 DBG("Thread channel poll started");
3050 /* Size is set to 1 for the consumer_channel pipe */
3051 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
3053 ERR("Poll set creation failed");
3057 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
3063 DBG("Channel main loop started");
3067 health_code_update();
3068 DBG("Channel poll wait");
3069 health_poll_entry();
3070 ret
= lttng_poll_wait(&events
, -1);
3071 DBG("Channel poll return from wait with %d fd(s)",
3072 LTTNG_POLL_GETNB(&events
));
3074 DBG("Channel event caught in thread");
3076 if (errno
== EINTR
) {
3077 ERR("Poll EINTR caught");
3080 if (LTTNG_POLL_GETNB(&events
) == 0) {
3081 err
= 0; /* All is OK */
3088 /* From here, the event is a channel wait fd */
3089 for (i
= 0; i
< nb_fd
; i
++) {
3090 health_code_update();
3092 revents
= LTTNG_POLL_GETEV(&events
, i
);
3093 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
3095 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
3096 if (revents
& LPOLLIN
) {
3097 enum consumer_channel_action action
;
3100 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
3103 ERR("Error reading channel pipe");
3105 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3110 case CONSUMER_CHANNEL_ADD
:
3111 DBG("Adding channel %d to poll set",
3114 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
3117 lttng_ht_add_unique_u64(channel_ht
,
3118 &chan
->wait_fd_node
);
3120 /* Add channel to the global poll events list */
3121 lttng_poll_add(&events
, chan
->wait_fd
,
3122 LPOLLERR
| LPOLLHUP
);
3124 case CONSUMER_CHANNEL_DEL
:
3127 * This command should never be called if the channel
3128 * has streams monitored by either the data or metadata
3129 * thread. The consumer only notify this thread with a
3130 * channel del. command if it receives a destroy
3131 * channel command from the session daemon that send it
3132 * if a command prior to the GET_CHANNEL failed.
3136 chan
= consumer_find_channel(key
);
3139 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
3142 lttng_poll_del(&events
, chan
->wait_fd
);
3143 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
3144 ret
= lttng_ht_del(channel_ht
, &iter
);
3147 switch (consumer_data
.type
) {
3148 case LTTNG_CONSUMER_KERNEL
:
3150 case LTTNG_CONSUMER32_UST
:
3151 case LTTNG_CONSUMER64_UST
:
3152 health_code_update();
3153 /* Destroy streams that might have been left in the stream list. */
3154 clean_channel_stream_list(chan
);
3157 ERR("Unknown consumer_data type");
3162 * Release our own refcount. Force channel deletion even if
3163 * streams were not initialized.
3165 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
3166 consumer_del_channel(chan
);
3171 case CONSUMER_CHANNEL_QUIT
:
3173 * Remove the pipe from the poll set and continue the loop
3174 * since their might be data to consume.
3176 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3179 ERR("Unknown action");
3182 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3183 DBG("Channel thread pipe hung up");
3185 * Remove the pipe from the poll set and continue the loop
3186 * since their might be data to consume.
3188 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3191 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3195 /* Handle other stream */
3201 uint64_t tmp_id
= (uint64_t) pollfd
;
3203 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3205 node
= lttng_ht_iter_get_node_u64(&iter
);
3208 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3211 /* Check for error event */
3212 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3213 DBG("Channel fd %d is hup|err.", pollfd
);
3215 lttng_poll_del(&events
, chan
->wait_fd
);
3216 ret
= lttng_ht_del(channel_ht
, &iter
);
3220 * This will close the wait fd for each stream associated to
3221 * this channel AND monitored by the data/metadata thread thus
3222 * will be clean by the right thread.
3224 consumer_close_channel_streams(chan
);
3226 /* Release our own refcount */
3227 if (!uatomic_sub_return(&chan
->refcount
, 1)
3228 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3229 consumer_del_channel(chan
);
3232 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3237 /* Release RCU lock for the channel looked up */
3245 lttng_poll_clean(&events
);
3247 destroy_channel_ht(channel_ht
);
3250 DBG("Channel poll thread exiting");
3253 ERR("Health error occurred in %s", __func__
);
3255 health_unregister(health_consumerd
);
3256 rcu_unregister_thread();
3260 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3261 struct pollfd
*sockpoll
, int client_socket
)
3268 ret
= lttng_consumer_poll_socket(sockpoll
);
3272 DBG("Metadata connection on client_socket");
3274 /* Blocking call, waiting for transmission */
3275 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3276 if (ctx
->consumer_metadata_socket
< 0) {
3277 WARN("On accept metadata");
3288 * This thread listens on the consumerd socket and receives the file
3289 * descriptors from the session daemon.
3291 void *consumer_thread_sessiond_poll(void *data
)
3293 int sock
= -1, client_socket
, ret
, err
= -1;
3295 * structure to poll for incoming data on communication socket avoids
3296 * making blocking sockets.
3298 struct pollfd consumer_sockpoll
[2];
3299 struct lttng_consumer_local_data
*ctx
= data
;
3301 rcu_register_thread();
3303 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3305 if (testpoint(consumerd_thread_sessiond
)) {
3306 goto error_testpoint
;
3309 health_code_update();
3311 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3312 unlink(ctx
->consumer_command_sock_path
);
3313 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3314 if (client_socket
< 0) {
3315 ERR("Cannot create command socket");
3319 ret
= lttcomm_listen_unix_sock(client_socket
);
3324 DBG("Sending ready command to lttng-sessiond");
3325 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3326 /* return < 0 on error, but == 0 is not fatal */
3328 ERR("Error sending ready command to lttng-sessiond");
3332 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3333 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3334 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3335 consumer_sockpoll
[1].fd
= client_socket
;
3336 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3338 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3346 DBG("Connection on client_socket");
3348 /* Blocking call, waiting for transmission */
3349 sock
= lttcomm_accept_unix_sock(client_socket
);
3356 * Setup metadata socket which is the second socket connection on the
3357 * command unix socket.
3359 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3368 /* This socket is not useful anymore. */
3369 ret
= close(client_socket
);
3371 PERROR("close client_socket");
3375 /* update the polling structure to poll on the established socket */
3376 consumer_sockpoll
[1].fd
= sock
;
3377 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3380 health_code_update();
3382 health_poll_entry();
3383 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3392 DBG("Incoming command on sock");
3393 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3396 * This could simply be a session daemon quitting. Don't output
3399 DBG("Communication interrupted on command socket");
3403 if (CMM_LOAD_SHARED(consumer_quit
)) {
3404 DBG("consumer_thread_receive_fds received quit from signal");
3405 err
= 0; /* All is OK */
3408 DBG("received command on sock");
3414 DBG("Consumer thread sessiond poll exiting");
3417 * Close metadata streams since the producer is the session daemon which
3420 * NOTE: for now, this only applies to the UST tracer.
3422 lttng_consumer_close_all_metadata();
3425 * when all fds have hung up, the polling thread
3428 CMM_STORE_SHARED(consumer_quit
, 1);
3431 * Notify the data poll thread to poll back again and test the
3432 * consumer_quit state that we just set so to quit gracefully.
3434 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3436 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3438 notify_health_quit_pipe(health_quit_pipe
);
3440 /* Cleaning up possibly open sockets. */
3444 PERROR("close sock sessiond poll");
3447 if (client_socket
>= 0) {
3448 ret
= close(client_socket
);
3450 PERROR("close client_socket sessiond poll");
3457 ERR("Health error occurred in %s", __func__
);
3459 health_unregister(health_consumerd
);
3461 rcu_unregister_thread();
3465 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3466 struct lttng_consumer_local_data
*ctx
)
3470 pthread_mutex_lock(&stream
->chan
->lock
);
3471 pthread_mutex_lock(&stream
->lock
);
3472 if (stream
->metadata_flag
) {
3473 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3476 switch (consumer_data
.type
) {
3477 case LTTNG_CONSUMER_KERNEL
:
3478 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3480 case LTTNG_CONSUMER32_UST
:
3481 case LTTNG_CONSUMER64_UST
:
3482 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3485 ERR("Unknown consumer_data type");
3491 if (stream
->metadata_flag
) {
3492 pthread_cond_broadcast(&stream
->metadata_rdv
);
3493 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3495 pthread_mutex_unlock(&stream
->lock
);
3496 pthread_mutex_unlock(&stream
->chan
->lock
);
3501 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3503 switch (consumer_data
.type
) {
3504 case LTTNG_CONSUMER_KERNEL
:
3505 return lttng_kconsumer_on_recv_stream(stream
);
3506 case LTTNG_CONSUMER32_UST
:
3507 case LTTNG_CONSUMER64_UST
:
3508 return lttng_ustconsumer_on_recv_stream(stream
);
3510 ERR("Unknown consumer_data type");
3517 * Allocate and set consumer data hash tables.
3519 int lttng_consumer_init(void)
3521 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3522 if (!consumer_data
.channel_ht
) {
3526 consumer_data
.channels_by_session_id_ht
=
3527 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3528 if (!consumer_data
.channels_by_session_id_ht
) {
3532 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3533 if (!consumer_data
.relayd_ht
) {
3537 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3538 if (!consumer_data
.stream_list_ht
) {
3542 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3543 if (!consumer_data
.stream_per_chan_id_ht
) {
3547 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3552 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3557 consumer_data
.chunk_registry
= lttng_trace_chunk_registry_create();
3558 if (!consumer_data
.chunk_registry
) {
3569 * Process the ADD_RELAYD command receive by a consumer.
3571 * This will create a relayd socket pair and add it to the relayd hash table.
3572 * The caller MUST acquire a RCU read side lock before calling it.
3574 void consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3575 struct lttng_consumer_local_data
*ctx
, int sock
,
3576 struct pollfd
*consumer_sockpoll
,
3577 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3578 uint64_t relayd_session_id
)
3580 int fd
= -1, ret
= -1, relayd_created
= 0;
3581 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3582 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3585 assert(relayd_sock
);
3587 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3589 /* Get relayd reference if exists. */
3590 relayd
= consumer_find_relayd(net_seq_idx
);
3591 if (relayd
== NULL
) {
3592 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3593 /* Not found. Allocate one. */
3594 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3595 if (relayd
== NULL
) {
3596 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3599 relayd
->sessiond_session_id
= sessiond_id
;
3604 * This code path MUST continue to the consumer send status message to
3605 * we can notify the session daemon and continue our work without
3606 * killing everything.
3610 * relayd key should never be found for control socket.
3612 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3615 /* First send a status message before receiving the fds. */
3616 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3618 /* Somehow, the session daemon is not responding anymore. */
3619 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3620 goto error_nosignal
;
3623 /* Poll on consumer socket. */
3624 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3626 /* Needing to exit in the middle of a command: error. */
3627 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3628 goto error_nosignal
;
3631 /* Get relayd socket from session daemon */
3632 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3633 if (ret
!= sizeof(fd
)) {
3634 fd
= -1; /* Just in case it gets set with an invalid value. */
3637 * Failing to receive FDs might indicate a major problem such as
3638 * reaching a fd limit during the receive where the kernel returns a
3639 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3640 * don't take any chances and stop everything.
3642 * XXX: Feature request #558 will fix that and avoid this possible
3643 * issue when reaching the fd limit.
3645 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3646 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3650 /* Copy socket information and received FD */
3651 switch (sock_type
) {
3652 case LTTNG_STREAM_CONTROL
:
3653 /* Copy received lttcomm socket */
3654 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3655 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3656 /* Handle create_sock error. */
3658 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3662 * Close the socket created internally by
3663 * lttcomm_create_sock, so we can replace it by the one
3664 * received from sessiond.
3666 if (close(relayd
->control_sock
.sock
.fd
)) {
3670 /* Assign new file descriptor */
3671 relayd
->control_sock
.sock
.fd
= fd
;
3672 /* Assign version values. */
3673 relayd
->control_sock
.major
= relayd_sock
->major
;
3674 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3676 relayd
->relayd_session_id
= relayd_session_id
;
3679 case LTTNG_STREAM_DATA
:
3680 /* Copy received lttcomm socket */
3681 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3682 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3683 /* Handle create_sock error. */
3685 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3689 * Close the socket created internally by
3690 * lttcomm_create_sock, so we can replace it by the one
3691 * received from sessiond.
3693 if (close(relayd
->data_sock
.sock
.fd
)) {
3697 /* Assign new file descriptor */
3698 relayd
->data_sock
.sock
.fd
= fd
;
3699 /* Assign version values. */
3700 relayd
->data_sock
.major
= relayd_sock
->major
;
3701 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3704 ERR("Unknown relayd socket type (%d)", sock_type
);
3705 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3709 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3710 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3711 relayd
->net_seq_idx
, fd
);
3713 * We gave the ownership of the fd to the relayd structure. Set the
3714 * fd to -1 so we don't call close() on it in the error path below.
3718 /* We successfully added the socket. Send status back. */
3719 ret
= consumer_send_status_msg(sock
, ret_code
);
3721 /* Somehow, the session daemon is not responding anymore. */
3722 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3723 goto error_nosignal
;
3727 * Add relayd socket pair to consumer data hashtable. If object already
3728 * exists or on error, the function gracefully returns.
3737 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3738 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3742 /* Close received socket if valid. */
3745 PERROR("close received socket");
3749 if (relayd_created
) {
3755 * Search for a relayd associated to the session id and return the reference.
3757 * A rcu read side lock MUST be acquire before calling this function and locked
3758 * until the relayd object is no longer necessary.
3760 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3762 struct lttng_ht_iter iter
;
3763 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3765 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3766 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3769 * Check by sessiond id which is unique here where the relayd session
3770 * id might not be when having multiple relayd.
3772 if (relayd
->sessiond_session_id
== id
) {
3773 /* Found the relayd. There can be only one per id. */
3785 * Check if for a given session id there is still data needed to be extract
3788 * Return 1 if data is pending or else 0 meaning ready to be read.
3790 int consumer_data_pending(uint64_t id
)
3793 struct lttng_ht_iter iter
;
3794 struct lttng_ht
*ht
;
3795 struct lttng_consumer_stream
*stream
;
3796 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3797 int (*data_pending
)(struct lttng_consumer_stream
*);
3799 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3802 pthread_mutex_lock(&consumer_data
.lock
);
3804 switch (consumer_data
.type
) {
3805 case LTTNG_CONSUMER_KERNEL
:
3806 data_pending
= lttng_kconsumer_data_pending
;
3808 case LTTNG_CONSUMER32_UST
:
3809 case LTTNG_CONSUMER64_UST
:
3810 data_pending
= lttng_ustconsumer_data_pending
;
3813 ERR("Unknown consumer data type");
3817 /* Ease our life a bit */
3818 ht
= consumer_data
.stream_list_ht
;
3820 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3821 ht
->hash_fct(&id
, lttng_ht_seed
),
3823 &iter
.iter
, stream
, node_session_id
.node
) {
3824 pthread_mutex_lock(&stream
->lock
);
3827 * A removed node from the hash table indicates that the stream has
3828 * been deleted thus having a guarantee that the buffers are closed
3829 * on the consumer side. However, data can still be transmitted
3830 * over the network so don't skip the relayd check.
3832 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3834 /* Check the stream if there is data in the buffers. */
3835 ret
= data_pending(stream
);
3837 pthread_mutex_unlock(&stream
->lock
);
3842 pthread_mutex_unlock(&stream
->lock
);
3845 relayd
= find_relayd_by_session_id(id
);
3847 unsigned int is_data_inflight
= 0;
3849 /* Send init command for data pending. */
3850 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3851 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3852 relayd
->relayd_session_id
);
3854 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3855 /* Communication error thus the relayd so no data pending. */
3856 goto data_not_pending
;
3859 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3860 ht
->hash_fct(&id
, lttng_ht_seed
),
3862 &iter
.iter
, stream
, node_session_id
.node
) {
3863 if (stream
->metadata_flag
) {
3864 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3865 stream
->relayd_stream_id
);
3867 ret
= relayd_data_pending(&relayd
->control_sock
,
3868 stream
->relayd_stream_id
,
3869 stream
->next_net_seq_num
- 1);
3873 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3875 } else if (ret
< 0) {
3876 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3877 lttng_consumer_cleanup_relayd(relayd
);
3878 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3879 goto data_not_pending
;
3883 /* Send end command for data pending. */
3884 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3885 relayd
->relayd_session_id
, &is_data_inflight
);
3886 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3888 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3889 lttng_consumer_cleanup_relayd(relayd
);
3890 goto data_not_pending
;
3892 if (is_data_inflight
) {
3898 * Finding _no_ node in the hash table and no inflight data means that the
3899 * stream(s) have been removed thus data is guaranteed to be available for
3900 * analysis from the trace files.
3904 /* Data is available to be read by a viewer. */
3905 pthread_mutex_unlock(&consumer_data
.lock
);
3910 /* Data is still being extracted from buffers. */
3911 pthread_mutex_unlock(&consumer_data
.lock
);
3917 * Send a ret code status message to the sessiond daemon.
3919 * Return the sendmsg() return value.
3921 int consumer_send_status_msg(int sock
, int ret_code
)
3923 struct lttcomm_consumer_status_msg msg
;
3925 memset(&msg
, 0, sizeof(msg
));
3926 msg
.ret_code
= ret_code
;
3928 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3932 * Send a channel status message to the sessiond daemon.
3934 * Return the sendmsg() return value.
3936 int consumer_send_status_channel(int sock
,
3937 struct lttng_consumer_channel
*channel
)
3939 struct lttcomm_consumer_status_channel msg
;
3943 memset(&msg
, 0, sizeof(msg
));
3945 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3947 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3948 msg
.key
= channel
->key
;
3949 msg
.stream_count
= channel
->streams
.count
;
3952 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3955 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3956 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3957 uint64_t max_sb_size
)
3959 unsigned long start_pos
;
3961 if (!nb_packets_per_stream
) {
3962 return consumed_pos
; /* Grab everything */
3964 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3965 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3966 if ((long) (start_pos
- consumed_pos
) < 0) {
3967 return consumed_pos
; /* Grab everything */
3973 int consumer_flush_buffer(struct lttng_consumer_stream
*stream
, int producer_active
)
3977 switch (consumer_data
.type
) {
3978 case LTTNG_CONSUMER_KERNEL
:
3979 ret
= kernctl_buffer_flush(stream
->wait_fd
);
3981 ERR("Failed to flush kernel stream");
3985 case LTTNG_CONSUMER32_UST
:
3986 case LTTNG_CONSUMER64_UST
:
3987 lttng_ustctl_flush_buffer(stream
, producer_active
);
3990 ERR("Unknown consumer_data type");
3999 * Sample the rotate position for all the streams of a channel. If a stream
4000 * is already at the rotate position (produced == consumed), we flag it as
4001 * ready for rotation. The rotation of ready streams occurs after we have
4002 * replied to the session daemon that we have finished sampling the positions.
4003 * Must be called with RCU read-side lock held to ensure existence of channel.
4005 * Returns 0 on success, < 0 on error
4007 int lttng_consumer_rotate_channel(struct lttng_consumer_channel
*channel
,
4008 uint64_t key
, uint64_t relayd_id
, uint32_t metadata
,
4009 struct lttng_consumer_local_data
*ctx
)
4012 struct lttng_consumer_stream
*stream
;
4013 struct lttng_ht_iter iter
;
4014 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
4016 DBG("Consumer sample rotate position for channel %" PRIu64
, key
);
4020 pthread_mutex_lock(&channel
->lock
);
4022 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4023 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4024 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4025 stream
, node_channel_id
.node
) {
4026 unsigned long consumed_pos
;
4028 health_code_update();
4031 * Lock stream because we are about to change its state.
4033 pthread_mutex_lock(&stream
->lock
);
4035 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4037 ERR("Failed to sample snapshot position during channel rotation");
4038 goto end_unlock_stream
;
4041 ret
= lttng_consumer_get_produced_snapshot(stream
,
4042 &stream
->rotate_position
);
4044 ERR("Failed to sample produced position during channel rotation");
4045 goto end_unlock_stream
;
4048 lttng_consumer_get_consumed_snapshot(stream
,
4050 if (consumed_pos
== stream
->rotate_position
) {
4051 stream
->rotate_ready
= true;
4054 ret
= consumer_flush_buffer(stream
, 1);
4056 ERR("Failed to flush stream %" PRIu64
" during channel rotation",
4058 goto end_unlock_stream
;
4061 pthread_mutex_unlock(&stream
->lock
);
4063 pthread_mutex_unlock(&channel
->lock
);
4069 pthread_mutex_unlock(&stream
->lock
);
4070 pthread_mutex_unlock(&channel
->lock
);
4077 * Check if a stream is ready to be rotated after extracting it.
4079 * Return 1 if it is ready for rotation, 0 if it is not, a negative value on
4080 * error. Stream lock must be held.
4082 int lttng_consumer_stream_is_rotate_ready(struct lttng_consumer_stream
*stream
)
4085 unsigned long consumed_pos
;
4087 if (!stream
->rotate_position
&& !stream
->rotate_ready
) {
4092 if (stream
->rotate_ready
) {
4098 * If we don't have the rotate_ready flag, check the consumed position
4099 * to determine if we need to rotate.
4101 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4103 ERR("Taking snapshot positions");
4107 ret
= lttng_consumer_get_consumed_snapshot(stream
, &consumed_pos
);
4109 ERR("Consumed snapshot position");
4113 /* Rotate position not reached yet (with check for overflow). */
4114 if ((long) (consumed_pos
- stream
->rotate_position
) < 0) {
4125 * Reset the state for a stream after a rotation occurred.
4127 void lttng_consumer_reset_stream_rotate_state(struct lttng_consumer_stream
*stream
)
4129 stream
->rotate_position
= 0;
4130 stream
->rotate_ready
= false;
4134 * Perform the rotation a local stream file.
4137 int rotate_local_stream(struct lttng_consumer_local_data
*ctx
,
4138 struct lttng_consumer_stream
*stream
)
4142 DBG("Rotate local stream: stream key %" PRIu64
", channel key %" PRIu64
,
4145 stream
->tracefile_size_current
= 0;
4146 stream
->tracefile_count_current
= 0;
4148 if (stream
->out_fd
>= 0) {
4149 ret
= close(stream
->out_fd
);
4151 PERROR("Failed to close stream out_fd of channel \"%s\"",
4152 stream
->chan
->name
);
4154 stream
->out_fd
= -1;
4157 if (stream
->index_file
) {
4158 lttng_index_file_put(stream
->index_file
);
4159 stream
->index_file
= NULL
;
4162 if (!stream
->trace_chunk
) {
4166 ret
= consumer_stream_create_output_files(stream
, true);
4172 * Perform the rotation a stream file on the relay.
4174 int rotate_relay_stream(struct lttng_consumer_local_data
*ctx
,
4175 struct lttng_consumer_stream
*stream
)
4178 struct consumer_relayd_sock_pair
*relayd
;
4180 enum lttng_trace_chunk_status chunk_status
;
4182 DBG("Rotate relay stream");
4183 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
4185 ERR("Failed to find relayd");
4190 chunk_status
= lttng_trace_chunk_get_id(stream
->chan
->trace_chunk
,
4192 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4193 ERR("Failed to retrieve the id of the current trace chunk of channel \"%s\"",
4194 stream
->chan
->name
);
4199 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4200 ret
= relayd_rotate_stream(&relayd
->control_sock
,
4201 stream
->relayd_stream_id
,
4203 stream
->last_sequence_number
);
4204 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4206 ERR("Relayd rotate stream failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
4207 lttng_consumer_cleanup_relayd(relayd
);
4210 ERR("Rotate relay stream");
4218 * Performs the stream rotation for the rotate session feature if needed.
4219 * It must be called with the channel and stream locks held.
4221 * Return 0 on success, a negative number of error.
4223 int lttng_consumer_rotate_stream(struct lttng_consumer_local_data
*ctx
,
4224 struct lttng_consumer_stream
*stream
)
4228 DBG("Consumer rotate stream %" PRIu64
, stream
->key
);
4231 * Update the stream's 'current' chunk to the session's (channel)
4232 * now-current chunk.
4234 lttng_trace_chunk_put(stream
->trace_chunk
);
4235 if (stream
->chan
->trace_chunk
== stream
->trace_chunk
) {
4237 * A channel can be rotated and not have a "next" chunk
4238 * to transition to. In that case, the channel's "current chunk"
4239 * has not been closed yet, but it has not been updated to
4240 * a "next" trace chunk either. Hence, the stream, like its
4241 * parent channel, becomes part of no chunk and can't output
4242 * anything until a new trace chunk is created.
4244 stream
->trace_chunk
= NULL
;
4245 } else if (stream
->chan
->trace_chunk
&&
4246 !lttng_trace_chunk_get(stream
->chan
->trace_chunk
)) {
4247 ERR("Failed to acquire a reference to channel's trace chunk during stream rotation");
4252 * Update the stream's trace chunk to its parent channel's
4253 * current trace chunk.
4255 stream
->trace_chunk
= stream
->chan
->trace_chunk
;
4258 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
4259 ret
= rotate_relay_stream(ctx
, stream
);
4261 ret
= rotate_local_stream(ctx
, stream
);
4264 ERR("Failed to rotate stream, ret = %i", ret
);
4268 if (stream
->metadata_flag
&& stream
->trace_chunk
) {
4270 * If the stream has transitioned to a new trace
4271 * chunk, the metadata should be re-dumped to the
4274 * However, it is possible for a stream to transition to
4275 * a "no-chunk" state. This can happen if a rotation
4276 * occurs on an inactive session. In such cases, the metadata
4277 * regeneration will happen when the next trace chunk is
4280 ret
= consumer_metadata_stream_dump(stream
);
4285 lttng_consumer_reset_stream_rotate_state(stream
);
4294 * Rotate all the ready streams now.
4296 * This is especially important for low throughput streams that have already
4297 * been consumed, we cannot wait for their next packet to perform the
4299 * Need to be called with RCU read-side lock held to ensure existence of
4302 * Returns 0 on success, < 0 on error
4304 int lttng_consumer_rotate_ready_streams(struct lttng_consumer_channel
*channel
,
4305 uint64_t key
, struct lttng_consumer_local_data
*ctx
)
4308 struct lttng_consumer_stream
*stream
;
4309 struct lttng_ht_iter iter
;
4310 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
4314 DBG("Consumer rotate ready streams in channel %" PRIu64
, key
);
4316 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4317 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4318 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4319 stream
, node_channel_id
.node
) {
4320 health_code_update();
4322 pthread_mutex_lock(&stream
->chan
->lock
);
4323 pthread_mutex_lock(&stream
->lock
);
4325 if (!stream
->rotate_ready
) {
4326 pthread_mutex_unlock(&stream
->lock
);
4327 pthread_mutex_unlock(&stream
->chan
->lock
);
4330 DBG("Consumer rotate ready stream %" PRIu64
, stream
->key
);
4332 ret
= lttng_consumer_rotate_stream(ctx
, stream
);
4333 pthread_mutex_unlock(&stream
->lock
);
4334 pthread_mutex_unlock(&stream
->chan
->lock
);
4347 enum lttcomm_return_code
lttng_consumer_init_command(
4348 struct lttng_consumer_local_data
*ctx
,
4349 const lttng_uuid sessiond_uuid
)
4351 enum lttcomm_return_code ret
;
4352 char uuid_str
[UUID_STR_LEN
];
4354 if (ctx
->sessiond_uuid
.is_set
) {
4355 ret
= LTTCOMM_CONSUMERD_ALREADY_SET
;
4359 ctx
->sessiond_uuid
.is_set
= true;
4360 memcpy(ctx
->sessiond_uuid
.value
, sessiond_uuid
, sizeof(lttng_uuid
));
4361 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4362 lttng_uuid_to_str(sessiond_uuid
, uuid_str
);
4363 DBG("Received session daemon UUID: %s", uuid_str
);
4368 enum lttcomm_return_code
lttng_consumer_create_trace_chunk(
4369 const uint64_t *relayd_id
, uint64_t session_id
,
4371 time_t chunk_creation_timestamp
,
4372 const char *chunk_override_name
,
4373 const struct lttng_credentials
*credentials
,
4374 struct lttng_directory_handle
*chunk_directory_handle
)
4377 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4378 struct lttng_trace_chunk
*created_chunk
, *published_chunk
;
4379 enum lttng_trace_chunk_status chunk_status
;
4380 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4381 char creation_timestamp_buffer
[ISO8601_STR_LEN
];
4382 const char *relayd_id_str
= "(none)";
4383 const char *creation_timestamp_str
;
4384 struct lttng_ht_iter iter
;
4385 struct lttng_consumer_channel
*channel
;
4388 /* Only used for logging purposes. */
4389 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4390 "%" PRIu64
, *relayd_id
);
4391 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4392 relayd_id_str
= relayd_id_buffer
;
4394 relayd_id_str
= "(formatting error)";
4398 /* Local protocol error. */
4399 assert(chunk_creation_timestamp
);
4400 ret
= time_to_iso8601_str(chunk_creation_timestamp
,
4401 creation_timestamp_buffer
,
4402 sizeof(creation_timestamp_buffer
));
4403 creation_timestamp_str
= !ret
? creation_timestamp_buffer
:
4404 "(formatting error)";
4406 DBG("Consumer create trace chunk command: relay_id = %s"
4407 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4408 ", chunk_override_name = %s"
4409 ", chunk_creation_timestamp = %s",
4410 relayd_id_str
, session_id
, chunk_id
,
4411 chunk_override_name
? : "(none)",
4412 creation_timestamp_str
);
4415 * The trace chunk registry, as used by the consumer daemon, implicitly
4416 * owns the trace chunks. This is only needed in the consumer since
4417 * the consumer has no notion of a session beyond session IDs being
4418 * used to identify other objects.
4420 * The lttng_trace_chunk_registry_publish() call below provides a
4421 * reference which is not released; it implicitly becomes the session
4422 * daemon's reference to the chunk in the consumer daemon.
4424 * The lifetime of trace chunks in the consumer daemon is managed by
4425 * the session daemon through the LTTNG_CONSUMER_CREATE_TRACE_CHUNK
4426 * and LTTNG_CONSUMER_DESTROY_TRACE_CHUNK commands.
4428 created_chunk
= lttng_trace_chunk_create(chunk_id
,
4429 chunk_creation_timestamp
);
4430 if (!created_chunk
) {
4431 ERR("Failed to create trace chunk");
4432 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4436 if (chunk_override_name
) {
4437 chunk_status
= lttng_trace_chunk_override_name(created_chunk
,
4438 chunk_override_name
);
4439 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4440 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4445 if (chunk_directory_handle
) {
4446 chunk_status
= lttng_trace_chunk_set_credentials(created_chunk
,
4448 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4449 ERR("Failed to set trace chunk credentials");
4450 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4454 * The consumer daemon has no ownership of the chunk output
4457 chunk_status
= lttng_trace_chunk_set_as_user(created_chunk
,
4458 chunk_directory_handle
);
4459 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4460 ERR("Failed to set trace chunk's directory handle");
4461 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4466 published_chunk
= lttng_trace_chunk_registry_publish_chunk(
4467 consumer_data
.chunk_registry
, session_id
,
4469 lttng_trace_chunk_put(created_chunk
);
4470 created_chunk
= NULL
;
4471 if (!published_chunk
) {
4472 ERR("Failed to publish trace chunk");
4473 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4478 cds_lfht_for_each_entry_duplicate(consumer_data
.channels_by_session_id_ht
->ht
,
4479 consumer_data
.channels_by_session_id_ht
->hash_fct(
4480 &session_id
, lttng_ht_seed
),
4481 consumer_data
.channels_by_session_id_ht
->match_fct
,
4482 &session_id
, &iter
.iter
, channel
,
4483 channels_by_session_id_ht_node
.node
) {
4484 ret
= lttng_consumer_channel_set_trace_chunk(channel
,
4488 * Roll-back the creation of this chunk.
4490 * This is important since the session daemon will
4491 * assume that the creation of this chunk failed and
4492 * will never ask for it to be closed, resulting
4493 * in a leak and an inconsistent state for some
4496 enum lttcomm_return_code close_ret
;
4498 DBG("Failed to set new trace chunk on existing channels, rolling back");
4499 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4500 session_id
, chunk_id
,
4501 chunk_creation_timestamp
, NULL
);
4502 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4503 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4504 session_id
, chunk_id
);
4507 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4513 struct consumer_relayd_sock_pair
*relayd
;
4515 relayd
= consumer_find_relayd(*relayd_id
);
4517 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4518 ret
= relayd_create_trace_chunk(
4519 &relayd
->control_sock
, published_chunk
);
4520 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4522 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
, *relayd_id
);
4525 if (!relayd
|| ret
) {
4526 enum lttcomm_return_code close_ret
;
4528 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4531 chunk_creation_timestamp
,
4533 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4534 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4539 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4545 /* Release the reference returned by the "publish" operation. */
4546 lttng_trace_chunk_put(published_chunk
);
4551 enum lttcomm_return_code
lttng_consumer_close_trace_chunk(
4552 const uint64_t *relayd_id
, uint64_t session_id
,
4553 uint64_t chunk_id
, time_t chunk_close_timestamp
,
4554 const enum lttng_trace_chunk_command_type
*close_command
)
4556 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4557 struct lttng_trace_chunk
*chunk
;
4558 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4559 const char *relayd_id_str
= "(none)";
4560 const char *close_command_name
= "none";
4561 struct lttng_ht_iter iter
;
4562 struct lttng_consumer_channel
*channel
;
4563 enum lttng_trace_chunk_status chunk_status
;
4568 /* Only used for logging purposes. */
4569 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4570 "%" PRIu64
, *relayd_id
);
4571 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4572 relayd_id_str
= relayd_id_buffer
;
4574 relayd_id_str
= "(formatting error)";
4577 if (close_command
) {
4578 close_command_name
= lttng_trace_chunk_command_type_get_name(
4582 DBG("Consumer close trace chunk command: relayd_id = %s"
4583 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4584 ", close command = %s",
4585 relayd_id_str
, session_id
, chunk_id
,
4586 close_command_name
);
4588 chunk
= lttng_trace_chunk_registry_find_chunk(
4589 consumer_data
.chunk_registry
, session_id
, chunk_id
);
4591 ERR("Failed to find chunk: session_id = %" PRIu64
4592 ", chunk_id = %" PRIu64
,
4593 session_id
, chunk_id
);
4594 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4598 chunk_status
= lttng_trace_chunk_set_close_timestamp(chunk
,
4599 chunk_close_timestamp
);
4600 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4601 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4605 if (close_command
) {
4606 chunk_status
= lttng_trace_chunk_set_close_command(
4607 chunk
, *close_command
);
4608 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4609 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4615 * chunk is now invalid to access as we no longer hold a reference to
4616 * it; it is only kept around to compare it (by address) to the
4617 * current chunk found in the session's channels.
4620 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
,
4621 channel
, node
.node
) {
4625 * Only change the channel's chunk to NULL if it still
4626 * references the chunk being closed. The channel may
4627 * reference a newer channel in the case of a session
4628 * rotation. When a session rotation occurs, the "next"
4629 * chunk is created before the "current" chunk is closed.
4631 if (channel
->trace_chunk
!= chunk
) {
4634 ret
= lttng_consumer_channel_set_trace_chunk(channel
, NULL
);
4637 * Attempt to close the chunk on as many channels as
4640 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4646 struct consumer_relayd_sock_pair
*relayd
;
4648 relayd
= consumer_find_relayd(*relayd_id
);
4650 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4651 ret
= relayd_close_trace_chunk(
4652 &relayd
->control_sock
, chunk
);
4653 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4655 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
,
4659 if (!relayd
|| ret
) {
4660 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4668 * Release the reference returned by the "find" operation and
4669 * the session daemon's implicit reference to the chunk.
4671 lttng_trace_chunk_put(chunk
);
4672 lttng_trace_chunk_put(chunk
);
4677 enum lttcomm_return_code
lttng_consumer_trace_chunk_exists(
4678 const uint64_t *relayd_id
, uint64_t session_id
,
4681 enum lttcomm_return_code ret_code
;
4682 struct lttng_trace_chunk
*chunk
;
4683 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4684 const char *relayd_id_str
= "(none)";
4689 /* Only used for logging purposes. */
4690 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4691 "%" PRIu64
, *relayd_id
);
4692 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4693 relayd_id_str
= relayd_id_buffer
;
4695 relayd_id_str
= "(formatting error)";
4699 DBG("Consumer trace chunk exists command: relayd_id = %s"
4700 ", session_id = %" PRIu64
4701 ", chunk_id = %" PRIu64
, relayd_id_str
,
4702 session_id
, chunk_id
);
4703 chunk
= lttng_trace_chunk_registry_find_chunk(
4704 consumer_data
.chunk_registry
, session_id
,
4706 DBG("Trace chunk %s locally", chunk
? "exists" : "does not exist");
4707 ret_code
= chunk
? LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_LOCAL
:
4708 LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4710 lttng_trace_chunk_put(chunk
);