2 * Copyright (C) 2011 EfficiOS Inc.
3 * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * Copyright (C) 2012 David Goulet <dgoulet@efficios.com>
6 * SPDX-License-Identifier: GPL-2.0-only
18 #include <sys/socket.h>
19 #include <sys/types.h>
22 #include <bin/lttng-consumerd/health-consumerd.hpp>
23 #include <common/align.hpp>
24 #include <common/common.hpp>
25 #include <common/compat/endian.hpp>
26 #include <common/compat/poll.hpp>
27 #include <common/consumer/consumer-metadata-cache.hpp>
28 #include <common/consumer/consumer-stream.hpp>
29 #include <common/consumer/consumer-testpoint.hpp>
30 #include <common/consumer/consumer-timer.hpp>
31 #include <common/consumer/consumer.hpp>
32 #include <common/dynamic-array.hpp>
33 #include <common/index/ctf-index.hpp>
34 #include <common/index/index.hpp>
35 #include <common/kernel-consumer/kernel-consumer.hpp>
36 #include <common/kernel-ctl/kernel-ctl.hpp>
37 #include <common/relayd/relayd.hpp>
38 #include <common/sessiond-comm/relayd.hpp>
39 #include <common/sessiond-comm/sessiond-comm.hpp>
40 #include <common/string-utils/format.hpp>
41 #include <common/time.hpp>
42 #include <common/trace-chunk-registry.hpp>
43 #include <common/trace-chunk.hpp>
44 #include <common/ust-consumer/ust-consumer.hpp>
45 #include <common/utils.hpp>
47 lttng_consumer_global_data the_consumer_data
;
49 enum consumer_channel_action
{
52 CONSUMER_CHANNEL_QUIT
,
56 struct consumer_channel_msg
{
57 enum consumer_channel_action action
;
58 struct lttng_consumer_channel
*chan
; /* add */
59 uint64_t key
; /* del */
63 * Global hash table containing respectively metadata and data streams. The
64 * stream element in this ht should only be updated by the metadata poll thread
65 * for the metadata and the data poll thread for the data.
67 struct lttng_ht
*metadata_ht
;
68 struct lttng_ht
*data_ht
;
71 /* Flag used to temporarily pause data consumption from testpoints. */
72 int data_consumption_paused
;
75 * Flag to inform the polling thread to quit when all fd hung up. Updated by
76 * the consumer_thread_receive_fds when it notices that all fds has hung up.
77 * Also updated by the signal handler (consumer_should_exit()). Read by the
82 static const char *get_consumer_domain(void)
84 switch (the_consumer_data
.type
) {
85 case LTTNG_CONSUMER_KERNEL
:
86 return DEFAULT_KERNEL_TRACE_DIR
;
87 case LTTNG_CONSUMER64_UST
:
89 case LTTNG_CONSUMER32_UST
:
90 return DEFAULT_UST_TRACE_DIR
;
97 * Notify a thread lttng pipe to poll back again. This usually means that some
98 * global state has changed so we just send back the thread in a poll wait
101 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
103 struct lttng_consumer_stream
*null_stream
= NULL
;
107 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
110 static void notify_health_quit_pipe(int *pipe
)
114 ret
= lttng_write(pipe
[1], "4", 1);
116 PERROR("write consumer health quit");
120 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
121 struct lttng_consumer_channel
*chan
,
123 enum consumer_channel_action action
)
125 struct consumer_channel_msg msg
;
128 memset(&msg
, 0, sizeof(msg
));
133 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
134 if (ret
< sizeof(msg
)) {
135 PERROR("notify_channel_pipe write error");
139 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
142 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
145 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
146 struct lttng_consumer_channel
**chan
,
148 enum consumer_channel_action
*action
)
150 struct consumer_channel_msg msg
;
153 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
154 if (ret
< sizeof(msg
)) {
158 *action
= msg
.action
;
166 * Cleanup the stream list of a channel. Those streams are not yet globally
169 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
171 struct lttng_consumer_stream
*stream
, *stmp
;
173 LTTNG_ASSERT(channel
);
175 /* Delete streams that might have been left in the stream list. */
176 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
179 * Once a stream is added to this list, the buffers were created so we
180 * have a guarantee that this call will succeed. Setting the monitor
181 * mode to 0 so we don't lock nor try to delete the stream from the
185 consumer_stream_destroy(stream
, NULL
);
190 * Find a stream. The consumer_data.lock must be locked during this
193 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
196 struct lttng_ht_iter iter
;
197 struct lttng_ht_node_u64
*node
;
198 struct lttng_consumer_stream
*stream
= NULL
;
202 /* -1ULL keys are lookup failures */
203 if (key
== (uint64_t) -1ULL) {
209 lttng_ht_lookup(ht
, &key
, &iter
);
210 node
= lttng_ht_iter_get_node_u64(&iter
);
212 stream
= lttng::utils::container_of(node
, <tng_consumer_stream::node
);
220 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
222 struct lttng_consumer_stream
*stream
;
225 stream
= find_stream(key
, ht
);
227 stream
->key
= (uint64_t) -1ULL;
229 * We don't want the lookup to match, but we still need
230 * to iterate on this stream when iterating over the hash table. Just
231 * change the node key.
233 stream
->node
.key
= (uint64_t) -1ULL;
239 * Return a channel object for the given key.
241 * RCU read side lock MUST be acquired before calling this function and
242 * protects the channel ptr.
244 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
246 struct lttng_ht_iter iter
;
247 struct lttng_ht_node_u64
*node
;
248 struct lttng_consumer_channel
*channel
= NULL
;
250 ASSERT_RCU_READ_LOCKED();
252 /* -1ULL keys are lookup failures */
253 if (key
== (uint64_t) -1ULL) {
257 lttng_ht_lookup(the_consumer_data
.channel_ht
, &key
, &iter
);
258 node
= lttng_ht_iter_get_node_u64(&iter
);
260 channel
= lttng::utils::container_of(node
, <tng_consumer_channel::node
);
267 * There is a possibility that the consumer does not have enough time between
268 * the close of the channel on the session daemon and the cleanup in here thus
269 * once we have a channel add with an existing key, we know for sure that this
270 * channel will eventually get cleaned up by all streams being closed.
272 * This function just nullifies the already existing channel key.
274 static void steal_channel_key(uint64_t key
)
276 struct lttng_consumer_channel
*channel
;
279 channel
= consumer_find_channel(key
);
281 channel
->key
= (uint64_t) -1ULL;
283 * We don't want the lookup to match, but we still need to iterate on
284 * this channel when iterating over the hash table. Just change the
287 channel
->node
.key
= (uint64_t) -1ULL;
292 static void free_channel_rcu(struct rcu_head
*head
)
294 struct lttng_ht_node_u64
*node
=
295 lttng::utils::container_of(head
, <tng_ht_node_u64::head
);
296 struct lttng_consumer_channel
*channel
=
297 lttng::utils::container_of(node
, <tng_consumer_channel::node
);
299 switch (the_consumer_data
.type
) {
300 case LTTNG_CONSUMER_KERNEL
:
302 case LTTNG_CONSUMER32_UST
:
303 case LTTNG_CONSUMER64_UST
:
304 lttng_ustconsumer_free_channel(channel
);
307 ERR("Unknown consumer_data type");
314 * RCU protected relayd socket pair free.
316 static void free_relayd_rcu(struct rcu_head
*head
)
318 struct lttng_ht_node_u64
*node
=
319 lttng::utils::container_of(head
, <tng_ht_node_u64::head
);
320 struct consumer_relayd_sock_pair
*relayd
=
321 lttng::utils::container_of(node
, &consumer_relayd_sock_pair::node
);
324 * Close all sockets. This is done in the call RCU since we don't want the
325 * socket fds to be reassigned thus potentially creating bad state of the
328 * We do not have to lock the control socket mutex here since at this stage
329 * there is no one referencing to this relayd object.
331 (void) relayd_close(&relayd
->control_sock
);
332 (void) relayd_close(&relayd
->data_sock
);
334 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
339 * Destroy and free relayd socket pair object.
341 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
344 struct lttng_ht_iter iter
;
346 if (relayd
== NULL
) {
350 DBG("Consumer destroy and close relayd socket pair");
352 iter
.iter
.node
= &relayd
->node
.node
;
353 ret
= lttng_ht_del(the_consumer_data
.relayd_ht
, &iter
);
355 /* We assume the relayd is being or is destroyed */
359 /* RCU free() call */
360 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
364 * Remove a channel from the global list protected by a mutex. This function is
365 * also responsible for freeing its data structures.
367 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
369 struct lttng_ht_iter iter
;
371 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
373 pthread_mutex_lock(&the_consumer_data
.lock
);
374 pthread_mutex_lock(&channel
->lock
);
376 /* Destroy streams that might have been left in the stream list. */
377 clean_channel_stream_list(channel
);
379 if (channel
->live_timer_enabled
== 1) {
380 consumer_timer_live_stop(channel
);
382 if (channel
->monitor_timer_enabled
== 1) {
383 consumer_timer_monitor_stop(channel
);
387 * Send a last buffer statistics sample to the session daemon
388 * to ensure it tracks the amount of data consumed by this channel.
390 sample_and_send_channel_buffer_stats(channel
);
392 switch (the_consumer_data
.type
) {
393 case LTTNG_CONSUMER_KERNEL
:
395 case LTTNG_CONSUMER32_UST
:
396 case LTTNG_CONSUMER64_UST
:
397 lttng_ustconsumer_del_channel(channel
);
400 ERR("Unknown consumer_data type");
405 lttng_trace_chunk_put(channel
->trace_chunk
);
406 channel
->trace_chunk
= NULL
;
408 if (channel
->is_published
) {
412 iter
.iter
.node
= &channel
->node
.node
;
413 ret
= lttng_ht_del(the_consumer_data
.channel_ht
, &iter
);
416 iter
.iter
.node
= &channel
->channels_by_session_id_ht_node
.node
;
417 ret
= lttng_ht_del(the_consumer_data
.channels_by_session_id_ht
,
423 channel
->is_deleted
= true;
424 call_rcu(&channel
->node
.head
, free_channel_rcu
);
426 pthread_mutex_unlock(&channel
->lock
);
427 pthread_mutex_unlock(&the_consumer_data
.lock
);
431 * Iterate over the relayd hash table and destroy each element. Finally,
432 * destroy the whole hash table.
434 static void cleanup_relayd_ht(void)
436 struct lttng_ht_iter iter
;
437 struct consumer_relayd_sock_pair
*relayd
;
441 cds_lfht_for_each_entry(the_consumer_data
.relayd_ht
->ht
, &iter
.iter
,
443 consumer_destroy_relayd(relayd
);
448 lttng_ht_destroy(the_consumer_data
.relayd_ht
);
452 * Update the end point status of all streams having the given network sequence
453 * index (relayd index).
455 * It's atomically set without having the stream mutex locked which is fine
456 * because we handle the write/read race with a pipe wakeup for each thread.
458 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
459 enum consumer_endpoint_status status
)
461 struct lttng_ht_iter iter
;
462 struct lttng_consumer_stream
*stream
;
464 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
468 /* Let's begin with metadata */
469 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
470 if (stream
->net_seq_idx
== net_seq_idx
) {
471 uatomic_set(&stream
->endpoint_status
, status
);
472 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
476 /* Follow up by the data streams */
477 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
478 if (stream
->net_seq_idx
== net_seq_idx
) {
479 uatomic_set(&stream
->endpoint_status
, status
);
480 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
487 * Cleanup a relayd object by flagging every associated streams for deletion,
488 * destroying the object meaning removing it from the relayd hash table,
489 * closing the sockets and freeing the memory in a RCU call.
491 * If a local data context is available, notify the threads that the streams'
492 * state have changed.
494 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
498 LTTNG_ASSERT(relayd
);
500 DBG("Cleaning up relayd object ID %" PRIu64
, relayd
->net_seq_idx
);
502 /* Save the net sequence index before destroying the object */
503 netidx
= relayd
->net_seq_idx
;
506 * Delete the relayd from the relayd hash table, close the sockets and free
507 * the object in a RCU call.
509 consumer_destroy_relayd(relayd
);
511 /* Set inactive endpoint to all streams */
512 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
515 * With a local data context, notify the threads that the streams' state
516 * have changed. The write() action on the pipe acts as an "implicit"
517 * memory barrier ordering the updates of the end point status from the
518 * read of this status which happens AFTER receiving this notify.
520 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
521 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
525 * Flag a relayd socket pair for destruction. Destroy it if the refcount
528 * RCU read side lock MUST be aquired before calling this function.
530 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
532 LTTNG_ASSERT(relayd
);
533 ASSERT_RCU_READ_LOCKED();
535 /* Set destroy flag for this object */
536 uatomic_set(&relayd
->destroy_flag
, 1);
538 /* Destroy the relayd if refcount is 0 */
539 if (uatomic_read(&relayd
->refcount
) == 0) {
540 consumer_destroy_relayd(relayd
);
545 * Completly destroy stream from every visiable data structure and the given
548 * One this call returns, the stream object is not longer usable nor visible.
550 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
553 consumer_stream_destroy(stream
, ht
);
557 * XXX naming of del vs destroy is all mixed up.
559 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
561 consumer_stream_destroy(stream
, data_ht
);
564 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
566 consumer_stream_destroy(stream
, metadata_ht
);
569 void consumer_stream_update_channel_attributes(
570 struct lttng_consumer_stream
*stream
,
571 struct lttng_consumer_channel
*channel
)
573 stream
->channel_read_only_attributes
.tracefile_size
=
574 channel
->tracefile_size
;
578 * Add a stream to the global list protected by a mutex.
580 void consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
582 struct lttng_ht
*ht
= data_ht
;
584 LTTNG_ASSERT(stream
);
587 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
589 pthread_mutex_lock(&the_consumer_data
.lock
);
590 pthread_mutex_lock(&stream
->chan
->lock
);
591 pthread_mutex_lock(&stream
->chan
->timer_lock
);
592 pthread_mutex_lock(&stream
->lock
);
595 /* Steal stream identifier to avoid having streams with the same key */
596 steal_stream_key(stream
->key
, ht
);
598 lttng_ht_add_unique_u64(ht
, &stream
->node
);
600 lttng_ht_add_u64(the_consumer_data
.stream_per_chan_id_ht
,
601 &stream
->node_channel_id
);
604 * Add stream to the stream_list_ht of the consumer data. No need to steal
605 * the key since the HT does not use it and we allow to add redundant keys
608 lttng_ht_add_u64(the_consumer_data
.stream_list_ht
,
609 &stream
->node_session_id
);
612 * When nb_init_stream_left reaches 0, we don't need to trigger any action
613 * in terms of destroying the associated channel, because the action that
614 * causes the count to become 0 also causes a stream to be added. The
615 * channel deletion will thus be triggered by the following removal of this
618 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
619 /* Increment refcount before decrementing nb_init_stream_left */
621 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
624 /* Update consumer data once the node is inserted. */
625 the_consumer_data
.stream_count
++;
626 the_consumer_data
.need_update
= 1;
629 pthread_mutex_unlock(&stream
->lock
);
630 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
631 pthread_mutex_unlock(&stream
->chan
->lock
);
632 pthread_mutex_unlock(&the_consumer_data
.lock
);
636 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
637 * be acquired before calling this.
639 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
642 struct lttng_ht_node_u64
*node
;
643 struct lttng_ht_iter iter
;
645 LTTNG_ASSERT(relayd
);
646 ASSERT_RCU_READ_LOCKED();
648 lttng_ht_lookup(the_consumer_data
.relayd_ht
, &relayd
->net_seq_idx
,
650 node
= lttng_ht_iter_get_node_u64(&iter
);
654 lttng_ht_add_unique_u64(the_consumer_data
.relayd_ht
, &relayd
->node
);
661 * Allocate and return a consumer relayd socket.
663 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
664 uint64_t net_seq_idx
)
666 struct consumer_relayd_sock_pair
*obj
= NULL
;
668 /* net sequence index of -1 is a failure */
669 if (net_seq_idx
== (uint64_t) -1ULL) {
673 obj
= zmalloc
<consumer_relayd_sock_pair
>();
675 PERROR("zmalloc relayd sock");
679 obj
->net_seq_idx
= net_seq_idx
;
681 obj
->destroy_flag
= 0;
682 obj
->control_sock
.sock
.fd
= -1;
683 obj
->data_sock
.sock
.fd
= -1;
684 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
685 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
692 * Find a relayd socket pair in the global consumer data.
694 * Return the object if found else NULL.
695 * RCU read-side lock must be held across this call and while using the
698 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
700 struct lttng_ht_iter iter
;
701 struct lttng_ht_node_u64
*node
;
702 struct consumer_relayd_sock_pair
*relayd
= NULL
;
704 ASSERT_RCU_READ_LOCKED();
706 /* Negative keys are lookup failures */
707 if (key
== (uint64_t) -1ULL) {
711 lttng_ht_lookup(the_consumer_data
.relayd_ht
, &key
, &iter
);
712 node
= lttng_ht_iter_get_node_u64(&iter
);
714 relayd
= lttng::utils::container_of(node
, &consumer_relayd_sock_pair::node
);
722 * Find a relayd and send the stream
724 * Returns 0 on success, < 0 on error
726 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
730 struct consumer_relayd_sock_pair
*relayd
;
732 LTTNG_ASSERT(stream
);
733 LTTNG_ASSERT(stream
->net_seq_idx
!= -1ULL);
736 /* The stream is not metadata. Get relayd reference if exists. */
738 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
739 if (relayd
!= NULL
) {
740 /* Add stream on the relayd */
741 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
742 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
743 get_consumer_domain(), path
, &stream
->relayd_stream_id
,
744 stream
->chan
->tracefile_size
,
745 stream
->chan
->tracefile_count
,
746 stream
->trace_chunk
);
747 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
749 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
750 lttng_consumer_cleanup_relayd(relayd
);
754 uatomic_inc(&relayd
->refcount
);
755 stream
->sent_to_relayd
= 1;
757 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
758 stream
->key
, stream
->net_seq_idx
);
763 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
764 stream
->name
, stream
->key
, stream
->net_seq_idx
);
772 * Find a relayd and send the streams sent message
774 * Returns 0 on success, < 0 on error
776 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
779 struct consumer_relayd_sock_pair
*relayd
;
781 LTTNG_ASSERT(net_seq_idx
!= -1ULL);
783 /* The stream is not metadata. Get relayd reference if exists. */
785 relayd
= consumer_find_relayd(net_seq_idx
);
786 if (relayd
!= NULL
) {
787 /* Add stream on the relayd */
788 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
789 ret
= relayd_streams_sent(&relayd
->control_sock
);
790 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
792 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
793 lttng_consumer_cleanup_relayd(relayd
);
797 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
804 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
812 * Find a relayd and close the stream
814 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
816 struct consumer_relayd_sock_pair
*relayd
;
818 /* The stream is not metadata. Get relayd reference if exists. */
820 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
822 consumer_stream_relayd_close(stream
, relayd
);
828 * Handle stream for relayd transmission if the stream applies for network
829 * streaming where the net sequence index is set.
831 * Return destination file descriptor or negative value on error.
833 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
834 size_t data_size
, unsigned long padding
,
835 struct consumer_relayd_sock_pair
*relayd
)
838 struct lttcomm_relayd_data_hdr data_hdr
;
841 LTTNG_ASSERT(stream
);
842 LTTNG_ASSERT(relayd
);
844 /* Reset data header */
845 memset(&data_hdr
, 0, sizeof(data_hdr
));
847 if (stream
->metadata_flag
) {
848 /* Caller MUST acquire the relayd control socket lock */
849 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
854 /* Metadata are always sent on the control socket. */
855 outfd
= relayd
->control_sock
.sock
.fd
;
857 /* Set header with stream information */
858 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
859 data_hdr
.data_size
= htobe32(data_size
);
860 data_hdr
.padding_size
= htobe32(padding
);
863 * Note that net_seq_num below is assigned with the *current* value of
864 * next_net_seq_num and only after that the next_net_seq_num will be
865 * increment. This is why when issuing a command on the relayd using
866 * this next value, 1 should always be substracted in order to compare
867 * the last seen sequence number on the relayd side to the last sent.
869 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
870 /* Other fields are zeroed previously */
872 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
878 ++stream
->next_net_seq_num
;
880 /* Set to go on data socket */
881 outfd
= relayd
->data_sock
.sock
.fd
;
889 * Write a character on the metadata poll pipe to wake the metadata thread.
890 * Returns 0 on success, -1 on error.
892 int consumer_metadata_wakeup_pipe(const struct lttng_consumer_channel
*channel
)
896 DBG("Waking up metadata poll thread (writing to pipe): channel name = '%s'",
898 if (channel
->monitor
&& channel
->metadata_stream
) {
899 const char dummy
= 'c';
900 const ssize_t write_ret
= lttng_write(
901 channel
->metadata_stream
->ust_metadata_poll_pipe
[1],
905 if (errno
== EWOULDBLOCK
) {
907 * This is fine, the metadata poll thread
908 * is having a hard time keeping-up, but
909 * it will eventually wake-up and consume
910 * the available data.
914 PERROR("Failed to write to UST metadata pipe while attempting to wake-up the metadata poll thread");
926 * Trigger a dump of the metadata content. Following/during the succesful
927 * completion of this call, the metadata poll thread will start receiving
928 * metadata packets to consume.
930 * The caller must hold the channel and stream locks.
933 int consumer_metadata_stream_dump(struct lttng_consumer_stream
*stream
)
937 ASSERT_LOCKED(stream
->chan
->lock
);
938 ASSERT_LOCKED(stream
->lock
);
939 LTTNG_ASSERT(stream
->metadata_flag
);
940 LTTNG_ASSERT(stream
->chan
->trace_chunk
);
942 switch (the_consumer_data
.type
) {
943 case LTTNG_CONSUMER_KERNEL
:
945 * Reset the position of what has been read from the
946 * metadata cache to 0 so we can dump it again.
948 ret
= kernctl_metadata_cache_dump(stream
->wait_fd
);
950 case LTTNG_CONSUMER32_UST
:
951 case LTTNG_CONSUMER64_UST
:
953 * Reset the position pushed from the metadata cache so it
954 * will write from the beginning on the next push.
956 stream
->ust_metadata_pushed
= 0;
957 ret
= consumer_metadata_wakeup_pipe(stream
->chan
);
960 ERR("Unknown consumer_data type");
964 ERR("Failed to dump the metadata cache");
970 int lttng_consumer_channel_set_trace_chunk(
971 struct lttng_consumer_channel
*channel
,
972 struct lttng_trace_chunk
*new_trace_chunk
)
974 pthread_mutex_lock(&channel
->lock
);
975 if (channel
->is_deleted
) {
977 * The channel has been logically deleted and should no longer
978 * be used. It has released its reference to its current trace
979 * chunk and should not acquire a new one.
981 * Return success as there is nothing for the caller to do.
987 * The acquisition of the reference cannot fail (barring
988 * a severe internal error) since a reference to the published
989 * chunk is already held by the caller.
991 if (new_trace_chunk
) {
992 const bool acquired_reference
= lttng_trace_chunk_get(
995 LTTNG_ASSERT(acquired_reference
);
998 lttng_trace_chunk_put(channel
->trace_chunk
);
999 channel
->trace_chunk
= new_trace_chunk
;
1001 pthread_mutex_unlock(&channel
->lock
);
1006 * Allocate and return a new lttng_consumer_channel object using the given key
1007 * to initialize the hash table node.
1009 * On error, return NULL.
1011 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
1012 uint64_t session_id
,
1013 const uint64_t *chunk_id
,
1014 const char *pathname
,
1017 enum lttng_event_output output
,
1018 uint64_t tracefile_size
,
1019 uint64_t tracefile_count
,
1020 uint64_t session_id_per_pid
,
1021 unsigned int monitor
,
1022 unsigned int live_timer_interval
,
1023 bool is_in_live_session
,
1024 const char *root_shm_path
,
1025 const char *shm_path
)
1027 struct lttng_consumer_channel
*channel
= NULL
;
1028 struct lttng_trace_chunk
*trace_chunk
= NULL
;
1031 trace_chunk
= lttng_trace_chunk_registry_find_chunk(
1032 the_consumer_data
.chunk_registry
, session_id
,
1035 ERR("Failed to find trace chunk reference during creation of channel");
1040 channel
= zmalloc
<lttng_consumer_channel
>();
1041 if (channel
== NULL
) {
1042 PERROR("malloc struct lttng_consumer_channel");
1047 channel
->refcount
= 0;
1048 channel
->session_id
= session_id
;
1049 channel
->session_id_per_pid
= session_id_per_pid
;
1050 channel
->relayd_id
= relayd_id
;
1051 channel
->tracefile_size
= tracefile_size
;
1052 channel
->tracefile_count
= tracefile_count
;
1053 channel
->monitor
= monitor
;
1054 channel
->live_timer_interval
= live_timer_interval
;
1055 channel
->is_live
= is_in_live_session
;
1056 pthread_mutex_init(&channel
->lock
, NULL
);
1057 pthread_mutex_init(&channel
->timer_lock
, NULL
);
1060 case LTTNG_EVENT_SPLICE
:
1061 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
1063 case LTTNG_EVENT_MMAP
:
1064 channel
->output
= CONSUMER_CHANNEL_MMAP
;
1074 * In monitor mode, the streams associated with the channel will be put in
1075 * a special list ONLY owned by this channel. So, the refcount is set to 1
1076 * here meaning that the channel itself has streams that are referenced.
1078 * On a channel deletion, once the channel is no longer visible, the
1079 * refcount is decremented and checked for a zero value to delete it. With
1080 * streams in no monitor mode, it will now be safe to destroy the channel.
1082 if (!channel
->monitor
) {
1083 channel
->refcount
= 1;
1086 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1087 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1089 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1090 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1092 if (root_shm_path
) {
1093 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1094 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1097 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1098 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1101 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1102 lttng_ht_node_init_u64(&channel
->channels_by_session_id_ht_node
,
1103 channel
->session_id
);
1105 channel
->wait_fd
= -1;
1106 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1109 int ret
= lttng_consumer_channel_set_trace_chunk(channel
,
1116 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1119 lttng_trace_chunk_put(trace_chunk
);
1122 consumer_del_channel(channel
);
1128 * Add a channel to the global list protected by a mutex.
1130 * Always return 0 indicating success.
1132 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1133 struct lttng_consumer_local_data
*ctx
)
1135 pthread_mutex_lock(&the_consumer_data
.lock
);
1136 pthread_mutex_lock(&channel
->lock
);
1137 pthread_mutex_lock(&channel
->timer_lock
);
1140 * This gives us a guarantee that the channel we are about to add to the
1141 * channel hash table will be unique. See this function comment on the why
1142 * we need to steel the channel key at this stage.
1144 steal_channel_key(channel
->key
);
1147 lttng_ht_add_unique_u64(the_consumer_data
.channel_ht
, &channel
->node
);
1148 lttng_ht_add_u64(the_consumer_data
.channels_by_session_id_ht
,
1149 &channel
->channels_by_session_id_ht_node
);
1151 channel
->is_published
= true;
1153 pthread_mutex_unlock(&channel
->timer_lock
);
1154 pthread_mutex_unlock(&channel
->lock
);
1155 pthread_mutex_unlock(&the_consumer_data
.lock
);
1157 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1158 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1165 * Allocate the pollfd structure and the local view of the out fds to avoid
1166 * doing a lookup in the linked list and concurrency issues when writing is
1167 * needed. Called with consumer_data.lock held.
1169 * Returns the number of fds in the structures.
1171 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1172 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1173 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1176 struct lttng_ht_iter iter
;
1177 struct lttng_consumer_stream
*stream
;
1181 LTTNG_ASSERT(pollfd
);
1182 LTTNG_ASSERT(local_stream
);
1184 DBG("Updating poll fd array");
1185 *nb_inactive_fd
= 0;
1187 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1189 * Only active streams with an active end point can be added to the
1190 * poll set and local stream storage of the thread.
1192 * There is a potential race here for endpoint_status to be updated
1193 * just after the check. However, this is OK since the stream(s) will
1194 * be deleted once the thread is notified that the end point state has
1195 * changed where this function will be called back again.
1197 * We track the number of inactive FDs because they still need to be
1198 * closed by the polling thread after a wakeup on the data_pipe or
1201 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1202 (*nb_inactive_fd
)++;
1206 * This clobbers way too much the debug output. Uncomment that if you
1207 * need it for debugging purposes.
1209 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1210 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1211 local_stream
[i
] = stream
;
1217 * Insert the consumer_data_pipe at the end of the array and don't
1218 * increment i so nb_fd is the number of real FD.
1220 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1221 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1223 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1224 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1229 * Poll on the should_quit pipe and the command socket return -1 on
1230 * error, 1 if should exit, 0 if data is available on the command socket
1232 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1237 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1238 if (num_rdy
== -1) {
1240 * Restart interrupted system call.
1242 if (errno
== EINTR
) {
1245 PERROR("Poll error");
1248 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1249 DBG("consumer_should_quit wake up");
1256 * Set the error socket.
1258 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1261 ctx
->consumer_error_socket
= sock
;
1265 * Set the command socket path.
1267 void lttng_consumer_set_command_sock_path(
1268 struct lttng_consumer_local_data
*ctx
, char *sock
)
1270 ctx
->consumer_command_sock_path
= sock
;
1274 * Send return code to the session daemon.
1275 * If the socket is not defined, we return 0, it is not a fatal error
1277 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1279 if (ctx
->consumer_error_socket
> 0) {
1280 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1281 sizeof(enum lttcomm_sessiond_command
));
1288 * Close all the tracefiles and stream fds and MUST be called when all
1289 * instances are destroyed i.e. when all threads were joined and are ended.
1291 void lttng_consumer_cleanup(void)
1293 struct lttng_ht_iter iter
;
1294 struct lttng_consumer_channel
*channel
;
1295 unsigned int trace_chunks_left
;
1299 cds_lfht_for_each_entry(the_consumer_data
.channel_ht
->ht
, &iter
.iter
,
1300 channel
, node
.node
) {
1301 consumer_del_channel(channel
);
1306 lttng_ht_destroy(the_consumer_data
.channel_ht
);
1307 lttng_ht_destroy(the_consumer_data
.channels_by_session_id_ht
);
1309 cleanup_relayd_ht();
1311 lttng_ht_destroy(the_consumer_data
.stream_per_chan_id_ht
);
1314 * This HT contains streams that are freed by either the metadata thread or
1315 * the data thread so we do *nothing* on the hash table and simply destroy
1318 lttng_ht_destroy(the_consumer_data
.stream_list_ht
);
1321 * Trace chunks in the registry may still exist if the session
1322 * daemon has encountered an internal error and could not
1323 * tear down its sessions and/or trace chunks properly.
1325 * Release the session daemon's implicit reference to any remaining
1326 * trace chunk and print an error if any trace chunk was found. Note
1327 * that there are _no_ legitimate cases for trace chunks to be left,
1328 * it is a leak. However, it can happen following a crash of the
1329 * session daemon and not emptying the registry would cause an assertion
1332 trace_chunks_left
= lttng_trace_chunk_registry_put_each_chunk(
1333 the_consumer_data
.chunk_registry
);
1334 if (trace_chunks_left
) {
1335 ERR("%u trace chunks are leaked by lttng-consumerd. "
1336 "This can be caused by an internal error of the session daemon.",
1339 /* Run all callbacks freeing each chunk. */
1341 lttng_trace_chunk_registry_destroy(the_consumer_data
.chunk_registry
);
1345 * Called from signal handler.
1347 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1351 CMM_STORE_SHARED(consumer_quit
, 1);
1352 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1354 PERROR("write consumer quit");
1357 DBG("Consumer flag that it should quit");
1362 * Flush pending writes to trace output disk file.
1365 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1369 int outfd
= stream
->out_fd
;
1372 * This does a blocking write-and-wait on any page that belongs to the
1373 * subbuffer prior to the one we just wrote.
1374 * Don't care about error values, as these are just hints and ways to
1375 * limit the amount of page cache used.
1377 if (orig_offset
< stream
->max_sb_size
) {
1380 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1381 stream
->max_sb_size
,
1382 SYNC_FILE_RANGE_WAIT_BEFORE
1383 | SYNC_FILE_RANGE_WRITE
1384 | SYNC_FILE_RANGE_WAIT_AFTER
);
1386 * Give hints to the kernel about how we access the file:
1387 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1390 * We need to call fadvise again after the file grows because the
1391 * kernel does not seem to apply fadvise to non-existing parts of the
1394 * Call fadvise _after_ having waited for the page writeback to
1395 * complete because the dirty page writeback semantic is not well
1396 * defined. So it can be expected to lead to lower throughput in
1399 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1400 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1401 if (ret
&& ret
!= -ENOSYS
) {
1403 PERROR("posix_fadvise on fd %i", outfd
);
1408 * Initialise the necessary environnement :
1409 * - create a new context
1410 * - create the poll_pipe
1411 * - create the should_quit pipe (for signal handler)
1412 * - create the thread pipe (for splice)
1414 * Takes a function pointer as argument, this function is called when data is
1415 * available on a buffer. This function is responsible to do the
1416 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1417 * buffer configuration and then kernctl_put_next_subbuf at the end.
1419 * Returns a pointer to the new context or NULL on error.
1421 struct lttng_consumer_local_data
*lttng_consumer_create(
1422 enum lttng_consumer_type type
,
1423 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1424 struct lttng_consumer_local_data
*ctx
, bool locked_by_caller
),
1425 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1426 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1427 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1430 struct lttng_consumer_local_data
*ctx
;
1432 LTTNG_ASSERT(the_consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1433 the_consumer_data
.type
== type
);
1434 the_consumer_data
.type
= type
;
1436 ctx
= zmalloc
<lttng_consumer_local_data
>();
1438 PERROR("allocating context");
1442 ctx
->consumer_error_socket
= -1;
1443 ctx
->consumer_metadata_socket
= -1;
1444 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1445 /* assign the callbacks */
1446 ctx
->on_buffer_ready
= buffer_ready
;
1447 ctx
->on_recv_channel
= recv_channel
;
1448 ctx
->on_recv_stream
= recv_stream
;
1449 ctx
->on_update_stream
= update_stream
;
1451 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1452 if (!ctx
->consumer_data_pipe
) {
1453 goto error_poll_pipe
;
1456 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1457 if (!ctx
->consumer_wakeup_pipe
) {
1458 goto error_wakeup_pipe
;
1461 ret
= pipe(ctx
->consumer_should_quit
);
1463 PERROR("Error creating recv pipe");
1464 goto error_quit_pipe
;
1467 ret
= pipe(ctx
->consumer_channel_pipe
);
1469 PERROR("Error creating channel pipe");
1470 goto error_channel_pipe
;
1473 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1474 if (!ctx
->consumer_metadata_pipe
) {
1475 goto error_metadata_pipe
;
1478 ctx
->channel_monitor_pipe
= -1;
1482 error_metadata_pipe
:
1483 utils_close_pipe(ctx
->consumer_channel_pipe
);
1485 utils_close_pipe(ctx
->consumer_should_quit
);
1487 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1489 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1497 * Iterate over all streams of the hashtable and free them properly.
1499 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1501 struct lttng_ht_iter iter
;
1502 struct lttng_consumer_stream
*stream
;
1509 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1511 * Ignore return value since we are currently cleaning up so any error
1514 (void) consumer_del_stream(stream
, ht
);
1518 lttng_ht_destroy(ht
);
1522 * Iterate over all streams of the metadata hashtable and free them
1525 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1527 struct lttng_ht_iter iter
;
1528 struct lttng_consumer_stream
*stream
;
1535 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1537 * Ignore return value since we are currently cleaning up so any error
1540 (void) consumer_del_metadata_stream(stream
, ht
);
1544 lttng_ht_destroy(ht
);
1548 * Close all fds associated with the instance and free the context.
1550 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1554 DBG("Consumer destroying it. Closing everything.");
1560 destroy_data_stream_ht(data_ht
);
1561 destroy_metadata_stream_ht(metadata_ht
);
1563 ret
= close(ctx
->consumer_error_socket
);
1567 ret
= close(ctx
->consumer_metadata_socket
);
1571 utils_close_pipe(ctx
->consumer_channel_pipe
);
1572 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1573 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1574 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1575 utils_close_pipe(ctx
->consumer_should_quit
);
1577 unlink(ctx
->consumer_command_sock_path
);
1582 * Write the metadata stream id on the specified file descriptor.
1584 static int write_relayd_metadata_id(int fd
,
1585 struct lttng_consumer_stream
*stream
,
1586 unsigned long padding
)
1589 struct lttcomm_relayd_metadata_payload hdr
;
1591 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1592 hdr
.padding_size
= htobe32(padding
);
1593 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1594 if (ret
< sizeof(hdr
)) {
1596 * This error means that the fd's end is closed so ignore the PERROR
1597 * not to clubber the error output since this can happen in a normal
1600 if (errno
!= EPIPE
) {
1601 PERROR("write metadata stream id");
1603 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1605 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1606 * handle writting the missing part so report that as an error and
1607 * don't lie to the caller.
1612 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1613 stream
->relayd_stream_id
, padding
);
1620 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1621 * core function for writing trace buffers to either the local filesystem or
1624 * It must be called with the stream and the channel lock held.
1626 * Careful review MUST be put if any changes occur!
1628 * Returns the number of bytes written
1630 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1631 struct lttng_consumer_stream
*stream
,
1632 const struct lttng_buffer_view
*buffer
,
1633 unsigned long padding
)
1636 off_t orig_offset
= stream
->out_fd_offset
;
1637 /* Default is on the disk */
1638 int outfd
= stream
->out_fd
;
1639 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1640 unsigned int relayd_hang_up
= 0;
1641 const size_t subbuf_content_size
= buffer
->size
- padding
;
1644 /* RCU lock for the relayd pointer */
1646 LTTNG_ASSERT(stream
->net_seq_idx
!= (uint64_t) -1ULL ||
1647 stream
->trace_chunk
);
1649 /* Flag that the current stream if set for network streaming. */
1650 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1651 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1652 if (relayd
== NULL
) {
1658 /* Handle stream on the relayd if the output is on the network */
1660 unsigned long netlen
= subbuf_content_size
;
1663 * Lock the control socket for the complete duration of the function
1664 * since from this point on we will use the socket.
1666 if (stream
->metadata_flag
) {
1667 /* Metadata requires the control socket. */
1668 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1669 if (stream
->reset_metadata_flag
) {
1670 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1671 stream
->relayd_stream_id
,
1672 stream
->metadata_version
);
1677 stream
->reset_metadata_flag
= 0;
1679 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1682 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1687 /* Use the returned socket. */
1690 /* Write metadata stream id before payload */
1691 if (stream
->metadata_flag
) {
1692 ret
= write_relayd_metadata_id(outfd
, stream
, padding
);
1699 write_len
= subbuf_content_size
;
1701 /* No streaming; we have to write the full padding. */
1702 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1703 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1705 ERR("Reset metadata file");
1708 stream
->reset_metadata_flag
= 0;
1712 * Check if we need to change the tracefile before writing the packet.
1714 if (stream
->chan
->tracefile_size
> 0 &&
1715 (stream
->tracefile_size_current
+ buffer
->size
) >
1716 stream
->chan
->tracefile_size
) {
1717 ret
= consumer_stream_rotate_output_files(stream
);
1721 outfd
= stream
->out_fd
;
1724 stream
->tracefile_size_current
+= buffer
->size
;
1725 write_len
= buffer
->size
;
1729 * This call guarantee that len or less is returned. It's impossible to
1730 * receive a ret value that is bigger than len.
1732 ret
= lttng_write(outfd
, buffer
->data
, write_len
);
1733 DBG("Consumer mmap write() ret %zd (len %zu)", ret
, write_len
);
1734 if (ret
< 0 || ((size_t) ret
!= write_len
)) {
1736 * Report error to caller if nothing was written else at least send the
1744 /* Socket operation failed. We consider the relayd dead */
1745 if (errno
== EPIPE
) {
1747 * This is possible if the fd is closed on the other side
1748 * (outfd) or any write problem. It can be verbose a bit for a
1749 * normal execution if for instance the relayd is stopped
1750 * abruptly. This can happen so set this to a DBG statement.
1752 DBG("Consumer mmap write detected relayd hang up");
1754 /* Unhandled error, print it and stop function right now. */
1755 PERROR("Error in write mmap (ret %zd != write_len %zu)", ret
,
1760 stream
->output_written
+= ret
;
1762 /* This call is useless on a socket so better save a syscall. */
1764 /* This won't block, but will start writeout asynchronously */
1765 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, write_len
,
1766 SYNC_FILE_RANGE_WRITE
);
1767 stream
->out_fd_offset
+= write_len
;
1768 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1773 * This is a special case that the relayd has closed its socket. Let's
1774 * cleanup the relayd object and all associated streams.
1776 if (relayd
&& relayd_hang_up
) {
1777 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1778 lttng_consumer_cleanup_relayd(relayd
);
1782 /* Unlock only if ctrl socket used */
1783 if (relayd
&& stream
->metadata_flag
) {
1784 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1792 * Splice the data from the ring buffer to the tracefile.
1794 * It must be called with the stream lock held.
1796 * Returns the number of bytes spliced.
1798 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1799 struct lttng_consumer_local_data
*ctx
,
1800 struct lttng_consumer_stream
*stream
, unsigned long len
,
1801 unsigned long padding
)
1803 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1805 off_t orig_offset
= stream
->out_fd_offset
;
1806 int fd
= stream
->wait_fd
;
1807 /* Default is on the disk */
1808 int outfd
= stream
->out_fd
;
1809 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1811 unsigned int relayd_hang_up
= 0;
1813 switch (the_consumer_data
.type
) {
1814 case LTTNG_CONSUMER_KERNEL
:
1816 case LTTNG_CONSUMER32_UST
:
1817 case LTTNG_CONSUMER64_UST
:
1818 /* Not supported for user space tracing */
1821 ERR("Unknown consumer_data type");
1825 /* RCU lock for the relayd pointer */
1828 /* Flag that the current stream if set for network streaming. */
1829 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1830 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1831 if (relayd
== NULL
) {
1836 splice_pipe
= stream
->splice_pipe
;
1838 /* Write metadata stream id before payload */
1840 unsigned long total_len
= len
;
1842 if (stream
->metadata_flag
) {
1844 * Lock the control socket for the complete duration of the function
1845 * since from this point on we will use the socket.
1847 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1849 if (stream
->reset_metadata_flag
) {
1850 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1851 stream
->relayd_stream_id
,
1852 stream
->metadata_version
);
1857 stream
->reset_metadata_flag
= 0;
1859 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
,
1867 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1870 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1876 /* Use the returned socket. */
1879 /* No streaming, we have to set the len with the full padding */
1882 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1883 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1885 ERR("Reset metadata file");
1888 stream
->reset_metadata_flag
= 0;
1891 * Check if we need to change the tracefile before writing the packet.
1893 if (stream
->chan
->tracefile_size
> 0 &&
1894 (stream
->tracefile_size_current
+ len
) >
1895 stream
->chan
->tracefile_size
) {
1896 ret
= consumer_stream_rotate_output_files(stream
);
1901 outfd
= stream
->out_fd
;
1904 stream
->tracefile_size_current
+= len
;
1908 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1909 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1910 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1911 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1912 DBG("splice chan to pipe, ret %zd", ret_splice
);
1913 if (ret_splice
< 0) {
1916 PERROR("Error in relay splice");
1920 /* Handle stream on the relayd if the output is on the network */
1921 if (relayd
&& stream
->metadata_flag
) {
1922 size_t metadata_payload_size
=
1923 sizeof(struct lttcomm_relayd_metadata_payload
);
1925 /* Update counter to fit the spliced data */
1926 ret_splice
+= metadata_payload_size
;
1927 len
+= metadata_payload_size
;
1929 * We do this so the return value can match the len passed as
1930 * argument to this function.
1932 written
-= metadata_payload_size
;
1935 /* Splice data out */
1936 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1937 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1938 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1940 if (ret_splice
< 0) {
1945 } else if (ret_splice
> len
) {
1947 * We don't expect this code path to be executed but you never know
1948 * so this is an extra protection agains a buggy splice().
1951 written
+= ret_splice
;
1952 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1956 /* All good, update current len and continue. */
1960 /* This call is useless on a socket so better save a syscall. */
1962 /* This won't block, but will start writeout asynchronously */
1963 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1964 SYNC_FILE_RANGE_WRITE
);
1965 stream
->out_fd_offset
+= ret_splice
;
1967 stream
->output_written
+= ret_splice
;
1968 written
+= ret_splice
;
1971 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1977 * This is a special case that the relayd has closed its socket. Let's
1978 * cleanup the relayd object and all associated streams.
1980 if (relayd
&& relayd_hang_up
) {
1981 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1982 lttng_consumer_cleanup_relayd(relayd
);
1983 /* Skip splice error so the consumer does not fail */
1988 /* send the appropriate error description to sessiond */
1991 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1994 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1997 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
2002 if (relayd
&& stream
->metadata_flag
) {
2003 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2011 * Sample the snapshot positions for a specific fd
2013 * Returns 0 on success, < 0 on error
2015 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream
*stream
)
2017 switch (the_consumer_data
.type
) {
2018 case LTTNG_CONSUMER_KERNEL
:
2019 return lttng_kconsumer_sample_snapshot_positions(stream
);
2020 case LTTNG_CONSUMER32_UST
:
2021 case LTTNG_CONSUMER64_UST
:
2022 return lttng_ustconsumer_sample_snapshot_positions(stream
);
2024 ERR("Unknown consumer_data type");
2030 * Take a snapshot for a specific fd
2032 * Returns 0 on success, < 0 on error
2034 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
2036 switch (the_consumer_data
.type
) {
2037 case LTTNG_CONSUMER_KERNEL
:
2038 return lttng_kconsumer_take_snapshot(stream
);
2039 case LTTNG_CONSUMER32_UST
:
2040 case LTTNG_CONSUMER64_UST
:
2041 return lttng_ustconsumer_take_snapshot(stream
);
2043 ERR("Unknown consumer_data type");
2050 * Get the produced position
2052 * Returns 0 on success, < 0 on error
2054 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
2057 switch (the_consumer_data
.type
) {
2058 case LTTNG_CONSUMER_KERNEL
:
2059 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
2060 case LTTNG_CONSUMER32_UST
:
2061 case LTTNG_CONSUMER64_UST
:
2062 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2064 ERR("Unknown consumer_data type");
2071 * Get the consumed position (free-running counter position in bytes).
2073 * Returns 0 on success, < 0 on error
2075 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream
*stream
,
2078 switch (the_consumer_data
.type
) {
2079 case LTTNG_CONSUMER_KERNEL
:
2080 return lttng_kconsumer_get_consumed_snapshot(stream
, pos
);
2081 case LTTNG_CONSUMER32_UST
:
2082 case LTTNG_CONSUMER64_UST
:
2083 return lttng_ustconsumer_get_consumed_snapshot(stream
, pos
);
2085 ERR("Unknown consumer_data type");
2091 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2092 int sock
, struct pollfd
*consumer_sockpoll
)
2094 switch (the_consumer_data
.type
) {
2095 case LTTNG_CONSUMER_KERNEL
:
2096 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2097 case LTTNG_CONSUMER32_UST
:
2098 case LTTNG_CONSUMER64_UST
:
2099 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2101 ERR("Unknown consumer_data type");
2108 void lttng_consumer_close_all_metadata(void)
2110 switch (the_consumer_data
.type
) {
2111 case LTTNG_CONSUMER_KERNEL
:
2113 * The Kernel consumer has a different metadata scheme so we don't
2114 * close anything because the stream will be closed by the session
2118 case LTTNG_CONSUMER32_UST
:
2119 case LTTNG_CONSUMER64_UST
:
2121 * Close all metadata streams. The metadata hash table is passed and
2122 * this call iterates over it by closing all wakeup fd. This is safe
2123 * because at this point we are sure that the metadata producer is
2124 * either dead or blocked.
2126 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2129 ERR("Unknown consumer_data type");
2135 * Clean up a metadata stream and free its memory.
2137 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2138 struct lttng_ht
*ht
)
2140 struct lttng_consumer_channel
*channel
= NULL
;
2141 bool free_channel
= false;
2143 LTTNG_ASSERT(stream
);
2145 * This call should NEVER receive regular stream. It must always be
2146 * metadata stream and this is crucial for data structure synchronization.
2148 LTTNG_ASSERT(stream
->metadata_flag
);
2150 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2152 pthread_mutex_lock(&the_consumer_data
.lock
);
2154 * Note that this assumes that a stream's channel is never changed and
2155 * that the stream's lock doesn't need to be taken to sample its
2158 channel
= stream
->chan
;
2159 pthread_mutex_lock(&channel
->lock
);
2160 pthread_mutex_lock(&stream
->lock
);
2161 if (channel
->metadata_cache
) {
2162 /* Only applicable to userspace consumers. */
2163 pthread_mutex_lock(&channel
->metadata_cache
->lock
);
2166 /* Remove any reference to that stream. */
2167 consumer_stream_delete(stream
, ht
);
2169 /* Close down everything including the relayd if one. */
2170 consumer_stream_close_output(stream
);
2171 /* Destroy tracer buffers of the stream. */
2172 consumer_stream_destroy_buffers(stream
);
2174 /* Atomically decrement channel refcount since other threads can use it. */
2175 if (!uatomic_sub_return(&channel
->refcount
, 1)
2176 && !uatomic_read(&channel
->nb_init_stream_left
)) {
2177 /* Go for channel deletion! */
2178 free_channel
= true;
2180 stream
->chan
= NULL
;
2183 * Nullify the stream reference so it is not used after deletion. The
2184 * channel lock MUST be acquired before being able to check for a NULL
2187 channel
->metadata_stream
= NULL
;
2189 if (channel
->metadata_cache
) {
2190 pthread_mutex_unlock(&channel
->metadata_cache
->lock
);
2192 pthread_mutex_unlock(&stream
->lock
);
2193 pthread_mutex_unlock(&channel
->lock
);
2194 pthread_mutex_unlock(&the_consumer_data
.lock
);
2197 consumer_del_channel(channel
);
2200 lttng_trace_chunk_put(stream
->trace_chunk
);
2201 stream
->trace_chunk
= NULL
;
2202 consumer_stream_free(stream
);
2206 * Action done with the metadata stream when adding it to the consumer internal
2207 * data structures to handle it.
2209 void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2211 struct lttng_ht
*ht
= metadata_ht
;
2212 struct lttng_ht_iter iter
;
2213 struct lttng_ht_node_u64
*node
;
2215 LTTNG_ASSERT(stream
);
2218 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2220 pthread_mutex_lock(&the_consumer_data
.lock
);
2221 pthread_mutex_lock(&stream
->chan
->lock
);
2222 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2223 pthread_mutex_lock(&stream
->lock
);
2226 * From here, refcounts are updated so be _careful_ when returning an error
2233 * Lookup the stream just to make sure it does not exist in our internal
2234 * state. This should NEVER happen.
2236 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2237 node
= lttng_ht_iter_get_node_u64(&iter
);
2238 LTTNG_ASSERT(!node
);
2241 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2242 * in terms of destroying the associated channel, because the action that
2243 * causes the count to become 0 also causes a stream to be added. The
2244 * channel deletion will thus be triggered by the following removal of this
2247 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2248 /* Increment refcount before decrementing nb_init_stream_left */
2250 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2253 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2255 lttng_ht_add_u64(the_consumer_data
.stream_per_chan_id_ht
,
2256 &stream
->node_channel_id
);
2259 * Add stream to the stream_list_ht of the consumer data. No need to steal
2260 * the key since the HT does not use it and we allow to add redundant keys
2263 lttng_ht_add_u64(the_consumer_data
.stream_list_ht
,
2264 &stream
->node_session_id
);
2268 pthread_mutex_unlock(&stream
->lock
);
2269 pthread_mutex_unlock(&stream
->chan
->lock
);
2270 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2271 pthread_mutex_unlock(&the_consumer_data
.lock
);
2275 * Delete data stream that are flagged for deletion (endpoint_status).
2277 static void validate_endpoint_status_data_stream(void)
2279 struct lttng_ht_iter iter
;
2280 struct lttng_consumer_stream
*stream
;
2282 DBG("Consumer delete flagged data stream");
2285 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2286 /* Validate delete flag of the stream */
2287 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2290 /* Delete it right now */
2291 consumer_del_stream(stream
, data_ht
);
2297 * Delete metadata stream that are flagged for deletion (endpoint_status).
2299 static void validate_endpoint_status_metadata_stream(
2300 struct lttng_poll_event
*pollset
)
2302 struct lttng_ht_iter iter
;
2303 struct lttng_consumer_stream
*stream
;
2305 DBG("Consumer delete flagged metadata stream");
2307 LTTNG_ASSERT(pollset
);
2310 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2311 /* Validate delete flag of the stream */
2312 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2316 * Remove from pollset so the metadata thread can continue without
2317 * blocking on a deleted stream.
2319 lttng_poll_del(pollset
, stream
->wait_fd
);
2321 /* Delete it right now */
2322 consumer_del_metadata_stream(stream
, metadata_ht
);
2328 * Thread polls on metadata file descriptor and write them on disk or on the
2331 void *consumer_thread_metadata_poll(void *data
)
2333 int ret
, i
, pollfd
, err
= -1;
2334 uint32_t revents
, nb_fd
;
2335 struct lttng_consumer_stream
*stream
= NULL
;
2336 struct lttng_ht_iter iter
;
2337 struct lttng_ht_node_u64
*node
;
2338 struct lttng_poll_event events
;
2339 struct lttng_consumer_local_data
*ctx
= (lttng_consumer_local_data
*) data
;
2342 rcu_register_thread();
2344 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2346 if (testpoint(consumerd_thread_metadata
)) {
2347 goto error_testpoint
;
2350 health_code_update();
2352 DBG("Thread metadata poll started");
2354 /* Size is set to 1 for the consumer_metadata pipe */
2355 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2357 ERR("Poll set creation failed");
2361 ret
= lttng_poll_add(&events
,
2362 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2368 DBG("Metadata main loop started");
2372 health_code_update();
2373 health_poll_entry();
2374 DBG("Metadata poll wait");
2375 ret
= lttng_poll_wait(&events
, -1);
2376 DBG("Metadata poll return from wait with %d fd(s)",
2377 LTTNG_POLL_GETNB(&events
));
2379 DBG("Metadata event caught in thread");
2381 if (errno
== EINTR
) {
2382 ERR("Poll EINTR caught");
2385 if (LTTNG_POLL_GETNB(&events
) == 0) {
2386 err
= 0; /* All is OK */
2393 /* From here, the event is a metadata wait fd */
2394 for (i
= 0; i
< nb_fd
; i
++) {
2395 health_code_update();
2397 revents
= LTTNG_POLL_GETEV(&events
, i
);
2398 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2400 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2401 if (revents
& LPOLLIN
) {
2404 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2405 &stream
, sizeof(stream
));
2406 if (pipe_len
< sizeof(stream
)) {
2408 PERROR("read metadata stream");
2411 * Remove the pipe from the poll set and continue the loop
2412 * since their might be data to consume.
2414 lttng_poll_del(&events
,
2415 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2416 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2420 /* A NULL stream means that the state has changed. */
2421 if (stream
== NULL
) {
2422 /* Check for deleted streams. */
2423 validate_endpoint_status_metadata_stream(&events
);
2427 DBG("Adding metadata stream %d to poll set",
2430 /* Add metadata stream to the global poll events list */
2431 lttng_poll_add(&events
, stream
->wait_fd
, LPOLLIN
| LPOLLPRI
);
2432 }else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2433 DBG("Metadata thread pipe hung up");
2435 * Remove the pipe from the poll set and continue the loop
2436 * since their might be data to consume.
2438 lttng_poll_del(&events
,
2439 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2440 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2443 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2447 /* Handle other stream */
2453 uint64_t tmp_id
= (uint64_t) pollfd
;
2455 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2457 node
= lttng_ht_iter_get_node_u64(&iter
);
2460 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2463 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2464 /* Get the data out of the metadata file descriptor */
2465 DBG("Metadata available on fd %d", pollfd
);
2466 LTTNG_ASSERT(stream
->wait_fd
== pollfd
);
2469 health_code_update();
2471 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2473 * We don't check the return value here since if we get
2474 * a negative len, it means an error occurred thus we
2475 * simply remove it from the poll set and free the
2480 /* It's ok to have an unavailable sub-buffer */
2481 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2482 /* Clean up stream from consumer and free it. */
2483 lttng_poll_del(&events
, stream
->wait_fd
);
2484 consumer_del_metadata_stream(stream
, metadata_ht
);
2486 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2487 DBG("Metadata fd %d is hup|err.", pollfd
);
2488 if (!stream
->hangup_flush_done
&&
2489 (the_consumer_data
.type
== LTTNG_CONSUMER32_UST
||
2490 the_consumer_data
.type
==
2491 LTTNG_CONSUMER64_UST
)) {
2492 DBG("Attempting to flush and consume the UST buffers");
2493 lttng_ustconsumer_on_stream_hangup(stream
);
2495 /* We just flushed the stream now read it. */
2497 health_code_update();
2499 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2501 * We don't check the return value here since if we get
2502 * a negative len, it means an error occurred thus we
2503 * simply remove it from the poll set and free the
2509 lttng_poll_del(&events
, stream
->wait_fd
);
2511 * This call update the channel states, closes file descriptors
2512 * and securely free the stream.
2514 consumer_del_metadata_stream(stream
, metadata_ht
);
2516 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2520 /* Release RCU lock for the stream looked up */
2528 DBG("Metadata poll thread exiting");
2530 lttng_poll_clean(&events
);
2535 ERR("Health error occurred in %s", __func__
);
2537 health_unregister(health_consumerd
);
2538 rcu_unregister_thread();
2543 * This thread polls the fds in the set to consume the data and write
2544 * it to tracefile if necessary.
2546 void *consumer_thread_data_poll(void *data
)
2548 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2549 struct pollfd
*pollfd
= NULL
;
2550 /* local view of the streams */
2551 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2552 /* local view of consumer_data.fds_count */
2554 /* 2 for the consumer_data_pipe and wake up pipe */
2555 const int nb_pipes_fd
= 2;
2556 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2557 int nb_inactive_fd
= 0;
2558 struct lttng_consumer_local_data
*ctx
= (lttng_consumer_local_data
*) data
;
2561 rcu_register_thread();
2563 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2565 if (testpoint(consumerd_thread_data
)) {
2566 goto error_testpoint
;
2569 health_code_update();
2571 local_stream
= zmalloc
<lttng_consumer_stream
*>();
2572 if (local_stream
== NULL
) {
2573 PERROR("local_stream malloc");
2578 health_code_update();
2584 * the fds set has been updated, we need to update our
2585 * local array as well
2587 pthread_mutex_lock(&the_consumer_data
.lock
);
2588 if (the_consumer_data
.need_update
) {
2593 local_stream
= NULL
;
2595 /* Allocate for all fds */
2596 pollfd
= calloc
<struct pollfd
>(the_consumer_data
.stream_count
+ nb_pipes_fd
);
2597 if (pollfd
== NULL
) {
2598 PERROR("pollfd malloc");
2599 pthread_mutex_unlock(&the_consumer_data
.lock
);
2603 local_stream
= calloc
<lttng_consumer_stream
*>(the_consumer_data
.stream_count
+ nb_pipes_fd
);
2604 if (local_stream
== NULL
) {
2605 PERROR("local_stream malloc");
2606 pthread_mutex_unlock(&the_consumer_data
.lock
);
2609 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2610 data_ht
, &nb_inactive_fd
);
2612 ERR("Error in allocating pollfd or local_outfds");
2613 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2614 pthread_mutex_unlock(&the_consumer_data
.lock
);
2618 the_consumer_data
.need_update
= 0;
2620 pthread_mutex_unlock(&the_consumer_data
.lock
);
2622 /* No FDs and consumer_quit, consumer_cleanup the thread */
2623 if (nb_fd
== 0 && nb_inactive_fd
== 0 &&
2624 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2625 err
= 0; /* All is OK */
2628 /* poll on the array of fds */
2630 DBG("polling on %d fd", nb_fd
+ nb_pipes_fd
);
2631 if (testpoint(consumerd_thread_data_poll
)) {
2634 health_poll_entry();
2635 num_rdy
= poll(pollfd
, nb_fd
+ nb_pipes_fd
, -1);
2637 DBG("poll num_rdy : %d", num_rdy
);
2638 if (num_rdy
== -1) {
2640 * Restart interrupted system call.
2642 if (errno
== EINTR
) {
2645 PERROR("Poll error");
2646 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2648 } else if (num_rdy
== 0) {
2649 DBG("Polling thread timed out");
2653 if (caa_unlikely(data_consumption_paused
)) {
2654 DBG("Data consumption paused, sleeping...");
2660 * If the consumer_data_pipe triggered poll go directly to the
2661 * beginning of the loop to update the array. We want to prioritize
2662 * array update over low-priority reads.
2664 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2665 ssize_t pipe_readlen
;
2667 DBG("consumer_data_pipe wake up");
2668 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2669 &new_stream
, sizeof(new_stream
));
2670 if (pipe_readlen
< sizeof(new_stream
)) {
2671 PERROR("Consumer data pipe");
2672 /* Continue so we can at least handle the current stream(s). */
2677 * If the stream is NULL, just ignore it. It's also possible that
2678 * the sessiond poll thread changed the consumer_quit state and is
2679 * waking us up to test it.
2681 if (new_stream
== NULL
) {
2682 validate_endpoint_status_data_stream();
2686 /* Continue to update the local streams and handle prio ones */
2690 /* Handle wakeup pipe. */
2691 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2693 ssize_t pipe_readlen
;
2695 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2697 if (pipe_readlen
< 0) {
2698 PERROR("Consumer data wakeup pipe");
2700 /* We've been awakened to handle stream(s). */
2701 ctx
->has_wakeup
= 0;
2704 /* Take care of high priority channels first. */
2705 for (i
= 0; i
< nb_fd
; i
++) {
2706 health_code_update();
2708 if (local_stream
[i
] == NULL
) {
2711 if (pollfd
[i
].revents
& POLLPRI
) {
2712 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2714 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2715 /* it's ok to have an unavailable sub-buffer */
2716 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2717 /* Clean the stream and free it. */
2718 consumer_del_stream(local_stream
[i
], data_ht
);
2719 local_stream
[i
] = NULL
;
2720 } else if (len
> 0) {
2721 local_stream
[i
]->has_data_left_to_be_read_before_teardown
= 1;
2727 * If we read high prio channel in this loop, try again
2728 * for more high prio data.
2734 /* Take care of low priority channels. */
2735 for (i
= 0; i
< nb_fd
; i
++) {
2736 health_code_update();
2738 if (local_stream
[i
] == NULL
) {
2741 if ((pollfd
[i
].revents
& POLLIN
) ||
2742 local_stream
[i
]->hangup_flush_done
||
2743 local_stream
[i
]->has_data
) {
2744 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2745 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2746 /* it's ok to have an unavailable sub-buffer */
2747 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2748 /* Clean the stream and free it. */
2749 consumer_del_stream(local_stream
[i
], data_ht
);
2750 local_stream
[i
] = NULL
;
2751 } else if (len
> 0) {
2752 local_stream
[i
]->has_data_left_to_be_read_before_teardown
= 1;
2757 /* Handle hangup and errors */
2758 for (i
= 0; i
< nb_fd
; i
++) {
2759 health_code_update();
2761 if (local_stream
[i
] == NULL
) {
2764 if (!local_stream
[i
]->hangup_flush_done
2765 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2766 && (the_consumer_data
.type
== LTTNG_CONSUMER32_UST
2767 || the_consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2768 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2770 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2771 /* Attempt read again, for the data we just flushed. */
2772 local_stream
[i
]->has_data_left_to_be_read_before_teardown
= 1;
2775 * When a stream's pipe dies (hup/err/nval), an "inactive producer" flush is
2776 * performed. This type of flush ensures that a new packet is produced no
2777 * matter the consumed/produced positions are.
2779 * This, in turn, causes the next pass to see that data available for the
2780 * stream. When we come back here, we can be assured that all available
2781 * data has been consumed and we can finally destroy the stream.
2783 * If the poll flag is HUP/ERR/NVAL and we have
2784 * read no data in this pass, we can remove the
2785 * stream from its hash table.
2787 if ((pollfd
[i
].revents
& POLLHUP
)) {
2788 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2789 if (!local_stream
[i
]->has_data_left_to_be_read_before_teardown
) {
2790 consumer_del_stream(local_stream
[i
], data_ht
);
2791 local_stream
[i
] = NULL
;
2794 } else if (pollfd
[i
].revents
& POLLERR
) {
2795 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2796 if (!local_stream
[i
]->has_data_left_to_be_read_before_teardown
) {
2797 consumer_del_stream(local_stream
[i
], data_ht
);
2798 local_stream
[i
] = NULL
;
2801 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2802 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2803 if (!local_stream
[i
]->has_data_left_to_be_read_before_teardown
) {
2804 consumer_del_stream(local_stream
[i
], data_ht
);
2805 local_stream
[i
] = NULL
;
2809 if (local_stream
[i
] != NULL
) {
2810 local_stream
[i
]->has_data_left_to_be_read_before_teardown
= 0;
2817 DBG("polling thread exiting");
2822 * Close the write side of the pipe so epoll_wait() in
2823 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2824 * read side of the pipe. If we close them both, epoll_wait strangely does
2825 * not return and could create a endless wait period if the pipe is the
2826 * only tracked fd in the poll set. The thread will take care of closing
2829 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2834 ERR("Health error occurred in %s", __func__
);
2836 health_unregister(health_consumerd
);
2838 rcu_unregister_thread();
2843 * Close wake-up end of each stream belonging to the channel. This will
2844 * allow the poll() on the stream read-side to detect when the
2845 * write-side (application) finally closes them.
2848 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2850 struct lttng_ht
*ht
;
2851 struct lttng_consumer_stream
*stream
;
2852 struct lttng_ht_iter iter
;
2854 ht
= the_consumer_data
.stream_per_chan_id_ht
;
2857 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2858 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2859 ht
->match_fct
, &channel
->key
,
2860 &iter
.iter
, stream
, node_channel_id
.node
) {
2862 * Protect against teardown with mutex.
2864 pthread_mutex_lock(&stream
->lock
);
2865 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2868 switch (the_consumer_data
.type
) {
2869 case LTTNG_CONSUMER_KERNEL
:
2871 case LTTNG_CONSUMER32_UST
:
2872 case LTTNG_CONSUMER64_UST
:
2873 if (stream
->metadata_flag
) {
2874 /* Safe and protected by the stream lock. */
2875 lttng_ustconsumer_close_metadata(stream
->chan
);
2878 * Note: a mutex is taken internally within
2879 * liblttng-ust-ctl to protect timer wakeup_fd
2880 * use from concurrent close.
2882 lttng_ustconsumer_close_stream_wakeup(stream
);
2886 ERR("Unknown consumer_data type");
2890 pthread_mutex_unlock(&stream
->lock
);
2895 static void destroy_channel_ht(struct lttng_ht
*ht
)
2897 struct lttng_ht_iter iter
;
2898 struct lttng_consumer_channel
*channel
;
2906 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2907 ret
= lttng_ht_del(ht
, &iter
);
2908 LTTNG_ASSERT(ret
!= 0);
2912 lttng_ht_destroy(ht
);
2916 * This thread polls the channel fds to detect when they are being
2917 * closed. It closes all related streams if the channel is detected as
2918 * closed. It is currently only used as a shim layer for UST because the
2919 * consumerd needs to keep the per-stream wakeup end of pipes open for
2922 void *consumer_thread_channel_poll(void *data
)
2924 int ret
, i
, pollfd
, err
= -1;
2925 uint32_t revents
, nb_fd
;
2926 struct lttng_consumer_channel
*chan
= NULL
;
2927 struct lttng_ht_iter iter
;
2928 struct lttng_ht_node_u64
*node
;
2929 struct lttng_poll_event events
;
2930 struct lttng_consumer_local_data
*ctx
= (lttng_consumer_local_data
*) data
;
2931 struct lttng_ht
*channel_ht
;
2933 rcu_register_thread();
2935 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2937 if (testpoint(consumerd_thread_channel
)) {
2938 goto error_testpoint
;
2941 health_code_update();
2943 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2945 /* ENOMEM at this point. Better to bail out. */
2949 DBG("Thread channel poll started");
2951 /* Size is set to 1 for the consumer_channel pipe */
2952 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2954 ERR("Poll set creation failed");
2958 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2964 DBG("Channel main loop started");
2968 health_code_update();
2969 DBG("Channel poll wait");
2970 health_poll_entry();
2971 ret
= lttng_poll_wait(&events
, -1);
2972 DBG("Channel poll return from wait with %d fd(s)",
2973 LTTNG_POLL_GETNB(&events
));
2975 DBG("Channel event caught in thread");
2977 if (errno
== EINTR
) {
2978 ERR("Poll EINTR caught");
2981 if (LTTNG_POLL_GETNB(&events
) == 0) {
2982 err
= 0; /* All is OK */
2989 /* From here, the event is a channel wait fd */
2990 for (i
= 0; i
< nb_fd
; i
++) {
2991 health_code_update();
2993 revents
= LTTNG_POLL_GETEV(&events
, i
);
2994 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2996 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2997 if (revents
& LPOLLIN
) {
2998 enum consumer_channel_action action
;
3001 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
3004 ERR("Error reading channel pipe");
3006 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3011 case CONSUMER_CHANNEL_ADD
:
3012 DBG("Adding channel %d to poll set",
3015 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
3018 lttng_ht_add_unique_u64(channel_ht
,
3019 &chan
->wait_fd_node
);
3021 /* Add channel to the global poll events list */
3022 // FIXME: Empty flag on a pipe pollset, this might hang on FreeBSD.
3023 lttng_poll_add(&events
, chan
->wait_fd
, 0);
3025 case CONSUMER_CHANNEL_DEL
:
3028 * This command should never be called if the channel
3029 * has streams monitored by either the data or metadata
3030 * thread. The consumer only notify this thread with a
3031 * channel del. command if it receives a destroy
3032 * channel command from the session daemon that send it
3033 * if a command prior to the GET_CHANNEL failed.
3037 chan
= consumer_find_channel(key
);
3040 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
3043 lttng_poll_del(&events
, chan
->wait_fd
);
3044 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
3045 ret
= lttng_ht_del(channel_ht
, &iter
);
3046 LTTNG_ASSERT(ret
== 0);
3048 switch (the_consumer_data
.type
) {
3049 case LTTNG_CONSUMER_KERNEL
:
3051 case LTTNG_CONSUMER32_UST
:
3052 case LTTNG_CONSUMER64_UST
:
3053 health_code_update();
3054 /* Destroy streams that might have been left in the stream list. */
3055 clean_channel_stream_list(chan
);
3058 ERR("Unknown consumer_data type");
3063 * Release our own refcount. Force channel deletion even if
3064 * streams were not initialized.
3066 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
3067 consumer_del_channel(chan
);
3072 case CONSUMER_CHANNEL_QUIT
:
3074 * Remove the pipe from the poll set and continue the loop
3075 * since their might be data to consume.
3077 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3080 ERR("Unknown action");
3083 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3084 DBG("Channel thread pipe hung up");
3086 * Remove the pipe from the poll set and continue the loop
3087 * since their might be data to consume.
3089 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3092 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3096 /* Handle other stream */
3102 uint64_t tmp_id
= (uint64_t) pollfd
;
3104 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3106 node
= lttng_ht_iter_get_node_u64(&iter
);
3109 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3112 /* Check for error event */
3113 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3114 DBG("Channel fd %d is hup|err.", pollfd
);
3116 lttng_poll_del(&events
, chan
->wait_fd
);
3117 ret
= lttng_ht_del(channel_ht
, &iter
);
3118 LTTNG_ASSERT(ret
== 0);
3121 * This will close the wait fd for each stream associated to
3122 * this channel AND monitored by the data/metadata thread thus
3123 * will be clean by the right thread.
3125 consumer_close_channel_streams(chan
);
3127 /* Release our own refcount */
3128 if (!uatomic_sub_return(&chan
->refcount
, 1)
3129 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3130 consumer_del_channel(chan
);
3133 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3138 /* Release RCU lock for the channel looked up */
3146 lttng_poll_clean(&events
);
3148 destroy_channel_ht(channel_ht
);
3151 DBG("Channel poll thread exiting");
3154 ERR("Health error occurred in %s", __func__
);
3156 health_unregister(health_consumerd
);
3157 rcu_unregister_thread();
3161 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3162 struct pollfd
*sockpoll
, int client_socket
)
3167 LTTNG_ASSERT(sockpoll
);
3169 ret
= lttng_consumer_poll_socket(sockpoll
);
3173 DBG("Metadata connection on client_socket");
3175 /* Blocking call, waiting for transmission */
3176 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3177 if (ctx
->consumer_metadata_socket
< 0) {
3178 WARN("On accept metadata");
3189 * This thread listens on the consumerd socket and receives the file
3190 * descriptors from the session daemon.
3192 void *consumer_thread_sessiond_poll(void *data
)
3194 int sock
= -1, client_socket
, ret
, err
= -1;
3196 * structure to poll for incoming data on communication socket avoids
3197 * making blocking sockets.
3199 struct pollfd consumer_sockpoll
[2];
3200 struct lttng_consumer_local_data
*ctx
= (lttng_consumer_local_data
*) data
;
3202 rcu_register_thread();
3204 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3206 if (testpoint(consumerd_thread_sessiond
)) {
3207 goto error_testpoint
;
3210 health_code_update();
3212 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3213 unlink(ctx
->consumer_command_sock_path
);
3214 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3215 if (client_socket
< 0) {
3216 ERR("Cannot create command socket");
3220 ret
= lttcomm_listen_unix_sock(client_socket
);
3225 DBG("Sending ready command to lttng-sessiond");
3226 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3227 /* return < 0 on error, but == 0 is not fatal */
3229 ERR("Error sending ready command to lttng-sessiond");
3233 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3234 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3235 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3236 consumer_sockpoll
[1].fd
= client_socket
;
3237 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3239 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3247 DBG("Connection on client_socket");
3249 /* Blocking call, waiting for transmission */
3250 sock
= lttcomm_accept_unix_sock(client_socket
);
3257 * Setup metadata socket which is the second socket connection on the
3258 * command unix socket.
3260 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3269 /* This socket is not useful anymore. */
3270 ret
= close(client_socket
);
3272 PERROR("close client_socket");
3276 /* update the polling structure to poll on the established socket */
3277 consumer_sockpoll
[1].fd
= sock
;
3278 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3281 health_code_update();
3283 health_poll_entry();
3284 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3293 DBG("Incoming command on sock");
3294 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3297 * This could simply be a session daemon quitting. Don't output
3300 DBG("Communication interrupted on command socket");
3304 if (CMM_LOAD_SHARED(consumer_quit
)) {
3305 DBG("consumer_thread_receive_fds received quit from signal");
3306 err
= 0; /* All is OK */
3309 DBG("Received command on sock");
3315 DBG("Consumer thread sessiond poll exiting");
3318 * Close metadata streams since the producer is the session daemon which
3321 * NOTE: for now, this only applies to the UST tracer.
3323 lttng_consumer_close_all_metadata();
3326 * when all fds have hung up, the polling thread
3329 CMM_STORE_SHARED(consumer_quit
, 1);
3332 * Notify the data poll thread to poll back again and test the
3333 * consumer_quit state that we just set so to quit gracefully.
3335 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3337 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3339 notify_health_quit_pipe(health_quit_pipe
);
3341 /* Cleaning up possibly open sockets. */
3345 PERROR("close sock sessiond poll");
3348 if (client_socket
>= 0) {
3349 ret
= close(client_socket
);
3351 PERROR("close client_socket sessiond poll");
3358 ERR("Health error occurred in %s", __func__
);
3360 health_unregister(health_consumerd
);
3362 rcu_unregister_thread();
3366 static int post_consume(struct lttng_consumer_stream
*stream
,
3367 const struct stream_subbuffer
*subbuffer
,
3368 struct lttng_consumer_local_data
*ctx
)
3372 const size_t count
= lttng_dynamic_array_get_count(
3373 &stream
->read_subbuffer_ops
.post_consume_cbs
);
3375 for (i
= 0; i
< count
; i
++) {
3376 const post_consume_cb op
= *(post_consume_cb
*) lttng_dynamic_array_get_element(
3377 &stream
->read_subbuffer_ops
.post_consume_cbs
,
3380 ret
= op(stream
, subbuffer
, ctx
);
3389 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3390 struct lttng_consumer_local_data
*ctx
,
3391 bool locked_by_caller
)
3393 ssize_t ret
, written_bytes
= 0;
3395 struct stream_subbuffer subbuffer
= {};
3396 enum get_next_subbuffer_status get_next_status
;
3398 if (!locked_by_caller
) {
3399 stream
->read_subbuffer_ops
.lock(stream
);
3401 stream
->read_subbuffer_ops
.assert_locked(stream
);
3404 if (stream
->read_subbuffer_ops
.on_wake_up
) {
3405 ret
= stream
->read_subbuffer_ops
.on_wake_up(stream
);
3412 * If the stream was flagged to be ready for rotation before we extract
3413 * the next packet, rotate it now.
3415 if (stream
->rotate_ready
) {
3416 DBG("Rotate stream before consuming data");
3417 ret
= lttng_consumer_rotate_stream(stream
);
3419 ERR("Stream rotation error before consuming data");
3424 get_next_status
= stream
->read_subbuffer_ops
.get_next_subbuffer(
3425 stream
, &subbuffer
);
3426 switch (get_next_status
) {
3427 case GET_NEXT_SUBBUFFER_STATUS_OK
:
3429 case GET_NEXT_SUBBUFFER_STATUS_NO_DATA
:
3433 case GET_NEXT_SUBBUFFER_STATUS_ERROR
:
3440 ret
= stream
->read_subbuffer_ops
.pre_consume_subbuffer(
3441 stream
, &subbuffer
);
3443 goto error_put_subbuf
;
3446 written_bytes
= stream
->read_subbuffer_ops
.consume_subbuffer(
3447 ctx
, stream
, &subbuffer
);
3448 if (written_bytes
<= 0) {
3449 ERR("Error consuming subbuffer: (%zd)", written_bytes
);
3450 ret
= (int) written_bytes
;
3451 goto error_put_subbuf
;
3454 ret
= stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3459 ret
= post_consume(stream
, &subbuffer
, ctx
);
3465 * After extracting the packet, we check if the stream is now ready to
3466 * be rotated and perform the action immediately.
3468 * Don't overwrite `ret` as callers expect the number of bytes
3469 * consumed to be returned on success.
3471 rotation_ret
= lttng_consumer_stream_is_rotate_ready(stream
);
3472 if (rotation_ret
== 1) {
3473 rotation_ret
= lttng_consumer_rotate_stream(stream
);
3474 if (rotation_ret
< 0) {
3476 ERR("Stream rotation error after consuming data");
3480 } else if (rotation_ret
< 0) {
3482 ERR("Failed to check if stream was ready to rotate after consuming data");
3487 if (stream
->read_subbuffer_ops
.on_sleep
) {
3488 stream
->read_subbuffer_ops
.on_sleep(stream
, ctx
);
3491 ret
= written_bytes
;
3493 if (!locked_by_caller
) {
3494 stream
->read_subbuffer_ops
.unlock(stream
);
3499 (void) stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3503 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3505 switch (the_consumer_data
.type
) {
3506 case LTTNG_CONSUMER_KERNEL
:
3507 return lttng_kconsumer_on_recv_stream(stream
);
3508 case LTTNG_CONSUMER32_UST
:
3509 case LTTNG_CONSUMER64_UST
:
3510 return lttng_ustconsumer_on_recv_stream(stream
);
3512 ERR("Unknown consumer_data type");
3519 * Allocate and set consumer data hash tables.
3521 int lttng_consumer_init(void)
3523 the_consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3524 if (!the_consumer_data
.channel_ht
) {
3528 the_consumer_data
.channels_by_session_id_ht
=
3529 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3530 if (!the_consumer_data
.channels_by_session_id_ht
) {
3534 the_consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3535 if (!the_consumer_data
.relayd_ht
) {
3539 the_consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3540 if (!the_consumer_data
.stream_list_ht
) {
3544 the_consumer_data
.stream_per_chan_id_ht
=
3545 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3546 if (!the_consumer_data
.stream_per_chan_id_ht
) {
3550 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3555 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3560 the_consumer_data
.chunk_registry
= lttng_trace_chunk_registry_create();
3561 if (!the_consumer_data
.chunk_registry
) {
3572 * Process the ADD_RELAYD command receive by a consumer.
3574 * This will create a relayd socket pair and add it to the relayd hash table.
3575 * The caller MUST acquire a RCU read side lock before calling it.
3577 void consumer_add_relayd_socket(uint64_t net_seq_idx
,
3579 struct lttng_consumer_local_data
*ctx
,
3581 struct pollfd
*consumer_sockpoll
,
3582 uint64_t sessiond_id
,
3583 uint64_t relayd_session_id
,
3584 uint32_t relayd_version_major
,
3585 uint32_t relayd_version_minor
,
3586 enum lttcomm_sock_proto relayd_socket_protocol
)
3588 int fd
= -1, ret
= -1, relayd_created
= 0;
3589 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3590 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3593 LTTNG_ASSERT(sock
>= 0);
3594 ASSERT_RCU_READ_LOCKED();
3596 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3598 /* Get relayd reference if exists. */
3599 relayd
= consumer_find_relayd(net_seq_idx
);
3600 if (relayd
== NULL
) {
3601 LTTNG_ASSERT(sock_type
== LTTNG_STREAM_CONTROL
);
3602 /* Not found. Allocate one. */
3603 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3604 if (relayd
== NULL
) {
3605 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3608 relayd
->sessiond_session_id
= sessiond_id
;
3613 * This code path MUST continue to the consumer send status message to
3614 * we can notify the session daemon and continue our work without
3615 * killing everything.
3619 * relayd key should never be found for control socket.
3621 LTTNG_ASSERT(sock_type
!= LTTNG_STREAM_CONTROL
);
3624 /* First send a status message before receiving the fds. */
3625 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3627 /* Somehow, the session daemon is not responding anymore. */
3628 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3629 goto error_nosignal
;
3632 /* Poll on consumer socket. */
3633 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3635 /* Needing to exit in the middle of a command: error. */
3636 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3637 goto error_nosignal
;
3640 /* Get relayd socket from session daemon */
3641 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3642 if (ret
!= sizeof(fd
)) {
3643 fd
= -1; /* Just in case it gets set with an invalid value. */
3646 * Failing to receive FDs might indicate a major problem such as
3647 * reaching a fd limit during the receive where the kernel returns a
3648 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3649 * don't take any chances and stop everything.
3651 * XXX: Feature request #558 will fix that and avoid this possible
3652 * issue when reaching the fd limit.
3654 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3655 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3659 /* Copy socket information and received FD */
3660 switch (sock_type
) {
3661 case LTTNG_STREAM_CONTROL
:
3662 /* Copy received lttcomm socket */
3663 ret
= lttcomm_populate_sock_from_open_socket(
3664 &relayd
->control_sock
.sock
, fd
,
3665 relayd_socket_protocol
);
3667 /* Assign version values. */
3668 relayd
->control_sock
.major
= relayd_version_major
;
3669 relayd
->control_sock
.minor
= relayd_version_minor
;
3671 relayd
->relayd_session_id
= relayd_session_id
;
3674 case LTTNG_STREAM_DATA
:
3675 /* Copy received lttcomm socket */
3676 ret
= lttcomm_populate_sock_from_open_socket(
3677 &relayd
->data_sock
.sock
, fd
,
3678 relayd_socket_protocol
);
3679 /* Assign version values. */
3680 relayd
->data_sock
.major
= relayd_version_major
;
3681 relayd
->data_sock
.minor
= relayd_version_minor
;
3684 ERR("Unknown relayd socket type (%d)", sock_type
);
3685 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3690 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3694 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3695 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3696 relayd
->net_seq_idx
, fd
);
3698 * We gave the ownership of the fd to the relayd structure. Set the
3699 * fd to -1 so we don't call close() on it in the error path below.
3703 /* We successfully added the socket. Send status back. */
3704 ret
= consumer_send_status_msg(sock
, ret_code
);
3706 /* Somehow, the session daemon is not responding anymore. */
3707 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3708 goto error_nosignal
;
3712 * Add relayd socket pair to consumer data hashtable. If object already
3713 * exists or on error, the function gracefully returns.
3722 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3723 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3727 /* Close received socket if valid. */
3730 PERROR("close received socket");
3734 if (relayd_created
) {
3740 * Search for a relayd associated to the session id and return the reference.
3742 * A rcu read side lock MUST be acquire before calling this function and locked
3743 * until the relayd object is no longer necessary.
3745 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3747 struct lttng_ht_iter iter
;
3748 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3750 ASSERT_RCU_READ_LOCKED();
3752 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3753 cds_lfht_for_each_entry(the_consumer_data
.relayd_ht
->ht
, &iter
.iter
,
3754 relayd
, node
.node
) {
3756 * Check by sessiond id which is unique here where the relayd session
3757 * id might not be when having multiple relayd.
3759 if (relayd
->sessiond_session_id
== id
) {
3760 /* Found the relayd. There can be only one per id. */
3772 * Check if for a given session id there is still data needed to be extract
3775 * Return 1 if data is pending or else 0 meaning ready to be read.
3777 int consumer_data_pending(uint64_t id
)
3780 struct lttng_ht_iter iter
;
3781 struct lttng_ht
*ht
;
3782 struct lttng_consumer_stream
*stream
;
3783 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3784 int (*data_pending
)(struct lttng_consumer_stream
*);
3786 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3789 pthread_mutex_lock(&the_consumer_data
.lock
);
3791 switch (the_consumer_data
.type
) {
3792 case LTTNG_CONSUMER_KERNEL
:
3793 data_pending
= lttng_kconsumer_data_pending
;
3795 case LTTNG_CONSUMER32_UST
:
3796 case LTTNG_CONSUMER64_UST
:
3797 data_pending
= lttng_ustconsumer_data_pending
;
3800 ERR("Unknown consumer data type");
3804 /* Ease our life a bit */
3805 ht
= the_consumer_data
.stream_list_ht
;
3807 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3808 ht
->hash_fct(&id
, lttng_ht_seed
),
3810 &iter
.iter
, stream
, node_session_id
.node
) {
3811 pthread_mutex_lock(&stream
->lock
);
3814 * A removed node from the hash table indicates that the stream has
3815 * been deleted thus having a guarantee that the buffers are closed
3816 * on the consumer side. However, data can still be transmitted
3817 * over the network so don't skip the relayd check.
3819 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3821 /* Check the stream if there is data in the buffers. */
3822 ret
= data_pending(stream
);
3824 pthread_mutex_unlock(&stream
->lock
);
3829 pthread_mutex_unlock(&stream
->lock
);
3832 relayd
= find_relayd_by_session_id(id
);
3834 unsigned int is_data_inflight
= 0;
3836 /* Send init command for data pending. */
3837 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3838 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3839 relayd
->relayd_session_id
);
3841 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3842 /* Communication error thus the relayd so no data pending. */
3843 goto data_not_pending
;
3846 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3847 ht
->hash_fct(&id
, lttng_ht_seed
),
3849 &iter
.iter
, stream
, node_session_id
.node
) {
3850 if (stream
->metadata_flag
) {
3851 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3852 stream
->relayd_stream_id
);
3854 ret
= relayd_data_pending(&relayd
->control_sock
,
3855 stream
->relayd_stream_id
,
3856 stream
->next_net_seq_num
- 1);
3860 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3862 } else if (ret
< 0) {
3863 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3864 lttng_consumer_cleanup_relayd(relayd
);
3865 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3866 goto data_not_pending
;
3870 /* Send end command for data pending. */
3871 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3872 relayd
->relayd_session_id
, &is_data_inflight
);
3873 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3875 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3876 lttng_consumer_cleanup_relayd(relayd
);
3877 goto data_not_pending
;
3879 if (is_data_inflight
) {
3885 * Finding _no_ node in the hash table and no inflight data means that the
3886 * stream(s) have been removed thus data is guaranteed to be available for
3887 * analysis from the trace files.
3891 /* Data is available to be read by a viewer. */
3892 pthread_mutex_unlock(&the_consumer_data
.lock
);
3897 /* Data is still being extracted from buffers. */
3898 pthread_mutex_unlock(&the_consumer_data
.lock
);
3904 * Send a ret code status message to the sessiond daemon.
3906 * Return the sendmsg() return value.
3908 int consumer_send_status_msg(int sock
, int ret_code
)
3910 struct lttcomm_consumer_status_msg msg
;
3912 memset(&msg
, 0, sizeof(msg
));
3913 msg
.ret_code
= (lttcomm_return_code
) ret_code
;
3915 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3919 * Send a channel status message to the sessiond daemon.
3921 * Return the sendmsg() return value.
3923 int consumer_send_status_channel(int sock
,
3924 struct lttng_consumer_channel
*channel
)
3926 struct lttcomm_consumer_status_channel msg
;
3928 LTTNG_ASSERT(sock
>= 0);
3930 memset(&msg
, 0, sizeof(msg
));
3932 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3934 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3935 msg
.key
= channel
->key
;
3936 msg
.stream_count
= channel
->streams
.count
;
3939 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3942 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3943 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3944 uint64_t max_sb_size
)
3946 unsigned long start_pos
;
3948 if (!nb_packets_per_stream
) {
3949 return consumed_pos
; /* Grab everything */
3951 start_pos
= produced_pos
- lttng_offset_align_floor(produced_pos
, max_sb_size
);
3952 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3953 if ((long) (start_pos
- consumed_pos
) < 0) {
3954 return consumed_pos
; /* Grab everything */
3959 /* Stream lock must be held by the caller. */
3960 static int sample_stream_positions(struct lttng_consumer_stream
*stream
,
3961 unsigned long *produced
, unsigned long *consumed
)
3965 ASSERT_LOCKED(stream
->lock
);
3967 ret
= lttng_consumer_sample_snapshot_positions(stream
);
3969 ERR("Failed to sample snapshot positions");
3973 ret
= lttng_consumer_get_produced_snapshot(stream
, produced
);
3975 ERR("Failed to sample produced position");
3979 ret
= lttng_consumer_get_consumed_snapshot(stream
, consumed
);
3981 ERR("Failed to sample consumed position");
3990 * Sample the rotate position for all the streams of a channel. If a stream
3991 * is already at the rotate position (produced == consumed), we flag it as
3992 * ready for rotation. The rotation of ready streams occurs after we have
3993 * replied to the session daemon that we have finished sampling the positions.
3994 * Must be called with RCU read-side lock held to ensure existence of channel.
3996 * Returns 0 on success, < 0 on error
3998 int lttng_consumer_rotate_channel(struct lttng_consumer_channel
*channel
,
3999 uint64_t key
, uint64_t relayd_id
)
4002 struct lttng_consumer_stream
*stream
;
4003 struct lttng_ht_iter iter
;
4004 struct lttng_ht
*ht
= the_consumer_data
.stream_per_chan_id_ht
;
4005 struct lttng_dynamic_array stream_rotation_positions
;
4006 uint64_t next_chunk_id
, stream_count
= 0;
4007 enum lttng_trace_chunk_status chunk_status
;
4008 const bool is_local_trace
= relayd_id
== -1ULL;
4009 struct consumer_relayd_sock_pair
*relayd
= NULL
;
4010 bool rotating_to_new_chunk
= true;
4011 /* Array of `struct lttng_consumer_stream *` */
4012 struct lttng_dynamic_pointer_array streams_packet_to_open
;
4015 ASSERT_RCU_READ_LOCKED();
4017 DBG("Consumer sample rotate position for channel %" PRIu64
, key
);
4019 lttng_dynamic_array_init(&stream_rotation_positions
,
4020 sizeof(struct relayd_stream_rotation_position
), NULL
);
4021 lttng_dynamic_pointer_array_init(&streams_packet_to_open
, NULL
);
4025 pthread_mutex_lock(&channel
->lock
);
4026 LTTNG_ASSERT(channel
->trace_chunk
);
4027 chunk_status
= lttng_trace_chunk_get_id(channel
->trace_chunk
,
4029 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4031 goto end_unlock_channel
;
4034 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4035 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4036 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4037 stream
, node_channel_id
.node
) {
4038 unsigned long produced_pos
= 0, consumed_pos
= 0;
4040 health_code_update();
4043 * Lock stream because we are about to change its state.
4045 pthread_mutex_lock(&stream
->lock
);
4047 if (stream
->trace_chunk
== stream
->chan
->trace_chunk
) {
4048 rotating_to_new_chunk
= false;
4052 * Do not flush a packet when rotating from a NULL trace
4053 * chunk. The stream has no means to output data, and the prior
4054 * rotation which rotated to NULL performed that side-effect
4055 * already. No new data can be produced when a stream has no
4056 * associated trace chunk (e.g. a stop followed by a rotate).
4058 if (stream
->trace_chunk
) {
4061 if (stream
->metadata_flag
) {
4063 * Don't produce an empty metadata packet,
4064 * simply close the current one.
4066 * Metadata is regenerated on every trace chunk
4067 * switch; there is no concern that no data was
4070 flush_active
= true;
4073 * Only flush an empty packet if the "packet
4074 * open" could not be performed on transition
4075 * to a new trace chunk and no packets were
4076 * consumed within the chunk's lifetime.
4078 if (stream
->opened_packet_in_current_trace_chunk
) {
4079 flush_active
= true;
4082 * Stream could have been full at the
4083 * time of rotation, but then have had
4084 * no activity at all.
4086 * It is important to flush a packet
4087 * to prevent 0-length files from being
4088 * produced as most viewers choke on
4091 * Unfortunately viewers will not be
4092 * able to know that tracing was active
4093 * for this stream during this trace
4096 ret
= sample_stream_positions(stream
, &produced_pos
, &consumed_pos
);
4098 goto end_unlock_stream
;
4102 * Don't flush an empty packet if data
4103 * was produced; it will be consumed
4104 * before the rotation completes.
4106 flush_active
= produced_pos
!= consumed_pos
;
4107 if (!flush_active
) {
4108 const char *trace_chunk_name
;
4109 uint64_t trace_chunk_id
;
4111 chunk_status
= lttng_trace_chunk_get_name(
4112 stream
->trace_chunk
,
4115 if (chunk_status
== LTTNG_TRACE_CHUNK_STATUS_NONE
) {
4116 trace_chunk_name
= "none";
4120 * Consumer trace chunks are
4123 chunk_status
= lttng_trace_chunk_get_id(
4124 stream
->trace_chunk
,
4126 LTTNG_ASSERT(chunk_status
==
4127 LTTNG_TRACE_CHUNK_STATUS_OK
);
4129 DBG("Unable to open packet for stream during trace chunk's lifetime. "
4130 "Flushing an empty packet to prevent an empty file from being created: "
4131 "stream id = %" PRIu64
", trace chunk name = `%s`, trace chunk id = %" PRIu64
,
4132 stream
->key
, trace_chunk_name
, trace_chunk_id
);
4138 * Close the current packet before sampling the
4139 * ring buffer positions.
4141 ret
= consumer_stream_flush_buffer(stream
, flush_active
);
4143 ERR("Failed to flush stream %" PRIu64
" during channel rotation",
4145 goto end_unlock_stream
;
4149 ret
= lttng_consumer_take_snapshot(stream
);
4150 if (ret
< 0 && ret
!= -ENODATA
&& ret
!= -EAGAIN
) {
4151 ERR("Failed to sample snapshot position during channel rotation");
4152 goto end_unlock_stream
;
4155 ret
= lttng_consumer_get_produced_snapshot(stream
,
4158 ERR("Failed to sample produced position during channel rotation");
4159 goto end_unlock_stream
;
4162 ret
= lttng_consumer_get_consumed_snapshot(stream
,
4165 ERR("Failed to sample consumed position during channel rotation");
4166 goto end_unlock_stream
;
4170 * Align produced position on the start-of-packet boundary of the first
4171 * packet going into the next trace chunk.
4173 produced_pos
= lttng_align_floor(produced_pos
, stream
->max_sb_size
);
4174 if (consumed_pos
== produced_pos
) {
4175 DBG("Set rotate ready for stream %" PRIu64
" produced = %lu consumed = %lu",
4176 stream
->key
, produced_pos
, consumed_pos
);
4177 stream
->rotate_ready
= true;
4179 DBG("Different consumed and produced positions "
4180 "for stream %" PRIu64
" produced = %lu consumed = %lu",
4181 stream
->key
, produced_pos
, consumed_pos
);
4184 * The rotation position is based on the packet_seq_num of the
4185 * packet following the last packet that was consumed for this
4186 * stream, incremented by the offset between produced and
4187 * consumed positions. This rotation position is a lower bound
4188 * (inclusive) at which the next trace chunk starts. Since it
4189 * is a lower bound, it is OK if the packet_seq_num does not
4190 * correspond exactly to the same packet identified by the
4191 * consumed_pos, which can happen in overwrite mode.
4193 if (stream
->sequence_number_unavailable
) {
4195 * Rotation should never be performed on a session which
4196 * interacts with a pre-2.8 lttng-modules, which does
4197 * not implement packet sequence number.
4199 ERR("Failure to rotate stream %" PRIu64
": sequence number unavailable",
4202 goto end_unlock_stream
;
4204 stream
->rotate_position
= stream
->last_sequence_number
+ 1 +
4205 ((produced_pos
- consumed_pos
) / stream
->max_sb_size
);
4206 DBG("Set rotation position for stream %" PRIu64
" at position %" PRIu64
,
4207 stream
->key
, stream
->rotate_position
);
4209 if (!is_local_trace
) {
4211 * The relay daemon control protocol expects a rotation
4212 * position as "the sequence number of the first packet
4213 * _after_ the current trace chunk".
4215 const struct relayd_stream_rotation_position position
= {
4216 .stream_id
= stream
->relayd_stream_id
,
4217 .rotate_at_seq_num
= stream
->rotate_position
,
4220 ret
= lttng_dynamic_array_add_element(
4221 &stream_rotation_positions
,
4224 ERR("Failed to allocate stream rotation position");
4225 goto end_unlock_stream
;
4230 stream
->opened_packet_in_current_trace_chunk
= false;
4232 if (rotating_to_new_chunk
&& !stream
->metadata_flag
) {
4234 * Attempt to flush an empty packet as close to the
4235 * rotation point as possible. In the event where a
4236 * stream remains inactive after the rotation point,
4237 * this ensures that the new trace chunk has a
4238 * beginning timestamp set at the begining of the
4239 * trace chunk instead of only creating an empty
4240 * packet when the trace chunk is stopped.
4242 * This indicates to the viewers that the stream
4243 * was being recorded, but more importantly it
4244 * allows viewers to determine a useable trace
4247 * This presents a problem in the case where the
4248 * ring-buffer is completely full.
4250 * Consider the following scenario:
4251 * - The consumption of data is slow (slow network,
4253 * - The ring buffer is full,
4254 * - A rotation is initiated,
4255 * - The flush below does nothing (no space left to
4256 * open a new packet),
4257 * - The other streams rotate very soon, and new
4258 * data is produced in the new chunk,
4259 * - This stream completes its rotation long after the
4260 * rotation was initiated
4261 * - The session is stopped before any event can be
4262 * produced in this stream's buffers.
4264 * The resulting trace chunk will have a single packet
4265 * temporaly at the end of the trace chunk for this
4266 * stream making the stream intersection more narrow
4267 * than it should be.
4269 * To work-around this, an empty flush is performed
4270 * after the first consumption of a packet during a
4271 * rotation if open_packet fails. The idea is that
4272 * consuming a packet frees enough space to switch
4273 * packets in this scenario and allows the tracer to
4274 * "stamp" the beginning of the new trace chunk at the
4275 * earliest possible point.
4277 * The packet open is performed after the channel
4278 * rotation to ensure that no attempt to open a packet
4279 * is performed in a stream that has no active trace
4282 ret
= lttng_dynamic_pointer_array_add_pointer(
4283 &streams_packet_to_open
, stream
);
4285 PERROR("Failed to add a stream pointer to array of streams in which to open a packet");
4287 goto end_unlock_stream
;
4291 pthread_mutex_unlock(&stream
->lock
);
4295 if (!is_local_trace
) {
4296 relayd
= consumer_find_relayd(relayd_id
);
4298 ERR("Failed to find relayd %" PRIu64
, relayd_id
);
4300 goto end_unlock_channel
;
4303 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4304 ret
= relayd_rotate_streams(&relayd
->control_sock
, stream_count
,
4305 rotating_to_new_chunk
? &next_chunk_id
: NULL
,
4306 (const struct relayd_stream_rotation_position
*)
4307 stream_rotation_positions
.buffer
4309 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4311 ERR("Relayd rotate stream failed. Cleaning up relayd %" PRIu64
,
4312 relayd
->net_seq_idx
);
4313 lttng_consumer_cleanup_relayd(relayd
);
4314 goto end_unlock_channel
;
4318 for (stream_idx
= 0;
4319 stream_idx
< lttng_dynamic_pointer_array_get_count(
4320 &streams_packet_to_open
);
4322 enum consumer_stream_open_packet_status status
;
4324 stream
= (lttng_consumer_stream
*) lttng_dynamic_pointer_array_get_pointer(
4325 &streams_packet_to_open
, stream_idx
);
4327 pthread_mutex_lock(&stream
->lock
);
4328 status
= consumer_stream_open_packet(stream
);
4329 pthread_mutex_unlock(&stream
->lock
);
4331 case CONSUMER_STREAM_OPEN_PACKET_STATUS_OPENED
:
4332 DBG("Opened a packet after a rotation: stream id = %" PRIu64
4333 ", channel name = %s, session id = %" PRIu64
,
4334 stream
->key
, stream
->chan
->name
,
4335 stream
->chan
->session_id
);
4337 case CONSUMER_STREAM_OPEN_PACKET_STATUS_NO_SPACE
:
4339 * Can't open a packet as there is no space left
4340 * in the buffer. A new packet will be opened
4341 * once one has been consumed.
4343 DBG("No space left to open a packet after a rotation: stream id = %" PRIu64
4344 ", channel name = %s, session id = %" PRIu64
,
4345 stream
->key
, stream
->chan
->name
,
4346 stream
->chan
->session_id
);
4348 case CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR
:
4349 /* Logged by callee. */
4351 goto end_unlock_channel
;
4357 pthread_mutex_unlock(&channel
->lock
);
4362 pthread_mutex_unlock(&stream
->lock
);
4364 pthread_mutex_unlock(&channel
->lock
);
4367 lttng_dynamic_array_reset(&stream_rotation_positions
);
4368 lttng_dynamic_pointer_array_reset(&streams_packet_to_open
);
4373 int consumer_clear_buffer(struct lttng_consumer_stream
*stream
)
4376 unsigned long consumed_pos_before
, consumed_pos_after
;
4378 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4380 ERR("Taking snapshot positions");
4384 ret
= lttng_consumer_get_consumed_snapshot(stream
, &consumed_pos_before
);
4386 ERR("Consumed snapshot position");
4390 switch (the_consumer_data
.type
) {
4391 case LTTNG_CONSUMER_KERNEL
:
4392 ret
= kernctl_buffer_clear(stream
->wait_fd
);
4394 ERR("Failed to clear kernel stream (ret = %d)", ret
);
4398 case LTTNG_CONSUMER32_UST
:
4399 case LTTNG_CONSUMER64_UST
:
4400 ret
= lttng_ustconsumer_clear_buffer(stream
);
4402 ERR("Failed to clear ust stream (ret = %d)", ret
);
4407 ERR("Unknown consumer_data type");
4411 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4413 ERR("Taking snapshot positions");
4416 ret
= lttng_consumer_get_consumed_snapshot(stream
, &consumed_pos_after
);
4418 ERR("Consumed snapshot position");
4421 DBG("clear: before: %lu after: %lu", consumed_pos_before
, consumed_pos_after
);
4427 int consumer_clear_stream(struct lttng_consumer_stream
*stream
)
4431 ret
= consumer_stream_flush_buffer(stream
, 1);
4433 ERR("Failed to flush stream %" PRIu64
" during channel clear",
4435 ret
= LTTCOMM_CONSUMERD_FATAL
;
4439 ret
= consumer_clear_buffer(stream
);
4441 ERR("Failed to clear stream %" PRIu64
" during channel clear",
4443 ret
= LTTCOMM_CONSUMERD_FATAL
;
4447 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4453 int consumer_clear_unmonitored_channel(struct lttng_consumer_channel
*channel
)
4456 struct lttng_consumer_stream
*stream
;
4459 pthread_mutex_lock(&channel
->lock
);
4460 cds_list_for_each_entry(stream
, &channel
->streams
.head
, send_node
) {
4461 health_code_update();
4462 pthread_mutex_lock(&stream
->lock
);
4463 ret
= consumer_clear_stream(stream
);
4467 pthread_mutex_unlock(&stream
->lock
);
4469 pthread_mutex_unlock(&channel
->lock
);
4474 pthread_mutex_unlock(&stream
->lock
);
4475 pthread_mutex_unlock(&channel
->lock
);
4481 * Check if a stream is ready to be rotated after extracting it.
4483 * Return 1 if it is ready for rotation, 0 if it is not, a negative value on
4484 * error. Stream lock must be held.
4486 int lttng_consumer_stream_is_rotate_ready(struct lttng_consumer_stream
*stream
)
4488 DBG("Check is rotate ready for stream %" PRIu64
4489 " ready %u rotate_position %" PRIu64
4490 " last_sequence_number %" PRIu64
,
4491 stream
->key
, stream
->rotate_ready
,
4492 stream
->rotate_position
, stream
->last_sequence_number
);
4493 if (stream
->rotate_ready
) {
4498 * If packet seq num is unavailable, it means we are interacting
4499 * with a pre-2.8 lttng-modules which does not implement the
4500 * sequence number. Rotation should never be used by sessiond in this
4503 if (stream
->sequence_number_unavailable
) {
4504 ERR("Internal error: rotation used on stream %" PRIu64
4505 " with unavailable sequence number",
4510 if (stream
->rotate_position
== -1ULL ||
4511 stream
->last_sequence_number
== -1ULL) {
4516 * Rotate position not reached yet. The stream rotate position is
4517 * the position of the next packet belonging to the next trace chunk,
4518 * but consumerd considers rotation ready when reaching the last
4519 * packet of the current chunk, hence the "rotate_position - 1".
4522 DBG("Check is rotate ready for stream %" PRIu64
4523 " last_sequence_number %" PRIu64
4524 " rotate_position %" PRIu64
,
4525 stream
->key
, stream
->last_sequence_number
,
4526 stream
->rotate_position
);
4527 if (stream
->last_sequence_number
>= stream
->rotate_position
- 1) {
4535 * Reset the state for a stream after a rotation occurred.
4537 void lttng_consumer_reset_stream_rotate_state(struct lttng_consumer_stream
*stream
)
4539 DBG("lttng_consumer_reset_stream_rotate_state for stream %" PRIu64
,
4541 stream
->rotate_position
= -1ULL;
4542 stream
->rotate_ready
= false;
4546 * Perform the rotation a local stream file.
4549 int rotate_local_stream(struct lttng_consumer_stream
*stream
)
4553 DBG("Rotate local stream: stream key %" PRIu64
", channel key %" PRIu64
,
4556 stream
->tracefile_size_current
= 0;
4557 stream
->tracefile_count_current
= 0;
4559 if (stream
->out_fd
>= 0) {
4560 ret
= close(stream
->out_fd
);
4562 PERROR("Failed to close stream out_fd of channel \"%s\"",
4563 stream
->chan
->name
);
4565 stream
->out_fd
= -1;
4568 if (stream
->index_file
) {
4569 lttng_index_file_put(stream
->index_file
);
4570 stream
->index_file
= NULL
;
4573 if (!stream
->trace_chunk
) {
4577 ret
= consumer_stream_create_output_files(stream
, true);
4583 * Performs the stream rotation for the rotate session feature if needed.
4584 * It must be called with the channel and stream locks held.
4586 * Return 0 on success, a negative number of error.
4588 int lttng_consumer_rotate_stream(struct lttng_consumer_stream
*stream
)
4592 DBG("Consumer rotate stream %" PRIu64
, stream
->key
);
4595 * Update the stream's 'current' chunk to the session's (channel)
4596 * now-current chunk.
4598 lttng_trace_chunk_put(stream
->trace_chunk
);
4599 if (stream
->chan
->trace_chunk
== stream
->trace_chunk
) {
4601 * A channel can be rotated and not have a "next" chunk
4602 * to transition to. In that case, the channel's "current chunk"
4603 * has not been closed yet, but it has not been updated to
4604 * a "next" trace chunk either. Hence, the stream, like its
4605 * parent channel, becomes part of no chunk and can't output
4606 * anything until a new trace chunk is created.
4608 stream
->trace_chunk
= NULL
;
4609 } else if (stream
->chan
->trace_chunk
&&
4610 !lttng_trace_chunk_get(stream
->chan
->trace_chunk
)) {
4611 ERR("Failed to acquire a reference to channel's trace chunk during stream rotation");
4616 * Update the stream's trace chunk to its parent channel's
4617 * current trace chunk.
4619 stream
->trace_chunk
= stream
->chan
->trace_chunk
;
4622 if (stream
->net_seq_idx
== (uint64_t) -1ULL) {
4623 ret
= rotate_local_stream(stream
);
4625 ERR("Failed to rotate stream, ret = %i", ret
);
4630 if (stream
->metadata_flag
&& stream
->trace_chunk
) {
4632 * If the stream has transitioned to a new trace
4633 * chunk, the metadata should be re-dumped to the
4636 * However, it is possible for a stream to transition to
4637 * a "no-chunk" state. This can happen if a rotation
4638 * occurs on an inactive session. In such cases, the metadata
4639 * regeneration will happen when the next trace chunk is
4642 ret
= consumer_metadata_stream_dump(stream
);
4647 lttng_consumer_reset_stream_rotate_state(stream
);
4656 * Rotate all the ready streams now.
4658 * This is especially important for low throughput streams that have already
4659 * been consumed, we cannot wait for their next packet to perform the
4661 * Need to be called with RCU read-side lock held to ensure existence of
4664 * Returns 0 on success, < 0 on error
4666 int lttng_consumer_rotate_ready_streams(struct lttng_consumer_channel
*channel
,
4670 struct lttng_consumer_stream
*stream
;
4671 struct lttng_ht_iter iter
;
4672 struct lttng_ht
*ht
= the_consumer_data
.stream_per_chan_id_ht
;
4674 ASSERT_RCU_READ_LOCKED();
4678 DBG("Consumer rotate ready streams in channel %" PRIu64
, key
);
4680 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4681 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4682 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4683 stream
, node_channel_id
.node
) {
4684 health_code_update();
4686 pthread_mutex_lock(&stream
->chan
->lock
);
4687 pthread_mutex_lock(&stream
->lock
);
4689 if (!stream
->rotate_ready
) {
4690 pthread_mutex_unlock(&stream
->lock
);
4691 pthread_mutex_unlock(&stream
->chan
->lock
);
4694 DBG("Consumer rotate ready stream %" PRIu64
, stream
->key
);
4696 ret
= lttng_consumer_rotate_stream(stream
);
4697 pthread_mutex_unlock(&stream
->lock
);
4698 pthread_mutex_unlock(&stream
->chan
->lock
);
4711 enum lttcomm_return_code
lttng_consumer_init_command(
4712 struct lttng_consumer_local_data
*ctx
,
4713 const lttng_uuid
& sessiond_uuid
)
4715 enum lttcomm_return_code ret
;
4716 char uuid_str
[LTTNG_UUID_STR_LEN
];
4718 if (ctx
->sessiond_uuid
.is_set
) {
4719 ret
= LTTCOMM_CONSUMERD_ALREADY_SET
;
4723 ctx
->sessiond_uuid
.is_set
= true;
4724 ctx
->sessiond_uuid
.value
= sessiond_uuid
;
4725 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4726 lttng_uuid_to_str(sessiond_uuid
, uuid_str
);
4727 DBG("Received session daemon UUID: %s", uuid_str
);
4732 enum lttcomm_return_code
lttng_consumer_create_trace_chunk(
4733 const uint64_t *relayd_id
, uint64_t session_id
,
4735 time_t chunk_creation_timestamp
,
4736 const char *chunk_override_name
,
4737 const struct lttng_credentials
*credentials
,
4738 struct lttng_directory_handle
*chunk_directory_handle
)
4741 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4742 struct lttng_trace_chunk
*created_chunk
= NULL
, *published_chunk
= NULL
;
4743 enum lttng_trace_chunk_status chunk_status
;
4744 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4745 char creation_timestamp_buffer
[ISO8601_STR_LEN
];
4746 const char *relayd_id_str
= "(none)";
4747 const char *creation_timestamp_str
;
4748 struct lttng_ht_iter iter
;
4749 struct lttng_consumer_channel
*channel
;
4752 /* Only used for logging purposes. */
4753 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4754 "%" PRIu64
, *relayd_id
);
4755 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4756 relayd_id_str
= relayd_id_buffer
;
4758 relayd_id_str
= "(formatting error)";
4762 /* Local protocol error. */
4763 LTTNG_ASSERT(chunk_creation_timestamp
);
4764 ret
= time_to_iso8601_str(chunk_creation_timestamp
,
4765 creation_timestamp_buffer
,
4766 sizeof(creation_timestamp_buffer
));
4767 creation_timestamp_str
= !ret
? creation_timestamp_buffer
:
4768 "(formatting error)";
4770 DBG("Consumer create trace chunk command: relay_id = %s"
4771 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4772 ", chunk_override_name = %s"
4773 ", chunk_creation_timestamp = %s",
4774 relayd_id_str
, session_id
, chunk_id
,
4775 chunk_override_name
? : "(none)",
4776 creation_timestamp_str
);
4779 * The trace chunk registry, as used by the consumer daemon, implicitly
4780 * owns the trace chunks. This is only needed in the consumer since
4781 * the consumer has no notion of a session beyond session IDs being
4782 * used to identify other objects.
4784 * The lttng_trace_chunk_registry_publish() call below provides a
4785 * reference which is not released; it implicitly becomes the session
4786 * daemon's reference to the chunk in the consumer daemon.
4788 * The lifetime of trace chunks in the consumer daemon is managed by
4789 * the session daemon through the LTTNG_CONSUMER_CREATE_TRACE_CHUNK
4790 * and LTTNG_CONSUMER_DESTROY_TRACE_CHUNK commands.
4792 created_chunk
= lttng_trace_chunk_create(chunk_id
,
4793 chunk_creation_timestamp
, NULL
);
4794 if (!created_chunk
) {
4795 ERR("Failed to create trace chunk");
4796 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4800 if (chunk_override_name
) {
4801 chunk_status
= lttng_trace_chunk_override_name(created_chunk
,
4802 chunk_override_name
);
4803 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4804 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4809 if (chunk_directory_handle
) {
4810 chunk_status
= lttng_trace_chunk_set_credentials(created_chunk
,
4812 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4813 ERR("Failed to set trace chunk credentials");
4814 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4818 * The consumer daemon has no ownership of the chunk output
4821 chunk_status
= lttng_trace_chunk_set_as_user(created_chunk
,
4822 chunk_directory_handle
);
4823 chunk_directory_handle
= NULL
;
4824 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4825 ERR("Failed to set trace chunk's directory handle");
4826 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4831 published_chunk
= lttng_trace_chunk_registry_publish_chunk(
4832 the_consumer_data
.chunk_registry
, session_id
,
4834 lttng_trace_chunk_put(created_chunk
);
4835 created_chunk
= NULL
;
4836 if (!published_chunk
) {
4837 ERR("Failed to publish trace chunk");
4838 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4843 cds_lfht_for_each_entry_duplicate(
4844 the_consumer_data
.channels_by_session_id_ht
->ht
,
4845 the_consumer_data
.channels_by_session_id_ht
->hash_fct(
4846 &session_id
, lttng_ht_seed
),
4847 the_consumer_data
.channels_by_session_id_ht
->match_fct
,
4848 &session_id
, &iter
.iter
, channel
,
4849 channels_by_session_id_ht_node
.node
) {
4850 ret
= lttng_consumer_channel_set_trace_chunk(channel
,
4854 * Roll-back the creation of this chunk.
4856 * This is important since the session daemon will
4857 * assume that the creation of this chunk failed and
4858 * will never ask for it to be closed, resulting
4859 * in a leak and an inconsistent state for some
4862 enum lttcomm_return_code close_ret
;
4863 char path
[LTTNG_PATH_MAX
];
4865 DBG("Failed to set new trace chunk on existing channels, rolling back");
4866 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4867 session_id
, chunk_id
,
4868 chunk_creation_timestamp
, NULL
,
4870 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4871 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4872 session_id
, chunk_id
);
4875 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4881 struct consumer_relayd_sock_pair
*relayd
;
4883 relayd
= consumer_find_relayd(*relayd_id
);
4885 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4886 ret
= relayd_create_trace_chunk(
4887 &relayd
->control_sock
, published_chunk
);
4888 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4890 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
, *relayd_id
);
4893 if (!relayd
|| ret
) {
4894 enum lttcomm_return_code close_ret
;
4895 char path
[LTTNG_PATH_MAX
];
4897 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4900 chunk_creation_timestamp
,
4902 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4903 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4908 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4915 /* Release the reference returned by the "publish" operation. */
4916 lttng_trace_chunk_put(published_chunk
);
4917 lttng_trace_chunk_put(created_chunk
);
4921 enum lttcomm_return_code
lttng_consumer_close_trace_chunk(
4922 const uint64_t *relayd_id
, uint64_t session_id
,
4923 uint64_t chunk_id
, time_t chunk_close_timestamp
,
4924 const enum lttng_trace_chunk_command_type
*close_command
,
4927 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4928 struct lttng_trace_chunk
*chunk
;
4929 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4930 const char *relayd_id_str
= "(none)";
4931 const char *close_command_name
= "none";
4932 struct lttng_ht_iter iter
;
4933 struct lttng_consumer_channel
*channel
;
4934 enum lttng_trace_chunk_status chunk_status
;
4939 /* Only used for logging purposes. */
4940 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4941 "%" PRIu64
, *relayd_id
);
4942 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4943 relayd_id_str
= relayd_id_buffer
;
4945 relayd_id_str
= "(formatting error)";
4948 if (close_command
) {
4949 close_command_name
= lttng_trace_chunk_command_type_get_name(
4953 DBG("Consumer close trace chunk command: relayd_id = %s"
4954 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4955 ", close command = %s",
4956 relayd_id_str
, session_id
, chunk_id
,
4957 close_command_name
);
4959 chunk
= lttng_trace_chunk_registry_find_chunk(
4960 the_consumer_data
.chunk_registry
, session_id
, chunk_id
);
4962 ERR("Failed to find chunk: session_id = %" PRIu64
4963 ", chunk_id = %" PRIu64
,
4964 session_id
, chunk_id
);
4965 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4969 chunk_status
= lttng_trace_chunk_set_close_timestamp(chunk
,
4970 chunk_close_timestamp
);
4971 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4972 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4976 if (close_command
) {
4977 chunk_status
= lttng_trace_chunk_set_close_command(
4978 chunk
, *close_command
);
4979 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4980 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4986 * chunk is now invalid to access as we no longer hold a reference to
4987 * it; it is only kept around to compare it (by address) to the
4988 * current chunk found in the session's channels.
4991 cds_lfht_for_each_entry(the_consumer_data
.channel_ht
->ht
, &iter
.iter
,
4992 channel
, node
.node
) {
4996 * Only change the channel's chunk to NULL if it still
4997 * references the chunk being closed. The channel may
4998 * reference a newer channel in the case of a session
4999 * rotation. When a session rotation occurs, the "next"
5000 * chunk is created before the "current" chunk is closed.
5002 if (channel
->trace_chunk
!= chunk
) {
5005 ret
= lttng_consumer_channel_set_trace_chunk(channel
, NULL
);
5008 * Attempt to close the chunk on as many channels as
5011 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
5017 struct consumer_relayd_sock_pair
*relayd
;
5019 relayd
= consumer_find_relayd(*relayd_id
);
5021 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
5022 ret
= relayd_close_trace_chunk(
5023 &relayd
->control_sock
, chunk
,
5025 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
5027 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
,
5031 if (!relayd
|| ret
) {
5032 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
5040 * Release the reference returned by the "find" operation and
5041 * the session daemon's implicit reference to the chunk.
5043 lttng_trace_chunk_put(chunk
);
5044 lttng_trace_chunk_put(chunk
);
5049 enum lttcomm_return_code
lttng_consumer_trace_chunk_exists(
5050 const uint64_t *relayd_id
, uint64_t session_id
,
5054 enum lttcomm_return_code ret_code
;
5055 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
5056 const char *relayd_id_str
= "(none)";
5057 const bool is_local_trace
= !relayd_id
;
5058 struct consumer_relayd_sock_pair
*relayd
= NULL
;
5059 bool chunk_exists_local
, chunk_exists_remote
;
5062 /* Only used for logging purposes. */
5063 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
5064 "%" PRIu64
, *relayd_id
);
5065 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
5066 relayd_id_str
= relayd_id_buffer
;
5068 relayd_id_str
= "(formatting error)";
5072 DBG("Consumer trace chunk exists command: relayd_id = %s"
5073 ", chunk_id = %" PRIu64
, relayd_id_str
,
5075 ret
= lttng_trace_chunk_registry_chunk_exists(
5076 the_consumer_data
.chunk_registry
, session_id
, chunk_id
,
5077 &chunk_exists_local
);
5079 /* Internal error. */
5080 ERR("Failed to query the existence of a trace chunk");
5081 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
5084 DBG("Trace chunk %s locally",
5085 chunk_exists_local
? "exists" : "does not exist");
5086 if (chunk_exists_local
) {
5087 ret_code
= LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_LOCAL
;
5089 } else if (is_local_trace
) {
5090 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
5095 relayd
= consumer_find_relayd(*relayd_id
);
5097 ERR("Failed to find relayd %" PRIu64
, *relayd_id
);
5098 ret_code
= LTTCOMM_CONSUMERD_INVALID_PARAMETERS
;
5099 goto end_rcu_unlock
;
5101 DBG("Looking up existence of trace chunk on relay daemon");
5102 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
5103 ret
= relayd_trace_chunk_exists(&relayd
->control_sock
, chunk_id
,
5104 &chunk_exists_remote
);
5105 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
5107 ERR("Failed to look-up the existence of trace chunk on relay daemon");
5108 ret_code
= LTTCOMM_CONSUMERD_RELAYD_FAIL
;
5109 goto end_rcu_unlock
;
5112 ret_code
= chunk_exists_remote
?
5113 LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_REMOTE
:
5114 LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
5115 DBG("Trace chunk %s on relay daemon",
5116 chunk_exists_remote
? "exists" : "does not exist");
5125 int consumer_clear_monitored_channel(struct lttng_consumer_channel
*channel
)
5127 struct lttng_ht
*ht
;
5128 struct lttng_consumer_stream
*stream
;
5129 struct lttng_ht_iter iter
;
5132 ht
= the_consumer_data
.stream_per_chan_id_ht
;
5135 cds_lfht_for_each_entry_duplicate(ht
->ht
,
5136 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
5137 ht
->match_fct
, &channel
->key
,
5138 &iter
.iter
, stream
, node_channel_id
.node
) {
5140 * Protect against teardown with mutex.
5142 pthread_mutex_lock(&stream
->lock
);
5143 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
5146 ret
= consumer_clear_stream(stream
);
5151 pthread_mutex_unlock(&stream
->lock
);
5154 return LTTCOMM_CONSUMERD_SUCCESS
;
5157 pthread_mutex_unlock(&stream
->lock
);
5162 int lttng_consumer_clear_channel(struct lttng_consumer_channel
*channel
)
5166 DBG("Consumer clear channel %" PRIu64
, channel
->key
);
5168 if (channel
->type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
5170 * Nothing to do for the metadata channel/stream.
5171 * Snapshot mechanism already take care of the metadata
5172 * handling/generation, and monitored channels only need to
5173 * have their data stream cleared..
5175 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
5179 if (!channel
->monitor
) {
5180 ret
= consumer_clear_unmonitored_channel(channel
);
5182 ret
= consumer_clear_monitored_channel(channel
);
5188 enum lttcomm_return_code
lttng_consumer_open_channel_packets(
5189 struct lttng_consumer_channel
*channel
)
5191 struct lttng_consumer_stream
*stream
;
5192 enum lttcomm_return_code ret
= LTTCOMM_CONSUMERD_SUCCESS
;
5194 if (channel
->metadata_stream
) {
5195 ERR("Open channel packets command attempted on a metadata channel");
5196 ret
= LTTCOMM_CONSUMERD_INVALID_PARAMETERS
;
5201 cds_list_for_each_entry(stream
, &channel
->streams
.head
, send_node
) {
5202 enum consumer_stream_open_packet_status status
;
5204 pthread_mutex_lock(&stream
->lock
);
5205 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
5209 status
= consumer_stream_open_packet(stream
);
5211 case CONSUMER_STREAM_OPEN_PACKET_STATUS_OPENED
:
5212 DBG("Opened a packet in \"open channel packets\" command: stream id = %" PRIu64
5213 ", channel name = %s, session id = %" PRIu64
,
5214 stream
->key
, stream
->chan
->name
,
5215 stream
->chan
->session_id
);
5216 stream
->opened_packet_in_current_trace_chunk
= true;
5218 case CONSUMER_STREAM_OPEN_PACKET_STATUS_NO_SPACE
:
5219 DBG("No space left to open a packet in \"open channel packets\" command: stream id = %" PRIu64
5220 ", channel name = %s, session id = %" PRIu64
,
5221 stream
->key
, stream
->chan
->name
,
5222 stream
->chan
->session_id
);
5224 case CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR
:
5226 * Only unexpected internal errors can lead to this
5227 * failing. Report an unknown error.
5229 ERR("Failed to flush empty buffer in \"open channel packets\" command: stream id = %" PRIu64
5230 ", channel id = %" PRIu64
5231 ", channel name = %s"
5232 ", session id = %" PRIu64
,
5233 stream
->key
, channel
->key
,
5234 channel
->name
, channel
->session_id
);
5235 ret
= LTTCOMM_CONSUMERD_UNKNOWN_ERROR
;
5242 pthread_mutex_unlock(&stream
->lock
);
5251 pthread_mutex_unlock(&stream
->lock
);
5252 goto end_rcu_unlock
;
5255 void lttng_consumer_sigbus_handle(void *addr
)
5257 lttng_ustconsumer_sigbus_handle(addr
);