2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 #include <sys/socket.h>
28 #include <sys/types.h>
33 #include <bin/lttng-consumerd/health-consumerd.h>
34 #include <common/common.h>
35 #include <common/utils.h>
36 #include <common/time.h>
37 #include <common/compat/poll.h>
38 #include <common/compat/endian.h>
39 #include <common/index/index.h>
40 #include <common/kernel-ctl/kernel-ctl.h>
41 #include <common/sessiond-comm/relayd.h>
42 #include <common/sessiond-comm/sessiond-comm.h>
43 #include <common/kernel-consumer/kernel-consumer.h>
44 #include <common/relayd/relayd.h>
45 #include <common/ust-consumer/ust-consumer.h>
46 #include <common/consumer/consumer-timer.h>
47 #include <common/consumer/consumer.h>
48 #include <common/consumer/consumer-stream.h>
49 #include <common/consumer/consumer-testpoint.h>
50 #include <common/align.h>
51 #include <common/consumer/consumer-metadata-cache.h>
52 #include <common/trace-chunk.h>
53 #include <common/trace-chunk-registry.h>
54 #include <common/string-utils/format.h>
55 #include <common/dynamic-array.h>
57 struct lttng_consumer_global_data consumer_data
= {
60 .type
= LTTNG_CONSUMER_UNKNOWN
,
63 enum consumer_channel_action
{
66 CONSUMER_CHANNEL_QUIT
,
69 struct consumer_channel_msg
{
70 enum consumer_channel_action action
;
71 struct lttng_consumer_channel
*chan
; /* add */
72 uint64_t key
; /* del */
75 /* Flag used to temporarily pause data consumption from testpoints. */
76 int data_consumption_paused
;
79 * Flag to inform the polling thread to quit when all fd hung up. Updated by
80 * the consumer_thread_receive_fds when it notices that all fds has hung up.
81 * Also updated by the signal handler (consumer_should_exit()). Read by the
87 * Global hash table containing respectively metadata and data streams. The
88 * stream element in this ht should only be updated by the metadata poll thread
89 * for the metadata and the data poll thread for the data.
91 static struct lttng_ht
*metadata_ht
;
92 static struct lttng_ht
*data_ht
;
95 * Notify a thread lttng pipe to poll back again. This usually means that some
96 * global state has changed so we just send back the thread in a poll wait
99 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
101 struct lttng_consumer_stream
*null_stream
= NULL
;
105 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
108 static void notify_health_quit_pipe(int *pipe
)
112 ret
= lttng_write(pipe
[1], "4", 1);
114 PERROR("write consumer health quit");
118 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
119 struct lttng_consumer_channel
*chan
,
121 enum consumer_channel_action action
)
123 struct consumer_channel_msg msg
;
126 memset(&msg
, 0, sizeof(msg
));
131 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
132 if (ret
< sizeof(msg
)) {
133 PERROR("notify_channel_pipe write error");
137 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
140 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
143 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
144 struct lttng_consumer_channel
**chan
,
146 enum consumer_channel_action
*action
)
148 struct consumer_channel_msg msg
;
151 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
152 if (ret
< sizeof(msg
)) {
156 *action
= msg
.action
;
164 * Cleanup the stream list of a channel. Those streams are not yet globally
167 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
169 struct lttng_consumer_stream
*stream
, *stmp
;
173 /* Delete streams that might have been left in the stream list. */
174 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
176 cds_list_del(&stream
->send_node
);
178 * Once a stream is added to this list, the buffers were created so we
179 * have a guarantee that this call will succeed. Setting the monitor
180 * mode to 0 so we don't lock nor try to delete the stream from the
184 consumer_stream_destroy(stream
, NULL
);
189 * Find a stream. The consumer_data.lock must be locked during this
192 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
195 struct lttng_ht_iter iter
;
196 struct lttng_ht_node_u64
*node
;
197 struct lttng_consumer_stream
*stream
= NULL
;
201 /* -1ULL keys are lookup failures */
202 if (key
== (uint64_t) -1ULL) {
208 lttng_ht_lookup(ht
, &key
, &iter
);
209 node
= lttng_ht_iter_get_node_u64(&iter
);
211 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
219 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
221 struct lttng_consumer_stream
*stream
;
224 stream
= find_stream(key
, ht
);
226 stream
->key
= (uint64_t) -1ULL;
228 * We don't want the lookup to match, but we still need
229 * to iterate on this stream when iterating over the hash table. Just
230 * change the node key.
232 stream
->node
.key
= (uint64_t) -1ULL;
238 * Return a channel object for the given key.
240 * RCU read side lock MUST be acquired before calling this function and
241 * protects the channel ptr.
243 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
245 struct lttng_ht_iter iter
;
246 struct lttng_ht_node_u64
*node
;
247 struct lttng_consumer_channel
*channel
= NULL
;
249 /* -1ULL keys are lookup failures */
250 if (key
== (uint64_t) -1ULL) {
254 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
255 node
= lttng_ht_iter_get_node_u64(&iter
);
257 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
264 * There is a possibility that the consumer does not have enough time between
265 * the close of the channel on the session daemon and the cleanup in here thus
266 * once we have a channel add with an existing key, we know for sure that this
267 * channel will eventually get cleaned up by all streams being closed.
269 * This function just nullifies the already existing channel key.
271 static void steal_channel_key(uint64_t key
)
273 struct lttng_consumer_channel
*channel
;
276 channel
= consumer_find_channel(key
);
278 channel
->key
= (uint64_t) -1ULL;
280 * We don't want the lookup to match, but we still need to iterate on
281 * this channel when iterating over the hash table. Just change the
284 channel
->node
.key
= (uint64_t) -1ULL;
289 static void free_channel_rcu(struct rcu_head
*head
)
291 struct lttng_ht_node_u64
*node
=
292 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
293 struct lttng_consumer_channel
*channel
=
294 caa_container_of(node
, struct lttng_consumer_channel
, node
);
296 switch (consumer_data
.type
) {
297 case LTTNG_CONSUMER_KERNEL
:
299 case LTTNG_CONSUMER32_UST
:
300 case LTTNG_CONSUMER64_UST
:
301 lttng_ustconsumer_free_channel(channel
);
304 ERR("Unknown consumer_data type");
311 * RCU protected relayd socket pair free.
313 static void free_relayd_rcu(struct rcu_head
*head
)
315 struct lttng_ht_node_u64
*node
=
316 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
317 struct consumer_relayd_sock_pair
*relayd
=
318 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
321 * Close all sockets. This is done in the call RCU since we don't want the
322 * socket fds to be reassigned thus potentially creating bad state of the
325 * We do not have to lock the control socket mutex here since at this stage
326 * there is no one referencing to this relayd object.
328 (void) relayd_close(&relayd
->control_sock
);
329 (void) relayd_close(&relayd
->data_sock
);
331 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
336 * Destroy and free relayd socket pair object.
338 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
341 struct lttng_ht_iter iter
;
343 if (relayd
== NULL
) {
347 DBG("Consumer destroy and close relayd socket pair");
349 iter
.iter
.node
= &relayd
->node
.node
;
350 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
352 /* We assume the relayd is being or is destroyed */
356 /* RCU free() call */
357 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
361 * Remove a channel from the global list protected by a mutex. This function is
362 * also responsible for freeing its data structures.
364 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
366 struct lttng_ht_iter iter
;
368 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
370 pthread_mutex_lock(&consumer_data
.lock
);
371 pthread_mutex_lock(&channel
->lock
);
373 /* Destroy streams that might have been left in the stream list. */
374 clean_channel_stream_list(channel
);
376 if (channel
->live_timer_enabled
== 1) {
377 consumer_timer_live_stop(channel
);
379 if (channel
->monitor_timer_enabled
== 1) {
380 consumer_timer_monitor_stop(channel
);
383 switch (consumer_data
.type
) {
384 case LTTNG_CONSUMER_KERNEL
:
386 case LTTNG_CONSUMER32_UST
:
387 case LTTNG_CONSUMER64_UST
:
388 lttng_ustconsumer_del_channel(channel
);
391 ERR("Unknown consumer_data type");
396 lttng_trace_chunk_put(channel
->trace_chunk
);
397 channel
->trace_chunk
= NULL
;
399 if (channel
->is_published
) {
403 iter
.iter
.node
= &channel
->node
.node
;
404 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
407 iter
.iter
.node
= &channel
->channels_by_session_id_ht_node
.node
;
408 ret
= lttng_ht_del(consumer_data
.channels_by_session_id_ht
,
414 channel
->is_deleted
= true;
415 call_rcu(&channel
->node
.head
, free_channel_rcu
);
417 pthread_mutex_unlock(&channel
->lock
);
418 pthread_mutex_unlock(&consumer_data
.lock
);
422 * Iterate over the relayd hash table and destroy each element. Finally,
423 * destroy the whole hash table.
425 static void cleanup_relayd_ht(void)
427 struct lttng_ht_iter iter
;
428 struct consumer_relayd_sock_pair
*relayd
;
432 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
434 consumer_destroy_relayd(relayd
);
439 lttng_ht_destroy(consumer_data
.relayd_ht
);
443 * Update the end point status of all streams having the given network sequence
444 * index (relayd index).
446 * It's atomically set without having the stream mutex locked which is fine
447 * because we handle the write/read race with a pipe wakeup for each thread.
449 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
450 enum consumer_endpoint_status status
)
452 struct lttng_ht_iter iter
;
453 struct lttng_consumer_stream
*stream
;
455 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
459 /* Let's begin with metadata */
460 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
461 if (stream
->net_seq_idx
== net_seq_idx
) {
462 uatomic_set(&stream
->endpoint_status
, status
);
463 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
467 /* Follow up by the data streams */
468 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
469 if (stream
->net_seq_idx
== net_seq_idx
) {
470 uatomic_set(&stream
->endpoint_status
, status
);
471 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
478 * Cleanup a relayd object by flagging every associated streams for deletion,
479 * destroying the object meaning removing it from the relayd hash table,
480 * closing the sockets and freeing the memory in a RCU call.
482 * If a local data context is available, notify the threads that the streams'
483 * state have changed.
485 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
491 DBG("Cleaning up relayd object ID %"PRIu64
, relayd
->net_seq_idx
);
493 /* Save the net sequence index before destroying the object */
494 netidx
= relayd
->net_seq_idx
;
497 * Delete the relayd from the relayd hash table, close the sockets and free
498 * the object in a RCU call.
500 consumer_destroy_relayd(relayd
);
502 /* Set inactive endpoint to all streams */
503 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
506 * With a local data context, notify the threads that the streams' state
507 * have changed. The write() action on the pipe acts as an "implicit"
508 * memory barrier ordering the updates of the end point status from the
509 * read of this status which happens AFTER receiving this notify.
511 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
512 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
516 * Flag a relayd socket pair for destruction. Destroy it if the refcount
519 * RCU read side lock MUST be aquired before calling this function.
521 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
525 /* Set destroy flag for this object */
526 uatomic_set(&relayd
->destroy_flag
, 1);
528 /* Destroy the relayd if refcount is 0 */
529 if (uatomic_read(&relayd
->refcount
) == 0) {
530 consumer_destroy_relayd(relayd
);
535 * Completly destroy stream from every visiable data structure and the given
538 * One this call returns, the stream object is not longer usable nor visible.
540 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
543 consumer_stream_destroy(stream
, ht
);
547 * XXX naming of del vs destroy is all mixed up.
549 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
551 consumer_stream_destroy(stream
, data_ht
);
554 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
556 consumer_stream_destroy(stream
, metadata_ht
);
559 void consumer_stream_update_channel_attributes(
560 struct lttng_consumer_stream
*stream
,
561 struct lttng_consumer_channel
*channel
)
563 stream
->channel_read_only_attributes
.tracefile_size
=
564 channel
->tracefile_size
;
567 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
569 const char *channel_name
,
572 struct lttng_trace_chunk
*trace_chunk
,
575 enum consumer_channel_type type
,
576 unsigned int monitor
)
579 struct lttng_consumer_stream
*stream
;
581 stream
= zmalloc(sizeof(*stream
));
582 if (stream
== NULL
) {
583 PERROR("malloc struct lttng_consumer_stream");
588 if (trace_chunk
&& !lttng_trace_chunk_get(trace_chunk
)) {
589 ERR("Failed to acquire trace chunk reference during the creation of a stream");
595 stream
->key
= stream_key
;
596 stream
->trace_chunk
= trace_chunk
;
598 stream
->out_fd_offset
= 0;
599 stream
->output_written
= 0;
600 stream
->net_seq_idx
= relayd_id
;
601 stream
->session_id
= session_id
;
602 stream
->monitor
= monitor
;
603 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
604 stream
->index_file
= NULL
;
605 stream
->last_sequence_number
= -1ULL;
606 pthread_mutex_init(&stream
->lock
, NULL
);
607 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
609 /* If channel is the metadata, flag this stream as metadata. */
610 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
611 stream
->metadata_flag
= 1;
612 /* Metadata is flat out. */
613 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
614 /* Live rendez-vous point. */
615 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
616 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
618 /* Format stream name to <channel_name>_<cpu_number> */
619 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
622 PERROR("snprintf stream name");
627 /* Key is always the wait_fd for streams. */
628 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
630 /* Init node per channel id key */
631 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
633 /* Init session id node with the stream session id */
634 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
636 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
637 " relayd_id %" PRIu64
", session_id %" PRIu64
,
638 stream
->name
, stream
->key
, channel_key
,
639 stream
->net_seq_idx
, stream
->session_id
);
646 lttng_trace_chunk_put(stream
->trace_chunk
);
656 * Add a stream to the global list protected by a mutex.
658 void consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
660 struct lttng_ht
*ht
= data_ht
;
665 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
667 pthread_mutex_lock(&consumer_data
.lock
);
668 pthread_mutex_lock(&stream
->chan
->lock
);
669 pthread_mutex_lock(&stream
->chan
->timer_lock
);
670 pthread_mutex_lock(&stream
->lock
);
673 /* Steal stream identifier to avoid having streams with the same key */
674 steal_stream_key(stream
->key
, ht
);
676 lttng_ht_add_unique_u64(ht
, &stream
->node
);
678 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
679 &stream
->node_channel_id
);
682 * Add stream to the stream_list_ht of the consumer data. No need to steal
683 * the key since the HT does not use it and we allow to add redundant keys
686 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
689 * When nb_init_stream_left reaches 0, we don't need to trigger any action
690 * in terms of destroying the associated channel, because the action that
691 * causes the count to become 0 also causes a stream to be added. The
692 * channel deletion will thus be triggered by the following removal of this
695 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
696 /* Increment refcount before decrementing nb_init_stream_left */
698 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
701 /* Update consumer data once the node is inserted. */
702 consumer_data
.stream_count
++;
703 consumer_data
.need_update
= 1;
706 pthread_mutex_unlock(&stream
->lock
);
707 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
708 pthread_mutex_unlock(&stream
->chan
->lock
);
709 pthread_mutex_unlock(&consumer_data
.lock
);
712 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
714 consumer_del_stream(stream
, data_ht
);
718 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
719 * be acquired before calling this.
721 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
724 struct lttng_ht_node_u64
*node
;
725 struct lttng_ht_iter iter
;
729 lttng_ht_lookup(consumer_data
.relayd_ht
,
730 &relayd
->net_seq_idx
, &iter
);
731 node
= lttng_ht_iter_get_node_u64(&iter
);
735 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
742 * Allocate and return a consumer relayd socket.
744 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
745 uint64_t net_seq_idx
)
747 struct consumer_relayd_sock_pair
*obj
= NULL
;
749 /* net sequence index of -1 is a failure */
750 if (net_seq_idx
== (uint64_t) -1ULL) {
754 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
756 PERROR("zmalloc relayd sock");
760 obj
->net_seq_idx
= net_seq_idx
;
762 obj
->destroy_flag
= 0;
763 obj
->control_sock
.sock
.fd
= -1;
764 obj
->data_sock
.sock
.fd
= -1;
765 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
766 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
773 * Find a relayd socket pair in the global consumer data.
775 * Return the object if found else NULL.
776 * RCU read-side lock must be held across this call and while using the
779 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
781 struct lttng_ht_iter iter
;
782 struct lttng_ht_node_u64
*node
;
783 struct consumer_relayd_sock_pair
*relayd
= NULL
;
785 /* Negative keys are lookup failures */
786 if (key
== (uint64_t) -1ULL) {
790 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
792 node
= lttng_ht_iter_get_node_u64(&iter
);
794 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
802 * Find a relayd and send the stream
804 * Returns 0 on success, < 0 on error
806 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
810 struct consumer_relayd_sock_pair
*relayd
;
813 assert(stream
->net_seq_idx
!= -1ULL);
816 /* The stream is not metadata. Get relayd reference if exists. */
818 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
819 if (relayd
!= NULL
) {
820 /* Add stream on the relayd */
821 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
822 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
823 path
, &stream
->relayd_stream_id
,
824 stream
->chan
->tracefile_size
,
825 stream
->chan
->tracefile_count
,
826 stream
->trace_chunk
);
827 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
829 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
830 lttng_consumer_cleanup_relayd(relayd
);
834 uatomic_inc(&relayd
->refcount
);
835 stream
->sent_to_relayd
= 1;
837 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
838 stream
->key
, stream
->net_seq_idx
);
843 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
844 stream
->name
, stream
->key
, stream
->net_seq_idx
);
852 * Find a relayd and send the streams sent message
854 * Returns 0 on success, < 0 on error
856 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
859 struct consumer_relayd_sock_pair
*relayd
;
861 assert(net_seq_idx
!= -1ULL);
863 /* The stream is not metadata. Get relayd reference if exists. */
865 relayd
= consumer_find_relayd(net_seq_idx
);
866 if (relayd
!= NULL
) {
867 /* Add stream on the relayd */
868 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
869 ret
= relayd_streams_sent(&relayd
->control_sock
);
870 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
872 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
873 lttng_consumer_cleanup_relayd(relayd
);
877 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
884 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
892 * Find a relayd and close the stream
894 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
896 struct consumer_relayd_sock_pair
*relayd
;
898 /* The stream is not metadata. Get relayd reference if exists. */
900 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
902 consumer_stream_relayd_close(stream
, relayd
);
908 * Handle stream for relayd transmission if the stream applies for network
909 * streaming where the net sequence index is set.
911 * Return destination file descriptor or negative value on error.
913 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
914 size_t data_size
, unsigned long padding
,
915 struct consumer_relayd_sock_pair
*relayd
)
918 struct lttcomm_relayd_data_hdr data_hdr
;
924 /* Reset data header */
925 memset(&data_hdr
, 0, sizeof(data_hdr
));
927 if (stream
->metadata_flag
) {
928 /* Caller MUST acquire the relayd control socket lock */
929 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
934 /* Metadata are always sent on the control socket. */
935 outfd
= relayd
->control_sock
.sock
.fd
;
937 /* Set header with stream information */
938 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
939 data_hdr
.data_size
= htobe32(data_size
);
940 data_hdr
.padding_size
= htobe32(padding
);
943 * Note that net_seq_num below is assigned with the *current* value of
944 * next_net_seq_num and only after that the next_net_seq_num will be
945 * increment. This is why when issuing a command on the relayd using
946 * this next value, 1 should always be substracted in order to compare
947 * the last seen sequence number on the relayd side to the last sent.
949 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
950 /* Other fields are zeroed previously */
952 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
958 ++stream
->next_net_seq_num
;
960 /* Set to go on data socket */
961 outfd
= relayd
->data_sock
.sock
.fd
;
969 * Trigger a dump of the metadata content. Following/during the succesful
970 * completion of this call, the metadata poll thread will start receiving
971 * metadata packets to consume.
973 * The caller must hold the channel and stream locks.
976 int consumer_metadata_stream_dump(struct lttng_consumer_stream
*stream
)
980 ASSERT_LOCKED(stream
->chan
->lock
);
981 ASSERT_LOCKED(stream
->lock
);
982 assert(stream
->metadata_flag
);
983 assert(stream
->chan
->trace_chunk
);
985 switch (consumer_data
.type
) {
986 case LTTNG_CONSUMER_KERNEL
:
988 * Reset the position of what has been read from the
989 * metadata cache to 0 so we can dump it again.
991 ret
= kernctl_metadata_cache_dump(stream
->wait_fd
);
993 case LTTNG_CONSUMER32_UST
:
994 case LTTNG_CONSUMER64_UST
:
996 * Reset the position pushed from the metadata cache so it
997 * will write from the beginning on the next push.
999 stream
->ust_metadata_pushed
= 0;
1000 ret
= consumer_metadata_wakeup_pipe(stream
->chan
);
1003 ERR("Unknown consumer_data type");
1007 ERR("Failed to dump the metadata cache");
1013 int lttng_consumer_channel_set_trace_chunk(
1014 struct lttng_consumer_channel
*channel
,
1015 struct lttng_trace_chunk
*new_trace_chunk
)
1018 const bool is_local_trace
= channel
->relayd_id
== -1ULL;
1019 bool update_stream_trace_chunk
;
1020 struct cds_lfht_iter iter
;
1021 struct lttng_consumer_stream
*stream
;
1022 unsigned long channel_hash
;
1024 pthread_mutex_lock(&channel
->lock
);
1025 if (channel
->is_deleted
) {
1027 * The channel has been logically deleted and should no longer
1028 * be used. It has released its reference to its current trace
1029 * chunk and should not acquire a new one.
1031 * Return success as there is nothing for the caller to do.
1036 * A stream can transition to a state where it and its channel
1037 * no longer belong to a trace chunk. For instance, this happens when
1038 * a session is rotated while it is inactive. After the rotation
1039 * of an inactive session completes, the channel and its streams no
1040 * longer belong to a trace chunk.
1042 * However, if a session is stopped, rotated, and started again,
1043 * the session daemon will create a new chunk and send it to its peers.
1044 * In that case, the streams' transition to a new chunk can be performed
1047 * This trace chunk transition could also be performed lazily when
1048 * a buffer is consumed. However, creating the files here allows the
1049 * consumer daemon to report any creation error to the session daemon
1050 * and cause the start of the tracing session to fail.
1052 update_stream_trace_chunk
= !channel
->trace_chunk
&& new_trace_chunk
;
1055 * The acquisition of the reference cannot fail (barring
1056 * a severe internal error) since a reference to the published
1057 * chunk is already held by the caller.
1059 if (new_trace_chunk
) {
1060 const bool acquired_reference
= lttng_trace_chunk_get(
1063 assert(acquired_reference
);
1066 lttng_trace_chunk_put(channel
->trace_chunk
);
1067 channel
->trace_chunk
= new_trace_chunk
;
1068 if (!is_local_trace
|| !new_trace_chunk
) {
1073 if (!update_stream_trace_chunk
) {
1077 channel_hash
= consumer_data
.stream_per_chan_id_ht
->hash_fct(
1078 &channel
->key
, lttng_ht_seed
);
1080 cds_lfht_for_each_entry_duplicate(consumer_data
.stream_per_chan_id_ht
->ht
,
1082 consumer_data
.stream_per_chan_id_ht
->match_fct
,
1083 &channel
->key
, &iter
, stream
, node_channel_id
.node
) {
1084 bool acquired_reference
, should_regenerate_metadata
= false;
1086 acquired_reference
= lttng_trace_chunk_get(channel
->trace_chunk
);
1087 assert(acquired_reference
);
1089 pthread_mutex_lock(&stream
->lock
);
1092 * On a transition from "no-chunk" to a new chunk, a metadata
1093 * stream's content must be entirely dumped. This must occcur
1094 * _after_ the creation of the metadata stream's output files
1095 * as the consumption thread (not necessarily the one executing
1096 * this) may start to consume during the call to
1097 * consumer_metadata_stream_dump().
1099 should_regenerate_metadata
=
1100 stream
->metadata_flag
&&
1101 !stream
->trace_chunk
&& channel
->trace_chunk
;
1102 stream
->trace_chunk
= channel
->trace_chunk
;
1103 ret
= consumer_stream_create_output_files(stream
, true);
1105 pthread_mutex_unlock(&stream
->lock
);
1106 goto end_rcu_unlock
;
1108 if (should_regenerate_metadata
) {
1109 ret
= consumer_metadata_stream_dump(stream
);
1111 pthread_mutex_unlock(&stream
->lock
);
1113 goto end_rcu_unlock
;
1119 pthread_mutex_unlock(&channel
->lock
);
1124 * Allocate and return a new lttng_consumer_channel object using the given key
1125 * to initialize the hash table node.
1127 * On error, return NULL.
1129 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
1130 uint64_t session_id
,
1131 const uint64_t *chunk_id
,
1132 const char *pathname
,
1135 enum lttng_event_output output
,
1136 uint64_t tracefile_size
,
1137 uint64_t tracefile_count
,
1138 uint64_t session_id_per_pid
,
1139 unsigned int monitor
,
1140 unsigned int live_timer_interval
,
1141 const char *root_shm_path
,
1142 const char *shm_path
)
1144 struct lttng_consumer_channel
*channel
= NULL
;
1145 struct lttng_trace_chunk
*trace_chunk
= NULL
;
1148 trace_chunk
= lttng_trace_chunk_registry_find_chunk(
1149 consumer_data
.chunk_registry
, session_id
,
1152 ERR("Failed to find trace chunk reference during creation of channel");
1157 channel
= zmalloc(sizeof(*channel
));
1158 if (channel
== NULL
) {
1159 PERROR("malloc struct lttng_consumer_channel");
1164 channel
->refcount
= 0;
1165 channel
->session_id
= session_id
;
1166 channel
->session_id_per_pid
= session_id_per_pid
;
1167 channel
->relayd_id
= relayd_id
;
1168 channel
->tracefile_size
= tracefile_size
;
1169 channel
->tracefile_count
= tracefile_count
;
1170 channel
->monitor
= monitor
;
1171 channel
->live_timer_interval
= live_timer_interval
;
1172 pthread_mutex_init(&channel
->lock
, NULL
);
1173 pthread_mutex_init(&channel
->timer_lock
, NULL
);
1176 case LTTNG_EVENT_SPLICE
:
1177 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
1179 case LTTNG_EVENT_MMAP
:
1180 channel
->output
= CONSUMER_CHANNEL_MMAP
;
1190 * In monitor mode, the streams associated with the channel will be put in
1191 * a special list ONLY owned by this channel. So, the refcount is set to 1
1192 * here meaning that the channel itself has streams that are referenced.
1194 * On a channel deletion, once the channel is no longer visible, the
1195 * refcount is decremented and checked for a zero value to delete it. With
1196 * streams in no monitor mode, it will now be safe to destroy the channel.
1198 if (!channel
->monitor
) {
1199 channel
->refcount
= 1;
1202 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1203 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1205 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1206 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1208 if (root_shm_path
) {
1209 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1210 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1213 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1214 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1217 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1218 lttng_ht_node_init_u64(&channel
->channels_by_session_id_ht_node
,
1219 channel
->session_id
);
1221 channel
->wait_fd
= -1;
1222 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1225 int ret
= lttng_consumer_channel_set_trace_chunk(channel
,
1232 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1235 lttng_trace_chunk_put(trace_chunk
);
1238 consumer_del_channel(channel
);
1244 * Add a channel to the global list protected by a mutex.
1246 * Always return 0 indicating success.
1248 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1249 struct lttng_consumer_local_data
*ctx
)
1251 pthread_mutex_lock(&consumer_data
.lock
);
1252 pthread_mutex_lock(&channel
->lock
);
1253 pthread_mutex_lock(&channel
->timer_lock
);
1256 * This gives us a guarantee that the channel we are about to add to the
1257 * channel hash table will be unique. See this function comment on the why
1258 * we need to steel the channel key at this stage.
1260 steal_channel_key(channel
->key
);
1263 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1264 lttng_ht_add_u64(consumer_data
.channels_by_session_id_ht
,
1265 &channel
->channels_by_session_id_ht_node
);
1267 channel
->is_published
= true;
1269 pthread_mutex_unlock(&channel
->timer_lock
);
1270 pthread_mutex_unlock(&channel
->lock
);
1271 pthread_mutex_unlock(&consumer_data
.lock
);
1273 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1274 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1281 * Allocate the pollfd structure and the local view of the out fds to avoid
1282 * doing a lookup in the linked list and concurrency issues when writing is
1283 * needed. Called with consumer_data.lock held.
1285 * Returns the number of fds in the structures.
1287 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1288 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1289 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1292 struct lttng_ht_iter iter
;
1293 struct lttng_consumer_stream
*stream
;
1298 assert(local_stream
);
1300 DBG("Updating poll fd array");
1301 *nb_inactive_fd
= 0;
1303 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1305 * Only active streams with an active end point can be added to the
1306 * poll set and local stream storage of the thread.
1308 * There is a potential race here for endpoint_status to be updated
1309 * just after the check. However, this is OK since the stream(s) will
1310 * be deleted once the thread is notified that the end point state has
1311 * changed where this function will be called back again.
1313 * We track the number of inactive FDs because they still need to be
1314 * closed by the polling thread after a wakeup on the data_pipe or
1317 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1318 (*nb_inactive_fd
)++;
1322 * This clobbers way too much the debug output. Uncomment that if you
1323 * need it for debugging purposes.
1325 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1326 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1327 local_stream
[i
] = stream
;
1333 * Insert the consumer_data_pipe at the end of the array and don't
1334 * increment i so nb_fd is the number of real FD.
1336 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1337 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1339 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1340 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1345 * Poll on the should_quit pipe and the command socket return -1 on
1346 * error, 1 if should exit, 0 if data is available on the command socket
1348 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1353 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1354 if (num_rdy
== -1) {
1356 * Restart interrupted system call.
1358 if (errno
== EINTR
) {
1361 PERROR("Poll error");
1364 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1365 DBG("consumer_should_quit wake up");
1372 * Set the error socket.
1374 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1377 ctx
->consumer_error_socket
= sock
;
1381 * Set the command socket path.
1383 void lttng_consumer_set_command_sock_path(
1384 struct lttng_consumer_local_data
*ctx
, char *sock
)
1386 ctx
->consumer_command_sock_path
= sock
;
1390 * Send return code to the session daemon.
1391 * If the socket is not defined, we return 0, it is not a fatal error
1393 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1395 if (ctx
->consumer_error_socket
> 0) {
1396 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1397 sizeof(enum lttcomm_sessiond_command
));
1404 * Close all the tracefiles and stream fds and MUST be called when all
1405 * instances are destroyed i.e. when all threads were joined and are ended.
1407 void lttng_consumer_cleanup(void)
1409 struct lttng_ht_iter iter
;
1410 struct lttng_consumer_channel
*channel
;
1414 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1416 consumer_del_channel(channel
);
1421 lttng_ht_destroy(consumer_data
.channel_ht
);
1422 lttng_ht_destroy(consumer_data
.channels_by_session_id_ht
);
1424 cleanup_relayd_ht();
1426 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1429 * This HT contains streams that are freed by either the metadata thread or
1430 * the data thread so we do *nothing* on the hash table and simply destroy
1433 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1435 lttng_trace_chunk_registry_destroy(consumer_data
.chunk_registry
);
1439 * Called from signal handler.
1441 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1445 CMM_STORE_SHARED(consumer_quit
, 1);
1446 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1448 PERROR("write consumer quit");
1451 DBG("Consumer flag that it should quit");
1456 * Flush pending writes to trace output disk file.
1459 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1463 int outfd
= stream
->out_fd
;
1466 * This does a blocking write-and-wait on any page that belongs to the
1467 * subbuffer prior to the one we just wrote.
1468 * Don't care about error values, as these are just hints and ways to
1469 * limit the amount of page cache used.
1471 if (orig_offset
< stream
->max_sb_size
) {
1474 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1475 stream
->max_sb_size
,
1476 SYNC_FILE_RANGE_WAIT_BEFORE
1477 | SYNC_FILE_RANGE_WRITE
1478 | SYNC_FILE_RANGE_WAIT_AFTER
);
1480 * Give hints to the kernel about how we access the file:
1481 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1484 * We need to call fadvise again after the file grows because the
1485 * kernel does not seem to apply fadvise to non-existing parts of the
1488 * Call fadvise _after_ having waited for the page writeback to
1489 * complete because the dirty page writeback semantic is not well
1490 * defined. So it can be expected to lead to lower throughput in
1493 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1494 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1495 if (ret
&& ret
!= -ENOSYS
) {
1497 PERROR("posix_fadvise on fd %i", outfd
);
1502 * Initialise the necessary environnement :
1503 * - create a new context
1504 * - create the poll_pipe
1505 * - create the should_quit pipe (for signal handler)
1506 * - create the thread pipe (for splice)
1508 * Takes a function pointer as argument, this function is called when data is
1509 * available on a buffer. This function is responsible to do the
1510 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1511 * buffer configuration and then kernctl_put_next_subbuf at the end.
1513 * Returns a pointer to the new context or NULL on error.
1515 struct lttng_consumer_local_data
*lttng_consumer_create(
1516 enum lttng_consumer_type type
,
1517 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1518 struct lttng_consumer_local_data
*ctx
),
1519 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1520 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1521 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1524 struct lttng_consumer_local_data
*ctx
;
1526 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1527 consumer_data
.type
== type
);
1528 consumer_data
.type
= type
;
1530 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1532 PERROR("allocating context");
1536 ctx
->consumer_error_socket
= -1;
1537 ctx
->consumer_metadata_socket
= -1;
1538 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1539 /* assign the callbacks */
1540 ctx
->on_buffer_ready
= buffer_ready
;
1541 ctx
->on_recv_channel
= recv_channel
;
1542 ctx
->on_recv_stream
= recv_stream
;
1543 ctx
->on_update_stream
= update_stream
;
1545 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1546 if (!ctx
->consumer_data_pipe
) {
1547 goto error_poll_pipe
;
1550 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1551 if (!ctx
->consumer_wakeup_pipe
) {
1552 goto error_wakeup_pipe
;
1555 ret
= pipe(ctx
->consumer_should_quit
);
1557 PERROR("Error creating recv pipe");
1558 goto error_quit_pipe
;
1561 ret
= pipe(ctx
->consumer_channel_pipe
);
1563 PERROR("Error creating channel pipe");
1564 goto error_channel_pipe
;
1567 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1568 if (!ctx
->consumer_metadata_pipe
) {
1569 goto error_metadata_pipe
;
1572 ctx
->channel_monitor_pipe
= -1;
1576 error_metadata_pipe
:
1577 utils_close_pipe(ctx
->consumer_channel_pipe
);
1579 utils_close_pipe(ctx
->consumer_should_quit
);
1581 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1583 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1591 * Iterate over all streams of the hashtable and free them properly.
1593 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1595 struct lttng_ht_iter iter
;
1596 struct lttng_consumer_stream
*stream
;
1603 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1605 * Ignore return value since we are currently cleaning up so any error
1608 (void) consumer_del_stream(stream
, ht
);
1612 lttng_ht_destroy(ht
);
1616 * Iterate over all streams of the metadata hashtable and free them
1619 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1621 struct lttng_ht_iter iter
;
1622 struct lttng_consumer_stream
*stream
;
1629 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1631 * Ignore return value since we are currently cleaning up so any error
1634 (void) consumer_del_metadata_stream(stream
, ht
);
1638 lttng_ht_destroy(ht
);
1642 * Close all fds associated with the instance and free the context.
1644 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1648 DBG("Consumer destroying it. Closing everything.");
1654 destroy_data_stream_ht(data_ht
);
1655 destroy_metadata_stream_ht(metadata_ht
);
1657 ret
= close(ctx
->consumer_error_socket
);
1661 ret
= close(ctx
->consumer_metadata_socket
);
1665 utils_close_pipe(ctx
->consumer_channel_pipe
);
1666 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1667 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1668 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1669 utils_close_pipe(ctx
->consumer_should_quit
);
1671 unlink(ctx
->consumer_command_sock_path
);
1676 * Write the metadata stream id on the specified file descriptor.
1678 static int write_relayd_metadata_id(int fd
,
1679 struct lttng_consumer_stream
*stream
,
1680 unsigned long padding
)
1683 struct lttcomm_relayd_metadata_payload hdr
;
1685 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1686 hdr
.padding_size
= htobe32(padding
);
1687 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1688 if (ret
< sizeof(hdr
)) {
1690 * This error means that the fd's end is closed so ignore the PERROR
1691 * not to clubber the error output since this can happen in a normal
1694 if (errno
!= EPIPE
) {
1695 PERROR("write metadata stream id");
1697 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1699 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1700 * handle writting the missing part so report that as an error and
1701 * don't lie to the caller.
1706 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1707 stream
->relayd_stream_id
, padding
);
1714 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1715 * core function for writing trace buffers to either the local filesystem or
1718 * It must be called with the stream and the channel lock held.
1720 * Careful review MUST be put if any changes occur!
1722 * Returns the number of bytes written
1724 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1725 struct lttng_consumer_local_data
*ctx
,
1726 struct lttng_consumer_stream
*stream
, unsigned long len
,
1727 unsigned long padding
,
1728 struct ctf_packet_index
*index
)
1730 unsigned long mmap_offset
;
1733 off_t orig_offset
= stream
->out_fd_offset
;
1734 /* Default is on the disk */
1735 int outfd
= stream
->out_fd
;
1736 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1737 unsigned int relayd_hang_up
= 0;
1739 /* RCU lock for the relayd pointer */
1741 assert(stream
->net_seq_idx
!= (uint64_t) -1ULL ||
1742 stream
->trace_chunk
);
1744 /* Flag that the current stream if set for network streaming. */
1745 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1746 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1747 if (relayd
== NULL
) {
1753 /* get the offset inside the fd to mmap */
1754 switch (consumer_data
.type
) {
1755 case LTTNG_CONSUMER_KERNEL
:
1756 mmap_base
= stream
->mmap_base
;
1757 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1759 PERROR("tracer ctl get_mmap_read_offset");
1763 case LTTNG_CONSUMER32_UST
:
1764 case LTTNG_CONSUMER64_UST
:
1765 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1767 ERR("read mmap get mmap base for stream %s", stream
->name
);
1771 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1773 PERROR("tracer ctl get_mmap_read_offset");
1779 ERR("Unknown consumer_data type");
1783 /* Handle stream on the relayd if the output is on the network */
1785 unsigned long netlen
= len
;
1788 * Lock the control socket for the complete duration of the function
1789 * since from this point on we will use the socket.
1791 if (stream
->metadata_flag
) {
1792 /* Metadata requires the control socket. */
1793 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1794 if (stream
->reset_metadata_flag
) {
1795 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1796 stream
->relayd_stream_id
,
1797 stream
->metadata_version
);
1802 stream
->reset_metadata_flag
= 0;
1804 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1807 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1812 /* Use the returned socket. */
1815 /* Write metadata stream id before payload */
1816 if (stream
->metadata_flag
) {
1817 ret
= write_relayd_metadata_id(outfd
, stream
, padding
);
1824 /* No streaming, we have to set the len with the full padding */
1827 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1828 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1830 ERR("Reset metadata file");
1833 stream
->reset_metadata_flag
= 0;
1837 * Check if we need to change the tracefile before writing the packet.
1839 if (stream
->chan
->tracefile_size
> 0 &&
1840 (stream
->tracefile_size_current
+ len
) >
1841 stream
->chan
->tracefile_size
) {
1842 ret
= consumer_stream_rotate_output_files(stream
);
1846 outfd
= stream
->out_fd
;
1849 stream
->tracefile_size_current
+= len
;
1851 index
->offset
= htobe64(stream
->out_fd_offset
);
1856 * This call guarantee that len or less is returned. It's impossible to
1857 * receive a ret value that is bigger than len.
1859 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1860 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1861 if (ret
< 0 || ((size_t) ret
!= len
)) {
1863 * Report error to caller if nothing was written else at least send the
1871 /* Socket operation failed. We consider the relayd dead */
1872 if (errno
== EPIPE
) {
1874 * This is possible if the fd is closed on the other side
1875 * (outfd) or any write problem. It can be verbose a bit for a
1876 * normal execution if for instance the relayd is stopped
1877 * abruptly. This can happen so set this to a DBG statement.
1879 DBG("Consumer mmap write detected relayd hang up");
1881 /* Unhandled error, print it and stop function right now. */
1882 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1886 stream
->output_written
+= ret
;
1888 /* This call is useless on a socket so better save a syscall. */
1890 /* This won't block, but will start writeout asynchronously */
1891 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1892 SYNC_FILE_RANGE_WRITE
);
1893 stream
->out_fd_offset
+= len
;
1894 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1899 * This is a special case that the relayd has closed its socket. Let's
1900 * cleanup the relayd object and all associated streams.
1902 if (relayd
&& relayd_hang_up
) {
1903 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1904 lttng_consumer_cleanup_relayd(relayd
);
1908 /* Unlock only if ctrl socket used */
1909 if (relayd
&& stream
->metadata_flag
) {
1910 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1918 * Splice the data from the ring buffer to the tracefile.
1920 * It must be called with the stream lock held.
1922 * Returns the number of bytes spliced.
1924 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1925 struct lttng_consumer_local_data
*ctx
,
1926 struct lttng_consumer_stream
*stream
, unsigned long len
,
1927 unsigned long padding
,
1928 struct ctf_packet_index
*index
)
1930 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1932 off_t orig_offset
= stream
->out_fd_offset
;
1933 int fd
= stream
->wait_fd
;
1934 /* Default is on the disk */
1935 int outfd
= stream
->out_fd
;
1936 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1938 unsigned int relayd_hang_up
= 0;
1940 switch (consumer_data
.type
) {
1941 case LTTNG_CONSUMER_KERNEL
:
1943 case LTTNG_CONSUMER32_UST
:
1944 case LTTNG_CONSUMER64_UST
:
1945 /* Not supported for user space tracing */
1948 ERR("Unknown consumer_data type");
1952 /* RCU lock for the relayd pointer */
1955 /* Flag that the current stream if set for network streaming. */
1956 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1957 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1958 if (relayd
== NULL
) {
1963 splice_pipe
= stream
->splice_pipe
;
1965 /* Write metadata stream id before payload */
1967 unsigned long total_len
= len
;
1969 if (stream
->metadata_flag
) {
1971 * Lock the control socket for the complete duration of the function
1972 * since from this point on we will use the socket.
1974 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1976 if (stream
->reset_metadata_flag
) {
1977 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1978 stream
->relayd_stream_id
,
1979 stream
->metadata_version
);
1984 stream
->reset_metadata_flag
= 0;
1986 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
,
1994 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1997 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
2003 /* Use the returned socket. */
2006 /* No streaming, we have to set the len with the full padding */
2009 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
2010 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
2012 ERR("Reset metadata file");
2015 stream
->reset_metadata_flag
= 0;
2018 * Check if we need to change the tracefile before writing the packet.
2020 if (stream
->chan
->tracefile_size
> 0 &&
2021 (stream
->tracefile_size_current
+ len
) >
2022 stream
->chan
->tracefile_size
) {
2023 ret
= consumer_stream_rotate_output_files(stream
);
2028 outfd
= stream
->out_fd
;
2031 stream
->tracefile_size_current
+= len
;
2032 index
->offset
= htobe64(stream
->out_fd_offset
);
2036 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
2037 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
2038 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
2039 SPLICE_F_MOVE
| SPLICE_F_MORE
);
2040 DBG("splice chan to pipe, ret %zd", ret_splice
);
2041 if (ret_splice
< 0) {
2044 PERROR("Error in relay splice");
2048 /* Handle stream on the relayd if the output is on the network */
2049 if (relayd
&& stream
->metadata_flag
) {
2050 size_t metadata_payload_size
=
2051 sizeof(struct lttcomm_relayd_metadata_payload
);
2053 /* Update counter to fit the spliced data */
2054 ret_splice
+= metadata_payload_size
;
2055 len
+= metadata_payload_size
;
2057 * We do this so the return value can match the len passed as
2058 * argument to this function.
2060 written
-= metadata_payload_size
;
2063 /* Splice data out */
2064 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
2065 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
2066 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
2068 if (ret_splice
< 0) {
2073 } else if (ret_splice
> len
) {
2075 * We don't expect this code path to be executed but you never know
2076 * so this is an extra protection agains a buggy splice().
2079 written
+= ret_splice
;
2080 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
2084 /* All good, update current len and continue. */
2088 /* This call is useless on a socket so better save a syscall. */
2090 /* This won't block, but will start writeout asynchronously */
2091 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
2092 SYNC_FILE_RANGE_WRITE
);
2093 stream
->out_fd_offset
+= ret_splice
;
2095 stream
->output_written
+= ret_splice
;
2096 written
+= ret_splice
;
2099 lttng_consumer_sync_trace_file(stream
, orig_offset
);
2105 * This is a special case that the relayd has closed its socket. Let's
2106 * cleanup the relayd object and all associated streams.
2108 if (relayd
&& relayd_hang_up
) {
2109 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
2110 lttng_consumer_cleanup_relayd(relayd
);
2111 /* Skip splice error so the consumer does not fail */
2116 /* send the appropriate error description to sessiond */
2119 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
2122 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
2125 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
2130 if (relayd
&& stream
->metadata_flag
) {
2131 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2139 * Sample the snapshot positions for a specific fd
2141 * Returns 0 on success, < 0 on error
2143 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream
*stream
)
2145 switch (consumer_data
.type
) {
2146 case LTTNG_CONSUMER_KERNEL
:
2147 return lttng_kconsumer_sample_snapshot_positions(stream
);
2148 case LTTNG_CONSUMER32_UST
:
2149 case LTTNG_CONSUMER64_UST
:
2150 return lttng_ustconsumer_sample_snapshot_positions(stream
);
2152 ERR("Unknown consumer_data type");
2158 * Take a snapshot for a specific fd
2160 * Returns 0 on success, < 0 on error
2162 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
2164 switch (consumer_data
.type
) {
2165 case LTTNG_CONSUMER_KERNEL
:
2166 return lttng_kconsumer_take_snapshot(stream
);
2167 case LTTNG_CONSUMER32_UST
:
2168 case LTTNG_CONSUMER64_UST
:
2169 return lttng_ustconsumer_take_snapshot(stream
);
2171 ERR("Unknown consumer_data type");
2178 * Get the produced position
2180 * Returns 0 on success, < 0 on error
2182 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
2185 switch (consumer_data
.type
) {
2186 case LTTNG_CONSUMER_KERNEL
:
2187 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
2188 case LTTNG_CONSUMER32_UST
:
2189 case LTTNG_CONSUMER64_UST
:
2190 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2192 ERR("Unknown consumer_data type");
2199 * Get the consumed position (free-running counter position in bytes).
2201 * Returns 0 on success, < 0 on error
2203 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream
*stream
,
2206 switch (consumer_data
.type
) {
2207 case LTTNG_CONSUMER_KERNEL
:
2208 return lttng_kconsumer_get_consumed_snapshot(stream
, pos
);
2209 case LTTNG_CONSUMER32_UST
:
2210 case LTTNG_CONSUMER64_UST
:
2211 return lttng_ustconsumer_get_consumed_snapshot(stream
, pos
);
2213 ERR("Unknown consumer_data type");
2219 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2220 int sock
, struct pollfd
*consumer_sockpoll
)
2222 switch (consumer_data
.type
) {
2223 case LTTNG_CONSUMER_KERNEL
:
2224 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2225 case LTTNG_CONSUMER32_UST
:
2226 case LTTNG_CONSUMER64_UST
:
2227 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2229 ERR("Unknown consumer_data type");
2235 void lttng_consumer_close_all_metadata(void)
2237 switch (consumer_data
.type
) {
2238 case LTTNG_CONSUMER_KERNEL
:
2240 * The Kernel consumer has a different metadata scheme so we don't
2241 * close anything because the stream will be closed by the session
2245 case LTTNG_CONSUMER32_UST
:
2246 case LTTNG_CONSUMER64_UST
:
2248 * Close all metadata streams. The metadata hash table is passed and
2249 * this call iterates over it by closing all wakeup fd. This is safe
2250 * because at this point we are sure that the metadata producer is
2251 * either dead or blocked.
2253 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2256 ERR("Unknown consumer_data type");
2262 * Clean up a metadata stream and free its memory.
2264 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2265 struct lttng_ht
*ht
)
2267 struct lttng_consumer_channel
*channel
= NULL
;
2268 bool free_channel
= false;
2272 * This call should NEVER receive regular stream. It must always be
2273 * metadata stream and this is crucial for data structure synchronization.
2275 assert(stream
->metadata_flag
);
2277 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2279 pthread_mutex_lock(&consumer_data
.lock
);
2281 * Note that this assumes that a stream's channel is never changed and
2282 * that the stream's lock doesn't need to be taken to sample its
2285 channel
= stream
->chan
;
2286 pthread_mutex_lock(&channel
->lock
);
2287 pthread_mutex_lock(&stream
->lock
);
2288 if (channel
->metadata_cache
) {
2289 /* Only applicable to userspace consumers. */
2290 pthread_mutex_lock(&channel
->metadata_cache
->lock
);
2293 /* Remove any reference to that stream. */
2294 consumer_stream_delete(stream
, ht
);
2296 /* Close down everything including the relayd if one. */
2297 consumer_stream_close(stream
);
2298 /* Destroy tracer buffers of the stream. */
2299 consumer_stream_destroy_buffers(stream
);
2301 /* Atomically decrement channel refcount since other threads can use it. */
2302 if (!uatomic_sub_return(&channel
->refcount
, 1)
2303 && !uatomic_read(&channel
->nb_init_stream_left
)) {
2304 /* Go for channel deletion! */
2305 free_channel
= true;
2307 stream
->chan
= NULL
;
2310 * Nullify the stream reference so it is not used after deletion. The
2311 * channel lock MUST be acquired before being able to check for a NULL
2314 channel
->metadata_stream
= NULL
;
2316 if (channel
->metadata_cache
) {
2317 pthread_mutex_unlock(&channel
->metadata_cache
->lock
);
2319 pthread_mutex_unlock(&stream
->lock
);
2320 pthread_mutex_unlock(&channel
->lock
);
2321 pthread_mutex_unlock(&consumer_data
.lock
);
2324 consumer_del_channel(channel
);
2327 lttng_trace_chunk_put(stream
->trace_chunk
);
2328 stream
->trace_chunk
= NULL
;
2329 consumer_stream_free(stream
);
2333 * Action done with the metadata stream when adding it to the consumer internal
2334 * data structures to handle it.
2336 void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2338 struct lttng_ht
*ht
= metadata_ht
;
2339 struct lttng_ht_iter iter
;
2340 struct lttng_ht_node_u64
*node
;
2345 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2347 pthread_mutex_lock(&consumer_data
.lock
);
2348 pthread_mutex_lock(&stream
->chan
->lock
);
2349 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2350 pthread_mutex_lock(&stream
->lock
);
2353 * From here, refcounts are updated so be _careful_ when returning an error
2360 * Lookup the stream just to make sure it does not exist in our internal
2361 * state. This should NEVER happen.
2363 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2364 node
= lttng_ht_iter_get_node_u64(&iter
);
2368 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2369 * in terms of destroying the associated channel, because the action that
2370 * causes the count to become 0 also causes a stream to be added. The
2371 * channel deletion will thus be triggered by the following removal of this
2374 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2375 /* Increment refcount before decrementing nb_init_stream_left */
2377 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2380 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2382 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2383 &stream
->node_channel_id
);
2386 * Add stream to the stream_list_ht of the consumer data. No need to steal
2387 * the key since the HT does not use it and we allow to add redundant keys
2390 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2394 pthread_mutex_unlock(&stream
->lock
);
2395 pthread_mutex_unlock(&stream
->chan
->lock
);
2396 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2397 pthread_mutex_unlock(&consumer_data
.lock
);
2401 * Delete data stream that are flagged for deletion (endpoint_status).
2403 static void validate_endpoint_status_data_stream(void)
2405 struct lttng_ht_iter iter
;
2406 struct lttng_consumer_stream
*stream
;
2408 DBG("Consumer delete flagged data stream");
2411 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2412 /* Validate delete flag of the stream */
2413 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2416 /* Delete it right now */
2417 consumer_del_stream(stream
, data_ht
);
2423 * Delete metadata stream that are flagged for deletion (endpoint_status).
2425 static void validate_endpoint_status_metadata_stream(
2426 struct lttng_poll_event
*pollset
)
2428 struct lttng_ht_iter iter
;
2429 struct lttng_consumer_stream
*stream
;
2431 DBG("Consumer delete flagged metadata stream");
2436 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2437 /* Validate delete flag of the stream */
2438 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2442 * Remove from pollset so the metadata thread can continue without
2443 * blocking on a deleted stream.
2445 lttng_poll_del(pollset
, stream
->wait_fd
);
2447 /* Delete it right now */
2448 consumer_del_metadata_stream(stream
, metadata_ht
);
2454 * Thread polls on metadata file descriptor and write them on disk or on the
2457 void *consumer_thread_metadata_poll(void *data
)
2459 int ret
, i
, pollfd
, err
= -1;
2460 uint32_t revents
, nb_fd
;
2461 struct lttng_consumer_stream
*stream
= NULL
;
2462 struct lttng_ht_iter iter
;
2463 struct lttng_ht_node_u64
*node
;
2464 struct lttng_poll_event events
;
2465 struct lttng_consumer_local_data
*ctx
= data
;
2468 rcu_register_thread();
2470 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2472 if (testpoint(consumerd_thread_metadata
)) {
2473 goto error_testpoint
;
2476 health_code_update();
2478 DBG("Thread metadata poll started");
2480 /* Size is set to 1 for the consumer_metadata pipe */
2481 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2483 ERR("Poll set creation failed");
2487 ret
= lttng_poll_add(&events
,
2488 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2494 DBG("Metadata main loop started");
2498 health_code_update();
2499 health_poll_entry();
2500 DBG("Metadata poll wait");
2501 ret
= lttng_poll_wait(&events
, -1);
2502 DBG("Metadata poll return from wait with %d fd(s)",
2503 LTTNG_POLL_GETNB(&events
));
2505 DBG("Metadata event caught in thread");
2507 if (errno
== EINTR
) {
2508 ERR("Poll EINTR caught");
2511 if (LTTNG_POLL_GETNB(&events
) == 0) {
2512 err
= 0; /* All is OK */
2519 /* From here, the event is a metadata wait fd */
2520 for (i
= 0; i
< nb_fd
; i
++) {
2521 health_code_update();
2523 revents
= LTTNG_POLL_GETEV(&events
, i
);
2524 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2526 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2527 if (revents
& LPOLLIN
) {
2530 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2531 &stream
, sizeof(stream
));
2532 if (pipe_len
< sizeof(stream
)) {
2534 PERROR("read metadata stream");
2537 * Remove the pipe from the poll set and continue the loop
2538 * since their might be data to consume.
2540 lttng_poll_del(&events
,
2541 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2542 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2546 /* A NULL stream means that the state has changed. */
2547 if (stream
== NULL
) {
2548 /* Check for deleted streams. */
2549 validate_endpoint_status_metadata_stream(&events
);
2553 DBG("Adding metadata stream %d to poll set",
2556 /* Add metadata stream to the global poll events list */
2557 lttng_poll_add(&events
, stream
->wait_fd
,
2558 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2559 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2560 DBG("Metadata thread pipe hung up");
2562 * Remove the pipe from the poll set and continue the loop
2563 * since their might be data to consume.
2565 lttng_poll_del(&events
,
2566 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2567 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2570 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2574 /* Handle other stream */
2580 uint64_t tmp_id
= (uint64_t) pollfd
;
2582 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2584 node
= lttng_ht_iter_get_node_u64(&iter
);
2587 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2590 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2591 /* Get the data out of the metadata file descriptor */
2592 DBG("Metadata available on fd %d", pollfd
);
2593 assert(stream
->wait_fd
== pollfd
);
2596 health_code_update();
2598 len
= ctx
->on_buffer_ready(stream
, ctx
);
2600 * We don't check the return value here since if we get
2601 * a negative len, it means an error occurred thus we
2602 * simply remove it from the poll set and free the
2607 /* It's ok to have an unavailable sub-buffer */
2608 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2609 /* Clean up stream from consumer and free it. */
2610 lttng_poll_del(&events
, stream
->wait_fd
);
2611 consumer_del_metadata_stream(stream
, metadata_ht
);
2613 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2614 DBG("Metadata fd %d is hup|err.", pollfd
);
2615 if (!stream
->hangup_flush_done
2616 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2617 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2618 DBG("Attempting to flush and consume the UST buffers");
2619 lttng_ustconsumer_on_stream_hangup(stream
);
2621 /* We just flushed the stream now read it. */
2623 health_code_update();
2625 len
= ctx
->on_buffer_ready(stream
, ctx
);
2627 * We don't check the return value here since if we get
2628 * a negative len, it means an error occurred thus we
2629 * simply remove it from the poll set and free the
2635 lttng_poll_del(&events
, stream
->wait_fd
);
2637 * This call update the channel states, closes file descriptors
2638 * and securely free the stream.
2640 consumer_del_metadata_stream(stream
, metadata_ht
);
2642 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2646 /* Release RCU lock for the stream looked up */
2654 DBG("Metadata poll thread exiting");
2656 lttng_poll_clean(&events
);
2661 ERR("Health error occurred in %s", __func__
);
2663 health_unregister(health_consumerd
);
2664 rcu_unregister_thread();
2669 * This thread polls the fds in the set to consume the data and write
2670 * it to tracefile if necessary.
2672 void *consumer_thread_data_poll(void *data
)
2674 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2675 struct pollfd
*pollfd
= NULL
;
2676 /* local view of the streams */
2677 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2678 /* local view of consumer_data.fds_count */
2680 /* 2 for the consumer_data_pipe and wake up pipe */
2681 const int nb_pipes_fd
= 2;
2682 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2683 int nb_inactive_fd
= 0;
2684 struct lttng_consumer_local_data
*ctx
= data
;
2687 rcu_register_thread();
2689 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2691 if (testpoint(consumerd_thread_data
)) {
2692 goto error_testpoint
;
2695 health_code_update();
2697 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2698 if (local_stream
== NULL
) {
2699 PERROR("local_stream malloc");
2704 health_code_update();
2710 * the fds set has been updated, we need to update our
2711 * local array as well
2713 pthread_mutex_lock(&consumer_data
.lock
);
2714 if (consumer_data
.need_update
) {
2719 local_stream
= NULL
;
2721 /* Allocate for all fds */
2722 pollfd
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) * sizeof(struct pollfd
));
2723 if (pollfd
== NULL
) {
2724 PERROR("pollfd malloc");
2725 pthread_mutex_unlock(&consumer_data
.lock
);
2729 local_stream
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) *
2730 sizeof(struct lttng_consumer_stream
*));
2731 if (local_stream
== NULL
) {
2732 PERROR("local_stream malloc");
2733 pthread_mutex_unlock(&consumer_data
.lock
);
2736 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2737 data_ht
, &nb_inactive_fd
);
2739 ERR("Error in allocating pollfd or local_outfds");
2740 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2741 pthread_mutex_unlock(&consumer_data
.lock
);
2745 consumer_data
.need_update
= 0;
2747 pthread_mutex_unlock(&consumer_data
.lock
);
2749 /* No FDs and consumer_quit, consumer_cleanup the thread */
2750 if (nb_fd
== 0 && nb_inactive_fd
== 0 &&
2751 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2752 err
= 0; /* All is OK */
2755 /* poll on the array of fds */
2757 DBG("polling on %d fd", nb_fd
+ nb_pipes_fd
);
2758 if (testpoint(consumerd_thread_data_poll
)) {
2761 health_poll_entry();
2762 num_rdy
= poll(pollfd
, nb_fd
+ nb_pipes_fd
, -1);
2764 DBG("poll num_rdy : %d", num_rdy
);
2765 if (num_rdy
== -1) {
2767 * Restart interrupted system call.
2769 if (errno
== EINTR
) {
2772 PERROR("Poll error");
2773 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2775 } else if (num_rdy
== 0) {
2776 DBG("Polling thread timed out");
2780 if (caa_unlikely(data_consumption_paused
)) {
2781 DBG("Data consumption paused, sleeping...");
2787 * If the consumer_data_pipe triggered poll go directly to the
2788 * beginning of the loop to update the array. We want to prioritize
2789 * array update over low-priority reads.
2791 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2792 ssize_t pipe_readlen
;
2794 DBG("consumer_data_pipe wake up");
2795 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2796 &new_stream
, sizeof(new_stream
));
2797 if (pipe_readlen
< sizeof(new_stream
)) {
2798 PERROR("Consumer data pipe");
2799 /* Continue so we can at least handle the current stream(s). */
2804 * If the stream is NULL, just ignore it. It's also possible that
2805 * the sessiond poll thread changed the consumer_quit state and is
2806 * waking us up to test it.
2808 if (new_stream
== NULL
) {
2809 validate_endpoint_status_data_stream();
2813 /* Continue to update the local streams and handle prio ones */
2817 /* Handle wakeup pipe. */
2818 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2820 ssize_t pipe_readlen
;
2822 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2824 if (pipe_readlen
< 0) {
2825 PERROR("Consumer data wakeup pipe");
2827 /* We've been awakened to handle stream(s). */
2828 ctx
->has_wakeup
= 0;
2831 /* Take care of high priority channels first. */
2832 for (i
= 0; i
< nb_fd
; i
++) {
2833 health_code_update();
2835 if (local_stream
[i
] == NULL
) {
2838 if (pollfd
[i
].revents
& POLLPRI
) {
2839 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2841 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2842 /* it's ok to have an unavailable sub-buffer */
2843 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2844 /* Clean the stream and free it. */
2845 consumer_del_stream(local_stream
[i
], data_ht
);
2846 local_stream
[i
] = NULL
;
2847 } else if (len
> 0) {
2848 local_stream
[i
]->data_read
= 1;
2854 * If we read high prio channel in this loop, try again
2855 * for more high prio data.
2861 /* Take care of low priority channels. */
2862 for (i
= 0; i
< nb_fd
; i
++) {
2863 health_code_update();
2865 if (local_stream
[i
] == NULL
) {
2868 if ((pollfd
[i
].revents
& POLLIN
) ||
2869 local_stream
[i
]->hangup_flush_done
||
2870 local_stream
[i
]->has_data
) {
2871 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2872 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2873 /* it's ok to have an unavailable sub-buffer */
2874 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2875 /* Clean the stream and free it. */
2876 consumer_del_stream(local_stream
[i
], data_ht
);
2877 local_stream
[i
] = NULL
;
2878 } else if (len
> 0) {
2879 local_stream
[i
]->data_read
= 1;
2884 /* Handle hangup and errors */
2885 for (i
= 0; i
< nb_fd
; i
++) {
2886 health_code_update();
2888 if (local_stream
[i
] == NULL
) {
2891 if (!local_stream
[i
]->hangup_flush_done
2892 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2893 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2894 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2895 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2897 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2898 /* Attempt read again, for the data we just flushed. */
2899 local_stream
[i
]->data_read
= 1;
2902 * If the poll flag is HUP/ERR/NVAL and we have
2903 * read no data in this pass, we can remove the
2904 * stream from its hash table.
2906 if ((pollfd
[i
].revents
& POLLHUP
)) {
2907 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2908 if (!local_stream
[i
]->data_read
) {
2909 consumer_del_stream(local_stream
[i
], data_ht
);
2910 local_stream
[i
] = NULL
;
2913 } else if (pollfd
[i
].revents
& POLLERR
) {
2914 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2915 if (!local_stream
[i
]->data_read
) {
2916 consumer_del_stream(local_stream
[i
], data_ht
);
2917 local_stream
[i
] = NULL
;
2920 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2921 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2922 if (!local_stream
[i
]->data_read
) {
2923 consumer_del_stream(local_stream
[i
], data_ht
);
2924 local_stream
[i
] = NULL
;
2928 if (local_stream
[i
] != NULL
) {
2929 local_stream
[i
]->data_read
= 0;
2936 DBG("polling thread exiting");
2941 * Close the write side of the pipe so epoll_wait() in
2942 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2943 * read side of the pipe. If we close them both, epoll_wait strangely does
2944 * not return and could create a endless wait period if the pipe is the
2945 * only tracked fd in the poll set. The thread will take care of closing
2948 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2953 ERR("Health error occurred in %s", __func__
);
2955 health_unregister(health_consumerd
);
2957 rcu_unregister_thread();
2962 * Close wake-up end of each stream belonging to the channel. This will
2963 * allow the poll() on the stream read-side to detect when the
2964 * write-side (application) finally closes them.
2967 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2969 struct lttng_ht
*ht
;
2970 struct lttng_consumer_stream
*stream
;
2971 struct lttng_ht_iter iter
;
2973 ht
= consumer_data
.stream_per_chan_id_ht
;
2976 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2977 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2978 ht
->match_fct
, &channel
->key
,
2979 &iter
.iter
, stream
, node_channel_id
.node
) {
2981 * Protect against teardown with mutex.
2983 pthread_mutex_lock(&stream
->lock
);
2984 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2987 switch (consumer_data
.type
) {
2988 case LTTNG_CONSUMER_KERNEL
:
2990 case LTTNG_CONSUMER32_UST
:
2991 case LTTNG_CONSUMER64_UST
:
2992 if (stream
->metadata_flag
) {
2993 /* Safe and protected by the stream lock. */
2994 lttng_ustconsumer_close_metadata(stream
->chan
);
2997 * Note: a mutex is taken internally within
2998 * liblttng-ust-ctl to protect timer wakeup_fd
2999 * use from concurrent close.
3001 lttng_ustconsumer_close_stream_wakeup(stream
);
3005 ERR("Unknown consumer_data type");
3009 pthread_mutex_unlock(&stream
->lock
);
3014 static void destroy_channel_ht(struct lttng_ht
*ht
)
3016 struct lttng_ht_iter iter
;
3017 struct lttng_consumer_channel
*channel
;
3025 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
3026 ret
= lttng_ht_del(ht
, &iter
);
3031 lttng_ht_destroy(ht
);
3035 * This thread polls the channel fds to detect when they are being
3036 * closed. It closes all related streams if the channel is detected as
3037 * closed. It is currently only used as a shim layer for UST because the
3038 * consumerd needs to keep the per-stream wakeup end of pipes open for
3041 void *consumer_thread_channel_poll(void *data
)
3043 int ret
, i
, pollfd
, err
= -1;
3044 uint32_t revents
, nb_fd
;
3045 struct lttng_consumer_channel
*chan
= NULL
;
3046 struct lttng_ht_iter iter
;
3047 struct lttng_ht_node_u64
*node
;
3048 struct lttng_poll_event events
;
3049 struct lttng_consumer_local_data
*ctx
= data
;
3050 struct lttng_ht
*channel_ht
;
3052 rcu_register_thread();
3054 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
3056 if (testpoint(consumerd_thread_channel
)) {
3057 goto error_testpoint
;
3060 health_code_update();
3062 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3064 /* ENOMEM at this point. Better to bail out. */
3068 DBG("Thread channel poll started");
3070 /* Size is set to 1 for the consumer_channel pipe */
3071 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
3073 ERR("Poll set creation failed");
3077 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
3083 DBG("Channel main loop started");
3087 health_code_update();
3088 DBG("Channel poll wait");
3089 health_poll_entry();
3090 ret
= lttng_poll_wait(&events
, -1);
3091 DBG("Channel poll return from wait with %d fd(s)",
3092 LTTNG_POLL_GETNB(&events
));
3094 DBG("Channel event caught in thread");
3096 if (errno
== EINTR
) {
3097 ERR("Poll EINTR caught");
3100 if (LTTNG_POLL_GETNB(&events
) == 0) {
3101 err
= 0; /* All is OK */
3108 /* From here, the event is a channel wait fd */
3109 for (i
= 0; i
< nb_fd
; i
++) {
3110 health_code_update();
3112 revents
= LTTNG_POLL_GETEV(&events
, i
);
3113 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
3115 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
3116 if (revents
& LPOLLIN
) {
3117 enum consumer_channel_action action
;
3120 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
3123 ERR("Error reading channel pipe");
3125 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3130 case CONSUMER_CHANNEL_ADD
:
3131 DBG("Adding channel %d to poll set",
3134 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
3137 lttng_ht_add_unique_u64(channel_ht
,
3138 &chan
->wait_fd_node
);
3140 /* Add channel to the global poll events list */
3141 lttng_poll_add(&events
, chan
->wait_fd
,
3142 LPOLLERR
| LPOLLHUP
);
3144 case CONSUMER_CHANNEL_DEL
:
3147 * This command should never be called if the channel
3148 * has streams monitored by either the data or metadata
3149 * thread. The consumer only notify this thread with a
3150 * channel del. command if it receives a destroy
3151 * channel command from the session daemon that send it
3152 * if a command prior to the GET_CHANNEL failed.
3156 chan
= consumer_find_channel(key
);
3159 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
3162 lttng_poll_del(&events
, chan
->wait_fd
);
3163 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
3164 ret
= lttng_ht_del(channel_ht
, &iter
);
3167 switch (consumer_data
.type
) {
3168 case LTTNG_CONSUMER_KERNEL
:
3170 case LTTNG_CONSUMER32_UST
:
3171 case LTTNG_CONSUMER64_UST
:
3172 health_code_update();
3173 /* Destroy streams that might have been left in the stream list. */
3174 clean_channel_stream_list(chan
);
3177 ERR("Unknown consumer_data type");
3182 * Release our own refcount. Force channel deletion even if
3183 * streams were not initialized.
3185 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
3186 consumer_del_channel(chan
);
3191 case CONSUMER_CHANNEL_QUIT
:
3193 * Remove the pipe from the poll set and continue the loop
3194 * since their might be data to consume.
3196 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3199 ERR("Unknown action");
3202 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3203 DBG("Channel thread pipe hung up");
3205 * Remove the pipe from the poll set and continue the loop
3206 * since their might be data to consume.
3208 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3211 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3215 /* Handle other stream */
3221 uint64_t tmp_id
= (uint64_t) pollfd
;
3223 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3225 node
= lttng_ht_iter_get_node_u64(&iter
);
3228 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3231 /* Check for error event */
3232 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3233 DBG("Channel fd %d is hup|err.", pollfd
);
3235 lttng_poll_del(&events
, chan
->wait_fd
);
3236 ret
= lttng_ht_del(channel_ht
, &iter
);
3240 * This will close the wait fd for each stream associated to
3241 * this channel AND monitored by the data/metadata thread thus
3242 * will be clean by the right thread.
3244 consumer_close_channel_streams(chan
);
3246 /* Release our own refcount */
3247 if (!uatomic_sub_return(&chan
->refcount
, 1)
3248 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3249 consumer_del_channel(chan
);
3252 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3257 /* Release RCU lock for the channel looked up */
3265 lttng_poll_clean(&events
);
3267 destroy_channel_ht(channel_ht
);
3270 DBG("Channel poll thread exiting");
3273 ERR("Health error occurred in %s", __func__
);
3275 health_unregister(health_consumerd
);
3276 rcu_unregister_thread();
3280 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3281 struct pollfd
*sockpoll
, int client_socket
)
3288 ret
= lttng_consumer_poll_socket(sockpoll
);
3292 DBG("Metadata connection on client_socket");
3294 /* Blocking call, waiting for transmission */
3295 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3296 if (ctx
->consumer_metadata_socket
< 0) {
3297 WARN("On accept metadata");
3308 * This thread listens on the consumerd socket and receives the file
3309 * descriptors from the session daemon.
3311 void *consumer_thread_sessiond_poll(void *data
)
3313 int sock
= -1, client_socket
, ret
, err
= -1;
3315 * structure to poll for incoming data on communication socket avoids
3316 * making blocking sockets.
3318 struct pollfd consumer_sockpoll
[2];
3319 struct lttng_consumer_local_data
*ctx
= data
;
3321 rcu_register_thread();
3323 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3325 if (testpoint(consumerd_thread_sessiond
)) {
3326 goto error_testpoint
;
3329 health_code_update();
3331 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3332 unlink(ctx
->consumer_command_sock_path
);
3333 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3334 if (client_socket
< 0) {
3335 ERR("Cannot create command socket");
3339 ret
= lttcomm_listen_unix_sock(client_socket
);
3344 DBG("Sending ready command to lttng-sessiond");
3345 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3346 /* return < 0 on error, but == 0 is not fatal */
3348 ERR("Error sending ready command to lttng-sessiond");
3352 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3353 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3354 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3355 consumer_sockpoll
[1].fd
= client_socket
;
3356 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3358 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3366 DBG("Connection on client_socket");
3368 /* Blocking call, waiting for transmission */
3369 sock
= lttcomm_accept_unix_sock(client_socket
);
3376 * Setup metadata socket which is the second socket connection on the
3377 * command unix socket.
3379 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3388 /* This socket is not useful anymore. */
3389 ret
= close(client_socket
);
3391 PERROR("close client_socket");
3395 /* update the polling structure to poll on the established socket */
3396 consumer_sockpoll
[1].fd
= sock
;
3397 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3400 health_code_update();
3402 health_poll_entry();
3403 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3412 DBG("Incoming command on sock");
3413 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3416 * This could simply be a session daemon quitting. Don't output
3419 DBG("Communication interrupted on command socket");
3423 if (CMM_LOAD_SHARED(consumer_quit
)) {
3424 DBG("consumer_thread_receive_fds received quit from signal");
3425 err
= 0; /* All is OK */
3428 DBG("received command on sock");
3434 DBG("Consumer thread sessiond poll exiting");
3437 * Close metadata streams since the producer is the session daemon which
3440 * NOTE: for now, this only applies to the UST tracer.
3442 lttng_consumer_close_all_metadata();
3445 * when all fds have hung up, the polling thread
3448 CMM_STORE_SHARED(consumer_quit
, 1);
3451 * Notify the data poll thread to poll back again and test the
3452 * consumer_quit state that we just set so to quit gracefully.
3454 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3456 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3458 notify_health_quit_pipe(health_quit_pipe
);
3460 /* Cleaning up possibly open sockets. */
3464 PERROR("close sock sessiond poll");
3467 if (client_socket
>= 0) {
3468 ret
= close(client_socket
);
3470 PERROR("close client_socket sessiond poll");
3477 ERR("Health error occurred in %s", __func__
);
3479 health_unregister(health_consumerd
);
3481 rcu_unregister_thread();
3485 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3486 struct lttng_consumer_local_data
*ctx
)
3490 pthread_mutex_lock(&stream
->chan
->lock
);
3491 pthread_mutex_lock(&stream
->lock
);
3492 if (stream
->metadata_flag
) {
3493 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3496 switch (consumer_data
.type
) {
3497 case LTTNG_CONSUMER_KERNEL
:
3498 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3500 case LTTNG_CONSUMER32_UST
:
3501 case LTTNG_CONSUMER64_UST
:
3502 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3505 ERR("Unknown consumer_data type");
3511 if (stream
->metadata_flag
) {
3512 pthread_cond_broadcast(&stream
->metadata_rdv
);
3513 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3515 pthread_mutex_unlock(&stream
->lock
);
3516 pthread_mutex_unlock(&stream
->chan
->lock
);
3521 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3523 switch (consumer_data
.type
) {
3524 case LTTNG_CONSUMER_KERNEL
:
3525 return lttng_kconsumer_on_recv_stream(stream
);
3526 case LTTNG_CONSUMER32_UST
:
3527 case LTTNG_CONSUMER64_UST
:
3528 return lttng_ustconsumer_on_recv_stream(stream
);
3530 ERR("Unknown consumer_data type");
3537 * Allocate and set consumer data hash tables.
3539 int lttng_consumer_init(void)
3541 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3542 if (!consumer_data
.channel_ht
) {
3546 consumer_data
.channels_by_session_id_ht
=
3547 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3548 if (!consumer_data
.channels_by_session_id_ht
) {
3552 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3553 if (!consumer_data
.relayd_ht
) {
3557 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3558 if (!consumer_data
.stream_list_ht
) {
3562 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3563 if (!consumer_data
.stream_per_chan_id_ht
) {
3567 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3572 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3577 consumer_data
.chunk_registry
= lttng_trace_chunk_registry_create();
3578 if (!consumer_data
.chunk_registry
) {
3589 * Process the ADD_RELAYD command receive by a consumer.
3591 * This will create a relayd socket pair and add it to the relayd hash table.
3592 * The caller MUST acquire a RCU read side lock before calling it.
3594 void consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3595 struct lttng_consumer_local_data
*ctx
, int sock
,
3596 struct pollfd
*consumer_sockpoll
,
3597 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3598 uint64_t relayd_session_id
)
3600 int fd
= -1, ret
= -1, relayd_created
= 0;
3601 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3602 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3605 assert(relayd_sock
);
3607 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3609 /* Get relayd reference if exists. */
3610 relayd
= consumer_find_relayd(net_seq_idx
);
3611 if (relayd
== NULL
) {
3612 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3613 /* Not found. Allocate one. */
3614 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3615 if (relayd
== NULL
) {
3616 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3619 relayd
->sessiond_session_id
= sessiond_id
;
3624 * This code path MUST continue to the consumer send status message to
3625 * we can notify the session daemon and continue our work without
3626 * killing everything.
3630 * relayd key should never be found for control socket.
3632 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3635 /* First send a status message before receiving the fds. */
3636 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3638 /* Somehow, the session daemon is not responding anymore. */
3639 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3640 goto error_nosignal
;
3643 /* Poll on consumer socket. */
3644 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3646 /* Needing to exit in the middle of a command: error. */
3647 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3648 goto error_nosignal
;
3651 /* Get relayd socket from session daemon */
3652 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3653 if (ret
!= sizeof(fd
)) {
3654 fd
= -1; /* Just in case it gets set with an invalid value. */
3657 * Failing to receive FDs might indicate a major problem such as
3658 * reaching a fd limit during the receive where the kernel returns a
3659 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3660 * don't take any chances and stop everything.
3662 * XXX: Feature request #558 will fix that and avoid this possible
3663 * issue when reaching the fd limit.
3665 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3666 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3670 /* Copy socket information and received FD */
3671 switch (sock_type
) {
3672 case LTTNG_STREAM_CONTROL
:
3673 /* Copy received lttcomm socket */
3674 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3675 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3676 /* Handle create_sock error. */
3678 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3682 * Close the socket created internally by
3683 * lttcomm_create_sock, so we can replace it by the one
3684 * received from sessiond.
3686 if (close(relayd
->control_sock
.sock
.fd
)) {
3690 /* Assign new file descriptor */
3691 relayd
->control_sock
.sock
.fd
= fd
;
3692 /* Assign version values. */
3693 relayd
->control_sock
.major
= relayd_sock
->major
;
3694 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3696 relayd
->relayd_session_id
= relayd_session_id
;
3699 case LTTNG_STREAM_DATA
:
3700 /* Copy received lttcomm socket */
3701 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3702 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3703 /* Handle create_sock error. */
3705 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3709 * Close the socket created internally by
3710 * lttcomm_create_sock, so we can replace it by the one
3711 * received from sessiond.
3713 if (close(relayd
->data_sock
.sock
.fd
)) {
3717 /* Assign new file descriptor */
3718 relayd
->data_sock
.sock
.fd
= fd
;
3719 /* Assign version values. */
3720 relayd
->data_sock
.major
= relayd_sock
->major
;
3721 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3724 ERR("Unknown relayd socket type (%d)", sock_type
);
3725 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3729 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3730 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3731 relayd
->net_seq_idx
, fd
);
3733 * We gave the ownership of the fd to the relayd structure. Set the
3734 * fd to -1 so we don't call close() on it in the error path below.
3738 /* We successfully added the socket. Send status back. */
3739 ret
= consumer_send_status_msg(sock
, ret_code
);
3741 /* Somehow, the session daemon is not responding anymore. */
3742 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3743 goto error_nosignal
;
3747 * Add relayd socket pair to consumer data hashtable. If object already
3748 * exists or on error, the function gracefully returns.
3757 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3758 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3762 /* Close received socket if valid. */
3765 PERROR("close received socket");
3769 if (relayd_created
) {
3775 * Search for a relayd associated to the session id and return the reference.
3777 * A rcu read side lock MUST be acquire before calling this function and locked
3778 * until the relayd object is no longer necessary.
3780 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3782 struct lttng_ht_iter iter
;
3783 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3785 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3786 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3789 * Check by sessiond id which is unique here where the relayd session
3790 * id might not be when having multiple relayd.
3792 if (relayd
->sessiond_session_id
== id
) {
3793 /* Found the relayd. There can be only one per id. */
3805 * Check if for a given session id there is still data needed to be extract
3808 * Return 1 if data is pending or else 0 meaning ready to be read.
3810 int consumer_data_pending(uint64_t id
)
3813 struct lttng_ht_iter iter
;
3814 struct lttng_ht
*ht
;
3815 struct lttng_consumer_stream
*stream
;
3816 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3817 int (*data_pending
)(struct lttng_consumer_stream
*);
3819 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3822 pthread_mutex_lock(&consumer_data
.lock
);
3824 switch (consumer_data
.type
) {
3825 case LTTNG_CONSUMER_KERNEL
:
3826 data_pending
= lttng_kconsumer_data_pending
;
3828 case LTTNG_CONSUMER32_UST
:
3829 case LTTNG_CONSUMER64_UST
:
3830 data_pending
= lttng_ustconsumer_data_pending
;
3833 ERR("Unknown consumer data type");
3837 /* Ease our life a bit */
3838 ht
= consumer_data
.stream_list_ht
;
3840 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3841 ht
->hash_fct(&id
, lttng_ht_seed
),
3843 &iter
.iter
, stream
, node_session_id
.node
) {
3844 pthread_mutex_lock(&stream
->lock
);
3847 * A removed node from the hash table indicates that the stream has
3848 * been deleted thus having a guarantee that the buffers are closed
3849 * on the consumer side. However, data can still be transmitted
3850 * over the network so don't skip the relayd check.
3852 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3854 /* Check the stream if there is data in the buffers. */
3855 ret
= data_pending(stream
);
3857 pthread_mutex_unlock(&stream
->lock
);
3862 pthread_mutex_unlock(&stream
->lock
);
3865 relayd
= find_relayd_by_session_id(id
);
3867 unsigned int is_data_inflight
= 0;
3869 /* Send init command for data pending. */
3870 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3871 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3872 relayd
->relayd_session_id
);
3874 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3875 /* Communication error thus the relayd so no data pending. */
3876 goto data_not_pending
;
3879 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3880 ht
->hash_fct(&id
, lttng_ht_seed
),
3882 &iter
.iter
, stream
, node_session_id
.node
) {
3883 if (stream
->metadata_flag
) {
3884 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3885 stream
->relayd_stream_id
);
3887 ret
= relayd_data_pending(&relayd
->control_sock
,
3888 stream
->relayd_stream_id
,
3889 stream
->next_net_seq_num
- 1);
3893 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3895 } else if (ret
< 0) {
3896 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3897 lttng_consumer_cleanup_relayd(relayd
);
3898 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3899 goto data_not_pending
;
3903 /* Send end command for data pending. */
3904 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3905 relayd
->relayd_session_id
, &is_data_inflight
);
3906 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3908 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3909 lttng_consumer_cleanup_relayd(relayd
);
3910 goto data_not_pending
;
3912 if (is_data_inflight
) {
3918 * Finding _no_ node in the hash table and no inflight data means that the
3919 * stream(s) have been removed thus data is guaranteed to be available for
3920 * analysis from the trace files.
3924 /* Data is available to be read by a viewer. */
3925 pthread_mutex_unlock(&consumer_data
.lock
);
3930 /* Data is still being extracted from buffers. */
3931 pthread_mutex_unlock(&consumer_data
.lock
);
3937 * Send a ret code status message to the sessiond daemon.
3939 * Return the sendmsg() return value.
3941 int consumer_send_status_msg(int sock
, int ret_code
)
3943 struct lttcomm_consumer_status_msg msg
;
3945 memset(&msg
, 0, sizeof(msg
));
3946 msg
.ret_code
= ret_code
;
3948 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3952 * Send a channel status message to the sessiond daemon.
3954 * Return the sendmsg() return value.
3956 int consumer_send_status_channel(int sock
,
3957 struct lttng_consumer_channel
*channel
)
3959 struct lttcomm_consumer_status_channel msg
;
3963 memset(&msg
, 0, sizeof(msg
));
3965 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3967 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3968 msg
.key
= channel
->key
;
3969 msg
.stream_count
= channel
->streams
.count
;
3972 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3975 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3976 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3977 uint64_t max_sb_size
)
3979 unsigned long start_pos
;
3981 if (!nb_packets_per_stream
) {
3982 return consumed_pos
; /* Grab everything */
3984 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3985 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3986 if ((long) (start_pos
- consumed_pos
) < 0) {
3987 return consumed_pos
; /* Grab everything */
3993 int consumer_flush_buffer(struct lttng_consumer_stream
*stream
, int producer_active
)
3997 switch (consumer_data
.type
) {
3998 case LTTNG_CONSUMER_KERNEL
:
3999 ret
= kernctl_buffer_flush(stream
->wait_fd
);
4001 ERR("Failed to flush kernel stream");
4005 case LTTNG_CONSUMER32_UST
:
4006 case LTTNG_CONSUMER64_UST
:
4007 lttng_ustctl_flush_buffer(stream
, producer_active
);
4010 ERR("Unknown consumer_data type");
4019 * Sample the rotate position for all the streams of a channel. If a stream
4020 * is already at the rotate position (produced == consumed), we flag it as
4021 * ready for rotation. The rotation of ready streams occurs after we have
4022 * replied to the session daemon that we have finished sampling the positions.
4023 * Must be called with RCU read-side lock held to ensure existence of channel.
4025 * Returns 0 on success, < 0 on error
4027 int lttng_consumer_rotate_channel(struct lttng_consumer_channel
*channel
,
4028 uint64_t key
, uint64_t relayd_id
, uint32_t metadata
,
4029 struct lttng_consumer_local_data
*ctx
)
4032 struct lttng_consumer_stream
*stream
;
4033 struct lttng_ht_iter iter
;
4034 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
4035 struct lttng_dynamic_array stream_rotation_positions
;
4036 uint64_t next_chunk_id
, stream_count
= 0;
4037 enum lttng_trace_chunk_status chunk_status
;
4038 const bool is_local_trace
= relayd_id
== -1ULL;
4039 struct consumer_relayd_sock_pair
*relayd
= NULL
;
4040 bool rotating_to_new_chunk
= true;
4042 DBG("Consumer sample rotate position for channel %" PRIu64
, key
);
4044 lttng_dynamic_array_init(&stream_rotation_positions
,
4045 sizeof(struct relayd_stream_rotation_position
), NULL
);
4049 pthread_mutex_lock(&channel
->lock
);
4050 assert(channel
->trace_chunk
);
4051 chunk_status
= lttng_trace_chunk_get_id(channel
->trace_chunk
,
4053 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4055 goto end_unlock_channel
;
4058 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4059 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4060 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4061 stream
, node_channel_id
.node
) {
4062 unsigned long consumed_pos
;
4064 health_code_update();
4067 * Lock stream because we are about to change its state.
4069 pthread_mutex_lock(&stream
->lock
);
4071 if (stream
->trace_chunk
== stream
->chan
->trace_chunk
) {
4072 rotating_to_new_chunk
= false;
4075 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4077 ERR("Failed to sample snapshot position during channel rotation");
4078 goto end_unlock_stream
;
4081 ret
= lttng_consumer_get_produced_snapshot(stream
,
4082 &stream
->rotate_position
);
4084 ERR("Failed to sample produced position during channel rotation");
4085 goto end_unlock_stream
;
4088 lttng_consumer_get_consumed_snapshot(stream
,
4090 if (consumed_pos
== stream
->rotate_position
) {
4091 stream
->rotate_ready
= true;
4095 * Active flush; has no effect if the production position
4096 * is at a packet boundary.
4098 ret
= consumer_flush_buffer(stream
, 1);
4100 ERR("Failed to flush stream %" PRIu64
" during channel rotation",
4102 goto end_unlock_stream
;
4105 if (!is_local_trace
) {
4107 * The relay daemon control protocol expects a rotation
4108 * position as "the sequence number of the first packet
4109 * _after_ the current trace chunk.
4111 * At the moment when the positions of the buffers are
4112 * sampled, the production position does not necessarily
4113 * sit at a packet boundary. The 'active' flush
4114 * operation above will push the production position to
4115 * the next packet boundary _if_ it is not already
4116 * sitting at such a boundary.
4118 * Assuming a current production position that is not
4119 * on the bound of a packet, the 'target' sequence
4121 * (consumed_pos / subbuffer_size) + 1
4122 * Note the '+ 1' to ensure the current packet is
4123 * part of the current trace chunk.
4125 * However, if the production position is already at
4126 * a packet boundary, the '+ 1' is not necessary as the
4127 * last packet of the current chunk is already
4130 const struct relayd_stream_rotation_position position
= {
4131 .stream_id
= stream
->relayd_stream_id
,
4132 .rotate_at_seq_num
= (stream
->rotate_position
/ stream
->max_sb_size
) +
4133 !!(stream
->rotate_position
% stream
->max_sb_size
),
4136 ret
= lttng_dynamic_array_add_element(
4137 &stream_rotation_positions
,
4140 ERR("Failed to allocate stream rotation position");
4141 goto end_unlock_stream
;
4145 pthread_mutex_unlock(&stream
->lock
);
4148 pthread_mutex_unlock(&channel
->lock
);
4150 if (is_local_trace
) {
4155 relayd
= consumer_find_relayd(relayd_id
);
4157 ERR("Failed to find relayd %" PRIu64
, relayd_id
);
4162 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4163 ret
= relayd_rotate_streams(&relayd
->control_sock
, stream_count
,
4164 rotating_to_new_chunk
? &next_chunk_id
: NULL
,
4165 (const struct relayd_stream_rotation_position
*)
4166 stream_rotation_positions
.buffer
.data
);
4167 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4169 ERR("Relayd rotate stream failed. Cleaning up relayd %" PRIu64
,
4170 relayd
->net_seq_idx
);
4171 lttng_consumer_cleanup_relayd(relayd
);
4179 pthread_mutex_unlock(&stream
->lock
);
4181 pthread_mutex_unlock(&channel
->lock
);
4184 lttng_dynamic_array_reset(&stream_rotation_positions
);
4189 * Check if a stream is ready to be rotated after extracting it.
4191 * Return 1 if it is ready for rotation, 0 if it is not, a negative value on
4192 * error. Stream lock must be held.
4194 int lttng_consumer_stream_is_rotate_ready(struct lttng_consumer_stream
*stream
)
4197 unsigned long consumed_pos
;
4199 if (!stream
->rotate_position
&& !stream
->rotate_ready
) {
4204 if (stream
->rotate_ready
) {
4210 * If we don't have the rotate_ready flag, check the consumed position
4211 * to determine if we need to rotate.
4213 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4215 ERR("Taking snapshot positions");
4219 ret
= lttng_consumer_get_consumed_snapshot(stream
, &consumed_pos
);
4221 ERR("Consumed snapshot position");
4225 /* Rotate position not reached yet (with check for overflow). */
4226 if ((long) (consumed_pos
- stream
->rotate_position
) < 0) {
4237 * Reset the state for a stream after a rotation occurred.
4239 void lttng_consumer_reset_stream_rotate_state(struct lttng_consumer_stream
*stream
)
4241 stream
->rotate_position
= 0;
4242 stream
->rotate_ready
= false;
4246 * Perform the rotation a local stream file.
4249 int rotate_local_stream(struct lttng_consumer_local_data
*ctx
,
4250 struct lttng_consumer_stream
*stream
)
4254 DBG("Rotate local stream: stream key %" PRIu64
", channel key %" PRIu64
,
4257 stream
->tracefile_size_current
= 0;
4258 stream
->tracefile_count_current
= 0;
4260 if (stream
->out_fd
>= 0) {
4261 ret
= close(stream
->out_fd
);
4263 PERROR("Failed to close stream out_fd of channel \"%s\"",
4264 stream
->chan
->name
);
4266 stream
->out_fd
= -1;
4269 if (stream
->index_file
) {
4270 lttng_index_file_put(stream
->index_file
);
4271 stream
->index_file
= NULL
;
4274 if (!stream
->trace_chunk
) {
4278 ret
= consumer_stream_create_output_files(stream
, true);
4284 * Performs the stream rotation for the rotate session feature if needed.
4285 * It must be called with the channel and stream locks held.
4287 * Return 0 on success, a negative number of error.
4289 int lttng_consumer_rotate_stream(struct lttng_consumer_local_data
*ctx
,
4290 struct lttng_consumer_stream
*stream
)
4294 DBG("Consumer rotate stream %" PRIu64
, stream
->key
);
4297 * Update the stream's 'current' chunk to the session's (channel)
4298 * now-current chunk.
4300 lttng_trace_chunk_put(stream
->trace_chunk
);
4301 if (stream
->chan
->trace_chunk
== stream
->trace_chunk
) {
4303 * A channel can be rotated and not have a "next" chunk
4304 * to transition to. In that case, the channel's "current chunk"
4305 * has not been closed yet, but it has not been updated to
4306 * a "next" trace chunk either. Hence, the stream, like its
4307 * parent channel, becomes part of no chunk and can't output
4308 * anything until a new trace chunk is created.
4310 stream
->trace_chunk
= NULL
;
4311 } else if (stream
->chan
->trace_chunk
&&
4312 !lttng_trace_chunk_get(stream
->chan
->trace_chunk
)) {
4313 ERR("Failed to acquire a reference to channel's trace chunk during stream rotation");
4318 * Update the stream's trace chunk to its parent channel's
4319 * current trace chunk.
4321 stream
->trace_chunk
= stream
->chan
->trace_chunk
;
4324 if (stream
->net_seq_idx
== (uint64_t) -1ULL) {
4325 ret
= rotate_local_stream(ctx
, stream
);
4327 ERR("Failed to rotate stream, ret = %i", ret
);
4332 if (stream
->metadata_flag
&& stream
->trace_chunk
) {
4334 * If the stream has transitioned to a new trace
4335 * chunk, the metadata should be re-dumped to the
4338 * However, it is possible for a stream to transition to
4339 * a "no-chunk" state. This can happen if a rotation
4340 * occurs on an inactive session. In such cases, the metadata
4341 * regeneration will happen when the next trace chunk is
4344 ret
= consumer_metadata_stream_dump(stream
);
4349 lttng_consumer_reset_stream_rotate_state(stream
);
4358 * Rotate all the ready streams now.
4360 * This is especially important for low throughput streams that have already
4361 * been consumed, we cannot wait for their next packet to perform the
4363 * Need to be called with RCU read-side lock held to ensure existence of
4366 * Returns 0 on success, < 0 on error
4368 int lttng_consumer_rotate_ready_streams(struct lttng_consumer_channel
*channel
,
4369 uint64_t key
, struct lttng_consumer_local_data
*ctx
)
4372 struct lttng_consumer_stream
*stream
;
4373 struct lttng_ht_iter iter
;
4374 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
4378 DBG("Consumer rotate ready streams in channel %" PRIu64
, key
);
4380 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4381 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4382 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4383 stream
, node_channel_id
.node
) {
4384 health_code_update();
4386 pthread_mutex_lock(&stream
->chan
->lock
);
4387 pthread_mutex_lock(&stream
->lock
);
4389 if (!stream
->rotate_ready
) {
4390 pthread_mutex_unlock(&stream
->lock
);
4391 pthread_mutex_unlock(&stream
->chan
->lock
);
4394 DBG("Consumer rotate ready stream %" PRIu64
, stream
->key
);
4396 ret
= lttng_consumer_rotate_stream(ctx
, stream
);
4397 pthread_mutex_unlock(&stream
->lock
);
4398 pthread_mutex_unlock(&stream
->chan
->lock
);
4411 enum lttcomm_return_code
lttng_consumer_init_command(
4412 struct lttng_consumer_local_data
*ctx
,
4413 const lttng_uuid sessiond_uuid
)
4415 enum lttcomm_return_code ret
;
4416 char uuid_str
[UUID_STR_LEN
];
4418 if (ctx
->sessiond_uuid
.is_set
) {
4419 ret
= LTTCOMM_CONSUMERD_ALREADY_SET
;
4423 ctx
->sessiond_uuid
.is_set
= true;
4424 memcpy(ctx
->sessiond_uuid
.value
, sessiond_uuid
, sizeof(lttng_uuid
));
4425 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4426 lttng_uuid_to_str(sessiond_uuid
, uuid_str
);
4427 DBG("Received session daemon UUID: %s", uuid_str
);
4432 enum lttcomm_return_code
lttng_consumer_create_trace_chunk(
4433 const uint64_t *relayd_id
, uint64_t session_id
,
4435 time_t chunk_creation_timestamp
,
4436 const char *chunk_override_name
,
4437 const struct lttng_credentials
*credentials
,
4438 struct lttng_directory_handle
*chunk_directory_handle
)
4441 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4442 struct lttng_trace_chunk
*created_chunk
, *published_chunk
;
4443 enum lttng_trace_chunk_status chunk_status
;
4444 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4445 char creation_timestamp_buffer
[ISO8601_STR_LEN
];
4446 const char *relayd_id_str
= "(none)";
4447 const char *creation_timestamp_str
;
4448 struct lttng_ht_iter iter
;
4449 struct lttng_consumer_channel
*channel
;
4452 /* Only used for logging purposes. */
4453 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4454 "%" PRIu64
, *relayd_id
);
4455 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4456 relayd_id_str
= relayd_id_buffer
;
4458 relayd_id_str
= "(formatting error)";
4462 /* Local protocol error. */
4463 assert(chunk_creation_timestamp
);
4464 ret
= time_to_iso8601_str(chunk_creation_timestamp
,
4465 creation_timestamp_buffer
,
4466 sizeof(creation_timestamp_buffer
));
4467 creation_timestamp_str
= !ret
? creation_timestamp_buffer
:
4468 "(formatting error)";
4470 DBG("Consumer create trace chunk command: relay_id = %s"
4471 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4472 ", chunk_override_name = %s"
4473 ", chunk_creation_timestamp = %s",
4474 relayd_id_str
, session_id
, chunk_id
,
4475 chunk_override_name
? : "(none)",
4476 creation_timestamp_str
);
4479 * The trace chunk registry, as used by the consumer daemon, implicitly
4480 * owns the trace chunks. This is only needed in the consumer since
4481 * the consumer has no notion of a session beyond session IDs being
4482 * used to identify other objects.
4484 * The lttng_trace_chunk_registry_publish() call below provides a
4485 * reference which is not released; it implicitly becomes the session
4486 * daemon's reference to the chunk in the consumer daemon.
4488 * The lifetime of trace chunks in the consumer daemon is managed by
4489 * the session daemon through the LTTNG_CONSUMER_CREATE_TRACE_CHUNK
4490 * and LTTNG_CONSUMER_DESTROY_TRACE_CHUNK commands.
4492 created_chunk
= lttng_trace_chunk_create(chunk_id
,
4493 chunk_creation_timestamp
);
4494 if (!created_chunk
) {
4495 ERR("Failed to create trace chunk");
4496 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4500 if (chunk_override_name
) {
4501 chunk_status
= lttng_trace_chunk_override_name(created_chunk
,
4502 chunk_override_name
);
4503 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4504 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4509 if (chunk_directory_handle
) {
4510 chunk_status
= lttng_trace_chunk_set_credentials(created_chunk
,
4512 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4513 ERR("Failed to set trace chunk credentials");
4514 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4518 * The consumer daemon has no ownership of the chunk output
4521 chunk_status
= lttng_trace_chunk_set_as_user(created_chunk
,
4522 chunk_directory_handle
);
4523 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4524 ERR("Failed to set trace chunk's directory handle");
4525 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4530 published_chunk
= lttng_trace_chunk_registry_publish_chunk(
4531 consumer_data
.chunk_registry
, session_id
,
4533 lttng_trace_chunk_put(created_chunk
);
4534 created_chunk
= NULL
;
4535 if (!published_chunk
) {
4536 ERR("Failed to publish trace chunk");
4537 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4542 cds_lfht_for_each_entry_duplicate(consumer_data
.channels_by_session_id_ht
->ht
,
4543 consumer_data
.channels_by_session_id_ht
->hash_fct(
4544 &session_id
, lttng_ht_seed
),
4545 consumer_data
.channels_by_session_id_ht
->match_fct
,
4546 &session_id
, &iter
.iter
, channel
,
4547 channels_by_session_id_ht_node
.node
) {
4548 ret
= lttng_consumer_channel_set_trace_chunk(channel
,
4552 * Roll-back the creation of this chunk.
4554 * This is important since the session daemon will
4555 * assume that the creation of this chunk failed and
4556 * will never ask for it to be closed, resulting
4557 * in a leak and an inconsistent state for some
4560 enum lttcomm_return_code close_ret
;
4562 DBG("Failed to set new trace chunk on existing channels, rolling back");
4563 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4564 session_id
, chunk_id
,
4565 chunk_creation_timestamp
, NULL
);
4566 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4567 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4568 session_id
, chunk_id
);
4571 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4577 struct consumer_relayd_sock_pair
*relayd
;
4579 relayd
= consumer_find_relayd(*relayd_id
);
4581 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4582 ret
= relayd_create_trace_chunk(
4583 &relayd
->control_sock
, published_chunk
);
4584 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4586 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
, *relayd_id
);
4589 if (!relayd
|| ret
) {
4590 enum lttcomm_return_code close_ret
;
4592 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4595 chunk_creation_timestamp
,
4597 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4598 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4603 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4609 /* Release the reference returned by the "publish" operation. */
4610 lttng_trace_chunk_put(published_chunk
);
4615 enum lttcomm_return_code
lttng_consumer_close_trace_chunk(
4616 const uint64_t *relayd_id
, uint64_t session_id
,
4617 uint64_t chunk_id
, time_t chunk_close_timestamp
,
4618 const enum lttng_trace_chunk_command_type
*close_command
)
4620 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4621 struct lttng_trace_chunk
*chunk
;
4622 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4623 const char *relayd_id_str
= "(none)";
4624 const char *close_command_name
= "none";
4625 struct lttng_ht_iter iter
;
4626 struct lttng_consumer_channel
*channel
;
4627 enum lttng_trace_chunk_status chunk_status
;
4632 /* Only used for logging purposes. */
4633 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4634 "%" PRIu64
, *relayd_id
);
4635 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4636 relayd_id_str
= relayd_id_buffer
;
4638 relayd_id_str
= "(formatting error)";
4641 if (close_command
) {
4642 close_command_name
= lttng_trace_chunk_command_type_get_name(
4646 DBG("Consumer close trace chunk command: relayd_id = %s"
4647 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4648 ", close command = %s",
4649 relayd_id_str
, session_id
, chunk_id
,
4650 close_command_name
);
4652 chunk
= lttng_trace_chunk_registry_find_chunk(
4653 consumer_data
.chunk_registry
, session_id
, chunk_id
);
4655 ERR("Failed to find chunk: session_id = %" PRIu64
4656 ", chunk_id = %" PRIu64
,
4657 session_id
, chunk_id
);
4658 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4662 chunk_status
= lttng_trace_chunk_set_close_timestamp(chunk
,
4663 chunk_close_timestamp
);
4664 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4665 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4669 if (close_command
) {
4670 chunk_status
= lttng_trace_chunk_set_close_command(
4671 chunk
, *close_command
);
4672 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4673 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4679 * chunk is now invalid to access as we no longer hold a reference to
4680 * it; it is only kept around to compare it (by address) to the
4681 * current chunk found in the session's channels.
4684 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
,
4685 channel
, node
.node
) {
4689 * Only change the channel's chunk to NULL if it still
4690 * references the chunk being closed. The channel may
4691 * reference a newer channel in the case of a session
4692 * rotation. When a session rotation occurs, the "next"
4693 * chunk is created before the "current" chunk is closed.
4695 if (channel
->trace_chunk
!= chunk
) {
4698 ret
= lttng_consumer_channel_set_trace_chunk(channel
, NULL
);
4701 * Attempt to close the chunk on as many channels as
4704 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4710 struct consumer_relayd_sock_pair
*relayd
;
4712 relayd
= consumer_find_relayd(*relayd_id
);
4714 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4715 ret
= relayd_close_trace_chunk(
4716 &relayd
->control_sock
, chunk
);
4717 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4719 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
,
4723 if (!relayd
|| ret
) {
4724 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4732 * Release the reference returned by the "find" operation and
4733 * the session daemon's implicit reference to the chunk.
4735 lttng_trace_chunk_put(chunk
);
4736 lttng_trace_chunk_put(chunk
);
4741 enum lttcomm_return_code
lttng_consumer_trace_chunk_exists(
4742 const uint64_t *relayd_id
, uint64_t session_id
,
4746 enum lttcomm_return_code ret_code
;
4747 struct lttng_trace_chunk
*chunk
;
4748 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4749 const char *relayd_id_str
= "(none)";
4750 const bool is_local_trace
= !relayd_id
;
4751 struct consumer_relayd_sock_pair
*relayd
= NULL
;
4752 bool chunk_exists_remote
;
4757 /* Only used for logging purposes. */
4758 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4759 "%" PRIu64
, *relayd_id
);
4760 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4761 relayd_id_str
= relayd_id_buffer
;
4763 relayd_id_str
= "(formatting error)";
4767 DBG("Consumer trace chunk exists command: relayd_id = %s"
4768 ", chunk_id = %" PRIu64
, relayd_id_str
,
4770 chunk
= lttng_trace_chunk_registry_find_chunk(
4771 consumer_data
.chunk_registry
, session_id
,
4773 DBG("Trace chunk %s locally", chunk
? "exists" : "does not exist");
4775 ret_code
= LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_LOCAL
;
4776 lttng_trace_chunk_put(chunk
);
4778 } else if (is_local_trace
) {
4779 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4784 relayd
= consumer_find_relayd(*relayd_id
);
4786 ERR("Failed to find relayd %" PRIu64
, *relayd_id
);
4787 ret_code
= LTTCOMM_CONSUMERD_INVALID_PARAMETERS
;
4788 goto end_rcu_unlock
;
4790 DBG("Looking up existence of trace chunk on relay daemon");
4791 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4792 ret
= relayd_trace_chunk_exists(&relayd
->control_sock
, chunk_id
,
4793 &chunk_exists_remote
);
4794 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4796 ERR("Failed to look-up the existence of trace chunk on relay daemon");
4797 ret_code
= LTTCOMM_CONSUMERD_RELAYD_FAIL
;
4798 goto end_rcu_unlock
;
4801 ret_code
= chunk_exists_remote
?
4802 LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_REMOTE
:
4803 LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4804 DBG("Trace chunk %s on relay daemon",
4805 chunk_exists_remote
? "exists" : "does not exist");