2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 #include <sys/socket.h>
28 #include <sys/types.h>
33 #include <bin/lttng-consumerd/health-consumerd.h>
34 #include <common/common.h>
35 #include <common/utils.h>
36 #include <common/time.h>
37 #include <common/compat/poll.h>
38 #include <common/compat/endian.h>
39 #include <common/index/index.h>
40 #include <common/kernel-ctl/kernel-ctl.h>
41 #include <common/sessiond-comm/relayd.h>
42 #include <common/sessiond-comm/sessiond-comm.h>
43 #include <common/kernel-consumer/kernel-consumer.h>
44 #include <common/relayd/relayd.h>
45 #include <common/ust-consumer/ust-consumer.h>
46 #include <common/consumer/consumer-timer.h>
47 #include <common/consumer/consumer.h>
48 #include <common/consumer/consumer-stream.h>
49 #include <common/consumer/consumer-testpoint.h>
50 #include <common/align.h>
51 #include <common/consumer/consumer-metadata-cache.h>
52 #include <common/trace-chunk.h>
53 #include <common/trace-chunk-registry.h>
54 #include <common/string-utils/format.h>
55 #include <common/dynamic-array.h>
57 struct lttng_consumer_global_data consumer_data
= {
60 .type
= LTTNG_CONSUMER_UNKNOWN
,
63 enum consumer_channel_action
{
66 CONSUMER_CHANNEL_QUIT
,
69 struct consumer_channel_msg
{
70 enum consumer_channel_action action
;
71 struct lttng_consumer_channel
*chan
; /* add */
72 uint64_t key
; /* del */
75 /* Flag used to temporarily pause data consumption from testpoints. */
76 int data_consumption_paused
;
79 * Flag to inform the polling thread to quit when all fd hung up. Updated by
80 * the consumer_thread_receive_fds when it notices that all fds has hung up.
81 * Also updated by the signal handler (consumer_should_exit()). Read by the
87 * Global hash table containing respectively metadata and data streams. The
88 * stream element in this ht should only be updated by the metadata poll thread
89 * for the metadata and the data poll thread for the data.
91 static struct lttng_ht
*metadata_ht
;
92 static struct lttng_ht
*data_ht
;
95 * Notify a thread lttng pipe to poll back again. This usually means that some
96 * global state has changed so we just send back the thread in a poll wait
99 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
101 struct lttng_consumer_stream
*null_stream
= NULL
;
105 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
108 static void notify_health_quit_pipe(int *pipe
)
112 ret
= lttng_write(pipe
[1], "4", 1);
114 PERROR("write consumer health quit");
118 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
119 struct lttng_consumer_channel
*chan
,
121 enum consumer_channel_action action
)
123 struct consumer_channel_msg msg
;
126 memset(&msg
, 0, sizeof(msg
));
131 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
132 if (ret
< sizeof(msg
)) {
133 PERROR("notify_channel_pipe write error");
137 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
140 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
143 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
144 struct lttng_consumer_channel
**chan
,
146 enum consumer_channel_action
*action
)
148 struct consumer_channel_msg msg
;
151 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
152 if (ret
< sizeof(msg
)) {
156 *action
= msg
.action
;
164 * Cleanup the stream list of a channel. Those streams are not yet globally
167 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
169 struct lttng_consumer_stream
*stream
, *stmp
;
173 /* Delete streams that might have been left in the stream list. */
174 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
176 cds_list_del(&stream
->send_node
);
178 * Once a stream is added to this list, the buffers were created so we
179 * have a guarantee that this call will succeed. Setting the monitor
180 * mode to 0 so we don't lock nor try to delete the stream from the
184 consumer_stream_destroy(stream
, NULL
);
189 * Find a stream. The consumer_data.lock must be locked during this
192 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
195 struct lttng_ht_iter iter
;
196 struct lttng_ht_node_u64
*node
;
197 struct lttng_consumer_stream
*stream
= NULL
;
201 /* -1ULL keys are lookup failures */
202 if (key
== (uint64_t) -1ULL) {
208 lttng_ht_lookup(ht
, &key
, &iter
);
209 node
= lttng_ht_iter_get_node_u64(&iter
);
211 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
219 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
221 struct lttng_consumer_stream
*stream
;
224 stream
= find_stream(key
, ht
);
226 stream
->key
= (uint64_t) -1ULL;
228 * We don't want the lookup to match, but we still need
229 * to iterate on this stream when iterating over the hash table. Just
230 * change the node key.
232 stream
->node
.key
= (uint64_t) -1ULL;
238 * Return a channel object for the given key.
240 * RCU read side lock MUST be acquired before calling this function and
241 * protects the channel ptr.
243 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
245 struct lttng_ht_iter iter
;
246 struct lttng_ht_node_u64
*node
;
247 struct lttng_consumer_channel
*channel
= NULL
;
249 /* -1ULL keys are lookup failures */
250 if (key
== (uint64_t) -1ULL) {
254 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
255 node
= lttng_ht_iter_get_node_u64(&iter
);
257 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
264 * There is a possibility that the consumer does not have enough time between
265 * the close of the channel on the session daemon and the cleanup in here thus
266 * once we have a channel add with an existing key, we know for sure that this
267 * channel will eventually get cleaned up by all streams being closed.
269 * This function just nullifies the already existing channel key.
271 static void steal_channel_key(uint64_t key
)
273 struct lttng_consumer_channel
*channel
;
276 channel
= consumer_find_channel(key
);
278 channel
->key
= (uint64_t) -1ULL;
280 * We don't want the lookup to match, but we still need to iterate on
281 * this channel when iterating over the hash table. Just change the
284 channel
->node
.key
= (uint64_t) -1ULL;
289 static void free_channel_rcu(struct rcu_head
*head
)
291 struct lttng_ht_node_u64
*node
=
292 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
293 struct lttng_consumer_channel
*channel
=
294 caa_container_of(node
, struct lttng_consumer_channel
, node
);
296 switch (consumer_data
.type
) {
297 case LTTNG_CONSUMER_KERNEL
:
299 case LTTNG_CONSUMER32_UST
:
300 case LTTNG_CONSUMER64_UST
:
301 lttng_ustconsumer_free_channel(channel
);
304 ERR("Unknown consumer_data type");
311 * RCU protected relayd socket pair free.
313 static void free_relayd_rcu(struct rcu_head
*head
)
315 struct lttng_ht_node_u64
*node
=
316 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
317 struct consumer_relayd_sock_pair
*relayd
=
318 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
321 * Close all sockets. This is done in the call RCU since we don't want the
322 * socket fds to be reassigned thus potentially creating bad state of the
325 * We do not have to lock the control socket mutex here since at this stage
326 * there is no one referencing to this relayd object.
328 (void) relayd_close(&relayd
->control_sock
);
329 (void) relayd_close(&relayd
->data_sock
);
331 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
336 * Destroy and free relayd socket pair object.
338 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
341 struct lttng_ht_iter iter
;
343 if (relayd
== NULL
) {
347 DBG("Consumer destroy and close relayd socket pair");
349 iter
.iter
.node
= &relayd
->node
.node
;
350 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
352 /* We assume the relayd is being or is destroyed */
356 /* RCU free() call */
357 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
361 * Remove a channel from the global list protected by a mutex. This function is
362 * also responsible for freeing its data structures.
364 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
366 struct lttng_ht_iter iter
;
368 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
370 pthread_mutex_lock(&consumer_data
.lock
);
371 pthread_mutex_lock(&channel
->lock
);
373 /* Destroy streams that might have been left in the stream list. */
374 clean_channel_stream_list(channel
);
376 if (channel
->live_timer_enabled
== 1) {
377 consumer_timer_live_stop(channel
);
379 if (channel
->monitor_timer_enabled
== 1) {
380 consumer_timer_monitor_stop(channel
);
383 switch (consumer_data
.type
) {
384 case LTTNG_CONSUMER_KERNEL
:
386 case LTTNG_CONSUMER32_UST
:
387 case LTTNG_CONSUMER64_UST
:
388 lttng_ustconsumer_del_channel(channel
);
391 ERR("Unknown consumer_data type");
396 lttng_trace_chunk_put(channel
->trace_chunk
);
397 channel
->trace_chunk
= NULL
;
399 if (channel
->is_published
) {
403 iter
.iter
.node
= &channel
->node
.node
;
404 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
407 iter
.iter
.node
= &channel
->channels_by_session_id_ht_node
.node
;
408 ret
= lttng_ht_del(consumer_data
.channels_by_session_id_ht
,
414 channel
->is_deleted
= true;
415 call_rcu(&channel
->node
.head
, free_channel_rcu
);
417 pthread_mutex_unlock(&channel
->lock
);
418 pthread_mutex_unlock(&consumer_data
.lock
);
422 * Iterate over the relayd hash table and destroy each element. Finally,
423 * destroy the whole hash table.
425 static void cleanup_relayd_ht(void)
427 struct lttng_ht_iter iter
;
428 struct consumer_relayd_sock_pair
*relayd
;
432 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
434 consumer_destroy_relayd(relayd
);
439 lttng_ht_destroy(consumer_data
.relayd_ht
);
443 * Update the end point status of all streams having the given network sequence
444 * index (relayd index).
446 * It's atomically set without having the stream mutex locked which is fine
447 * because we handle the write/read race with a pipe wakeup for each thread.
449 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
450 enum consumer_endpoint_status status
)
452 struct lttng_ht_iter iter
;
453 struct lttng_consumer_stream
*stream
;
455 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
459 /* Let's begin with metadata */
460 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
461 if (stream
->net_seq_idx
== net_seq_idx
) {
462 uatomic_set(&stream
->endpoint_status
, status
);
463 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
467 /* Follow up by the data streams */
468 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
469 if (stream
->net_seq_idx
== net_seq_idx
) {
470 uatomic_set(&stream
->endpoint_status
, status
);
471 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
478 * Cleanup a relayd object by flagging every associated streams for deletion,
479 * destroying the object meaning removing it from the relayd hash table,
480 * closing the sockets and freeing the memory in a RCU call.
482 * If a local data context is available, notify the threads that the streams'
483 * state have changed.
485 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
491 DBG("Cleaning up relayd object ID %"PRIu64
, relayd
->net_seq_idx
);
493 /* Save the net sequence index before destroying the object */
494 netidx
= relayd
->net_seq_idx
;
497 * Delete the relayd from the relayd hash table, close the sockets and free
498 * the object in a RCU call.
500 consumer_destroy_relayd(relayd
);
502 /* Set inactive endpoint to all streams */
503 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
506 * With a local data context, notify the threads that the streams' state
507 * have changed. The write() action on the pipe acts as an "implicit"
508 * memory barrier ordering the updates of the end point status from the
509 * read of this status which happens AFTER receiving this notify.
511 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
512 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
516 * Flag a relayd socket pair for destruction. Destroy it if the refcount
519 * RCU read side lock MUST be aquired before calling this function.
521 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
525 /* Set destroy flag for this object */
526 uatomic_set(&relayd
->destroy_flag
, 1);
528 /* Destroy the relayd if refcount is 0 */
529 if (uatomic_read(&relayd
->refcount
) == 0) {
530 consumer_destroy_relayd(relayd
);
535 * Completly destroy stream from every visiable data structure and the given
538 * One this call returns, the stream object is not longer usable nor visible.
540 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
543 consumer_stream_destroy(stream
, ht
);
547 * XXX naming of del vs destroy is all mixed up.
549 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
551 consumer_stream_destroy(stream
, data_ht
);
554 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
556 consumer_stream_destroy(stream
, metadata_ht
);
559 void consumer_stream_update_channel_attributes(
560 struct lttng_consumer_stream
*stream
,
561 struct lttng_consumer_channel
*channel
)
563 stream
->channel_read_only_attributes
.tracefile_size
=
564 channel
->tracefile_size
;
567 struct lttng_consumer_stream
*consumer_allocate_stream(
568 struct lttng_consumer_channel
*channel
,
569 uint64_t channel_key
,
571 const char *channel_name
,
574 struct lttng_trace_chunk
*trace_chunk
,
577 enum consumer_channel_type type
,
578 unsigned int monitor
)
581 struct lttng_consumer_stream
*stream
;
583 stream
= zmalloc(sizeof(*stream
));
584 if (stream
== NULL
) {
585 PERROR("malloc struct lttng_consumer_stream");
590 if (trace_chunk
&& !lttng_trace_chunk_get(trace_chunk
)) {
591 ERR("Failed to acquire trace chunk reference during the creation of a stream");
597 stream
->chan
= channel
;
598 stream
->key
= stream_key
;
599 stream
->trace_chunk
= trace_chunk
;
601 stream
->out_fd_offset
= 0;
602 stream
->output_written
= 0;
603 stream
->net_seq_idx
= relayd_id
;
604 stream
->session_id
= session_id
;
605 stream
->monitor
= monitor
;
606 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
607 stream
->index_file
= NULL
;
608 stream
->last_sequence_number
= -1ULL;
609 stream
->rotate_position
= -1ULL;
610 pthread_mutex_init(&stream
->lock
, NULL
);
611 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
613 /* If channel is the metadata, flag this stream as metadata. */
614 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
615 stream
->metadata_flag
= 1;
616 /* Metadata is flat out. */
617 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
618 /* Live rendez-vous point. */
619 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
620 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
622 /* Format stream name to <channel_name>_<cpu_number> */
623 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
626 PERROR("snprintf stream name");
631 /* Key is always the wait_fd for streams. */
632 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
634 /* Init node per channel id key */
635 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
637 /* Init session id node with the stream session id */
638 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
640 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
641 " relayd_id %" PRIu64
", session_id %" PRIu64
,
642 stream
->name
, stream
->key
, channel_key
,
643 stream
->net_seq_idx
, stream
->session_id
);
650 lttng_trace_chunk_put(stream
->trace_chunk
);
660 * Add a stream to the global list protected by a mutex.
662 void consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
664 struct lttng_ht
*ht
= data_ht
;
669 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
671 pthread_mutex_lock(&consumer_data
.lock
);
672 pthread_mutex_lock(&stream
->chan
->lock
);
673 pthread_mutex_lock(&stream
->chan
->timer_lock
);
674 pthread_mutex_lock(&stream
->lock
);
677 /* Steal stream identifier to avoid having streams with the same key */
678 steal_stream_key(stream
->key
, ht
);
680 lttng_ht_add_unique_u64(ht
, &stream
->node
);
682 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
683 &stream
->node_channel_id
);
686 * Add stream to the stream_list_ht of the consumer data. No need to steal
687 * the key since the HT does not use it and we allow to add redundant keys
690 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
693 * When nb_init_stream_left reaches 0, we don't need to trigger any action
694 * in terms of destroying the associated channel, because the action that
695 * causes the count to become 0 also causes a stream to be added. The
696 * channel deletion will thus be triggered by the following removal of this
699 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
700 /* Increment refcount before decrementing nb_init_stream_left */
702 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
705 /* Update consumer data once the node is inserted. */
706 consumer_data
.stream_count
++;
707 consumer_data
.need_update
= 1;
710 pthread_mutex_unlock(&stream
->lock
);
711 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
712 pthread_mutex_unlock(&stream
->chan
->lock
);
713 pthread_mutex_unlock(&consumer_data
.lock
);
716 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
718 consumer_del_stream(stream
, data_ht
);
722 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
723 * be acquired before calling this.
725 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
728 struct lttng_ht_node_u64
*node
;
729 struct lttng_ht_iter iter
;
733 lttng_ht_lookup(consumer_data
.relayd_ht
,
734 &relayd
->net_seq_idx
, &iter
);
735 node
= lttng_ht_iter_get_node_u64(&iter
);
739 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
746 * Allocate and return a consumer relayd socket.
748 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
749 uint64_t net_seq_idx
)
751 struct consumer_relayd_sock_pair
*obj
= NULL
;
753 /* net sequence index of -1 is a failure */
754 if (net_seq_idx
== (uint64_t) -1ULL) {
758 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
760 PERROR("zmalloc relayd sock");
764 obj
->net_seq_idx
= net_seq_idx
;
766 obj
->destroy_flag
= 0;
767 obj
->control_sock
.sock
.fd
= -1;
768 obj
->data_sock
.sock
.fd
= -1;
769 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
770 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
777 * Find a relayd socket pair in the global consumer data.
779 * Return the object if found else NULL.
780 * RCU read-side lock must be held across this call and while using the
783 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
785 struct lttng_ht_iter iter
;
786 struct lttng_ht_node_u64
*node
;
787 struct consumer_relayd_sock_pair
*relayd
= NULL
;
789 /* Negative keys are lookup failures */
790 if (key
== (uint64_t) -1ULL) {
794 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
796 node
= lttng_ht_iter_get_node_u64(&iter
);
798 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
806 * Find a relayd and send the stream
808 * Returns 0 on success, < 0 on error
810 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
814 struct consumer_relayd_sock_pair
*relayd
;
817 assert(stream
->net_seq_idx
!= -1ULL);
820 /* The stream is not metadata. Get relayd reference if exists. */
822 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
823 if (relayd
!= NULL
) {
824 /* Add stream on the relayd */
825 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
826 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
827 path
, &stream
->relayd_stream_id
,
828 stream
->chan
->tracefile_size
,
829 stream
->chan
->tracefile_count
,
830 stream
->trace_chunk
);
831 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
833 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
834 lttng_consumer_cleanup_relayd(relayd
);
838 uatomic_inc(&relayd
->refcount
);
839 stream
->sent_to_relayd
= 1;
841 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
842 stream
->key
, stream
->net_seq_idx
);
847 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
848 stream
->name
, stream
->key
, stream
->net_seq_idx
);
856 * Find a relayd and send the streams sent message
858 * Returns 0 on success, < 0 on error
860 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
863 struct consumer_relayd_sock_pair
*relayd
;
865 assert(net_seq_idx
!= -1ULL);
867 /* The stream is not metadata. Get relayd reference if exists. */
869 relayd
= consumer_find_relayd(net_seq_idx
);
870 if (relayd
!= NULL
) {
871 /* Add stream on the relayd */
872 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
873 ret
= relayd_streams_sent(&relayd
->control_sock
);
874 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
876 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
877 lttng_consumer_cleanup_relayd(relayd
);
881 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
888 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
896 * Find a relayd and close the stream
898 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
900 struct consumer_relayd_sock_pair
*relayd
;
902 /* The stream is not metadata. Get relayd reference if exists. */
904 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
906 consumer_stream_relayd_close(stream
, relayd
);
912 * Handle stream for relayd transmission if the stream applies for network
913 * streaming where the net sequence index is set.
915 * Return destination file descriptor or negative value on error.
917 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
918 size_t data_size
, unsigned long padding
,
919 struct consumer_relayd_sock_pair
*relayd
)
922 struct lttcomm_relayd_data_hdr data_hdr
;
928 /* Reset data header */
929 memset(&data_hdr
, 0, sizeof(data_hdr
));
931 if (stream
->metadata_flag
) {
932 /* Caller MUST acquire the relayd control socket lock */
933 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
938 /* Metadata are always sent on the control socket. */
939 outfd
= relayd
->control_sock
.sock
.fd
;
941 /* Set header with stream information */
942 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
943 data_hdr
.data_size
= htobe32(data_size
);
944 data_hdr
.padding_size
= htobe32(padding
);
947 * Note that net_seq_num below is assigned with the *current* value of
948 * next_net_seq_num and only after that the next_net_seq_num will be
949 * increment. This is why when issuing a command on the relayd using
950 * this next value, 1 should always be substracted in order to compare
951 * the last seen sequence number on the relayd side to the last sent.
953 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
954 /* Other fields are zeroed previously */
956 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
962 ++stream
->next_net_seq_num
;
964 /* Set to go on data socket */
965 outfd
= relayd
->data_sock
.sock
.fd
;
973 * Trigger a dump of the metadata content. Following/during the succesful
974 * completion of this call, the metadata poll thread will start receiving
975 * metadata packets to consume.
977 * The caller must hold the channel and stream locks.
980 int consumer_metadata_stream_dump(struct lttng_consumer_stream
*stream
)
984 ASSERT_LOCKED(stream
->chan
->lock
);
985 ASSERT_LOCKED(stream
->lock
);
986 assert(stream
->metadata_flag
);
987 assert(stream
->chan
->trace_chunk
);
989 switch (consumer_data
.type
) {
990 case LTTNG_CONSUMER_KERNEL
:
992 * Reset the position of what has been read from the
993 * metadata cache to 0 so we can dump it again.
995 ret
= kernctl_metadata_cache_dump(stream
->wait_fd
);
997 case LTTNG_CONSUMER32_UST
:
998 case LTTNG_CONSUMER64_UST
:
1000 * Reset the position pushed from the metadata cache so it
1001 * will write from the beginning on the next push.
1003 stream
->ust_metadata_pushed
= 0;
1004 ret
= consumer_metadata_wakeup_pipe(stream
->chan
);
1007 ERR("Unknown consumer_data type");
1011 ERR("Failed to dump the metadata cache");
1017 int lttng_consumer_channel_set_trace_chunk(
1018 struct lttng_consumer_channel
*channel
,
1019 struct lttng_trace_chunk
*new_trace_chunk
)
1021 pthread_mutex_lock(&channel
->lock
);
1022 if (channel
->is_deleted
) {
1024 * The channel has been logically deleted and should no longer
1025 * be used. It has released its reference to its current trace
1026 * chunk and should not acquire a new one.
1028 * Return success as there is nothing for the caller to do.
1034 * The acquisition of the reference cannot fail (barring
1035 * a severe internal error) since a reference to the published
1036 * chunk is already held by the caller.
1038 if (new_trace_chunk
) {
1039 const bool acquired_reference
= lttng_trace_chunk_get(
1042 assert(acquired_reference
);
1045 lttng_trace_chunk_put(channel
->trace_chunk
);
1046 channel
->trace_chunk
= new_trace_chunk
;
1048 pthread_mutex_unlock(&channel
->lock
);
1053 * Allocate and return a new lttng_consumer_channel object using the given key
1054 * to initialize the hash table node.
1056 * On error, return NULL.
1058 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
1059 uint64_t session_id
,
1060 const uint64_t *chunk_id
,
1061 const char *pathname
,
1064 enum lttng_event_output output
,
1065 uint64_t tracefile_size
,
1066 uint64_t tracefile_count
,
1067 uint64_t session_id_per_pid
,
1068 unsigned int monitor
,
1069 unsigned int live_timer_interval
,
1070 const char *root_shm_path
,
1071 const char *shm_path
)
1073 struct lttng_consumer_channel
*channel
= NULL
;
1074 struct lttng_trace_chunk
*trace_chunk
= NULL
;
1077 trace_chunk
= lttng_trace_chunk_registry_find_chunk(
1078 consumer_data
.chunk_registry
, session_id
,
1081 ERR("Failed to find trace chunk reference during creation of channel");
1086 channel
= zmalloc(sizeof(*channel
));
1087 if (channel
== NULL
) {
1088 PERROR("malloc struct lttng_consumer_channel");
1093 channel
->refcount
= 0;
1094 channel
->session_id
= session_id
;
1095 channel
->session_id_per_pid
= session_id_per_pid
;
1096 channel
->relayd_id
= relayd_id
;
1097 channel
->tracefile_size
= tracefile_size
;
1098 channel
->tracefile_count
= tracefile_count
;
1099 channel
->monitor
= monitor
;
1100 channel
->live_timer_interval
= live_timer_interval
;
1101 pthread_mutex_init(&channel
->lock
, NULL
);
1102 pthread_mutex_init(&channel
->timer_lock
, NULL
);
1105 case LTTNG_EVENT_SPLICE
:
1106 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
1108 case LTTNG_EVENT_MMAP
:
1109 channel
->output
= CONSUMER_CHANNEL_MMAP
;
1119 * In monitor mode, the streams associated with the channel will be put in
1120 * a special list ONLY owned by this channel. So, the refcount is set to 1
1121 * here meaning that the channel itself has streams that are referenced.
1123 * On a channel deletion, once the channel is no longer visible, the
1124 * refcount is decremented and checked for a zero value to delete it. With
1125 * streams in no monitor mode, it will now be safe to destroy the channel.
1127 if (!channel
->monitor
) {
1128 channel
->refcount
= 1;
1131 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1132 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1134 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1135 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1137 if (root_shm_path
) {
1138 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1139 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1142 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1143 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1146 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1147 lttng_ht_node_init_u64(&channel
->channels_by_session_id_ht_node
,
1148 channel
->session_id
);
1150 channel
->wait_fd
= -1;
1151 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1154 int ret
= lttng_consumer_channel_set_trace_chunk(channel
,
1161 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1164 lttng_trace_chunk_put(trace_chunk
);
1167 consumer_del_channel(channel
);
1173 * Add a channel to the global list protected by a mutex.
1175 * Always return 0 indicating success.
1177 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1178 struct lttng_consumer_local_data
*ctx
)
1180 pthread_mutex_lock(&consumer_data
.lock
);
1181 pthread_mutex_lock(&channel
->lock
);
1182 pthread_mutex_lock(&channel
->timer_lock
);
1185 * This gives us a guarantee that the channel we are about to add to the
1186 * channel hash table will be unique. See this function comment on the why
1187 * we need to steel the channel key at this stage.
1189 steal_channel_key(channel
->key
);
1192 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1193 lttng_ht_add_u64(consumer_data
.channels_by_session_id_ht
,
1194 &channel
->channels_by_session_id_ht_node
);
1196 channel
->is_published
= true;
1198 pthread_mutex_unlock(&channel
->timer_lock
);
1199 pthread_mutex_unlock(&channel
->lock
);
1200 pthread_mutex_unlock(&consumer_data
.lock
);
1202 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1203 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1210 * Allocate the pollfd structure and the local view of the out fds to avoid
1211 * doing a lookup in the linked list and concurrency issues when writing is
1212 * needed. Called with consumer_data.lock held.
1214 * Returns the number of fds in the structures.
1216 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1217 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1218 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1221 struct lttng_ht_iter iter
;
1222 struct lttng_consumer_stream
*stream
;
1227 assert(local_stream
);
1229 DBG("Updating poll fd array");
1230 *nb_inactive_fd
= 0;
1232 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1234 * Only active streams with an active end point can be added to the
1235 * poll set and local stream storage of the thread.
1237 * There is a potential race here for endpoint_status to be updated
1238 * just after the check. However, this is OK since the stream(s) will
1239 * be deleted once the thread is notified that the end point state has
1240 * changed where this function will be called back again.
1242 * We track the number of inactive FDs because they still need to be
1243 * closed by the polling thread after a wakeup on the data_pipe or
1246 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1247 (*nb_inactive_fd
)++;
1251 * This clobbers way too much the debug output. Uncomment that if you
1252 * need it for debugging purposes.
1254 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1255 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1256 local_stream
[i
] = stream
;
1262 * Insert the consumer_data_pipe at the end of the array and don't
1263 * increment i so nb_fd is the number of real FD.
1265 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1266 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1268 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1269 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1274 * Poll on the should_quit pipe and the command socket return -1 on
1275 * error, 1 if should exit, 0 if data is available on the command socket
1277 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1282 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1283 if (num_rdy
== -1) {
1285 * Restart interrupted system call.
1287 if (errno
== EINTR
) {
1290 PERROR("Poll error");
1293 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1294 DBG("consumer_should_quit wake up");
1301 * Set the error socket.
1303 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1306 ctx
->consumer_error_socket
= sock
;
1310 * Set the command socket path.
1312 void lttng_consumer_set_command_sock_path(
1313 struct lttng_consumer_local_data
*ctx
, char *sock
)
1315 ctx
->consumer_command_sock_path
= sock
;
1319 * Send return code to the session daemon.
1320 * If the socket is not defined, we return 0, it is not a fatal error
1322 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1324 if (ctx
->consumer_error_socket
> 0) {
1325 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1326 sizeof(enum lttcomm_sessiond_command
));
1333 * Close all the tracefiles and stream fds and MUST be called when all
1334 * instances are destroyed i.e. when all threads were joined and are ended.
1336 void lttng_consumer_cleanup(void)
1338 struct lttng_ht_iter iter
;
1339 struct lttng_consumer_channel
*channel
;
1340 unsigned int trace_chunks_left
;
1344 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1346 consumer_del_channel(channel
);
1351 lttng_ht_destroy(consumer_data
.channel_ht
);
1352 lttng_ht_destroy(consumer_data
.channels_by_session_id_ht
);
1354 cleanup_relayd_ht();
1356 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1359 * This HT contains streams that are freed by either the metadata thread or
1360 * the data thread so we do *nothing* on the hash table and simply destroy
1363 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1366 * Trace chunks in the registry may still exist if the session
1367 * daemon has encountered an internal error and could not
1368 * tear down its sessions and/or trace chunks properly.
1370 * Release the session daemon's implicit reference to any remaining
1371 * trace chunk and print an error if any trace chunk was found. Note
1372 * that there are _no_ legitimate cases for trace chunks to be left,
1373 * it is a leak. However, it can happen following a crash of the
1374 * session daemon and not emptying the registry would cause an assertion
1377 trace_chunks_left
= lttng_trace_chunk_registry_put_each_chunk(
1378 consumer_data
.chunk_registry
);
1379 if (trace_chunks_left
) {
1380 ERR("%u trace chunks are leaked by lttng-consumerd. "
1381 "This can be caused by an internal error of the session daemon.",
1384 /* Run all callbacks freeing each chunk. */
1386 lttng_trace_chunk_registry_destroy(consumer_data
.chunk_registry
);
1390 * Called from signal handler.
1392 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1396 CMM_STORE_SHARED(consumer_quit
, 1);
1397 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1399 PERROR("write consumer quit");
1402 DBG("Consumer flag that it should quit");
1407 * Flush pending writes to trace output disk file.
1410 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1414 int outfd
= stream
->out_fd
;
1417 * This does a blocking write-and-wait on any page that belongs to the
1418 * subbuffer prior to the one we just wrote.
1419 * Don't care about error values, as these are just hints and ways to
1420 * limit the amount of page cache used.
1422 if (orig_offset
< stream
->max_sb_size
) {
1425 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1426 stream
->max_sb_size
,
1427 SYNC_FILE_RANGE_WAIT_BEFORE
1428 | SYNC_FILE_RANGE_WRITE
1429 | SYNC_FILE_RANGE_WAIT_AFTER
);
1431 * Give hints to the kernel about how we access the file:
1432 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1435 * We need to call fadvise again after the file grows because the
1436 * kernel does not seem to apply fadvise to non-existing parts of the
1439 * Call fadvise _after_ having waited for the page writeback to
1440 * complete because the dirty page writeback semantic is not well
1441 * defined. So it can be expected to lead to lower throughput in
1444 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1445 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1446 if (ret
&& ret
!= -ENOSYS
) {
1448 PERROR("posix_fadvise on fd %i", outfd
);
1453 * Initialise the necessary environnement :
1454 * - create a new context
1455 * - create the poll_pipe
1456 * - create the should_quit pipe (for signal handler)
1457 * - create the thread pipe (for splice)
1459 * Takes a function pointer as argument, this function is called when data is
1460 * available on a buffer. This function is responsible to do the
1461 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1462 * buffer configuration and then kernctl_put_next_subbuf at the end.
1464 * Returns a pointer to the new context or NULL on error.
1466 struct lttng_consumer_local_data
*lttng_consumer_create(
1467 enum lttng_consumer_type type
,
1468 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1469 struct lttng_consumer_local_data
*ctx
),
1470 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1471 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1472 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1475 struct lttng_consumer_local_data
*ctx
;
1477 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1478 consumer_data
.type
== type
);
1479 consumer_data
.type
= type
;
1481 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1483 PERROR("allocating context");
1487 ctx
->consumer_error_socket
= -1;
1488 ctx
->consumer_metadata_socket
= -1;
1489 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1490 /* assign the callbacks */
1491 ctx
->on_buffer_ready
= buffer_ready
;
1492 ctx
->on_recv_channel
= recv_channel
;
1493 ctx
->on_recv_stream
= recv_stream
;
1494 ctx
->on_update_stream
= update_stream
;
1496 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1497 if (!ctx
->consumer_data_pipe
) {
1498 goto error_poll_pipe
;
1501 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1502 if (!ctx
->consumer_wakeup_pipe
) {
1503 goto error_wakeup_pipe
;
1506 ret
= pipe(ctx
->consumer_should_quit
);
1508 PERROR("Error creating recv pipe");
1509 goto error_quit_pipe
;
1512 ret
= pipe(ctx
->consumer_channel_pipe
);
1514 PERROR("Error creating channel pipe");
1515 goto error_channel_pipe
;
1518 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1519 if (!ctx
->consumer_metadata_pipe
) {
1520 goto error_metadata_pipe
;
1523 ctx
->channel_monitor_pipe
= -1;
1527 error_metadata_pipe
:
1528 utils_close_pipe(ctx
->consumer_channel_pipe
);
1530 utils_close_pipe(ctx
->consumer_should_quit
);
1532 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1534 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1542 * Iterate over all streams of the hashtable and free them properly.
1544 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1546 struct lttng_ht_iter iter
;
1547 struct lttng_consumer_stream
*stream
;
1554 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1556 * Ignore return value since we are currently cleaning up so any error
1559 (void) consumer_del_stream(stream
, ht
);
1563 lttng_ht_destroy(ht
);
1567 * Iterate over all streams of the metadata hashtable and free them
1570 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1572 struct lttng_ht_iter iter
;
1573 struct lttng_consumer_stream
*stream
;
1580 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1582 * Ignore return value since we are currently cleaning up so any error
1585 (void) consumer_del_metadata_stream(stream
, ht
);
1589 lttng_ht_destroy(ht
);
1593 * Close all fds associated with the instance and free the context.
1595 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1599 DBG("Consumer destroying it. Closing everything.");
1605 destroy_data_stream_ht(data_ht
);
1606 destroy_metadata_stream_ht(metadata_ht
);
1608 ret
= close(ctx
->consumer_error_socket
);
1612 ret
= close(ctx
->consumer_metadata_socket
);
1616 utils_close_pipe(ctx
->consumer_channel_pipe
);
1617 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1618 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1619 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1620 utils_close_pipe(ctx
->consumer_should_quit
);
1622 unlink(ctx
->consumer_command_sock_path
);
1627 * Write the metadata stream id on the specified file descriptor.
1629 static int write_relayd_metadata_id(int fd
,
1630 struct lttng_consumer_stream
*stream
,
1631 unsigned long padding
)
1634 struct lttcomm_relayd_metadata_payload hdr
;
1636 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1637 hdr
.padding_size
= htobe32(padding
);
1638 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1639 if (ret
< sizeof(hdr
)) {
1641 * This error means that the fd's end is closed so ignore the PERROR
1642 * not to clubber the error output since this can happen in a normal
1645 if (errno
!= EPIPE
) {
1646 PERROR("write metadata stream id");
1648 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1650 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1651 * handle writting the missing part so report that as an error and
1652 * don't lie to the caller.
1657 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1658 stream
->relayd_stream_id
, padding
);
1665 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1666 * core function for writing trace buffers to either the local filesystem or
1669 * It must be called with the stream and the channel lock held.
1671 * Careful review MUST be put if any changes occur!
1673 * Returns the number of bytes written
1675 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1676 struct lttng_consumer_local_data
*ctx
,
1677 struct lttng_consumer_stream
*stream
,
1678 const struct lttng_buffer_view
*buffer
,
1679 unsigned long padding
,
1680 struct ctf_packet_index
*index
)
1683 off_t orig_offset
= stream
->out_fd_offset
;
1684 /* Default is on the disk */
1685 int outfd
= stream
->out_fd
;
1686 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1687 unsigned int relayd_hang_up
= 0;
1688 const size_t subbuf_content_size
= buffer
->size
- padding
;
1691 /* RCU lock for the relayd pointer */
1693 assert(stream
->net_seq_idx
!= (uint64_t) -1ULL ||
1694 stream
->trace_chunk
);
1696 /* Flag that the current stream if set for network streaming. */
1697 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1698 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1699 if (relayd
== NULL
) {
1705 /* Handle stream on the relayd if the output is on the network */
1707 unsigned long netlen
= subbuf_content_size
;
1710 * Lock the control socket for the complete duration of the function
1711 * since from this point on we will use the socket.
1713 if (stream
->metadata_flag
) {
1714 /* Metadata requires the control socket. */
1715 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1716 if (stream
->reset_metadata_flag
) {
1717 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1718 stream
->relayd_stream_id
,
1719 stream
->metadata_version
);
1724 stream
->reset_metadata_flag
= 0;
1726 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1729 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1734 /* Use the returned socket. */
1737 /* Write metadata stream id before payload */
1738 if (stream
->metadata_flag
) {
1739 ret
= write_relayd_metadata_id(outfd
, stream
, padding
);
1746 write_len
= subbuf_content_size
;
1748 /* No streaming; we have to write the full padding. */
1749 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1750 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1752 ERR("Reset metadata file");
1755 stream
->reset_metadata_flag
= 0;
1759 * Check if we need to change the tracefile before writing the packet.
1761 if (stream
->chan
->tracefile_size
> 0 &&
1762 (stream
->tracefile_size_current
+ buffer
->size
) >
1763 stream
->chan
->tracefile_size
) {
1764 ret
= consumer_stream_rotate_output_files(stream
);
1768 outfd
= stream
->out_fd
;
1771 stream
->tracefile_size_current
+= buffer
->size
;
1773 index
->offset
= htobe64(stream
->out_fd_offset
);
1776 write_len
= buffer
->size
;
1780 * This call guarantee that len or less is returned. It's impossible to
1781 * receive a ret value that is bigger than len.
1783 ret
= lttng_write(outfd
, buffer
->data
, write_len
);
1784 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, write_len
);
1785 if (ret
< 0 || ((size_t) ret
!= write_len
)) {
1787 * Report error to caller if nothing was written else at least send the
1795 /* Socket operation failed. We consider the relayd dead */
1796 if (errno
== EPIPE
) {
1798 * This is possible if the fd is closed on the other side
1799 * (outfd) or any write problem. It can be verbose a bit for a
1800 * normal execution if for instance the relayd is stopped
1801 * abruptly. This can happen so set this to a DBG statement.
1803 DBG("Consumer mmap write detected relayd hang up");
1805 /* Unhandled error, print it and stop function right now. */
1806 PERROR("Error in write mmap (ret %zd != write_len %zu)", ret
,
1811 stream
->output_written
+= ret
;
1813 /* This call is useless on a socket so better save a syscall. */
1815 /* This won't block, but will start writeout asynchronously */
1816 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, write_len
,
1817 SYNC_FILE_RANGE_WRITE
);
1818 stream
->out_fd_offset
+= write_len
;
1819 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1824 * This is a special case that the relayd has closed its socket. Let's
1825 * cleanup the relayd object and all associated streams.
1827 if (relayd
&& relayd_hang_up
) {
1828 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1829 lttng_consumer_cleanup_relayd(relayd
);
1833 /* Unlock only if ctrl socket used */
1834 if (relayd
&& stream
->metadata_flag
) {
1835 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1843 * Splice the data from the ring buffer to the tracefile.
1845 * It must be called with the stream lock held.
1847 * Returns the number of bytes spliced.
1849 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1850 struct lttng_consumer_local_data
*ctx
,
1851 struct lttng_consumer_stream
*stream
, unsigned long len
,
1852 unsigned long padding
,
1853 struct ctf_packet_index
*index
)
1855 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1857 off_t orig_offset
= stream
->out_fd_offset
;
1858 int fd
= stream
->wait_fd
;
1859 /* Default is on the disk */
1860 int outfd
= stream
->out_fd
;
1861 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1863 unsigned int relayd_hang_up
= 0;
1865 switch (consumer_data
.type
) {
1866 case LTTNG_CONSUMER_KERNEL
:
1868 case LTTNG_CONSUMER32_UST
:
1869 case LTTNG_CONSUMER64_UST
:
1870 /* Not supported for user space tracing */
1873 ERR("Unknown consumer_data type");
1877 /* RCU lock for the relayd pointer */
1880 /* Flag that the current stream if set for network streaming. */
1881 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1882 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1883 if (relayd
== NULL
) {
1888 splice_pipe
= stream
->splice_pipe
;
1890 /* Write metadata stream id before payload */
1892 unsigned long total_len
= len
;
1894 if (stream
->metadata_flag
) {
1896 * Lock the control socket for the complete duration of the function
1897 * since from this point on we will use the socket.
1899 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1901 if (stream
->reset_metadata_flag
) {
1902 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1903 stream
->relayd_stream_id
,
1904 stream
->metadata_version
);
1909 stream
->reset_metadata_flag
= 0;
1911 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
,
1919 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1922 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1928 /* Use the returned socket. */
1931 /* No streaming, we have to set the len with the full padding */
1934 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1935 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1937 ERR("Reset metadata file");
1940 stream
->reset_metadata_flag
= 0;
1943 * Check if we need to change the tracefile before writing the packet.
1945 if (stream
->chan
->tracefile_size
> 0 &&
1946 (stream
->tracefile_size_current
+ len
) >
1947 stream
->chan
->tracefile_size
) {
1948 ret
= consumer_stream_rotate_output_files(stream
);
1953 outfd
= stream
->out_fd
;
1956 stream
->tracefile_size_current
+= len
;
1957 index
->offset
= htobe64(stream
->out_fd_offset
);
1961 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1962 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1963 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1964 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1965 DBG("splice chan to pipe, ret %zd", ret_splice
);
1966 if (ret_splice
< 0) {
1969 PERROR("Error in relay splice");
1973 /* Handle stream on the relayd if the output is on the network */
1974 if (relayd
&& stream
->metadata_flag
) {
1975 size_t metadata_payload_size
=
1976 sizeof(struct lttcomm_relayd_metadata_payload
);
1978 /* Update counter to fit the spliced data */
1979 ret_splice
+= metadata_payload_size
;
1980 len
+= metadata_payload_size
;
1982 * We do this so the return value can match the len passed as
1983 * argument to this function.
1985 written
-= metadata_payload_size
;
1988 /* Splice data out */
1989 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1990 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1991 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1993 if (ret_splice
< 0) {
1998 } else if (ret_splice
> len
) {
2000 * We don't expect this code path to be executed but you never know
2001 * so this is an extra protection agains a buggy splice().
2004 written
+= ret_splice
;
2005 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
2009 /* All good, update current len and continue. */
2013 /* This call is useless on a socket so better save a syscall. */
2015 /* This won't block, but will start writeout asynchronously */
2016 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
2017 SYNC_FILE_RANGE_WRITE
);
2018 stream
->out_fd_offset
+= ret_splice
;
2020 stream
->output_written
+= ret_splice
;
2021 written
+= ret_splice
;
2024 lttng_consumer_sync_trace_file(stream
, orig_offset
);
2030 * This is a special case that the relayd has closed its socket. Let's
2031 * cleanup the relayd object and all associated streams.
2033 if (relayd
&& relayd_hang_up
) {
2034 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
2035 lttng_consumer_cleanup_relayd(relayd
);
2036 /* Skip splice error so the consumer does not fail */
2041 /* send the appropriate error description to sessiond */
2044 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
2047 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
2050 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
2055 if (relayd
&& stream
->metadata_flag
) {
2056 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2064 * Sample the snapshot positions for a specific fd
2066 * Returns 0 on success, < 0 on error
2068 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream
*stream
)
2070 switch (consumer_data
.type
) {
2071 case LTTNG_CONSUMER_KERNEL
:
2072 return lttng_kconsumer_sample_snapshot_positions(stream
);
2073 case LTTNG_CONSUMER32_UST
:
2074 case LTTNG_CONSUMER64_UST
:
2075 return lttng_ustconsumer_sample_snapshot_positions(stream
);
2077 ERR("Unknown consumer_data type");
2083 * Take a snapshot for a specific fd
2085 * Returns 0 on success, < 0 on error
2087 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
2089 switch (consumer_data
.type
) {
2090 case LTTNG_CONSUMER_KERNEL
:
2091 return lttng_kconsumer_take_snapshot(stream
);
2092 case LTTNG_CONSUMER32_UST
:
2093 case LTTNG_CONSUMER64_UST
:
2094 return lttng_ustconsumer_take_snapshot(stream
);
2096 ERR("Unknown consumer_data type");
2103 * Get the produced position
2105 * Returns 0 on success, < 0 on error
2107 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
2110 switch (consumer_data
.type
) {
2111 case LTTNG_CONSUMER_KERNEL
:
2112 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
2113 case LTTNG_CONSUMER32_UST
:
2114 case LTTNG_CONSUMER64_UST
:
2115 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2117 ERR("Unknown consumer_data type");
2124 * Get the consumed position (free-running counter position in bytes).
2126 * Returns 0 on success, < 0 on error
2128 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream
*stream
,
2131 switch (consumer_data
.type
) {
2132 case LTTNG_CONSUMER_KERNEL
:
2133 return lttng_kconsumer_get_consumed_snapshot(stream
, pos
);
2134 case LTTNG_CONSUMER32_UST
:
2135 case LTTNG_CONSUMER64_UST
:
2136 return lttng_ustconsumer_get_consumed_snapshot(stream
, pos
);
2138 ERR("Unknown consumer_data type");
2144 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2145 int sock
, struct pollfd
*consumer_sockpoll
)
2147 switch (consumer_data
.type
) {
2148 case LTTNG_CONSUMER_KERNEL
:
2149 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2150 case LTTNG_CONSUMER32_UST
:
2151 case LTTNG_CONSUMER64_UST
:
2152 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2154 ERR("Unknown consumer_data type");
2160 void lttng_consumer_close_all_metadata(void)
2162 switch (consumer_data
.type
) {
2163 case LTTNG_CONSUMER_KERNEL
:
2165 * The Kernel consumer has a different metadata scheme so we don't
2166 * close anything because the stream will be closed by the session
2170 case LTTNG_CONSUMER32_UST
:
2171 case LTTNG_CONSUMER64_UST
:
2173 * Close all metadata streams. The metadata hash table is passed and
2174 * this call iterates over it by closing all wakeup fd. This is safe
2175 * because at this point we are sure that the metadata producer is
2176 * either dead or blocked.
2178 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2181 ERR("Unknown consumer_data type");
2187 * Clean up a metadata stream and free its memory.
2189 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2190 struct lttng_ht
*ht
)
2192 struct lttng_consumer_channel
*channel
= NULL
;
2193 bool free_channel
= false;
2197 * This call should NEVER receive regular stream. It must always be
2198 * metadata stream and this is crucial for data structure synchronization.
2200 assert(stream
->metadata_flag
);
2202 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2204 pthread_mutex_lock(&consumer_data
.lock
);
2206 * Note that this assumes that a stream's channel is never changed and
2207 * that the stream's lock doesn't need to be taken to sample its
2210 channel
= stream
->chan
;
2211 pthread_mutex_lock(&channel
->lock
);
2212 pthread_mutex_lock(&stream
->lock
);
2213 if (channel
->metadata_cache
) {
2214 /* Only applicable to userspace consumers. */
2215 pthread_mutex_lock(&channel
->metadata_cache
->lock
);
2218 /* Remove any reference to that stream. */
2219 consumer_stream_delete(stream
, ht
);
2221 /* Close down everything including the relayd if one. */
2222 consumer_stream_close(stream
);
2223 /* Destroy tracer buffers of the stream. */
2224 consumer_stream_destroy_buffers(stream
);
2226 /* Atomically decrement channel refcount since other threads can use it. */
2227 if (!uatomic_sub_return(&channel
->refcount
, 1)
2228 && !uatomic_read(&channel
->nb_init_stream_left
)) {
2229 /* Go for channel deletion! */
2230 free_channel
= true;
2232 stream
->chan
= NULL
;
2235 * Nullify the stream reference so it is not used after deletion. The
2236 * channel lock MUST be acquired before being able to check for a NULL
2239 channel
->metadata_stream
= NULL
;
2241 if (channel
->metadata_cache
) {
2242 pthread_mutex_unlock(&channel
->metadata_cache
->lock
);
2244 pthread_mutex_unlock(&stream
->lock
);
2245 pthread_mutex_unlock(&channel
->lock
);
2246 pthread_mutex_unlock(&consumer_data
.lock
);
2249 consumer_del_channel(channel
);
2252 lttng_trace_chunk_put(stream
->trace_chunk
);
2253 stream
->trace_chunk
= NULL
;
2254 consumer_stream_free(stream
);
2258 * Action done with the metadata stream when adding it to the consumer internal
2259 * data structures to handle it.
2261 void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2263 struct lttng_ht
*ht
= metadata_ht
;
2264 struct lttng_ht_iter iter
;
2265 struct lttng_ht_node_u64
*node
;
2270 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2272 pthread_mutex_lock(&consumer_data
.lock
);
2273 pthread_mutex_lock(&stream
->chan
->lock
);
2274 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2275 pthread_mutex_lock(&stream
->lock
);
2278 * From here, refcounts are updated so be _careful_ when returning an error
2285 * Lookup the stream just to make sure it does not exist in our internal
2286 * state. This should NEVER happen.
2288 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2289 node
= lttng_ht_iter_get_node_u64(&iter
);
2293 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2294 * in terms of destroying the associated channel, because the action that
2295 * causes the count to become 0 also causes a stream to be added. The
2296 * channel deletion will thus be triggered by the following removal of this
2299 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2300 /* Increment refcount before decrementing nb_init_stream_left */
2302 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2305 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2307 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2308 &stream
->node_channel_id
);
2311 * Add stream to the stream_list_ht of the consumer data. No need to steal
2312 * the key since the HT does not use it and we allow to add redundant keys
2315 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2319 pthread_mutex_unlock(&stream
->lock
);
2320 pthread_mutex_unlock(&stream
->chan
->lock
);
2321 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2322 pthread_mutex_unlock(&consumer_data
.lock
);
2326 * Delete data stream that are flagged for deletion (endpoint_status).
2328 static void validate_endpoint_status_data_stream(void)
2330 struct lttng_ht_iter iter
;
2331 struct lttng_consumer_stream
*stream
;
2333 DBG("Consumer delete flagged data stream");
2336 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2337 /* Validate delete flag of the stream */
2338 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2341 /* Delete it right now */
2342 consumer_del_stream(stream
, data_ht
);
2348 * Delete metadata stream that are flagged for deletion (endpoint_status).
2350 static void validate_endpoint_status_metadata_stream(
2351 struct lttng_poll_event
*pollset
)
2353 struct lttng_ht_iter iter
;
2354 struct lttng_consumer_stream
*stream
;
2356 DBG("Consumer delete flagged metadata stream");
2361 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2362 /* Validate delete flag of the stream */
2363 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2367 * Remove from pollset so the metadata thread can continue without
2368 * blocking on a deleted stream.
2370 lttng_poll_del(pollset
, stream
->wait_fd
);
2372 /* Delete it right now */
2373 consumer_del_metadata_stream(stream
, metadata_ht
);
2379 * Thread polls on metadata file descriptor and write them on disk or on the
2382 void *consumer_thread_metadata_poll(void *data
)
2384 int ret
, i
, pollfd
, err
= -1;
2385 uint32_t revents
, nb_fd
;
2386 struct lttng_consumer_stream
*stream
= NULL
;
2387 struct lttng_ht_iter iter
;
2388 struct lttng_ht_node_u64
*node
;
2389 struct lttng_poll_event events
;
2390 struct lttng_consumer_local_data
*ctx
= data
;
2393 rcu_register_thread();
2395 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2397 if (testpoint(consumerd_thread_metadata
)) {
2398 goto error_testpoint
;
2401 health_code_update();
2403 DBG("Thread metadata poll started");
2405 /* Size is set to 1 for the consumer_metadata pipe */
2406 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2408 ERR("Poll set creation failed");
2412 ret
= lttng_poll_add(&events
,
2413 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2419 DBG("Metadata main loop started");
2423 health_code_update();
2424 health_poll_entry();
2425 DBG("Metadata poll wait");
2426 ret
= lttng_poll_wait(&events
, -1);
2427 DBG("Metadata poll return from wait with %d fd(s)",
2428 LTTNG_POLL_GETNB(&events
));
2430 DBG("Metadata event caught in thread");
2432 if (errno
== EINTR
) {
2433 ERR("Poll EINTR caught");
2436 if (LTTNG_POLL_GETNB(&events
) == 0) {
2437 err
= 0; /* All is OK */
2444 /* From here, the event is a metadata wait fd */
2445 for (i
= 0; i
< nb_fd
; i
++) {
2446 health_code_update();
2448 revents
= LTTNG_POLL_GETEV(&events
, i
);
2449 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2451 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2452 if (revents
& LPOLLIN
) {
2455 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2456 &stream
, sizeof(stream
));
2457 if (pipe_len
< sizeof(stream
)) {
2459 PERROR("read metadata stream");
2462 * Remove the pipe from the poll set and continue the loop
2463 * since their might be data to consume.
2465 lttng_poll_del(&events
,
2466 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2467 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2471 /* A NULL stream means that the state has changed. */
2472 if (stream
== NULL
) {
2473 /* Check for deleted streams. */
2474 validate_endpoint_status_metadata_stream(&events
);
2478 DBG("Adding metadata stream %d to poll set",
2481 /* Add metadata stream to the global poll events list */
2482 lttng_poll_add(&events
, stream
->wait_fd
,
2483 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2484 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2485 DBG("Metadata thread pipe hung up");
2487 * Remove the pipe from the poll set and continue the loop
2488 * since their might be data to consume.
2490 lttng_poll_del(&events
,
2491 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2492 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2495 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2499 /* Handle other stream */
2505 uint64_t tmp_id
= (uint64_t) pollfd
;
2507 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2509 node
= lttng_ht_iter_get_node_u64(&iter
);
2512 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2515 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2516 /* Get the data out of the metadata file descriptor */
2517 DBG("Metadata available on fd %d", pollfd
);
2518 assert(stream
->wait_fd
== pollfd
);
2521 health_code_update();
2523 len
= ctx
->on_buffer_ready(stream
, ctx
);
2525 * We don't check the return value here since if we get
2526 * a negative len, it means an error occurred thus we
2527 * simply remove it from the poll set and free the
2532 /* It's ok to have an unavailable sub-buffer */
2533 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2534 /* Clean up stream from consumer and free it. */
2535 lttng_poll_del(&events
, stream
->wait_fd
);
2536 consumer_del_metadata_stream(stream
, metadata_ht
);
2538 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2539 DBG("Metadata fd %d is hup|err.", pollfd
);
2540 if (!stream
->hangup_flush_done
2541 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2542 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2543 DBG("Attempting to flush and consume the UST buffers");
2544 lttng_ustconsumer_on_stream_hangup(stream
);
2546 /* We just flushed the stream now read it. */
2548 health_code_update();
2550 len
= ctx
->on_buffer_ready(stream
, ctx
);
2552 * We don't check the return value here since if we get
2553 * a negative len, it means an error occurred thus we
2554 * simply remove it from the poll set and free the
2560 lttng_poll_del(&events
, stream
->wait_fd
);
2562 * This call update the channel states, closes file descriptors
2563 * and securely free the stream.
2565 consumer_del_metadata_stream(stream
, metadata_ht
);
2567 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2571 /* Release RCU lock for the stream looked up */
2579 DBG("Metadata poll thread exiting");
2581 lttng_poll_clean(&events
);
2586 ERR("Health error occurred in %s", __func__
);
2588 health_unregister(health_consumerd
);
2589 rcu_unregister_thread();
2594 * This thread polls the fds in the set to consume the data and write
2595 * it to tracefile if necessary.
2597 void *consumer_thread_data_poll(void *data
)
2599 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2600 struct pollfd
*pollfd
= NULL
;
2601 /* local view of the streams */
2602 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2603 /* local view of consumer_data.fds_count */
2605 /* 2 for the consumer_data_pipe and wake up pipe */
2606 const int nb_pipes_fd
= 2;
2607 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2608 int nb_inactive_fd
= 0;
2609 struct lttng_consumer_local_data
*ctx
= data
;
2612 rcu_register_thread();
2614 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2616 if (testpoint(consumerd_thread_data
)) {
2617 goto error_testpoint
;
2620 health_code_update();
2622 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2623 if (local_stream
== NULL
) {
2624 PERROR("local_stream malloc");
2629 health_code_update();
2635 * the fds set has been updated, we need to update our
2636 * local array as well
2638 pthread_mutex_lock(&consumer_data
.lock
);
2639 if (consumer_data
.need_update
) {
2644 local_stream
= NULL
;
2646 /* Allocate for all fds */
2647 pollfd
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) * sizeof(struct pollfd
));
2648 if (pollfd
== NULL
) {
2649 PERROR("pollfd malloc");
2650 pthread_mutex_unlock(&consumer_data
.lock
);
2654 local_stream
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) *
2655 sizeof(struct lttng_consumer_stream
*));
2656 if (local_stream
== NULL
) {
2657 PERROR("local_stream malloc");
2658 pthread_mutex_unlock(&consumer_data
.lock
);
2661 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2662 data_ht
, &nb_inactive_fd
);
2664 ERR("Error in allocating pollfd or local_outfds");
2665 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2666 pthread_mutex_unlock(&consumer_data
.lock
);
2670 consumer_data
.need_update
= 0;
2672 pthread_mutex_unlock(&consumer_data
.lock
);
2674 /* No FDs and consumer_quit, consumer_cleanup the thread */
2675 if (nb_fd
== 0 && nb_inactive_fd
== 0 &&
2676 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2677 err
= 0; /* All is OK */
2680 /* poll on the array of fds */
2682 DBG("polling on %d fd", nb_fd
+ nb_pipes_fd
);
2683 if (testpoint(consumerd_thread_data_poll
)) {
2686 health_poll_entry();
2687 num_rdy
= poll(pollfd
, nb_fd
+ nb_pipes_fd
, -1);
2689 DBG("poll num_rdy : %d", num_rdy
);
2690 if (num_rdy
== -1) {
2692 * Restart interrupted system call.
2694 if (errno
== EINTR
) {
2697 PERROR("Poll error");
2698 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2700 } else if (num_rdy
== 0) {
2701 DBG("Polling thread timed out");
2705 if (caa_unlikely(data_consumption_paused
)) {
2706 DBG("Data consumption paused, sleeping...");
2712 * If the consumer_data_pipe triggered poll go directly to the
2713 * beginning of the loop to update the array. We want to prioritize
2714 * array update over low-priority reads.
2716 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2717 ssize_t pipe_readlen
;
2719 DBG("consumer_data_pipe wake up");
2720 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2721 &new_stream
, sizeof(new_stream
));
2722 if (pipe_readlen
< sizeof(new_stream
)) {
2723 PERROR("Consumer data pipe");
2724 /* Continue so we can at least handle the current stream(s). */
2729 * If the stream is NULL, just ignore it. It's also possible that
2730 * the sessiond poll thread changed the consumer_quit state and is
2731 * waking us up to test it.
2733 if (new_stream
== NULL
) {
2734 validate_endpoint_status_data_stream();
2738 /* Continue to update the local streams and handle prio ones */
2742 /* Handle wakeup pipe. */
2743 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2745 ssize_t pipe_readlen
;
2747 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2749 if (pipe_readlen
< 0) {
2750 PERROR("Consumer data wakeup pipe");
2752 /* We've been awakened to handle stream(s). */
2753 ctx
->has_wakeup
= 0;
2756 /* Take care of high priority channels first. */
2757 for (i
= 0; i
< nb_fd
; i
++) {
2758 health_code_update();
2760 if (local_stream
[i
] == NULL
) {
2763 if (pollfd
[i
].revents
& POLLPRI
) {
2764 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2766 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2767 /* it's ok to have an unavailable sub-buffer */
2768 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2769 /* Clean the stream and free it. */
2770 consumer_del_stream(local_stream
[i
], data_ht
);
2771 local_stream
[i
] = NULL
;
2772 } else if (len
> 0) {
2773 local_stream
[i
]->data_read
= 1;
2779 * If we read high prio channel in this loop, try again
2780 * for more high prio data.
2786 /* Take care of low priority channels. */
2787 for (i
= 0; i
< nb_fd
; i
++) {
2788 health_code_update();
2790 if (local_stream
[i
] == NULL
) {
2793 if ((pollfd
[i
].revents
& POLLIN
) ||
2794 local_stream
[i
]->hangup_flush_done
||
2795 local_stream
[i
]->has_data
) {
2796 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2797 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2798 /* it's ok to have an unavailable sub-buffer */
2799 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2800 /* Clean the stream and free it. */
2801 consumer_del_stream(local_stream
[i
], data_ht
);
2802 local_stream
[i
] = NULL
;
2803 } else if (len
> 0) {
2804 local_stream
[i
]->data_read
= 1;
2809 /* Handle hangup and errors */
2810 for (i
= 0; i
< nb_fd
; i
++) {
2811 health_code_update();
2813 if (local_stream
[i
] == NULL
) {
2816 if (!local_stream
[i
]->hangup_flush_done
2817 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2818 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2819 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2820 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2822 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2823 /* Attempt read again, for the data we just flushed. */
2824 local_stream
[i
]->data_read
= 1;
2827 * If the poll flag is HUP/ERR/NVAL and we have
2828 * read no data in this pass, we can remove the
2829 * stream from its hash table.
2831 if ((pollfd
[i
].revents
& POLLHUP
)) {
2832 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2833 if (!local_stream
[i
]->data_read
) {
2834 consumer_del_stream(local_stream
[i
], data_ht
);
2835 local_stream
[i
] = NULL
;
2838 } else if (pollfd
[i
].revents
& POLLERR
) {
2839 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2840 if (!local_stream
[i
]->data_read
) {
2841 consumer_del_stream(local_stream
[i
], data_ht
);
2842 local_stream
[i
] = NULL
;
2845 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2846 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2847 if (!local_stream
[i
]->data_read
) {
2848 consumer_del_stream(local_stream
[i
], data_ht
);
2849 local_stream
[i
] = NULL
;
2853 if (local_stream
[i
] != NULL
) {
2854 local_stream
[i
]->data_read
= 0;
2861 DBG("polling thread exiting");
2866 * Close the write side of the pipe so epoll_wait() in
2867 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2868 * read side of the pipe. If we close them both, epoll_wait strangely does
2869 * not return and could create a endless wait period if the pipe is the
2870 * only tracked fd in the poll set. The thread will take care of closing
2873 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2878 ERR("Health error occurred in %s", __func__
);
2880 health_unregister(health_consumerd
);
2882 rcu_unregister_thread();
2887 * Close wake-up end of each stream belonging to the channel. This will
2888 * allow the poll() on the stream read-side to detect when the
2889 * write-side (application) finally closes them.
2892 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2894 struct lttng_ht
*ht
;
2895 struct lttng_consumer_stream
*stream
;
2896 struct lttng_ht_iter iter
;
2898 ht
= consumer_data
.stream_per_chan_id_ht
;
2901 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2902 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2903 ht
->match_fct
, &channel
->key
,
2904 &iter
.iter
, stream
, node_channel_id
.node
) {
2906 * Protect against teardown with mutex.
2908 pthread_mutex_lock(&stream
->lock
);
2909 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2912 switch (consumer_data
.type
) {
2913 case LTTNG_CONSUMER_KERNEL
:
2915 case LTTNG_CONSUMER32_UST
:
2916 case LTTNG_CONSUMER64_UST
:
2917 if (stream
->metadata_flag
) {
2918 /* Safe and protected by the stream lock. */
2919 lttng_ustconsumer_close_metadata(stream
->chan
);
2922 * Note: a mutex is taken internally within
2923 * liblttng-ust-ctl to protect timer wakeup_fd
2924 * use from concurrent close.
2926 lttng_ustconsumer_close_stream_wakeup(stream
);
2930 ERR("Unknown consumer_data type");
2934 pthread_mutex_unlock(&stream
->lock
);
2939 static void destroy_channel_ht(struct lttng_ht
*ht
)
2941 struct lttng_ht_iter iter
;
2942 struct lttng_consumer_channel
*channel
;
2950 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2951 ret
= lttng_ht_del(ht
, &iter
);
2956 lttng_ht_destroy(ht
);
2960 * This thread polls the channel fds to detect when they are being
2961 * closed. It closes all related streams if the channel is detected as
2962 * closed. It is currently only used as a shim layer for UST because the
2963 * consumerd needs to keep the per-stream wakeup end of pipes open for
2966 void *consumer_thread_channel_poll(void *data
)
2968 int ret
, i
, pollfd
, err
= -1;
2969 uint32_t revents
, nb_fd
;
2970 struct lttng_consumer_channel
*chan
= NULL
;
2971 struct lttng_ht_iter iter
;
2972 struct lttng_ht_node_u64
*node
;
2973 struct lttng_poll_event events
;
2974 struct lttng_consumer_local_data
*ctx
= data
;
2975 struct lttng_ht
*channel_ht
;
2977 rcu_register_thread();
2979 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2981 if (testpoint(consumerd_thread_channel
)) {
2982 goto error_testpoint
;
2985 health_code_update();
2987 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2989 /* ENOMEM at this point. Better to bail out. */
2993 DBG("Thread channel poll started");
2995 /* Size is set to 1 for the consumer_channel pipe */
2996 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2998 ERR("Poll set creation failed");
3002 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
3008 DBG("Channel main loop started");
3012 health_code_update();
3013 DBG("Channel poll wait");
3014 health_poll_entry();
3015 ret
= lttng_poll_wait(&events
, -1);
3016 DBG("Channel poll return from wait with %d fd(s)",
3017 LTTNG_POLL_GETNB(&events
));
3019 DBG("Channel event caught in thread");
3021 if (errno
== EINTR
) {
3022 ERR("Poll EINTR caught");
3025 if (LTTNG_POLL_GETNB(&events
) == 0) {
3026 err
= 0; /* All is OK */
3033 /* From here, the event is a channel wait fd */
3034 for (i
= 0; i
< nb_fd
; i
++) {
3035 health_code_update();
3037 revents
= LTTNG_POLL_GETEV(&events
, i
);
3038 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
3040 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
3041 if (revents
& LPOLLIN
) {
3042 enum consumer_channel_action action
;
3045 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
3048 ERR("Error reading channel pipe");
3050 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3055 case CONSUMER_CHANNEL_ADD
:
3056 DBG("Adding channel %d to poll set",
3059 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
3062 lttng_ht_add_unique_u64(channel_ht
,
3063 &chan
->wait_fd_node
);
3065 /* Add channel to the global poll events list */
3066 lttng_poll_add(&events
, chan
->wait_fd
,
3067 LPOLLERR
| LPOLLHUP
);
3069 case CONSUMER_CHANNEL_DEL
:
3072 * This command should never be called if the channel
3073 * has streams monitored by either the data or metadata
3074 * thread. The consumer only notify this thread with a
3075 * channel del. command if it receives a destroy
3076 * channel command from the session daemon that send it
3077 * if a command prior to the GET_CHANNEL failed.
3081 chan
= consumer_find_channel(key
);
3084 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
3087 lttng_poll_del(&events
, chan
->wait_fd
);
3088 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
3089 ret
= lttng_ht_del(channel_ht
, &iter
);
3092 switch (consumer_data
.type
) {
3093 case LTTNG_CONSUMER_KERNEL
:
3095 case LTTNG_CONSUMER32_UST
:
3096 case LTTNG_CONSUMER64_UST
:
3097 health_code_update();
3098 /* Destroy streams that might have been left in the stream list. */
3099 clean_channel_stream_list(chan
);
3102 ERR("Unknown consumer_data type");
3107 * Release our own refcount. Force channel deletion even if
3108 * streams were not initialized.
3110 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
3111 consumer_del_channel(chan
);
3116 case CONSUMER_CHANNEL_QUIT
:
3118 * Remove the pipe from the poll set and continue the loop
3119 * since their might be data to consume.
3121 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3124 ERR("Unknown action");
3127 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3128 DBG("Channel thread pipe hung up");
3130 * Remove the pipe from the poll set and continue the loop
3131 * since their might be data to consume.
3133 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3136 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3140 /* Handle other stream */
3146 uint64_t tmp_id
= (uint64_t) pollfd
;
3148 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3150 node
= lttng_ht_iter_get_node_u64(&iter
);
3153 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3156 /* Check for error event */
3157 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3158 DBG("Channel fd %d is hup|err.", pollfd
);
3160 lttng_poll_del(&events
, chan
->wait_fd
);
3161 ret
= lttng_ht_del(channel_ht
, &iter
);
3165 * This will close the wait fd for each stream associated to
3166 * this channel AND monitored by the data/metadata thread thus
3167 * will be clean by the right thread.
3169 consumer_close_channel_streams(chan
);
3171 /* Release our own refcount */
3172 if (!uatomic_sub_return(&chan
->refcount
, 1)
3173 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3174 consumer_del_channel(chan
);
3177 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3182 /* Release RCU lock for the channel looked up */
3190 lttng_poll_clean(&events
);
3192 destroy_channel_ht(channel_ht
);
3195 DBG("Channel poll thread exiting");
3198 ERR("Health error occurred in %s", __func__
);
3200 health_unregister(health_consumerd
);
3201 rcu_unregister_thread();
3205 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3206 struct pollfd
*sockpoll
, int client_socket
)
3213 ret
= lttng_consumer_poll_socket(sockpoll
);
3217 DBG("Metadata connection on client_socket");
3219 /* Blocking call, waiting for transmission */
3220 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3221 if (ctx
->consumer_metadata_socket
< 0) {
3222 WARN("On accept metadata");
3233 * This thread listens on the consumerd socket and receives the file
3234 * descriptors from the session daemon.
3236 void *consumer_thread_sessiond_poll(void *data
)
3238 int sock
= -1, client_socket
, ret
, err
= -1;
3240 * structure to poll for incoming data on communication socket avoids
3241 * making blocking sockets.
3243 struct pollfd consumer_sockpoll
[2];
3244 struct lttng_consumer_local_data
*ctx
= data
;
3246 rcu_register_thread();
3248 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3250 if (testpoint(consumerd_thread_sessiond
)) {
3251 goto error_testpoint
;
3254 health_code_update();
3256 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3257 unlink(ctx
->consumer_command_sock_path
);
3258 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3259 if (client_socket
< 0) {
3260 ERR("Cannot create command socket");
3264 ret
= lttcomm_listen_unix_sock(client_socket
);
3269 DBG("Sending ready command to lttng-sessiond");
3270 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3271 /* return < 0 on error, but == 0 is not fatal */
3273 ERR("Error sending ready command to lttng-sessiond");
3277 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3278 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3279 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3280 consumer_sockpoll
[1].fd
= client_socket
;
3281 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3283 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3291 DBG("Connection on client_socket");
3293 /* Blocking call, waiting for transmission */
3294 sock
= lttcomm_accept_unix_sock(client_socket
);
3301 * Setup metadata socket which is the second socket connection on the
3302 * command unix socket.
3304 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3313 /* This socket is not useful anymore. */
3314 ret
= close(client_socket
);
3316 PERROR("close client_socket");
3320 /* update the polling structure to poll on the established socket */
3321 consumer_sockpoll
[1].fd
= sock
;
3322 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3325 health_code_update();
3327 health_poll_entry();
3328 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3337 DBG("Incoming command on sock");
3338 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3341 * This could simply be a session daemon quitting. Don't output
3344 DBG("Communication interrupted on command socket");
3348 if (CMM_LOAD_SHARED(consumer_quit
)) {
3349 DBG("consumer_thread_receive_fds received quit from signal");
3350 err
= 0; /* All is OK */
3353 DBG("received command on sock");
3359 DBG("Consumer thread sessiond poll exiting");
3362 * Close metadata streams since the producer is the session daemon which
3365 * NOTE: for now, this only applies to the UST tracer.
3367 lttng_consumer_close_all_metadata();
3370 * when all fds have hung up, the polling thread
3373 CMM_STORE_SHARED(consumer_quit
, 1);
3376 * Notify the data poll thread to poll back again and test the
3377 * consumer_quit state that we just set so to quit gracefully.
3379 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3381 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3383 notify_health_quit_pipe(health_quit_pipe
);
3385 /* Cleaning up possibly open sockets. */
3389 PERROR("close sock sessiond poll");
3392 if (client_socket
>= 0) {
3393 ret
= close(client_socket
);
3395 PERROR("close client_socket sessiond poll");
3402 ERR("Health error occurred in %s", __func__
);
3404 health_unregister(health_consumerd
);
3406 rcu_unregister_thread();
3410 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3411 struct lttng_consumer_local_data
*ctx
)
3415 pthread_mutex_lock(&stream
->chan
->lock
);
3416 pthread_mutex_lock(&stream
->lock
);
3417 if (stream
->metadata_flag
) {
3418 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3421 switch (consumer_data
.type
) {
3422 case LTTNG_CONSUMER_KERNEL
:
3423 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3425 case LTTNG_CONSUMER32_UST
:
3426 case LTTNG_CONSUMER64_UST
:
3427 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3430 ERR("Unknown consumer_data type");
3436 if (stream
->metadata_flag
) {
3437 pthread_cond_broadcast(&stream
->metadata_rdv
);
3438 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3440 pthread_mutex_unlock(&stream
->lock
);
3441 pthread_mutex_unlock(&stream
->chan
->lock
);
3446 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3448 switch (consumer_data
.type
) {
3449 case LTTNG_CONSUMER_KERNEL
:
3450 return lttng_kconsumer_on_recv_stream(stream
);
3451 case LTTNG_CONSUMER32_UST
:
3452 case LTTNG_CONSUMER64_UST
:
3453 return lttng_ustconsumer_on_recv_stream(stream
);
3455 ERR("Unknown consumer_data type");
3462 * Allocate and set consumer data hash tables.
3464 int lttng_consumer_init(void)
3466 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3467 if (!consumer_data
.channel_ht
) {
3471 consumer_data
.channels_by_session_id_ht
=
3472 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3473 if (!consumer_data
.channels_by_session_id_ht
) {
3477 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3478 if (!consumer_data
.relayd_ht
) {
3482 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3483 if (!consumer_data
.stream_list_ht
) {
3487 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3488 if (!consumer_data
.stream_per_chan_id_ht
) {
3492 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3497 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3502 consumer_data
.chunk_registry
= lttng_trace_chunk_registry_create();
3503 if (!consumer_data
.chunk_registry
) {
3514 * Process the ADD_RELAYD command receive by a consumer.
3516 * This will create a relayd socket pair and add it to the relayd hash table.
3517 * The caller MUST acquire a RCU read side lock before calling it.
3519 void consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3520 struct lttng_consumer_local_data
*ctx
, int sock
,
3521 struct pollfd
*consumer_sockpoll
,
3522 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3523 uint64_t relayd_session_id
)
3525 int fd
= -1, ret
= -1, relayd_created
= 0;
3526 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3527 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3530 assert(relayd_sock
);
3532 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3534 /* Get relayd reference if exists. */
3535 relayd
= consumer_find_relayd(net_seq_idx
);
3536 if (relayd
== NULL
) {
3537 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3538 /* Not found. Allocate one. */
3539 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3540 if (relayd
== NULL
) {
3541 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3544 relayd
->sessiond_session_id
= sessiond_id
;
3549 * This code path MUST continue to the consumer send status message to
3550 * we can notify the session daemon and continue our work without
3551 * killing everything.
3555 * relayd key should never be found for control socket.
3557 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3560 /* First send a status message before receiving the fds. */
3561 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3563 /* Somehow, the session daemon is not responding anymore. */
3564 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3565 goto error_nosignal
;
3568 /* Poll on consumer socket. */
3569 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3571 /* Needing to exit in the middle of a command: error. */
3572 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3573 goto error_nosignal
;
3576 /* Get relayd socket from session daemon */
3577 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3578 if (ret
!= sizeof(fd
)) {
3579 fd
= -1; /* Just in case it gets set with an invalid value. */
3582 * Failing to receive FDs might indicate a major problem such as
3583 * reaching a fd limit during the receive where the kernel returns a
3584 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3585 * don't take any chances and stop everything.
3587 * XXX: Feature request #558 will fix that and avoid this possible
3588 * issue when reaching the fd limit.
3590 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3591 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3595 /* Copy socket information and received FD */
3596 switch (sock_type
) {
3597 case LTTNG_STREAM_CONTROL
:
3598 /* Copy received lttcomm socket */
3599 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3600 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3601 /* Handle create_sock error. */
3603 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3607 * Close the socket created internally by
3608 * lttcomm_create_sock, so we can replace it by the one
3609 * received from sessiond.
3611 if (close(relayd
->control_sock
.sock
.fd
)) {
3615 /* Assign new file descriptor */
3616 relayd
->control_sock
.sock
.fd
= fd
;
3617 /* Assign version values. */
3618 relayd
->control_sock
.major
= relayd_sock
->major
;
3619 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3621 relayd
->relayd_session_id
= relayd_session_id
;
3624 case LTTNG_STREAM_DATA
:
3625 /* Copy received lttcomm socket */
3626 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3627 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3628 /* Handle create_sock error. */
3630 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3634 * Close the socket created internally by
3635 * lttcomm_create_sock, so we can replace it by the one
3636 * received from sessiond.
3638 if (close(relayd
->data_sock
.sock
.fd
)) {
3642 /* Assign new file descriptor */
3643 relayd
->data_sock
.sock
.fd
= fd
;
3644 /* Assign version values. */
3645 relayd
->data_sock
.major
= relayd_sock
->major
;
3646 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3649 ERR("Unknown relayd socket type (%d)", sock_type
);
3650 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3654 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3655 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3656 relayd
->net_seq_idx
, fd
);
3658 * We gave the ownership of the fd to the relayd structure. Set the
3659 * fd to -1 so we don't call close() on it in the error path below.
3663 /* We successfully added the socket. Send status back. */
3664 ret
= consumer_send_status_msg(sock
, ret_code
);
3666 /* Somehow, the session daemon is not responding anymore. */
3667 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3668 goto error_nosignal
;
3672 * Add relayd socket pair to consumer data hashtable. If object already
3673 * exists or on error, the function gracefully returns.
3682 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3683 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3687 /* Close received socket if valid. */
3690 PERROR("close received socket");
3694 if (relayd_created
) {
3700 * Search for a relayd associated to the session id and return the reference.
3702 * A rcu read side lock MUST be acquire before calling this function and locked
3703 * until the relayd object is no longer necessary.
3705 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3707 struct lttng_ht_iter iter
;
3708 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3710 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3711 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3714 * Check by sessiond id which is unique here where the relayd session
3715 * id might not be when having multiple relayd.
3717 if (relayd
->sessiond_session_id
== id
) {
3718 /* Found the relayd. There can be only one per id. */
3730 * Check if for a given session id there is still data needed to be extract
3733 * Return 1 if data is pending or else 0 meaning ready to be read.
3735 int consumer_data_pending(uint64_t id
)
3738 struct lttng_ht_iter iter
;
3739 struct lttng_ht
*ht
;
3740 struct lttng_consumer_stream
*stream
;
3741 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3742 int (*data_pending
)(struct lttng_consumer_stream
*);
3744 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3747 pthread_mutex_lock(&consumer_data
.lock
);
3749 switch (consumer_data
.type
) {
3750 case LTTNG_CONSUMER_KERNEL
:
3751 data_pending
= lttng_kconsumer_data_pending
;
3753 case LTTNG_CONSUMER32_UST
:
3754 case LTTNG_CONSUMER64_UST
:
3755 data_pending
= lttng_ustconsumer_data_pending
;
3758 ERR("Unknown consumer data type");
3762 /* Ease our life a bit */
3763 ht
= consumer_data
.stream_list_ht
;
3765 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3766 ht
->hash_fct(&id
, lttng_ht_seed
),
3768 &iter
.iter
, stream
, node_session_id
.node
) {
3769 pthread_mutex_lock(&stream
->lock
);
3772 * A removed node from the hash table indicates that the stream has
3773 * been deleted thus having a guarantee that the buffers are closed
3774 * on the consumer side. However, data can still be transmitted
3775 * over the network so don't skip the relayd check.
3777 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3779 /* Check the stream if there is data in the buffers. */
3780 ret
= data_pending(stream
);
3782 pthread_mutex_unlock(&stream
->lock
);
3787 pthread_mutex_unlock(&stream
->lock
);
3790 relayd
= find_relayd_by_session_id(id
);
3792 unsigned int is_data_inflight
= 0;
3794 /* Send init command for data pending. */
3795 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3796 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3797 relayd
->relayd_session_id
);
3799 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3800 /* Communication error thus the relayd so no data pending. */
3801 goto data_not_pending
;
3804 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3805 ht
->hash_fct(&id
, lttng_ht_seed
),
3807 &iter
.iter
, stream
, node_session_id
.node
) {
3808 if (stream
->metadata_flag
) {
3809 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3810 stream
->relayd_stream_id
);
3812 ret
= relayd_data_pending(&relayd
->control_sock
,
3813 stream
->relayd_stream_id
,
3814 stream
->next_net_seq_num
- 1);
3818 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3820 } else if (ret
< 0) {
3821 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3822 lttng_consumer_cleanup_relayd(relayd
);
3823 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3824 goto data_not_pending
;
3828 /* Send end command for data pending. */
3829 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3830 relayd
->relayd_session_id
, &is_data_inflight
);
3831 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3833 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3834 lttng_consumer_cleanup_relayd(relayd
);
3835 goto data_not_pending
;
3837 if (is_data_inflight
) {
3843 * Finding _no_ node in the hash table and no inflight data means that the
3844 * stream(s) have been removed thus data is guaranteed to be available for
3845 * analysis from the trace files.
3849 /* Data is available to be read by a viewer. */
3850 pthread_mutex_unlock(&consumer_data
.lock
);
3855 /* Data is still being extracted from buffers. */
3856 pthread_mutex_unlock(&consumer_data
.lock
);
3862 * Send a ret code status message to the sessiond daemon.
3864 * Return the sendmsg() return value.
3866 int consumer_send_status_msg(int sock
, int ret_code
)
3868 struct lttcomm_consumer_status_msg msg
;
3870 memset(&msg
, 0, sizeof(msg
));
3871 msg
.ret_code
= ret_code
;
3873 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3877 * Send a channel status message to the sessiond daemon.
3879 * Return the sendmsg() return value.
3881 int consumer_send_status_channel(int sock
,
3882 struct lttng_consumer_channel
*channel
)
3884 struct lttcomm_consumer_status_channel msg
;
3888 memset(&msg
, 0, sizeof(msg
));
3890 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3892 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3893 msg
.key
= channel
->key
;
3894 msg
.stream_count
= channel
->streams
.count
;
3897 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3900 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3901 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3902 uint64_t max_sb_size
)
3904 unsigned long start_pos
;
3906 if (!nb_packets_per_stream
) {
3907 return consumed_pos
; /* Grab everything */
3909 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3910 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3911 if ((long) (start_pos
- consumed_pos
) < 0) {
3912 return consumed_pos
; /* Grab everything */
3918 int consumer_flush_buffer(struct lttng_consumer_stream
*stream
, int producer_active
)
3922 switch (consumer_data
.type
) {
3923 case LTTNG_CONSUMER_KERNEL
:
3924 if (producer_active
) {
3925 ret
= kernctl_buffer_flush(stream
->wait_fd
);
3927 ERR("Failed to flush kernel stream");
3931 ret
= kernctl_buffer_flush_empty(stream
->wait_fd
);
3933 ERR("Failed to flush kernel stream");
3938 case LTTNG_CONSUMER32_UST
:
3939 case LTTNG_CONSUMER64_UST
:
3940 lttng_ustconsumer_flush_buffer(stream
, producer_active
);
3943 ERR("Unknown consumer_data type");
3952 * Sample the rotate position for all the streams of a channel. If a stream
3953 * is already at the rotate position (produced == consumed), we flag it as
3954 * ready for rotation. The rotation of ready streams occurs after we have
3955 * replied to the session daemon that we have finished sampling the positions.
3956 * Must be called with RCU read-side lock held to ensure existence of channel.
3958 * Returns 0 on success, < 0 on error
3960 int lttng_consumer_rotate_channel(struct lttng_consumer_channel
*channel
,
3961 uint64_t key
, uint64_t relayd_id
, uint32_t metadata
,
3962 struct lttng_consumer_local_data
*ctx
)
3965 struct lttng_consumer_stream
*stream
;
3966 struct lttng_ht_iter iter
;
3967 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
3968 struct lttng_dynamic_array stream_rotation_positions
;
3969 uint64_t next_chunk_id
, stream_count
= 0;
3970 enum lttng_trace_chunk_status chunk_status
;
3971 const bool is_local_trace
= relayd_id
== -1ULL;
3972 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3973 bool rotating_to_new_chunk
= true;
3975 DBG("Consumer sample rotate position for channel %" PRIu64
, key
);
3977 lttng_dynamic_array_init(&stream_rotation_positions
,
3978 sizeof(struct relayd_stream_rotation_position
), NULL
);
3982 pthread_mutex_lock(&channel
->lock
);
3983 assert(channel
->trace_chunk
);
3984 chunk_status
= lttng_trace_chunk_get_id(channel
->trace_chunk
,
3986 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
3988 goto end_unlock_channel
;
3991 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3992 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
3993 ht
->match_fct
, &channel
->key
, &iter
.iter
,
3994 stream
, node_channel_id
.node
) {
3995 unsigned long produced_pos
= 0, consumed_pos
= 0;
3997 health_code_update();
4000 * Lock stream because we are about to change its state.
4002 pthread_mutex_lock(&stream
->lock
);
4004 if (stream
->trace_chunk
== stream
->chan
->trace_chunk
) {
4005 rotating_to_new_chunk
= false;
4009 * Do not flush an empty packet when rotating from a NULL trace
4010 * chunk. The stream has no means to output data, and the prior
4011 * rotation which rotated to NULL performed that side-effect already.
4013 if (stream
->trace_chunk
) {
4015 * For metadata stream, do an active flush, which does not
4016 * produce empty packets. For data streams, empty-flush;
4017 * ensures we have at least one packet in each stream per trace
4018 * chunk, even if no data was produced.
4020 ret
= consumer_flush_buffer(stream
, stream
->metadata_flag
? 1 : 0);
4022 ERR("Failed to flush stream %" PRIu64
" during channel rotation",
4024 goto end_unlock_stream
;
4028 ret
= lttng_consumer_take_snapshot(stream
);
4029 if (ret
< 0 && ret
!= -ENODATA
&& ret
!= -EAGAIN
) {
4030 ERR("Failed to sample snapshot position during channel rotation");
4031 goto end_unlock_stream
;
4034 ret
= lttng_consumer_get_produced_snapshot(stream
,
4037 ERR("Failed to sample produced position during channel rotation");
4038 goto end_unlock_stream
;
4041 ret
= lttng_consumer_get_consumed_snapshot(stream
,
4044 ERR("Failed to sample consumed position during channel rotation");
4045 goto end_unlock_stream
;
4049 * Align produced position on the start-of-packet boundary of the first
4050 * packet going into the next trace chunk.
4052 produced_pos
= ALIGN_FLOOR(produced_pos
, stream
->max_sb_size
);
4053 if (consumed_pos
== produced_pos
) {
4054 stream
->rotate_ready
= true;
4057 * The rotation position is based on the packet_seq_num of the
4058 * packet following the last packet that was consumed for this
4059 * stream, incremented by the offset between produced and
4060 * consumed positions. This rotation position is a lower bound
4061 * (inclusive) at which the next trace chunk starts. Since it
4062 * is a lower bound, it is OK if the packet_seq_num does not
4063 * correspond exactly to the same packet identified by the
4064 * consumed_pos, which can happen in overwrite mode.
4066 if (stream
->sequence_number_unavailable
) {
4068 * Rotation should never be performed on a session which
4069 * interacts with a pre-2.8 lttng-modules, which does
4070 * not implement packet sequence number.
4072 ERR("Failure to rotate stream %" PRIu64
": sequence number unavailable",
4075 goto end_unlock_stream
;
4077 stream
->rotate_position
= stream
->last_sequence_number
+ 1 +
4078 ((produced_pos
- consumed_pos
) / stream
->max_sb_size
);
4080 if (!is_local_trace
) {
4082 * The relay daemon control protocol expects a rotation
4083 * position as "the sequence number of the first packet
4084 * _after_ the current trace chunk".
4086 const struct relayd_stream_rotation_position position
= {
4087 .stream_id
= stream
->relayd_stream_id
,
4088 .rotate_at_seq_num
= stream
->rotate_position
,
4091 ret
= lttng_dynamic_array_add_element(
4092 &stream_rotation_positions
,
4095 ERR("Failed to allocate stream rotation position");
4096 goto end_unlock_stream
;
4100 pthread_mutex_unlock(&stream
->lock
);
4103 pthread_mutex_unlock(&channel
->lock
);
4105 if (is_local_trace
) {
4110 relayd
= consumer_find_relayd(relayd_id
);
4112 ERR("Failed to find relayd %" PRIu64
, relayd_id
);
4117 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4118 ret
= relayd_rotate_streams(&relayd
->control_sock
, stream_count
,
4119 rotating_to_new_chunk
? &next_chunk_id
: NULL
,
4120 (const struct relayd_stream_rotation_position
*)
4121 stream_rotation_positions
.buffer
.data
);
4122 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4124 ERR("Relayd rotate stream failed. Cleaning up relayd %" PRIu64
,
4125 relayd
->net_seq_idx
);
4126 lttng_consumer_cleanup_relayd(relayd
);
4134 pthread_mutex_unlock(&stream
->lock
);
4136 pthread_mutex_unlock(&channel
->lock
);
4139 lttng_dynamic_array_reset(&stream_rotation_positions
);
4144 * Check if a stream is ready to be rotated after extracting it.
4146 * Return 1 if it is ready for rotation, 0 if it is not, a negative value on
4147 * error. Stream lock must be held.
4149 int lttng_consumer_stream_is_rotate_ready(struct lttng_consumer_stream
*stream
)
4151 if (stream
->rotate_ready
) {
4156 * If packet seq num is unavailable, it means we are interacting
4157 * with a pre-2.8 lttng-modules which does not implement the
4158 * sequence number. Rotation should never be used by sessiond in this
4161 if (stream
->sequence_number_unavailable
) {
4162 ERR("Internal error: rotation used on stream %" PRIu64
4163 " with unavailable sequence number",
4168 if (stream
->rotate_position
== -1ULL ||
4169 stream
->last_sequence_number
== -1ULL) {
4174 * Rotate position not reached yet. The stream rotate position is
4175 * the position of the next packet belonging to the next trace chunk,
4176 * but consumerd considers rotation ready when reaching the last
4177 * packet of the current chunk, hence the "rotate_position - 1".
4179 if (stream
->last_sequence_number
>= stream
->rotate_position
- 1) {
4187 * Reset the state for a stream after a rotation occurred.
4189 void lttng_consumer_reset_stream_rotate_state(struct lttng_consumer_stream
*stream
)
4191 stream
->rotate_position
= -1ULL;
4192 stream
->rotate_ready
= false;
4196 * Perform the rotation a local stream file.
4199 int rotate_local_stream(struct lttng_consumer_local_data
*ctx
,
4200 struct lttng_consumer_stream
*stream
)
4204 DBG("Rotate local stream: stream key %" PRIu64
", channel key %" PRIu64
,
4207 stream
->tracefile_size_current
= 0;
4208 stream
->tracefile_count_current
= 0;
4210 if (stream
->out_fd
>= 0) {
4211 ret
= close(stream
->out_fd
);
4213 PERROR("Failed to close stream out_fd of channel \"%s\"",
4214 stream
->chan
->name
);
4216 stream
->out_fd
= -1;
4219 if (stream
->index_file
) {
4220 lttng_index_file_put(stream
->index_file
);
4221 stream
->index_file
= NULL
;
4224 if (!stream
->trace_chunk
) {
4228 ret
= consumer_stream_create_output_files(stream
, true);
4234 * Performs the stream rotation for the rotate session feature if needed.
4235 * It must be called with the channel and stream locks held.
4237 * Return 0 on success, a negative number of error.
4239 int lttng_consumer_rotate_stream(struct lttng_consumer_local_data
*ctx
,
4240 struct lttng_consumer_stream
*stream
)
4244 DBG("Consumer rotate stream %" PRIu64
, stream
->key
);
4247 * Update the stream's 'current' chunk to the session's (channel)
4248 * now-current chunk.
4250 lttng_trace_chunk_put(stream
->trace_chunk
);
4251 if (stream
->chan
->trace_chunk
== stream
->trace_chunk
) {
4253 * A channel can be rotated and not have a "next" chunk
4254 * to transition to. In that case, the channel's "current chunk"
4255 * has not been closed yet, but it has not been updated to
4256 * a "next" trace chunk either. Hence, the stream, like its
4257 * parent channel, becomes part of no chunk and can't output
4258 * anything until a new trace chunk is created.
4260 stream
->trace_chunk
= NULL
;
4261 } else if (stream
->chan
->trace_chunk
&&
4262 !lttng_trace_chunk_get(stream
->chan
->trace_chunk
)) {
4263 ERR("Failed to acquire a reference to channel's trace chunk during stream rotation");
4268 * Update the stream's trace chunk to its parent channel's
4269 * current trace chunk.
4271 stream
->trace_chunk
= stream
->chan
->trace_chunk
;
4274 if (stream
->net_seq_idx
== (uint64_t) -1ULL) {
4275 ret
= rotate_local_stream(ctx
, stream
);
4277 ERR("Failed to rotate stream, ret = %i", ret
);
4282 if (stream
->metadata_flag
&& stream
->trace_chunk
) {
4284 * If the stream has transitioned to a new trace
4285 * chunk, the metadata should be re-dumped to the
4288 * However, it is possible for a stream to transition to
4289 * a "no-chunk" state. This can happen if a rotation
4290 * occurs on an inactive session. In such cases, the metadata
4291 * regeneration will happen when the next trace chunk is
4294 ret
= consumer_metadata_stream_dump(stream
);
4299 lttng_consumer_reset_stream_rotate_state(stream
);
4308 * Rotate all the ready streams now.
4310 * This is especially important for low throughput streams that have already
4311 * been consumed, we cannot wait for their next packet to perform the
4313 * Need to be called with RCU read-side lock held to ensure existence of
4316 * Returns 0 on success, < 0 on error
4318 int lttng_consumer_rotate_ready_streams(struct lttng_consumer_channel
*channel
,
4319 uint64_t key
, struct lttng_consumer_local_data
*ctx
)
4322 struct lttng_consumer_stream
*stream
;
4323 struct lttng_ht_iter iter
;
4324 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
4328 DBG("Consumer rotate ready streams in channel %" PRIu64
, key
);
4330 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4331 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4332 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4333 stream
, node_channel_id
.node
) {
4334 health_code_update();
4336 pthread_mutex_lock(&stream
->chan
->lock
);
4337 pthread_mutex_lock(&stream
->lock
);
4339 if (!stream
->rotate_ready
) {
4340 pthread_mutex_unlock(&stream
->lock
);
4341 pthread_mutex_unlock(&stream
->chan
->lock
);
4344 DBG("Consumer rotate ready stream %" PRIu64
, stream
->key
);
4346 ret
= lttng_consumer_rotate_stream(ctx
, stream
);
4347 pthread_mutex_unlock(&stream
->lock
);
4348 pthread_mutex_unlock(&stream
->chan
->lock
);
4361 enum lttcomm_return_code
lttng_consumer_init_command(
4362 struct lttng_consumer_local_data
*ctx
,
4363 const lttng_uuid sessiond_uuid
)
4365 enum lttcomm_return_code ret
;
4366 char uuid_str
[UUID_STR_LEN
];
4368 if (ctx
->sessiond_uuid
.is_set
) {
4369 ret
= LTTCOMM_CONSUMERD_ALREADY_SET
;
4373 ctx
->sessiond_uuid
.is_set
= true;
4374 memcpy(ctx
->sessiond_uuid
.value
, sessiond_uuid
, sizeof(lttng_uuid
));
4375 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4376 lttng_uuid_to_str(sessiond_uuid
, uuid_str
);
4377 DBG("Received session daemon UUID: %s", uuid_str
);
4382 enum lttcomm_return_code
lttng_consumer_create_trace_chunk(
4383 const uint64_t *relayd_id
, uint64_t session_id
,
4385 time_t chunk_creation_timestamp
,
4386 const char *chunk_override_name
,
4387 const struct lttng_credentials
*credentials
,
4388 struct lttng_directory_handle
*chunk_directory_handle
)
4391 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4392 struct lttng_trace_chunk
*created_chunk
= NULL
, *published_chunk
= NULL
;
4393 enum lttng_trace_chunk_status chunk_status
;
4394 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4395 char creation_timestamp_buffer
[ISO8601_STR_LEN
];
4396 const char *relayd_id_str
= "(none)";
4397 const char *creation_timestamp_str
;
4398 struct lttng_ht_iter iter
;
4399 struct lttng_consumer_channel
*channel
;
4402 /* Only used for logging purposes. */
4403 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4404 "%" PRIu64
, *relayd_id
);
4405 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4406 relayd_id_str
= relayd_id_buffer
;
4408 relayd_id_str
= "(formatting error)";
4412 /* Local protocol error. */
4413 assert(chunk_creation_timestamp
);
4414 ret
= time_to_iso8601_str(chunk_creation_timestamp
,
4415 creation_timestamp_buffer
,
4416 sizeof(creation_timestamp_buffer
));
4417 creation_timestamp_str
= !ret
? creation_timestamp_buffer
:
4418 "(formatting error)";
4420 DBG("Consumer create trace chunk command: relay_id = %s"
4421 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4422 ", chunk_override_name = %s"
4423 ", chunk_creation_timestamp = %s",
4424 relayd_id_str
, session_id
, chunk_id
,
4425 chunk_override_name
? : "(none)",
4426 creation_timestamp_str
);
4429 * The trace chunk registry, as used by the consumer daemon, implicitly
4430 * owns the trace chunks. This is only needed in the consumer since
4431 * the consumer has no notion of a session beyond session IDs being
4432 * used to identify other objects.
4434 * The lttng_trace_chunk_registry_publish() call below provides a
4435 * reference which is not released; it implicitly becomes the session
4436 * daemon's reference to the chunk in the consumer daemon.
4438 * The lifetime of trace chunks in the consumer daemon is managed by
4439 * the session daemon through the LTTNG_CONSUMER_CREATE_TRACE_CHUNK
4440 * and LTTNG_CONSUMER_DESTROY_TRACE_CHUNK commands.
4442 created_chunk
= lttng_trace_chunk_create(chunk_id
,
4443 chunk_creation_timestamp
);
4444 if (!created_chunk
) {
4445 ERR("Failed to create trace chunk");
4446 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4450 if (chunk_override_name
) {
4451 chunk_status
= lttng_trace_chunk_override_name(created_chunk
,
4452 chunk_override_name
);
4453 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4454 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4459 if (chunk_directory_handle
) {
4460 chunk_status
= lttng_trace_chunk_set_credentials(created_chunk
,
4462 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4463 ERR("Failed to set trace chunk credentials");
4464 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4468 * The consumer daemon has no ownership of the chunk output
4471 chunk_status
= lttng_trace_chunk_set_as_user(created_chunk
,
4472 chunk_directory_handle
);
4473 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4474 ERR("Failed to set trace chunk's directory handle");
4475 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4480 published_chunk
= lttng_trace_chunk_registry_publish_chunk(
4481 consumer_data
.chunk_registry
, session_id
,
4483 lttng_trace_chunk_put(created_chunk
);
4484 created_chunk
= NULL
;
4485 if (!published_chunk
) {
4486 ERR("Failed to publish trace chunk");
4487 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4492 cds_lfht_for_each_entry_duplicate(consumer_data
.channels_by_session_id_ht
->ht
,
4493 consumer_data
.channels_by_session_id_ht
->hash_fct(
4494 &session_id
, lttng_ht_seed
),
4495 consumer_data
.channels_by_session_id_ht
->match_fct
,
4496 &session_id
, &iter
.iter
, channel
,
4497 channels_by_session_id_ht_node
.node
) {
4498 ret
= lttng_consumer_channel_set_trace_chunk(channel
,
4502 * Roll-back the creation of this chunk.
4504 * This is important since the session daemon will
4505 * assume that the creation of this chunk failed and
4506 * will never ask for it to be closed, resulting
4507 * in a leak and an inconsistent state for some
4510 enum lttcomm_return_code close_ret
;
4511 char path
[LTTNG_PATH_MAX
];
4513 DBG("Failed to set new trace chunk on existing channels, rolling back");
4514 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4515 session_id
, chunk_id
,
4516 chunk_creation_timestamp
, NULL
,
4518 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4519 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4520 session_id
, chunk_id
);
4523 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4529 struct consumer_relayd_sock_pair
*relayd
;
4531 relayd
= consumer_find_relayd(*relayd_id
);
4533 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4534 ret
= relayd_create_trace_chunk(
4535 &relayd
->control_sock
, published_chunk
);
4536 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4538 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
, *relayd_id
);
4541 if (!relayd
|| ret
) {
4542 enum lttcomm_return_code close_ret
;
4543 char path
[LTTNG_PATH_MAX
];
4545 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4548 chunk_creation_timestamp
,
4550 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4551 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4556 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4563 /* Release the reference returned by the "publish" operation. */
4564 lttng_trace_chunk_put(published_chunk
);
4565 lttng_trace_chunk_put(created_chunk
);
4569 enum lttcomm_return_code
lttng_consumer_close_trace_chunk(
4570 const uint64_t *relayd_id
, uint64_t session_id
,
4571 uint64_t chunk_id
, time_t chunk_close_timestamp
,
4572 const enum lttng_trace_chunk_command_type
*close_command
,
4575 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4576 struct lttng_trace_chunk
*chunk
;
4577 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4578 const char *relayd_id_str
= "(none)";
4579 const char *close_command_name
= "none";
4580 struct lttng_ht_iter iter
;
4581 struct lttng_consumer_channel
*channel
;
4582 enum lttng_trace_chunk_status chunk_status
;
4587 /* Only used for logging purposes. */
4588 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4589 "%" PRIu64
, *relayd_id
);
4590 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4591 relayd_id_str
= relayd_id_buffer
;
4593 relayd_id_str
= "(formatting error)";
4596 if (close_command
) {
4597 close_command_name
= lttng_trace_chunk_command_type_get_name(
4601 DBG("Consumer close trace chunk command: relayd_id = %s"
4602 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4603 ", close command = %s",
4604 relayd_id_str
, session_id
, chunk_id
,
4605 close_command_name
);
4607 chunk
= lttng_trace_chunk_registry_find_chunk(
4608 consumer_data
.chunk_registry
, session_id
, chunk_id
);
4610 ERR("Failed to find chunk: session_id = %" PRIu64
4611 ", chunk_id = %" PRIu64
,
4612 session_id
, chunk_id
);
4613 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4617 chunk_status
= lttng_trace_chunk_set_close_timestamp(chunk
,
4618 chunk_close_timestamp
);
4619 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4620 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4624 if (close_command
) {
4625 chunk_status
= lttng_trace_chunk_set_close_command(
4626 chunk
, *close_command
);
4627 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4628 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4634 * chunk is now invalid to access as we no longer hold a reference to
4635 * it; it is only kept around to compare it (by address) to the
4636 * current chunk found in the session's channels.
4639 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
,
4640 channel
, node
.node
) {
4644 * Only change the channel's chunk to NULL if it still
4645 * references the chunk being closed. The channel may
4646 * reference a newer channel in the case of a session
4647 * rotation. When a session rotation occurs, the "next"
4648 * chunk is created before the "current" chunk is closed.
4650 if (channel
->trace_chunk
!= chunk
) {
4653 ret
= lttng_consumer_channel_set_trace_chunk(channel
, NULL
);
4656 * Attempt to close the chunk on as many channels as
4659 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4665 struct consumer_relayd_sock_pair
*relayd
;
4667 relayd
= consumer_find_relayd(*relayd_id
);
4669 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4670 ret
= relayd_close_trace_chunk(
4671 &relayd
->control_sock
, chunk
,
4673 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4675 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
,
4679 if (!relayd
|| ret
) {
4680 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4688 * Release the reference returned by the "find" operation and
4689 * the session daemon's implicit reference to the chunk.
4691 lttng_trace_chunk_put(chunk
);
4692 lttng_trace_chunk_put(chunk
);
4697 enum lttcomm_return_code
lttng_consumer_trace_chunk_exists(
4698 const uint64_t *relayd_id
, uint64_t session_id
,
4702 enum lttcomm_return_code ret_code
;
4703 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4704 const char *relayd_id_str
= "(none)";
4705 const bool is_local_trace
= !relayd_id
;
4706 struct consumer_relayd_sock_pair
*relayd
= NULL
;
4707 bool chunk_exists_local
, chunk_exists_remote
;
4712 /* Only used for logging purposes. */
4713 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4714 "%" PRIu64
, *relayd_id
);
4715 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4716 relayd_id_str
= relayd_id_buffer
;
4718 relayd_id_str
= "(formatting error)";
4722 DBG("Consumer trace chunk exists command: relayd_id = %s"
4723 ", chunk_id = %" PRIu64
, relayd_id_str
,
4725 ret
= lttng_trace_chunk_registry_chunk_exists(
4726 consumer_data
.chunk_registry
, session_id
,
4727 chunk_id
, &chunk_exists_local
);
4729 /* Internal error. */
4730 ERR("Failed to query the existence of a trace chunk");
4731 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
4734 DBG("Trace chunk %s locally",
4735 chunk_exists_local
? "exists" : "does not exist");
4736 if (chunk_exists_local
) {
4737 ret_code
= LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_LOCAL
;
4739 } else if (is_local_trace
) {
4740 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4745 relayd
= consumer_find_relayd(*relayd_id
);
4747 ERR("Failed to find relayd %" PRIu64
, *relayd_id
);
4748 ret_code
= LTTCOMM_CONSUMERD_INVALID_PARAMETERS
;
4749 goto end_rcu_unlock
;
4751 DBG("Looking up existence of trace chunk on relay daemon");
4752 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4753 ret
= relayd_trace_chunk_exists(&relayd
->control_sock
, chunk_id
,
4754 &chunk_exists_remote
);
4755 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4757 ERR("Failed to look-up the existence of trace chunk on relay daemon");
4758 ret_code
= LTTCOMM_CONSUMERD_RELAYD_FAIL
;
4759 goto end_rcu_unlock
;
4762 ret_code
= chunk_exists_remote
?
4763 LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_REMOTE
:
4764 LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4765 DBG("Trace chunk %s on relay daemon",
4766 chunk_exists_remote
? "exists" : "does not exist");