2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
20 #include "common/index/ctf-index.h"
28 #include <sys/socket.h>
29 #include <sys/types.h>
34 #include <bin/lttng-consumerd/health-consumerd.h>
35 #include <common/common.h>
36 #include <common/utils.h>
37 #include <common/time.h>
38 #include <common/compat/poll.h>
39 #include <common/compat/endian.h>
40 #include <common/index/index.h>
41 #include <common/kernel-ctl/kernel-ctl.h>
42 #include <common/sessiond-comm/relayd.h>
43 #include <common/sessiond-comm/sessiond-comm.h>
44 #include <common/kernel-consumer/kernel-consumer.h>
45 #include <common/relayd/relayd.h>
46 #include <common/ust-consumer/ust-consumer.h>
47 #include <common/consumer/consumer-timer.h>
48 #include <common/consumer/consumer.h>
49 #include <common/consumer/consumer-stream.h>
50 #include <common/consumer/consumer-testpoint.h>
51 #include <common/align.h>
52 #include <common/consumer/consumer-metadata-cache.h>
53 #include <common/trace-chunk.h>
54 #include <common/trace-chunk-registry.h>
55 #include <common/string-utils/format.h>
56 #include <common/dynamic-array.h>
58 struct lttng_consumer_global_data consumer_data
= {
61 .type
= LTTNG_CONSUMER_UNKNOWN
,
64 enum consumer_channel_action
{
67 CONSUMER_CHANNEL_QUIT
,
70 struct consumer_channel_msg
{
71 enum consumer_channel_action action
;
72 struct lttng_consumer_channel
*chan
; /* add */
73 uint64_t key
; /* del */
76 /* Flag used to temporarily pause data consumption from testpoints. */
77 int data_consumption_paused
;
80 * Flag to inform the polling thread to quit when all fd hung up. Updated by
81 * the consumer_thread_receive_fds when it notices that all fds has hung up.
82 * Also updated by the signal handler (consumer_should_exit()). Read by the
88 * Global hash table containing respectively metadata and data streams. The
89 * stream element in this ht should only be updated by the metadata poll thread
90 * for the metadata and the data poll thread for the data.
92 static struct lttng_ht
*metadata_ht
;
93 static struct lttng_ht
*data_ht
;
96 * Notify a thread lttng pipe to poll back again. This usually means that some
97 * global state has changed so we just send back the thread in a poll wait
100 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
102 struct lttng_consumer_stream
*null_stream
= NULL
;
106 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
109 static void notify_health_quit_pipe(int *pipe
)
113 ret
= lttng_write(pipe
[1], "4", 1);
115 PERROR("write consumer health quit");
119 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
120 struct lttng_consumer_channel
*chan
,
122 enum consumer_channel_action action
)
124 struct consumer_channel_msg msg
;
127 memset(&msg
, 0, sizeof(msg
));
132 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
133 if (ret
< sizeof(msg
)) {
134 PERROR("notify_channel_pipe write error");
138 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
141 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
144 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
145 struct lttng_consumer_channel
**chan
,
147 enum consumer_channel_action
*action
)
149 struct consumer_channel_msg msg
;
152 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
153 if (ret
< sizeof(msg
)) {
157 *action
= msg
.action
;
165 * Cleanup the stream list of a channel. Those streams are not yet globally
168 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
170 struct lttng_consumer_stream
*stream
, *stmp
;
174 /* Delete streams that might have been left in the stream list. */
175 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
177 cds_list_del(&stream
->send_node
);
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
= caa_container_of(node
, struct lttng_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 /* -1ULL keys are lookup failures */
251 if (key
== (uint64_t) -1ULL) {
255 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
256 node
= lttng_ht_iter_get_node_u64(&iter
);
258 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
265 * There is a possibility that the consumer does not have enough time between
266 * the close of the channel on the session daemon and the cleanup in here thus
267 * once we have a channel add with an existing key, we know for sure that this
268 * channel will eventually get cleaned up by all streams being closed.
270 * This function just nullifies the already existing channel key.
272 static void steal_channel_key(uint64_t key
)
274 struct lttng_consumer_channel
*channel
;
277 channel
= consumer_find_channel(key
);
279 channel
->key
= (uint64_t) -1ULL;
281 * We don't want the lookup to match, but we still need to iterate on
282 * this channel when iterating over the hash table. Just change the
285 channel
->node
.key
= (uint64_t) -1ULL;
290 static void free_channel_rcu(struct rcu_head
*head
)
292 struct lttng_ht_node_u64
*node
=
293 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
294 struct lttng_consumer_channel
*channel
=
295 caa_container_of(node
, struct lttng_consumer_channel
, node
);
297 switch (consumer_data
.type
) {
298 case LTTNG_CONSUMER_KERNEL
:
300 case LTTNG_CONSUMER32_UST
:
301 case LTTNG_CONSUMER64_UST
:
302 lttng_ustconsumer_free_channel(channel
);
305 ERR("Unknown consumer_data type");
312 * RCU protected relayd socket pair free.
314 static void free_relayd_rcu(struct rcu_head
*head
)
316 struct lttng_ht_node_u64
*node
=
317 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
318 struct consumer_relayd_sock_pair
*relayd
=
319 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
322 * Close all sockets. This is done in the call RCU since we don't want the
323 * socket fds to be reassigned thus potentially creating bad state of the
326 * We do not have to lock the control socket mutex here since at this stage
327 * there is no one referencing to this relayd object.
329 (void) relayd_close(&relayd
->control_sock
);
330 (void) relayd_close(&relayd
->data_sock
);
332 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
337 * Destroy and free relayd socket pair object.
339 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
342 struct lttng_ht_iter iter
;
344 if (relayd
== NULL
) {
348 DBG("Consumer destroy and close relayd socket pair");
350 iter
.iter
.node
= &relayd
->node
.node
;
351 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
353 /* We assume the relayd is being or is destroyed */
357 /* RCU free() call */
358 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
362 * Remove a channel from the global list protected by a mutex. This function is
363 * also responsible for freeing its data structures.
365 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
367 struct lttng_ht_iter iter
;
369 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
371 pthread_mutex_lock(&consumer_data
.lock
);
372 pthread_mutex_lock(&channel
->lock
);
374 /* Destroy streams that might have been left in the stream list. */
375 clean_channel_stream_list(channel
);
377 if (channel
->live_timer_enabled
== 1) {
378 consumer_timer_live_stop(channel
);
380 if (channel
->monitor_timer_enabled
== 1) {
381 consumer_timer_monitor_stop(channel
);
384 switch (consumer_data
.type
) {
385 case LTTNG_CONSUMER_KERNEL
:
387 case LTTNG_CONSUMER32_UST
:
388 case LTTNG_CONSUMER64_UST
:
389 lttng_ustconsumer_del_channel(channel
);
392 ERR("Unknown consumer_data type");
397 lttng_trace_chunk_put(channel
->trace_chunk
);
398 channel
->trace_chunk
= NULL
;
400 if (channel
->is_published
) {
404 iter
.iter
.node
= &channel
->node
.node
;
405 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
408 iter
.iter
.node
= &channel
->channels_by_session_id_ht_node
.node
;
409 ret
= lttng_ht_del(consumer_data
.channels_by_session_id_ht
,
415 channel
->is_deleted
= true;
416 call_rcu(&channel
->node
.head
, free_channel_rcu
);
418 pthread_mutex_unlock(&channel
->lock
);
419 pthread_mutex_unlock(&consumer_data
.lock
);
423 * Iterate over the relayd hash table and destroy each element. Finally,
424 * destroy the whole hash table.
426 static void cleanup_relayd_ht(void)
428 struct lttng_ht_iter iter
;
429 struct consumer_relayd_sock_pair
*relayd
;
433 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
435 consumer_destroy_relayd(relayd
);
440 lttng_ht_destroy(consumer_data
.relayd_ht
);
444 * Update the end point status of all streams having the given network sequence
445 * index (relayd index).
447 * It's atomically set without having the stream mutex locked which is fine
448 * because we handle the write/read race with a pipe wakeup for each thread.
450 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
451 enum consumer_endpoint_status status
)
453 struct lttng_ht_iter iter
;
454 struct lttng_consumer_stream
*stream
;
456 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
460 /* Let's begin with metadata */
461 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
462 if (stream
->net_seq_idx
== net_seq_idx
) {
463 uatomic_set(&stream
->endpoint_status
, status
);
464 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
468 /* Follow up by the data streams */
469 cds_lfht_for_each_entry(data_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 data stream %d", stream
->wait_fd
);
479 * Cleanup a relayd object by flagging every associated streams for deletion,
480 * destroying the object meaning removing it from the relayd hash table,
481 * closing the sockets and freeing the memory in a RCU call.
483 * If a local data context is available, notify the threads that the streams'
484 * state have changed.
486 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
492 DBG("Cleaning up relayd object ID %"PRIu64
, relayd
->net_seq_idx
);
494 /* Save the net sequence index before destroying the object */
495 netidx
= relayd
->net_seq_idx
;
498 * Delete the relayd from the relayd hash table, close the sockets and free
499 * the object in a RCU call.
501 consumer_destroy_relayd(relayd
);
503 /* Set inactive endpoint to all streams */
504 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
507 * With a local data context, notify the threads that the streams' state
508 * have changed. The write() action on the pipe acts as an "implicit"
509 * memory barrier ordering the updates of the end point status from the
510 * read of this status which happens AFTER receiving this notify.
512 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
513 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
517 * Flag a relayd socket pair for destruction. Destroy it if the refcount
520 * RCU read side lock MUST be aquired before calling this function.
522 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
526 /* Set destroy flag for this object */
527 uatomic_set(&relayd
->destroy_flag
, 1);
529 /* Destroy the relayd if refcount is 0 */
530 if (uatomic_read(&relayd
->refcount
) == 0) {
531 consumer_destroy_relayd(relayd
);
536 * Completly destroy stream from every visiable data structure and the given
539 * One this call returns, the stream object is not longer usable nor visible.
541 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
544 consumer_stream_destroy(stream
, ht
);
548 * XXX naming of del vs destroy is all mixed up.
550 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
552 consumer_stream_destroy(stream
, data_ht
);
555 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
557 consumer_stream_destroy(stream
, metadata_ht
);
560 void consumer_stream_update_channel_attributes(
561 struct lttng_consumer_stream
*stream
,
562 struct lttng_consumer_channel
*channel
)
564 stream
->channel_read_only_attributes
.tracefile_size
=
565 channel
->tracefile_size
;
569 * Add a stream to the global list protected by a mutex.
571 void consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
573 struct lttng_ht
*ht
= data_ht
;
578 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
580 pthread_mutex_lock(&consumer_data
.lock
);
581 pthread_mutex_lock(&stream
->chan
->lock
);
582 pthread_mutex_lock(&stream
->chan
->timer_lock
);
583 pthread_mutex_lock(&stream
->lock
);
586 /* Steal stream identifier to avoid having streams with the same key */
587 steal_stream_key(stream
->key
, ht
);
589 lttng_ht_add_unique_u64(ht
, &stream
->node
);
591 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
592 &stream
->node_channel_id
);
595 * Add stream to the stream_list_ht of the consumer data. No need to steal
596 * the key since the HT does not use it and we allow to add redundant keys
599 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
602 * When nb_init_stream_left reaches 0, we don't need to trigger any action
603 * in terms of destroying the associated channel, because the action that
604 * causes the count to become 0 also causes a stream to be added. The
605 * channel deletion will thus be triggered by the following removal of this
608 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
609 /* Increment refcount before decrementing nb_init_stream_left */
611 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
614 /* Update consumer data once the node is inserted. */
615 consumer_data
.stream_count
++;
616 consumer_data
.need_update
= 1;
619 pthread_mutex_unlock(&stream
->lock
);
620 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
621 pthread_mutex_unlock(&stream
->chan
->lock
);
622 pthread_mutex_unlock(&consumer_data
.lock
);
625 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
627 consumer_del_stream(stream
, data_ht
);
631 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
632 * be acquired before calling this.
634 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
637 struct lttng_ht_node_u64
*node
;
638 struct lttng_ht_iter iter
;
642 lttng_ht_lookup(consumer_data
.relayd_ht
,
643 &relayd
->net_seq_idx
, &iter
);
644 node
= lttng_ht_iter_get_node_u64(&iter
);
648 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
655 * Allocate and return a consumer relayd socket.
657 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
658 uint64_t net_seq_idx
)
660 struct consumer_relayd_sock_pair
*obj
= NULL
;
662 /* net sequence index of -1 is a failure */
663 if (net_seq_idx
== (uint64_t) -1ULL) {
667 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
669 PERROR("zmalloc relayd sock");
673 obj
->net_seq_idx
= net_seq_idx
;
675 obj
->destroy_flag
= 0;
676 obj
->control_sock
.sock
.fd
= -1;
677 obj
->data_sock
.sock
.fd
= -1;
678 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
679 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
686 * Find a relayd socket pair in the global consumer data.
688 * Return the object if found else NULL.
689 * RCU read-side lock must be held across this call and while using the
692 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
694 struct lttng_ht_iter iter
;
695 struct lttng_ht_node_u64
*node
;
696 struct consumer_relayd_sock_pair
*relayd
= NULL
;
698 /* Negative keys are lookup failures */
699 if (key
== (uint64_t) -1ULL) {
703 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
705 node
= lttng_ht_iter_get_node_u64(&iter
);
707 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
715 * Find a relayd and send the stream
717 * Returns 0 on success, < 0 on error
719 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
723 struct consumer_relayd_sock_pair
*relayd
;
726 assert(stream
->net_seq_idx
!= -1ULL);
729 /* The stream is not metadata. Get relayd reference if exists. */
731 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
732 if (relayd
!= NULL
) {
733 /* Add stream on the relayd */
734 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
735 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
736 path
, &stream
->relayd_stream_id
,
737 stream
->chan
->tracefile_size
,
738 stream
->chan
->tracefile_count
,
739 stream
->trace_chunk
);
740 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
742 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
743 lttng_consumer_cleanup_relayd(relayd
);
747 uatomic_inc(&relayd
->refcount
);
748 stream
->sent_to_relayd
= 1;
750 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
751 stream
->key
, stream
->net_seq_idx
);
756 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
757 stream
->name
, stream
->key
, stream
->net_seq_idx
);
765 * Find a relayd and send the streams sent message
767 * Returns 0 on success, < 0 on error
769 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
772 struct consumer_relayd_sock_pair
*relayd
;
774 assert(net_seq_idx
!= -1ULL);
776 /* The stream is not metadata. Get relayd reference if exists. */
778 relayd
= consumer_find_relayd(net_seq_idx
);
779 if (relayd
!= NULL
) {
780 /* Add stream on the relayd */
781 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
782 ret
= relayd_streams_sent(&relayd
->control_sock
);
783 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
785 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
786 lttng_consumer_cleanup_relayd(relayd
);
790 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
797 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
805 * Find a relayd and close the stream
807 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
809 struct consumer_relayd_sock_pair
*relayd
;
811 /* The stream is not metadata. Get relayd reference if exists. */
813 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
815 consumer_stream_relayd_close(stream
, relayd
);
821 * Handle stream for relayd transmission if the stream applies for network
822 * streaming where the net sequence index is set.
824 * Return destination file descriptor or negative value on error.
826 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
827 size_t data_size
, unsigned long padding
,
828 struct consumer_relayd_sock_pair
*relayd
)
831 struct lttcomm_relayd_data_hdr data_hdr
;
837 /* Reset data header */
838 memset(&data_hdr
, 0, sizeof(data_hdr
));
840 if (stream
->metadata_flag
) {
841 /* Caller MUST acquire the relayd control socket lock */
842 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
847 /* Metadata are always sent on the control socket. */
848 outfd
= relayd
->control_sock
.sock
.fd
;
850 /* Set header with stream information */
851 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
852 data_hdr
.data_size
= htobe32(data_size
);
853 data_hdr
.padding_size
= htobe32(padding
);
856 * Note that net_seq_num below is assigned with the *current* value of
857 * next_net_seq_num and only after that the next_net_seq_num will be
858 * increment. This is why when issuing a command on the relayd using
859 * this next value, 1 should always be substracted in order to compare
860 * the last seen sequence number on the relayd side to the last sent.
862 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
863 /* Other fields are zeroed previously */
865 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
871 ++stream
->next_net_seq_num
;
873 /* Set to go on data socket */
874 outfd
= relayd
->data_sock
.sock
.fd
;
882 * Write a character on the metadata poll pipe to wake the metadata thread.
883 * Returns 0 on success, -1 on error.
885 int consumer_metadata_wakeup_pipe(const struct lttng_consumer_channel
*channel
)
889 DBG("Waking up metadata poll thread (writing to pipe): channel name = '%s'",
891 if (channel
->monitor
&& channel
->metadata_stream
) {
892 const char dummy
= 'c';
893 const ssize_t write_ret
= lttng_write(
894 channel
->metadata_stream
->ust_metadata_poll_pipe
[1],
898 if (errno
== EWOULDBLOCK
) {
900 * This is fine, the metadata poll thread
901 * is having a hard time keeping-up, but
902 * it will eventually wake-up and consume
903 * the available data.
907 PERROR("Failed to write to UST metadata pipe while attempting to wake-up the metadata poll thread");
919 * Trigger a dump of the metadata content. Following/during the succesful
920 * completion of this call, the metadata poll thread will start receiving
921 * metadata packets to consume.
923 * The caller must hold the channel and stream locks.
926 int consumer_metadata_stream_dump(struct lttng_consumer_stream
*stream
)
930 ASSERT_LOCKED(stream
->chan
->lock
);
931 ASSERT_LOCKED(stream
->lock
);
932 assert(stream
->metadata_flag
);
933 assert(stream
->chan
->trace_chunk
);
935 switch (consumer_data
.type
) {
936 case LTTNG_CONSUMER_KERNEL
:
938 * Reset the position of what has been read from the
939 * metadata cache to 0 so we can dump it again.
941 ret
= kernctl_metadata_cache_dump(stream
->wait_fd
);
943 case LTTNG_CONSUMER32_UST
:
944 case LTTNG_CONSUMER64_UST
:
946 * Reset the position pushed from the metadata cache so it
947 * will write from the beginning on the next push.
949 stream
->ust_metadata_pushed
= 0;
950 ret
= consumer_metadata_wakeup_pipe(stream
->chan
);
953 ERR("Unknown consumer_data type");
957 ERR("Failed to dump the metadata cache");
963 int lttng_consumer_channel_set_trace_chunk(
964 struct lttng_consumer_channel
*channel
,
965 struct lttng_trace_chunk
*new_trace_chunk
)
967 pthread_mutex_lock(&channel
->lock
);
968 if (channel
->is_deleted
) {
970 * The channel has been logically deleted and should no longer
971 * be used. It has released its reference to its current trace
972 * chunk and should not acquire a new one.
974 * Return success as there is nothing for the caller to do.
980 * The acquisition of the reference cannot fail (barring
981 * a severe internal error) since a reference to the published
982 * chunk is already held by the caller.
984 if (new_trace_chunk
) {
985 const bool acquired_reference
= lttng_trace_chunk_get(
988 assert(acquired_reference
);
991 lttng_trace_chunk_put(channel
->trace_chunk
);
992 channel
->trace_chunk
= new_trace_chunk
;
994 pthread_mutex_unlock(&channel
->lock
);
999 * Allocate and return a new lttng_consumer_channel object using the given key
1000 * to initialize the hash table node.
1002 * On error, return NULL.
1004 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
1005 uint64_t session_id
,
1006 const uint64_t *chunk_id
,
1007 const char *pathname
,
1010 enum lttng_event_output output
,
1011 uint64_t tracefile_size
,
1012 uint64_t tracefile_count
,
1013 uint64_t session_id_per_pid
,
1014 unsigned int monitor
,
1015 unsigned int live_timer_interval
,
1016 bool is_in_live_session
,
1017 const char *root_shm_path
,
1018 const char *shm_path
)
1020 struct lttng_consumer_channel
*channel
= NULL
;
1021 struct lttng_trace_chunk
*trace_chunk
= NULL
;
1024 trace_chunk
= lttng_trace_chunk_registry_find_chunk(
1025 consumer_data
.chunk_registry
, session_id
,
1028 ERR("Failed to find trace chunk reference during creation of channel");
1033 channel
= zmalloc(sizeof(*channel
));
1034 if (channel
== NULL
) {
1035 PERROR("malloc struct lttng_consumer_channel");
1040 channel
->refcount
= 0;
1041 channel
->session_id
= session_id
;
1042 channel
->session_id_per_pid
= session_id_per_pid
;
1043 channel
->relayd_id
= relayd_id
;
1044 channel
->tracefile_size
= tracefile_size
;
1045 channel
->tracefile_count
= tracefile_count
;
1046 channel
->monitor
= monitor
;
1047 channel
->live_timer_interval
= live_timer_interval
;
1048 channel
->is_live
= is_in_live_session
;
1049 pthread_mutex_init(&channel
->lock
, NULL
);
1050 pthread_mutex_init(&channel
->timer_lock
, NULL
);
1053 case LTTNG_EVENT_SPLICE
:
1054 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
1056 case LTTNG_EVENT_MMAP
:
1057 channel
->output
= CONSUMER_CHANNEL_MMAP
;
1067 * In monitor mode, the streams associated with the channel will be put in
1068 * a special list ONLY owned by this channel. So, the refcount is set to 1
1069 * here meaning that the channel itself has streams that are referenced.
1071 * On a channel deletion, once the channel is no longer visible, the
1072 * refcount is decremented and checked for a zero value to delete it. With
1073 * streams in no monitor mode, it will now be safe to destroy the channel.
1075 if (!channel
->monitor
) {
1076 channel
->refcount
= 1;
1079 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1080 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1082 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1083 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1085 if (root_shm_path
) {
1086 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1087 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1090 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1091 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1094 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1095 lttng_ht_node_init_u64(&channel
->channels_by_session_id_ht_node
,
1096 channel
->session_id
);
1098 channel
->wait_fd
= -1;
1099 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1102 int ret
= lttng_consumer_channel_set_trace_chunk(channel
,
1109 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1112 lttng_trace_chunk_put(trace_chunk
);
1115 consumer_del_channel(channel
);
1121 * Add a channel to the global list protected by a mutex.
1123 * Always return 0 indicating success.
1125 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1126 struct lttng_consumer_local_data
*ctx
)
1128 pthread_mutex_lock(&consumer_data
.lock
);
1129 pthread_mutex_lock(&channel
->lock
);
1130 pthread_mutex_lock(&channel
->timer_lock
);
1133 * This gives us a guarantee that the channel we are about to add to the
1134 * channel hash table will be unique. See this function comment on the why
1135 * we need to steel the channel key at this stage.
1137 steal_channel_key(channel
->key
);
1140 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1141 lttng_ht_add_u64(consumer_data
.channels_by_session_id_ht
,
1142 &channel
->channels_by_session_id_ht_node
);
1144 channel
->is_published
= true;
1146 pthread_mutex_unlock(&channel
->timer_lock
);
1147 pthread_mutex_unlock(&channel
->lock
);
1148 pthread_mutex_unlock(&consumer_data
.lock
);
1150 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1151 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1158 * Allocate the pollfd structure and the local view of the out fds to avoid
1159 * doing a lookup in the linked list and concurrency issues when writing is
1160 * needed. Called with consumer_data.lock held.
1162 * Returns the number of fds in the structures.
1164 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1165 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1166 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1169 struct lttng_ht_iter iter
;
1170 struct lttng_consumer_stream
*stream
;
1175 assert(local_stream
);
1177 DBG("Updating poll fd array");
1178 *nb_inactive_fd
= 0;
1180 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1182 * Only active streams with an active end point can be added to the
1183 * poll set and local stream storage of the thread.
1185 * There is a potential race here for endpoint_status to be updated
1186 * just after the check. However, this is OK since the stream(s) will
1187 * be deleted once the thread is notified that the end point state has
1188 * changed where this function will be called back again.
1190 * We track the number of inactive FDs because they still need to be
1191 * closed by the polling thread after a wakeup on the data_pipe or
1194 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1195 (*nb_inactive_fd
)++;
1199 * This clobbers way too much the debug output. Uncomment that if you
1200 * need it for debugging purposes.
1202 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1203 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1204 local_stream
[i
] = stream
;
1210 * Insert the consumer_data_pipe at the end of the array and don't
1211 * increment i so nb_fd is the number of real FD.
1213 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1214 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1216 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1217 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1222 * Poll on the should_quit pipe and the command socket return -1 on
1223 * error, 1 if should exit, 0 if data is available on the command socket
1225 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1230 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1231 if (num_rdy
== -1) {
1233 * Restart interrupted system call.
1235 if (errno
== EINTR
) {
1238 PERROR("Poll error");
1241 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1242 DBG("consumer_should_quit wake up");
1249 * Set the error socket.
1251 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1254 ctx
->consumer_error_socket
= sock
;
1258 * Set the command socket path.
1260 void lttng_consumer_set_command_sock_path(
1261 struct lttng_consumer_local_data
*ctx
, char *sock
)
1263 ctx
->consumer_command_sock_path
= sock
;
1267 * Send return code to the session daemon.
1268 * If the socket is not defined, we return 0, it is not a fatal error
1270 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1272 if (ctx
->consumer_error_socket
> 0) {
1273 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1274 sizeof(enum lttcomm_sessiond_command
));
1281 * Close all the tracefiles and stream fds and MUST be called when all
1282 * instances are destroyed i.e. when all threads were joined and are ended.
1284 void lttng_consumer_cleanup(void)
1286 struct lttng_ht_iter iter
;
1287 struct lttng_consumer_channel
*channel
;
1288 unsigned int trace_chunks_left
;
1292 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1294 consumer_del_channel(channel
);
1299 lttng_ht_destroy(consumer_data
.channel_ht
);
1300 lttng_ht_destroy(consumer_data
.channels_by_session_id_ht
);
1302 cleanup_relayd_ht();
1304 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1307 * This HT contains streams that are freed by either the metadata thread or
1308 * the data thread so we do *nothing* on the hash table and simply destroy
1311 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1314 * Trace chunks in the registry may still exist if the session
1315 * daemon has encountered an internal error and could not
1316 * tear down its sessions and/or trace chunks properly.
1318 * Release the session daemon's implicit reference to any remaining
1319 * trace chunk and print an error if any trace chunk was found. Note
1320 * that there are _no_ legitimate cases for trace chunks to be left,
1321 * it is a leak. However, it can happen following a crash of the
1322 * session daemon and not emptying the registry would cause an assertion
1325 trace_chunks_left
= lttng_trace_chunk_registry_put_each_chunk(
1326 consumer_data
.chunk_registry
);
1327 if (trace_chunks_left
) {
1328 ERR("%u trace chunks are leaked by lttng-consumerd. "
1329 "This can be caused by an internal error of the session daemon.",
1332 /* Run all callbacks freeing each chunk. */
1334 lttng_trace_chunk_registry_destroy(consumer_data
.chunk_registry
);
1338 * Called from signal handler.
1340 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1344 CMM_STORE_SHARED(consumer_quit
, 1);
1345 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1347 PERROR("write consumer quit");
1350 DBG("Consumer flag that it should quit");
1355 * Flush pending writes to trace output disk file.
1358 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1362 int outfd
= stream
->out_fd
;
1365 * This does a blocking write-and-wait on any page that belongs to the
1366 * subbuffer prior to the one we just wrote.
1367 * Don't care about error values, as these are just hints and ways to
1368 * limit the amount of page cache used.
1370 if (orig_offset
< stream
->max_sb_size
) {
1373 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1374 stream
->max_sb_size
,
1375 SYNC_FILE_RANGE_WAIT_BEFORE
1376 | SYNC_FILE_RANGE_WRITE
1377 | SYNC_FILE_RANGE_WAIT_AFTER
);
1379 * Give hints to the kernel about how we access the file:
1380 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1383 * We need to call fadvise again after the file grows because the
1384 * kernel does not seem to apply fadvise to non-existing parts of the
1387 * Call fadvise _after_ having waited for the page writeback to
1388 * complete because the dirty page writeback semantic is not well
1389 * defined. So it can be expected to lead to lower throughput in
1392 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1393 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1394 if (ret
&& ret
!= -ENOSYS
) {
1396 PERROR("posix_fadvise on fd %i", outfd
);
1401 * Initialise the necessary environnement :
1402 * - create a new context
1403 * - create the poll_pipe
1404 * - create the should_quit pipe (for signal handler)
1405 * - create the thread pipe (for splice)
1407 * Takes a function pointer as argument, this function is called when data is
1408 * available on a buffer. This function is responsible to do the
1409 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1410 * buffer configuration and then kernctl_put_next_subbuf at the end.
1412 * Returns a pointer to the new context or NULL on error.
1414 struct lttng_consumer_local_data
*lttng_consumer_create(
1415 enum lttng_consumer_type type
,
1416 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1417 struct lttng_consumer_local_data
*ctx
, bool locked_by_caller
),
1418 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1419 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1420 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1423 struct lttng_consumer_local_data
*ctx
;
1425 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1426 consumer_data
.type
== type
);
1427 consumer_data
.type
= type
;
1429 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1431 PERROR("allocating context");
1435 ctx
->consumer_error_socket
= -1;
1436 ctx
->consumer_metadata_socket
= -1;
1437 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1438 /* assign the callbacks */
1439 ctx
->on_buffer_ready
= buffer_ready
;
1440 ctx
->on_recv_channel
= recv_channel
;
1441 ctx
->on_recv_stream
= recv_stream
;
1442 ctx
->on_update_stream
= update_stream
;
1444 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1445 if (!ctx
->consumer_data_pipe
) {
1446 goto error_poll_pipe
;
1449 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1450 if (!ctx
->consumer_wakeup_pipe
) {
1451 goto error_wakeup_pipe
;
1454 ret
= pipe(ctx
->consumer_should_quit
);
1456 PERROR("Error creating recv pipe");
1457 goto error_quit_pipe
;
1460 ret
= pipe(ctx
->consumer_channel_pipe
);
1462 PERROR("Error creating channel pipe");
1463 goto error_channel_pipe
;
1466 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1467 if (!ctx
->consumer_metadata_pipe
) {
1468 goto error_metadata_pipe
;
1471 ctx
->channel_monitor_pipe
= -1;
1475 error_metadata_pipe
:
1476 utils_close_pipe(ctx
->consumer_channel_pipe
);
1478 utils_close_pipe(ctx
->consumer_should_quit
);
1480 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1482 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1490 * Iterate over all streams of the hashtable and free them properly.
1492 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1494 struct lttng_ht_iter iter
;
1495 struct lttng_consumer_stream
*stream
;
1502 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1504 * Ignore return value since we are currently cleaning up so any error
1507 (void) consumer_del_stream(stream
, ht
);
1511 lttng_ht_destroy(ht
);
1515 * Iterate over all streams of the metadata hashtable and free them
1518 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1520 struct lttng_ht_iter iter
;
1521 struct lttng_consumer_stream
*stream
;
1528 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1530 * Ignore return value since we are currently cleaning up so any error
1533 (void) consumer_del_metadata_stream(stream
, ht
);
1537 lttng_ht_destroy(ht
);
1541 * Close all fds associated with the instance and free the context.
1543 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1547 DBG("Consumer destroying it. Closing everything.");
1553 destroy_data_stream_ht(data_ht
);
1554 destroy_metadata_stream_ht(metadata_ht
);
1556 ret
= close(ctx
->consumer_error_socket
);
1560 ret
= close(ctx
->consumer_metadata_socket
);
1564 utils_close_pipe(ctx
->consumer_channel_pipe
);
1565 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1566 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1567 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1568 utils_close_pipe(ctx
->consumer_should_quit
);
1570 unlink(ctx
->consumer_command_sock_path
);
1575 * Write the metadata stream id on the specified file descriptor.
1577 static int write_relayd_metadata_id(int fd
,
1578 struct lttng_consumer_stream
*stream
,
1579 unsigned long padding
)
1582 struct lttcomm_relayd_metadata_payload hdr
;
1584 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1585 hdr
.padding_size
= htobe32(padding
);
1586 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1587 if (ret
< sizeof(hdr
)) {
1589 * This error means that the fd's end is closed so ignore the PERROR
1590 * not to clubber the error output since this can happen in a normal
1593 if (errno
!= EPIPE
) {
1594 PERROR("write metadata stream id");
1596 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1598 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1599 * handle writting the missing part so report that as an error and
1600 * don't lie to the caller.
1605 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1606 stream
->relayd_stream_id
, padding
);
1613 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1614 * core function for writing trace buffers to either the local filesystem or
1617 * It must be called with the stream and the channel lock held.
1619 * Careful review MUST be put if any changes occur!
1621 * Returns the number of bytes written
1623 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1624 struct lttng_consumer_stream
*stream
,
1625 const struct lttng_buffer_view
*buffer
,
1626 unsigned long padding
)
1629 off_t orig_offset
= stream
->out_fd_offset
;
1630 /* Default is on the disk */
1631 int outfd
= stream
->out_fd
;
1632 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1633 unsigned int relayd_hang_up
= 0;
1634 const size_t subbuf_content_size
= buffer
->size
- padding
;
1637 /* RCU lock for the relayd pointer */
1639 assert(stream
->net_seq_idx
!= (uint64_t) -1ULL ||
1640 stream
->trace_chunk
);
1642 /* Flag that the current stream if set for network streaming. */
1643 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1644 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1645 if (relayd
== NULL
) {
1651 /* Handle stream on the relayd if the output is on the network */
1653 unsigned long netlen
= subbuf_content_size
;
1656 * Lock the control socket for the complete duration of the function
1657 * since from this point on we will use the socket.
1659 if (stream
->metadata_flag
) {
1660 /* Metadata requires the control socket. */
1661 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1662 if (stream
->reset_metadata_flag
) {
1663 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1664 stream
->relayd_stream_id
,
1665 stream
->metadata_version
);
1670 stream
->reset_metadata_flag
= 0;
1672 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1675 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1680 /* Use the returned socket. */
1683 /* Write metadata stream id before payload */
1684 if (stream
->metadata_flag
) {
1685 ret
= write_relayd_metadata_id(outfd
, stream
, padding
);
1692 write_len
= subbuf_content_size
;
1694 /* No streaming; we have to write the full padding. */
1695 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1696 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1698 ERR("Reset metadata file");
1701 stream
->reset_metadata_flag
= 0;
1705 * Check if we need to change the tracefile before writing the packet.
1707 if (stream
->chan
->tracefile_size
> 0 &&
1708 (stream
->tracefile_size_current
+ buffer
->size
) >
1709 stream
->chan
->tracefile_size
) {
1710 ret
= consumer_stream_rotate_output_files(stream
);
1714 outfd
= stream
->out_fd
;
1717 stream
->tracefile_size_current
+= buffer
->size
;
1718 write_len
= buffer
->size
;
1722 * This call guarantee that len or less is returned. It's impossible to
1723 * receive a ret value that is bigger than len.
1725 ret
= lttng_write(outfd
, buffer
->data
, write_len
);
1726 DBG("Consumer mmap write() ret %zd (len %zu)", ret
, write_len
);
1727 if (ret
< 0 || ((size_t) ret
!= write_len
)) {
1729 * Report error to caller if nothing was written else at least send the
1737 /* Socket operation failed. We consider the relayd dead */
1738 if (errno
== EPIPE
) {
1740 * This is possible if the fd is closed on the other side
1741 * (outfd) or any write problem. It can be verbose a bit for a
1742 * normal execution if for instance the relayd is stopped
1743 * abruptly. This can happen so set this to a DBG statement.
1745 DBG("Consumer mmap write detected relayd hang up");
1747 /* Unhandled error, print it and stop function right now. */
1748 PERROR("Error in write mmap (ret %zd != write_len %zu)", ret
,
1753 stream
->output_written
+= ret
;
1755 /* This call is useless on a socket so better save a syscall. */
1757 /* This won't block, but will start writeout asynchronously */
1758 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, write_len
,
1759 SYNC_FILE_RANGE_WRITE
);
1760 stream
->out_fd_offset
+= write_len
;
1761 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1766 * This is a special case that the relayd has closed its socket. Let's
1767 * cleanup the relayd object and all associated streams.
1769 if (relayd
&& relayd_hang_up
) {
1770 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1771 lttng_consumer_cleanup_relayd(relayd
);
1775 /* Unlock only if ctrl socket used */
1776 if (relayd
&& stream
->metadata_flag
) {
1777 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1785 * Splice the data from the ring buffer to the tracefile.
1787 * It must be called with the stream lock held.
1789 * Returns the number of bytes spliced.
1791 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1792 struct lttng_consumer_local_data
*ctx
,
1793 struct lttng_consumer_stream
*stream
, unsigned long len
,
1794 unsigned long padding
)
1796 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1798 off_t orig_offset
= stream
->out_fd_offset
;
1799 int fd
= stream
->wait_fd
;
1800 /* Default is on the disk */
1801 int outfd
= stream
->out_fd
;
1802 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1804 unsigned int relayd_hang_up
= 0;
1806 switch (consumer_data
.type
) {
1807 case LTTNG_CONSUMER_KERNEL
:
1809 case LTTNG_CONSUMER32_UST
:
1810 case LTTNG_CONSUMER64_UST
:
1811 /* Not supported for user space tracing */
1814 ERR("Unknown consumer_data type");
1818 /* RCU lock for the relayd pointer */
1821 /* Flag that the current stream if set for network streaming. */
1822 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1823 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1824 if (relayd
== NULL
) {
1829 splice_pipe
= stream
->splice_pipe
;
1831 /* Write metadata stream id before payload */
1833 unsigned long total_len
= len
;
1835 if (stream
->metadata_flag
) {
1837 * Lock the control socket for the complete duration of the function
1838 * since from this point on we will use the socket.
1840 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1842 if (stream
->reset_metadata_flag
) {
1843 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1844 stream
->relayd_stream_id
,
1845 stream
->metadata_version
);
1850 stream
->reset_metadata_flag
= 0;
1852 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
,
1860 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1863 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1869 /* Use the returned socket. */
1872 /* No streaming, we have to set the len with the full padding */
1875 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1876 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1878 ERR("Reset metadata file");
1881 stream
->reset_metadata_flag
= 0;
1884 * Check if we need to change the tracefile before writing the packet.
1886 if (stream
->chan
->tracefile_size
> 0 &&
1887 (stream
->tracefile_size_current
+ len
) >
1888 stream
->chan
->tracefile_size
) {
1889 ret
= consumer_stream_rotate_output_files(stream
);
1894 outfd
= stream
->out_fd
;
1897 stream
->tracefile_size_current
+= len
;
1901 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1902 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1903 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1904 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1905 DBG("splice chan to pipe, ret %zd", ret_splice
);
1906 if (ret_splice
< 0) {
1909 PERROR("Error in relay splice");
1913 /* Handle stream on the relayd if the output is on the network */
1914 if (relayd
&& stream
->metadata_flag
) {
1915 size_t metadata_payload_size
=
1916 sizeof(struct lttcomm_relayd_metadata_payload
);
1918 /* Update counter to fit the spliced data */
1919 ret_splice
+= metadata_payload_size
;
1920 len
+= metadata_payload_size
;
1922 * We do this so the return value can match the len passed as
1923 * argument to this function.
1925 written
-= metadata_payload_size
;
1928 /* Splice data out */
1929 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1930 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1931 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1933 if (ret_splice
< 0) {
1938 } else if (ret_splice
> len
) {
1940 * We don't expect this code path to be executed but you never know
1941 * so this is an extra protection agains a buggy splice().
1944 written
+= ret_splice
;
1945 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1949 /* All good, update current len and continue. */
1953 /* This call is useless on a socket so better save a syscall. */
1955 /* This won't block, but will start writeout asynchronously */
1956 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1957 SYNC_FILE_RANGE_WRITE
);
1958 stream
->out_fd_offset
+= ret_splice
;
1960 stream
->output_written
+= ret_splice
;
1961 written
+= ret_splice
;
1964 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1970 * This is a special case that the relayd has closed its socket. Let's
1971 * cleanup the relayd object and all associated streams.
1973 if (relayd
&& relayd_hang_up
) {
1974 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1975 lttng_consumer_cleanup_relayd(relayd
);
1976 /* Skip splice error so the consumer does not fail */
1981 /* send the appropriate error description to sessiond */
1984 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1987 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1990 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1995 if (relayd
&& stream
->metadata_flag
) {
1996 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2004 * Sample the snapshot positions for a specific fd
2006 * Returns 0 on success, < 0 on error
2008 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream
*stream
)
2010 switch (consumer_data
.type
) {
2011 case LTTNG_CONSUMER_KERNEL
:
2012 return lttng_kconsumer_sample_snapshot_positions(stream
);
2013 case LTTNG_CONSUMER32_UST
:
2014 case LTTNG_CONSUMER64_UST
:
2015 return lttng_ustconsumer_sample_snapshot_positions(stream
);
2017 ERR("Unknown consumer_data type");
2023 * Take a snapshot for a specific fd
2025 * Returns 0 on success, < 0 on error
2027 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
2029 switch (consumer_data
.type
) {
2030 case LTTNG_CONSUMER_KERNEL
:
2031 return lttng_kconsumer_take_snapshot(stream
);
2032 case LTTNG_CONSUMER32_UST
:
2033 case LTTNG_CONSUMER64_UST
:
2034 return lttng_ustconsumer_take_snapshot(stream
);
2036 ERR("Unknown consumer_data type");
2043 * Get the produced position
2045 * Returns 0 on success, < 0 on error
2047 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
2050 switch (consumer_data
.type
) {
2051 case LTTNG_CONSUMER_KERNEL
:
2052 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
2053 case LTTNG_CONSUMER32_UST
:
2054 case LTTNG_CONSUMER64_UST
:
2055 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2057 ERR("Unknown consumer_data type");
2064 * Get the consumed position (free-running counter position in bytes).
2066 * Returns 0 on success, < 0 on error
2068 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream
*stream
,
2071 switch (consumer_data
.type
) {
2072 case LTTNG_CONSUMER_KERNEL
:
2073 return lttng_kconsumer_get_consumed_snapshot(stream
, pos
);
2074 case LTTNG_CONSUMER32_UST
:
2075 case LTTNG_CONSUMER64_UST
:
2076 return lttng_ustconsumer_get_consumed_snapshot(stream
, pos
);
2078 ERR("Unknown consumer_data type");
2084 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2085 int sock
, struct pollfd
*consumer_sockpoll
)
2087 switch (consumer_data
.type
) {
2088 case LTTNG_CONSUMER_KERNEL
:
2089 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2090 case LTTNG_CONSUMER32_UST
:
2091 case LTTNG_CONSUMER64_UST
:
2092 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2094 ERR("Unknown consumer_data type");
2100 void lttng_consumer_close_all_metadata(void)
2102 switch (consumer_data
.type
) {
2103 case LTTNG_CONSUMER_KERNEL
:
2105 * The Kernel consumer has a different metadata scheme so we don't
2106 * close anything because the stream will be closed by the session
2110 case LTTNG_CONSUMER32_UST
:
2111 case LTTNG_CONSUMER64_UST
:
2113 * Close all metadata streams. The metadata hash table is passed and
2114 * this call iterates over it by closing all wakeup fd. This is safe
2115 * because at this point we are sure that the metadata producer is
2116 * either dead or blocked.
2118 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2121 ERR("Unknown consumer_data type");
2127 * Clean up a metadata stream and free its memory.
2129 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2130 struct lttng_ht
*ht
)
2132 struct lttng_consumer_channel
*channel
= NULL
;
2133 bool free_channel
= false;
2137 * This call should NEVER receive regular stream. It must always be
2138 * metadata stream and this is crucial for data structure synchronization.
2140 assert(stream
->metadata_flag
);
2142 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2144 pthread_mutex_lock(&consumer_data
.lock
);
2146 * Note that this assumes that a stream's channel is never changed and
2147 * that the stream's lock doesn't need to be taken to sample its
2150 channel
= stream
->chan
;
2151 pthread_mutex_lock(&channel
->lock
);
2152 pthread_mutex_lock(&stream
->lock
);
2153 if (channel
->metadata_cache
) {
2154 /* Only applicable to userspace consumers. */
2155 pthread_mutex_lock(&channel
->metadata_cache
->lock
);
2158 /* Remove any reference to that stream. */
2159 consumer_stream_delete(stream
, ht
);
2161 /* Close down everything including the relayd if one. */
2162 consumer_stream_close(stream
);
2163 /* Destroy tracer buffers of the stream. */
2164 consumer_stream_destroy_buffers(stream
);
2166 /* Atomically decrement channel refcount since other threads can use it. */
2167 if (!uatomic_sub_return(&channel
->refcount
, 1)
2168 && !uatomic_read(&channel
->nb_init_stream_left
)) {
2169 /* Go for channel deletion! */
2170 free_channel
= true;
2172 stream
->chan
= NULL
;
2175 * Nullify the stream reference so it is not used after deletion. The
2176 * channel lock MUST be acquired before being able to check for a NULL
2179 channel
->metadata_stream
= NULL
;
2181 if (channel
->metadata_cache
) {
2182 pthread_mutex_unlock(&channel
->metadata_cache
->lock
);
2184 pthread_mutex_unlock(&stream
->lock
);
2185 pthread_mutex_unlock(&channel
->lock
);
2186 pthread_mutex_unlock(&consumer_data
.lock
);
2189 consumer_del_channel(channel
);
2192 lttng_trace_chunk_put(stream
->trace_chunk
);
2193 stream
->trace_chunk
= NULL
;
2194 consumer_stream_free(stream
);
2198 * Action done with the metadata stream when adding it to the consumer internal
2199 * data structures to handle it.
2201 void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2203 struct lttng_ht
*ht
= metadata_ht
;
2204 struct lttng_ht_iter iter
;
2205 struct lttng_ht_node_u64
*node
;
2210 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2212 pthread_mutex_lock(&consumer_data
.lock
);
2213 pthread_mutex_lock(&stream
->chan
->lock
);
2214 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2215 pthread_mutex_lock(&stream
->lock
);
2218 * From here, refcounts are updated so be _careful_ when returning an error
2225 * Lookup the stream just to make sure it does not exist in our internal
2226 * state. This should NEVER happen.
2228 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2229 node
= lttng_ht_iter_get_node_u64(&iter
);
2233 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2234 * in terms of destroying the associated channel, because the action that
2235 * causes the count to become 0 also causes a stream to be added. The
2236 * channel deletion will thus be triggered by the following removal of this
2239 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2240 /* Increment refcount before decrementing nb_init_stream_left */
2242 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2245 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2247 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2248 &stream
->node_channel_id
);
2251 * Add stream to the stream_list_ht of the consumer data. No need to steal
2252 * the key since the HT does not use it and we allow to add redundant keys
2255 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2259 pthread_mutex_unlock(&stream
->lock
);
2260 pthread_mutex_unlock(&stream
->chan
->lock
);
2261 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2262 pthread_mutex_unlock(&consumer_data
.lock
);
2266 * Delete data stream that are flagged for deletion (endpoint_status).
2268 static void validate_endpoint_status_data_stream(void)
2270 struct lttng_ht_iter iter
;
2271 struct lttng_consumer_stream
*stream
;
2273 DBG("Consumer delete flagged data stream");
2276 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2277 /* Validate delete flag of the stream */
2278 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2281 /* Delete it right now */
2282 consumer_del_stream(stream
, data_ht
);
2288 * Delete metadata stream that are flagged for deletion (endpoint_status).
2290 static void validate_endpoint_status_metadata_stream(
2291 struct lttng_poll_event
*pollset
)
2293 struct lttng_ht_iter iter
;
2294 struct lttng_consumer_stream
*stream
;
2296 DBG("Consumer delete flagged metadata stream");
2301 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2302 /* Validate delete flag of the stream */
2303 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2307 * Remove from pollset so the metadata thread can continue without
2308 * blocking on a deleted stream.
2310 lttng_poll_del(pollset
, stream
->wait_fd
);
2312 /* Delete it right now */
2313 consumer_del_metadata_stream(stream
, metadata_ht
);
2319 * Thread polls on metadata file descriptor and write them on disk or on the
2322 void *consumer_thread_metadata_poll(void *data
)
2324 int ret
, i
, pollfd
, err
= -1;
2325 uint32_t revents
, nb_fd
;
2326 struct lttng_consumer_stream
*stream
= NULL
;
2327 struct lttng_ht_iter iter
;
2328 struct lttng_ht_node_u64
*node
;
2329 struct lttng_poll_event events
;
2330 struct lttng_consumer_local_data
*ctx
= data
;
2333 rcu_register_thread();
2335 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2337 if (testpoint(consumerd_thread_metadata
)) {
2338 goto error_testpoint
;
2341 health_code_update();
2343 DBG("Thread metadata poll started");
2345 /* Size is set to 1 for the consumer_metadata pipe */
2346 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2348 ERR("Poll set creation failed");
2352 ret
= lttng_poll_add(&events
,
2353 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2359 DBG("Metadata main loop started");
2363 health_code_update();
2364 health_poll_entry();
2365 DBG("Metadata poll wait");
2366 ret
= lttng_poll_wait(&events
, -1);
2367 DBG("Metadata poll return from wait with %d fd(s)",
2368 LTTNG_POLL_GETNB(&events
));
2370 DBG("Metadata event caught in thread");
2372 if (errno
== EINTR
) {
2373 ERR("Poll EINTR caught");
2376 if (LTTNG_POLL_GETNB(&events
) == 0) {
2377 err
= 0; /* All is OK */
2384 /* From here, the event is a metadata wait fd */
2385 for (i
= 0; i
< nb_fd
; i
++) {
2386 health_code_update();
2388 revents
= LTTNG_POLL_GETEV(&events
, i
);
2389 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2391 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2392 if (revents
& LPOLLIN
) {
2395 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2396 &stream
, sizeof(stream
));
2397 if (pipe_len
< sizeof(stream
)) {
2399 PERROR("read metadata stream");
2402 * Remove the pipe from the poll set and continue the loop
2403 * since their might be data to consume.
2405 lttng_poll_del(&events
,
2406 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2407 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2411 /* A NULL stream means that the state has changed. */
2412 if (stream
== NULL
) {
2413 /* Check for deleted streams. */
2414 validate_endpoint_status_metadata_stream(&events
);
2418 DBG("Adding metadata stream %d to poll set",
2421 /* Add metadata stream to the global poll events list */
2422 lttng_poll_add(&events
, stream
->wait_fd
,
2423 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2424 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2425 DBG("Metadata thread pipe hung up");
2427 * Remove the pipe from the poll set and continue the loop
2428 * since their might be data to consume.
2430 lttng_poll_del(&events
,
2431 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2432 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2435 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2439 /* Handle other stream */
2445 uint64_t tmp_id
= (uint64_t) pollfd
;
2447 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2449 node
= lttng_ht_iter_get_node_u64(&iter
);
2452 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2455 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2456 /* Get the data out of the metadata file descriptor */
2457 DBG("Metadata available on fd %d", pollfd
);
2458 assert(stream
->wait_fd
== pollfd
);
2461 health_code_update();
2463 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2465 * We don't check the return value here since if we get
2466 * a negative len, it means an error occurred thus we
2467 * simply remove it from the poll set and free the
2472 /* It's ok to have an unavailable sub-buffer */
2473 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2474 /* Clean up stream from consumer and free it. */
2475 lttng_poll_del(&events
, stream
->wait_fd
);
2476 consumer_del_metadata_stream(stream
, metadata_ht
);
2478 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2479 DBG("Metadata fd %d is hup|err.", pollfd
);
2480 if (!stream
->hangup_flush_done
2481 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2482 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2483 DBG("Attempting to flush and consume the UST buffers");
2484 lttng_ustconsumer_on_stream_hangup(stream
);
2486 /* We just flushed the stream now read it. */
2488 health_code_update();
2490 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2492 * We don't check the return value here since if we get
2493 * a negative len, it means an error occurred thus we
2494 * simply remove it from the poll set and free the
2500 lttng_poll_del(&events
, stream
->wait_fd
);
2502 * This call update the channel states, closes file descriptors
2503 * and securely free the stream.
2505 consumer_del_metadata_stream(stream
, metadata_ht
);
2507 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2511 /* Release RCU lock for the stream looked up */
2519 DBG("Metadata poll thread exiting");
2521 lttng_poll_clean(&events
);
2526 ERR("Health error occurred in %s", __func__
);
2528 health_unregister(health_consumerd
);
2529 rcu_unregister_thread();
2534 * This thread polls the fds in the set to consume the data and write
2535 * it to tracefile if necessary.
2537 void *consumer_thread_data_poll(void *data
)
2539 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2540 struct pollfd
*pollfd
= NULL
;
2541 /* local view of the streams */
2542 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2543 /* local view of consumer_data.fds_count */
2545 /* 2 for the consumer_data_pipe and wake up pipe */
2546 const int nb_pipes_fd
= 2;
2547 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2548 int nb_inactive_fd
= 0;
2549 struct lttng_consumer_local_data
*ctx
= data
;
2552 rcu_register_thread();
2554 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2556 if (testpoint(consumerd_thread_data
)) {
2557 goto error_testpoint
;
2560 health_code_update();
2562 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2563 if (local_stream
== NULL
) {
2564 PERROR("local_stream malloc");
2569 health_code_update();
2575 * the fds set has been updated, we need to update our
2576 * local array as well
2578 pthread_mutex_lock(&consumer_data
.lock
);
2579 if (consumer_data
.need_update
) {
2584 local_stream
= NULL
;
2586 /* Allocate for all fds */
2587 pollfd
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) * sizeof(struct pollfd
));
2588 if (pollfd
== NULL
) {
2589 PERROR("pollfd malloc");
2590 pthread_mutex_unlock(&consumer_data
.lock
);
2594 local_stream
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) *
2595 sizeof(struct lttng_consumer_stream
*));
2596 if (local_stream
== NULL
) {
2597 PERROR("local_stream malloc");
2598 pthread_mutex_unlock(&consumer_data
.lock
);
2601 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2602 data_ht
, &nb_inactive_fd
);
2604 ERR("Error in allocating pollfd or local_outfds");
2605 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2606 pthread_mutex_unlock(&consumer_data
.lock
);
2610 consumer_data
.need_update
= 0;
2612 pthread_mutex_unlock(&consumer_data
.lock
);
2614 /* No FDs and consumer_quit, consumer_cleanup the thread */
2615 if (nb_fd
== 0 && nb_inactive_fd
== 0 &&
2616 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2617 err
= 0; /* All is OK */
2620 /* poll on the array of fds */
2622 DBG("polling on %d fd", nb_fd
+ nb_pipes_fd
);
2623 if (testpoint(consumerd_thread_data_poll
)) {
2626 health_poll_entry();
2627 num_rdy
= poll(pollfd
, nb_fd
+ nb_pipes_fd
, -1);
2629 DBG("poll num_rdy : %d", num_rdy
);
2630 if (num_rdy
== -1) {
2632 * Restart interrupted system call.
2634 if (errno
== EINTR
) {
2637 PERROR("Poll error");
2638 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2640 } else if (num_rdy
== 0) {
2641 DBG("Polling thread timed out");
2645 if (caa_unlikely(data_consumption_paused
)) {
2646 DBG("Data consumption paused, sleeping...");
2652 * If the consumer_data_pipe triggered poll go directly to the
2653 * beginning of the loop to update the array. We want to prioritize
2654 * array update over low-priority reads.
2656 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2657 ssize_t pipe_readlen
;
2659 DBG("consumer_data_pipe wake up");
2660 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2661 &new_stream
, sizeof(new_stream
));
2662 if (pipe_readlen
< sizeof(new_stream
)) {
2663 PERROR("Consumer data pipe");
2664 /* Continue so we can at least handle the current stream(s). */
2669 * If the stream is NULL, just ignore it. It's also possible that
2670 * the sessiond poll thread changed the consumer_quit state and is
2671 * waking us up to test it.
2673 if (new_stream
== NULL
) {
2674 validate_endpoint_status_data_stream();
2678 /* Continue to update the local streams and handle prio ones */
2682 /* Handle wakeup pipe. */
2683 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2685 ssize_t pipe_readlen
;
2687 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2689 if (pipe_readlen
< 0) {
2690 PERROR("Consumer data wakeup pipe");
2692 /* We've been awakened to handle stream(s). */
2693 ctx
->has_wakeup
= 0;
2696 /* Take care of high priority channels first. */
2697 for (i
= 0; i
< nb_fd
; i
++) {
2698 health_code_update();
2700 if (local_stream
[i
] == NULL
) {
2703 if (pollfd
[i
].revents
& POLLPRI
) {
2704 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2706 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2707 /* it's ok to have an unavailable sub-buffer */
2708 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2709 /* Clean the stream and free it. */
2710 consumer_del_stream(local_stream
[i
], data_ht
);
2711 local_stream
[i
] = NULL
;
2712 } else if (len
> 0) {
2713 local_stream
[i
]->data_read
= 1;
2719 * If we read high prio channel in this loop, try again
2720 * for more high prio data.
2726 /* Take care of low priority channels. */
2727 for (i
= 0; i
< nb_fd
; i
++) {
2728 health_code_update();
2730 if (local_stream
[i
] == NULL
) {
2733 if ((pollfd
[i
].revents
& POLLIN
) ||
2734 local_stream
[i
]->hangup_flush_done
||
2735 local_stream
[i
]->has_data
) {
2736 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2737 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2738 /* it's ok to have an unavailable sub-buffer */
2739 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2740 /* Clean the stream and free it. */
2741 consumer_del_stream(local_stream
[i
], data_ht
);
2742 local_stream
[i
] = NULL
;
2743 } else if (len
> 0) {
2744 local_stream
[i
]->data_read
= 1;
2749 /* Handle hangup and errors */
2750 for (i
= 0; i
< nb_fd
; i
++) {
2751 health_code_update();
2753 if (local_stream
[i
] == NULL
) {
2756 if (!local_stream
[i
]->hangup_flush_done
2757 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2758 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2759 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2760 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2762 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2763 /* Attempt read again, for the data we just flushed. */
2764 local_stream
[i
]->data_read
= 1;
2767 * If the poll flag is HUP/ERR/NVAL and we have
2768 * read no data in this pass, we can remove the
2769 * stream from its hash table.
2771 if ((pollfd
[i
].revents
& POLLHUP
)) {
2772 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2773 if (!local_stream
[i
]->data_read
) {
2774 consumer_del_stream(local_stream
[i
], data_ht
);
2775 local_stream
[i
] = NULL
;
2778 } else if (pollfd
[i
].revents
& POLLERR
) {
2779 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2780 if (!local_stream
[i
]->data_read
) {
2781 consumer_del_stream(local_stream
[i
], data_ht
);
2782 local_stream
[i
] = NULL
;
2785 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2786 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2787 if (!local_stream
[i
]->data_read
) {
2788 consumer_del_stream(local_stream
[i
], data_ht
);
2789 local_stream
[i
] = NULL
;
2793 if (local_stream
[i
] != NULL
) {
2794 local_stream
[i
]->data_read
= 0;
2801 DBG("polling thread exiting");
2806 * Close the write side of the pipe so epoll_wait() in
2807 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2808 * read side of the pipe. If we close them both, epoll_wait strangely does
2809 * not return and could create a endless wait period if the pipe is the
2810 * only tracked fd in the poll set. The thread will take care of closing
2813 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2818 ERR("Health error occurred in %s", __func__
);
2820 health_unregister(health_consumerd
);
2822 rcu_unregister_thread();
2827 * Close wake-up end of each stream belonging to the channel. This will
2828 * allow the poll() on the stream read-side to detect when the
2829 * write-side (application) finally closes them.
2832 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2834 struct lttng_ht
*ht
;
2835 struct lttng_consumer_stream
*stream
;
2836 struct lttng_ht_iter iter
;
2838 ht
= consumer_data
.stream_per_chan_id_ht
;
2841 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2842 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2843 ht
->match_fct
, &channel
->key
,
2844 &iter
.iter
, stream
, node_channel_id
.node
) {
2846 * Protect against teardown with mutex.
2848 pthread_mutex_lock(&stream
->lock
);
2849 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2852 switch (consumer_data
.type
) {
2853 case LTTNG_CONSUMER_KERNEL
:
2855 case LTTNG_CONSUMER32_UST
:
2856 case LTTNG_CONSUMER64_UST
:
2857 if (stream
->metadata_flag
) {
2858 /* Safe and protected by the stream lock. */
2859 lttng_ustconsumer_close_metadata(stream
->chan
);
2862 * Note: a mutex is taken internally within
2863 * liblttng-ust-ctl to protect timer wakeup_fd
2864 * use from concurrent close.
2866 lttng_ustconsumer_close_stream_wakeup(stream
);
2870 ERR("Unknown consumer_data type");
2874 pthread_mutex_unlock(&stream
->lock
);
2879 static void destroy_channel_ht(struct lttng_ht
*ht
)
2881 struct lttng_ht_iter iter
;
2882 struct lttng_consumer_channel
*channel
;
2890 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2891 ret
= lttng_ht_del(ht
, &iter
);
2896 lttng_ht_destroy(ht
);
2900 * This thread polls the channel fds to detect when they are being
2901 * closed. It closes all related streams if the channel is detected as
2902 * closed. It is currently only used as a shim layer for UST because the
2903 * consumerd needs to keep the per-stream wakeup end of pipes open for
2906 void *consumer_thread_channel_poll(void *data
)
2908 int ret
, i
, pollfd
, err
= -1;
2909 uint32_t revents
, nb_fd
;
2910 struct lttng_consumer_channel
*chan
= NULL
;
2911 struct lttng_ht_iter iter
;
2912 struct lttng_ht_node_u64
*node
;
2913 struct lttng_poll_event events
;
2914 struct lttng_consumer_local_data
*ctx
= data
;
2915 struct lttng_ht
*channel_ht
;
2917 rcu_register_thread();
2919 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2921 if (testpoint(consumerd_thread_channel
)) {
2922 goto error_testpoint
;
2925 health_code_update();
2927 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2929 /* ENOMEM at this point. Better to bail out. */
2933 DBG("Thread channel poll started");
2935 /* Size is set to 1 for the consumer_channel pipe */
2936 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2938 ERR("Poll set creation failed");
2942 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2948 DBG("Channel main loop started");
2952 health_code_update();
2953 DBG("Channel poll wait");
2954 health_poll_entry();
2955 ret
= lttng_poll_wait(&events
, -1);
2956 DBG("Channel poll return from wait with %d fd(s)",
2957 LTTNG_POLL_GETNB(&events
));
2959 DBG("Channel event caught in thread");
2961 if (errno
== EINTR
) {
2962 ERR("Poll EINTR caught");
2965 if (LTTNG_POLL_GETNB(&events
) == 0) {
2966 err
= 0; /* All is OK */
2973 /* From here, the event is a channel wait fd */
2974 for (i
= 0; i
< nb_fd
; i
++) {
2975 health_code_update();
2977 revents
= LTTNG_POLL_GETEV(&events
, i
);
2978 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2980 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2981 if (revents
& LPOLLIN
) {
2982 enum consumer_channel_action action
;
2985 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2988 ERR("Error reading channel pipe");
2990 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2995 case CONSUMER_CHANNEL_ADD
:
2996 DBG("Adding channel %d to poll set",
2999 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
3002 lttng_ht_add_unique_u64(channel_ht
,
3003 &chan
->wait_fd_node
);
3005 /* Add channel to the global poll events list */
3006 lttng_poll_add(&events
, chan
->wait_fd
,
3007 LPOLLERR
| LPOLLHUP
);
3009 case CONSUMER_CHANNEL_DEL
:
3012 * This command should never be called if the channel
3013 * has streams monitored by either the data or metadata
3014 * thread. The consumer only notify this thread with a
3015 * channel del. command if it receives a destroy
3016 * channel command from the session daemon that send it
3017 * if a command prior to the GET_CHANNEL failed.
3021 chan
= consumer_find_channel(key
);
3024 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
3027 lttng_poll_del(&events
, chan
->wait_fd
);
3028 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
3029 ret
= lttng_ht_del(channel_ht
, &iter
);
3032 switch (consumer_data
.type
) {
3033 case LTTNG_CONSUMER_KERNEL
:
3035 case LTTNG_CONSUMER32_UST
:
3036 case LTTNG_CONSUMER64_UST
:
3037 health_code_update();
3038 /* Destroy streams that might have been left in the stream list. */
3039 clean_channel_stream_list(chan
);
3042 ERR("Unknown consumer_data type");
3047 * Release our own refcount. Force channel deletion even if
3048 * streams were not initialized.
3050 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
3051 consumer_del_channel(chan
);
3056 case CONSUMER_CHANNEL_QUIT
:
3058 * Remove the pipe from the poll set and continue the loop
3059 * since their might be data to consume.
3061 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3064 ERR("Unknown action");
3067 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3068 DBG("Channel thread pipe hung up");
3070 * Remove the pipe from the poll set and continue the loop
3071 * since their might be data to consume.
3073 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3076 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3080 /* Handle other stream */
3086 uint64_t tmp_id
= (uint64_t) pollfd
;
3088 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3090 node
= lttng_ht_iter_get_node_u64(&iter
);
3093 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3096 /* Check for error event */
3097 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3098 DBG("Channel fd %d is hup|err.", pollfd
);
3100 lttng_poll_del(&events
, chan
->wait_fd
);
3101 ret
= lttng_ht_del(channel_ht
, &iter
);
3105 * This will close the wait fd for each stream associated to
3106 * this channel AND monitored by the data/metadata thread thus
3107 * will be clean by the right thread.
3109 consumer_close_channel_streams(chan
);
3111 /* Release our own refcount */
3112 if (!uatomic_sub_return(&chan
->refcount
, 1)
3113 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3114 consumer_del_channel(chan
);
3117 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3122 /* Release RCU lock for the channel looked up */
3130 lttng_poll_clean(&events
);
3132 destroy_channel_ht(channel_ht
);
3135 DBG("Channel poll thread exiting");
3138 ERR("Health error occurred in %s", __func__
);
3140 health_unregister(health_consumerd
);
3141 rcu_unregister_thread();
3145 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3146 struct pollfd
*sockpoll
, int client_socket
)
3153 ret
= lttng_consumer_poll_socket(sockpoll
);
3157 DBG("Metadata connection on client_socket");
3159 /* Blocking call, waiting for transmission */
3160 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3161 if (ctx
->consumer_metadata_socket
< 0) {
3162 WARN("On accept metadata");
3173 * This thread listens on the consumerd socket and receives the file
3174 * descriptors from the session daemon.
3176 void *consumer_thread_sessiond_poll(void *data
)
3178 int sock
= -1, client_socket
, ret
, err
= -1;
3180 * structure to poll for incoming data on communication socket avoids
3181 * making blocking sockets.
3183 struct pollfd consumer_sockpoll
[2];
3184 struct lttng_consumer_local_data
*ctx
= data
;
3186 rcu_register_thread();
3188 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3190 if (testpoint(consumerd_thread_sessiond
)) {
3191 goto error_testpoint
;
3194 health_code_update();
3196 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3197 unlink(ctx
->consumer_command_sock_path
);
3198 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3199 if (client_socket
< 0) {
3200 ERR("Cannot create command socket");
3204 ret
= lttcomm_listen_unix_sock(client_socket
);
3209 DBG("Sending ready command to lttng-sessiond");
3210 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3211 /* return < 0 on error, but == 0 is not fatal */
3213 ERR("Error sending ready command to lttng-sessiond");
3217 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3218 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3219 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3220 consumer_sockpoll
[1].fd
= client_socket
;
3221 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3223 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3231 DBG("Connection on client_socket");
3233 /* Blocking call, waiting for transmission */
3234 sock
= lttcomm_accept_unix_sock(client_socket
);
3241 * Setup metadata socket which is the second socket connection on the
3242 * command unix socket.
3244 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3253 /* This socket is not useful anymore. */
3254 ret
= close(client_socket
);
3256 PERROR("close client_socket");
3260 /* update the polling structure to poll on the established socket */
3261 consumer_sockpoll
[1].fd
= sock
;
3262 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3265 health_code_update();
3267 health_poll_entry();
3268 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3277 DBG("Incoming command on sock");
3278 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3281 * This could simply be a session daemon quitting. Don't output
3284 DBG("Communication interrupted on command socket");
3288 if (CMM_LOAD_SHARED(consumer_quit
)) {
3289 DBG("consumer_thread_receive_fds received quit from signal");
3290 err
= 0; /* All is OK */
3293 DBG("Received command on sock");
3299 DBG("Consumer thread sessiond poll exiting");
3302 * Close metadata streams since the producer is the session daemon which
3305 * NOTE: for now, this only applies to the UST tracer.
3307 lttng_consumer_close_all_metadata();
3310 * when all fds have hung up, the polling thread
3313 CMM_STORE_SHARED(consumer_quit
, 1);
3316 * Notify the data poll thread to poll back again and test the
3317 * consumer_quit state that we just set so to quit gracefully.
3319 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3321 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3323 notify_health_quit_pipe(health_quit_pipe
);
3325 /* Cleaning up possibly open sockets. */
3329 PERROR("close sock sessiond poll");
3332 if (client_socket
>= 0) {
3333 ret
= close(client_socket
);
3335 PERROR("close client_socket sessiond poll");
3342 ERR("Health error occurred in %s", __func__
);
3344 health_unregister(health_consumerd
);
3346 rcu_unregister_thread();
3350 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3351 struct lttng_consumer_local_data
*ctx
,
3352 bool locked_by_caller
)
3354 ssize_t ret
, written_bytes
= 0;
3356 struct stream_subbuffer subbuffer
= {};
3358 if (!locked_by_caller
) {
3359 stream
->read_subbuffer_ops
.lock(stream
);
3362 if (stream
->read_subbuffer_ops
.on_wake_up
) {
3363 ret
= stream
->read_subbuffer_ops
.on_wake_up(stream
);
3370 * If the stream was flagged to be ready for rotation before we extract
3371 * the next packet, rotate it now.
3373 if (stream
->rotate_ready
) {
3374 DBG("Rotate stream before consuming data");
3375 ret
= lttng_consumer_rotate_stream(ctx
, stream
);
3377 ERR("Stream rotation error before consuming data");
3382 ret
= stream
->read_subbuffer_ops
.get_next_subbuffer(stream
, &subbuffer
);
3384 if (ret
== -ENODATA
) {
3392 ret
= stream
->read_subbuffer_ops
.pre_consume_subbuffer(
3393 stream
, &subbuffer
);
3395 goto error_put_subbuf
;
3398 written_bytes
= stream
->read_subbuffer_ops
.consume_subbuffer(
3399 ctx
, stream
, &subbuffer
);
3400 if (written_bytes
<= 0) {
3401 ERR("Error consuming subbuffer: (%zd)", written_bytes
);
3402 ret
= (int) written_bytes
;
3403 goto error_put_subbuf
;
3406 ret
= stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3411 if (stream
->read_subbuffer_ops
.post_consume
) {
3412 ret
= stream
->read_subbuffer_ops
.post_consume(stream
, &subbuffer
, ctx
);
3419 * After extracting the packet, we check if the stream is now ready to
3420 * be rotated and perform the action immediately.
3422 * Don't overwrite `ret` as callers expect the number of bytes
3423 * consumed to be returned on success.
3425 rotation_ret
= lttng_consumer_stream_is_rotate_ready(stream
);
3426 if (rotation_ret
== 1) {
3427 rotation_ret
= lttng_consumer_rotate_stream(ctx
, stream
);
3428 if (rotation_ret
< 0) {
3430 ERR("Stream rotation error after consuming data");
3433 } else if (rotation_ret
< 0) {
3435 ERR("Failed to check if stream was ready to rotate after consuming data");
3440 if (stream
->read_subbuffer_ops
.on_sleep
) {
3441 stream
->read_subbuffer_ops
.on_sleep(stream
, ctx
);
3444 ret
= written_bytes
;
3446 if (!locked_by_caller
) {
3447 stream
->read_subbuffer_ops
.unlock(stream
);
3452 (void) stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3456 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3458 switch (consumer_data
.type
) {
3459 case LTTNG_CONSUMER_KERNEL
:
3460 return lttng_kconsumer_on_recv_stream(stream
);
3461 case LTTNG_CONSUMER32_UST
:
3462 case LTTNG_CONSUMER64_UST
:
3463 return lttng_ustconsumer_on_recv_stream(stream
);
3465 ERR("Unknown consumer_data type");
3472 * Allocate and set consumer data hash tables.
3474 int lttng_consumer_init(void)
3476 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3477 if (!consumer_data
.channel_ht
) {
3481 consumer_data
.channels_by_session_id_ht
=
3482 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3483 if (!consumer_data
.channels_by_session_id_ht
) {
3487 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3488 if (!consumer_data
.relayd_ht
) {
3492 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3493 if (!consumer_data
.stream_list_ht
) {
3497 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3498 if (!consumer_data
.stream_per_chan_id_ht
) {
3502 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3507 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3512 consumer_data
.chunk_registry
= lttng_trace_chunk_registry_create();
3513 if (!consumer_data
.chunk_registry
) {
3524 * Process the ADD_RELAYD command receive by a consumer.
3526 * This will create a relayd socket pair and add it to the relayd hash table.
3527 * The caller MUST acquire a RCU read side lock before calling it.
3529 void consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3530 struct lttng_consumer_local_data
*ctx
, int sock
,
3531 struct pollfd
*consumer_sockpoll
,
3532 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3533 uint64_t relayd_session_id
)
3535 int fd
= -1, ret
= -1, relayd_created
= 0;
3536 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3537 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3540 assert(relayd_sock
);
3542 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3544 /* Get relayd reference if exists. */
3545 relayd
= consumer_find_relayd(net_seq_idx
);
3546 if (relayd
== NULL
) {
3547 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3548 /* Not found. Allocate one. */
3549 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3550 if (relayd
== NULL
) {
3551 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3554 relayd
->sessiond_session_id
= sessiond_id
;
3559 * This code path MUST continue to the consumer send status message to
3560 * we can notify the session daemon and continue our work without
3561 * killing everything.
3565 * relayd key should never be found for control socket.
3567 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3570 /* First send a status message before receiving the fds. */
3571 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3573 /* Somehow, the session daemon is not responding anymore. */
3574 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3575 goto error_nosignal
;
3578 /* Poll on consumer socket. */
3579 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3581 /* Needing to exit in the middle of a command: error. */
3582 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3583 goto error_nosignal
;
3586 /* Get relayd socket from session daemon */
3587 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3588 if (ret
!= sizeof(fd
)) {
3589 fd
= -1; /* Just in case it gets set with an invalid value. */
3592 * Failing to receive FDs might indicate a major problem such as
3593 * reaching a fd limit during the receive where the kernel returns a
3594 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3595 * don't take any chances and stop everything.
3597 * XXX: Feature request #558 will fix that and avoid this possible
3598 * issue when reaching the fd limit.
3600 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3601 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3605 /* Copy socket information and received FD */
3606 switch (sock_type
) {
3607 case LTTNG_STREAM_CONTROL
:
3608 /* Copy received lttcomm socket */
3609 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3610 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3611 /* Handle create_sock error. */
3613 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3617 * Close the socket created internally by
3618 * lttcomm_create_sock, so we can replace it by the one
3619 * received from sessiond.
3621 if (close(relayd
->control_sock
.sock
.fd
)) {
3625 /* Assign new file descriptor */
3626 relayd
->control_sock
.sock
.fd
= fd
;
3627 /* Assign version values. */
3628 relayd
->control_sock
.major
= relayd_sock
->major
;
3629 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3631 relayd
->relayd_session_id
= relayd_session_id
;
3634 case LTTNG_STREAM_DATA
:
3635 /* Copy received lttcomm socket */
3636 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3637 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3638 /* Handle create_sock error. */
3640 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3644 * Close the socket created internally by
3645 * lttcomm_create_sock, so we can replace it by the one
3646 * received from sessiond.
3648 if (close(relayd
->data_sock
.sock
.fd
)) {
3652 /* Assign new file descriptor */
3653 relayd
->data_sock
.sock
.fd
= fd
;
3654 /* Assign version values. */
3655 relayd
->data_sock
.major
= relayd_sock
->major
;
3656 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3659 ERR("Unknown relayd socket type (%d)", sock_type
);
3660 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3664 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3665 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3666 relayd
->net_seq_idx
, fd
);
3668 * We gave the ownership of the fd to the relayd structure. Set the
3669 * fd to -1 so we don't call close() on it in the error path below.
3673 /* We successfully added the socket. Send status back. */
3674 ret
= consumer_send_status_msg(sock
, ret_code
);
3676 /* Somehow, the session daemon is not responding anymore. */
3677 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3678 goto error_nosignal
;
3682 * Add relayd socket pair to consumer data hashtable. If object already
3683 * exists or on error, the function gracefully returns.
3692 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3693 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3697 /* Close received socket if valid. */
3700 PERROR("close received socket");
3704 if (relayd_created
) {
3710 * Search for a relayd associated to the session id and return the reference.
3712 * A rcu read side lock MUST be acquire before calling this function and locked
3713 * until the relayd object is no longer necessary.
3715 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3717 struct lttng_ht_iter iter
;
3718 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3720 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3721 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3724 * Check by sessiond id which is unique here where the relayd session
3725 * id might not be when having multiple relayd.
3727 if (relayd
->sessiond_session_id
== id
) {
3728 /* Found the relayd. There can be only one per id. */
3740 * Check if for a given session id there is still data needed to be extract
3743 * Return 1 if data is pending or else 0 meaning ready to be read.
3745 int consumer_data_pending(uint64_t id
)
3748 struct lttng_ht_iter iter
;
3749 struct lttng_ht
*ht
;
3750 struct lttng_consumer_stream
*stream
;
3751 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3752 int (*data_pending
)(struct lttng_consumer_stream
*);
3754 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3757 pthread_mutex_lock(&consumer_data
.lock
);
3759 switch (consumer_data
.type
) {
3760 case LTTNG_CONSUMER_KERNEL
:
3761 data_pending
= lttng_kconsumer_data_pending
;
3763 case LTTNG_CONSUMER32_UST
:
3764 case LTTNG_CONSUMER64_UST
:
3765 data_pending
= lttng_ustconsumer_data_pending
;
3768 ERR("Unknown consumer data type");
3772 /* Ease our life a bit */
3773 ht
= consumer_data
.stream_list_ht
;
3775 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3776 ht
->hash_fct(&id
, lttng_ht_seed
),
3778 &iter
.iter
, stream
, node_session_id
.node
) {
3779 pthread_mutex_lock(&stream
->lock
);
3782 * A removed node from the hash table indicates that the stream has
3783 * been deleted thus having a guarantee that the buffers are closed
3784 * on the consumer side. However, data can still be transmitted
3785 * over the network so don't skip the relayd check.
3787 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3789 /* Check the stream if there is data in the buffers. */
3790 ret
= data_pending(stream
);
3792 pthread_mutex_unlock(&stream
->lock
);
3797 pthread_mutex_unlock(&stream
->lock
);
3800 relayd
= find_relayd_by_session_id(id
);
3802 unsigned int is_data_inflight
= 0;
3804 /* Send init command for data pending. */
3805 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3806 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3807 relayd
->relayd_session_id
);
3809 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3810 /* Communication error thus the relayd so no data pending. */
3811 goto data_not_pending
;
3814 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3815 ht
->hash_fct(&id
, lttng_ht_seed
),
3817 &iter
.iter
, stream
, node_session_id
.node
) {
3818 if (stream
->metadata_flag
) {
3819 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3820 stream
->relayd_stream_id
);
3822 ret
= relayd_data_pending(&relayd
->control_sock
,
3823 stream
->relayd_stream_id
,
3824 stream
->next_net_seq_num
- 1);
3828 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3830 } else if (ret
< 0) {
3831 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3832 lttng_consumer_cleanup_relayd(relayd
);
3833 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3834 goto data_not_pending
;
3838 /* Send end command for data pending. */
3839 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3840 relayd
->relayd_session_id
, &is_data_inflight
);
3841 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3843 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3844 lttng_consumer_cleanup_relayd(relayd
);
3845 goto data_not_pending
;
3847 if (is_data_inflight
) {
3853 * Finding _no_ node in the hash table and no inflight data means that the
3854 * stream(s) have been removed thus data is guaranteed to be available for
3855 * analysis from the trace files.
3859 /* Data is available to be read by a viewer. */
3860 pthread_mutex_unlock(&consumer_data
.lock
);
3865 /* Data is still being extracted from buffers. */
3866 pthread_mutex_unlock(&consumer_data
.lock
);
3872 * Send a ret code status message to the sessiond daemon.
3874 * Return the sendmsg() return value.
3876 int consumer_send_status_msg(int sock
, int ret_code
)
3878 struct lttcomm_consumer_status_msg msg
;
3880 memset(&msg
, 0, sizeof(msg
));
3881 msg
.ret_code
= ret_code
;
3883 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3887 * Send a channel status message to the sessiond daemon.
3889 * Return the sendmsg() return value.
3891 int consumer_send_status_channel(int sock
,
3892 struct lttng_consumer_channel
*channel
)
3894 struct lttcomm_consumer_status_channel msg
;
3898 memset(&msg
, 0, sizeof(msg
));
3900 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3902 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3903 msg
.key
= channel
->key
;
3904 msg
.stream_count
= channel
->streams
.count
;
3907 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3910 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3911 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3912 uint64_t max_sb_size
)
3914 unsigned long start_pos
;
3916 if (!nb_packets_per_stream
) {
3917 return consumed_pos
; /* Grab everything */
3919 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3920 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3921 if ((long) (start_pos
- consumed_pos
) < 0) {
3922 return consumed_pos
; /* Grab everything */
3928 int consumer_flush_buffer(struct lttng_consumer_stream
*stream
, int producer_active
)
3932 switch (consumer_data
.type
) {
3933 case LTTNG_CONSUMER_KERNEL
:
3934 if (producer_active
) {
3935 ret
= kernctl_buffer_flush(stream
->wait_fd
);
3937 ERR("Failed to flush kernel stream");
3941 ret
= kernctl_buffer_flush_empty(stream
->wait_fd
);
3943 ERR("Failed to flush kernel stream");
3948 case LTTNG_CONSUMER32_UST
:
3949 case LTTNG_CONSUMER64_UST
:
3950 lttng_ustconsumer_flush_buffer(stream
, producer_active
);
3953 ERR("Unknown consumer_data type");
3962 * Sample the rotate position for all the streams of a channel. If a stream
3963 * is already at the rotate position (produced == consumed), we flag it as
3964 * ready for rotation. The rotation of ready streams occurs after we have
3965 * replied to the session daemon that we have finished sampling the positions.
3966 * Must be called with RCU read-side lock held to ensure existence of channel.
3968 * Returns 0 on success, < 0 on error
3970 int lttng_consumer_rotate_channel(struct lttng_consumer_channel
*channel
,
3971 uint64_t key
, uint64_t relayd_id
, uint32_t metadata
,
3972 struct lttng_consumer_local_data
*ctx
)
3975 struct lttng_consumer_stream
*stream
;
3976 struct lttng_ht_iter iter
;
3977 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
3978 struct lttng_dynamic_array stream_rotation_positions
;
3979 uint64_t next_chunk_id
, stream_count
= 0;
3980 enum lttng_trace_chunk_status chunk_status
;
3981 const bool is_local_trace
= relayd_id
== -1ULL;
3982 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3983 bool rotating_to_new_chunk
= true;
3985 DBG("Consumer sample rotate position for channel %" PRIu64
, key
);
3987 lttng_dynamic_array_init(&stream_rotation_positions
,
3988 sizeof(struct relayd_stream_rotation_position
), NULL
);
3992 pthread_mutex_lock(&channel
->lock
);
3993 assert(channel
->trace_chunk
);
3994 chunk_status
= lttng_trace_chunk_get_id(channel
->trace_chunk
,
3996 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
3998 goto end_unlock_channel
;
4001 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4002 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4003 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4004 stream
, node_channel_id
.node
) {
4005 unsigned long produced_pos
= 0, consumed_pos
= 0;
4007 health_code_update();
4010 * Lock stream because we are about to change its state.
4012 pthread_mutex_lock(&stream
->lock
);
4014 if (stream
->trace_chunk
== stream
->chan
->trace_chunk
) {
4015 rotating_to_new_chunk
= false;
4019 * Do not flush an empty packet when rotating from a NULL trace
4020 * chunk. The stream has no means to output data, and the prior
4021 * rotation which rotated to NULL performed that side-effect already.
4023 if (stream
->trace_chunk
) {
4025 * For metadata stream, do an active flush, which does not
4026 * produce empty packets. For data streams, empty-flush;
4027 * ensures we have at least one packet in each stream per trace
4028 * chunk, even if no data was produced.
4030 ret
= consumer_flush_buffer(stream
, stream
->metadata_flag
? 1 : 0);
4032 ERR("Failed to flush stream %" PRIu64
" during channel rotation",
4034 goto end_unlock_stream
;
4038 ret
= lttng_consumer_take_snapshot(stream
);
4039 if (ret
< 0 && ret
!= -ENODATA
&& ret
!= -EAGAIN
) {
4040 ERR("Failed to sample snapshot position during channel rotation");
4041 goto end_unlock_stream
;
4044 ret
= lttng_consumer_get_produced_snapshot(stream
,
4047 ERR("Failed to sample produced position during channel rotation");
4048 goto end_unlock_stream
;
4051 ret
= lttng_consumer_get_consumed_snapshot(stream
,
4054 ERR("Failed to sample consumed position during channel rotation");
4055 goto end_unlock_stream
;
4059 * Align produced position on the start-of-packet boundary of the first
4060 * packet going into the next trace chunk.
4062 produced_pos
= ALIGN_FLOOR(produced_pos
, stream
->max_sb_size
);
4063 if (consumed_pos
== produced_pos
) {
4064 stream
->rotate_ready
= true;
4067 * The rotation position is based on the packet_seq_num of the
4068 * packet following the last packet that was consumed for this
4069 * stream, incremented by the offset between produced and
4070 * consumed positions. This rotation position is a lower bound
4071 * (inclusive) at which the next trace chunk starts. Since it
4072 * is a lower bound, it is OK if the packet_seq_num does not
4073 * correspond exactly to the same packet identified by the
4074 * consumed_pos, which can happen in overwrite mode.
4076 if (stream
->sequence_number_unavailable
) {
4078 * Rotation should never be performed on a session which
4079 * interacts with a pre-2.8 lttng-modules, which does
4080 * not implement packet sequence number.
4082 ERR("Failure to rotate stream %" PRIu64
": sequence number unavailable",
4085 goto end_unlock_stream
;
4087 stream
->rotate_position
= stream
->last_sequence_number
+ 1 +
4088 ((produced_pos
- consumed_pos
) / stream
->max_sb_size
);
4090 if (!is_local_trace
) {
4092 * The relay daemon control protocol expects a rotation
4093 * position as "the sequence number of the first packet
4094 * _after_ the current trace chunk".
4096 const struct relayd_stream_rotation_position position
= {
4097 .stream_id
= stream
->relayd_stream_id
,
4098 .rotate_at_seq_num
= stream
->rotate_position
,
4101 ret
= lttng_dynamic_array_add_element(
4102 &stream_rotation_positions
,
4105 ERR("Failed to allocate stream rotation position");
4106 goto end_unlock_stream
;
4110 pthread_mutex_unlock(&stream
->lock
);
4113 pthread_mutex_unlock(&channel
->lock
);
4115 if (is_local_trace
) {
4120 relayd
= consumer_find_relayd(relayd_id
);
4122 ERR("Failed to find relayd %" PRIu64
, relayd_id
);
4127 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4128 ret
= relayd_rotate_streams(&relayd
->control_sock
, stream_count
,
4129 rotating_to_new_chunk
? &next_chunk_id
: NULL
,
4130 (const struct relayd_stream_rotation_position
*)
4131 stream_rotation_positions
.buffer
.data
);
4132 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4134 ERR("Relayd rotate stream failed. Cleaning up relayd %" PRIu64
,
4135 relayd
->net_seq_idx
);
4136 lttng_consumer_cleanup_relayd(relayd
);
4144 pthread_mutex_unlock(&stream
->lock
);
4146 pthread_mutex_unlock(&channel
->lock
);
4149 lttng_dynamic_array_reset(&stream_rotation_positions
);
4154 * Check if a stream is ready to be rotated after extracting it.
4156 * Return 1 if it is ready for rotation, 0 if it is not, a negative value on
4157 * error. Stream lock must be held.
4159 int lttng_consumer_stream_is_rotate_ready(struct lttng_consumer_stream
*stream
)
4161 if (stream
->rotate_ready
) {
4166 * If packet seq num is unavailable, it means we are interacting
4167 * with a pre-2.8 lttng-modules which does not implement the
4168 * sequence number. Rotation should never be used by sessiond in this
4171 if (stream
->sequence_number_unavailable
) {
4172 ERR("Internal error: rotation used on stream %" PRIu64
4173 " with unavailable sequence number",
4178 if (stream
->rotate_position
== -1ULL ||
4179 stream
->last_sequence_number
== -1ULL) {
4184 * Rotate position not reached yet. The stream rotate position is
4185 * the position of the next packet belonging to the next trace chunk,
4186 * but consumerd considers rotation ready when reaching the last
4187 * packet of the current chunk, hence the "rotate_position - 1".
4189 if (stream
->last_sequence_number
>= stream
->rotate_position
- 1) {
4197 * Reset the state for a stream after a rotation occurred.
4199 void lttng_consumer_reset_stream_rotate_state(struct lttng_consumer_stream
*stream
)
4201 stream
->rotate_position
= -1ULL;
4202 stream
->rotate_ready
= false;
4206 * Perform the rotation a local stream file.
4209 int rotate_local_stream(struct lttng_consumer_local_data
*ctx
,
4210 struct lttng_consumer_stream
*stream
)
4214 DBG("Rotate local stream: stream key %" PRIu64
", channel key %" PRIu64
,
4217 stream
->tracefile_size_current
= 0;
4218 stream
->tracefile_count_current
= 0;
4220 if (stream
->out_fd
>= 0) {
4221 ret
= close(stream
->out_fd
);
4223 PERROR("Failed to close stream out_fd of channel \"%s\"",
4224 stream
->chan
->name
);
4226 stream
->out_fd
= -1;
4229 if (stream
->index_file
) {
4230 lttng_index_file_put(stream
->index_file
);
4231 stream
->index_file
= NULL
;
4234 if (!stream
->trace_chunk
) {
4238 ret
= consumer_stream_create_output_files(stream
, true);
4244 * Performs the stream rotation for the rotate session feature if needed.
4245 * It must be called with the channel and stream locks held.
4247 * Return 0 on success, a negative number of error.
4249 int lttng_consumer_rotate_stream(struct lttng_consumer_local_data
*ctx
,
4250 struct lttng_consumer_stream
*stream
)
4254 DBG("Consumer rotate stream %" PRIu64
, stream
->key
);
4257 * Update the stream's 'current' chunk to the session's (channel)
4258 * now-current chunk.
4260 lttng_trace_chunk_put(stream
->trace_chunk
);
4261 if (stream
->chan
->trace_chunk
== stream
->trace_chunk
) {
4263 * A channel can be rotated and not have a "next" chunk
4264 * to transition to. In that case, the channel's "current chunk"
4265 * has not been closed yet, but it has not been updated to
4266 * a "next" trace chunk either. Hence, the stream, like its
4267 * parent channel, becomes part of no chunk and can't output
4268 * anything until a new trace chunk is created.
4270 stream
->trace_chunk
= NULL
;
4271 } else if (stream
->chan
->trace_chunk
&&
4272 !lttng_trace_chunk_get(stream
->chan
->trace_chunk
)) {
4273 ERR("Failed to acquire a reference to channel's trace chunk during stream rotation");
4278 * Update the stream's trace chunk to its parent channel's
4279 * current trace chunk.
4281 stream
->trace_chunk
= stream
->chan
->trace_chunk
;
4284 if (stream
->net_seq_idx
== (uint64_t) -1ULL) {
4285 ret
= rotate_local_stream(ctx
, stream
);
4287 ERR("Failed to rotate stream, ret = %i", ret
);
4292 if (stream
->metadata_flag
&& stream
->trace_chunk
) {
4294 * If the stream has transitioned to a new trace
4295 * chunk, the metadata should be re-dumped to the
4298 * However, it is possible for a stream to transition to
4299 * a "no-chunk" state. This can happen if a rotation
4300 * occurs on an inactive session. In such cases, the metadata
4301 * regeneration will happen when the next trace chunk is
4304 ret
= consumer_metadata_stream_dump(stream
);
4309 lttng_consumer_reset_stream_rotate_state(stream
);
4318 * Rotate all the ready streams now.
4320 * This is especially important for low throughput streams that have already
4321 * been consumed, we cannot wait for their next packet to perform the
4323 * Need to be called with RCU read-side lock held to ensure existence of
4326 * Returns 0 on success, < 0 on error
4328 int lttng_consumer_rotate_ready_streams(struct lttng_consumer_channel
*channel
,
4329 uint64_t key
, struct lttng_consumer_local_data
*ctx
)
4332 struct lttng_consumer_stream
*stream
;
4333 struct lttng_ht_iter iter
;
4334 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
4338 DBG("Consumer rotate ready streams in channel %" PRIu64
, key
);
4340 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4341 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4342 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4343 stream
, node_channel_id
.node
) {
4344 health_code_update();
4346 pthread_mutex_lock(&stream
->chan
->lock
);
4347 pthread_mutex_lock(&stream
->lock
);
4349 if (!stream
->rotate_ready
) {
4350 pthread_mutex_unlock(&stream
->lock
);
4351 pthread_mutex_unlock(&stream
->chan
->lock
);
4354 DBG("Consumer rotate ready stream %" PRIu64
, stream
->key
);
4356 ret
= lttng_consumer_rotate_stream(ctx
, stream
);
4357 pthread_mutex_unlock(&stream
->lock
);
4358 pthread_mutex_unlock(&stream
->chan
->lock
);
4371 enum lttcomm_return_code
lttng_consumer_init_command(
4372 struct lttng_consumer_local_data
*ctx
,
4373 const lttng_uuid sessiond_uuid
)
4375 enum lttcomm_return_code ret
;
4376 char uuid_str
[UUID_STR_LEN
];
4378 if (ctx
->sessiond_uuid
.is_set
) {
4379 ret
= LTTCOMM_CONSUMERD_ALREADY_SET
;
4383 ctx
->sessiond_uuid
.is_set
= true;
4384 memcpy(ctx
->sessiond_uuid
.value
, sessiond_uuid
, sizeof(lttng_uuid
));
4385 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4386 lttng_uuid_to_str(sessiond_uuid
, uuid_str
);
4387 DBG("Received session daemon UUID: %s", uuid_str
);
4392 enum lttcomm_return_code
lttng_consumer_create_trace_chunk(
4393 const uint64_t *relayd_id
, uint64_t session_id
,
4395 time_t chunk_creation_timestamp
,
4396 const char *chunk_override_name
,
4397 const struct lttng_credentials
*credentials
,
4398 struct lttng_directory_handle
*chunk_directory_handle
)
4401 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4402 struct lttng_trace_chunk
*created_chunk
= NULL
, *published_chunk
= NULL
;
4403 enum lttng_trace_chunk_status chunk_status
;
4404 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4405 char creation_timestamp_buffer
[ISO8601_STR_LEN
];
4406 const char *relayd_id_str
= "(none)";
4407 const char *creation_timestamp_str
;
4408 struct lttng_ht_iter iter
;
4409 struct lttng_consumer_channel
*channel
;
4412 /* Only used for logging purposes. */
4413 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4414 "%" PRIu64
, *relayd_id
);
4415 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4416 relayd_id_str
= relayd_id_buffer
;
4418 relayd_id_str
= "(formatting error)";
4422 /* Local protocol error. */
4423 assert(chunk_creation_timestamp
);
4424 ret
= time_to_iso8601_str(chunk_creation_timestamp
,
4425 creation_timestamp_buffer
,
4426 sizeof(creation_timestamp_buffer
));
4427 creation_timestamp_str
= !ret
? creation_timestamp_buffer
:
4428 "(formatting error)";
4430 DBG("Consumer create trace chunk command: relay_id = %s"
4431 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4432 ", chunk_override_name = %s"
4433 ", chunk_creation_timestamp = %s",
4434 relayd_id_str
, session_id
, chunk_id
,
4435 chunk_override_name
? : "(none)",
4436 creation_timestamp_str
);
4439 * The trace chunk registry, as used by the consumer daemon, implicitly
4440 * owns the trace chunks. This is only needed in the consumer since
4441 * the consumer has no notion of a session beyond session IDs being
4442 * used to identify other objects.
4444 * The lttng_trace_chunk_registry_publish() call below provides a
4445 * reference which is not released; it implicitly becomes the session
4446 * daemon's reference to the chunk in the consumer daemon.
4448 * The lifetime of trace chunks in the consumer daemon is managed by
4449 * the session daemon through the LTTNG_CONSUMER_CREATE_TRACE_CHUNK
4450 * and LTTNG_CONSUMER_DESTROY_TRACE_CHUNK commands.
4452 created_chunk
= lttng_trace_chunk_create(chunk_id
,
4453 chunk_creation_timestamp
);
4454 if (!created_chunk
) {
4455 ERR("Failed to create trace chunk");
4456 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4460 if (chunk_override_name
) {
4461 chunk_status
= lttng_trace_chunk_override_name(created_chunk
,
4462 chunk_override_name
);
4463 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4464 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4469 if (chunk_directory_handle
) {
4470 chunk_status
= lttng_trace_chunk_set_credentials(created_chunk
,
4472 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4473 ERR("Failed to set trace chunk credentials");
4474 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4478 * The consumer daemon has no ownership of the chunk output
4481 chunk_status
= lttng_trace_chunk_set_as_user(created_chunk
,
4482 chunk_directory_handle
);
4483 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4484 ERR("Failed to set trace chunk's directory handle");
4485 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4490 published_chunk
= lttng_trace_chunk_registry_publish_chunk(
4491 consumer_data
.chunk_registry
, session_id
,
4493 lttng_trace_chunk_put(created_chunk
);
4494 created_chunk
= NULL
;
4495 if (!published_chunk
) {
4496 ERR("Failed to publish trace chunk");
4497 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4502 cds_lfht_for_each_entry_duplicate(consumer_data
.channels_by_session_id_ht
->ht
,
4503 consumer_data
.channels_by_session_id_ht
->hash_fct(
4504 &session_id
, lttng_ht_seed
),
4505 consumer_data
.channels_by_session_id_ht
->match_fct
,
4506 &session_id
, &iter
.iter
, channel
,
4507 channels_by_session_id_ht_node
.node
) {
4508 ret
= lttng_consumer_channel_set_trace_chunk(channel
,
4512 * Roll-back the creation of this chunk.
4514 * This is important since the session daemon will
4515 * assume that the creation of this chunk failed and
4516 * will never ask for it to be closed, resulting
4517 * in a leak and an inconsistent state for some
4520 enum lttcomm_return_code close_ret
;
4521 char path
[LTTNG_PATH_MAX
];
4523 DBG("Failed to set new trace chunk on existing channels, rolling back");
4524 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4525 session_id
, chunk_id
,
4526 chunk_creation_timestamp
, NULL
,
4528 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4529 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4530 session_id
, chunk_id
);
4533 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4539 struct consumer_relayd_sock_pair
*relayd
;
4541 relayd
= consumer_find_relayd(*relayd_id
);
4543 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4544 ret
= relayd_create_trace_chunk(
4545 &relayd
->control_sock
, published_chunk
);
4546 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4548 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
, *relayd_id
);
4551 if (!relayd
|| ret
) {
4552 enum lttcomm_return_code close_ret
;
4553 char path
[LTTNG_PATH_MAX
];
4555 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4558 chunk_creation_timestamp
,
4560 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4561 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4566 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4573 /* Release the reference returned by the "publish" operation. */
4574 lttng_trace_chunk_put(published_chunk
);
4575 lttng_trace_chunk_put(created_chunk
);
4579 enum lttcomm_return_code
lttng_consumer_close_trace_chunk(
4580 const uint64_t *relayd_id
, uint64_t session_id
,
4581 uint64_t chunk_id
, time_t chunk_close_timestamp
,
4582 const enum lttng_trace_chunk_command_type
*close_command
,
4585 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4586 struct lttng_trace_chunk
*chunk
;
4587 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4588 const char *relayd_id_str
= "(none)";
4589 const char *close_command_name
= "none";
4590 struct lttng_ht_iter iter
;
4591 struct lttng_consumer_channel
*channel
;
4592 enum lttng_trace_chunk_status chunk_status
;
4597 /* Only used for logging purposes. */
4598 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4599 "%" PRIu64
, *relayd_id
);
4600 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4601 relayd_id_str
= relayd_id_buffer
;
4603 relayd_id_str
= "(formatting error)";
4606 if (close_command
) {
4607 close_command_name
= lttng_trace_chunk_command_type_get_name(
4611 DBG("Consumer close trace chunk command: relayd_id = %s"
4612 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4613 ", close command = %s",
4614 relayd_id_str
, session_id
, chunk_id
,
4615 close_command_name
);
4617 chunk
= lttng_trace_chunk_registry_find_chunk(
4618 consumer_data
.chunk_registry
, session_id
, chunk_id
);
4620 ERR("Failed to find chunk: session_id = %" PRIu64
4621 ", chunk_id = %" PRIu64
,
4622 session_id
, chunk_id
);
4623 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4627 chunk_status
= lttng_trace_chunk_set_close_timestamp(chunk
,
4628 chunk_close_timestamp
);
4629 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4630 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4634 if (close_command
) {
4635 chunk_status
= lttng_trace_chunk_set_close_command(
4636 chunk
, *close_command
);
4637 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4638 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4644 * chunk is now invalid to access as we no longer hold a reference to
4645 * it; it is only kept around to compare it (by address) to the
4646 * current chunk found in the session's channels.
4649 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
,
4650 channel
, node
.node
) {
4654 * Only change the channel's chunk to NULL if it still
4655 * references the chunk being closed. The channel may
4656 * reference a newer channel in the case of a session
4657 * rotation. When a session rotation occurs, the "next"
4658 * chunk is created before the "current" chunk is closed.
4660 if (channel
->trace_chunk
!= chunk
) {
4663 ret
= lttng_consumer_channel_set_trace_chunk(channel
, NULL
);
4666 * Attempt to close the chunk on as many channels as
4669 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4675 struct consumer_relayd_sock_pair
*relayd
;
4677 relayd
= consumer_find_relayd(*relayd_id
);
4679 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4680 ret
= relayd_close_trace_chunk(
4681 &relayd
->control_sock
, chunk
,
4683 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4685 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
,
4689 if (!relayd
|| ret
) {
4690 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4698 * Release the reference returned by the "find" operation and
4699 * the session daemon's implicit reference to the chunk.
4701 lttng_trace_chunk_put(chunk
);
4702 lttng_trace_chunk_put(chunk
);
4707 enum lttcomm_return_code
lttng_consumer_trace_chunk_exists(
4708 const uint64_t *relayd_id
, uint64_t session_id
,
4712 enum lttcomm_return_code ret_code
;
4713 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4714 const char *relayd_id_str
= "(none)";
4715 const bool is_local_trace
= !relayd_id
;
4716 struct consumer_relayd_sock_pair
*relayd
= NULL
;
4717 bool chunk_exists_local
, chunk_exists_remote
;
4722 /* Only used for logging purposes. */
4723 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4724 "%" PRIu64
, *relayd_id
);
4725 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4726 relayd_id_str
= relayd_id_buffer
;
4728 relayd_id_str
= "(formatting error)";
4732 DBG("Consumer trace chunk exists command: relayd_id = %s"
4733 ", chunk_id = %" PRIu64
, relayd_id_str
,
4735 ret
= lttng_trace_chunk_registry_chunk_exists(
4736 consumer_data
.chunk_registry
, session_id
,
4737 chunk_id
, &chunk_exists_local
);
4739 /* Internal error. */
4740 ERR("Failed to query the existence of a trace chunk");
4741 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
4744 DBG("Trace chunk %s locally",
4745 chunk_exists_local
? "exists" : "does not exist");
4746 if (chunk_exists_local
) {
4747 ret_code
= LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_LOCAL
;
4749 } else if (is_local_trace
) {
4750 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4755 relayd
= consumer_find_relayd(*relayd_id
);
4757 ERR("Failed to find relayd %" PRIu64
, *relayd_id
);
4758 ret_code
= LTTCOMM_CONSUMERD_INVALID_PARAMETERS
;
4759 goto end_rcu_unlock
;
4761 DBG("Looking up existence of trace chunk on relay daemon");
4762 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4763 ret
= relayd_trace_chunk_exists(&relayd
->control_sock
, chunk_id
,
4764 &chunk_exists_remote
);
4765 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4767 ERR("Failed to look-up the existence of trace chunk on relay daemon");
4768 ret_code
= LTTCOMM_CONSUMERD_RELAYD_FAIL
;
4769 goto end_rcu_unlock
;
4772 ret_code
= chunk_exists_remote
?
4773 LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_REMOTE
:
4774 LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4775 DBG("Trace chunk %s on relay daemon",
4776 chunk_exists_remote
? "exists" : "does not exist");